diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_0/index_tensor([7446, 7446])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_0/index_tensor([7446, 7446])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1ff7ec4ec67a8756d1eb49623bc2831b4ee96f1a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_0/index_tensor([7446, 7446])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, deliberate, low-angle tracking shot slides from left to right across a meticulously displayed custom chopper motorcycle on a sun-drenched lawn at an outdoor motorcycle show. The camera first captures the tall, spoked front wheel, its chrome spokes gleaming in the bright sunlight. As it moves, it reveals the heart of the machine: a massive, gleaming chrome V-twin engine with polished cooling fins and intricate brass hardware. The camera continues along the bike's low-slung, sleek black hardtail frame, lingering on the custom teardrop fuel tank, painted a deep Candy Apple Red with a high-gloss finish and embellished with delicate hand-painted gold leaf pinstripe. The light captures the rich texture of the handcrafted brown leather solo seat before sweeping over the fat rear tire and the long, upward-sweeping chrome fishtail exhaust. In the background, the vibrant yet softly blurred scene of the Born-Free Motorcycle Show unfolds, with the distant crowd of enthusiasts and the tops of vendor tents visible against the clear blue sky. The blazing midday sun cast sharp black shadows on the dry grass beneath the bike and created brilliant mirror highlights on every polished surface, creating an atmosphere of reverence and admiration for the motorcycle as a pristine work of mechanical art.", + "seed": 6033, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_0/index_7446_seed_6033_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.00942925401031971, + "vlm_logprob": -0.00942925401031971, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.00942925401031971, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009279691614210606 + ], + "opposite_logprob_tokens": [ + -5.384279727935791 + ], + "logprob_reward": -0.009279691614210606, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.009279691614210606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4045545160770416, + 0.02609570138156414, + 0.05478602647781372, + 0.035170186311006546, + 0.1342451572418213, + 0.05413445085287094, + 0.17508544027805328, + 0.02219552733004093, + 0.05474676936864853, + 0.038986239582300186 + ], + "credit_map": [ + 0.024427736178040504, + 0.019761722534894943, + 0.06429566442966461, + 0.019042862579226494, + 0.1468660682439804, + 0.0926838368177414, + 0.01813550479710102, + 0.019341133534908295, + 0.003594552632421255, + 0.004924000706523657, + 0.00420660525560379, + 0.0026526835281401873, + 0.002797025488689542, + 0.0023320484906435013, + 0.003470705123618245, + 0.0021180787589401007, + 0.0051710461266338825, + 0.009768474847078323, + 0.007506169844418764, + 0.005539475474506617, + 0.005224449094384909, + 0.005571674555540085, + 0.0064285919070243835, + 0.009576142765581608, + 0.004319294821470976, + 0.002390381647273898, + 0.004752254579216242, + 0.003628784092143178, + 0.0043686735443770885, + 0.005519654136151075, + 0.006252429448068142, + 0.0039387112483382225, + 0.011277816258370876, + 0.006870063487440348, + 0.04089741408824921, + 0.020039282739162445, + 0.007799429818987846, + 0.013749809004366398, + 0.01648530922830105, + 0.01712602935731411, + 0.007323102559894323, + 0.004052124451845884, + 0.006173526868224144, + 0.01952270232141018, + 0.003901474177837372, + 0.005849375389516354, + 0.004180435091257095, + 0.0031317078974097967, + 0.05686759948730469, + 0.013138615526258945, + 0.058329302817583084, + 0.015025530010461807, + 0.007544206455349922, + 0.009094905108213425, + 0.008081107400357723, + 0.007004173006862402, + 0.004282630048692226, + 0.0029136622324585915, + 0.0027006925083696842, + 0.00209675170481205, + 0.0037183447275310755, + 0.002421730197966099, + 0.0017908592708408833, + 0.002270856173709035, + 0.0091142812743783, + 0.007352212909609079, + 0.01024770736694336, + 0.005030057393014431, + 0.008225471712648869, + 0.006559822242707014, + 0.003947806544601917, + 0.004269408527761698, + 0.007859149016439915, + 0.0062284329906105995, + 0.0058871968649327755, + 0.004101892933249474, + 0.004238778259605169, + 0.002652734750881791, + 0.0036645852960646152, + 0.004353465046733618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.009279691614210606 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650568313896656 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.008650568313896656, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650568313896656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16378982365131378, + 0.06955623626708984, + 0.052207157015800476, + 0.02927970513701439, + 0.14146137237548828, + 0.06301621347665787, + 0.3719809949398041, + 0.03263319283723831, + 0.03753982484340668, + 0.03853539749979973 + ], + "credit_map": [ + 0.012183928862214088, + 0.012898179702460766, + 0.04715079069137573, + 0.01229154597967863, + 0.03017730824649334, + 0.02881047874689102, + 0.010939251631498337, + 0.009338350035250187, + 0.008640552870929241, + 0.017288701608777046, + 0.0143237030133605, + 0.005617471877485514, + 0.006895830854773521, + 0.006623433902859688, + 0.005478724837303162, + 0.004687820561230183, + 0.007565738167613745, + 0.004760554060339928, + 0.007681561633944511, + 0.007386591751128435, + 0.007176938466727734, + 0.005157926585525274, + 0.006523443385958672, + 0.005954408552497625, + 0.0035598687827587128, + 0.0024419401306658983, + 0.003589097410440445, + 0.00217924895696342, + 0.0037995351012796164, + 0.005800562910735607, + 0.0042873225174844265, + 0.0036221309565007687, + 0.014552156440913677, + 0.007146165706217289, + 0.045912448316812515, + 0.016700828447937965, + 0.011906281113624573, + 0.011295152828097343, + 0.023468315601348877, + 0.01048003975301981, + 0.00771932490170002, + 0.005097660701721907, + 0.0069516790099442005, + 0.007028402294963598, + 0.008366831578314304, + 0.014980518259108067, + 0.005492568947374821, + 0.007379232440143824, + 0.13581860065460205, + 0.017467418685555458, + 0.14149802923202515, + 0.026191307231783867, + 0.012179397977888584, + 0.01278116088360548, + 0.012245253659784794, + 0.01379983127117157, + 0.0065630218014121056, + 0.0041284323669970036, + 0.003394329221919179, + 0.0034489522222429514, + 0.0048082307912409306, + 0.0029999716207385063, + 0.003354061394929886, + 0.003936195746064186, + 0.005488638766109943, + 0.004759134259074926, + 0.0060579185374081135, + 0.006913638673722744, + 0.004187747836112976, + 0.0031524498481303453, + 0.003821015590801835, + 0.003159286454319954, + 0.007573723327368498, + 0.0052503133192658424, + 0.004542545415461063, + 0.0041530560702085495, + 0.006700665690004826, + 0.0027084846515208483, + 0.003677857806906104, + 0.003928752150386572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.008650568313896656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251897498965263 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251897498965263, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251897498965263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1431257128715515, + 0.04553842917084694, + 0.029728079214692116, + 0.01850166730582714, + 0.38089725375175476, + 0.03788594901561737, + 0.249985471367836, + 0.03892005980014801, + 0.0222071073949337, + 0.03321024775505066 + ], + "credit_map": [ + 0.008061299100518227, + 0.008279870264232159, + 0.020177211612462997, + 0.008812579326331615, + 0.05213393643498421, + 0.03259528800845146, + 0.00606840243563056, + 0.006997125688940287, + 0.005629035644233227, + 0.013958240859210491, + 0.007979737594723701, + 0.003193541197106242, + 0.002671370282769203, + 0.005100460723042488, + 0.003884600242599845, + 0.0031214396003633738, + 0.004186473321169615, + 0.003424869617447257, + 0.008023320697247982, + 0.0032428305130451918, + 0.002332911593839526, + 0.0019581331871449947, + 0.003765241475775838, + 0.0027942953165620565, + 0.0019263086142018437, + 0.0015527608338743448, + 0.0028342055156826973, + 0.0021578592713922262, + 0.002699097152799368, + 0.0027549860533326864, + 0.0031407091300934553, + 0.0014357400359585881, + 0.02352854423224926, + 0.011976810172200203, + 0.1552029252052307, + 0.05962139740586281, + 0.01233907975256443, + 0.02994740940630436, + 0.05973803624510765, + 0.028543071821331978, + 0.004388273227959871, + 0.0031980935018509626, + 0.004164057783782482, + 0.009022573009133339, + 0.003831434529274702, + 0.006580804008990526, + 0.0032439082860946655, + 0.0034568069968372583, + 0.09186885505914688, + 0.011271918192505836, + 0.10110978037118912, + 0.019464081153273582, + 0.004150017164647579, + 0.008540346287190914, + 0.0056759063154459, + 0.007904580794274807, + 0.006553316488862038, + 0.003726227907463908, + 0.003270277986302972, + 0.004309278447180986, + 0.003456795820966363, + 0.006684829946607351, + 0.0038161384873092175, + 0.007103196810930967, + 0.00343640916980803, + 0.004457730334252119, + 0.003242835635319352, + 0.0034629686269909143, + 0.0023457426577806473, + 0.0017295639263466, + 0.0016949550481513143, + 0.0018369023455306888, + 0.006505455821752548, + 0.004102184437215328, + 0.004724840167909861, + 0.005363631062209606, + 0.003815474221482873, + 0.002680511213839054, + 0.0032001477666199207, + 0.002818005159497261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.005251897498965263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187190681695938 + ], + "opposite_logprob_tokens": [ + -4.264187335968018 + ], + "logprob_reward": -0.014187190681695938, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187190681695938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21920043230056763, + 0.04198725149035454, + 0.025789113715291023, + 0.034243252128362656, + 0.3521478772163391, + 0.06457322835922241, + 0.10099037736654282, + 0.05794918164610863, + 0.040104448795318604, + 0.06301487237215042 + ], + "credit_map": [ + 0.015559297986328602, + 0.01491018570959568, + 0.04394039139151573, + 0.01261072140187025, + 0.07887077331542969, + 0.0302496999502182, + 0.011693413369357586, + 0.011365938931703568, + 0.006185231730341911, + 0.01012121606618166, + 0.005822389852255583, + 0.0036191206891089678, + 0.003033621469512582, + 0.004862617235630751, + 0.0038684867322444916, + 0.004474562592804432, + 0.0039807576686143875, + 0.0022957869805395603, + 0.004558958113193512, + 0.0029751118272542953, + 0.0025396074634045362, + 0.0024285686668008566, + 0.0036684954538941383, + 0.003341827541589737, + 0.004537028260529041, + 0.0020450844895094633, + 0.0042441668920218945, + 0.0037076023872941732, + 0.004338814876973629, + 0.00488152215257287, + 0.0054236482828855515, + 0.005065383389592171, + 0.022585665807127953, + 0.010950365103781223, + 0.1442297101020813, + 0.048116546124219894, + 0.011826383881270885, + 0.026580380275845528, + 0.06317350268363953, + 0.02468530833721161, + 0.009701212868094444, + 0.006631926167756319, + 0.007443875540047884, + 0.008238263428211212, + 0.006332800257951021, + 0.011552194133400917, + 0.007016867399215698, + 0.007656086701899767, + 0.026694102212786674, + 0.006234359461814165, + 0.03271932527422905, + 0.01443142257630825, + 0.00459117628633976, + 0.006538693327456713, + 0.004694210831075907, + 0.005087079480290413, + 0.009401003830134869, + 0.006869107019156218, + 0.006547584198415279, + 0.008904202841222286, + 0.007727984804660082, + 0.0056882319040596485, + 0.00565701350569725, + 0.007154046557843685, + 0.006442597601562738, + 0.0054999408312141895, + 0.00615993607789278, + 0.005109978839755058, + 0.006761840078979731, + 0.0036701099015772343, + 0.0031383130699396133, + 0.0033217319287359715, + 0.013010813854634762, + 0.00812099315226078, + 0.009279242716729641, + 0.00655398890376091, + 0.007734052371233702, + 0.0056948596611619, + 0.00650821952149272, + 0.006112694274634123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.014187190681695938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776921942830086 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009776921942830086, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776921942830086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16614896059036255, + 0.165265291929245, + 0.03301762416958809, + 0.05873839184641838, + 0.2669983506202698, + 0.14132225513458252, + 0.10837788134813309, + 0.022053362801671028, + 0.018188895657658577, + 0.01988893374800682 + ], + "credit_map": [ + 0.010315337218344212, + 0.008237817324697971, + 0.015987277030944824, + 0.009030850604176521, + 0.05987466126680374, + 0.04753934592008591, + 0.007318409159779549, + 0.007845268584787846, + 0.027868665754795074, + 0.041067592799663544, + 0.0211588554084301, + 0.014120997861027718, + 0.012853049673140049, + 0.019489215686917305, + 0.011929390020668507, + 0.016777537763118744, + 0.004212912637740374, + 0.0052436660043895245, + 0.003099905326962471, + 0.002840756205841899, + 0.0035761091858148575, + 0.004114658106118441, + 0.005100473761558533, + 0.004829145967960358, + 0.007527275010943413, + 0.004215423483401537, + 0.007728387136012316, + 0.006918730214238167, + 0.0071804821491241455, + 0.008730312809348106, + 0.01054341346025467, + 0.0058943722397089005, + 0.01810608245432377, + 0.008140577003359795, + 0.10465038567781448, + 0.03920116275548935, + 0.008805334568023682, + 0.020710306242108345, + 0.04039425775408745, + 0.02699025720357895, + 0.02540232054889202, + 0.01678434945642948, + 0.017856206744909286, + 0.038211751729249954, + 0.010533198714256287, + 0.013909329660236835, + 0.010368047282099724, + 0.008257048204541206, + 0.03959398716688156, + 0.005408659111708403, + 0.0432601198554039, + 0.007726428564637899, + 0.0021804168354719877, + 0.0038449582643806934, + 0.002859110478311777, + 0.003504212247207761, + 0.004055415745824575, + 0.002662064041942358, + 0.0022533026058226824, + 0.002783342031762004, + 0.0032254932448267937, + 0.0024680441711097956, + 0.0022661546245217323, + 0.002339546335861087, + 0.0027961961459368467, + 0.0027260209899395704, + 0.0025733758229762316, + 0.0021331750322133303, + 0.00276414817199111, + 0.001932104816660285, + 0.0017326376400887966, + 0.0015312362229451537, + 0.0037392235826700926, + 0.0025001876056194305, + 0.0031857446301728487, + 0.002503905910998583, + 0.002807637443765998, + 0.0015702696982771158, + 0.0020952667109668255, + 0.0014866983983665705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.009776921942830086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate, low-angle tracking shot slides from left to right across a meticulously displayed custom chopper motorcycle on a sun-drenched lawn at an outdoor motorcycle show. The camera first captures the tall, spoked front wheel, its chrome spokes gleaming in the bright sunlight. As it moves, it reveals the heart of the machine: a massive, gleaming chrome V-twin engine with polished cooling fins and intricate brass hardware. The camera continues along the bike's low-slung, sleek black hardtail frame, lingering on the custom teardrop fuel tank, painted a deep Candy Apple Red with a high-gloss finish and embellished with delicate hand-painted gold leaf pinstripe. The light captures the rich texture of the handcrafted brown leather solo seat before sweeping over the fat rear tire and the long, upward-sweeping chrome fishtail exhaust. In the background, the vibrant yet softly blurred scene of the Born-Free Motorcycle Show unfolds, with the distant crowd of enthusiasts and the tops of vendor tents visible against the clear blue sky. The blazing midday sun cast sharp black shadows on the dry grass beneath the bike and created brilliant mirror highlights on every polished surface, creating an atmosphere of reverence and admiration for the motorcycle as a pristine work of mechanical art.", + "seed": 6034, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_0/index_7446_seed_6034_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01291375132277608, + "vlm_logprob": -0.01291375132277608, + "vlm_yes_no_margin": 4.499999952316284, + "vlm_token_credit": -0.01291375132277608, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022281773388385773 + ], + "opposite_logprob_tokens": [ + -4.022281646728516 + ], + "logprob_reward": -0.022281773388385773, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.022281773388385773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16021552681922913, + 0.16074497997760773, + 0.044101424515247345, + 0.05093064159154892, + 0.04210633784532547, + 0.09708187729120255, + 0.13742052018642426, + 0.03263738006353378, + 0.17660653591156006, + 0.09815486520528793 + ], + "credit_map": [ + 0.02634446695446968, + 0.014857775531709194, + 0.01578572951257229, + 0.012672876007854939, + 0.03390631079673767, + 0.0192634928971529, + 0.02227916568517685, + 0.015105714090168476, + 0.014222715049982071, + 0.03460334986448288, + 0.018678518012166023, + 0.016599804162979126, + 0.022098073735833168, + 0.015578176826238632, + 0.01718919351696968, + 0.021775148808956146, + 0.0032717802096158266, + 0.004557167645543814, + 0.002810381818562746, + 0.005163284484297037, + 0.006408157292753458, + 0.009313970804214478, + 0.005138137377798557, + 0.007438544183969498, + 0.00386013463139534, + 0.0031663423869758844, + 0.006698778830468655, + 0.008396744728088379, + 0.00980384275317192, + 0.012211139313876629, + 0.0033515505492687225, + 0.003442106768488884, + 0.005617761518806219, + 0.004938882309943438, + 0.00639098696410656, + 0.003920393064618111, + 0.004214223939925432, + 0.007340073119848967, + 0.0054886480793356895, + 0.004195365589112043, + 0.008970950730144978, + 0.010660849511623383, + 0.009466108866035938, + 0.007526351138949394, + 0.021152280271053314, + 0.006192762870341539, + 0.023342246189713478, + 0.009770325385034084, + 0.011733748950064182, + 0.012502970173954964, + 0.011071329936385155, + 0.02303863875567913, + 0.014413765631616116, + 0.012590697035193443, + 0.037747573107481, + 0.014321798458695412, + 0.0034476101864129305, + 0.005151391960680485, + 0.004674144089221954, + 0.005874615628272295, + 0.0029536366928368807, + 0.002898121252655983, + 0.00341996643692255, + 0.004217895679175854, + 0.023325257003307343, + 0.017263010144233704, + 0.02792114019393921, + 0.03709683194756508, + 0.014105298556387424, + 0.0164791252464056, + 0.01665123924612999, + 0.02376464381814003, + 0.015386560000479221, + 0.010902601294219494, + 0.015508823096752167, + 0.016147978603839874, + 0.0076575144194066525, + 0.008962485939264297, + 0.013498560525476933, + 0.010090342722833157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.022281773388385773 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656122721731663 + ], + "opposite_logprob_tokens": [ + -4.758656024932861 + ], + "logprob_reward": -0.008656122721731663, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656122721731663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05224613845348358, + 0.17041683197021484, + 0.07240279018878937, + 0.0408005453646183, + 0.06348767131567001, + 0.06646613031625748, + 0.12064427137374878, + 0.15342681109905243, + 0.1309116780757904, + 0.1291971653699875 + ], + "credit_map": [ + 0.008893897756934166, + 0.00561579829081893, + 0.005994030274450779, + 0.004379312042146921, + 0.00862845778465271, + 0.005514681339263916, + 0.006843773648142815, + 0.006376183591783047, + 0.014412147924304008, + 0.04283943399786949, + 0.014557305723428726, + 0.01847253367304802, + 0.017634863033890724, + 0.026686180382966995, + 0.014730066992342472, + 0.02108430117368698, + 0.006440075114369392, + 0.005610269494354725, + 0.0075003523379564285, + 0.006833174265921116, + 0.007805254776030779, + 0.01835007220506668, + 0.00727890944108367, + 0.012584684416651726, + 0.003948875702917576, + 0.004534941166639328, + 0.00589393638074398, + 0.004294659476727247, + 0.00836262945085764, + 0.007575375027954578, + 0.0031314948573708534, + 0.0030586370266973972, + 0.008575057610869408, + 0.006065069697797298, + 0.008921694941818714, + 0.005429020617157221, + 0.009432090446352959, + 0.008372158743441105, + 0.010620052926242352, + 0.006072521675378084, + 0.006588716525584459, + 0.006953416392207146, + 0.005732431076467037, + 0.0065935333259403706, + 0.014090740121901035, + 0.005537357181310654, + 0.014108474366366863, + 0.006861460395157337, + 0.01995377615094185, + 0.01045313011854887, + 0.013198335655033588, + 0.017944233492016792, + 0.01147961150854826, + 0.011280333623290062, + 0.02303367666900158, + 0.013301165774464607, + 0.012359526939690113, + 0.03342504799365997, + 0.015761949121952057, + 0.02564607560634613, + 0.019362563267350197, + 0.009859403595328331, + 0.021115295588970184, + 0.015896951779723167, + 0.02154841646552086, + 0.015459118410944939, + 0.02092542126774788, + 0.029245026409626007, + 0.010349125601351261, + 0.011036972515285015, + 0.009327231906354427, + 0.013020367361605167, + 0.01996018923819065, + 0.015039712190628052, + 0.020034363493323326, + 0.025203222408890724, + 0.008539034985005856, + 0.010035672225058079, + 0.017687810584902763, + 0.012697171419858932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.008656122721731663 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251778755337 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251778755337, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251778755337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12698715925216675, + 0.1851239651441574, + 0.04571533203125, + 0.03155224025249481, + 0.0472804494202137, + 0.060845207422971725, + 0.1727895438671112, + 0.053754713386297226, + 0.20533493161201477, + 0.0706164613366127 + ], + "credit_map": [ + 0.022721799090504646, + 0.012100418098270893, + 0.011428053490817547, + 0.009697919711470604, + 0.025949645787477493, + 0.01451477687805891, + 0.01882977969944477, + 0.01174477394670248, + 0.015906119719147682, + 0.033536165952682495, + 0.023922888562083244, + 0.019760357216000557, + 0.027867646887898445, + 0.019488120451569557, + 0.02317177690565586, + 0.021470898762345314, + 0.003957370296120644, + 0.004740349482744932, + 0.0037773437798023224, + 0.004398523364216089, + 0.004603604320436716, + 0.011901052668690681, + 0.004982561804354191, + 0.00735452538356185, + 0.0033618141897022724, + 0.004990297369658947, + 0.004755458328872919, + 0.00380946253426373, + 0.00289570027962327, + 0.005407090298831463, + 0.0027855271473526955, + 0.0035468940623104572, + 0.005854632705450058, + 0.005545920692384243, + 0.00722464919090271, + 0.004478273913264275, + 0.00474079605191946, + 0.008509260602295399, + 0.005753047298640013, + 0.005173872224986553, + 0.0074000488966703415, + 0.006077628582715988, + 0.006558999419212341, + 0.005170222371816635, + 0.012916275300085545, + 0.004677883815020323, + 0.012397552840411663, + 0.005646595731377602, + 0.014728804118931293, + 0.014524675905704498, + 0.01491909846663475, + 0.02952343225479126, + 0.017258502542972565, + 0.015039291232824326, + 0.04606664553284645, + 0.020729100331664085, + 0.00504327192902565, + 0.006330662872642279, + 0.007091496605426073, + 0.00999946054071188, + 0.005490632262080908, + 0.004442378878593445, + 0.007715205196291208, + 0.0076416125521063805, + 0.02812836691737175, + 0.019776511937379837, + 0.03211328014731407, + 0.04041210934519768, + 0.017739394679665565, + 0.01872790977358818, + 0.019750501960515976, + 0.028686869889497757, + 0.011504174210131168, + 0.00712581118568778, + 0.011607439257204533, + 0.010086104273796082, + 0.0063727861270308495, + 0.008329520002007484, + 0.00811680220067501, + 0.007473827339708805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.005251778755337 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188013039529324 + ], + "opposite_logprob_tokens": [ + -4.264187812805176 + ], + "logprob_reward": -0.014188013039529324, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188013039529324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08187208324670792, + 0.11776675283908844, + 0.05137907713651657, + 0.07349094748497009, + 0.06505835056304932, + 0.1574210226535797, + 0.18950068950653076, + 0.1092153936624527, + 0.08942711353302002, + 0.06486856937408447 + ], + "credit_map": [ + 0.013566245324909687, + 0.007340895943343639, + 0.00783658679574728, + 0.007275435607880354, + 0.015045352280139923, + 0.008594180457293987, + 0.012867513112723827, + 0.009345869533717632, + 0.012913673184812069, + 0.014447329565882683, + 0.019715577363967896, + 0.01287279836833477, + 0.014873960986733437, + 0.012102590873837471, + 0.014628586359322071, + 0.01621222123503685, + 0.0040297009982168674, + 0.004709851462393999, + 0.0038968429435044527, + 0.004813882987946272, + 0.004552867263555527, + 0.012379043735563755, + 0.007017245050519705, + 0.009979639202356339, + 0.005830172915011644, + 0.006666676606982946, + 0.006250970996916294, + 0.008409292437136173, + 0.006544589996337891, + 0.02200940065085888, + 0.00447867764160037, + 0.013301166705787182, + 0.007942864671349525, + 0.007114236243069172, + 0.009465521201491356, + 0.005811713170260191, + 0.007048964034765959, + 0.010486574843525887, + 0.010653487406671047, + 0.006534980610013008, + 0.012266318313777447, + 0.016849825158715248, + 0.01367438305169344, + 0.012235840782523155, + 0.03534984588623047, + 0.009867065586149693, + 0.04047123342752457, + 0.016706500202417374, + 0.029010217636823654, + 0.01456805132329464, + 0.019593222066760063, + 0.028914321213960648, + 0.016061600297689438, + 0.0158827044069767, + 0.04198629409074783, + 0.023484263569116592, + 0.010510980151593685, + 0.014644689857959747, + 0.013618054799735546, + 0.020872989669442177, + 0.013099604286253452, + 0.007556702941656113, + 0.016212115064263344, + 0.012700236402451992, + 0.012554340064525604, + 0.00944664515554905, + 0.01383323036134243, + 0.018844738602638245, + 0.007531074341386557, + 0.008740101009607315, + 0.008339934051036835, + 0.010137044824659824, + 0.008404303342103958, + 0.0048035490326583385, + 0.010404053144156933, + 0.012800523079931736, + 0.006246961187571287, + 0.0058087194338440895, + 0.007946234196424484, + 0.008454222232103348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.014188013039529324 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191068708896637 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191068708896637, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191068708896637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11962267011404037, + 0.13696445524692535, + 0.108268603682518, + 0.034968532621860504, + 0.04092417284846306, + 0.033473338931798935, + 0.21932806074619293, + 0.07954557240009308, + 0.16971024870872498, + 0.05719441547989845 + ], + "credit_map": [ + 0.022884638980031013, + 0.012608347460627556, + 0.0113888094201684, + 0.008897453546524048, + 0.024679066613316536, + 0.011840112507343292, + 0.017206495627760887, + 0.010117731057107449, + 0.012362228706479073, + 0.03694849833846092, + 0.015416809357702732, + 0.01574847102165222, + 0.014674627222120762, + 0.016804520040750504, + 0.010448756627738476, + 0.014560521580278873, + 0.011450821533799171, + 0.011825528927147388, + 0.013324601575732231, + 0.012216071598231792, + 0.011670681647956371, + 0.018300002440810204, + 0.011189433746039867, + 0.018291443586349487, + 0.004072755575180054, + 0.0040782541036605835, + 0.005999238695949316, + 0.0033820336684584618, + 0.00416948925703764, + 0.00610234634950757, + 0.0026639383286237717, + 0.00450047105550766, + 0.004427470266819, + 0.004201347474008799, + 0.006240759510546923, + 0.004485594108700752, + 0.00470617925748229, + 0.006745012942701578, + 0.005475780460983515, + 0.004642023239284754, + 0.005003245081752539, + 0.0034517021849751472, + 0.004558349493891001, + 0.002943586790934205, + 0.006497373804450035, + 0.0023136972449719906, + 0.0056008812971413136, + 0.0031044965144246817, + 0.03191933408379555, + 0.018663998693227768, + 0.023158656433224678, + 0.034183964133262634, + 0.019373677670955658, + 0.018977925181388855, + 0.045801494270563126, + 0.027248980477452278, + 0.006270471028983593, + 0.009892408736050129, + 0.010758464224636555, + 0.012476148083806038, + 0.007180158980190754, + 0.010178763419389725, + 0.007714409381151199, + 0.015074738301336765, + 0.01939004473388195, + 0.013374838046729565, + 0.02427363581955433, + 0.028598995879292488, + 0.012115377001464367, + 0.01923495903611183, + 0.015930719673633575, + 0.03679167106747627, + 0.009538544341921806, + 0.005983521696180105, + 0.009587532840669155, + 0.008700612932443619, + 0.004593250807374716, + 0.0052890232764184475, + 0.007559381425380707, + 0.005942541640251875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.014191068708896637 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate, low-angle tracking shot slides from left to right across a meticulously displayed custom chopper motorcycle on a sun-drenched lawn at an outdoor motorcycle show. The camera first captures the tall, spoked front wheel, its chrome spokes gleaming in the bright sunlight. As it moves, it reveals the heart of the machine: a massive, gleaming chrome V-twin engine with polished cooling fins and intricate brass hardware. The camera continues along the bike's low-slung, sleek black hardtail frame, lingering on the custom teardrop fuel tank, painted a deep Candy Apple Red with a high-gloss finish and embellished with delicate hand-painted gold leaf pinstripe. The light captures the rich texture of the handcrafted brown leather solo seat before sweeping over the fat rear tire and the long, upward-sweeping chrome fishtail exhaust. In the background, the vibrant yet softly blurred scene of the Born-Free Motorcycle Show unfolds, with the distant crowd of enthusiasts and the tops of vendor tents visible against the clear blue sky. The blazing midday sun cast sharp black shadows on the dry grass beneath the bike and created brilliant mirror highlights on every polished surface, creating an atmosphere of reverence and admiration for the motorcycle as a pristine work of mechanical art.", + "seed": 6035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_0/index_7446_seed_6035_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020165111776441334, + "vlm_logprob": -0.020165111776441334, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.020165111776441334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011458891443908215 + ], + "opposite_logprob_tokens": [ + -4.761458873748779 + ], + "logprob_reward": -0.011458891443908215, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011458891443908215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.241657555103302, + 0.043563373386859894, + 0.04736439138650894, + 0.022908328101038933, + 0.2312742918729782, + 0.050651922821998596, + 0.04786939546465874, + 0.21322689950466156, + 0.02716529555618763, + 0.07431856542825699 + ], + "credit_map": [ + 0.03634783625602722, + 0.022055640816688538, + 0.03026513382792473, + 0.04488470032811165, + 0.026308923959732056, + 0.03555795177817345, + 0.024893341585993767, + 0.021344022825360298, + 0.004129140172153711, + 0.004774521104991436, + 0.005014414433389902, + 0.006472373381257057, + 0.007859311997890472, + 0.004877917934209108, + 0.005825529806315899, + 0.004610165022313595, + 0.0047419932670891285, + 0.005771473981440067, + 0.006599690765142441, + 0.005304198712110519, + 0.0038019423373043537, + 0.006192764267325401, + 0.005399434361606836, + 0.009552893228828907, + 0.0025192664470523596, + 0.004760615527629852, + 0.0025292118079960346, + 0.002584812231361866, + 0.002146404469385743, + 0.003370843129232526, + 0.0017590642673894763, + 0.003238110337406397, + 0.07233325392007828, + 0.019099531695246696, + 0.022492187097668648, + 0.05920732393860817, + 0.014970360323786736, + 0.010527513921260834, + 0.01782531477510929, + 0.014818817377090454, + 0.006368428003042936, + 0.006177052855491638, + 0.007827458903193474, + 0.002842411631718278, + 0.005979078356176615, + 0.00492226704955101, + 0.010882147587835789, + 0.005653077270835638, + 0.006105835549533367, + 0.005020300857722759, + 0.003703327616676688, + 0.005755000747740269, + 0.008261573500931263, + 0.006869670934975147, + 0.007067703641951084, + 0.0050859819166362286, + 0.02569066360592842, + 0.03258277848362923, + 0.023986991494894028, + 0.0290196742862463, + 0.03562811017036438, + 0.02023039571940899, + 0.0180968027561903, + 0.02799147367477417, + 0.0045442404225468636, + 0.0034257695078849792, + 0.0032767646480351686, + 0.0038351521361619234, + 0.0027972296811640263, + 0.003261694684624672, + 0.0032688111532479525, + 0.0027556323911994696, + 0.015983089804649353, + 0.007325795013457537, + 0.01061796210706234, + 0.012338957749307156, + 0.007565471809357405, + 0.006983403582125902, + 0.006359308026731014, + 0.007144583854824305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.011458891443908215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794392623007298 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.009794392623007298, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794392623007298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16066981852054596, + 0.0943753644824028, + 0.08938270807266235, + 0.034538690000772476, + 0.1817554384469986, + 0.10434840619564056, + 0.0858350545167923, + 0.13726137578487396, + 0.03947019949555397, + 0.07236298173666 + ], + "credit_map": [ + 0.02386830747127533, + 0.01722736470401287, + 0.018262751400470734, + 0.03217197582125664, + 0.0152736259624362, + 0.022620826959609985, + 0.01638091169297695, + 0.014864049851894379, + 0.007902677170932293, + 0.01185075007379055, + 0.011562608182430267, + 0.008772184140980244, + 0.016738442704081535, + 0.015240705572068691, + 0.01234709657728672, + 0.009960893541574478, + 0.007071705535054207, + 0.009605242870748043, + 0.008326626382768154, + 0.010900446213781834, + 0.008300579153001308, + 0.020155958831310272, + 0.011988905258476734, + 0.013033243827521801, + 0.003176263766363263, + 0.005024927668273449, + 0.005936486646533012, + 0.003500873688608408, + 0.002985671628266573, + 0.005059448536485434, + 0.002273757476359606, + 0.0065812598913908005, + 0.019321313127875328, + 0.025388874113559723, + 0.02222421206533909, + 0.01688159443438053, + 0.022222042083740234, + 0.02284461073577404, + 0.022809859365224838, + 0.030062921345233917, + 0.009932312183082104, + 0.010791658423841, + 0.010502619668841362, + 0.008440553210675716, + 0.02196074277162552, + 0.01268294733017683, + 0.015523159876465797, + 0.014514414593577385, + 0.010840079747140408, + 0.010212033055722713, + 0.006367540918290615, + 0.009686385281383991, + 0.0150475287809968, + 0.011208009906113148, + 0.016063908115029335, + 0.00640956312417984, + 0.01743612438440323, + 0.0200045108795166, + 0.017103753983974457, + 0.02005484141409397, + 0.021584032103419304, + 0.011671729385852814, + 0.011697366833686829, + 0.017709005624055862, + 0.004971500486135483, + 0.004171099979430437, + 0.0053265271708369255, + 0.005406178068369627, + 0.003763548331335187, + 0.00583722535520792, + 0.0044683003798127174, + 0.00552581949159503, + 0.01310928724706173, + 0.006966746412217617, + 0.012405283749103546, + 0.010596285574138165, + 0.00606390368193388, + 0.00760932732373476, + 0.007000199519097805, + 0.008611949160695076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.009794392623007298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735715549439192 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735715549439192, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735715549439192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1688956618309021, + 0.06644457578659058, + 0.1069151908159256, + 0.037698034197092056, + 0.1650882363319397, + 0.06637730449438095, + 0.037200815975666046, + 0.21843171119689941, + 0.047032274305820465, + 0.0859161987900734 + ], + "credit_map": [ + 0.017328528687357903, + 0.01938888244330883, + 0.025437215343117714, + 0.022814342752099037, + 0.020179683342576027, + 0.019211331382393837, + 0.028880828991532326, + 0.015654845163226128, + 0.006473429035395384, + 0.007828814908862114, + 0.00895893108099699, + 0.007574469782412052, + 0.00903990026563406, + 0.008749041706323624, + 0.009252768941223621, + 0.008567222394049168, + 0.010307136923074722, + 0.011694472283124924, + 0.013768672943115234, + 0.011903947219252586, + 0.008770469576120377, + 0.019719095900654793, + 0.013375544920563698, + 0.017375849187374115, + 0.003176230238750577, + 0.004541620146483183, + 0.006161045283079147, + 0.003169994568452239, + 0.00394083559513092, + 0.005342516582459211, + 0.0030331488233059645, + 0.00833264272660017, + 0.04526609182357788, + 0.011801216751337051, + 0.01682681404054165, + 0.04637482762336731, + 0.01071250531822443, + 0.007473221980035305, + 0.015515844337642193, + 0.011117721907794476, + 0.006937889847904444, + 0.006212842185050249, + 0.009010166861116886, + 0.006062062922865152, + 0.009531664662063122, + 0.006203201599419117, + 0.01369889359921217, + 0.008720576763153076, + 0.0048644402995705605, + 0.0037122275680303574, + 0.0033564034383744, + 0.004345770925283432, + 0.0065744491294026375, + 0.005061475094407797, + 0.005162353627383709, + 0.004123696591705084, + 0.02459407038986683, + 0.033455174416303635, + 0.024830564856529236, + 0.02938896417617798, + 0.03639788553118706, + 0.020701458677649498, + 0.018591923639178276, + 0.030471662059426308, + 0.00564952939748764, + 0.005967298522591591, + 0.005003012716770172, + 0.006075023207813501, + 0.004733345936983824, + 0.007288346067070961, + 0.0062568639405071735, + 0.006058855447918177, + 0.01736214943230152, + 0.008887670934200287, + 0.0129415113478899, + 0.013284704647958279, + 0.00686809653416276, + 0.009099440649151802, + 0.00819985382258892, + 0.009272774681448936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.006735715549439192 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029786404222249985 + ], + "opposite_logprob_tokens": [ + -3.5297863483428955 + ], + "logprob_reward": -0.029786404222249985, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029786404222249985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1645248383283615, + 0.06400315463542938, + 0.09819437563419342, + 0.05944514647126198, + 0.1429426372051239, + 0.1014670580625534, + 0.05981513112783432, + 0.05506160482764244, + 0.10039442032575607, + 0.15415167808532715 + ], + "credit_map": [ + 0.029425695538520813, + 0.01703903079032898, + 0.018125019967556, + 0.02188250981271267, + 0.016253996640443802, + 0.022667471319437027, + 0.020632673054933548, + 0.01849842071533203, + 0.004603042732924223, + 0.007523379288613796, + 0.005622954107820988, + 0.006230527069419622, + 0.01628231257200241, + 0.00859472993761301, + 0.008091983385384083, + 0.007054222747683525, + 0.009004884399473667, + 0.010997208766639233, + 0.010063739493489265, + 0.010794137604534626, + 0.010292275808751583, + 0.022544626146554947, + 0.012774338014423847, + 0.01172315701842308, + 0.006943065207451582, + 0.00602328684180975, + 0.011281277984380722, + 0.0048981900326907635, + 0.0062196883372962475, + 0.0076004816219210625, + 0.004414631985127926, + 0.012064523994922638, + 0.043224457651376724, + 0.014673668891191483, + 0.010937802493572235, + 0.033961620181798935, + 0.01189209520816803, + 0.009679458104074001, + 0.009710293263196945, + 0.008863221853971481, + 0.012651970610022545, + 0.009870716370642185, + 0.016069334000349045, + 0.008174474351108074, + 0.015262790024280548, + 0.012545834295451641, + 0.014094831421971321, + 0.012797094881534576, + 0.005218850448727608, + 0.007013470865786076, + 0.0040769730694592, + 0.008814448490738869, + 0.00899476371705532, + 0.006905870512127876, + 0.014142035506665707, + 0.004648715257644653, + 0.005548716057091951, + 0.005674690008163452, + 0.006499742157757282, + 0.008127361536026001, + 0.007425167132169008, + 0.006488769315183163, + 0.00865519605576992, + 0.0066419560462236404, + 0.014530297368764877, + 0.012318945489823818, + 0.011541697196662426, + 0.014215187169611454, + 0.012261764146387577, + 0.012785308994352818, + 0.01074658427387476, + 0.011994622647762299, + 0.03319336101412773, + 0.01599881425499916, + 0.02179088070988655, + 0.026846803724765778, + 0.015231925062835217, + 0.01040273904800415, + 0.013080699369311333, + 0.01760643906891346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.029786404222249985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305015504360199 + ], + "opposite_logprob_tokens": [ + -3.1680500507354736 + ], + "logprob_reward": -0.04305015504360199, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305015504360199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22710229456424713, + 0.07498718053102493, + 0.26620882749557495, + 0.0586414597928524, + 0.08217393606901169, + 0.046059343963861465, + 0.04388824850320816, + 0.02263660356402397, + 0.03399424999952316, + 0.14430789649486542 + ], + "credit_map": [ + 0.02631557732820511, + 0.019430730491876602, + 0.038316041231155396, + 0.02404840476810932, + 0.03533630073070526, + 0.03794233128428459, + 0.02479107677936554, + 0.020921804010868073, + 0.007834740914404392, + 0.009156377986073494, + 0.006929522845894098, + 0.00769350491464138, + 0.017958778887987137, + 0.008164766244590282, + 0.008537961170077324, + 0.008711517788469791, + 0.02131526730954647, + 0.028411414474248886, + 0.026871593669056892, + 0.0281510129570961, + 0.025944748893380165, + 0.0630870833992958, + 0.03648700192570686, + 0.03594067692756653, + 0.0039694225415587425, + 0.004982918966561556, + 0.007733756210654974, + 0.006263528019189835, + 0.007368885446339846, + 0.01138926949352026, + 0.006416777614504099, + 0.010516894049942493, + 0.02089313417673111, + 0.01018326822668314, + 0.007250955794006586, + 0.01997586339712143, + 0.005256787408143282, + 0.004218666814267635, + 0.006754317786544561, + 0.007640935480594635, + 0.004341893829405308, + 0.006317197345197201, + 0.005934237502515316, + 0.00414553377777338, + 0.009746949188411236, + 0.005081831477582455, + 0.006309412885457277, + 0.004182286560535431, + 0.004998250864446163, + 0.005046629346907139, + 0.0035330005921423435, + 0.003651535836979747, + 0.005877980962395668, + 0.0049263061955571175, + 0.011366033926606178, + 0.0044885072857141495, + 0.0028069219551980495, + 0.003502548672258854, + 0.0017102912534028292, + 0.0036742077209055424, + 0.0030458918772637844, + 0.003150637960061431, + 0.0023778744507580996, + 0.0023682275786995888, + 0.003786479588598013, + 0.004365952685475349, + 0.003579501062631607, + 0.004820450209081173, + 0.00338525278493762, + 0.004176496993750334, + 0.004495093133300543, + 0.005385018419474363, + 0.02706466242671013, + 0.013740353286266327, + 0.01932118460536003, + 0.022287024185061455, + 0.02035270817577839, + 0.01362666580826044, + 0.014245884492993355, + 0.013669388368725777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.04305015504360199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, deliberate, low-angle tracking shot slides from left to right across a meticulously displayed custom chopper motorcycle on a sun-drenched lawn at an outdoor motorcycle show. The camera first captures the tall, spoked front wheel, its chrome spokes gleaming in the bright sunlight. As it moves, it reveals the heart of the machine: a massive, gleaming chrome V-twin engine with polished cooling fins and intricate brass hardware. The camera continues along the bike's low-slung, sleek black hardtail frame, lingering on the custom teardrop fuel tank, painted a deep Candy Apple Red with a high-gloss finish and embellished with delicate hand-painted gold leaf pinstripe. The light captures the rich texture of the handcrafted brown leather solo seat before sweeping over the fat rear tire and the long, upward-sweeping chrome fishtail exhaust. In the background, the vibrant yet softly blurred scene of the Born-Free Motorcycle Show unfolds, with the distant crowd of enthusiasts and the tops of vendor tents visible against the clear blue sky. The blazing midday sun cast sharp black shadows on the dry grass beneath the bike and created brilliant mirror highlights on every polished surface, creating an atmosphere of reverence and admiration for the motorcycle as a pristine work of mechanical art.", + "seed": 6036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_0/index_7446_seed_6036_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021401446685194968, + "vlm_logprob": -0.021401446685194968, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.021401446685194968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013288152404129505 + ], + "opposite_logprob_tokens": [ + -4.7632880210876465 + ], + "logprob_reward": -0.013288152404129505, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013288152404129505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057494595646858215, + 0.06470882892608643, + 0.06826601177453995, + 0.02950677089393139, + 0.12754742801189423, + 0.03523533418774605, + 0.013087804429233074, + 0.044942017644643784, + 0.5276004672050476, + 0.03161083906888962 + ], + "credit_map": [ + 0.008367377333343029, + 0.00618999358266592, + 0.010723796673119068, + 0.00605864729732275, + 0.006044934503734112, + 0.004869938362389803, + 0.007279828190803528, + 0.00796007551252842, + 0.006212524138391018, + 0.005630372557789087, + 0.005392769817262888, + 0.006712061353027821, + 0.012165149673819542, + 0.005622988101094961, + 0.012292956933379173, + 0.010679995641112328, + 0.008586264215409756, + 0.006745625287294388, + 0.008257243782281876, + 0.006456098519265652, + 0.013110700994729996, + 0.0059934877790510654, + 0.010556667111814022, + 0.00855991616845131, + 0.002524781972169876, + 0.0029821963980793953, + 0.002541401656344533, + 0.00283526792190969, + 0.004029958974570036, + 0.003907986916601658, + 0.0062649245373904705, + 0.004420251119881868, + 0.009753519669175148, + 0.0038485215045511723, + 0.006073107942938805, + 0.031153425574302673, + 0.003930612932890654, + 0.018967686221003532, + 0.008117198012769222, + 0.04570334777235985, + 0.0023489645682275295, + 0.004307955503463745, + 0.002845727140083909, + 0.0071724411100149155, + 0.0028326972387731075, + 0.004009008873254061, + 0.007802980951964855, + 0.003915552049875259, + 0.0014571959618479013, + 0.0014621318550780416, + 0.0016892411513254046, + 0.0013568439753726125, + 0.0013380857417359948, + 0.0019743535667657852, + 0.0018598880851641297, + 0.0019500632770359516, + 0.002884723711758852, + 0.0036991967353969812, + 0.010166870430111885, + 0.004466206766664982, + 0.002656428376212716, + 0.0039051841013133526, + 0.012233325280249119, + 0.0049300785176455975, + 0.047409866005182266, + 0.08411292731761932, + 0.0756484642624855, + 0.03387145698070526, + 0.09827949851751328, + 0.0401066429913044, + 0.1013922169804573, + 0.046779315918684006, + 0.002904523629695177, + 0.003294703783467412, + 0.0069368574768304825, + 0.0035167094320058823, + 0.0036083741579204798, + 0.0040643722750246525, + 0.004558331333100796, + 0.0027269672136753798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.013288152404129505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006753003224730492 + ], + "opposite_logprob_tokens": [ + -5.006752967834473 + ], + "logprob_reward": -0.006753003224730492, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006753003224730492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03608131408691406, + 0.04450571909546852, + 0.06718259304761887, + 0.018618358299136162, + 0.13255351781845093, + 0.040116921067237854, + 0.012375212274491787, + 0.09131256490945816, + 0.41617903113365173, + 0.14107480645179749 + ], + "credit_map": [ + 0.004703267943114042, + 0.003952653147280216, + 0.007381531409919262, + 0.00422497047111392, + 0.004349170718342066, + 0.0031374599784612656, + 0.003951431252062321, + 0.004380827769637108, + 0.005902847275137901, + 0.005021028220653534, + 0.0033950211945921183, + 0.00732108810916543, + 0.004468909930437803, + 0.0030993481632322073, + 0.004478610586374998, + 0.010818865150213242, + 0.010325098410248756, + 0.0077625117264688015, + 0.00698874844238162, + 0.0064852493815124035, + 0.012098263949155807, + 0.006353927310556173, + 0.009035518392920494, + 0.008133269846439362, + 0.002541422378271818, + 0.00166860013268888, + 0.0015970434760674834, + 0.0016431263647973537, + 0.002122710458934307, + 0.0027431566268205643, + 0.0037596896290779114, + 0.0025426079519093037, + 0.0031420583836734295, + 0.002904816996306181, + 0.004089741036295891, + 0.01769903674721718, + 0.005565440282225609, + 0.0539846308529377, + 0.005763829220086336, + 0.039403948932886124, + 0.002494438085705042, + 0.003218196565285325, + 0.0059338645078241825, + 0.004139901138842106, + 0.004043189343065023, + 0.007149641867727041, + 0.00697200046852231, + 0.006165690720081329, + 0.0014253678964450955, + 0.0016584339318796992, + 0.002119787037372589, + 0.0012475881958380342, + 0.0007878291071392596, + 0.001361469505354762, + 0.001746625523082912, + 0.0020281097386032343, + 0.004528536461293697, + 0.005848631262779236, + 0.029257049784064293, + 0.005334069486707449, + 0.0038122469559311867, + 0.003637586021795869, + 0.030870070680975914, + 0.008024372160434723, + 0.03544997423887253, + 0.06553550064563751, + 0.06145430728793144, + 0.02656658925116062, + 0.0761040449142456, + 0.03125043958425522, + 0.08176888525485992, + 0.03804924339056015, + 0.0063598607666790485, + 0.005757027305662632, + 0.03981456533074379, + 0.008829875849187374, + 0.012506307102739811, + 0.007557891309261322, + 0.047750074416399, + 0.012499206699430943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.006753003224730492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637676291167736 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637676291167736, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637676291167736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04980926588177681, + 0.02588396519422531, + 0.07951243966817856, + 0.020346565172076225, + 0.16484211385250092, + 0.04158271104097366, + 0.01953108049929142, + 0.12866966426372528, + 0.32712826132774353, + 0.14269395172595978 + ], + "credit_map": [ + 0.006513054482638836, + 0.005495324265211821, + 0.006458966061472893, + 0.004900939762592316, + 0.006447929423302412, + 0.004837241023778915, + 0.0054106330499053, + 0.009745176881551743, + 0.0032856494653970003, + 0.0025854669511318207, + 0.0021326676942408085, + 0.00397301884368062, + 0.002713684691116214, + 0.002241977257654071, + 0.0029192969668656588, + 0.0060322037898004055, + 0.009084499441087246, + 0.008510761894285679, + 0.008002573624253273, + 0.006586845498532057, + 0.018649432808160782, + 0.007253933232277632, + 0.013228456489741802, + 0.008195934817194939, + 0.001685678493231535, + 0.0014785828534513712, + 0.0028611558955162764, + 0.0014323870418593287, + 0.0023392685689032078, + 0.002868517069146037, + 0.003507749643176794, + 0.00417322525754571, + 0.012606366537511349, + 0.006880572997033596, + 0.01225875411182642, + 0.050489604473114014, + 0.00984587986022234, + 0.025875987485051155, + 0.012451420538127422, + 0.03443353623151779, + 0.0023719158489257097, + 0.003430909477174282, + 0.006286852993071079, + 0.004340149462223053, + 0.0068998741917312145, + 0.005293024703860283, + 0.007527939975261688, + 0.005432044621556997, + 0.0015164146898314357, + 0.0023655795957893133, + 0.0023014377802610397, + 0.0015756723005324602, + 0.0017298995517194271, + 0.0035580075345933437, + 0.003181294072419405, + 0.0033027746248990297, + 0.005927225109189749, + 0.007741900160908699, + 0.042528484016656876, + 0.007203579880297184, + 0.005394287873059511, + 0.005148440133780241, + 0.04360583797097206, + 0.011119917035102844, + 0.028728436678647995, + 0.052445679903030396, + 0.046349719166755676, + 0.02147379331290722, + 0.06110622361302376, + 0.024206383153796196, + 0.062263812869787216, + 0.030554194003343582, + 0.005905691534280777, + 0.005102654453366995, + 0.044335369020700455, + 0.008756269700825214, + 0.009459932334721088, + 0.007065434008836746, + 0.04968474060297012, + 0.012383862398564816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.004637676291167736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048634566366672516 + ], + "opposite_logprob_tokens": [ + -3.0486345291137695 + ], + "logprob_reward": -0.048634566366672516, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048634566366672516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058380551636219025, + 0.03360886499285698, + 0.04607158899307251, + 0.0617070235311985, + 0.12923485040664673, + 0.07224654406309128, + 0.02945450320839882, + 0.22233033180236816, + 0.22120891511440277, + 0.1257568597793579 + ], + "credit_map": [ + 0.006987891625612974, + 0.0062989722937345505, + 0.012497121468186378, + 0.006460996810346842, + 0.006337325554341078, + 0.005157969892024994, + 0.006765040569007397, + 0.007875227369368076, + 0.0038269245997071266, + 0.0028107305988669395, + 0.003152872435748577, + 0.0038704865146428347, + 0.007457391358911991, + 0.0027158698067069054, + 0.00625254912301898, + 0.0035220393911004066, + 0.0062364237383008, + 0.005070648156106472, + 0.005205185618251562, + 0.004215584136545658, + 0.007434824947267771, + 0.005422359332442284, + 0.0062677254900336266, + 0.006218831054866314, + 0.008326499722898006, + 0.005926699377596378, + 0.009991123341023922, + 0.0050648292526602745, + 0.004378749523311853, + 0.0053933593444526196, + 0.008901186287403107, + 0.013724570162594318, + 0.010234341956675053, + 0.005783967673778534, + 0.010739092715084553, + 0.03507484495639801, + 0.008976642973721027, + 0.029135504737496376, + 0.010639458894729614, + 0.018650975078344345, + 0.005019864533096552, + 0.005980056710541248, + 0.011714089661836624, + 0.006807425990700722, + 0.009356515482068062, + 0.013484656810760498, + 0.011167609132826328, + 0.008716316893696785, + 0.0023815552704036236, + 0.003539095167070627, + 0.0031231623142957687, + 0.002929758047685027, + 0.0024063759483397007, + 0.005115301348268986, + 0.004188103135675192, + 0.005771149881184101, + 0.010300671681761742, + 0.013424666598439217, + 0.07302772998809814, + 0.013189461082220078, + 0.009378463961184025, + 0.009963013231754303, + 0.073699451982975, + 0.019346853718161583, + 0.014286358840763569, + 0.029515547677874565, + 0.04245389625430107, + 0.01234443113207817, + 0.03455524146556854, + 0.01446473691612482, + 0.054556019604206085, + 0.0190326739102602, + 0.011320220306515694, + 0.012828119099140167, + 0.017439022660255432, + 0.015383826568722725, + 0.026350077241659164, + 0.01295178011059761, + 0.015186309814453125, + 0.014297490008175373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.048634566366672516 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom chopper motorcycle with a low-slung black hardtail frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the motorcycle's fuel tank painted a deep Candy Apple Red with gold leaf pinstripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a massive chrome V-twin engine with polished cooling fins and brass hardware?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a slow, deliberate low-angle tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a sun-drenched outdoor motorcycle show with a blurred crowd and vendor tents under a clear blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336938351392746 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.0336938351392746, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336938351392746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031508609652519226, + 0.04394088685512543, + 0.022454097867012024, + 0.011629444546997547, + 0.018370138481259346, + 0.013752730563282967, + 0.012548508122563362, + 0.015766235068440437, + 0.625831127166748, + 0.20419824123382568 + ], + "credit_map": [ + 0.0035424616653472185, + 0.003319414332509041, + 0.0037961124908179045, + 0.003275177674368024, + 0.004473888780921698, + 0.003096900414675474, + 0.0036527460906654596, + 0.006351910065859556, + 0.0075396145693957806, + 0.005823501385748386, + 0.004748216830193996, + 0.005244626197963953, + 0.004525364842265844, + 0.003353957086801529, + 0.0059790294617414474, + 0.006726579740643501, + 0.001936045940965414, + 0.0020439091604202986, + 0.00304635101929307, + 0.0025678493548184633, + 0.00431338744238019, + 0.001995619386434555, + 0.003764751832932234, + 0.002786184661090374, + 0.0011375312460586429, + 0.0012713490286841989, + 0.0012909919023513794, + 0.0012668733252212405, + 0.0011356595205143094, + 0.001784862601198256, + 0.002158343093469739, + 0.0015838340623304248, + 0.0011449281591922045, + 0.00076540611917153, + 0.0012557414593175054, + 0.0029761693440377712, + 0.0012821287382394075, + 0.0032069783192127943, + 0.001472107949666679, + 0.0062666768208146095, + 0.0008509788895025849, + 0.0012090790551155806, + 0.0014431572053581476, + 0.001504614483565092, + 0.00240387674421072, + 0.002376336371526122, + 0.0016653828788548708, + 0.002299304585903883, + 0.0013023242354393005, + 0.001913838554173708, + 0.0017473280895501375, + 0.0013817467261105776, + 0.0009546038927510381, + 0.0022598374634981155, + 0.0013828931841999292, + 0.0016059366753324866, + 0.0014554510125890374, + 0.0020885516423732042, + 0.003761870786547661, + 0.0017591897631064057, + 0.0012689467985183, + 0.0017995028756558895, + 0.0022922896314412355, + 0.001340433838777244, + 0.05008109658956528, + 0.09417630732059479, + 0.09792225062847137, + 0.040280863642692566, + 0.10851528495550156, + 0.04572466015815735, + 0.12996776401996613, + 0.059162888675928116, + 0.007511184550821781, + 0.006636882200837135, + 0.06307242065668106, + 0.012025941163301468, + 0.013499557040631771, + 0.008504804223775864, + 0.07370812445878983, + 0.01923932507634163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.0336938351392746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_1/index_tensor([9023, 9023])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_1/index_tensor([9023, 9023])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..96c65ce24516718cbef35add23070857d916671c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_1/index_tensor([9023, 9023])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extremely low-angle wide-angle shot captures the raw power of a massive steam locomotive as it roars relentlessly forward along a straight track, directly approaching the viewer's perspective. The train's massive form is rendered as a stark, imposing black silhouette, dramatically illuminated by a blazing sunset that fills the horizon with gradients of fiery red, crimson, and deep purple. The intense light from behind perfectly outlines the locomotive's massive boiler, tall smokestack, and the sharp, industrial edges of its heavy metal structure. A single, piercingly bright headlight in front of the engine cuts through the deepening dusk, casting a concentrated, brilliant white beam straight down the gleaming steel rails. Thick plumes of black smoke and white steam billow violently from the smokestack, catching the last rays of sunlight as they pour into the cool night air. The massive driving wheels are a blur of unstoppable motion, conveying incredible speed and force, creating a breathtaking and tense atmosphere of raw mechanical power against the vast, empty landscape.", + "seed": 382, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_1/index_9023_seed_382_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10323317935690283, + "vlm_logprob": -0.10323317935690283, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.10323317935690283, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004208636935800314 + ], + "opposite_logprob_tokens": [ + -6.379208564758301 + ], + "logprob_reward": -0.004208636935800314, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.004208636935800314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10019197314977646, + 0.07704368233680725, + 0.04060273990035057, + 0.06599973887205124, + 0.11707691848278046, + 0.1349533051252365, + 0.2455533891916275, + 0.08183181285858154, + 0.044185902923345566, + 0.09256050735712051 + ], + "credit_map": [ + 0.012294832617044449, + 0.013156067579984665, + 0.015499113127589226, + 0.015675906091928482, + 0.008887158706784248, + 0.010857127606868744, + 0.010958576574921608, + 0.01286319550126791, + 0.009910994209349155, + 0.01239749789237976, + 0.009383026510477066, + 0.011356930248439312, + 0.007646716665476561, + 0.008941550739109516, + 0.00979396142065525, + 0.007613007910549641, + 0.0037769705522805452, + 0.0056836288422346115, + 0.005312812980264425, + 0.005929457023739815, + 0.006814175751060247, + 0.0036029040347784758, + 0.0053986962884664536, + 0.004084092099219561, + 0.009769312106072903, + 0.005811059847474098, + 0.004551937337964773, + 0.005939091090112925, + 0.00937571283429861, + 0.019371991977095604, + 0.0074601611122488976, + 0.00372047140263021, + 0.012836981564760208, + 0.020040348172187805, + 0.012723727151751518, + 0.01854669116437435, + 0.014295280911028385, + 0.013020732440054417, + 0.014261523261666298, + 0.011351635679602623, + 0.019371651113033295, + 0.019332071766257286, + 0.012212220579385757, + 0.013809246011078358, + 0.0172889307141304, + 0.02524632401764393, + 0.012817752547562122, + 0.014875108376145363, + 0.057793471962213516, + 0.012189115397632122, + 0.019284376874566078, + 0.013620077632367611, + 0.01978745497763157, + 0.017697911709547043, + 0.019444387406110764, + 0.08573660254478455, + 0.008385516703128815, + 0.008144425228238106, + 0.007468376308679581, + 0.007973124273121357, + 0.008773721754550934, + 0.017719129100441933, + 0.011151334270834923, + 0.012216187082231045, + 0.004211405757814646, + 0.009394492022693157, + 0.00508256908506155, + 0.005043479148298502, + 0.005126073956489563, + 0.0059105223044753075, + 0.004744437988847494, + 0.004672922659665346, + 0.008500603958964348, + 0.00952509418129921, + 0.00924534723162651, + 0.012747404165565968, + 0.011606846936047077, + 0.014660283923149109, + 0.015010838396847248, + 0.011264094151556492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.004208636935800314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009802302345633507 + ], + "opposite_logprob_tokens": [ + -4.634802341461182 + ], + "logprob_reward": -0.009802302345633507, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009802302345633507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05000242590904236, + 0.10615437477827072, + 0.0957895964384079, + 0.05092858895659447, + 0.11861299723386765, + 0.05802644044160843, + 0.24550306797027588, + 0.04685990884900093, + 0.08936240524053574, + 0.13876017928123474 + ], + "credit_map": [ + 0.005774916615337133, + 0.006794811692088842, + 0.006583810783922672, + 0.006482246331870556, + 0.006410532630980015, + 0.0044336989521980286, + 0.0047557419165968895, + 0.008766666986048222, + 0.011165831238031387, + 0.016644246876239777, + 0.013943835161626339, + 0.01868121698498726, + 0.012227325700223446, + 0.011939341202378273, + 0.01478426344692707, + 0.006768313702195883, + 0.008552879095077515, + 0.01669994741678238, + 0.011920791119337082, + 0.015013828873634338, + 0.014917198568582535, + 0.008051666431128979, + 0.012214278802275658, + 0.008419002406299114, + 0.007507868576794863, + 0.005753704346716404, + 0.005111632868647575, + 0.00604309281334281, + 0.005516469478607178, + 0.011757742613554, + 0.005324123427271843, + 0.003913953434675932, + 0.0129073616117239, + 0.012441979721188545, + 0.01214938797056675, + 0.016588859260082245, + 0.019989266991615295, + 0.015736356377601624, + 0.014598824083805084, + 0.014200963079929352, + 0.007235981058329344, + 0.0066675287671387196, + 0.007672903593629599, + 0.00562615692615509, + 0.008847801014780998, + 0.009135139174759388, + 0.005952610168606043, + 0.006888318806886673, + 0.027664972469210625, + 0.03561641648411751, + 0.02400103397667408, + 0.026735765859484673, + 0.03985683247447014, + 0.01692923903465271, + 0.02796856313943863, + 0.04673023894429207, + 0.004866361152380705, + 0.007018279284238815, + 0.0052024973556399345, + 0.005264553241431713, + 0.005693353712558746, + 0.005080929026007652, + 0.006719021126627922, + 0.00701491255313158, + 0.0087372912093997, + 0.00912169087678194, + 0.009473664686083794, + 0.010912483558058739, + 0.01094207912683487, + 0.01536120381206274, + 0.01496761292219162, + 0.009846384637057781, + 0.011845030821859837, + 0.014898969791829586, + 0.015855778008699417, + 0.017057906836271286, + 0.01513019111007452, + 0.020077168941497803, + 0.02664845436811447, + 0.017246684059500694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.009802302345633507 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201516181230545 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201516181230545, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201516181230545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08439686894416809, + 0.09963972866535187, + 0.054901737719774246, + 0.05880965664982796, + 0.1605900377035141, + 0.1523781269788742, + 0.21545757353305817, + 0.09670238941907883, + 0.027636496350169182, + 0.04948735237121582 + ], + "credit_map": [ + 0.011461361311376095, + 0.010317519307136536, + 0.010738939046859741, + 0.013445088639855385, + 0.010883786715567112, + 0.008882217109203339, + 0.00809292308986187, + 0.01057503279298544, + 0.013075272552669048, + 0.01329705398529768, + 0.013568561524152756, + 0.01696104370057583, + 0.012052117846906185, + 0.007967443205416203, + 0.013199182227253914, + 0.009519054554402828, + 0.005358115304261446, + 0.008436759933829308, + 0.008224236778914928, + 0.00794830359518528, + 0.006924684625118971, + 0.004570413380861282, + 0.007341009099036455, + 0.006098215468227863, + 0.008474418893456459, + 0.00432199239730835, + 0.008971360512077808, + 0.005101117305457592, + 0.006427147425711155, + 0.015246210619807243, + 0.0053591039031744, + 0.004908308852463961, + 0.019443301483988762, + 0.032058022916316986, + 0.01679481379687786, + 0.015154005959630013, + 0.017307842150330544, + 0.02488831989467144, + 0.01777002587914467, + 0.017173700034618378, + 0.022618846967816353, + 0.02240077778697014, + 0.023918090388178825, + 0.021595289930701256, + 0.018096299842000008, + 0.01489284634590149, + 0.016241850331425667, + 0.01261412538588047, + 0.05368252471089363, + 0.015370775945484638, + 0.013224481604993343, + 0.013423432596027851, + 0.016962124034762383, + 0.016359824687242508, + 0.018379319459199905, + 0.06805510073900223, + 0.007808629889041185, + 0.020841391757130623, + 0.009150059893727303, + 0.011925450526177883, + 0.014446287415921688, + 0.007466723211109638, + 0.013456706888973713, + 0.011607134714722633, + 0.003109734272584319, + 0.0040687378495931625, + 0.003162308596074581, + 0.0032193593215197325, + 0.0036610569804906845, + 0.003973581362515688, + 0.003638867288827896, + 0.002802849980071187, + 0.005315996240824461, + 0.008487672545015812, + 0.005254260264337063, + 0.005226518027484417, + 0.0067811687476933, + 0.00666421465575695, + 0.006266605108976364, + 0.005490916781127453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.002201516181230545 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291339248418808 + ], + "opposite_logprob_tokens": [ + -2.2379133701324463 + ], + "logprob_reward": -0.11291339248418808, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291339248418808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10911179333925247, + 0.07766304165124893, + 0.04664634168148041, + 0.06981740891933441, + 0.1637386977672577, + 0.06666261702775955, + 0.22012585401535034, + 0.048185113817453384, + 0.09940310567617416, + 0.09864604473114014 + ], + "credit_map": [ + 0.01245237234979868, + 0.014231179840862751, + 0.01691836304962635, + 0.014914889819920063, + 0.007744537200778723, + 0.011673757806420326, + 0.009748755022883415, + 0.021427933126688004, + 0.0065191215835511684, + 0.01371712051331997, + 0.00925177987664938, + 0.013957373797893524, + 0.010602377355098724, + 0.006504570133984089, + 0.010971145704388618, + 0.00613955594599247, + 0.005829315632581711, + 0.006235102191567421, + 0.006310891360044479, + 0.007470674812793732, + 0.006206077989190817, + 0.004679008387029171, + 0.0054406700655817986, + 0.004474600777029991, + 0.007826484739780426, + 0.006716375704854727, + 0.009376510046422482, + 0.009703565388917923, + 0.0118379732593894, + 0.009986777789890766, + 0.008057341910898685, + 0.006312382873147726, + 0.018629012629389763, + 0.029896330088377, + 0.014852874912321568, + 0.026556065306067467, + 0.016544274985790253, + 0.01996113918721676, + 0.0189712792634964, + 0.01832771860063076, + 0.010140953585505486, + 0.0079668452963233, + 0.009916954673826694, + 0.00990824494510889, + 0.007621018681675196, + 0.006482122000306845, + 0.007670809514820576, + 0.006955669727176428, + 0.0337563157081604, + 0.02012920007109642, + 0.018515007570385933, + 0.01916719600558281, + 0.02679583430290222, + 0.016386142000555992, + 0.020770341157913208, + 0.06460582464933395, + 0.004935909528285265, + 0.006783320568501949, + 0.006706962361931801, + 0.0055963583290576935, + 0.005933656822890043, + 0.005037298426032066, + 0.006301039829850197, + 0.00689056608825922, + 0.009654116816818714, + 0.014584378339350224, + 0.010451030917465687, + 0.009878499433398247, + 0.010936781764030457, + 0.015849480405449867, + 0.015267184004187584, + 0.012781637720763683, + 0.009385420009493828, + 0.014225373975932598, + 0.010369591414928436, + 0.011604204773902893, + 0.012599322013556957, + 0.012932959012687206, + 0.01581811159849167, + 0.011711062863469124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.11291339248418808 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38704004883766174 + ], + "opposite_logprob_tokens": [ + -1.1370400190353394 + ], + "logprob_reward": -0.38704004883766174, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38704004883766174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09704776853322983, + 0.04310733079910278, + 0.05718911811709404, + 0.07138849049806595, + 0.36585527658462524, + 0.04899172484874725, + 0.06005160138010979, + 0.07037600129842758, + 0.14212657511234283, + 0.04386608302593231 + ], + "credit_map": [ + 0.012133986689150333, + 0.0154579421505332, + 0.013555982150137424, + 0.013441333547234535, + 0.00915502104908228, + 0.00982905738055706, + 0.008573855273425579, + 0.014900592155754566, + 0.0049267057329416275, + 0.005136354360729456, + 0.0061612422578036785, + 0.007397590205073357, + 0.005031357053667307, + 0.004940108396112919, + 0.005779882427304983, + 0.0037340850103646517, + 0.004067763686180115, + 0.010158286429941654, + 0.0062155043706297874, + 0.009124316275119781, + 0.009199610911309719, + 0.004780401475727558, + 0.008196053095161915, + 0.0054471855983138084, + 0.011126713827252388, + 0.006146942265331745, + 0.004588990472257137, + 0.0061917295679450035, + 0.008201022632420063, + 0.023004377260804176, + 0.00765783479437232, + 0.004470880143344402, + 0.0528486967086792, + 0.0789601281285286, + 0.028101898729801178, + 0.05378498136997223, + 0.03597258776426315, + 0.04675373435020447, + 0.04162736237049103, + 0.02780590020120144, + 0.004278834443539381, + 0.006829768884927034, + 0.005462888162583113, + 0.005776836536824703, + 0.006800956092774868, + 0.008553705178201199, + 0.005495216231793165, + 0.005793518852442503, + 0.005690025165677071, + 0.007212773896753788, + 0.007888436317443848, + 0.00785614550113678, + 0.009952347725629807, + 0.0050355191342532635, + 0.007736447267234325, + 0.008679906837642193, + 0.0061090243980288506, + 0.011129590682685375, + 0.007294125854969025, + 0.009647042490541935, + 0.010510668158531189, + 0.003960276488214731, + 0.01073378138244152, + 0.010991495102643967, + 0.012557050213217735, + 0.017494970932602882, + 0.014257479459047318, + 0.0185291338711977, + 0.015260916203260422, + 0.02608850598335266, + 0.0210266076028347, + 0.016911914572119713, + 0.0038104017730802298, + 0.005747051909565926, + 0.005554206669330597, + 0.005920362658798695, + 0.007132128346711397, + 0.006199388764798641, + 0.003519917605444789, + 0.005982627626508474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.38704004883766174 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extremely low-angle wide-angle shot captures the raw power of a massive steam locomotive as it roars relentlessly forward along a straight track, directly approaching the viewer's perspective. The train's massive form is rendered as a stark, imposing black silhouette, dramatically illuminated by a blazing sunset that fills the horizon with gradients of fiery red, crimson, and deep purple. The intense light from behind perfectly outlines the locomotive's massive boiler, tall smokestack, and the sharp, industrial edges of its heavy metal structure. A single, piercingly bright headlight in front of the engine cuts through the deepening dusk, casting a concentrated, brilliant white beam straight down the gleaming steel rails. Thick plumes of black smoke and white steam billow violently from the smokestack, catching the last rays of sunlight as they pour into the cool night air. The massive driving wheels are a blur of unstoppable motion, conveying incredible speed and force, creating a breathtaking and tense atmosphere of raw mechanical power against the vast, empty landscape.", + "seed": 383, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_1/index_9023_seed_383_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07338304091244936, + "vlm_logprob": -0.07338304091244936, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.07338304091244936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03928864747285843 + ], + "opposite_logprob_tokens": [ + -3.2892887592315674 + ], + "logprob_reward": -0.03928864747285843, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03928864747285843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10659080743789673, + 0.043282002210617065, + 0.03993472829461098, + 0.1839735358953476, + 0.029046310111880302, + 0.07947976142168045, + 0.10037834942340851, + 0.09611871838569641, + 0.19550207257270813, + 0.1256936937570572 + ], + "credit_map": [ + 0.013030965812504292, + 0.013197350315749645, + 0.016590749844908714, + 0.010215874761343002, + 0.015285742469131947, + 0.013482928276062012, + 0.00756514398381114, + 0.017222048714756966, + 0.004371002782136202, + 0.0049535478465259075, + 0.007588827516883612, + 0.006078528705984354, + 0.004826374351978302, + 0.005393337924033403, + 0.0029885016847401857, + 0.00708188209682703, + 0.007243430707603693, + 0.005126988515257835, + 0.005886706057935953, + 0.003300743643194437, + 0.004910200834274292, + 0.006248243618756533, + 0.002549846889451146, + 0.004668569192290306, + 0.04953719303011894, + 0.01678289659321308, + 0.028232451528310776, + 0.012206082232296467, + 0.015537266619503498, + 0.01901882328093052, + 0.007547024637460709, + 0.03511178866028786, + 0.004172442015260458, + 0.005943229887634516, + 0.004643104039132595, + 0.0018722879467532039, + 0.0031933635473251343, + 0.004117956850677729, + 0.0013039595214650035, + 0.0037999670021235943, + 0.00879234541207552, + 0.009495733305811882, + 0.010242775082588196, + 0.010249105282127857, + 0.013841480948030949, + 0.010264504700899124, + 0.007440563756972551, + 0.009153255261480808, + 0.0126231350004673, + 0.015028683468699455, + 0.012512287124991417, + 0.012873928993940353, + 0.017066214233636856, + 0.011895867995917797, + 0.00835021585226059, + 0.010028012096881866, + 0.020360078662633896, + 0.010549642145633698, + 0.014617729932069778, + 0.008392243646085262, + 0.009378204122185707, + 0.01456949021667242, + 0.004986914340406656, + 0.013264419510960579, + 0.02056799829006195, + 0.021421855315566063, + 0.02937410958111286, + 0.025797557085752487, + 0.025249378755688667, + 0.025545688346028328, + 0.017608772963285446, + 0.029936717823147774, + 0.009657147340476513, + 0.014369010925292969, + 0.010900747962296009, + 0.01974978670477867, + 0.01984364725649357, + 0.01652112789452076, + 0.01929376646876335, + 0.015358458273112774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.03928864747285843 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010509926360100508 + ], + "opposite_logprob_tokens": [ + -6.87605094909668 + ], + "logprob_reward": -0.0010509926360100508, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010509926360100508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10172276198863983, + 0.15009434521198273, + 0.09352981299161911, + 0.039859574288129807, + 0.05168271064758301, + 0.09658639878034592, + 0.07820692658424377, + 0.08384325355291367, + 0.11002635955810547, + 0.19444790482521057 + ], + "credit_map": [ + 0.013035527430474758, + 0.013136194087564945, + 0.017196258530020714, + 0.009593620896339417, + 0.008027439005672932, + 0.012032028287649155, + 0.008614806458353996, + 0.02008689008653164, + 0.022779108956456184, + 0.02352224849164486, + 0.02079085446894169, + 0.01483615767210722, + 0.019855882972478867, + 0.01650897227227688, + 0.010610159486532211, + 0.021190965548157692, + 0.020827854052186012, + 0.015214028768241405, + 0.012029443867504597, + 0.006078757345676422, + 0.008528322912752628, + 0.012554286979138851, + 0.005104889627546072, + 0.013192233629524708, + 0.008562175557017326, + 0.003556290175765753, + 0.008394586853682995, + 0.0029939021915197372, + 0.005600868724286556, + 0.0027523478493094444, + 0.0025541121140122414, + 0.005445292219519615, + 0.007344869431108236, + 0.008146725594997406, + 0.008440122939646244, + 0.005534794647246599, + 0.006146798841655254, + 0.0068971384316682816, + 0.0029042111709713936, + 0.006268047261983156, + 0.01195904053747654, + 0.00936855562031269, + 0.013550219126045704, + 0.014771643094718456, + 0.013427279889583588, + 0.012296764180064201, + 0.008940675295889378, + 0.012272217310965061, + 0.011283072642982006, + 0.009053652174770832, + 0.011693986132740974, + 0.0099942060187459, + 0.011416386812925339, + 0.009192057885229588, + 0.00640675937756896, + 0.009166805073618889, + 0.017284492030739784, + 0.008905495516955853, + 0.01290698442608118, + 0.007816843688488007, + 0.007291666232049465, + 0.013095616362988949, + 0.006626063957810402, + 0.009916087612509727, + 0.017329178750514984, + 0.014393935911357403, + 0.020536838099360466, + 0.009663005359470844, + 0.00928304623812437, + 0.020718565210700035, + 0.005501171108335257, + 0.012600621208548546, + 0.018956439569592476, + 0.024338452145457268, + 0.01921161450445652, + 0.024615595117211342, + 0.02839807979762554, + 0.033667974174022675, + 0.02216419391334057, + 0.023095564916729927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.0010509926360100508 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002202229807153344 + ], + "opposite_logprob_tokens": [ + -6.127202033996582 + ], + "logprob_reward": -0.002202229807153344, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002202229807153344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1291041225194931, + 0.20118868350982666, + 0.07874850928783417, + 0.20299683511257172, + 0.048832353204488754, + 0.02636021003127098, + 0.03900785744190216, + 0.04535245522856712, + 0.044507645070552826, + 0.1839013695716858 + ], + "credit_map": [ + 0.021387724205851555, + 0.012120116502046585, + 0.03177551552653313, + 0.00997005682438612, + 0.00740289781242609, + 0.015822917222976685, + 0.005119809415191412, + 0.025505075231194496, + 0.034431472420692444, + 0.03516789898276329, + 0.03495066985487938, + 0.018588313832879066, + 0.017161229625344276, + 0.023778917267918587, + 0.008145076222717762, + 0.02896510437130928, + 0.021910343319177628, + 0.010988705791532993, + 0.0112695824354887, + 0.004643306136131287, + 0.005746209528297186, + 0.00883826520293951, + 0.0031639691442251205, + 0.01218812819570303, + 0.05561324954032898, + 0.01775968261063099, + 0.033293336629867554, + 0.012396540492773056, + 0.011772584170103073, + 0.02023213542997837, + 0.007518349215388298, + 0.04441095516085625, + 0.008588716387748718, + 0.004200157709419727, + 0.009696640074253082, + 0.0038559201639145613, + 0.004876083228737116, + 0.00437106704339385, + 0.0029903054237365723, + 0.010253466665744781, + 0.0035488493740558624, + 0.002831578953191638, + 0.004814657848328352, + 0.002588334959000349, + 0.002739277435466647, + 0.00419284962117672, + 0.0014274825807660818, + 0.004217179026454687, + 0.004957403987646103, + 0.00425791647285223, + 0.0045343125239014626, + 0.0049496544525027275, + 0.006453889887779951, + 0.003707283176481724, + 0.0035008909180760384, + 0.006646506953984499, + 0.005417074542492628, + 0.005361846182495356, + 0.006118634715676308, + 0.0069428994320333, + 0.005754595156759024, + 0.005828141700476408, + 0.00390091841109097, + 0.006028346251696348, + 0.004976646974682808, + 0.006126519292593002, + 0.006392895709723234, + 0.005117580760270357, + 0.0049895173870027065, + 0.008129755035042763, + 0.0028196671046316624, + 0.00595506839454174, + 0.020813483744859695, + 0.0200624018907547, + 0.01799365133047104, + 0.02783002145588398, + 0.029510200023651123, + 0.021298939362168312, + 0.024216080084443092, + 0.0221765898168087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.002202229807153344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305586218833923 + ], + "opposite_logprob_tokens": [ + -3.168055772781372 + ], + "logprob_reward": -0.04305586218833923, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305586218833923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0539604052901268, + 0.06903848052024841, + 0.03549780696630478, + 0.06894456595182419, + 0.02478320710361004, + 0.021205959841609, + 0.055206410586833954, + 0.08861761540174484, + 0.08185818791389465, + 0.5008873343467712 + ], + "credit_map": [ + 0.009306223131716251, + 0.01000240445137024, + 0.007920070551335812, + 0.005102802999317646, + 0.004143062513321638, + 0.0066461749374866486, + 0.003854577662423253, + 0.006985094398260117, + 0.010256117209792137, + 0.014511818997561932, + 0.00876810122281313, + 0.00619889423251152, + 0.006963267456740141, + 0.009977412410080433, + 0.0035895495675504208, + 0.008773325942456722, + 0.007094120141118765, + 0.007053819485008717, + 0.004369976464658976, + 0.002775319619104266, + 0.003877498907968402, + 0.004799054469913244, + 0.0018299005459994078, + 0.003698117798194289, + 0.01287668477743864, + 0.006797959562391043, + 0.01013623084872961, + 0.00582778500393033, + 0.00981121975928545, + 0.009015973657369614, + 0.0049268403090536594, + 0.009551872499287128, + 0.0034753340296447277, + 0.003944948315620422, + 0.0031418250873684883, + 0.0024176626466214657, + 0.0031446681823581457, + 0.003586005186662078, + 0.0014530635671690106, + 0.003619701834395528, + 0.0028790398500859737, + 0.002108173444867134, + 0.002556517021730542, + 0.0020558726973831654, + 0.004212753847241402, + 0.002470399020239711, + 0.0017730799736455083, + 0.0031501250341534615, + 0.007786828093230724, + 0.006001506000757217, + 0.007828565314412117, + 0.007417954970151186, + 0.00667426735162735, + 0.006012663710862398, + 0.004108753055334091, + 0.009375879541039467, + 0.008354159072041512, + 0.008322485722601414, + 0.013081308454275131, + 0.012003638781607151, + 0.009240430779755116, + 0.012543866410851479, + 0.007421406451612711, + 0.01765032485127449, + 0.00682849669829011, + 0.012095032259821892, + 0.008224144577980042, + 0.009929127991199493, + 0.012801350094377995, + 0.016280312091112137, + 0.005263552535325289, + 0.010436175391077995, + 0.055368177592754364, + 0.05527350679039955, + 0.04743819311261177, + 0.07187105715274811, + 0.0837370976805687, + 0.06289476156234741, + 0.06606858223676682, + 0.05823599174618721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.04305586218833923 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28131747245788574 + ], + "opposite_logprob_tokens": [ + -1.4063174724578857 + ], + "logprob_reward": -0.28131747245788574, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28131747245788574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17460490763187408, + 0.14177988469600677, + 0.0742914155125618, + 0.04007143899798393, + 0.028031455352902412, + 0.08877509087324142, + 0.07712820172309875, + 0.09603714197874069, + 0.12375154346227646, + 0.1555289477109909 + ], + "credit_map": [ + 0.027059655636548996, + 0.026048174127936363, + 0.02594843879342079, + 0.017318373546004295, + 0.01682215742766857, + 0.021118443459272385, + 0.013889215886592865, + 0.02640043944120407, + 0.019071096554398537, + 0.01968562789261341, + 0.022850923240184784, + 0.016821084544062614, + 0.015858493745326996, + 0.016228409484028816, + 0.012664491310715675, + 0.018599750474095345, + 0.01623232662677765, + 0.015184313990175724, + 0.009017403237521648, + 0.005075106862932444, + 0.007194889709353447, + 0.009333429858088493, + 0.0038764062337577343, + 0.008377542719244957, + 0.010179543867707253, + 0.0036151621025055647, + 0.0050829495303332806, + 0.0027632464189082384, + 0.004863332957029343, + 0.00411907397210598, + 0.002661721082404256, + 0.006786407437175512, + 0.003552333451807499, + 0.004918993450701237, + 0.0038745817728340626, + 0.002809707773849368, + 0.0036993289832025766, + 0.0039687613025307655, + 0.0020616017282009125, + 0.0031461475882679224, + 0.008894343860447407, + 0.009573018178343773, + 0.011041116900742054, + 0.012434499338269234, + 0.017169062048196793, + 0.011385328136384487, + 0.00804078858345747, + 0.010236934758722782, + 0.014386817812919617, + 0.007003337144851685, + 0.0187982190400362, + 0.005328309256583452, + 0.006127478554844856, + 0.008354616351425648, + 0.004412897862493992, + 0.01271652802824974, + 0.015934206545352936, + 0.013302847743034363, + 0.01421892549842596, + 0.008461371064186096, + 0.008455152623355389, + 0.01612597145140171, + 0.006902455352246761, + 0.0126362144947052, + 0.012239424511790276, + 0.013301624916493893, + 0.01710343360900879, + 0.017884818837046623, + 0.01990547776222229, + 0.01452147401869297, + 0.011460025794804096, + 0.017335262149572372, + 0.014983167871832848, + 0.02136840671300888, + 0.013688537292182446, + 0.01730847731232643, + 0.02220476046204567, + 0.034888096153736115, + 0.013425543904304504, + 0.01766195334494114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.28131747245788574 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extremely low-angle wide-angle shot captures the raw power of a massive steam locomotive as it roars relentlessly forward along a straight track, directly approaching the viewer's perspective. The train's massive form is rendered as a stark, imposing black silhouette, dramatically illuminated by a blazing sunset that fills the horizon with gradients of fiery red, crimson, and deep purple. The intense light from behind perfectly outlines the locomotive's massive boiler, tall smokestack, and the sharp, industrial edges of its heavy metal structure. A single, piercingly bright headlight in front of the engine cuts through the deepening dusk, casting a concentrated, brilliant white beam straight down the gleaming steel rails. Thick plumes of black smoke and white steam billow violently from the smokestack, catching the last rays of sunlight as they pour into the cool night air. The massive driving wheels are a blur of unstoppable motion, conveying incredible speed and force, creating a breathtaking and tense atmosphere of raw mechanical power against the vast, empty landscape.", + "seed": 384, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_1/index_9023_seed_384_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03771400444675237, + "vlm_logprob": -0.03771400444675237, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.03771400444675237, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01554902270436287 + ], + "opposite_logprob_tokens": [ + -4.265549182891846 + ], + "logprob_reward": -0.01554902270436287, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01554902270436287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25827279686927795, + 0.05991839990019798, + 0.05331325903534889, + 0.07704564183950424, + 0.06738922744989395, + 0.05879523977637291, + 0.09509261697530746, + 0.17721213400363922, + 0.10171651095151901, + 0.05124412104487419 + ], + "credit_map": [ + 0.025837628170847893, + 0.027476102113723755, + 0.027302199974656105, + 0.02185545675456524, + 0.042190928012132645, + 0.029363689944148064, + 0.031021786853671074, + 0.05322500318288803, + 0.008340721018612385, + 0.007293050643056631, + 0.008921374566853046, + 0.006535951513797045, + 0.006207807455211878, + 0.007859578356146812, + 0.006111242342740297, + 0.008648673072457314, + 0.007251358591020107, + 0.00735614774748683, + 0.008347596041858196, + 0.006412927992641926, + 0.005427050404250622, + 0.0055748894810676575, + 0.008047802373766899, + 0.0048954845406115055, + 0.0068703279830515385, + 0.009465282782912254, + 0.011966480873525143, + 0.010708210058510303, + 0.011049740016460419, + 0.007024737074971199, + 0.012091019190847874, + 0.0078698405995965, + 0.006720797158777714, + 0.008410451002418995, + 0.008304767310619354, + 0.010382949374616146, + 0.007791294250637293, + 0.008193666115403175, + 0.011253580451011658, + 0.006331725511699915, + 0.005121000111103058, + 0.006410254165530205, + 0.008803864941000938, + 0.008843515999615192, + 0.009125935845077038, + 0.00583876296877861, + 0.00841318629682064, + 0.006238716654479504, + 0.006261693313717842, + 0.010886036790907383, + 0.016420288011431694, + 0.014315224252641201, + 0.01653386279940605, + 0.010358041152358055, + 0.013703401200473309, + 0.0066140745766460896, + 0.013648811727762222, + 0.018163606524467468, + 0.019916659221053123, + 0.03412218391895294, + 0.02648274600505829, + 0.0177552942186594, + 0.029067188501358032, + 0.018055662512779236, + 0.01054046768695116, + 0.010991442948579788, + 0.012413746677339077, + 0.014449299313127995, + 0.01129978895187378, + 0.00986392144113779, + 0.018687404692173004, + 0.013470441102981567, + 0.004412337206304073, + 0.0068884645588696, + 0.011468570679426193, + 0.00746317021548748, + 0.004745054058730602, + 0.0053732730448246, + 0.006230837199836969, + 0.004662413615733385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.01554902270436287 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003203738247975707 + ], + "opposite_logprob_tokens": [ + -5.753203868865967 + ], + "logprob_reward": -0.003203738247975707, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003203738247975707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14805720746517181, + 0.09783951193094254, + 0.044509921222925186, + 0.1107504665851593, + 0.0684451013803482, + 0.05136460065841675, + 0.05777648836374283, + 0.12812063097953796, + 0.14177650213241577, + 0.15135960280895233 + ], + "credit_map": [ + 0.023653220385313034, + 0.022059567272663116, + 0.01333389151841402, + 0.015957847237586975, + 0.018763534724712372, + 0.021381143480539322, + 0.019901711493730545, + 0.013006286695599556, + 0.01305725984275341, + 0.007965222001075745, + 0.01171433087438345, + 0.010984426364302635, + 0.005741072818636894, + 0.014483477920293808, + 0.008496943861246109, + 0.025396784767508507, + 0.006599031388759613, + 0.0070865219458937645, + 0.007849207147955894, + 0.005918185226619244, + 0.003901214338839054, + 0.005203584674745798, + 0.004511849954724312, + 0.0034403251484036446, + 0.01054297387599945, + 0.01384028047323227, + 0.012716913595795631, + 0.019141944125294685, + 0.013425901532173157, + 0.010794434696435928, + 0.02258477918803692, + 0.007703248877078295, + 0.010578954592347145, + 0.006216163747012615, + 0.009359914809465408, + 0.006841357797384262, + 0.007603998761624098, + 0.008963217958807945, + 0.008081131614744663, + 0.01080036349594593, + 0.004366279114037752, + 0.004794461186975241, + 0.00578940799459815, + 0.008763279765844345, + 0.006032482720911503, + 0.00620897626504302, + 0.009759233333170414, + 0.0056504798121750355, + 0.003830656874924898, + 0.006436747498810291, + 0.006538885645568371, + 0.009477062150835991, + 0.011176629923284054, + 0.005789129063487053, + 0.010743974708020687, + 0.0037834018003195524, + 0.011626734398305416, + 0.012881984934210777, + 0.01721886359155178, + 0.023663129657506943, + 0.014397352933883667, + 0.01536258403211832, + 0.0186251662671566, + 0.014344818890094757, + 0.020680662244558334, + 0.01172857079654932, + 0.01448419876396656, + 0.018028683960437775, + 0.0165543369948864, + 0.019982967525720596, + 0.015743480995297432, + 0.024573596194386482, + 0.010762038640677929, + 0.014217192307114601, + 0.01806255616247654, + 0.01979674957692623, + 0.025438517332077026, + 0.014112100936472416, + 0.02215740829706192, + 0.02681303210556507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.003203738247975707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019468179671093822 + ], + "opposite_logprob_tokens": [ + -6.251946926116943 + ], + "logprob_reward": -0.0019468179671093822, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019468179671093822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.358060359954834, + 0.07192761451005936, + 0.05185689404606819, + 0.056203048676252365, + 0.034710370004177094, + 0.04291541129350662, + 0.03654993698000908, + 0.1127469465136528, + 0.10027720779180527, + 0.13475219905376434 + ], + "credit_map": [ + 0.03370554745197296, + 0.03445565328001976, + 0.0364827923476696, + 0.026217514649033546, + 0.05839836969971657, + 0.04381809011101723, + 0.051666341722011566, + 0.07331603765487671, + 0.010572965256869793, + 0.0075051444582641125, + 0.010118328966200352, + 0.00641329400241375, + 0.005379802547395229, + 0.010577385313808918, + 0.006797921843826771, + 0.014562775380909443, + 0.006060462910681963, + 0.004459415562450886, + 0.007255971897393465, + 0.004850722383707762, + 0.0027086620684713125, + 0.0067875925451517105, + 0.0050090462900698185, + 0.01472502388060093, + 0.010578020475804806, + 0.00648767314851284, + 0.007784543093293905, + 0.005630785599350929, + 0.0059722005389630795, + 0.007971036247909069, + 0.006459680385887623, + 0.005319110117852688, + 0.006576161365956068, + 0.004783017560839653, + 0.003148997202515602, + 0.00468455720692873, + 0.0026931017637252808, + 0.00396373588591814, + 0.004748230800032616, + 0.004112569149583578, + 0.005519827827811241, + 0.0037923650816082954, + 0.0034889208618551493, + 0.0071155778132379055, + 0.004230317659676075, + 0.006129675544798374, + 0.008073150180280209, + 0.004565579351037741, + 0.0038628962356597185, + 0.003480328479781747, + 0.007206220645457506, + 0.006061687134206295, + 0.0042441789992153645, + 0.004024791996926069, + 0.00452701561152935, + 0.0031428211368620396, + 0.011393898166716099, + 0.011072101071476936, + 0.015165818855166435, + 0.019899586215615273, + 0.015754109248518944, + 0.01154024712741375, + 0.016317840665578842, + 0.011603341437876225, + 0.0058425855822861195, + 0.007905960083007812, + 0.011416318826377392, + 0.021689582616090775, + 0.010239196941256523, + 0.009344683960080147, + 0.024429863318800926, + 0.00940900668501854, + 0.013402204029262066, + 0.012173653580248356, + 0.01646723411977291, + 0.02200412005186081, + 0.012275852262973785, + 0.016244130209088326, + 0.020189540460705757, + 0.021995460614562035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0019468179671093822 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495023354887962 + ], + "opposite_logprob_tokens": [ + -2.92995023727417 + ], + "logprob_reward": -0.05495023354887962, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495023354887962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17817041277885437, + 0.07633090019226074, + 0.04443226382136345, + 0.07460079342126846, + 0.08549049496650696, + 0.04141761362552643, + 0.05172732472419739, + 0.09912043064832687, + 0.3034953474998474, + 0.045214395970106125 + ], + "credit_map": [ + 0.019322074949741364, + 0.0282696932554245, + 0.021784760057926178, + 0.021452058106660843, + 0.022048626095056534, + 0.023373007774353027, + 0.02704665996134281, + 0.014873542822897434, + 0.010830639861524105, + 0.010952528566122055, + 0.008562523871660233, + 0.008485384285449982, + 0.007443148642778397, + 0.00951769482344389, + 0.009956276044249535, + 0.010582707822322845, + 0.004711804445832968, + 0.005890904925763607, + 0.005568931810557842, + 0.004605882801115513, + 0.005640123970806599, + 0.005555802024900913, + 0.006544256582856178, + 0.005914553999900818, + 0.008214681409299374, + 0.008221464231610298, + 0.007541295140981674, + 0.012686665169894695, + 0.008541451767086983, + 0.008254767395555973, + 0.015443998388946056, + 0.005696470383554697, + 0.008885223418474197, + 0.010015861131250858, + 0.008963174186646938, + 0.01200602576136589, + 0.008695072494447231, + 0.010541500523686409, + 0.011169612407684326, + 0.015214024111628532, + 0.005103590432554483, + 0.0038558212108910084, + 0.005168278235942125, + 0.006712775211781263, + 0.0033861701376736164, + 0.005773107521235943, + 0.0059507545083761215, + 0.005467114504426718, + 0.004618442617356777, + 0.005814334377646446, + 0.009098859503865242, + 0.007374473847448826, + 0.00928966049104929, + 0.006037990562617779, + 0.006245016120374203, + 0.0032485476695001125, + 0.009100163355469704, + 0.009333104826509953, + 0.013995989225804806, + 0.018487311899662018, + 0.012694736942648888, + 0.010234441608190536, + 0.013723061420023441, + 0.011551626026630402, + 0.01908918283879757, + 0.023864615708589554, + 0.025689201429486275, + 0.05843300744891167, + 0.02858683466911316, + 0.025641996413469315, + 0.08363458514213562, + 0.03855593129992485, + 0.0038516041822731495, + 0.004174055531620979, + 0.008654982782900333, + 0.0058562858030200005, + 0.006199315190315247, + 0.004911922849714756, + 0.005618099123239517, + 0.0059481291100382805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.05495023354887962 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292020976543427 + ], + "opposite_logprob_tokens": [ + -2.2379202842712402 + ], + "logprob_reward": -0.11292020976543427, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292020976543427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23304222524166107, + 0.06124258413910866, + 0.04696383699774742, + 0.06539402157068253, + 0.124672070145607, + 0.06712184846401215, + 0.06858877837657928, + 0.13486775755882263, + 0.1257047802209854, + 0.07240207493305206 + ], + "credit_map": [ + 0.038518328219652176, + 0.03691776096820831, + 0.027006778866052628, + 0.02365296706557274, + 0.020088477060198784, + 0.03634151071310043, + 0.021768197417259216, + 0.028748219832777977, + 0.006018254440277815, + 0.007324468810111284, + 0.011309485882520676, + 0.0046581607311964035, + 0.0045204805210232735, + 0.008487911894917488, + 0.006394281517714262, + 0.012529539875686169, + 0.007780439220368862, + 0.005960854701697826, + 0.00589214451611042, + 0.006329755764454603, + 0.003383068135008216, + 0.005026055965572596, + 0.006224981043487787, + 0.006366537418216467, + 0.00569388410076499, + 0.00920517835766077, + 0.006050495896488428, + 0.008902938105165958, + 0.009116534143686295, + 0.008330002427101135, + 0.010717839002609253, + 0.007377155125141144, + 0.014606855809688568, + 0.013229855336248875, + 0.011323963291943073, + 0.015828991308808327, + 0.011075066402554512, + 0.019709717482328415, + 0.014341764152050018, + 0.024555858224630356, + 0.006196524947881699, + 0.007037088740617037, + 0.008801628835499287, + 0.009868196211755276, + 0.006889731157571077, + 0.009274665266275406, + 0.010403599590063095, + 0.008650409057736397, + 0.0055771684274077415, + 0.008288628421723843, + 0.010972448624670506, + 0.010633363388478756, + 0.010012280195951462, + 0.007830488495528698, + 0.011247831396758556, + 0.004026572685688734, + 0.012207643128931522, + 0.01737738959491253, + 0.011840802617371082, + 0.0244075246155262, + 0.024193737655878067, + 0.014606871642172337, + 0.021919840946793556, + 0.008313948288559914, + 0.010263157077133656, + 0.014197844080626965, + 0.020137125626206398, + 0.027555067092180252, + 0.011389713734388351, + 0.013447966426610947, + 0.016773346811532974, + 0.01194055937230587, + 0.006696407683193684, + 0.009245886467397213, + 0.013382048346102238, + 0.009306538850069046, + 0.007612976245582104, + 0.007533682975918055, + 0.009049749001860619, + 0.009574786759912968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.11292020976543427 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extremely low-angle wide-angle shot captures the raw power of a massive steam locomotive as it roars relentlessly forward along a straight track, directly approaching the viewer's perspective. The train's massive form is rendered as a stark, imposing black silhouette, dramatically illuminated by a blazing sunset that fills the horizon with gradients of fiery red, crimson, and deep purple. The intense light from behind perfectly outlines the locomotive's massive boiler, tall smokestack, and the sharp, industrial edges of its heavy metal structure. A single, piercingly bright headlight in front of the engine cuts through the deepening dusk, casting a concentrated, brilliant white beam straight down the gleaming steel rails. Thick plumes of black smoke and white steam billow violently from the smokestack, catching the last rays of sunlight as they pour into the cool night air. The massive driving wheels are a blur of unstoppable motion, conveying incredible speed and force, creating a breathtaking and tense atmosphere of raw mechanical power against the vast, empty landscape.", + "seed": 385, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_1/index_9023_seed_385_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08165637517813593, + "vlm_logprob": -0.08165637517813593, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.08165637517813593, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006302244495600462 + ], + "opposite_logprob_tokens": [ + -5.506302356719971 + ], + "logprob_reward": -0.006302244495600462, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006302244495600462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3882838189601898, + 0.1608661264181137, + 0.0961223691701889, + 0.048676278442144394, + 0.02948703244328499, + 0.061402902007102966, + 0.07641655206680298, + 0.04639153182506561, + 0.03571368381381035, + 0.05663970112800598 + ], + "credit_map": [ + 0.0436008982360363, + 0.08025944232940674, + 0.055226489901542664, + 0.07890579849481583, + 0.04476850479841232, + 0.03122548572719097, + 0.03466115891933441, + 0.01963602751493454, + 0.0209336094558239, + 0.02206491306424141, + 0.020777717232704163, + 0.02471175603568554, + 0.023455727845430374, + 0.02032688446342945, + 0.01675017736852169, + 0.011845339089632034, + 0.009984639473259449, + 0.014088273979723454, + 0.0143178291618824, + 0.016266020014882088, + 0.01336966548115015, + 0.010545413941144943, + 0.0115185072645545, + 0.006032022647559643, + 0.006891606841236353, + 0.0098599037155509, + 0.006057690363377333, + 0.008126122877001762, + 0.005050225183367729, + 0.004031486343592405, + 0.004161715041846037, + 0.004497528076171875, + 0.0035317116416990757, + 0.0038030173163861036, + 0.004705141764134169, + 0.004674089606851339, + 0.0032607074826955795, + 0.0027248214464634657, + 0.0034633942414075136, + 0.0033241494093090296, + 0.009495703503489494, + 0.007837063632905483, + 0.009058532305061817, + 0.009641353972256184, + 0.006416863761842251, + 0.0061737853102386, + 0.006233497057110071, + 0.006546105723828077, + 0.011078081093728542, + 0.009200869128108025, + 0.01088087446987629, + 0.013198541477322578, + 0.007768896874040365, + 0.009174799546599388, + 0.007913153618574142, + 0.007201328873634338, + 0.007053917273879051, + 0.006054229103028774, + 0.006817406974732876, + 0.008509061299264431, + 0.004338321276009083, + 0.005578948650509119, + 0.004402304533869028, + 0.0036373420152813196, + 0.004268219694495201, + 0.005158453714102507, + 0.0058918120339512825, + 0.0069337706081569195, + 0.0034998448099941015, + 0.003003879450261593, + 0.003912372514605522, + 0.0030453314539045095, + 0.00651350524276495, + 0.007599063217639923, + 0.008016283623874187, + 0.011042289435863495, + 0.006521489936858416, + 0.005889293737709522, + 0.006097333040088415, + 0.004960442427545786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.006302244495600462 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003202668856829405 + ], + "opposite_logprob_tokens": [ + -5.753202438354492 + ], + "logprob_reward": -0.003202668856829405, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003202668856829405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28711551427841187, + 0.1913902908563614, + 0.06996087729930878, + 0.037799302488565445, + 0.11132083088159561, + 0.09883355349302292, + 0.04824046045541763, + 0.04509454593062401, + 0.031758613884449005, + 0.07848598062992096 + ], + "credit_map": [ + 0.026217598468065262, + 0.06029027700424194, + 0.04499715194106102, + 0.054295189678668976, + 0.02784702368080616, + 0.02753639593720436, + 0.02632308192551136, + 0.019608793780207634, + 0.023024553433060646, + 0.034925878047943115, + 0.030073869973421097, + 0.0322096012532711, + 0.014418648555874825, + 0.01776966266334057, + 0.02511615678668022, + 0.013851922005414963, + 0.00996438879519701, + 0.009625502862036228, + 0.008080428466200829, + 0.010515276342630386, + 0.007628633175045252, + 0.008024044334888458, + 0.008064029738307, + 0.008058574050664902, + 0.00622998271137476, + 0.007540919352322817, + 0.004600777756422758, + 0.005797550082206726, + 0.004398387856781483, + 0.003938158042728901, + 0.0028405890334397554, + 0.002452937653288245, + 0.016271594911813736, + 0.013662269338965416, + 0.013475945219397545, + 0.01767602376639843, + 0.016164828091859818, + 0.012832875363528728, + 0.010575915686786175, + 0.01066137757152319, + 0.013307906687259674, + 0.01364994514733553, + 0.015554734505712986, + 0.015551278367638588, + 0.009896530769765377, + 0.008945743553340435, + 0.011159098707139492, + 0.010768312960863113, + 0.006147994194179773, + 0.005943640135228634, + 0.00680103525519371, + 0.008460469543933868, + 0.004566007759422064, + 0.006316373590379953, + 0.006185830105096102, + 0.003819105913862586, + 0.008803808130323887, + 0.005642599426209927, + 0.00473167235031724, + 0.0070227948017418385, + 0.006283175200223923, + 0.004517646040767431, + 0.004121527075767517, + 0.003971324302256107, + 0.004859660752117634, + 0.00484331464394927, + 0.00451685581356287, + 0.006001031957566738, + 0.0029217901173979044, + 0.0035514968913048506, + 0.0029684228356927633, + 0.002096042037010193, + 0.016707129776477814, + 0.015163080766797066, + 0.006891279481351376, + 0.01079215481877327, + 0.005507880821824074, + 0.0080747464671731, + 0.005206536501646042, + 0.010143164545297623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.003202668856829405 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015174552099779248 + ], + "opposite_logprob_tokens": [ + -6.501517295837402 + ], + "logprob_reward": -0.0015174552099779248, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015174552099779248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.426428884267807, + 0.13477084040641785, + 0.07989366352558136, + 0.049042027443647385, + 0.056604623794555664, + 0.05393405258655548, + 0.06273774802684784, + 0.04267073795199394, + 0.03675870969891548, + 0.05715873837471008 + ], + "credit_map": [ + 0.040410857647657394, + 0.07094486802816391, + 0.05809609219431877, + 0.07194690406322479, + 0.0420883446931839, + 0.040011048316955566, + 0.04355984926223755, + 0.05937090516090393, + 0.019836869090795517, + 0.027706962078809738, + 0.016913391649723053, + 0.02035476826131344, + 0.012228590436279774, + 0.013219223357737064, + 0.010601353831589222, + 0.013909676112234592, + 0.011054592207074165, + 0.014078369364142418, + 0.007999644614756107, + 0.011842580512166023, + 0.012239376083016396, + 0.006878437474370003, + 0.007420916110277176, + 0.00837975088506937, + 0.008029624819755554, + 0.007846323773264885, + 0.006107914261519909, + 0.008717553690075874, + 0.004307515919208527, + 0.004968374967575073, + 0.004719193559139967, + 0.004345526918768883, + 0.006473212037235498, + 0.004972411785274744, + 0.006360411178320646, + 0.007195235695689917, + 0.004588402342051268, + 0.00846317782998085, + 0.005825487431138754, + 0.012726286426186562, + 0.006375982891768217, + 0.006300924811512232, + 0.008235890418291092, + 0.009545156732201576, + 0.007149116601794958, + 0.005321784410625696, + 0.005200391169637442, + 0.005804805085062981, + 0.006857713218778372, + 0.007426855154335499, + 0.009014376439154148, + 0.01014364417642355, + 0.0061534252017736435, + 0.007630449719727039, + 0.006620882544666529, + 0.008890405297279358, + 0.0063279131427407265, + 0.007514272350817919, + 0.005115607287734747, + 0.007731559686362743, + 0.004889774601906538, + 0.00424899673089385, + 0.003842466976493597, + 0.003000151365995407, + 0.005226156674325466, + 0.007912811823189259, + 0.004801704082638025, + 0.006323809735476971, + 0.0029190750792622566, + 0.003353661624714732, + 0.002399793826043606, + 0.003821695689111948, + 0.007960482500493526, + 0.00933078397065401, + 0.006483265198767185, + 0.008862236514687538, + 0.007648613303899765, + 0.0059264106675982475, + 0.006120613776147366, + 0.004826327785849571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0015174552099779248 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048643652349710464 + ], + "opposite_logprob_tokens": [ + -3.0486435890197754 + ], + "logprob_reward": -0.048643652349710464, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048643652349710464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24523480236530304, + 0.11977700144052505, + 0.077361099421978, + 0.10964721441268921, + 0.16065962612628937, + 0.09192970395088196, + 0.06963515281677246, + 0.03741567209362984, + 0.033076003193855286, + 0.05526374280452728 + ], + "credit_map": [ + 0.02524058148264885, + 0.042534105479717255, + 0.03463176265358925, + 0.04372558742761612, + 0.029609091579914093, + 0.021634269505739212, + 0.019034363329410553, + 0.028825027868151665, + 0.01498593483120203, + 0.016107887029647827, + 0.019286012277007103, + 0.02256702072918415, + 0.01053799781948328, + 0.014454169198870659, + 0.01390712708234787, + 0.007930852472782135, + 0.014912940561771393, + 0.014334323816001415, + 0.008638973347842693, + 0.012168826535344124, + 0.005900510121136904, + 0.007780216634273529, + 0.007285391446202993, + 0.006339912302792072, + 0.01450569648295641, + 0.018006475642323494, + 0.014480690471827984, + 0.019170237705111504, + 0.011175129562616348, + 0.009650233201682568, + 0.01279127411544323, + 0.009867485612630844, + 0.01707080751657486, + 0.020441962406039238, + 0.02611469104886055, + 0.03105389140546322, + 0.01699357107281685, + 0.01793980970978737, + 0.01844591647386551, + 0.01259897742420435, + 0.008604215458035469, + 0.011990955099463463, + 0.012315278872847557, + 0.018113620579242706, + 0.011461446061730385, + 0.008983192034065723, + 0.012321957387030125, + 0.008139031939208508, + 0.00975615344941616, + 0.008461755700409412, + 0.009783348068594933, + 0.012164165265858173, + 0.005552238784730434, + 0.008442552760243416, + 0.007845988497138023, + 0.007628956343978643, + 0.007156734820455313, + 0.004771815612912178, + 0.003985338378697634, + 0.0057686069048941135, + 0.0056840600445866585, + 0.0038442090153694153, + 0.00317607750184834, + 0.003028832608833909, + 0.004616251215338707, + 0.00629842234775424, + 0.004547661170363426, + 0.0050401766784489155, + 0.0030410082545131445, + 0.0034595532342791557, + 0.0032130111940205097, + 0.002859919099137187, + 0.009670997969806194, + 0.009008591994643211, + 0.0058104549534618855, + 0.008210374973714352, + 0.004312523640692234, + 0.005423286464065313, + 0.005198284517973661, + 0.0076292273588478565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.048643652349710464 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive steam locomotive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the locomotive rendered as a stark black silhouette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a blazing sunset with fiery red, crimson, and deep purple gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the locomotive have a single bright headlight casting a white beam down the rails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486158549785614 + ], + "opposite_logprob_tokens": [ + -1.2236158847808838 + ], + "logprob_reward": -0.3486158549785614, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486158549785614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25216397643089294, + 0.2085874229669571, + 0.08872661739587784, + 0.11882372200489044, + 0.06260750442743301, + 0.05883970484137535, + 0.03376135975122452, + 0.06376421451568604, + 0.051726218312978745, + 0.06099925935268402 + ], + "credit_map": [ + 0.02500319667160511, + 0.0445735901594162, + 0.031724631786346436, + 0.039655089378356934, + 0.024981604889035225, + 0.02572903037071228, + 0.029359916225075722, + 0.03113693743944168, + 0.023236120119690895, + 0.02496480755507946, + 0.02551634982228279, + 0.030503494665026665, + 0.023516783490777016, + 0.022755037993192673, + 0.027822021394968033, + 0.03027281165122986, + 0.01163218542933464, + 0.013725216500461102, + 0.010240664705634117, + 0.014571248553693295, + 0.011665526777505875, + 0.009604877792298794, + 0.010501737706363201, + 0.006785161793231964, + 0.014612085185945034, + 0.017452888190746307, + 0.015571572817862034, + 0.019655847921967506, + 0.010934071615338326, + 0.011798703111708164, + 0.012787736020982265, + 0.016010817140340805, + 0.008685454726219177, + 0.009031563065946102, + 0.007142951712012291, + 0.009212217293679714, + 0.007929369807243347, + 0.00721843121573329, + 0.006041106767952442, + 0.007346406579017639, + 0.007702155504375696, + 0.00987321324646473, + 0.0068842265754938126, + 0.009444084949791431, + 0.007096899673342705, + 0.006742644589394331, + 0.006327916868031025, + 0.004768564831465483, + 0.004065420012921095, + 0.004735158756375313, + 0.004236722830682993, + 0.00561733590438962, + 0.003634480759501457, + 0.003340488998219371, + 0.003955323249101639, + 0.004176429472863674, + 0.008435104973614216, + 0.009696520864963531, + 0.009625514037907124, + 0.011481330730021, + 0.007292954251170158, + 0.007172401994466782, + 0.006715476047247648, + 0.0033449125476181507, + 0.006775461602956057, + 0.008467149920761585, + 0.007576813921332359, + 0.010188638232648373, + 0.0044601378031075, + 0.005020884331315756, + 0.005068187136203051, + 0.004168943967670202, + 0.006792277097702026, + 0.007085782941430807, + 0.009308667853474617, + 0.011312460526823997, + 0.006566233467310667, + 0.005958008114248514, + 0.008192379958927631, + 0.0057834540493786335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.3486158549785614 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_10/index_tensor([5505, 5505])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_10/index_tensor([5505, 5505])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..35d87452060753ce5cf1e545f3621648425ba081 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_10/index_tensor([5505, 5505])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly zooming wide-angle shot initially focuses on a portion of a vast gray planet. Ancient, massive craters loom in shadow on its rocky surface, their edges softened by age. The planet rotates at a majestic, almost imperceptible pace, suspended within a swirling, luminous nebula of deep purple, fiery magenta, and brilliant cyan particles. As the camera steadily pans backward, the planet's full presence gradually emerges in the center of the frame, surrounded by colorful particles that float, swirl, and intertwine like living dust. These self-luminous particles, the sole source of light in the scene, cast a vibrant, flowing hue across the planet's gray surface, outlining its majestic contours. Light and shadow slowly shift across the vast, sunken features, creating a dynamic play of light and dark. The entire scene is imbued with a sense of cosmic tranquility and grandeur, evoking an atmosphere of mysterious and epic proportions.", + "seed": 9829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_10/index_5505_seed_9829_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1661898092366755, + "vlm_logprob": -0.1661898092366755, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.1661898092366755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11691714823246002 + ], + "opposite_logprob_tokens": [ + -2.241917133331299 + ], + "logprob_reward": -0.11691714823246002, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11691714823246002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11607973277568817, + 0.2018318623304367, + 0.04257343336939812, + 0.1567942351102829, + 0.17975781857967377, + 0.07662647217512131, + 0.09303472936153412, + 0.042962055653333664, + 0.062403567135334015, + 0.027936024591326714 + ], + "credit_map": [ + 0.008466155268251896, + 0.00922918226569891, + 0.01946721225976944, + 0.02171538770198822, + 0.007014623377472162, + 0.016079846769571304, + 0.020990876480937004, + 0.013116437941789627, + 0.014108824543654919, + 0.0217826459556818, + 0.0302138514816761, + 0.03364177420735359, + 0.02234974317252636, + 0.02562829665839672, + 0.02953561767935753, + 0.0245711300522089, + 0.005238120909780264, + 0.00471618864685297, + 0.005025643855333328, + 0.00547663914039731, + 0.0054620797745883465, + 0.003482240252196789, + 0.007163355126976967, + 0.006009168457239866, + 0.009230446070432663, + 0.04706277325749397, + 0.050595398992300034, + 0.015453516505658627, + 0.013120779767632484, + 0.005181562155485153, + 0.009545537643134594, + 0.006604225840419531, + 0.019441446289420128, + 0.014010242186486721, + 0.038285624235868454, + 0.030298519879579544, + 0.027105603367090225, + 0.012885541655123234, + 0.02073671482503414, + 0.016994129866361618, + 0.005446753930300474, + 0.0029012078884989023, + 0.008665553294122219, + 0.012146996334195137, + 0.00420245248824358, + 0.0064005241729319096, + 0.02244766615331173, + 0.014415320940315723, + 0.002033261815086007, + 0.002912038704380393, + 0.0034922249615192413, + 0.03927135467529297, + 0.0027751275338232517, + 0.0016302724834531546, + 0.03531889244914055, + 0.005601550918072462, + 0.002422216348350048, + 0.0036069771740585566, + 0.004349015187472105, + 0.012903536669909954, + 0.004629787988960743, + 0.003281186567619443, + 0.006884801667183638, + 0.0048845321871340275, + 0.0063753873109817505, + 0.003968976438045502, + 0.0057916902005672455, + 0.007965156808495522, + 0.009385256096720695, + 0.010307853110134602, + 0.008544808253645897, + 0.0100644426420331, + 0.003112072590738535, + 0.0031448937952518463, + 0.003608817933127284, + 0.004236121196299791, + 0.0034401006996631622, + 0.003472063457593322, + 0.0037714862264692783, + 0.0031504672951996326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.11691714823246002 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487192392349243 + ], + "opposite_logprob_tokens": [ + -1.2237192392349243 + ], + "logprob_reward": -0.3487192392349243, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3487192392349243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08299685269594193, + 0.07799934595823288, + 0.06238633766770363, + 0.3093380928039551, + 0.0660770833492279, + 0.08681850880384445, + 0.20825020968914032, + 0.031546834856271744, + 0.05110270529985428, + 0.023484058678150177 + ], + "credit_map": [ + 0.007026192732155323, + 0.006455110851675272, + 0.01201609242707491, + 0.01346560101956129, + 0.007668735459446907, + 0.011862710118293762, + 0.012295759283006191, + 0.012206648476421833, + 0.0062896208837628365, + 0.010322337038815022, + 0.011641080491244793, + 0.013752644881606102, + 0.007468339987099171, + 0.008979390375316143, + 0.012661147862672806, + 0.006884785834699869, + 0.004907557740807533, + 0.007027805782854557, + 0.010554117150604725, + 0.01626451313495636, + 0.004123202059417963, + 0.004385452251881361, + 0.008819825015962124, + 0.006303862202912569, + 0.018261339515447617, + 0.09754280745983124, + 0.10941362380981445, + 0.029384879395365715, + 0.021810201928019524, + 0.008849941194057465, + 0.01315020676702261, + 0.010925108566880226, + 0.007386167533695698, + 0.005625774152576923, + 0.014381499961018562, + 0.009889772161841393, + 0.01025842223316431, + 0.004403274040669203, + 0.008075031451880932, + 0.006057141348719597, + 0.008934966288506985, + 0.00394835090264678, + 0.007743001915514469, + 0.012385567650198936, + 0.005702983122318983, + 0.009229486808180809, + 0.02286945842206478, + 0.016004696488380432, + 0.010828294791281223, + 0.007415576372295618, + 0.009211346507072449, + 0.0910973772406578, + 0.0057157110422849655, + 0.004283867310732603, + 0.06870583444833755, + 0.010992196388542652, + 0.0024731899611651897, + 0.0030280749779194593, + 0.005209445953369141, + 0.0063962372951209545, + 0.0029026460833847523, + 0.003718999680131674, + 0.004172667860984802, + 0.0036455723457038403, + 0.00453310227021575, + 0.002882138593122363, + 0.005477133672684431, + 0.0073002781718969345, + 0.006022526416927576, + 0.0044463989324867725, + 0.014148393645882607, + 0.0062927319668233395, + 0.0020787299145013094, + 0.002840230241417885, + 0.0028922618366777897, + 0.0031454286072403193, + 0.0027143650222569704, + 0.0036710300482809544, + 0.0036737232003360987, + 0.002468290738761425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.3487192392349243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024926564656198025 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024926564656198025, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024926564656198025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1360965371131897, + 0.13134242594242096, + 0.16300784051418304, + 0.24535566568374634, + 0.08480416238307953, + 0.1263408064842224, + 0.023633532226085663, + 0.0422978512942791, + 0.0287823174148798, + 0.018338831141591072 + ], + "credit_map": [ + 0.009093114174902439, + 0.008191471919417381, + 0.026564693078398705, + 0.029590237885713577, + 0.007114404812455177, + 0.012366443872451782, + 0.02452576905488968, + 0.018650421872735023, + 0.010294031351804733, + 0.014243530109524727, + 0.01898091845214367, + 0.020209284499287605, + 0.015471369959414005, + 0.02103319577872753, + 0.016944291070103645, + 0.014165808446705341, + 0.013151487335562706, + 0.016301168128848076, + 0.023025954142212868, + 0.03876025974750519, + 0.017715035006403923, + 0.01110615860670805, + 0.02527398057281971, + 0.017673803493380547, + 0.014941890724003315, + 0.07958289980888367, + 0.0838145986199379, + 0.022128211334347725, + 0.018653573468327522, + 0.007010621018707752, + 0.010760226286947727, + 0.008463667705655098, + 0.009882096201181412, + 0.007376675494015217, + 0.01890522986650467, + 0.013973529450595379, + 0.011955534107983112, + 0.006103834602981806, + 0.009382819756865501, + 0.007224447559565306, + 0.008232020772993565, + 0.007326566614210606, + 0.011217036284506321, + 0.027119100093841553, + 0.007753984536975622, + 0.008847915567457676, + 0.03404871001839638, + 0.021795492619276047, + 0.0020598648115992546, + 0.0017778489273041487, + 0.0035631260834634304, + 0.0032014960888773203, + 0.0015248729614540935, + 0.0025902222841978073, + 0.004718819633126259, + 0.004197284113615751, + 0.004805079195648432, + 0.0025270183105021715, + 0.006237095221877098, + 0.0072426884435117245, + 0.0032770042307674885, + 0.002939534140750766, + 0.01037319004535675, + 0.004896247293800116, + 0.0022050391416996717, + 0.0022618635557591915, + 0.0034091670531779528, + 0.0034810472279787064, + 0.003441475797444582, + 0.005285825114697218, + 0.0038287693168967962, + 0.004869131837040186, + 0.0018225823296234012, + 0.0018813767237588763, + 0.002937404904514551, + 0.0022663504350930452, + 0.0022355106193572283, + 0.0018387018935754895, + 0.002612696960568428, + 0.0027442078571766615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0024926564656198025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0142059950158 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.0142059950158, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0142059950158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0701301246881485, + 0.09607481211423874, + 0.061723742634058, + 0.05697636678814888, + 0.05339880287647247, + 0.0617142915725708, + 0.27164551615715027, + 0.035856474190950394, + 0.27455562353134155, + 0.017924286425113678 + ], + "credit_map": [ + 0.0055759926326572895, + 0.0065008364617824554, + 0.010336663573980331, + 0.013617724180221558, + 0.006191651802510023, + 0.007595358416438103, + 0.009311002679169178, + 0.01100088655948639, + 0.007374491076916456, + 0.008870914578437805, + 0.016079440712928772, + 0.015512012876570225, + 0.011271321214735508, + 0.012113625183701515, + 0.010423053987324238, + 0.01442995760589838, + 0.0043497951701283455, + 0.006067456677556038, + 0.00942301657050848, + 0.01282592210918665, + 0.00638162437826395, + 0.006733379792422056, + 0.009284889325499535, + 0.006657659541815519, + 0.003279902972280979, + 0.01254437305033207, + 0.01786775141954422, + 0.0064536635763943195, + 0.00534905306994915, + 0.0027256945613771677, + 0.004382229410111904, + 0.004373702686280012, + 0.0053699263371527195, + 0.00350003968924284, + 0.01274818368256092, + 0.00975202675908804, + 0.004251815844327211, + 0.004762711934745312, + 0.006917120423167944, + 0.006096976343542337, + 0.009347253479063511, + 0.0048958477564156055, + 0.006420781835913658, + 0.013305115513503551, + 0.004754744470119476, + 0.007958580739796162, + 0.007231524214148521, + 0.007800441700965166, + 0.01398898009210825, + 0.009519090875983238, + 0.011806556954979897, + 0.11909754574298859, + 0.007604448590427637, + 0.005631112027913332, + 0.0888991728425026, + 0.015098602510988712, + 0.003559043863788247, + 0.0029517451766878366, + 0.006446775048971176, + 0.005909782834351063, + 0.0030713167507201433, + 0.004502816591411829, + 0.005252993665635586, + 0.004161998629570007, + 0.0197360347956419, + 0.013634777627885342, + 0.02149001508951187, + 0.04725688323378563, + 0.017564168199896812, + 0.018679508939385414, + 0.09696672111749649, + 0.03922750800848007, + 0.001681576482951641, + 0.0020861595403403044, + 0.002675428753718734, + 0.001999571220949292, + 0.0023476877249777317, + 0.0018716448685154319, + 0.0026481603272259235, + 0.0026140569243580103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.0142059950158 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34861400723457336 + ], + "opposite_logprob_tokens": [ + -1.223613977432251 + ], + "logprob_reward": -0.34861400723457336, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34861400723457336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1358049213886261, + 0.0474587120115757, + 0.04826473817229271, + 0.15305444598197937, + 0.04799477756023407, + 0.20026633143424988, + 0.11686205863952637, + 0.026645300909876823, + 0.205875426530838, + 0.017773298546671867 + ], + "credit_map": [ + 0.010906383395195007, + 0.008989976719021797, + 0.017740199342370033, + 0.02736993134021759, + 0.009785144589841366, + 0.014519862830638885, + 0.024305779486894608, + 0.02218763530254364, + 0.004003338515758514, + 0.005230128299444914, + 0.007374524604529142, + 0.008435473777353764, + 0.005184748210012913, + 0.004343958105891943, + 0.007803307846188545, + 0.005083232652395964, + 0.003896905342116952, + 0.005055909976363182, + 0.0070368205197155476, + 0.009830002672970295, + 0.003977652173489332, + 0.0029076074715703726, + 0.010622064583003521, + 0.004937771707773209, + 0.008425655774772167, + 0.045777007937431335, + 0.05348403751850128, + 0.01541913952678442, + 0.010413587093353271, + 0.0044692885130643845, + 0.008280877023935318, + 0.006784848868846893, + 0.006458432879298925, + 0.0038194789085537195, + 0.008770822547376156, + 0.00755902286618948, + 0.005914398934692144, + 0.00496588135138154, + 0.00566429877653718, + 0.004842438269406557, + 0.020084679126739502, + 0.010123750194907188, + 0.018336258828639984, + 0.02987808920443058, + 0.011122902855277061, + 0.021891573444008827, + 0.05242586508393288, + 0.03640321269631386, + 0.01616932451725006, + 0.007751784287393093, + 0.009899957105517387, + 0.03680333495140076, + 0.0073437774553895, + 0.005886372644454241, + 0.016818854957818985, + 0.01618865877389908, + 0.003068623598664999, + 0.0023123500868678093, + 0.004090342670679092, + 0.005319064948707819, + 0.002597155747935176, + 0.0031682769767940044, + 0.003135510953143239, + 0.0029539777897298336, + 0.01648852787911892, + 0.008221268653869629, + 0.01287170685827732, + 0.035690631717443466, + 0.010761160403490067, + 0.009337490424513817, + 0.08189670741558075, + 0.030607936903834343, + 0.0018380745314061642, + 0.0023237091954797506, + 0.0019879210740327835, + 0.0029209323693066835, + 0.0025154161266982555, + 0.0018594356952235103, + 0.0025459269527345896, + 0.00178188132122159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.34861400723457336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly zooming wide-angle shot initially focuses on a portion of a vast gray planet. Ancient, massive craters loom in shadow on its rocky surface, their edges softened by age. The planet rotates at a majestic, almost imperceptible pace, suspended within a swirling, luminous nebula of deep purple, fiery magenta, and brilliant cyan particles. As the camera steadily pans backward, the planet's full presence gradually emerges in the center of the frame, surrounded by colorful particles that float, swirl, and intertwine like living dust. These self-luminous particles, the sole source of light in the scene, cast a vibrant, flowing hue across the planet's gray surface, outlining its majestic contours. Light and shadow slowly shift across the vast, sunken features, creating a dynamic play of light and dark. The entire scene is imbued with a sense of cosmic tranquility and grandeur, evoking an atmosphere of mysterious and epic proportions.", + "seed": 9830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_10/index_5505_seed_9830_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11721157012507319, + "vlm_logprob": -0.11721157012507319, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.11721157012507319, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1622217297554016 + ], + "opposite_logprob_tokens": [ + -1.9122216701507568 + ], + "logprob_reward": -0.1622217297554016, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1622217297554016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.487540602684021, + 0.03996735066175461, + 0.3053146004676819, + 0.031441036611795425, + 0.04089777171611786, + 0.02594495378434658, + 0.014549587853252888, + 0.022636117413640022, + 0.01198708638548851, + 0.019720861688256264 + ], + "credit_map": [ + 0.03157978877425194, + 0.07984650880098343, + 0.026334963738918304, + 0.027578867971897125, + 0.14652670919895172, + 0.04405655339360237, + 0.10341960936784744, + 0.02819761261343956, + 0.00291457655839622, + 0.009019232355058193, + 0.004114390350878239, + 0.004085004795342684, + 0.008054396137595177, + 0.0037917448207736015, + 0.0037602437660098076, + 0.004227765370160341, + 0.04102124273777008, + 0.022987378761172295, + 0.13051626086235046, + 0.02074986882507801, + 0.05410866066813469, + 0.010707151144742966, + 0.012064882554113865, + 0.01315915584564209, + 0.00720116775482893, + 0.0024292413145303726, + 0.010212494060397148, + 0.003351375227794051, + 0.0028250881005078554, + 0.002047992777079344, + 0.0018727007554844022, + 0.0015009789494797587, + 0.0022034423891454935, + 0.004551401361823082, + 0.005433248355984688, + 0.013514550402760506, + 0.004902753513306379, + 0.0021387157030403614, + 0.004287953022867441, + 0.003865711158141494, + 0.001817892654798925, + 0.002727090148255229, + 0.002618937287479639, + 0.007070852909237146, + 0.0027067754417657852, + 0.0017361327772960067, + 0.0027785575948655605, + 0.004488715901970863, + 0.001356587978079915, + 0.0017040955135598779, + 0.0017655156552791595, + 0.002425064332783222, + 0.0016983808018267155, + 0.0018366152653470635, + 0.0016930261626839638, + 0.0020703033078461885, + 0.001638678484596312, + 0.0019247904419898987, + 0.0014252607943490148, + 0.0022661106195300817, + 0.002229514764621854, + 0.005294835194945335, + 0.0020018971990793943, + 0.005855030845850706, + 0.001088624121621251, + 0.001412976416759193, + 0.0013264541048556566, + 0.002071079798042774, + 0.0017323070205748081, + 0.0013023181818425655, + 0.001470278250053525, + 0.0015830491902306676, + 0.0014586206525564194, + 0.0017824958777055144, + 0.0024023503065109253, + 0.0034498826134949923, + 0.002472959691658616, + 0.0017521982081234455, + 0.0031314510852098465, + 0.0032709066290408373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.1622217297554016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28142642974853516 + ], + "opposite_logprob_tokens": [ + -1.4064264297485352 + ], + "logprob_reward": -0.28142642974853516, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28142642974853516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19579675793647766, + 0.04128209501504898, + 0.4663832187652588, + 0.02940881811082363, + 0.09559884667396545, + 0.0361122190952301, + 0.04679779335856438, + 0.03270880877971649, + 0.03902362659573555, + 0.016887793317437172 + ], + "credit_map": [ + 0.013674083165824413, + 0.03466206043958664, + 0.00841454230248928, + 0.009735717438161373, + 0.06355997174978256, + 0.01651906967163086, + 0.03694666922092438, + 0.012284637428820133, + 0.0028846210334450006, + 0.006855101324617863, + 0.0034805452451109886, + 0.004349072929471731, + 0.011387871578335762, + 0.005783344618976116, + 0.003295843955129385, + 0.003245691768825054, + 0.06253950297832489, + 0.03594636917114258, + 0.19919441640377045, + 0.03127771615982056, + 0.08133076131343842, + 0.017188357189297676, + 0.01928524300456047, + 0.019620850682258606, + 0.0036887500900775194, + 0.0032864839304238558, + 0.00671533914282918, + 0.005821587983518839, + 0.0037010684609413147, + 0.001953395549207926, + 0.002496054396033287, + 0.0017461390234529972, + 0.005886847153306007, + 0.015143697150051594, + 0.00918723363429308, + 0.03170658275485039, + 0.00841667503118515, + 0.00700443796813488, + 0.009670380502939224, + 0.008582992479205132, + 0.0035776267759501934, + 0.004127140622586012, + 0.005642017815262079, + 0.006704859435558319, + 0.004157163202762604, + 0.0021331659518182278, + 0.003746248781681061, + 0.006023995578289032, + 0.0035061019007116556, + 0.0023058431688696146, + 0.0023369514383375645, + 0.0036250008270144463, + 0.0029864481184631586, + 0.01273383293300867, + 0.003342997981235385, + 0.01596061699092388, + 0.0023605315946042538, + 0.0023660913575440645, + 0.0021780866663903, + 0.006759740877896547, + 0.0030355441849678755, + 0.0026803589425981045, + 0.00305155455134809, + 0.01027690153568983, + 0.002922783372923732, + 0.003944564610719681, + 0.004666729364544153, + 0.004260397516191006, + 0.006563908886164427, + 0.005654266569763422, + 0.004753636196255684, + 0.006257344502955675, + 0.001410381868481636, + 0.0017334333388134837, + 0.0024478815030306578, + 0.0018397185485810041, + 0.0022987318225204945, + 0.0022538104094564915, + 0.0024222121573984623, + 0.0024816233199089766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.28142642974853516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028236540965735912 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.0028236540965735912, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028236540965735912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41213053464889526, + 0.08691086620092392, + 0.07777518779039383, + 0.09413517266511917, + 0.14651942253112793, + 0.030234316363930702, + 0.028482601046562195, + 0.03920986130833626, + 0.030204666778445244, + 0.05439741536974907 + ], + "credit_map": [ + 0.02042992226779461, + 0.07402485609054565, + 0.019504887983202934, + 0.0214399341493845, + 0.13235417008399963, + 0.04001111909747124, + 0.07774804532527924, + 0.026617586612701416, + 0.00786597654223442, + 0.010597278364002705, + 0.010758495889604092, + 0.013446571305394173, + 0.01813327707350254, + 0.011364833451807499, + 0.007790819741785526, + 0.006953612435609102, + 0.0077600968070328236, + 0.013640007004141808, + 0.024694066494703293, + 0.007802823092788458, + 0.0069501278921961784, + 0.005962016060948372, + 0.005573699250817299, + 0.005392353981733322, + 0.02011510170996189, + 0.009065997786819935, + 0.029774118214845657, + 0.009782562963664532, + 0.007978588342666626, + 0.006406654603779316, + 0.007206535432487726, + 0.0038056173361837864, + 0.008925423957407475, + 0.0163080096244812, + 0.011472037993371487, + 0.053747642785310745, + 0.013701784424483776, + 0.00800994411110878, + 0.01493245828896761, + 0.019422125071287155, + 0.002329122042283416, + 0.003302413271740079, + 0.0032057110220193863, + 0.007037373259663582, + 0.0032980693504214287, + 0.002301981672644615, + 0.002546374686062336, + 0.006213271990418434, + 0.0025689946487545967, + 0.003749808529391885, + 0.0032919838558882475, + 0.007558665703982115, + 0.0032173199579119682, + 0.002358783734962344, + 0.0026484692934900522, + 0.0030885767191648483, + 0.002270708791911602, + 0.0035386476665735245, + 0.0025151909794658422, + 0.00614943215623498, + 0.005926085170358419, + 0.007721433416008949, + 0.0035008566919714212, + 0.007587505038827658, + 0.0034037132281810045, + 0.003804802196100354, + 0.005639677867293358, + 0.0036372067406773567, + 0.003775297198444605, + 0.003398885251954198, + 0.002425639657303691, + 0.0041194455698132515, + 0.0041212718933820724, + 0.004868361633270979, + 0.008934595622122288, + 0.010632147081196308, + 0.006016146391630173, + 0.005306598264724016, + 0.006916888058185577, + 0.007601404562592506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0028236540965735912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551834806799889 + ], + "opposite_logprob_tokens": [ + -4.387551784515381 + ], + "logprob_reward": -0.012551834806799889, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551834806799889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1377037763595581, + 0.06093825027346611, + 0.4367091655731201, + 0.047896139323711395, + 0.16031140089035034, + 0.037577614188194275, + 0.02191862091422081, + 0.017451046034693718, + 0.054169777780771255, + 0.025324221700429916 + ], + "credit_map": [ + 0.011491886340081692, + 0.025432854890823364, + 0.006541553419083357, + 0.008234472945332527, + 0.03782074898481369, + 0.015257672406733036, + 0.02152971737086773, + 0.011394866742193699, + 0.00620036106556654, + 0.01127151120454073, + 0.005487713497132063, + 0.0054922630079090595, + 0.016826381906867027, + 0.006717629265040159, + 0.0036021326668560505, + 0.005340261850506067, + 0.06006976589560509, + 0.03361009061336517, + 0.1845638006925583, + 0.03000989556312561, + 0.07447995990514755, + 0.016628239303827286, + 0.017776789143681526, + 0.0195706095546484, + 0.007585553918033838, + 0.003395173465833068, + 0.01601448282599449, + 0.004233859479427338, + 0.00395192438736558, + 0.002192837418988347, + 0.008029289543628693, + 0.00249302014708519, + 0.008191501721739769, + 0.021386519074440002, + 0.014152091927826405, + 0.06043575704097748, + 0.01202310249209404, + 0.009921654127538204, + 0.01501715462654829, + 0.01918361335992813, + 0.004239980597048998, + 0.002538137836381793, + 0.003663616254925728, + 0.006402450148016214, + 0.002803298644721508, + 0.0026293678674846888, + 0.003980143461376429, + 0.011320620775222778, + 0.001456590136513114, + 0.0017960242694243789, + 0.0015956817660480738, + 0.004217562265694141, + 0.0017667554784566164, + 0.006230670027434826, + 0.0013661952689290047, + 0.003489139722660184, + 0.002522703493013978, + 0.0022598798386752605, + 0.0019670783076435328, + 0.002256570616737008, + 0.002765193348750472, + 0.0015823491849005222, + 0.0016738022677600384, + 0.0024234678130596876, + 0.005699648056179285, + 0.0072977482341229916, + 0.007428412325680256, + 0.0064233457669615746, + 0.008926157839596272, + 0.0048555838875472546, + 0.00617213174700737, + 0.007366752717643976, + 0.00170719379093498, + 0.002599814673885703, + 0.004317154176533222, + 0.0022726948373019695, + 0.0024438935797661543, + 0.003172716125845909, + 0.0029706615023314953, + 0.005840094294399023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.012551834806799889 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703420221805573 + ], + "opposite_logprob_tokens": [ + -2.1270341873168945 + ], + "logprob_reward": -0.12703420221805573, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703420221805573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06367634236812592, + 0.07196714729070663, + 0.4407815933227539, + 0.032746151089668274, + 0.1400282084941864, + 0.06333062797784805, + 0.0646875724196434, + 0.038815904408693314, + 0.06996698677539825, + 0.013999495655298233 + ], + "credit_map": [ + 0.00862252339720726, + 0.008851866237819195, + 0.006182858720421791, + 0.006469763815402985, + 0.011644313111901283, + 0.0070228129625320435, + 0.010297334752976894, + 0.00458486657589674, + 0.004230675287544727, + 0.019241271540522575, + 0.003609833074733615, + 0.005197376012802124, + 0.023535819724202156, + 0.006439806893467903, + 0.004366603214293718, + 0.005345769692212343, + 0.061974313110113144, + 0.03518776595592499, + 0.1861274242401123, + 0.029894623905420303, + 0.07066277414560318, + 0.017887171357870102, + 0.018953561782836914, + 0.020093951374292374, + 0.004280942492187023, + 0.0025112552102655172, + 0.006825485732406378, + 0.005149869713932276, + 0.00439565721899271, + 0.003916773479431868, + 0.0033539829310029745, + 0.002312183380126953, + 0.010453407652676105, + 0.02047009952366352, + 0.015447419136762619, + 0.044993482530117035, + 0.01143287867307663, + 0.006602690555155277, + 0.010395986028015614, + 0.020232241600751877, + 0.0031430863309651613, + 0.006416813936084509, + 0.0050844415090978146, + 0.019206609576940536, + 0.007600968703627586, + 0.004045906476676464, + 0.0035684804897755384, + 0.014264323748648167, + 0.005470889154821634, + 0.0034310973715037107, + 0.0026639478746801615, + 0.005279805511236191, + 0.004208793397992849, + 0.020431580021977425, + 0.00443294970318675, + 0.01876850798726082, + 0.0031584210228174925, + 0.004264554940164089, + 0.0035753007978200912, + 0.009531293995678425, + 0.004567509051412344, + 0.004404096398502588, + 0.004819901194423437, + 0.0044948263093829155, + 0.005090726539492607, + 0.009992691688239574, + 0.008127766661345959, + 0.007678864523768425, + 0.013696419075131416, + 0.00762166827917099, + 0.00897609256207943, + 0.008782759308815002, + 0.001141347805969417, + 0.0014314694562926888, + 0.0019556996412575245, + 0.0022908407263457775, + 0.0017691717948764563, + 0.0014554362278431654, + 0.001865353318862617, + 0.0020901772659271955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.12703420221805573 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly zooming wide-angle shot initially focuses on a portion of a vast gray planet. Ancient, massive craters loom in shadow on its rocky surface, their edges softened by age. The planet rotates at a majestic, almost imperceptible pace, suspended within a swirling, luminous nebula of deep purple, fiery magenta, and brilliant cyan particles. As the camera steadily pans backward, the planet's full presence gradually emerges in the center of the frame, surrounded by colorful particles that float, swirl, and intertwine like living dust. These self-luminous particles, the sole source of light in the scene, cast a vibrant, flowing hue across the planet's gray surface, outlining its majestic contours. Light and shadow slowly shift across the vast, sunken features, creating a dynamic play of light and dark. The entire scene is imbued with a sense of cosmic tranquility and grandeur, evoking an atmosphere of mysterious and epic proportions.", + "seed": 9831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_10/index_5505_seed_9831_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.23178313616663218, + "vlm_logprob": -0.23178313616663218, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.23178313616663218, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13103745877742767 + ], + "opposite_logprob_tokens": [ + -2.131037473678589 + ], + "logprob_reward": -0.13103745877742767, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13103745877742767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33834055066108704, + 0.03640136495232582, + 0.08842289447784424, + 0.063356414437294, + 0.03707382455468178, + 0.28260794281959534, + 0.03131682798266411, + 0.031567253172397614, + 0.05020495131611824, + 0.040708042681217194 + ], + "credit_map": [ + 0.04332888126373291, + 0.06449039280414581, + 0.04651779308915138, + 0.057618800550699234, + 0.047831952571868896, + 0.021222352981567383, + 0.021725548431277275, + 0.0356048084795475, + 0.006187719292938709, + 0.004462821409106255, + 0.00698093743994832, + 0.0036282348446547985, + 0.0029935978818684816, + 0.003528157016262412, + 0.00450873700901866, + 0.004111155867576599, + 0.01218978501856327, + 0.008884836919605732, + 0.012358246371150017, + 0.013003893196582794, + 0.009916791692376137, + 0.007303292863070965, + 0.01136546116322279, + 0.01340058259665966, + 0.008322772569954395, + 0.0066394624300301075, + 0.006777054164558649, + 0.008717401884496212, + 0.004939311649650335, + 0.006368795409798622, + 0.009267184883356094, + 0.012324429117143154, + 0.004660563077777624, + 0.004312219098210335, + 0.004006323870271444, + 0.006078310776501894, + 0.003943618852645159, + 0.004094278905540705, + 0.004140619188547134, + 0.005837889853864908, + 0.0388203039765358, + 0.02769564464688301, + 0.0494086854159832, + 0.03153560683131218, + 0.025377720594406128, + 0.015919407829642296, + 0.041213568300008774, + 0.05263698846101761, + 0.0035797159653156996, + 0.004469117149710655, + 0.004750330001115799, + 0.004828948527574539, + 0.0034320310223847628, + 0.0030743747483938932, + 0.0034575052559375763, + 0.0037248036824166775, + 0.0029917419888079166, + 0.004154878668487072, + 0.005310429725795984, + 0.004780429881066084, + 0.002920919330790639, + 0.0019845645874738693, + 0.006790250074118376, + 0.002634036121889949, + 0.005094844847917557, + 0.005319604184478521, + 0.011043265461921692, + 0.011821117252111435, + 0.0030452958308160305, + 0.00233284174464643, + 0.008930501528084278, + 0.002617478836327791, + 0.004046412650495768, + 0.006124705541878939, + 0.008233495987951756, + 0.006505631376057863, + 0.0036307265982031822, + 0.004284570924937725, + 0.004748577252030373, + 0.0031339169945567846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.13103745877742767 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135237693786621 + ], + "opposite_logprob_tokens": [ + -1.313523769378662 + ], + "logprob_reward": -0.3135237693786621, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135237693786621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23378781974315643, + 0.10178276896476746, + 0.05924231931567192, + 0.06206785887479782, + 0.059288542717695236, + 0.1428779512643814, + 0.11618640273809433, + 0.05986235663294792, + 0.08649396151304245, + 0.07841003686189651 + ], + "credit_map": [ + 0.030811524018645287, + 0.04754357784986496, + 0.03366387262940407, + 0.039097439497709274, + 0.029689563438296318, + 0.012028086930513382, + 0.014193873852491379, + 0.02675984986126423, + 0.012801206670701504, + 0.01173434779047966, + 0.02414797432720661, + 0.011602756567299366, + 0.008228910155594349, + 0.009323431178927422, + 0.01431487686932087, + 0.009629257954657078, + 0.0067558325827121735, + 0.005878125317394733, + 0.008524294942617416, + 0.009963762946426868, + 0.00524958735331893, + 0.005463236477226019, + 0.006808120291680098, + 0.010599350556731224, + 0.007219273597002029, + 0.007441462017595768, + 0.006167283281683922, + 0.011668143793940544, + 0.005914436187595129, + 0.005089664366096258, + 0.008730272762477398, + 0.009837319143116474, + 0.006962671410292387, + 0.008068601600825787, + 0.007201181724667549, + 0.008406565524637699, + 0.005123250652104616, + 0.005106153432279825, + 0.010236026719212532, + 0.008184084668755531, + 0.027000438421964645, + 0.018078681081533432, + 0.016963599249720573, + 0.02073495276272297, + 0.010936136357486248, + 0.009059332311153412, + 0.019198454916477203, + 0.02090635895729065, + 0.017997559159994125, + 0.009248913265764713, + 0.0186674315482378, + 0.02044166438281536, + 0.010331577621400356, + 0.007345080841332674, + 0.012643329799175262, + 0.019510839134454727, + 0.005542743485420942, + 0.007408235687762499, + 0.009333574213087559, + 0.010532751679420471, + 0.007171190343797207, + 0.004447732586413622, + 0.007548490539193153, + 0.007877631112933159, + 0.00853453204035759, + 0.011386088095605373, + 0.013956603594124317, + 0.01627235859632492, + 0.0053588817827403545, + 0.004114819224923849, + 0.02023984305560589, + 0.006630823481827974, + 0.006875922437757254, + 0.00901518203318119, + 0.02183881402015686, + 0.014952006749808788, + 0.005421051289886236, + 0.00404844805598259, + 0.009967557154595852, + 0.0062910523265600204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.3135237693786621 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028211576864123344 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028211576864123344, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028211576864123344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14911872148513794, + 0.1435612440109253, + 0.17574840784072876, + 0.04211416468024254, + 0.041063111275434494, + 0.1699725091457367, + 0.12980306148529053, + 0.04117127135396004, + 0.03383797034621239, + 0.07360956072807312 + ], + "credit_map": [ + 0.0176223237067461, + 0.020690500736236572, + 0.02269422449171543, + 0.022698869928717613, + 0.01677628979086876, + 0.019313707947731018, + 0.00956652220338583, + 0.01975627802312374, + 0.013166536577045918, + 0.00721445307135582, + 0.05280385911464691, + 0.015237501822412014, + 0.006310286000370979, + 0.010318753309547901, + 0.017962364479899406, + 0.020547494292259216, + 0.017466222867369652, + 0.018088774755597115, + 0.025591541081666946, + 0.03018759936094284, + 0.021565619856119156, + 0.014655612409114838, + 0.02603670209646225, + 0.022156324237585068, + 0.004045042209327221, + 0.005084825213998556, + 0.00414511701092124, + 0.00774585222825408, + 0.004257090389728546, + 0.003423883579671383, + 0.006630748976022005, + 0.006781603209674358, + 0.004135445691645145, + 0.00333274994045496, + 0.0046998984180390835, + 0.00601618317887187, + 0.005800756625831127, + 0.004398337099701166, + 0.006099408026784658, + 0.006580328568816185, + 0.018624840304255486, + 0.01801358535885811, + 0.025968849658966064, + 0.021539388224482536, + 0.018528858199715614, + 0.015269382856786251, + 0.022722167894244194, + 0.029305433854460716, + 0.008538205176591873, + 0.007414677646011114, + 0.012822480872273445, + 0.03419550508260727, + 0.014904736541211605, + 0.007965476252138615, + 0.013920122757554054, + 0.030041860416531563, + 0.0028351235669106245, + 0.004402720835059881, + 0.004774156957864761, + 0.0064392476342618465, + 0.004190768115222454, + 0.0030313662718981504, + 0.011429183185100555, + 0.004068706650286913, + 0.0032800063490867615, + 0.003142520086839795, + 0.007582905702292919, + 0.006264770403504372, + 0.0027278230991214514, + 0.0019167539430782199, + 0.0066194431856274605, + 0.002303747460246086, + 0.005010813940316439, + 0.010595696046948433, + 0.01237417571246624, + 0.011949921026825905, + 0.009303472936153412, + 0.007705895230174065, + 0.009385723620653152, + 0.007283864542841911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.0028211576864123344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195616081357002 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195616081357002, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195616081357002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20079545676708221, + 0.1520586460828781, + 0.2351798713207245, + 0.05023438110947609, + 0.03107411414384842, + 0.05064768344163895, + 0.07886837422847748, + 0.0528753288090229, + 0.027146542444825172, + 0.12111961841583252 + ], + "credit_map": [ + 0.031228138133883476, + 0.017096567898988724, + 0.03174544870853424, + 0.020815130323171616, + 0.021318672224879265, + 0.023020504042506218, + 0.025121530517935753, + 0.030449453741312027, + 0.017280906438827515, + 0.008539415895938873, + 0.05285777524113655, + 0.014451893977820873, + 0.006566649302840233, + 0.008899470791220665, + 0.016887059435248375, + 0.026575474068522453, + 0.032849185168743134, + 0.020904406905174255, + 0.03091651014983654, + 0.03573072701692581, + 0.027717342600226402, + 0.017024148255586624, + 0.03037809208035469, + 0.03965943306684494, + 0.005529389251023531, + 0.005071318242698908, + 0.005776474252343178, + 0.008529827930033207, + 0.005495023913681507, + 0.005628729704767466, + 0.0066305301152169704, + 0.007573089562356472, + 0.0039809285663068295, + 0.003016118658706546, + 0.0033264642115682364, + 0.0034203743562102318, + 0.0032724810298532248, + 0.005073616746813059, + 0.0035483897663652897, + 0.005435740575194359, + 0.008864657953381538, + 0.004262902773916721, + 0.006521780509501696, + 0.006877515465021133, + 0.004296039696782827, + 0.004643636289983988, + 0.007459329906851053, + 0.007721824571490288, + 0.006656330544501543, + 0.006020731758326292, + 0.007597631774842739, + 0.02045275643467903, + 0.009512756951153278, + 0.0047359224408864975, + 0.009031559340655804, + 0.014860684052109718, + 0.005447207018733025, + 0.005933291278779507, + 0.007933417335152626, + 0.00978558138012886, + 0.005629049148410559, + 0.003975893836468458, + 0.00814084429293871, + 0.006030041258782148, + 0.003082012990489602, + 0.002825349336490035, + 0.005479435436427593, + 0.003986585419625044, + 0.002839241875335574, + 0.002215448534116149, + 0.0038199611008167267, + 0.0028985082171857357, + 0.009621722623705864, + 0.021676819771528244, + 0.01688709296286106, + 0.019827356562018394, + 0.013853740878403187, + 0.012314947322010994, + 0.01402969378978014, + 0.012908244505524635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.018195616081357002 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933376789093018 + ], + "opposite_logprob_tokens": [ + -0.6933376789093018 + ], + "logprob_reward": -0.6933376789093018, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933376789093018, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08006326854228973, + 0.26166993379592896, + 0.15050114691257477, + 0.02236246131360531, + 0.03463383764028549, + 0.050701066851615906, + 0.07223039865493774, + 0.042250148952007294, + 0.06809492409229279, + 0.21749277412891388 + ], + "credit_map": [ + 0.01106562465429306, + 0.01410039234906435, + 0.011072518303990364, + 0.01617506705224514, + 0.007690048310905695, + 0.003394919214770198, + 0.006384450476616621, + 0.010180247947573662, + 0.0243709497153759, + 0.010008834302425385, + 0.10905911773443222, + 0.02441471442580223, + 0.007630121894180775, + 0.00837704911828041, + 0.03464723378419876, + 0.043161921203136444, + 0.020067084580659866, + 0.016866464167833328, + 0.017428908497095108, + 0.019501792266964912, + 0.02028173767030239, + 0.011358454823493958, + 0.025563277304172516, + 0.019433431327342987, + 0.0026149533223360777, + 0.002675458323210478, + 0.0026202057488262653, + 0.0033184820786118507, + 0.002136927330866456, + 0.004000327084213495, + 0.002399063901975751, + 0.002597043290734291, + 0.006252103950828314, + 0.003440354485064745, + 0.004414880182594061, + 0.0049914284609258175, + 0.003113861195743084, + 0.0033178336452692747, + 0.004115157760679722, + 0.004988217260688543, + 0.007108438294380903, + 0.005702456459403038, + 0.005769319832324982, + 0.005633784923702478, + 0.008418554440140724, + 0.005563957151025534, + 0.007230803370475769, + 0.005273751448839903, + 0.006464701611548662, + 0.006658799014985561, + 0.008373727090656757, + 0.018127156421542168, + 0.005505471955984831, + 0.0044581154361367226, + 0.005853675771504641, + 0.016788756474852562, + 0.004373844712972641, + 0.005147200543433428, + 0.007912951521575451, + 0.00665390957146883, + 0.003931595012545586, + 0.0029407385736703873, + 0.005698471795767546, + 0.005591437686234713, + 0.006416712887585163, + 0.009822414256632328, + 0.011436793953180313, + 0.014460602775216103, + 0.004048987757414579, + 0.0026278067380189896, + 0.014672008343040943, + 0.00460960203781724, + 0.01686951518058777, + 0.017535226419568062, + 0.07201530039310455, + 0.04089079052209854, + 0.009607244282960892, + 0.013086761347949505, + 0.02674090303480625, + 0.02074703574180603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.6933376789093018 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slowly zooming wide-angle shot initially focuses on a portion of a vast gray planet. Ancient, massive craters loom in shadow on its rocky surface, their edges softened by age. The planet rotates at a majestic, almost imperceptible pace, suspended within a swirling, luminous nebula of deep purple, fiery magenta, and brilliant cyan particles. As the camera steadily pans backward, the planet's full presence gradually emerges in the center of the frame, surrounded by colorful particles that float, swirl, and intertwine like living dust. These self-luminous particles, the sole source of light in the scene, cast a vibrant, flowing hue across the planet's gray surface, outlining its majestic contours. Light and shadow slowly shift across the vast, sunken features, creating a dynamic play of light and dark. The entire scene is imbued with a sense of cosmic tranquility and grandeur, evoking an atmosphere of mysterious and epic proportions.", + "seed": 9832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_10/index_5505_seed_9832_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17385860639624298, + "vlm_logprob": -0.17385860639624298, + "vlm_yes_no_margin": 2.724999988079071, + "vlm_token_credit": -0.17385860639624298, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1034526601433754 + ], + "opposite_logprob_tokens": [ + -2.353452682495117 + ], + "logprob_reward": -0.1034526601433754, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1034526601433754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15556120872497559, + 0.18019381165504456, + 0.0949060469865799, + 0.1912703663110733, + 0.04281168058514595, + 0.21790483593940735, + 0.04720832034945488, + 0.024014972150325775, + 0.021178629249334335, + 0.02495020627975464 + ], + "credit_map": [ + 0.022080842405557632, + 0.017397256568074226, + 0.02723490074276924, + 0.02367832511663437, + 0.01303848996758461, + 0.02842031791806221, + 0.013741920702159405, + 0.009969144128262997, + 0.024046244099736214, + 0.012798543088138103, + 0.056840941309928894, + 0.05674263462424278, + 0.006870463490486145, + 0.008648018352687359, + 0.008795452304184437, + 0.005451489705592394, + 0.007551030721515417, + 0.013324720785021782, + 0.013164433650672436, + 0.01991930790245533, + 0.00712260976433754, + 0.008598771877586842, + 0.016164593398571014, + 0.009060573764145374, + 0.028958294540643692, + 0.027677569538354874, + 0.03855162113904953, + 0.025224702432751656, + 0.015968473628163338, + 0.015732405707240105, + 0.01956137642264366, + 0.019595900550484657, + 0.004501238465309143, + 0.004520390182733536, + 0.00971737876534462, + 0.003908015321940184, + 0.004254504106938839, + 0.0032633175142109394, + 0.007355322130024433, + 0.005291510373353958, + 0.024099459871649742, + 0.011662816628813744, + 0.08563315123319626, + 0.030621124431490898, + 0.007511382922530174, + 0.009601505473256111, + 0.018709108233451843, + 0.030066274106502533, + 0.0042570810765028, + 0.0031475620344281197, + 0.010155542753636837, + 0.007631688844412565, + 0.004275678191334009, + 0.005189043935388327, + 0.006989568937569857, + 0.005562150850892067, + 0.002356714801862836, + 0.002772973384708166, + 0.0024269232526421547, + 0.005491817835718393, + 0.002497229026630521, + 0.00283436244353652, + 0.0032656933180987835, + 0.002369253197684884, + 0.002138051437214017, + 0.001932498998939991, + 0.002443313132971525, + 0.004405300598591566, + 0.0021440996788442135, + 0.0019759973511099815, + 0.0032573742792010307, + 0.002881991444155574, + 0.0030261932406574488, + 0.0031508193351328373, + 0.002655474469065666, + 0.004116537980735302, + 0.0022636069916188717, + 0.002313703764230013, + 0.003112680511549115, + 0.004311188589781523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.1034526601433754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289868175983429 + ], + "opposite_logprob_tokens": [ + -1.0539867877960205 + ], + "logprob_reward": -0.4289868175983429, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289868175983429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06617069989442825, + 0.08765581995248795, + 0.06407453864812851, + 0.0777139812707901, + 0.1314059942960739, + 0.39250147342681885, + 0.08762122690677643, + 0.04663689061999321, + 0.019561581313610077, + 0.026657769456505775 + ], + "credit_map": [ + 0.008807028643786907, + 0.005136079620569944, + 0.013218049891293049, + 0.009256714023649693, + 0.005383976735174656, + 0.01434006355702877, + 0.004729283042252064, + 0.005299499724060297, + 0.007917402312159538, + 0.01117014605551958, + 0.01420271210372448, + 0.018563171848654747, + 0.007300267461687326, + 0.011054376140236855, + 0.006553482264280319, + 0.010894260369241238, + 0.007127231918275356, + 0.0075151859782636166, + 0.008602320216596127, + 0.01069510355591774, + 0.006305997725576162, + 0.008161567151546478, + 0.009195854887366295, + 0.006471275817602873, + 0.008241523057222366, + 0.00943834986537695, + 0.011815962381660938, + 0.008132225833833218, + 0.008081679232418537, + 0.008907723240554333, + 0.011439774185419083, + 0.011656747199594975, + 0.015071633271872997, + 0.016502268612384796, + 0.02855021506547928, + 0.010849068872630596, + 0.014800908043980598, + 0.011512038297951221, + 0.017676854506134987, + 0.01644301414489746, + 0.04190048202872276, + 0.01838494837284088, + 0.15014269948005676, + 0.05142572894692421, + 0.01690363883972168, + 0.01735013723373413, + 0.02998974919319153, + 0.0664040595293045, + 0.006502486299723387, + 0.004121686331927776, + 0.03201090171933174, + 0.020887499675154686, + 0.004304991103708744, + 0.0049295914359390736, + 0.006690111476927996, + 0.008173947222530842, + 0.003658975474536419, + 0.005061060655862093, + 0.004586873576045036, + 0.011255261488258839, + 0.004235231317579746, + 0.003897425951436162, + 0.00722450902685523, + 0.0067175570875406265, + 0.0021500096190720797, + 0.0018502047751098871, + 0.002392124617472291, + 0.0036225952208042145, + 0.0018826250452548265, + 0.0019333608215674758, + 0.0029037061613053083, + 0.002826954238116741, + 0.0028984795790165663, + 0.0036421017721295357, + 0.002505543641746044, + 0.004528139717876911, + 0.0023985893931239843, + 0.002457220572978258, + 0.0039680046029388905, + 0.0042596906423568726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.4289868175983429 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028255560901015997 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028255560901015997, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028255560901015997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1955879181623459, + 0.20257647335529327, + 0.08403946459293365, + 0.09011726826429367, + 0.07523566484451294, + 0.09049984812736511, + 0.05868802219629288, + 0.07348598539829254, + 0.07023939490318298, + 0.059529975056648254 + ], + "credit_map": [ + 0.03537707030773163, + 0.023873116821050644, + 0.04123777151107788, + 0.033347055315971375, + 0.014953075908124447, + 0.017442742362618446, + 0.017163341864943504, + 0.012193736620247364, + 0.011477883905172348, + 0.012035712599754333, + 0.06389744579792023, + 0.06972776353359222, + 0.007310258224606514, + 0.013184864073991776, + 0.014280385337769985, + 0.01066216267645359, + 0.005731264594942331, + 0.010844291187822819, + 0.011213723570108414, + 0.011950332671403885, + 0.006635445635765791, + 0.010961470194160938, + 0.016900911927223206, + 0.009802022017538548, + 0.011953353881835938, + 0.016339126974344254, + 0.013962633907794952, + 0.01068937312811613, + 0.009490836411714554, + 0.007671218831092119, + 0.009523372165858746, + 0.01048735436052084, + 0.009925113059580326, + 0.005667359568178654, + 0.01660669781267643, + 0.0058903819881379604, + 0.009123704396188259, + 0.00916161946952343, + 0.009237479418516159, + 0.009623309597373009, + 0.010201506316661835, + 0.006851399317383766, + 0.03047904558479786, + 0.011776669882237911, + 0.006530944723635912, + 0.005221338476985693, + 0.01030720304697752, + 0.00913174543529749, + 0.0050677708350121975, + 0.002537966938689351, + 0.020254580304026604, + 0.013251826167106628, + 0.0032669627107679844, + 0.0028937316965311766, + 0.0067906808108091354, + 0.004624501336365938, + 0.0050948732532560825, + 0.009529010392725468, + 0.011306576430797577, + 0.010497843846678734, + 0.007763444911688566, + 0.006432769820094109, + 0.011602540500462055, + 0.011258924379944801, + 0.00659392261877656, + 0.006606728304177523, + 0.008523711934685707, + 0.016726333647966385, + 0.007216155529022217, + 0.006510537583380938, + 0.010162480175495148, + 0.007899528369307518, + 0.006607424467802048, + 0.00795686338096857, + 0.005859484430402517, + 0.008614611811935902, + 0.00485061714425683, + 0.006355794612318277, + 0.009536424651741982, + 0.00974875595420599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0028255560901015997 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059643715620041 + ], + "opposite_logprob_tokens": [ + -3.895596504211426 + ], + "logprob_reward": -0.02059643715620041, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059643715620041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08381394296884537, + 0.4826698899269104, + 0.07413012534379959, + 0.11298362910747528, + 0.061569564044475555, + 0.04354758933186531, + 0.023396598175168037, + 0.06068137288093567, + 0.030277181416749954, + 0.02693014219403267 + ], + "credit_map": [ + 0.01384765189141035, + 0.007547261193394661, + 0.011282800696790218, + 0.011103264056146145, + 0.00828530266880989, + 0.014219572767615318, + 0.008554561994969845, + 0.008973520249128342, + 0.020934052765369415, + 0.01658220961689949, + 0.19898846745491028, + 0.19862577319145203, + 0.009426764212548733, + 0.011764922179281712, + 0.016104919835925102, + 0.010242770425975323, + 0.008015758357942104, + 0.00614449055865407, + 0.013830387033522129, + 0.011581212282180786, + 0.007479665335267782, + 0.00933796912431717, + 0.00803932175040245, + 0.009701316244900227, + 0.016317149624228477, + 0.01639407128095627, + 0.01525819581001997, + 0.009916696697473526, + 0.011084426194429398, + 0.012689139693975449, + 0.015573468990623951, + 0.015750475227832794, + 0.005316701717674732, + 0.008571338839828968, + 0.012203902006149292, + 0.00742427259683609, + 0.008458307944238186, + 0.005857438780367374, + 0.0068487608805298805, + 0.0068888356909155846, + 0.005181033629924059, + 0.00425704475492239, + 0.010565844364464283, + 0.006279127672314644, + 0.004289909731596708, + 0.0034277408849447966, + 0.004684958606958389, + 0.004861923865973949, + 0.002445218386128545, + 0.0014838242204859853, + 0.005409757141023874, + 0.004192121792584658, + 0.0024021416902542114, + 0.002242563758045435, + 0.002791121369227767, + 0.00242984713986516, + 0.005058642942458391, + 0.0061525520868599415, + 0.008065251633524895, + 0.017950888723134995, + 0.0044806948862969875, + 0.005106588359922171, + 0.00783957913517952, + 0.006027172319591045, + 0.0024989121593534946, + 0.003472837619483471, + 0.003399982349947095, + 0.00477784126996994, + 0.0032651827204972506, + 0.0028528182301670313, + 0.005299960728734732, + 0.00470964377745986, + 0.002470148028805852, + 0.0029099646490067244, + 0.003103566588833928, + 0.005662036594003439, + 0.002858675317838788, + 0.001877973205409944, + 0.004088467452675104, + 0.003959308378398418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.02059643715620041 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vast gray planet with ancient, massive craters visible on its rocky surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the planet rotating at a slow, almost imperceptible pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a swirling nebula containing deep purple, fiery magenta, and brilliant cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles act as the sole source of light, casting vibrant hues and outlining the planet's contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide-angle focus on a portion of the planet and then steadily pan backward to reveal the full planet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134315609931946 + ], + "opposite_logprob_tokens": [ + -1.3134315013885498 + ], + "logprob_reward": -0.3134315609931946, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3134315609931946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10796383768320084, + 0.3958764374256134, + 0.09542950242757797, + 0.0784028172492981, + 0.05337531492114067, + 0.15149079263210297, + 0.023388225585222244, + 0.032061848789453506, + 0.016814587637782097, + 0.045196615159511566 + ], + "credit_map": [ + 0.017761874943971634, + 0.009773303754627705, + 0.023103361949324608, + 0.019006989896297455, + 0.009155652485787868, + 0.009222641587257385, + 0.011692575179040432, + 0.008247449062764645, + 0.019694175571203232, + 0.02068595588207245, + 0.15236593782901764, + 0.15367844700813293, + 0.011538640595972538, + 0.009715041145682335, + 0.017662040889263153, + 0.010536238551139832, + 0.009146120399236679, + 0.011921473778784275, + 0.012129058130085468, + 0.01789913885295391, + 0.007960769347846508, + 0.010599208064377308, + 0.015893565490841866, + 0.00988017302006483, + 0.010392316617071629, + 0.01107796374708414, + 0.013535890728235245, + 0.007905257865786552, + 0.006140347570180893, + 0.009334519505500793, + 0.012459397315979004, + 0.007557125762104988, + 0.007779757492244244, + 0.006839197129011154, + 0.011104529723525047, + 0.005118828732520342, + 0.0054439823143184185, + 0.0034966168459504843, + 0.006730993278324604, + 0.006861413363367319, + 0.014757268130779266, + 0.007712865248322487, + 0.05731195956468582, + 0.019613105803728104, + 0.006228117272257805, + 0.006731241941452026, + 0.012807472608983517, + 0.026328790932893753, + 0.0036113709211349487, + 0.0018488771747797728, + 0.0046573299914598465, + 0.0035832675639539957, + 0.0018747820286080241, + 0.00193928147200495, + 0.0024979086592793465, + 0.0033754094038158655, + 0.003071030369028449, + 0.004015322308987379, + 0.0040267291478812695, + 0.004911768715828657, + 0.0038864987436681986, + 0.004108107183128595, + 0.0038782297633588314, + 0.00416416535153985, + 0.002054802607744932, + 0.0018195969751104712, + 0.0021621552295982838, + 0.0031107068061828613, + 0.0016801143065094948, + 0.0012364548165351152, + 0.002465867903083563, + 0.0022848909720778465, + 0.004916775040328503, + 0.0037038037553429604, + 0.005616090726107359, + 0.01436144020408392, + 0.0030677439644932747, + 0.0034720231778919697, + 0.00607544369995594, + 0.003983300179243088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.3134315609931946 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_11/index_tensor([3672, 3672])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_11/index_tensor([3672, 3672])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c716d806a131b3c52a8e0e6fd354201bd4de2270 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_11/index_tensor([3672, 3672])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up macro lens, employing an extremely shallow depth of field, slowly tilts the camera downward, precisely tracking a magnificent, branching snowflake. The snowflake displays a perfectly hexagonal symmetry, its six main branches sprouting countless finer, fern-like arms of ice crystals. Covered in a microscopic frost texture, its edges are outlined by a soft, cold backlight, creating an ethereal halo that shimmers like a diamond against a deep, completely out-of-focus background. It descends with a slow, graceful ballet-like movement, gently rotating around its central axis, revealing intricate detail from every angle. As it approaches, the dark ground beneath the lens—perhaps a rough, textured dark wool fabric or a patch of frozen black soil—gradually comes into focus. Finally, with feather-lightness, the snowflake silently touches the ground, its intricate crystalline structure miraculously intact at the moment of contact. The entire image evokes a magical atmosphere of utter tranquility, a focus on the beauty of the microcosm, framed by a cool, pure palette.", + "seed": 32785, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_11/index_3672_seed_32785_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03449278511106968, + "vlm_logprob": -0.03449278511106968, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03449278511106968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0039175208657979965 + ], + "opposite_logprob_tokens": [ + -5.753917694091797 + ], + "logprob_reward": -0.0039175208657979965, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0039175208657979965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24725691974163055, + 0.059388820081949234, + 0.05246369168162346, + 0.18276816606521606, + 0.08787919580936432, + 0.07958769798278809, + 0.06356263905763626, + 0.13216890394687653, + 0.023821262642741203, + 0.07110268622636795 + ], + "credit_map": [ + 0.013086897321045399, + 0.0039008434396237135, + 0.0881253108382225, + 0.012792659923434258, + 0.00453144870698452, + 0.01191143412142992, + 0.013722090050578117, + 0.0991862565279007, + 0.004404472652822733, + 0.0016397683648392558, + 0.02305532433092594, + 0.00271317339502275, + 0.001818289514631033, + 0.015730183571577072, + 0.0021292669698596, + 0.00789833813905716, + 0.006891612429171801, + 0.002680659294128418, + 0.009952283464372158, + 0.006119396071881056, + 0.0030308773275464773, + 0.009258696809411049, + 0.005003687925636768, + 0.009526479057967663, + 0.01847108267247677, + 0.010103421285748482, + 0.049004144966602325, + 0.018094448372721672, + 0.010094501078128815, + 0.03438837453722954, + 0.02233481965959072, + 0.020277364179491997, + 0.012189940549433231, + 0.009354682639241219, + 0.013348174281418324, + 0.008522579446434975, + 0.009817329235374928, + 0.012959090061485767, + 0.009544525295495987, + 0.012142880819737911, + 0.0068723042495548725, + 0.004816648084670305, + 0.020990679040551186, + 0.007212168071419001, + 0.007054930552840233, + 0.009430606849491596, + 0.006847442593425512, + 0.01636291667819023, + 0.006607561372220516, + 0.004878618288785219, + 0.008645500056445599, + 0.007859607227146626, + 0.007412800099700689, + 0.01186996977776289, + 0.008285034447908401, + 0.008003555238246918, + 0.019589070230722427, + 0.011756529100239277, + 0.01896919496357441, + 0.018306242302060127, + 0.019587654620409012, + 0.012651030905544758, + 0.016243327409029007, + 0.015065860003232956, + 0.0034548251423984766, + 0.002021291758865118, + 0.003255892777815461, + 0.002097055083140731, + 0.0019064965890720487, + 0.0030969432555139065, + 0.0024574617855250835, + 0.005531297530978918, + 0.007293407339602709, + 0.005101098213344812, + 0.012265517376363277, + 0.007254324853420258, + 0.006684985477477312, + 0.011864971369504929, + 0.008031012490391731, + 0.012607365846633911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0039175208657979965 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638650014996529 + ], + "opposite_logprob_tokens": [ + -4.882638454437256 + ], + "logprob_reward": -0.007638650014996529, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638650014996529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14996378123760223, + 0.05168366804718971, + 0.03711579740047455, + 0.1485625058412552, + 0.1975926160812378, + 0.12701497972011566, + 0.0627850741147995, + 0.0630340725183487, + 0.07224610447883606, + 0.0900014266371727 + ], + "credit_map": [ + 0.010778231546282768, + 0.003472445299848914, + 0.04881224408745766, + 0.007385075092315674, + 0.00499924598261714, + 0.010311710648238659, + 0.009067777544260025, + 0.0551370270550251, + 0.004639134276658297, + 0.002039206214249134, + 0.008176849223673344, + 0.0026095700450241566, + 0.0019761419389396906, + 0.010747467167675495, + 0.00404771463945508, + 0.01744757778942585, + 0.0038671831134706736, + 0.001999764936044812, + 0.009045066311955452, + 0.0033621303737163544, + 0.002613705350086093, + 0.006078089587390423, + 0.003576535265892744, + 0.006573317106813192, + 0.01572670042514801, + 0.005102868191897869, + 0.05067424476146698, + 0.011110397055745125, + 0.00555080734193325, + 0.03515226021409035, + 0.01533610187470913, + 0.009909115731716156, + 0.01639619842171669, + 0.011998179368674755, + 0.0424746535718441, + 0.01668575592339039, + 0.01382050383836031, + 0.023354196920990944, + 0.0191049762070179, + 0.05375814065337181, + 0.009771136566996574, + 0.007267007604241371, + 0.037114467471838, + 0.012444313615560532, + 0.007969220168888569, + 0.014627030119299889, + 0.011190823279321194, + 0.026630962267518044, + 0.008257289417088032, + 0.00413603987544775, + 0.013006710447371006, + 0.007431435398757458, + 0.004547481890767813, + 0.009712517261505127, + 0.005600388161838055, + 0.010093208402395248, + 0.0073556555435061455, + 0.005773002747446299, + 0.008269530721008778, + 0.008994178846478462, + 0.010740933939814568, + 0.0071037267334759235, + 0.006492658518254757, + 0.008304381743073463, + 0.009100085124373436, + 0.004398754332214594, + 0.018672974780201912, + 0.006630018353462219, + 0.005120635498315096, + 0.007154800463467836, + 0.0066485218703746796, + 0.014520313590765, + 0.00642541516572237, + 0.006529603153467178, + 0.02097361534833908, + 0.011417062021791935, + 0.008204366080462933, + 0.013389444909989834, + 0.010366889648139477, + 0.012695026583969593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.007638650014996529 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191635608673096 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.018191635608673096, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191635608673096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10907135903835297, + 0.07893971353769302, + 0.15940818190574646, + 0.19872552156448364, + 0.06498003005981445, + 0.040270425379276276, + 0.045082587748765945, + 0.08877740055322647, + 0.06169259548187256, + 0.15305215120315552 + ], + "credit_map": [ + 0.006606756243854761, + 0.0017642717575654387, + 0.03807005658745766, + 0.005981155671179295, + 0.0021034611854702234, + 0.005775978323072195, + 0.006128012668341398, + 0.04264167323708534, + 0.011103298515081406, + 0.004066765774041414, + 0.02088872902095318, + 0.004519753623753786, + 0.0036345780827105045, + 0.01897471956908703, + 0.007469194009900093, + 0.00828267727047205, + 0.021514171734452248, + 0.006930155213922262, + 0.03405549377202988, + 0.015284774824976921, + 0.008965788409113884, + 0.037071939557790756, + 0.013608857989311218, + 0.021977005526423454, + 0.021387849003076553, + 0.006772333290427923, + 0.062041282653808594, + 0.017830930650234222, + 0.007665927521884441, + 0.04318033531308174, + 0.019744284451007843, + 0.020102612674236298, + 0.0078169209882617, + 0.0028802966699004173, + 0.012404381297528744, + 0.004605086520314217, + 0.004357736557722092, + 0.01166713610291481, + 0.005780806299299002, + 0.015467675402760506, + 0.004047816619277, + 0.003382869530469179, + 0.006535265129059553, + 0.0050298552960157394, + 0.0041308775544166565, + 0.005200810264796019, + 0.004262091591954231, + 0.007680845446884632, + 0.004811627324670553, + 0.004316167905926704, + 0.007648896425962448, + 0.006042604800313711, + 0.0050211879424750805, + 0.005955818109214306, + 0.0056708697229623795, + 0.005615414585918188, + 0.018369734287261963, + 0.005726804491132498, + 0.01531289890408516, + 0.010550425387918949, + 0.007191120181232691, + 0.00949211698025465, + 0.010797799564898014, + 0.011336510069668293, + 0.005576363764703274, + 0.004155539441853762, + 0.012380634434521198, + 0.007551048882305622, + 0.005638389382511377, + 0.00711477966979146, + 0.007483731023967266, + 0.011792115867137909, + 0.012011216022074223, + 0.011600884608924389, + 0.037499211728572845, + 0.01769823394715786, + 0.015040463767945766, + 0.020126495510339737, + 0.014808055013418198, + 0.024267613887786865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.018191635608673096 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292670667171478 + ], + "opposite_logprob_tokens": [ + -2.237926721572876 + ], + "logprob_reward": -0.11292670667171478, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292670667171478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10296668857336044, + 0.04704904928803444, + 0.14121626317501068, + 0.10092181712388992, + 0.15160439908504486, + 0.09795930236577988, + 0.09779602289199829, + 0.08001653850078583, + 0.07465945929288864, + 0.10581053793430328 + ], + "credit_map": [ + 0.008201519958674908, + 0.003084600903093815, + 0.032212551683187485, + 0.006035275757312775, + 0.004112944472581148, + 0.007345525082200766, + 0.006937630940228701, + 0.03503663092851639, + 0.005319869611412287, + 0.0021035566460341215, + 0.004288606345653534, + 0.0026311513502150774, + 0.002320335479453206, + 0.012771700508892536, + 0.0037278919480741024, + 0.01388593390583992, + 0.01673804223537445, + 0.006617483217269182, + 0.03179250657558441, + 0.015082007274031639, + 0.0072935158386826515, + 0.030222907662391663, + 0.012809772044420242, + 0.02066001296043396, + 0.008457177318632603, + 0.003722503548488021, + 0.03539358079433441, + 0.009094489738345146, + 0.003797197248786688, + 0.020338863134384155, + 0.0115117896348238, + 0.008606203831732273, + 0.013409508392214775, + 0.010131367482244968, + 0.03132086619734764, + 0.013124414719641209, + 0.011669379658997059, + 0.018023597076535225, + 0.01459069550037384, + 0.039334554225206375, + 0.01038005668669939, + 0.005664501339197159, + 0.023633969947695732, + 0.009496916085481644, + 0.006876252591609955, + 0.013448105193674564, + 0.006940333638340235, + 0.021519159898161888, + 0.013526690192520618, + 0.00659287441521883, + 0.018136687576770782, + 0.010653621517121792, + 0.007620302494615316, + 0.020473049953579903, + 0.009462026879191399, + 0.01133076660335064, + 0.015201461501419544, + 0.005972094833850861, + 0.01044603530317545, + 0.00978311337530613, + 0.011084354482591152, + 0.009082576259970665, + 0.009539484977722168, + 0.008907405659556389, + 0.01336771808564663, + 0.005231250077486038, + 0.015043536201119423, + 0.006449608132243156, + 0.004919198341667652, + 0.009688840247690678, + 0.0058408766053617, + 0.014118429273366928, + 0.00962256733328104, + 0.00845437590032816, + 0.024384640157222748, + 0.011053684167563915, + 0.007864134386181831, + 0.016730554401874542, + 0.010052626021206379, + 0.01764794997870922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.11292670667171478 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789412394165993 + ], + "opposite_logprob_tokens": [ + -3.529789447784424 + ], + "logprob_reward": -0.029789412394165993, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789412394165993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040011532604694366, + 0.018855493515729904, + 0.12632036209106445, + 0.12394961714744568, + 0.08945699036121368, + 0.10361483693122864, + 0.07624883949756622, + 0.083955317735672, + 0.07901500910520554, + 0.258571982383728 + ], + "credit_map": [ + 0.003335877088829875, + 0.0010645120637491345, + 0.01194525696337223, + 0.002642429666593671, + 0.0014927476877346635, + 0.0034109442494809628, + 0.0025915605947375298, + 0.013528207316994667, + 0.0023476663045585155, + 0.0012476552510634065, + 0.0019897136371582747, + 0.0016400478780269623, + 0.001310149789787829, + 0.004798030015081167, + 0.0021071394439786673, + 0.0034150893334299326, + 0.014474193565547466, + 0.007361540105193853, + 0.020170805975794792, + 0.009855730459094048, + 0.008984499610960484, + 0.02060631476342678, + 0.014493998140096664, + 0.030373280867934227, + 0.012661750428378582, + 0.003902493743225932, + 0.04246765002608299, + 0.008690995164215565, + 0.004413073416799307, + 0.02754126489162445, + 0.012706390582025051, + 0.011566005647182465, + 0.008207679726183414, + 0.005768209230154753, + 0.018429197371006012, + 0.00754866749048233, + 0.006386599503457546, + 0.011517397128045559, + 0.008189503103494644, + 0.02340974099934101, + 0.008015237748622894, + 0.006036432925611734, + 0.031080031767487526, + 0.010360883548855782, + 0.006071567535400391, + 0.010490440763533115, + 0.007172002457082272, + 0.024388238787651062, + 0.009097544476389885, + 0.005731531418859959, + 0.010683253407478333, + 0.009835748933255672, + 0.0080699622631073, + 0.016105178743600845, + 0.009256996214389801, + 0.007468621712177992, + 0.00955123733729124, + 0.0050876559689641, + 0.01385557185858488, + 0.01912076026201248, + 0.006194045767188072, + 0.006387558300048113, + 0.01130620576441288, + 0.012452276423573494, + 0.01075337827205658, + 0.00478350929915905, + 0.013674579560756683, + 0.00749922776594758, + 0.005194143392145634, + 0.013046115636825562, + 0.006075895857065916, + 0.017988160252571106, + 0.022631842643022537, + 0.016478050500154495, + 0.07676906883716583, + 0.03284193575382233, + 0.017711374908685684, + 0.029549632221460342, + 0.023403620347380638, + 0.03918645530939102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.029789412394165993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro lens, employing an extremely shallow depth of field, slowly tilts the camera downward, precisely tracking a magnificent, branching snowflake. The snowflake displays a perfectly hexagonal symmetry, its six main branches sprouting countless finer, fern-like arms of ice crystals. Covered in a microscopic frost texture, its edges are outlined by a soft, cold backlight, creating an ethereal halo that shimmers like a diamond against a deep, completely out-of-focus background. It descends with a slow, graceful ballet-like movement, gently rotating around its central axis, revealing intricate detail from every angle. As it approaches, the dark ground beneath the lens—perhaps a rough, textured dark wool fabric or a patch of frozen black soil—gradually comes into focus. Finally, with feather-lightness, the snowflake silently touches the ground, its intricate crystalline structure miraculously intact at the moment of contact. The entire image evokes a magical atmosphere of utter tranquility, a focus on the beauty of the microcosm, framed by a cool, pure palette.", + "seed": 32786, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_11/index_3672_seed_32786_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04116527726873755, + "vlm_logprob": -0.04116527726873755, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04116527726873755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004541796166449785 + ], + "opposite_logprob_tokens": [ + -5.754541873931885 + ], + "logprob_reward": -0.004541796166449785, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004541796166449785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047356776893138885, + 0.08397477865219116, + 0.12322744727134705, + 0.12219155579805374, + 0.05358569696545601, + 0.20254577696323395, + 0.09677372127771378, + 0.04510506987571716, + 0.08549091219902039, + 0.1397482007741928 + ], + "credit_map": [ + 0.006648157257586718, + 0.004568108357489109, + 0.008196836337447166, + 0.004702404141426086, + 0.004328024107962847, + 0.008049355819821358, + 0.003404476447030902, + 0.007459415588527918, + 0.013029706664383411, + 0.004435492679476738, + 0.016545282676815987, + 0.005285111255943775, + 0.005772137548774481, + 0.014947982504963875, + 0.009099792689085007, + 0.014859281480312347, + 0.0168521236628294, + 0.008605328388512135, + 0.012746643275022507, + 0.013339921832084656, + 0.006885594222694635, + 0.014905817806720734, + 0.017712973058223724, + 0.03217906132340431, + 0.015244177542626858, + 0.01363453734666109, + 0.01889886148273945, + 0.01567068323493004, + 0.013008611276745796, + 0.019995765760540962, + 0.015511400066316128, + 0.01022752933204174, + 0.00882150512188673, + 0.0072500440292060375, + 0.005642625503242016, + 0.007980170659720898, + 0.006300533190369606, + 0.005171998869627714, + 0.00601342273876071, + 0.0064054024405777454, + 0.028381019830703735, + 0.05201631784439087, + 0.014515585266053677, + 0.01837898977100849, + 0.025767449289560318, + 0.014147762209177017, + 0.017423486337065697, + 0.03191518783569336, + 0.009991545230150223, + 0.014327348209917545, + 0.018110865727066994, + 0.015744013711810112, + 0.007994275540113449, + 0.007325825747102499, + 0.013166752643883228, + 0.01011310052126646, + 0.006120331585407257, + 0.0038058545906096697, + 0.0047890073619782925, + 0.005904007703065872, + 0.00887615792453289, + 0.00436430424451828, + 0.006834970787167549, + 0.004410437773913145, + 0.008090625517070293, + 0.007089570630341768, + 0.008653267286717892, + 0.014202963560819626, + 0.0133917685598135, + 0.01215356308966875, + 0.013815172016620636, + 0.008093982934951782, + 0.010020690970122814, + 0.009302089922130108, + 0.012118135578930378, + 0.009967460297048092, + 0.037687066942453384, + 0.039280008524656296, + 0.010717938654124737, + 0.010654808022081852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.004541796166449785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865198578685522 + ], + "opposite_logprob_tokens": [ + -4.758652210235596 + ], + "logprob_reward": -0.00865198578685522, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865198578685522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06556494534015656, + 0.038853421807289124, + 0.0771365538239479, + 0.12624897062778473, + 0.06664492189884186, + 0.1934196650981903, + 0.10057324171066284, + 0.10723219066858292, + 0.10690270364284515, + 0.11742337048053741 + ], + "credit_map": [ + 0.008356591686606407, + 0.0043570538982748985, + 0.010651905089616776, + 0.010977990925312042, + 0.006085102912038565, + 0.00887121818959713, + 0.007286753971129656, + 0.008978325873613358, + 0.005065080709755421, + 0.0025539842899888754, + 0.006705426145344973, + 0.002792482730001211, + 0.0029276448767632246, + 0.005071836523711681, + 0.0042217811569571495, + 0.009515183977782726, + 0.012018434703350067, + 0.003902064636349678, + 0.007837814278900623, + 0.007134396117180586, + 0.005031661596149206, + 0.00924560148268938, + 0.010519620962440968, + 0.021446960046887398, + 0.018062790855765343, + 0.010506875813007355, + 0.015237701125442982, + 0.018712420016527176, + 0.012344103306531906, + 0.020709650591015816, + 0.015003493055701256, + 0.01567193679511547, + 0.007458768784999847, + 0.007165464572608471, + 0.014774249866604805, + 0.007679452188313007, + 0.006519846152514219, + 0.008045965805649757, + 0.0073894658125936985, + 0.007611712906509638, + 0.026340393349528313, + 0.06206147000193596, + 0.011935087852180004, + 0.013429397717118263, + 0.03252671658992767, + 0.012873484753072262, + 0.010403952561318874, + 0.023849163204431534, + 0.011829772032797337, + 0.014251286163926125, + 0.017146846279501915, + 0.015433554537594318, + 0.00837092474102974, + 0.0063816457986831665, + 0.0161051694303751, + 0.01105404831469059, + 0.017101874575018883, + 0.00870532263070345, + 0.009341117925941944, + 0.011806776747107506, + 0.024641072377562523, + 0.014863775111734867, + 0.011641503311693668, + 0.00913074892014265, + 0.00943736545741558, + 0.009946174919605255, + 0.009542128071188927, + 0.011604679748415947, + 0.02328677661716938, + 0.017695926129817963, + 0.01348546426743269, + 0.011904189363121986, + 0.011859041638672352, + 0.0071531031280756, + 0.010684177279472351, + 0.0062608057633042336, + 0.030751371756196022, + 0.035648804157972336, + 0.007826016284525394, + 0.007240049075335264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.00865198578685522 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048662953078746796 + ], + "opposite_logprob_tokens": [ + -3.0486629009246826 + ], + "logprob_reward": -0.048662953078746796, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048662953078746796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04786832630634308, + 0.036824747920036316, + 0.05970624461770058, + 0.17505450546741486, + 0.04306790605187416, + 0.35981234908103943, + 0.06375207006931305, + 0.025440607219934464, + 0.12286520004272461, + 0.06560802459716797 + ], + "credit_map": [ + 0.008683035150170326, + 0.003439430147409439, + 0.008688393048942089, + 0.0047904206439852715, + 0.0042951651848852634, + 0.0071778809651732445, + 0.003671458223834634, + 0.007122538983821869, + 0.004884996451437473, + 0.003176713129505515, + 0.0054223318584263325, + 0.004750675987452269, + 0.004034822341054678, + 0.004017178434878588, + 0.005677443463355303, + 0.004860586952418089, + 0.009892904199659824, + 0.003216381184756756, + 0.009409346617758274, + 0.004901227541267872, + 0.004148680251091719, + 0.010473222471773624, + 0.007460635155439377, + 0.010203849524259567, + 0.01990828849375248, + 0.009988841600716114, + 0.050411421805620193, + 0.021872704848647118, + 0.010772740468382835, + 0.036939896643161774, + 0.014097223989665508, + 0.011063391342759132, + 0.006499660667032003, + 0.008633910678327084, + 0.003953319042921066, + 0.005291301291435957, + 0.006131969857960939, + 0.003701198613271117, + 0.004730515647679567, + 0.004126028623431921, + 0.05235963687300682, + 0.08151667565107346, + 0.026703301817178726, + 0.03628432750701904, + 0.040630992501974106, + 0.029574207961559296, + 0.03521358594298363, + 0.05752963945269585, + 0.005611305125057697, + 0.006117725279182196, + 0.010565838776528835, + 0.01080561988055706, + 0.005898099858313799, + 0.006325871217995882, + 0.011179647408425808, + 0.00724796624854207, + 0.002998508047312498, + 0.0027540968731045723, + 0.003169652074575424, + 0.0034245671704411507, + 0.0030396964866667986, + 0.0026774881407618523, + 0.004217559937387705, + 0.0031590373255312443, + 0.011478645727038383, + 0.008845214731991291, + 0.007027113810181618, + 0.009347223676741123, + 0.04220221936702728, + 0.02697306126356125, + 0.0103982612490654, + 0.0065934667363762856, + 0.006720879580825567, + 0.0038186314050108194, + 0.008765818551182747, + 0.0045973435044288635, + 0.013646276667714119, + 0.01746722310781479, + 0.004559921100735664, + 0.006031929049640894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.048662953078746796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903183043003082 + ], + "opposite_logprob_tokens": [ + -2.464031934738159 + ], + "logprob_reward": -0.08903183043003082, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903183043003082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051089800894260406, + 0.040057625621557236, + 0.07769953459501266, + 0.2142839878797531, + 0.11140619218349457, + 0.14438018202781677, + 0.16134141385555267, + 0.07113979756832123, + 0.05304013565182686, + 0.0755612701177597 + ], + "credit_map": [ + 0.007829656824469566, + 0.0027932487428188324, + 0.011026675812900066, + 0.004244986921548843, + 0.004017635714262724, + 0.008429297246038914, + 0.0037330384366214275, + 0.009015260264277458, + 0.0052267396822571754, + 0.002303815446794033, + 0.007372044958174229, + 0.0023384029045701027, + 0.0030567629728466272, + 0.0073197977617383, + 0.003167301882058382, + 0.009272758848965168, + 0.010792997665703297, + 0.0030124569311738014, + 0.01959720253944397, + 0.006421573925763369, + 0.0038687658961862326, + 0.0176224485039711, + 0.005936221219599247, + 0.010447870008647442, + 0.025140631943941116, + 0.024571608752012253, + 0.0445677787065506, + 0.022302355617284775, + 0.02161712571978569, + 0.03943225368857384, + 0.025548644363880157, + 0.011103593744337559, + 0.028617186471819878, + 0.009826535359025002, + 0.010516541078686714, + 0.00809566117823124, + 0.010209516622126102, + 0.02118869684636593, + 0.010201203636825085, + 0.012750848196446896, + 0.01986679621040821, + 0.02869778871536255, + 0.011306926608085632, + 0.014805457554757595, + 0.016703203320503235, + 0.014530496671795845, + 0.014786163344979286, + 0.02368335984647274, + 0.01462896354496479, + 0.016021709889173508, + 0.021780092269182205, + 0.031201234087347984, + 0.014822193421423435, + 0.009411044418811798, + 0.03365267440676689, + 0.019823500886559486, + 0.00914438534528017, + 0.006916029844433069, + 0.006347036920487881, + 0.00797965470701456, + 0.012662450782954693, + 0.01372316014021635, + 0.008657344616949558, + 0.00570973614230752, + 0.005912782624363899, + 0.0039586592465639114, + 0.006596957799047232, + 0.007298230659216642, + 0.008430366404354572, + 0.006826993077993393, + 0.006938256323337555, + 0.007077894639223814, + 0.00810557696968317, + 0.00526277394965291, + 0.00967175792902708, + 0.006011537276208401, + 0.015375332906842232, + 0.01843428984284401, + 0.004563117399811745, + 0.008136887103319168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.08903183043003082 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493782088160515 + ], + "opposite_logprob_tokens": [ + -2.9299378395080566 + ], + "logprob_reward": -0.05493782088160515, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493782088160515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028568558394908905, + 0.010050179436802864, + 0.04157157614827156, + 0.24163557589054108, + 0.051706429570913315, + 0.16149286925792694, + 0.0617484487593174, + 0.051636677235364914, + 0.12442225217819214, + 0.2271675020456314 + ], + "credit_map": [ + 0.004705087281763554, + 0.001715344376862049, + 0.004003952257335186, + 0.004709905479103327, + 0.002223791554570198, + 0.004165878519415855, + 0.0028870422393083572, + 0.004157555755227804, + 0.0013035814044997096, + 0.0006065793568268418, + 0.0022221312392503023, + 0.0009152452694252133, + 0.0007149895536713302, + 0.0017592079238966107, + 0.0009071773383766413, + 0.001621265895664692, + 0.005967533215880394, + 0.0027867371682077646, + 0.004891906399279833, + 0.004120645113289356, + 0.0024744223337620497, + 0.005962652619928122, + 0.005336708854883909, + 0.01003097090870142, + 0.02526448853313923, + 0.02615799941122532, + 0.037928394973278046, + 0.043630558997392654, + 0.03052215836942196, + 0.0287749283015728, + 0.03742377087473869, + 0.011933248490095139, + 0.011723216623067856, + 0.005632841028273106, + 0.005663969554007053, + 0.004409074317663908, + 0.004796912427991629, + 0.00820955354720354, + 0.004796702414751053, + 0.006474150810390711, + 0.026755163446068764, + 0.02676839753985405, + 0.014668845571577549, + 0.01674053631722927, + 0.017647894099354744, + 0.016479672864079475, + 0.01720510423183441, + 0.02522723376750946, + 0.005967347417026758, + 0.006894264370203018, + 0.011490684933960438, + 0.009702949784696102, + 0.006605681963264942, + 0.005635428242385387, + 0.009263519197702408, + 0.006188561208546162, + 0.009431017562747002, + 0.003912067506462336, + 0.004168947227299213, + 0.005610964726656675, + 0.010308338329195976, + 0.006922880187630653, + 0.007090984843671322, + 0.004191470332443714, + 0.009970750659704208, + 0.012073321267962456, + 0.011599644087255001, + 0.010650553740561008, + 0.03617779165506363, + 0.022889163345098495, + 0.009261342696845531, + 0.011799677275121212, + 0.023616736754775047, + 0.013017753139138222, + 0.027216583490371704, + 0.014310847036540508, + 0.05091700702905655, + 0.06535957753658295, + 0.013764600269496441, + 0.018964355811476707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.05493782088160515 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro lens, employing an extremely shallow depth of field, slowly tilts the camera downward, precisely tracking a magnificent, branching snowflake. The snowflake displays a perfectly hexagonal symmetry, its six main branches sprouting countless finer, fern-like arms of ice crystals. Covered in a microscopic frost texture, its edges are outlined by a soft, cold backlight, creating an ethereal halo that shimmers like a diamond against a deep, completely out-of-focus background. It descends with a slow, graceful ballet-like movement, gently rotating around its central axis, revealing intricate detail from every angle. As it approaches, the dark ground beneath the lens—perhaps a rough, textured dark wool fabric or a patch of frozen black soil—gradually comes into focus. Finally, with feather-lightness, the snowflake silently touches the ground, its intricate crystalline structure miraculously intact at the moment of contact. The entire image evokes a magical atmosphere of utter tranquility, a focus on the beauty of the microcosm, framed by a cool, pure palette.", + "seed": 32787, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_11/index_3672_seed_32787_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02508743219077587, + "vlm_logprob": -0.02508743219077587, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.02508743219077587, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004340394400060177 + ], + "opposite_logprob_tokens": [ + -5.879340171813965 + ], + "logprob_reward": -0.004340394400060177, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004340394400060177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059061601758003235, + 0.03120013140141964, + 0.0713295042514801, + 0.37168121337890625, + 0.02169661782681942, + 0.03249218314886093, + 0.02176177129149437, + 0.08138677477836609, + 0.10102005302906036, + 0.2083701640367508 + ], + "credit_map": [ + 0.007254006341099739, + 0.009957184083759785, + 0.007161190267652273, + 0.006963054183870554, + 0.007361677475273609, + 0.004783377051353455, + 0.009557993151247501, + 0.006023111753165722, + 0.0038107007276266813, + 0.0030235894955694675, + 0.0028997326735407114, + 0.005351629573851824, + 0.003995017148554325, + 0.0059022861532866955, + 0.003269789507612586, + 0.0029473842587321997, + 0.006286695133894682, + 0.007461938541382551, + 0.012216160073876381, + 0.00705003971233964, + 0.01349441148340702, + 0.01295399572700262, + 0.0075100320391356945, + 0.004356229677796364, + 0.060428373515605927, + 0.035381220281124115, + 0.05453198403120041, + 0.02928604744374752, + 0.060759708285331726, + 0.06528937071561813, + 0.025142822414636612, + 0.040861669927835464, + 0.0021821078844368458, + 0.002949950285255909, + 0.0025670717004686594, + 0.0026194360107183456, + 0.003609882900491357, + 0.0030093095265328884, + 0.0027482989244163036, + 0.002010558033362031, + 0.008403614163398743, + 0.004836368840187788, + 0.0030611383263021708, + 0.0037732215132564306, + 0.0033703239168971777, + 0.004483127035200596, + 0.002788771642372012, + 0.0017756137531250715, + 0.0034241846296936274, + 0.001831560512073338, + 0.002544299466535449, + 0.0018789507448673248, + 0.0028258529491722584, + 0.0028363496530801058, + 0.0035225884057581425, + 0.0028979824855923653, + 0.009602864272892475, + 0.009752188809216022, + 0.010096791200339794, + 0.01129057165235281, + 0.012448812834918499, + 0.013022112660109997, + 0.007360814604908228, + 0.00781261082738638, + 0.013097616843879223, + 0.016577670350670815, + 0.013504899106919765, + 0.00968839880079031, + 0.008526789024472237, + 0.012714093551039696, + 0.015016665682196617, + 0.011893908493220806, + 0.007862080819904804, + 0.05788816139101982, + 0.01905536651611328, + 0.01208208967000246, + 0.015478197485208511, + 0.017193203791975975, + 0.06178181245923042, + 0.017029231414198875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.004340394400060177 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647141046822071 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008647141046822071, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647141046822071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0913524255156517, + 0.047619596123695374, + 0.09797899425029755, + 0.0807856023311615, + 0.032376017421483994, + 0.025197526440024376, + 0.02283453568816185, + 0.26055651903152466, + 0.17701560258865356, + 0.16428327560424805 + ], + "credit_map": [ + 0.01094760186970234, + 0.012447178363800049, + 0.00927822757512331, + 0.010116019286215305, + 0.011326530948281288, + 0.007707587443292141, + 0.0171325895935297, + 0.012396679259836674, + 0.009916813112795353, + 0.004308069124817848, + 0.005420522764325142, + 0.007565315347164869, + 0.00581524008885026, + 0.003982124850153923, + 0.005664613097906113, + 0.004946892615407705, + 0.006715834140777588, + 0.008468328975141048, + 0.011452826671302319, + 0.0076271407306194305, + 0.025991223752498627, + 0.023758748546242714, + 0.00926855020225048, + 0.004696335177868605, + 0.008807952515780926, + 0.008229607716202736, + 0.008323629386723042, + 0.01100530568510294, + 0.014809914864599705, + 0.014302852563560009, + 0.007713017053902149, + 0.007593317423015833, + 0.0033632328268140554, + 0.0034480192698538303, + 0.0035427226684987545, + 0.0032383904326707125, + 0.004990092478692532, + 0.005833593662828207, + 0.0038793424610048532, + 0.004080622922629118, + 0.0045856898650527, + 0.003300504991784692, + 0.002016096143051982, + 0.0035420663189142942, + 0.0033773216418921947, + 0.003627604804933071, + 0.002738011535257101, + 0.0020102295093238354, + 0.0027841508854180574, + 0.0036612136755138636, + 0.0023509319871664047, + 0.0023666624911129475, + 0.0031845010817050934, + 0.003201585728675127, + 0.002763018710538745, + 0.002522470895200968, + 0.021496156230568886, + 0.018189558759331703, + 0.0222214013338089, + 0.016329962760210037, + 0.024965358898043633, + 0.07525776326656342, + 0.017273657023906708, + 0.06482263654470444, + 0.020880352705717087, + 0.032259345054626465, + 0.023648153990507126, + 0.01757206954061985, + 0.018438642844557762, + 0.0245558749884367, + 0.020025815814733505, + 0.01963534764945507, + 0.008113092742860317, + 0.037280935794115067, + 0.018115002661943436, + 0.011545775458216667, + 0.012905409559607506, + 0.013093207031488419, + 0.04713751748204231, + 0.016092319041490555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.008647141046822071 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634863369166851 + ], + "opposite_logprob_tokens": [ + -3.65134859085083 + ], + "logprob_reward": -0.02634863369166851, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634863369166851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0635155439376831, + 0.03234843164682388, + 0.03138745203614235, + 0.35913848876953125, + 0.02019161358475685, + 0.02408468723297119, + 0.029537027701735497, + 0.2552182972431183, + 0.10582223534584045, + 0.07875623553991318 + ], + "credit_map": [ + 0.007319942116737366, + 0.0059777358546853065, + 0.005241999868303537, + 0.006910374388098717, + 0.007205181289464235, + 0.008614598773419857, + 0.01505163125693798, + 0.007194074336439371, + 0.0074911704286932945, + 0.0035593591164797544, + 0.0031917777378112078, + 0.004793859552592039, + 0.004727718885987997, + 0.0034448886290192604, + 0.0025024048518389463, + 0.0026372564025223255, + 0.0026067723520100117, + 0.002344145905226469, + 0.0039601377211511135, + 0.001752043841406703, + 0.009217983111739159, + 0.0064867581240832806, + 0.002985537750646472, + 0.0020340727642178535, + 0.05795690417289734, + 0.032466448843479156, + 0.04908972978591919, + 0.025682708248496056, + 0.05800991132855415, + 0.06392961740493774, + 0.025854753330349922, + 0.04614841938018799, + 0.0016759666614234447, + 0.002686993218958378, + 0.0019472942221909761, + 0.0020516947843134403, + 0.0033227549865841866, + 0.003137608291581273, + 0.0027458604890853167, + 0.0026234404649585485, + 0.004864055663347244, + 0.004184281453490257, + 0.002471502171829343, + 0.0015997488517314196, + 0.0025155821349471807, + 0.005204873625189066, + 0.0013775378465652466, + 0.001867105020210147, + 0.004653770476579666, + 0.0034958140458911657, + 0.004011582117527723, + 0.002183258533477783, + 0.003688880940899253, + 0.0035777080338448286, + 0.0038656597025692463, + 0.004060354549437761, + 0.019655819982290268, + 0.019528813660144806, + 0.020477840676903725, + 0.022351130843162537, + 0.02326778508722782, + 0.07818113267421722, + 0.015458425506949425, + 0.05629736930131912, + 0.014756159856915474, + 0.014603547751903534, + 0.012916610576212406, + 0.011400601826608181, + 0.010511894710361958, + 0.014362823218107224, + 0.011379742063581944, + 0.015890857204794884, + 0.0063004884868860245, + 0.009811149910092354, + 0.01156522985547781, + 0.00785736832767725, + 0.008162743411958218, + 0.006712323985993862, + 0.017138540744781494, + 0.011208392679691315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.02634863369166851 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305529221892357 + ], + "opposite_logprob_tokens": [ + -3.168055295944214 + ], + "logprob_reward": -0.04305529221892357, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305529221892357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058092519640922546, + 0.04125967621803284, + 0.07015490531921387, + 0.2587641775608063, + 0.020065797492861748, + 0.02553984895348549, + 0.054577216506004333, + 0.2237919270992279, + 0.19376248121261597, + 0.05399145185947418 + ], + "credit_map": [ + 0.006896267645061016, + 0.009050673805177212, + 0.005391295067965984, + 0.006148173473775387, + 0.005710814148187637, + 0.004986614920198917, + 0.011236030608415604, + 0.008672650903463364, + 0.007744638714939356, + 0.003977445885539055, + 0.00401076627895236, + 0.004497216548770666, + 0.005735136568546295, + 0.005874750204384327, + 0.004329327028244734, + 0.005090396385639906, + 0.0027717738412320614, + 0.0034727519378066063, + 0.004269603174179792, + 0.0030727218836545944, + 0.026441950350999832, + 0.02338564768433571, + 0.003547950414940715, + 0.003192502073943615, + 0.04187430068850517, + 0.025141729041934013, + 0.03794720768928528, + 0.023690631613135338, + 0.03743933141231537, + 0.04078815132379532, + 0.019846489652991295, + 0.032036345452070236, + 0.0018856264650821686, + 0.0032586127053946257, + 0.001967416610568762, + 0.0023559362161904573, + 0.002599027706310153, + 0.003415099810808897, + 0.0028371603693813086, + 0.0017469179583713412, + 0.005856744479387999, + 0.004133862443268299, + 0.002594104502350092, + 0.002121214522048831, + 0.002810300327837467, + 0.004502349067479372, + 0.00165157753508538, + 0.0018696957267820835, + 0.009097686037421227, + 0.007429556455463171, + 0.0070311762392520905, + 0.004161536227911711, + 0.00618104450404644, + 0.007716318592429161, + 0.005183715373277664, + 0.007776178885251284, + 0.015138349495828152, + 0.017988741397857666, + 0.014551674015820026, + 0.02054428681731224, + 0.019277973100543022, + 0.06707485020160675, + 0.01404653862118721, + 0.055169519037008286, + 0.02396601438522339, + 0.030172551050782204, + 0.02746044285595417, + 0.019916214048862457, + 0.02066376619040966, + 0.027835367247462273, + 0.021087680011987686, + 0.02266044169664383, + 0.003972786944359541, + 0.007320206146687269, + 0.00919707678258419, + 0.004944218322634697, + 0.005771681200712919, + 0.0048782965168356895, + 0.011672718450427055, + 0.006234467029571533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.04305529221892357 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304569959640503 + ], + "opposite_logprob_tokens": [ + -3.16804575920105 + ], + "logprob_reward": -0.04304569959640503, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304569959640503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09057475626468658, + 0.01994938775897026, + 0.10147048532962799, + 0.20174361765384674, + 0.02065717615187168, + 0.022174270823597908, + 0.016681699082255363, + 0.10424034297466278, + 0.23674574494361877, + 0.18576249480247498 + ], + "credit_map": [ + 0.0067716664634644985, + 0.016547489911317825, + 0.007095749955624342, + 0.01085756067186594, + 0.016305897384881973, + 0.010347188450396061, + 0.016325170174241066, + 0.006324032787233591, + 0.003494713455438614, + 0.0025331934448331594, + 0.0023170323111116886, + 0.0023642946034669876, + 0.002524434356018901, + 0.0021004334557801485, + 0.001975405029952526, + 0.002639881568029523, + 0.006189478561282158, + 0.006511937361210585, + 0.012170360423624516, + 0.007107596844434738, + 0.03027763031423092, + 0.028075315058231354, + 0.006676022429019213, + 0.004462144337594509, + 0.034420035779476166, + 0.019082659855484962, + 0.026963043957948685, + 0.014934731647372246, + 0.032610680907964706, + 0.03512837365269661, + 0.014928814955055714, + 0.023675277829170227, + 0.002577323466539383, + 0.00265660067088902, + 0.001553666079416871, + 0.0022409057710319757, + 0.002874441910535097, + 0.004306969698518515, + 0.0021338947117328644, + 0.0023133736103773117, + 0.004135428462177515, + 0.003224477404728532, + 0.00223405659198761, + 0.0019976338371634483, + 0.0023158302064985037, + 0.00460992194712162, + 0.0017919405363500118, + 0.0018649804405868053, + 0.002128831809386611, + 0.0017324094660580158, + 0.0021518294233828783, + 0.0016261470736935735, + 0.0019177417270839214, + 0.002557287225499749, + 0.002179448725655675, + 0.00238800304941833, + 0.007946048863232136, + 0.011029664427042007, + 0.008273273706436157, + 0.013496942818164825, + 0.0111995292827487, + 0.022964658215641975, + 0.008201648481190205, + 0.02112858183681965, + 0.03288164734840393, + 0.03901468962430954, + 0.030264414846897125, + 0.02390618994832039, + 0.023625578731298447, + 0.03011375293135643, + 0.029123350977897644, + 0.02781612053513527, + 0.017006246373057365, + 0.017796218395233154, + 0.034603338688611984, + 0.019491741433739662, + 0.020458552986383438, + 0.016137870028614998, + 0.03551075980067253, + 0.02475776895880699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.04304569959640503 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro lens, employing an extremely shallow depth of field, slowly tilts the camera downward, precisely tracking a magnificent, branching snowflake. The snowflake displays a perfectly hexagonal symmetry, its six main branches sprouting countless finer, fern-like arms of ice crystals. Covered in a microscopic frost texture, its edges are outlined by a soft, cold backlight, creating an ethereal halo that shimmers like a diamond against a deep, completely out-of-focus background. It descends with a slow, graceful ballet-like movement, gently rotating around its central axis, revealing intricate detail from every angle. As it approaches, the dark ground beneath the lens—perhaps a rough, textured dark wool fabric or a patch of frozen black soil—gradually comes into focus. Finally, with feather-lightness, the snowflake silently touches the ground, its intricate crystalline structure miraculously intact at the moment of contact. The entire image evokes a magical atmosphere of utter tranquility, a focus on the beauty of the microcosm, framed by a cool, pure palette.", + "seed": 32788, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_11/index_3672_seed_32788_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.053743588645011185, + "vlm_logprob": -0.053743588645011185, + "vlm_yes_no_margin": 3.5999999761581423, + "vlm_token_credit": -0.053743588645011185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041607958264648914 + ], + "opposite_logprob_tokens": [ + -5.8791608810424805 + ], + "logprob_reward": -0.0041607958264648914, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0041607958264648914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054855309426784515, + 0.053676679730415344, + 0.04864213988184929, + 0.0425102598965168, + 0.023337718099355698, + 0.026609811931848526, + 0.06101416051387787, + 0.13132961094379425, + 0.05204132944345474, + 0.5059830546379089 + ], + "credit_map": [ + 0.007319728843867779, + 0.00470120320096612, + 0.010280665010213852, + 0.005002988036721945, + 0.005467098671942949, + 0.005397467873990536, + 0.010243618860840797, + 0.006442535202950239, + 0.00651851249858737, + 0.003577473573386669, + 0.005072377156466246, + 0.0031521038617938757, + 0.006580121349543333, + 0.015056492760777473, + 0.005373004358261824, + 0.008346588350832462, + 0.004771376494318247, + 0.004830436781048775, + 0.006594516336917877, + 0.004104448948055506, + 0.006925057619810104, + 0.01051362045109272, + 0.004863596986979246, + 0.006039082538336515, + 0.003790279384702444, + 0.0029755171854048967, + 0.007495023310184479, + 0.0029256013222038746, + 0.0040292623452842236, + 0.01092467550188303, + 0.0030873825307935476, + 0.007282514590770006, + 0.002249110257253051, + 0.0025248159654438496, + 0.004144848790019751, + 0.00383597519248724, + 0.002123673912137747, + 0.0033333622850477695, + 0.0020505287684500217, + 0.0030754003673791885, + 0.0028372458182275295, + 0.0032361221965402365, + 0.0030839734245091677, + 0.004517814610153437, + 0.003534815739840269, + 0.0029767530504614115, + 0.0036052910145372152, + 0.002817792585119605, + 0.0073262290097773075, + 0.0065524079836905, + 0.006252642720937729, + 0.008914170786738396, + 0.0055537596344947815, + 0.007158303167670965, + 0.008301684632897377, + 0.010954955592751503, + 0.014709312468767166, + 0.0128830811008811, + 0.033424388617277145, + 0.011654827743768692, + 0.011073670350015163, + 0.019892454147338867, + 0.015395371243357658, + 0.01229647547006607, + 0.00533212861046195, + 0.005550270434468985, + 0.009046312421560287, + 0.006372631527483463, + 0.006051093805581331, + 0.0063576172105968, + 0.006326448637992144, + 0.007004819344729185, + 0.04865279048681259, + 0.16128762066364288, + 0.04104900360107422, + 0.017651580274105072, + 0.037382327020168304, + 0.1456080824136734, + 0.014236430637538433, + 0.04011518135666847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.0041607958264648914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080989614129066 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011080989614129066, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080989614129066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17623256146907806, + 0.16529425978660583, + 0.06256453692913055, + 0.027135400101542473, + 0.049028076231479645, + 0.018171586096286774, + 0.09883351624011993, + 0.14655008912086487, + 0.08333156257867813, + 0.17285841703414917 + ], + "credit_map": [ + 0.02479051984846592, + 0.00656937574967742, + 0.007384628523141146, + 0.009331733919680119, + 0.008657321333885193, + 0.05794486403465271, + 0.012566610239446163, + 0.048987507820129395, + 0.018334386870265007, + 0.009200467728078365, + 0.011788939125835896, + 0.0077703120186924934, + 0.01886293664574623, + 0.053237639367580414, + 0.016415027901530266, + 0.029684554785490036, + 0.005453893914818764, + 0.00650250818580389, + 0.00811198353767395, + 0.005797383841127157, + 0.00870539527386427, + 0.012953992001712322, + 0.007052133791148663, + 0.007987244985997677, + 0.002581266453489661, + 0.0022450622636824846, + 0.00531470961868763, + 0.00239549670368433, + 0.0022797556594014168, + 0.005380003713071346, + 0.0024474121164530516, + 0.004491693340241909, + 0.0034811352379620075, + 0.002857346786186099, + 0.02023421972990036, + 0.003539362223818898, + 0.002269453601911664, + 0.008643506094813347, + 0.002671043621376157, + 0.00533200753852725, + 0.002380774822086096, + 0.002067637862637639, + 0.0037091549020260572, + 0.002413216046988964, + 0.001917739980854094, + 0.001843422302044928, + 0.0016522094374522567, + 0.002187429927289486, + 0.010790287517011166, + 0.009534177370369434, + 0.011964373290538788, + 0.013015777803957462, + 0.01020965725183487, + 0.011236490681767464, + 0.013936039060354233, + 0.018146710470318794, + 0.021249037235975266, + 0.01787526346743107, + 0.019461002200841904, + 0.010768790729343891, + 0.025705086067318916, + 0.026049485430121422, + 0.010899076238274574, + 0.014542348682880402, + 0.010145549662411213, + 0.011727975681424141, + 0.011428700760006905, + 0.011488955467939377, + 0.011540164239704609, + 0.010870303027331829, + 0.007753622718155384, + 0.008376284502446651, + 0.01584813930094242, + 0.053013693541288376, + 0.015390501357614994, + 0.007638806942850351, + 0.012224829755723476, + 0.04186229407787323, + 0.006332420278340578, + 0.02054772898554802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.011080989614129066 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865648224949837 + ], + "opposite_logprob_tokens": [ + -3.048656463623047 + ], + "logprob_reward": -0.04865648224949837, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865648224949837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06283213198184967, + 0.19206105172634125, + 0.03204505890607834, + 0.032174013555049896, + 0.019743163138628006, + 0.024676086381077766, + 0.10569178313016891, + 0.12729568779468536, + 0.13257147371768951, + 0.27090954780578613 + ], + "credit_map": [ + 0.008312348276376724, + 0.0032532978802919388, + 0.004337385296821594, + 0.004026056732982397, + 0.004162543453276157, + 0.018606573343276978, + 0.004821505397558212, + 0.015312422066926956, + 0.021421324461698532, + 0.01164174173027277, + 0.01526364590972662, + 0.009166338481009007, + 0.022308003157377243, + 0.06046716123819351, + 0.018978793174028397, + 0.03281405195593834, + 0.002898610197007656, + 0.004499499686062336, + 0.005040021613240242, + 0.003115482861176133, + 0.00463102338835597, + 0.004559747409075499, + 0.003786645829677582, + 0.003514029085636139, + 0.005202446132898331, + 0.0033283850643783808, + 0.004834648221731186, + 0.0030874088406562805, + 0.003389600198715925, + 0.005364648066461086, + 0.0033613017294555902, + 0.003605575766414404, + 0.0012414662633091211, + 0.0015131556428968906, + 0.0029967196751385927, + 0.002328438451513648, + 0.001582754892297089, + 0.004122201353311539, + 0.0017226840136572719, + 0.004235742148011923, + 0.0019647139124572277, + 0.0020833357702940702, + 0.005050096195191145, + 0.003403993556275964, + 0.0021100742742419243, + 0.003441906301304698, + 0.0025685743894428015, + 0.004053392447531223, + 0.011658702045679092, + 0.006657938938587904, + 0.019422762095928192, + 0.00811473373323679, + 0.007921963930130005, + 0.013523471541702747, + 0.008615977130830288, + 0.02977623976767063, + 0.020531117916107178, + 0.019112857058644295, + 0.022036323323845863, + 0.010774140246212482, + 0.015560749918222427, + 0.01709437184035778, + 0.00978692714124918, + 0.01239919662475586, + 0.015708627179265022, + 0.016512220725417137, + 0.019340885803103447, + 0.016325438395142555, + 0.01897089183330536, + 0.013947643339633942, + 0.020379850640892982, + 0.011385920457541943, + 0.02552775666117668, + 0.08205365389585495, + 0.02437528409063816, + 0.012225746177136898, + 0.019337141886353493, + 0.06901750713586807, + 0.010795640759170055, + 0.02757682465016842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.04865648224949837 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279043674468994 + ], + "opposite_logprob_tokens": [ + -2.0177903175354004 + ], + "logprob_reward": -0.14279043674468994, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14279043674468994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13437621295452118, + 0.16530627012252808, + 0.07851797342300415, + 0.02758086286485195, + 0.04596438631415367, + 0.06844223290681839, + 0.09300778061151505, + 0.16103284060955048, + 0.09861872345209122, + 0.1271527111530304 + ], + "credit_map": [ + 0.01573440246284008, + 0.006415897980332375, + 0.013116818852722645, + 0.007620308548212051, + 0.009280598722398281, + 0.0360284186899662, + 0.01460559293627739, + 0.031574178487062454, + 0.017601780593395233, + 0.009886354207992554, + 0.013712355867028236, + 0.008215618319809437, + 0.020286185666918755, + 0.049759577959775925, + 0.015742605552077293, + 0.030101796612143517, + 0.0070586311630904675, + 0.008021434769034386, + 0.009680314920842648, + 0.006324453745037317, + 0.011146333999931812, + 0.017618592828512192, + 0.008853278122842312, + 0.009814929217100143, + 0.0032403019722551107, + 0.002691523637622595, + 0.005254732444882393, + 0.002540541347116232, + 0.002477680565789342, + 0.004654659423977137, + 0.0026256171986460686, + 0.0040958053432404995, + 0.0029046935960650444, + 0.0021064395550638437, + 0.01528971828520298, + 0.004261412192136049, + 0.002407014137133956, + 0.012606238946318626, + 0.0026381530333310366, + 0.0037507168017327785, + 0.006430991925299168, + 0.0064978511072695255, + 0.014791798777878284, + 0.010178966447710991, + 0.007886985316872597, + 0.008566209115087986, + 0.0074060894548892975, + 0.006683345418423414, + 0.007890991866588593, + 0.005723521579056978, + 0.015669574961066246, + 0.0085776187479496, + 0.006484957877546549, + 0.013728736899793148, + 0.01067894883453846, + 0.02425343357026577, + 0.022544175386428833, + 0.024314548820257187, + 0.02143033966422081, + 0.015543036162853241, + 0.020262153819203377, + 0.02198900654911995, + 0.015077897347509861, + 0.0198716688901186, + 0.009608225896954536, + 0.013848483562469482, + 0.013920237310230732, + 0.012448174878954887, + 0.014884484000504017, + 0.010784476064145565, + 0.015046441927552223, + 0.008078201673924923, + 0.012992018833756447, + 0.019241981208324432, + 0.018198475241661072, + 0.01520969532430172, + 0.009443388320505619, + 0.010486260056495667, + 0.014857731759548187, + 0.026723146438598633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.14279043674468994 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a snowflake with perfect hexagonal symmetry and fern-like ice crystal arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake covered in a microscopic frost texture with edges outlined by a soft, cold backlight creating an ethereal halo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the snowflake descend with a slow, graceful movement while gently rotating around its central axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with an extreme close-up and extremely shallow depth of field, then slowly tilt downward to track the snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the snowflake touching a dark, textured ground (like wool fabric or frozen soil) while remaining intact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062029238790273666 + ], + "opposite_logprob_tokens": [ + -2.8120291233062744 + ], + "logprob_reward": -0.062029238790273666, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062029238790273666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03320299834012985, + 0.24987195432186127, + 0.04313896596431732, + 0.016209224238991737, + 0.028512252494692802, + 0.023580946028232574, + 0.020456550642848015, + 0.1212463527917862, + 0.1762939989566803, + 0.2874867618083954 + ], + "credit_map": [ + 0.004315681289881468, + 0.001458593993447721, + 0.002205716446042061, + 0.0018635719316080213, + 0.0018767733126878738, + 0.010008972138166428, + 0.002498706104233861, + 0.008974982425570488, + 0.028150087222456932, + 0.015645651146769524, + 0.019776135683059692, + 0.01210743933916092, + 0.028888436034321785, + 0.07813332974910736, + 0.024586744606494904, + 0.0425841361284256, + 0.003708387492224574, + 0.004957450088113546, + 0.005676965694874525, + 0.0030515019316226244, + 0.006647010333836079, + 0.00908207893371582, + 0.004708101041615009, + 0.005307470448315144, + 0.0016218783566728234, + 0.0012968031223863363, + 0.0032945075072348118, + 0.0015674654860049486, + 0.0013597251381725073, + 0.0033010768238455057, + 0.0015443038428202271, + 0.002223462797701359, + 0.002074409509077668, + 0.001590144936926663, + 0.010965718887746334, + 0.0016741642029955983, + 0.001279125572182238, + 0.0033611583057790995, + 0.0017943049315363169, + 0.005773227196186781, + 0.0022775616962462664, + 0.0024128358345478773, + 0.004503011703491211, + 0.003000532742589712, + 0.0022620870731770992, + 0.003011158434674144, + 0.002636608202010393, + 0.0034771510399878025, + 0.0013847160153090954, + 0.0017451185267418623, + 0.0033354240003973246, + 0.003017266048118472, + 0.0015688998391851783, + 0.002205559052526951, + 0.002480478258803487, + 0.004719090182334185, + 0.013262977823615074, + 0.011636071838438511, + 0.031497176736593246, + 0.008745557628571987, + 0.014980016276240349, + 0.018862536177039146, + 0.010385992005467415, + 0.011876026168465614, + 0.021989166736602783, + 0.024532873183488846, + 0.025359153747558594, + 0.020723901689052582, + 0.026709316298365593, + 0.02300160750746727, + 0.017389917746186256, + 0.016588058322668076, + 0.028048355132341385, + 0.09259220212697983, + 0.023113500326871872, + 0.010256392881274223, + 0.02154943346977234, + 0.07851848006248474, + 0.008355721831321716, + 0.02505265735089779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.062029238790273666 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_12/index_tensor([6514, 6514])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_12/index_tensor([6514, 6514])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3a5e21a9a024eb2c130458a2ef3dbea762ceb392 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_12/index_tensor([6514, 6514])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning macro close-up, captured from a slightly overhead angle, captures a thick, white ceramic coffee cup resting on a deeply textured wooden desk. A soft, morning light slants down from a side window, casting a bright spot on the surface of the rich, almost black coffee in the cup. Initially, the surface is as calm as a mirror. Suddenly, a subtle vibration begins to transmit, and the desk begins to shake slightly but steadily. This tremor instantly shatters the tranquility, completely disrupting the smooth, mirror-like surface of the coffee. Concentric, fine ripples radiate rapidly and rhythmically from the center. The sidelight precisely captures the undulating edges of each ripple, giving them a fleeting golden highlight that contrasts sharply with the dark liquid. The cup slides ever so slightly across the tabletop, emitting a barely audible scraping sound. The entire image, in exquisite detail, conveys a sense of tension and suspense, a tranquility shattered by an external force, creating an atmosphere of intense concentration and tension.", + "seed": 27365, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_12/index_6514_seed_27365_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14741534353233873, + "vlm_logprob": -0.14741534353233873, + "vlm_yes_no_margin": 2.749999988079071, + "vlm_token_credit": -0.14741534353233873, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05626165121793747 + ], + "opposite_logprob_tokens": [ + -3.0562615394592285 + ], + "logprob_reward": -0.05626165121793747, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05626165121793747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13401152193546295, + 0.0771905779838562, + 0.08949733525514603, + 0.16373275220394135, + 0.15756723284721375, + 0.05221422761678696, + 0.08430664986371994, + 0.08474405854940414, + 0.0661846324801445, + 0.09055103361606598 + ], + "credit_map": [ + 0.014242023229598999, + 0.013530733063817024, + 0.01608959212899208, + 0.01752350479364395, + 0.01828925870358944, + 0.01634296588599682, + 0.012537993490695953, + 0.02545543946325779, + 0.00847605150192976, + 0.008832099847495556, + 0.007817025296390057, + 0.0073668924160301685, + 0.007096957881003618, + 0.00993100181221962, + 0.012875965796411037, + 0.01479458436369896, + 0.008262678049504757, + 0.010174435563385487, + 0.012246912345290184, + 0.011167541146278381, + 0.012041923590004444, + 0.008205833844840527, + 0.015064528211951256, + 0.012333485297858715, + 0.01638094335794449, + 0.0318770557641983, + 0.02494632452726364, + 0.018992597237229347, + 0.015682589262723923, + 0.01801575906574726, + 0.011805376037955284, + 0.02603210136294365, + 0.018699128180742264, + 0.021726300939917564, + 0.023698681965470314, + 0.015400196425616741, + 0.01975594088435173, + 0.011758138425648212, + 0.021514790132641792, + 0.02501405030488968, + 0.004500105045735836, + 0.008336353115737438, + 0.0055468203499913216, + 0.006311134900897741, + 0.008229435421526432, + 0.005016406998038292, + 0.004479298368096352, + 0.009794672951102257, + 0.011647460050880909, + 0.010168825276196003, + 0.01274268887937069, + 0.00998623389750719, + 0.008654589764773846, + 0.00809432938694954, + 0.011801902204751968, + 0.01121062133461237, + 0.01158665306866169, + 0.007606418337672949, + 0.012380185537040234, + 0.009822340682148933, + 0.011520336382091045, + 0.007873998954892159, + 0.012012246996164322, + 0.011941884644329548, + 0.009345518425107002, + 0.008283666335046291, + 0.009538572281599045, + 0.009554184973239899, + 0.00917865801602602, + 0.005875121336430311, + 0.006828102748841047, + 0.007580812554806471, + 0.010820961557328701, + 0.010327593423426151, + 0.012753003276884556, + 0.009733195416629314, + 0.009478054009377956, + 0.014842805452644825, + 0.010529836639761925, + 0.0120655857026577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.05626165121793747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42888861894607544 + ], + "opposite_logprob_tokens": [ + -1.0538885593414307 + ], + "logprob_reward": -0.42888861894607544, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42888861894607544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1975366324186325, + 0.15203510224819183, + 0.17943739891052246, + 0.06689467281103134, + 0.10057391971349716, + 0.11294123530387878, + 0.02452186308801174, + 0.06679688394069672, + 0.03878599405288696, + 0.06047637760639191 + ], + "credit_map": [ + 0.014530405402183533, + 0.015867240726947784, + 0.024038122966885567, + 0.01878206431865692, + 0.015701942145824432, + 0.02837860956788063, + 0.023614555597305298, + 0.056623686105012894, + 0.017546122893691063, + 0.023428067564964294, + 0.01724471151828766, + 0.012679202482104301, + 0.020645348355174065, + 0.014868819154798985, + 0.017616398632526398, + 0.028006428852677345, + 0.023454604670405388, + 0.016642972826957703, + 0.024743540212512016, + 0.010931168682873249, + 0.015333029441535473, + 0.023543404415249825, + 0.012669253163039684, + 0.05211944133043289, + 0.006689033005386591, + 0.012394782155752182, + 0.007489020004868507, + 0.012049120850861073, + 0.0062667811289429665, + 0.005444920156151056, + 0.005549918860197067, + 0.011011100374162197, + 0.015423756092786789, + 0.010926496237516403, + 0.012112545780837536, + 0.007681506220251322, + 0.009317822754383087, + 0.01079355925321579, + 0.008588720113039017, + 0.02572951465845108, + 0.006286082323640585, + 0.02310173772275448, + 0.009584508836269379, + 0.012345720082521439, + 0.02672683820128441, + 0.00706677557900548, + 0.016726985573768616, + 0.011102575808763504, + 0.0036619389429688454, + 0.002595372498035431, + 0.0037065765354782343, + 0.001743496861308813, + 0.002358129946514964, + 0.0019562721718102694, + 0.0021290727891027927, + 0.006371001712977886, + 0.007131796330213547, + 0.0066460855305194855, + 0.007609811145812273, + 0.007811744697391987, + 0.007659982889890671, + 0.008091293275356293, + 0.010027226991951466, + 0.011818946339190006, + 0.005285276565700769, + 0.004931063391268253, + 0.004684280604124069, + 0.005135662388056517, + 0.004980245139449835, + 0.0058993506245315075, + 0.0034575480967760086, + 0.0044125644490122795, + 0.006585748866200447, + 0.005077314097434282, + 0.00509498780593276, + 0.006767383310943842, + 0.007151483092457056, + 0.013715394772589207, + 0.004977129399776459, + 0.011106935329735279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.42888861894607544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024994343984872103 + ], + "opposite_logprob_tokens": [ + -6.002499580383301 + ], + "logprob_reward": -0.0024994343984872103, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024994343984872103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2720138132572174, + 0.08458464592695236, + 0.07662535458803177, + 0.1950588971376419, + 0.09374027699232101, + 0.1440325677394867, + 0.0379803329706192, + 0.037424370646476746, + 0.03386782854795456, + 0.024671902880072594 + ], + "credit_map": [ + 0.02073063515126705, + 0.03979663550853729, + 0.027072928845882416, + 0.033971160650253296, + 0.034854400902986526, + 0.0352727510035038, + 0.0260071512311697, + 0.05430816486477852, + 0.011437106877565384, + 0.014152989722788334, + 0.009527613408863544, + 0.007124474737793207, + 0.011822559870779514, + 0.008706217631697655, + 0.009184570051729679, + 0.01262911967933178, + 0.009269461035728455, + 0.00950164720416069, + 0.007978160865604877, + 0.009495275095105171, + 0.009935759007930756, + 0.007509187329560518, + 0.009968891739845276, + 0.012966972775757313, + 0.019577106460928917, + 0.04411174729466438, + 0.023448359221220016, + 0.023379921913146973, + 0.02695556916296482, + 0.01746351085603237, + 0.015718704089522362, + 0.02440398931503296, + 0.007924072444438934, + 0.013990333303809166, + 0.00964860338717699, + 0.010926659218966961, + 0.012600723654031754, + 0.009297232143580914, + 0.01233590953052044, + 0.017016740515828133, + 0.006654804572463036, + 0.02984718047082424, + 0.009335312992334366, + 0.017329860478639603, + 0.03720545396208763, + 0.011574886739253998, + 0.0202970951795578, + 0.011787974275648594, + 0.00385446660220623, + 0.004305353853851557, + 0.0052484734915196896, + 0.0052901338785886765, + 0.0037934286519885063, + 0.003742275293916464, + 0.0063674855045974255, + 0.005378717090934515, + 0.0040983338840305805, + 0.0036757478956133127, + 0.0048484718427062035, + 0.004406463820487261, + 0.004415844567120075, + 0.0031267802696675062, + 0.00720073701813817, + 0.005651992745697498, + 0.004377009347081184, + 0.004517401102930307, + 0.004539300687611103, + 0.00451618293300271, + 0.0047658514231443405, + 0.002860451815649867, + 0.003929951693862677, + 0.0043616811744868755, + 0.002678079064935446, + 0.003112252103164792, + 0.0029942800756543875, + 0.0027771838940680027, + 0.002964780665934086, + 0.003617594251409173, + 0.002592056291177869, + 0.0039356779307127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0024994343984872103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905047923326492 + ], + "opposite_logprob_tokens": [ + -2.464050531387329 + ], + "logprob_reward": -0.08905047923326492, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905047923326492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11810872703790665, + 0.17735259234905243, + 0.0762009546160698, + 0.40550699830055237, + 0.03370377793908119, + 0.10258986055850983, + 0.019045526161789894, + 0.026997007429599762, + 0.028297968208789825, + 0.012196616269648075 + ], + "credit_map": [ + 0.013735041953623295, + 0.014309221878647804, + 0.010240910574793816, + 0.015059689059853554, + 0.015338760800659657, + 0.014781029894948006, + 0.00934416800737381, + 0.025299903005361557, + 0.015100236982107162, + 0.015805896371603012, + 0.021463239565491676, + 0.020004190504550934, + 0.019487418234348297, + 0.020093340426683426, + 0.04391852766275406, + 0.02147972583770752, + 0.00770022626966238, + 0.008792456239461899, + 0.010214281268417835, + 0.011216754093766212, + 0.008622751571238041, + 0.006797888316214085, + 0.01271252054721117, + 0.010144071653485298, + 0.035402920097112656, + 0.08451986312866211, + 0.044761911034584045, + 0.053359754383563995, + 0.06423040479421616, + 0.03932633250951767, + 0.03673635050654411, + 0.04716942459344864, + 0.004742739722132683, + 0.005493843927979469, + 0.005435842089354992, + 0.0032374386209994555, + 0.0040036034770309925, + 0.002762051299214363, + 0.0034855338744819164, + 0.004542723763734102, + 0.005502793937921524, + 0.0205690898001194, + 0.0062212515622377396, + 0.011592091992497444, + 0.025297753512859344, + 0.007253653835505247, + 0.016413381323218346, + 0.009739840403199196, + 0.002667748834937811, + 0.001965529518201947, + 0.0029710468370467424, + 0.0023059870582073927, + 0.0016945492243394256, + 0.0014286702498793602, + 0.0027056641411036253, + 0.003306328784674406, + 0.001992798177525401, + 0.002720735501497984, + 0.0032784156501293182, + 0.0031040157191455364, + 0.004171041306108236, + 0.00223402283154428, + 0.0032191677019000053, + 0.006276809144765139, + 0.003822934115305543, + 0.00486537953838706, + 0.0026763214264065027, + 0.004398053977638483, + 0.0030700843781232834, + 0.003931745421141386, + 0.002388729713857174, + 0.003144718473777175, + 0.000965243496466428, + 0.0009683571988716722, + 0.0013183954870328307, + 0.0012350903125479817, + 0.0018118451116606593, + 0.0023068697191774845, + 0.0014152098447084427, + 0.002175604924559593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.08905047923326492 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037653386592865 + ], + "opposite_logprob_tokens": [ + -1.9103765487670898 + ], + "logprob_reward": -0.16037653386592865, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037653386592865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1702054738998413, + 0.11405728757381439, + 0.04766887426376343, + 0.14105646312236786, + 0.04865216836333275, + 0.3921981155872345, + 0.021443121135234833, + 0.014505697414278984, + 0.0298761036247015, + 0.020336702466011047 + ], + "credit_map": [ + 0.013525496236979961, + 0.010837347246706486, + 0.024978265166282654, + 0.01698935031890869, + 0.013568599708378315, + 0.024249520152807236, + 0.03412497416138649, + 0.03193191438913345, + 0.012690499424934387, + 0.012013896368443966, + 0.011224851943552494, + 0.010368011891841888, + 0.010718363337218761, + 0.014276851899921894, + 0.024071726948022842, + 0.01869307830929756, + 0.007638530805706978, + 0.00770342443138361, + 0.00566730834543705, + 0.004091781098395586, + 0.0044358642771840096, + 0.004521609749644995, + 0.004400602076202631, + 0.009209753014147282, + 0.015218595042824745, + 0.03243159502744675, + 0.012316725216805935, + 0.019601808860898018, + 0.02023140713572502, + 0.012311285361647606, + 0.012338913977146149, + 0.016606131568551064, + 0.005600779317319393, + 0.009753253310918808, + 0.005482617300003767, + 0.004209491424262524, + 0.006145884282886982, + 0.0042361365631222725, + 0.004103183746337891, + 0.00912082102149725, + 0.02178514562547207, + 0.08693085610866547, + 0.028201093897223473, + 0.04567024111747742, + 0.091973677277565, + 0.0326029434800148, + 0.047811444848775864, + 0.037222713232040405, + 0.0015546021750196815, + 0.002603167435154319, + 0.00289493752643466, + 0.0033804038539528847, + 0.002431403147056699, + 0.0021407026797533035, + 0.00351272476837039, + 0.0029251796659082174, + 0.0018506812630221248, + 0.0015307720750570297, + 0.0022626526188105345, + 0.0016790236113592982, + 0.0014389435527846217, + 0.0014402326196432114, + 0.0015184376388788223, + 0.0027849539183080196, + 0.002469189465045929, + 0.006684455089271069, + 0.0030044561717659235, + 0.004626929294317961, + 0.0034015511628240347, + 0.0028714295476675034, + 0.0036231596022844315, + 0.0031949321273714304, + 0.001762222032994032, + 0.002079073106870055, + 0.001979921944439411, + 0.0040278080850839615, + 0.0027425300795584917, + 0.002876375336199999, + 0.0019419968593865633, + 0.0029267759528011084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.16037653386592865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, captured from a slightly overhead angle, captures a thick, white ceramic coffee cup resting on a deeply textured wooden desk. A soft, morning light slants down from a side window, casting a bright spot on the surface of the rich, almost black coffee in the cup. Initially, the surface is as calm as a mirror. Suddenly, a subtle vibration begins to transmit, and the desk begins to shake slightly but steadily. This tremor instantly shatters the tranquility, completely disrupting the smooth, mirror-like surface of the coffee. Concentric, fine ripples radiate rapidly and rhythmically from the center. The sidelight precisely captures the undulating edges of each ripple, giving them a fleeting golden highlight that contrasts sharply with the dark liquid. The cup slides ever so slightly across the tabletop, emitting a barely audible scraping sound. The entire image, in exquisite detail, conveys a sense of tension and suspense, a tranquility shattered by an external force, creating an atmosphere of intense concentration and tension.", + "seed": 27366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_12/index_6514_seed_27366_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.19462927249260248, + "vlm_logprob": -0.19462927249260248, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.19462927249260248, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020175477489829063 + ], + "opposite_logprob_tokens": [ + -4.270175457000732 + ], + "logprob_reward": -0.020175477489829063, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020175477489829063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42298606038093567, + 0.045770347118377686, + 0.031365733593702316, + 0.019449803978204727, + 0.020656438544392586, + 0.02429237775504589, + 0.011783100664615631, + 0.17201319336891174, + 0.06558246910572052, + 0.18610049784183502 + ], + "credit_map": [ + 0.08718614280223846, + 0.03821423277258873, + 0.025880323722958565, + 0.013846473768353462, + 0.12304819375276566, + 0.05378749594092369, + 0.031744252890348434, + 0.04927893728017807, + 0.004902731627225876, + 0.004243847914040089, + 0.003293572925031185, + 0.002233474515378475, + 0.0150286965072155, + 0.0067299772053956985, + 0.0020592701621353626, + 0.007278775330632925, + 0.005108875222504139, + 0.001585867372341454, + 0.00941762886941433, + 0.0026288151275366545, + 0.002268170239403844, + 0.0042143831960856915, + 0.0023599108681082726, + 0.00378208514302969, + 0.005915097892284393, + 0.0009945629863068461, + 0.0025667143054306507, + 0.0014011553721502423, + 0.0013524624519050121, + 0.002862951485440135, + 0.001737087732180953, + 0.002619770122691989, + 0.004112160298973322, + 0.001168098533526063, + 0.005134806502610445, + 0.0012433236697688699, + 0.0031282478012144566, + 0.0014102035202085972, + 0.00120151462033391, + 0.0032580839470028877, + 0.00816055666655302, + 0.0008815667242743075, + 0.0028607917483896017, + 0.001504052197560668, + 0.0010978247737511992, + 0.006825849413871765, + 0.001059182919561863, + 0.0019025530200451612, + 0.002106743399053812, + 0.0011442886898294091, + 0.001981015084311366, + 0.001451132702641189, + 0.0011239944724366069, + 0.0011008689180016518, + 0.0011323890648782253, + 0.0017426690319553018, + 0.019987404346466064, + 0.01575455628335476, + 0.023307982832193375, + 0.01713593304157257, + 0.031165096908807755, + 0.023062458261847496, + 0.028263883665204048, + 0.013335872441530228, + 0.0065560066141188145, + 0.0013850703835487366, + 0.0037688324227929115, + 0.00265313358977437, + 0.003602590411901474, + 0.02429276891052723, + 0.0035261346492916346, + 0.019797930493950844, + 0.025752779096364975, + 0.005308371502906084, + 0.009262841194868088, + 0.006172930356115103, + 0.009060772135853767, + 0.08220411837100983, + 0.004297747276723385, + 0.044040948152542114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.020175477489829063 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578235864639282 + ], + "opposite_logprob_tokens": [ + -0.6328235864639282 + ], + "logprob_reward": -0.7578235864639282, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578235864639282, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.29462260007858276, + 0.018611859530210495, + 0.07192132622003555, + 0.09339940547943115, + 0.04499555751681328, + 0.0876595750451088, + 0.08443410694599152, + 0.0493616983294487, + 0.06942485272884369, + 0.18556898832321167 + ], + "credit_map": [ + 0.08439195901155472, + 0.01560431718826294, + 0.021559521555900574, + 0.011249668896198273, + 0.03073861636221409, + 0.02962963841855526, + 0.022492334246635437, + 0.07895652949810028, + 0.0025512389838695526, + 0.002063447842374444, + 0.0020729436073452234, + 0.0018198693869635463, + 0.0036797858774662018, + 0.0022442368790507317, + 0.0011638787109404802, + 0.003016458824276924, + 0.008353791199624538, + 0.005631870590150356, + 0.01413123682141304, + 0.0067043667659163475, + 0.010415509343147278, + 0.007594887632876635, + 0.006366641726344824, + 0.012723026797175407, + 0.027313806116580963, + 0.0044432468712329865, + 0.01434625219553709, + 0.00527274701744318, + 0.00561566324904561, + 0.012674265541136265, + 0.008907339535653591, + 0.014826090075075626, + 0.006929654162377119, + 0.002272733487188816, + 0.004528502933681011, + 0.0027629954274743795, + 0.008424251340329647, + 0.01127558946609497, + 0.002347719157114625, + 0.006454113870859146, + 0.03086368925869465, + 0.0034236141946166754, + 0.021503996104002, + 0.005014801397919655, + 0.003091734368354082, + 0.01506955549120903, + 0.0030603166669607162, + 0.005631868727505207, + 0.014361871406435966, + 0.0035282904282212257, + 0.009341846220195293, + 0.0042226905934512615, + 0.00403946777805686, + 0.017998050898313522, + 0.005631823092699051, + 0.025310073047876358, + 0.006967224180698395, + 0.0037345392629504204, + 0.007441339548677206, + 0.00449792854487896, + 0.007008435670286417, + 0.00535626383498311, + 0.009440268389880657, + 0.004915700759738684, + 0.02310597151517868, + 0.003912870306521654, + 0.006895115599036217, + 0.00506881158798933, + 0.0029199181590229273, + 0.009135222993791103, + 0.005170601885765791, + 0.013216341845691204, + 0.025595203042030334, + 0.006070743314921856, + 0.010120292194187641, + 0.006442046258598566, + 0.010429682210087776, + 0.08051096647977829, + 0.0045106676407158375, + 0.04188939929008484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.7578235864639282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003207065397873521 + ], + "opposite_logprob_tokens": [ + -5.753207206726074 + ], + "logprob_reward": -0.003207065397873521, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003207065397873521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14304368197917938, + 0.05903488025069237, + 0.06546992063522339, + 0.04250573366880417, + 0.014075382612645626, + 0.030476244166493416, + 0.20076781511306763, + 0.04561299458146095, + 0.3178453743457794, + 0.08116806298494339 + ], + "credit_map": [ + 0.05092586949467659, + 0.006813125219196081, + 0.01094142533838749, + 0.006545887794345617, + 0.008580025285482407, + 0.014230466447770596, + 0.011623253114521503, + 0.033383600413799286, + 0.007535591255873442, + 0.003934288863092661, + 0.006849209312349558, + 0.004323534667491913, + 0.005099357571452856, + 0.015346501022577286, + 0.002711306791752577, + 0.01323508471250534, + 0.011618559248745441, + 0.004531377460807562, + 0.010753351263701916, + 0.005501643288880587, + 0.007466793991625309, + 0.005108530633151531, + 0.005828627850860357, + 0.014661022461950779, + 0.003873193170875311, + 0.0019163276301696897, + 0.00471236277371645, + 0.003330898704007268, + 0.005393898114562035, + 0.01008297223597765, + 0.0026487342547625303, + 0.01054733619093895, + 0.003830415429547429, + 0.0005991280195303261, + 0.003291650675237179, + 0.0007627466693520546, + 0.001384189585223794, + 0.00267422734759748, + 0.0006176350289024413, + 0.0009153873543255031, + 0.009527334943413734, + 0.0012976296711713076, + 0.0032576164230704308, + 0.0027139768935739994, + 0.0013916968600824475, + 0.008200781419873238, + 0.0013300514547154307, + 0.002757153008133173, + 0.018447333946824074, + 0.008792062290012836, + 0.017637824639678, + 0.011365181766450405, + 0.011563608422875404, + 0.06336741894483566, + 0.014680042862892151, + 0.05491430684924126, + 0.00838471483439207, + 0.004530222620815039, + 0.008843842893838882, + 0.005316923372447491, + 0.0054159266874194145, + 0.0037351828068494797, + 0.003941579721868038, + 0.005444593261927366, + 0.046002812683582306, + 0.007586554624140263, + 0.019908659160137177, + 0.01219896599650383, + 0.01585349440574646, + 0.12375838309526443, + 0.013529115356504917, + 0.0790073424577713, + 0.014664123766124249, + 0.002334947930648923, + 0.006151281297206879, + 0.003196433885022998, + 0.004622922744601965, + 0.034741222858428955, + 0.002132062567397952, + 0.013325055129826069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.003207065397873521 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899010181427002 + ], + "opposite_logprob_tokens": [ + -2.5789899826049805 + ], + "logprob_reward": -0.07899010181427002, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899010181427002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17229892313480377, + 0.05190245434641838, + 0.030936816707253456, + 0.08031592518091202, + 0.024982847273349762, + 0.03679484501481056, + 0.06350447237491608, + 0.05824558064341545, + 0.1455586701631546, + 0.3354593813419342 + ], + "credit_map": [ + 0.051716338843107224, + 0.009611455723643303, + 0.013724349439144135, + 0.007188023999333382, + 0.01523674838244915, + 0.013723908923566341, + 0.014421935193240643, + 0.046676184982061386, + 0.0068389493972063065, + 0.004390990361571312, + 0.0064343116246163845, + 0.004214114509522915, + 0.011177926324307919, + 0.011567346751689911, + 0.0026037523057311773, + 0.004675067029893398, + 0.006169210188090801, + 0.002270276192575693, + 0.007301660254597664, + 0.0031175739131867886, + 0.0029065203852951527, + 0.002542658941820264, + 0.002355123171582818, + 0.0042737992480397224, + 0.010651377961039543, + 0.0031881737522780895, + 0.00957670621573925, + 0.004212685860693455, + 0.009802981279790401, + 0.022624937817454338, + 0.004507558420300484, + 0.015751514583826065, + 0.00602702796459198, + 0.0012029302306473255, + 0.006973897572606802, + 0.0015446458710357547, + 0.0021081138402223587, + 0.0027161978650838137, + 0.0011089323088526726, + 0.003301102900877595, + 0.012786445207893848, + 0.0013904842780902982, + 0.007673026528209448, + 0.0022477335296571255, + 0.0012701088562607765, + 0.007458377163857222, + 0.0016366641502827406, + 0.002332010306417942, + 0.009465430863201618, + 0.0023764553479850292, + 0.007667460013180971, + 0.0035168032627552748, + 0.0031353977974504232, + 0.013738121837377548, + 0.0034895259886980057, + 0.020115286111831665, + 0.007127008400857449, + 0.004205312114208937, + 0.0091185187920928, + 0.005450008437037468, + 0.006620112806558609, + 0.006050093099474907, + 0.013372709043323994, + 0.0063018216751515865, + 0.03491710498929024, + 0.004326837603002787, + 0.011363212019205093, + 0.00545017933472991, + 0.007079049479216337, + 0.053909894078969955, + 0.005316603928804398, + 0.02319580316543579, + 0.04711578041315079, + 0.009656637907028198, + 0.01654035598039627, + 0.011451000347733498, + 0.01620570197701454, + 0.1465793401002884, + 0.007833855226635933, + 0.08007675409317017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.07899010181427002 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295013129711151 + ], + "opposite_logprob_tokens": [ + -2.237950086593628 + ], + "logprob_reward": -0.11295013129711151, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295013129711151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11394643038511276, + 0.04610530659556389, + 0.11307092010974884, + 0.06541614979505539, + 0.0704522356390953, + 0.14412444829940796, + 0.16961202025413513, + 0.052300579845905304, + 0.20176218450069427, + 0.023209797218441963 + ], + "credit_map": [ + 0.0301890317350626, + 0.007276143878698349, + 0.007201740983873606, + 0.004086422733962536, + 0.018914077430963516, + 0.010935964062809944, + 0.00981894601136446, + 0.025524098426103592, + 0.005913862027227879, + 0.0028811325319111347, + 0.004943201784044504, + 0.003983320202678442, + 0.005876010283827782, + 0.011619961820542812, + 0.0018731107702478766, + 0.009014706127345562, + 0.017149532213807106, + 0.00719344662502408, + 0.019040191546082497, + 0.008683809079229832, + 0.012775765731930733, + 0.013109312392771244, + 0.00962830986827612, + 0.025490541011095047, + 0.008968413807451725, + 0.0028231495525687933, + 0.006692006718367338, + 0.003293914720416069, + 0.009333640336990356, + 0.015086663886904716, + 0.006028691306710243, + 0.013189663179218769, + 0.01785898581147194, + 0.0034539606422185898, + 0.024896958842873573, + 0.004528122488409281, + 0.006435135379433632, + 0.004700168501585722, + 0.0028647149447351694, + 0.005714184604585171, + 0.05051326006650925, + 0.005251556634902954, + 0.028991984203457832, + 0.007824174128472805, + 0.005220963153988123, + 0.03138227388262749, + 0.0049845133908092976, + 0.00995571818202734, + 0.026542432606220245, + 0.006413660477846861, + 0.013190189376473427, + 0.007988732308149338, + 0.007757863495498896, + 0.04399421066045761, + 0.010134483687579632, + 0.05359043926000595, + 0.007572751492261887, + 0.003987533506006002, + 0.009128368459641933, + 0.005180128384381533, + 0.0047401925548911095, + 0.005091712344437838, + 0.010188369080424309, + 0.006411523092538118, + 0.05133352801203728, + 0.006498861592262983, + 0.01391121931374073, + 0.007718937937170267, + 0.009954131208360195, + 0.0762762501835823, + 0.0066512348130345345, + 0.029418013989925385, + 0.004834078252315521, + 0.0013341435696929693, + 0.001957022352144122, + 0.0010651936754584312, + 0.0014457093784585595, + 0.008458641357719898, + 0.0007803001790307462, + 0.0033347078133374453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.11295013129711151 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, captured from a slightly overhead angle, captures a thick, white ceramic coffee cup resting on a deeply textured wooden desk. A soft, morning light slants down from a side window, casting a bright spot on the surface of the rich, almost black coffee in the cup. Initially, the surface is as calm as a mirror. Suddenly, a subtle vibration begins to transmit, and the desk begins to shake slightly but steadily. This tremor instantly shatters the tranquility, completely disrupting the smooth, mirror-like surface of the coffee. Concentric, fine ripples radiate rapidly and rhythmically from the center. The sidelight precisely captures the undulating edges of each ripple, giving them a fleeting golden highlight that contrasts sharply with the dark liquid. The cup slides ever so slightly across the tabletop, emitting a barely audible scraping sound. The entire image, in exquisite detail, conveys a sense of tension and suspense, a tranquility shattered by an external force, creating an atmosphere of intense concentration and tension.", + "seed": 27367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_12/index_6514_seed_27367_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.38194725196808577, + "vlm_logprob": -0.38194725196808577, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.38194725196808577, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02281453087925911 + ], + "opposite_logprob_tokens": [ + -4.147814750671387 + ], + "logprob_reward": -0.02281453087925911, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02281453087925911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18996548652648926, + 0.08192754536867142, + 0.03760972246527672, + 0.04584493115544319, + 0.06576407700777054, + 0.17099317908287048, + 0.13451099395751953, + 0.11912388354539871, + 0.10277567058801651, + 0.051484543830156326 + ], + "credit_map": [ + 0.036306705325841904, + 0.018832456320524216, + 0.013449856080114841, + 0.017426973208785057, + 0.024887703359127045, + 0.03929096460342407, + 0.018960611894726753, + 0.02081022411584854, + 0.0165749229490757, + 0.007080685812979937, + 0.0070932707749307156, + 0.009167603217065334, + 0.007074043620377779, + 0.013055544346570969, + 0.007096285466104746, + 0.01478518731892109, + 0.007016709074378014, + 0.003711218014359474, + 0.004102199804037809, + 0.004462814424186945, + 0.0036246725358068943, + 0.007018697913736105, + 0.0035418833140283823, + 0.004131529945880175, + 0.008887216448783875, + 0.0043077352456748486, + 0.005196493584662676, + 0.004946413915604353, + 0.0046300399117171764, + 0.007305419538170099, + 0.004573086276650429, + 0.005998523905873299, + 0.012411939911544323, + 0.009429153054952621, + 0.007585994433611631, + 0.008982925675809383, + 0.007770369295030832, + 0.00896614883095026, + 0.005263152997940779, + 0.005354393273591995, + 0.026203934103250504, + 0.018518861383199692, + 0.019779762253165245, + 0.013854117132723331, + 0.017851902171969414, + 0.033020682632923126, + 0.01804923824965954, + 0.02371467836201191, + 0.021257920190691948, + 0.01535891368985176, + 0.01577734388411045, + 0.012574502266943455, + 0.016068236902356148, + 0.01776060275733471, + 0.01590898633003235, + 0.019804485142230988, + 0.029014980420470238, + 0.014065351337194443, + 0.008504146710038185, + 0.01103671919554472, + 0.010469458997249603, + 0.02570338174700737, + 0.010038724169135094, + 0.010291130281984806, + 0.015268963761627674, + 0.012063248082995415, + 0.009916456416249275, + 0.009433132596313953, + 0.011689659208059311, + 0.022998737171292305, + 0.009151545353233814, + 0.012253927066922188, + 0.010876969434320927, + 0.005690320394933224, + 0.005201592110097408, + 0.005145636852830648, + 0.005002978257834911, + 0.0075841303914785385, + 0.005822398699820042, + 0.006160520017147064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.02281453087925911 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063289165496826 + ], + "opposite_logprob_tokens": [ + -0.2813289165496826 + ], + "logprob_reward": -1.4063289165496826, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063289165496826, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1885259747505188, + 0.14314721524715424, + 0.09819720685482025, + 0.07457288354635239, + 0.0971703827381134, + 0.08628397434949875, + 0.08533450216054916, + 0.08573025465011597, + 0.036339156329631805, + 0.1046985313296318 + ], + "credit_map": [ + 0.031032824888825417, + 0.030588246881961823, + 0.020674971863627434, + 0.02339128963649273, + 0.025301748886704445, + 0.02650940604507923, + 0.01396037358790636, + 0.01706709712743759, + 0.022989124059677124, + 0.016046302393078804, + 0.017015470191836357, + 0.015494029968976974, + 0.019817547872662544, + 0.021804112941026688, + 0.0175422802567482, + 0.012438330799341202, + 0.01873394101858139, + 0.011904675513505936, + 0.011848818510770798, + 0.010530027560889721, + 0.008931398391723633, + 0.01553547102957964, + 0.011063508689403534, + 0.009649354964494705, + 0.019662952050566673, + 0.006745385471731424, + 0.00741912703961134, + 0.006467249244451523, + 0.007256545592099428, + 0.012761763297021389, + 0.0076277912594377995, + 0.006632063537836075, + 0.02277318388223648, + 0.012023979797959328, + 0.0071692387573421, + 0.009722241200506687, + 0.0159755926579237, + 0.014795325696468353, + 0.006812586914747953, + 0.00789822917431593, + 0.028747333213686943, + 0.006627433001995087, + 0.004371292889118195, + 0.005075425375252962, + 0.006196594797074795, + 0.022086529061198235, + 0.006685751490294933, + 0.006493601948022842, + 0.024535350501537323, + 0.007523961365222931, + 0.008951304480433464, + 0.005093268118798733, + 0.007322114892303944, + 0.013354839757084846, + 0.005679137073457241, + 0.012874515727162361, + 0.016136957332491875, + 0.012423713691532612, + 0.0082699004560709, + 0.0085486164316535, + 0.009691654704511166, + 0.013581284321844578, + 0.008468646556138992, + 0.008609466254711151, + 0.007161651737987995, + 0.004116344265639782, + 0.003525816136971116, + 0.0026300116442143917, + 0.005421363282948732, + 0.006258212961256504, + 0.003244915744289756, + 0.0039808363653719425, + 0.01885875128209591, + 0.012147712521255016, + 0.009916219860315323, + 0.00921949278563261, + 0.012575424276292324, + 0.020327813923358917, + 0.009187199175357819, + 0.01246591005474329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -1.4063289165496826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028317375108599663 + ], + "opposite_logprob_tokens": [ + -5.877831935882568 + ], + "logprob_reward": -0.0028317375108599663, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028317375108599663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11312855780124664, + 0.11599486321210861, + 0.09810581058263779, + 0.03793712332844734, + 0.07147429883480072, + 0.10879510641098022, + 0.05929601937532425, + 0.09457208961248398, + 0.058354321867227554, + 0.24234183132648468 + ], + "credit_map": [ + 0.0152526143938303, + 0.016572916880249977, + 0.010405400767922401, + 0.012323047034442425, + 0.01663200557231903, + 0.019623300060629845, + 0.013718691654503345, + 0.008600576780736446, + 0.021968863904476166, + 0.0139484703540802, + 0.011501056142151356, + 0.01192107331007719, + 0.0185173898935318, + 0.016394145786762238, + 0.013437402434647083, + 0.008306463249027729, + 0.026452157646417618, + 0.009466509334743023, + 0.011116674169898033, + 0.007253190036863089, + 0.008813807740807533, + 0.016620440408587456, + 0.007997940294444561, + 0.010385091416537762, + 0.005962728522717953, + 0.004031843040138483, + 0.002864915644749999, + 0.004464710596948862, + 0.005585985258221626, + 0.008949578739702702, + 0.0033716277685016394, + 0.0027057339902967215, + 0.011492743156850338, + 0.008393957279622555, + 0.005529750604182482, + 0.006562451831996441, + 0.013787975534796715, + 0.011986291036009789, + 0.006829333957284689, + 0.006891792640089989, + 0.03023896925151348, + 0.010029507800936699, + 0.005773542448878288, + 0.007028013467788696, + 0.01105935126543045, + 0.023861512541770935, + 0.013401811011135578, + 0.007402407005429268, + 0.01649468205869198, + 0.004660394974052906, + 0.004423816688358784, + 0.0038551792968064547, + 0.00823759101331234, + 0.010471281595528126, + 0.00470918882638216, + 0.006443885620683432, + 0.027864161878824234, + 0.00801013596355915, + 0.005590362474322319, + 0.007582102436572313, + 0.009306414052844048, + 0.024574778974056244, + 0.00634803669527173, + 0.005296097137033939, + 0.014138641767203808, + 0.005416114814579487, + 0.00526442239060998, + 0.004161417949944735, + 0.007505676709115505, + 0.014114304445683956, + 0.0033283785451203585, + 0.00442536873742938, + 0.04613117873668671, + 0.024862270802259445, + 0.01607665792107582, + 0.01694055087864399, + 0.03018614649772644, + 0.06125005707144737, + 0.01687029004096985, + 0.030024677515029907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0028317375108599663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521442770957947 + ], + "opposite_logprob_tokens": [ + -1.5021443367004395 + ], + "logprob_reward": -0.2521442770957947, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521442770957947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12601785361766815, + 0.14844441413879395, + 0.07837534695863724, + 0.06315860152244568, + 0.10847961157560349, + 0.2570782005786896, + 0.0664014220237732, + 0.05782270431518555, + 0.05267653986811638, + 0.04154527932405472 + ], + "credit_map": [ + 0.021320287138223648, + 0.01620359905064106, + 0.01105814054608345, + 0.010580123402178288, + 0.020414205268025398, + 0.019639980047941208, + 0.013606333173811436, + 0.013195185922086239, + 0.03500901162624359, + 0.014976096339523792, + 0.012883220799267292, + 0.012728927657008171, + 0.014874400570988655, + 0.024220777675509453, + 0.01032046414911747, + 0.023431526497006416, + 0.0123446611687541, + 0.010176874697208405, + 0.00929865799844265, + 0.00841576885432005, + 0.009262099862098694, + 0.011745907366275787, + 0.009470004588365555, + 0.007661369163542986, + 0.013247963041067123, + 0.005070772022008896, + 0.0077803912572562695, + 0.006176089867949486, + 0.007477249018847942, + 0.008784912526607513, + 0.006482418160885572, + 0.008138804696500301, + 0.02150840312242508, + 0.018307775259017944, + 0.008796271868050098, + 0.00966703798621893, + 0.015802571550011635, + 0.019435113295912743, + 0.007795513141900301, + 0.007166928146034479, + 0.07951582223176956, + 0.020240548998117447, + 0.011021293699741364, + 0.014838182367384434, + 0.02201777882874012, + 0.05881958082318306, + 0.030153583735227585, + 0.02047143317759037, + 0.018695512786507607, + 0.005594306625425816, + 0.004673969931900501, + 0.004299071151763201, + 0.00769583135843277, + 0.010643143206834793, + 0.005539307836443186, + 0.00926027912646532, + 0.013003116473555565, + 0.008400410413742065, + 0.004835961852222681, + 0.005070067476481199, + 0.006148825865238905, + 0.011106535792350769, + 0.0049324906431138515, + 0.004325294401496649, + 0.010312295518815517, + 0.003964256960898638, + 0.005496259778738022, + 0.004574738442897797, + 0.004607864655554295, + 0.01293126866221428, + 0.003603401593863964, + 0.007186455186456442, + 0.008681639097630978, + 0.003174158977344632, + 0.0038695107214152813, + 0.0027179408352822065, + 0.004425249993801117, + 0.009599023498594761, + 0.0030385032296180725, + 0.006039252504706383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.2521442770957947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561679780483246 + ], + "opposite_logprob_tokens": [ + -1.6006168127059937 + ], + "logprob_reward": -0.22561679780483246, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561679780483246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1385669857263565, + 0.05606870353221893, + 0.038400694727897644, + 0.0575687400996685, + 0.26910457015037537, + 0.28810805082321167, + 0.05543409287929535, + 0.04929041117429733, + 0.02909153513610363, + 0.018366225063800812 + ], + "credit_map": [ + 0.02115851454436779, + 0.020445192232728004, + 0.01918148249387741, + 0.016378050670027733, + 0.013698833994567394, + 0.021540923044085503, + 0.010576740838587284, + 0.01558726280927658, + 0.011405176483094692, + 0.005491324234753847, + 0.006831108592450619, + 0.004025018308311701, + 0.004912159405648708, + 0.010802216827869415, + 0.0032695394475013018, + 0.009332161396741867, + 0.007583913393318653, + 0.0035643656738102436, + 0.0031917912419885397, + 0.004561927635222673, + 0.0038096397183835506, + 0.006899910047650337, + 0.0049699232913553715, + 0.0038192267529666424, + 0.010395387187600136, + 0.005914610344916582, + 0.006287810858339071, + 0.00591941736638546, + 0.0059801433235406876, + 0.0123271644115448, + 0.0048713237047195435, + 0.0058728791773319244, + 0.06940466165542603, + 0.02725374512374401, + 0.017960838973522186, + 0.023073330521583557, + 0.038791775703430176, + 0.04427340626716614, + 0.02158074639737606, + 0.026766061782836914, + 0.0915287509560585, + 0.02273155003786087, + 0.014094982296228409, + 0.01555809285491705, + 0.023409754037857056, + 0.06504173576831818, + 0.03060392290353775, + 0.02513927035033703, + 0.011734431609511375, + 0.005346958991140127, + 0.005226473789662123, + 0.0033843691926449537, + 0.005505826789885759, + 0.013583382591605186, + 0.0037398729473352432, + 0.0069127739407122135, + 0.009717867709696293, + 0.006963334046304226, + 0.004761165473610163, + 0.004502232186496258, + 0.005692014470696449, + 0.009913520887494087, + 0.0036609950475394726, + 0.004079284146428108, + 0.0041363355703651905, + 0.0033971434459090233, + 0.0026949576567858458, + 0.002501533832401037, + 0.004172065295279026, + 0.0044896663166582584, + 0.002866756869480014, + 0.004833076149225235, + 0.003173548961058259, + 0.0016227778978645802, + 0.002182255731895566, + 0.0013760821893811226, + 0.001484479638747871, + 0.00459990045055747, + 0.0017445482080802321, + 0.002182631753385067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.22561679780483246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, captured from a slightly overhead angle, captures a thick, white ceramic coffee cup resting on a deeply textured wooden desk. A soft, morning light slants down from a side window, casting a bright spot on the surface of the rich, almost black coffee in the cup. Initially, the surface is as calm as a mirror. Suddenly, a subtle vibration begins to transmit, and the desk begins to shake slightly but steadily. This tremor instantly shatters the tranquility, completely disrupting the smooth, mirror-like surface of the coffee. Concentric, fine ripples radiate rapidly and rhythmically from the center. The sidelight precisely captures the undulating edges of each ripple, giving them a fleeting golden highlight that contrasts sharply with the dark liquid. The cup slides ever so slightly across the tabletop, emitting a barely audible scraping sound. The entire image, in exquisite detail, conveys a sense of tension and suspense, a tranquility shattered by an external force, creating an atmosphere of intense concentration and tension.", + "seed": 27368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_12/index_6514_seed_27368_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07439505474176258, + "vlm_logprob": -0.07439505474176258, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.07439505474176258, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032265324145555496 + ], + "opposite_logprob_tokens": [ + -3.6572654247283936 + ], + "logprob_reward": -0.032265324145555496, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.032265324145555496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18446983397006989, + 0.0754307433962822, + 0.06118301674723625, + 0.040372997522354126, + 0.06885740160942078, + 0.15424925088882446, + 0.09028536081314087, + 0.08366775512695312, + 0.14684265851974487, + 0.09464094042778015 + ], + "credit_map": [ + 0.026952648535370827, + 0.020973334088921547, + 0.02673075534403324, + 0.021100573241710663, + 0.013514655642211437, + 0.017607063055038452, + 0.01603519730269909, + 0.041555602103471756, + 0.015446889214217663, + 0.008778966031968594, + 0.010985841043293476, + 0.007110833190381527, + 0.00572916679084301, + 0.005048317834734917, + 0.005503878928720951, + 0.016826851293444633, + 0.013307249173521996, + 0.00568185281008482, + 0.008907469920814037, + 0.007363747805356979, + 0.003337927395477891, + 0.0070924824103713036, + 0.0036151614040136337, + 0.011877124197781086, + 0.010482280515134335, + 0.004383530002087355, + 0.0064080459997057915, + 0.005774446297436953, + 0.0020046501886099577, + 0.0027072655502706766, + 0.002490560756996274, + 0.006122217979282141, + 0.0086979940533638, + 0.008175469003617764, + 0.011669673956930637, + 0.023680970072746277, + 0.002999435644596815, + 0.004714251961559057, + 0.004106704145669937, + 0.0048129018396139145, + 0.02937401831150055, + 0.015392576344311237, + 0.02838204614818096, + 0.024737536907196045, + 0.008998578414320946, + 0.013200893998146057, + 0.01415861863642931, + 0.020004985854029655, + 0.023141222074627876, + 0.006792067084461451, + 0.014623787254095078, + 0.008376316167414188, + 0.009351084940135479, + 0.010895665735006332, + 0.007292427588254213, + 0.00981278344988823, + 0.014125416055321693, + 0.009340421296656132, + 0.01091631781309843, + 0.008260814473032951, + 0.01100259181112051, + 0.012757495976984501, + 0.008891788311302662, + 0.008372907526791096, + 0.020434904843568802, + 0.014836019836366177, + 0.018444057554006577, + 0.014904994517564774, + 0.01716649904847145, + 0.02480580471456051, + 0.021695686504244804, + 0.014554697088897228, + 0.01122510340064764, + 0.014320919290184975, + 0.009853902272880077, + 0.012127147987484932, + 0.014088909141719341, + 0.014149167574942112, + 0.009972311556339264, + 0.008903480134904385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.032265324145555496 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608303189277649 + ], + "opposite_logprob_tokens": [ + -4.141083240509033 + ], + "logprob_reward": -0.01608303189277649, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608303189277649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08640814572572708, + 0.05499238520860672, + 0.024869421496987343, + 0.04712989926338196, + 0.08652002364397049, + 0.18557490408420563, + 0.2184237390756607, + 0.1293574571609497, + 0.10588879883289337, + 0.06083526834845543 + ], + "credit_map": [ + 0.026636922731995583, + 0.006627542432397604, + 0.015193103812634945, + 0.009853244759142399, + 0.004264955408871174, + 0.005662872456014156, + 0.005706709809601307, + 0.012462790124118328, + 0.010579518042504787, + 0.0043906159698963165, + 0.007483719848096371, + 0.005861785728484392, + 0.0024969361256808043, + 0.004020129330456257, + 0.004820580128580332, + 0.015339097008109093, + 0.006137455347925425, + 0.001732315169647336, + 0.0036209416575729847, + 0.002794839208945632, + 0.0009068429353646934, + 0.001581442542374134, + 0.0017896007047966123, + 0.0063059828244149685, + 0.012821287848055363, + 0.004211330320686102, + 0.008901875466108322, + 0.0052598947659134865, + 0.0016111071454361081, + 0.0030519741121679544, + 0.003220795886591077, + 0.008051632903516293, + 0.010502999648451805, + 0.010302097536623478, + 0.013541443273425102, + 0.031149929389357567, + 0.0021866993047297, + 0.0036524992901831865, + 0.004795580171048641, + 0.010388779453933239, + 0.03236300125718117, + 0.023340502753853798, + 0.036122653633356094, + 0.03715571016073227, + 0.010149158537387848, + 0.012861699797213078, + 0.0181854497641325, + 0.015396724455058575, + 0.07201313972473145, + 0.016931431367993355, + 0.04791522026062012, + 0.022486526519060135, + 0.014714080840349197, + 0.012586475349962711, + 0.014402817003428936, + 0.017374036833643913, + 0.025466911494731903, + 0.012554229237139225, + 0.028518488630652428, + 0.01510702632367611, + 0.011380302719771862, + 0.012818277813494205, + 0.012112738564610481, + 0.011399490758776665, + 0.015574276447296143, + 0.01137879490852356, + 0.01330574881285429, + 0.011337069794535637, + 0.01339966244995594, + 0.013555708341300488, + 0.017537374049425125, + 0.00980016216635704, + 0.008025983348488808, + 0.005550904665142298, + 0.007180755026638508, + 0.01023882906883955, + 0.007082859519869089, + 0.007934738881886005, + 0.008057664148509502, + 0.00676353694871068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.01608303189277649 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015212640864774585 + ], + "opposite_logprob_tokens": [ + -6.501521110534668 + ], + "logprob_reward": -0.0015212640864774585, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015212640864774585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15749205648899078, + 0.0796690285205841, + 0.05757863074541092, + 0.127027228474617, + 0.06281580030918121, + 0.11582008749246597, + 0.19403508305549622, + 0.04677562415599823, + 0.07216320931911469, + 0.08662322908639908 + ], + "credit_map": [ + 0.024431446567177773, + 0.02075190655887127, + 0.02061602473258972, + 0.01645846664905548, + 0.02006063610315323, + 0.01890348456799984, + 0.012364155612885952, + 0.02390594594180584, + 0.016681917011737823, + 0.009069751016795635, + 0.010420639999210835, + 0.00773773156106472, + 0.010280236601829529, + 0.009691355749964714, + 0.007616874296218157, + 0.008170519024133682, + 0.01194650586694479, + 0.004390327259898186, + 0.008967168629169464, + 0.006878313142806292, + 0.0032971275504678488, + 0.004386087879538536, + 0.004600616171956062, + 0.013112486340105534, + 0.022425740957260132, + 0.011372067034244537, + 0.01941688358783722, + 0.014979478903114796, + 0.0106782466173172, + 0.01315084658563137, + 0.012283924967050552, + 0.02272004634141922, + 0.007936137728393078, + 0.006875395774841309, + 0.008842338807880878, + 0.018360914662480354, + 0.0033961348235607147, + 0.006058549974113703, + 0.004751299042254686, + 0.006595028098672628, + 0.01390527281910181, + 0.010583823546767235, + 0.010141590610146523, + 0.011698570102453232, + 0.011659670621156693, + 0.024309681728482246, + 0.015725402161478996, + 0.017796074971556664, + 0.021375099197030067, + 0.018992800265550613, + 0.017514541745185852, + 0.010084191337227821, + 0.030549725517630577, + 0.05240778997540474, + 0.022116942331194878, + 0.020993994548916817, + 0.008962951600551605, + 0.004538442939519882, + 0.006193318869918585, + 0.00452924519777298, + 0.005518515128642321, + 0.00939114298671484, + 0.003800232894718647, + 0.003841772675514221, + 0.012348571792244911, + 0.007854030467569828, + 0.010617941617965698, + 0.008547852747142315, + 0.007422180846333504, + 0.008814916014671326, + 0.008866425603628159, + 0.007691289763897657, + 0.008304053917527199, + 0.011446318589150906, + 0.007651447784155607, + 0.0071182227693498135, + 0.009464148432016373, + 0.01956411451101303, + 0.007666419725865126, + 0.015408513136208057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.0015212640864774585 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001326233148575 + ], + "opposite_logprob_tokens": [ + -2.6950132846832275 + ], + "logprob_reward": -0.07001326233148575, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001326233148575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10048317164182663, + 0.03722784295678139, + 0.021122854202985764, + 0.03634230047464371, + 0.15528611838817596, + 0.11830652505159378, + 0.08022195845842361, + 0.06320422142744064, + 0.21215632557868958, + 0.17564868927001953 + ], + "credit_map": [ + 0.017384033650159836, + 0.007938067428767681, + 0.015390202403068542, + 0.011137913912534714, + 0.0066510774195194244, + 0.006806167773902416, + 0.008391116745769978, + 0.026784593239426613, + 0.010630470700562, + 0.003939244896173477, + 0.005547644570469856, + 0.0036096093244850636, + 0.0037509414833039045, + 0.002711860928684473, + 0.0038415025919675827, + 0.0031965691596269608, + 0.002756870351731777, + 0.0023173694498836994, + 0.0030205303337424994, + 0.0022654198110103607, + 0.001852926448918879, + 0.0016231224872171879, + 0.0026994848158210516, + 0.004587129689753056, + 0.007092072628438473, + 0.0043787541799247265, + 0.005532888229936361, + 0.0044700270518660545, + 0.0027053493540734053, + 0.003121498040854931, + 0.002968535292893648, + 0.006073172204196453, + 0.017447542399168015, + 0.022124988958239555, + 0.023798784241080284, + 0.05738959088921547, + 0.0050347414799034595, + 0.008246449753642082, + 0.007766869384795427, + 0.0134771429002285, + 0.02323707938194275, + 0.015007912181317806, + 0.019865388050675392, + 0.02340911142528057, + 0.006733780261129141, + 0.006898482795804739, + 0.012259053066372871, + 0.010895715095102787, + 0.013361990451812744, + 0.007944542914628983, + 0.013558498583734035, + 0.007737276144325733, + 0.010823342017829418, + 0.012130901217460632, + 0.007051445078104734, + 0.00761395925655961, + 0.014757542870938778, + 0.005749810952693224, + 0.010050210170447826, + 0.0054942769929766655, + 0.005383648909628391, + 0.007788681425154209, + 0.006759020499885082, + 0.0072210305370390415, + 0.03130931034684181, + 0.020900197327136993, + 0.03292547166347504, + 0.019155079498887062, + 0.021285487338900566, + 0.026781408116221428, + 0.0383208766579628, + 0.021478496491909027, + 0.017398836091160774, + 0.01824326626956463, + 0.013939104042947292, + 0.01720806583762169, + 0.01877695322036743, + 0.042455241084098816, + 0.016430286690592766, + 0.031196914613246918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.07001326233148575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a thick, white ceramic coffee cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cup resting on a deeply textured wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the coffee in the cup appear rich and almost black with a bright spot from sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the desk shake slightly and steadily, causing concentric ripples to radiate from the coffee surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cup slide slightly across the tabletop while the camera maintains a slightly overhead angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520923912525177 + ], + "opposite_logprob_tokens": [ + -1.5020923614501953 + ], + "logprob_reward": -0.2520923912525177, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520923912525177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1081160232424736, + 0.028222406283020973, + 0.026601463556289673, + 0.04587239399552345, + 0.15611997246742249, + 0.0842825174331665, + 0.09338950365781784, + 0.06634073704481125, + 0.050237853080034256, + 0.3408171236515045 + ], + "credit_map": [ + 0.019930485635995865, + 0.007955233566462994, + 0.016073662787675858, + 0.010102174244821072, + 0.009077651426196098, + 0.007826478220522404, + 0.009931265376508236, + 0.027219071984291077, + 0.004957543686032295, + 0.003309916937723756, + 0.003884835634380579, + 0.0034211266320198774, + 0.0036391618195921183, + 0.0024751476012170315, + 0.002840090775862336, + 0.003694582497701049, + 0.006454460322856903, + 0.0024930895306169987, + 0.003469442017376423, + 0.0032179581467062235, + 0.002567938296124339, + 0.0021896613761782646, + 0.00296239135786891, + 0.00324652181006968, + 0.007973730564117432, + 0.004798957612365484, + 0.0058034369722008705, + 0.006843900308012962, + 0.004306717310100794, + 0.004610052797943354, + 0.004267578013241291, + 0.0072680204175412655, + 0.025968031957745552, + 0.0168690774589777, + 0.025652652606368065, + 0.05505342781543732, + 0.005924123805016279, + 0.008337586186826229, + 0.00828317180275917, + 0.010031909681856632, + 0.01248210296034813, + 0.008024041540920734, + 0.015041317790746689, + 0.012502024881541729, + 0.009840455837547779, + 0.010081935673952103, + 0.007850644178688526, + 0.008460002020001411, + 0.019356876611709595, + 0.008421449922025204, + 0.01329967100173235, + 0.008540987968444824, + 0.012393007054924965, + 0.014134258963167667, + 0.007955794222652912, + 0.00928745698183775, + 0.010151647962629795, + 0.007232697680592537, + 0.006448494736105204, + 0.003409263212233782, + 0.009961656294763088, + 0.013934073969721794, + 0.006336137652397156, + 0.00886676087975502, + 0.00888088345527649, + 0.004772374872118235, + 0.007312154863029718, + 0.005698167718946934, + 0.00495025934651494, + 0.00629865936934948, + 0.0081561841070652, + 0.0041691698133945465, + 0.026870235800743103, + 0.0714697539806366, + 0.021236363798379898, + 0.028168505057692528, + 0.06370115280151367, + 0.06891485303640366, + 0.031732697039842606, + 0.028723575174808502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.2520923912525177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_13/index_tensor([3387, 3387])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_13/index_tensor([3387, 3387])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e57c5430f02b3798051c1dd9f24cdd9126b05629 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_13/index_tensor([3387, 3387])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic close-up shot from a slightly overhead angle locks the viewer's gaze firmly on a cauldron of bubbling Sichuan hot pot. The vibrant, scalding broth is coated in a layer of shimmering, alluring red oil. Countless giant bubbles surge upward from the bottom, violently bursting on the surface, sending up tiny droplets of oil and releasing wisps of white steam. Freshly sliced, tender, tender beef rolls, freshly simmered in the broth, unfurl and tumble, their edges rapidly transforming from raw red to off-white. A few emerald green leaves, slightly wilting in the heat, join golden, saturated tofu and sliced shiitake mushrooms, bobbing and colliding with each surge of the broth, as if engaging in a passionate and chaotic dance. Warm lighting from above casts bright highlights on the glossy broth, imbuing the scene with a warm, rich, and earthy atmosphere, creating a mouth-watering and deeply impactful sensory experience.", + "seed": 27190, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_13/index_3387_seed_27190_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009684605686925352, + "vlm_logprob": -0.009684605686925352, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.009684605686925352, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005128561519086361 + ], + "opposite_logprob_tokens": [ + -6.255128383636475 + ], + "logprob_reward": -0.005128561519086361, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.005128561519086361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053966224193573, + 0.03245985507965088, + 0.07288992404937744, + 0.01653679832816124, + 0.058351367712020874, + 0.06854171305894852, + 0.04443904757499695, + 0.02486235648393631, + 0.09466470032930374, + 0.5332880020141602 + ], + "credit_map": [ + 0.0049920775927603245, + 0.005806851200759411, + 0.011207642033696175, + 0.007787137757986784, + 0.006428712047636509, + 0.00471294904127717, + 0.006769157014787197, + 0.0062616909854114056, + 0.002923182211816311, + 0.004530067555606365, + 0.00358145241625607, + 0.003380248788744211, + 0.0072501529939472675, + 0.004170430824160576, + 0.0035558072850108147, + 0.0030685116071254015, + 0.007124180439859629, + 0.022545121610164642, + 0.0037493170239031315, + 0.004525961820036173, + 0.005670655984431505, + 0.002592621138319373, + 0.019767096266150475, + 0.006914969068020582, + 0.003400494810193777, + 0.0018308727303519845, + 0.0014289749087765813, + 0.0018163103377446532, + 0.0019108575070276856, + 0.0017574115190654993, + 0.002498786896467209, + 0.0018930882215499878, + 0.007636279799044132, + 0.005978915374726057, + 0.00828082486987114, + 0.005826674401760101, + 0.007737467531114817, + 0.006737714633345604, + 0.008226004429161549, + 0.0079274857416749, + 0.006460152100771666, + 0.007931464351713657, + 0.0077151767909526825, + 0.009345684200525284, + 0.008692308329045773, + 0.007176950108259916, + 0.010040193796157837, + 0.011179784312844276, + 0.0028521502390503883, + 0.0049243648536503315, + 0.0027652543503791094, + 0.004342551343142986, + 0.009549073874950409, + 0.004318867344409227, + 0.00859811156988144, + 0.0070886751636862755, + 0.002994273789227009, + 0.00407244311645627, + 0.0026134506333619356, + 0.003727985778823495, + 0.0034459049347788095, + 0.002040027640759945, + 0.0032022728119045496, + 0.0027659956831485033, + 0.003907335922122002, + 0.028585048392415047, + 0.005829957313835621, + 0.005293654277920723, + 0.008984576910734177, + 0.005013147369027138, + 0.03077082708477974, + 0.006280158180743456, + 0.019810575991868973, + 0.19983789324760437, + 0.009993428364396095, + 0.199366956949234, + 0.02435251511633396, + 0.02131914719939232, + 0.04551433026790619, + 0.0130931930616498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.005128561519086361 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010495636379346251 + ], + "opposite_logprob_tokens": [ + -6.876049518585205 + ], + "logprob_reward": -0.0010495636379346251, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010495636379346251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07225523144006729, + 0.09575839340686798, + 0.11637237668037415, + 0.04316532984375954, + 0.03745116665959358, + 0.09205308556556702, + 0.04852036386728287, + 0.01804691180586815, + 0.041322145611047745, + 0.4350549578666687 + ], + "credit_map": [ + 0.006143728271126747, + 0.012374958023428917, + 0.011275764554738998, + 0.008731778711080551, + 0.008677003905177116, + 0.006498625036329031, + 0.010173666290938854, + 0.008379722014069557, + 0.007797889411449432, + 0.009832791052758694, + 0.010061088018119335, + 0.012637102045118809, + 0.010605871677398682, + 0.012691340409219265, + 0.019631240516901016, + 0.012501075863838196, + 0.01062942948192358, + 0.052203912287950516, + 0.005548020359128714, + 0.0048613520339131355, + 0.016557028517127037, + 0.003932307008653879, + 0.013696686364710331, + 0.008943652734160423, + 0.0054860240779817104, + 0.006788693368434906, + 0.004772108513861895, + 0.005390039179474115, + 0.00459868460893631, + 0.005773492157459259, + 0.004659767262637615, + 0.005696527194231749, + 0.003916150890290737, + 0.005059496965259314, + 0.004044208209961653, + 0.004132389556616545, + 0.005408850032836199, + 0.00479850172996521, + 0.005065517965704203, + 0.005026055965572596, + 0.00889023207128048, + 0.014393409714102745, + 0.008996536955237389, + 0.011844939552247524, + 0.012095787562429905, + 0.010337116196751595, + 0.010999190621078014, + 0.014495879411697388, + 0.005922503769397736, + 0.005287115927785635, + 0.005089982412755489, + 0.00626634294167161, + 0.005247117485851049, + 0.005812143906950951, + 0.007270826492458582, + 0.007624332327395678, + 0.0023189366329461336, + 0.0025860269088298082, + 0.0020690117962658405, + 0.0017856743652373552, + 0.002369510941207409, + 0.0023474448826164007, + 0.002163964556530118, + 0.002406343352049589, + 0.002420349046587944, + 0.011987738311290741, + 0.002332853153347969, + 0.0025102272629737854, + 0.00350409303791821, + 0.002709816675633192, + 0.012862439267337322, + 0.0029946311842650175, + 0.01676330156624317, + 0.17362801730632782, + 0.008198242634534836, + 0.12410498410463333, + 0.05563173070549965, + 0.014343301765620708, + 0.031643301248550415, + 0.010742095299065113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.0010495636379346251 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343639940023422 + ], + "opposite_logprob_tokens": [ + -3.651343584060669 + ], + "logprob_reward": -0.026343639940023422, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343639940023422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1713845282793045, + 0.0708257332444191, + 0.08058803528547287, + 0.06374885141849518, + 0.04216371104121208, + 0.06357276439666748, + 0.20964469015598297, + 0.028284767642617226, + 0.01971728913486004, + 0.25006967782974243 + ], + "credit_map": [ + 0.011949730105698109, + 0.032052550464868546, + 0.018260963261127472, + 0.020671667531132698, + 0.022006476297974586, + 0.021570341661572456, + 0.023189857602119446, + 0.021682940423488617, + 0.006691050250083208, + 0.00855360459536314, + 0.0067008016631007195, + 0.007313242647796869, + 0.011577563360333443, + 0.007892235182225704, + 0.014713347889482975, + 0.007383886259049177, + 0.005710723344236612, + 0.03306368365883827, + 0.003292292356491089, + 0.00374016584828496, + 0.009071589447557926, + 0.0030223927460610867, + 0.017741816118359566, + 0.004945366643369198, + 0.008516359142959118, + 0.006556341890245676, + 0.008135216310620308, + 0.00991069432348013, + 0.006799414753913879, + 0.008001204580068588, + 0.007205781061202288, + 0.00862384308129549, + 0.008108505979180336, + 0.005138666834682226, + 0.0036198380403220654, + 0.0046623568050563335, + 0.0053284652531147, + 0.005458339117467403, + 0.005517227575182915, + 0.0043303100392222404, + 0.005706814583390951, + 0.007440953049808741, + 0.006901155225932598, + 0.009001165628433228, + 0.008736563846468925, + 0.006420580670237541, + 0.01007077656686306, + 0.009294752962887287, + 0.0120552321895957, + 0.027035826817154884, + 0.019849885255098343, + 0.037961445748806, + 0.02581934444606304, + 0.022456694394350052, + 0.03246459364891052, + 0.032001662999391556, + 0.002886841306462884, + 0.0036931289359927177, + 0.002884363289922476, + 0.0033414687495678663, + 0.0030767079442739487, + 0.0029519558884203434, + 0.00557843828573823, + 0.0038718634750694036, + 0.001438428065739572, + 0.002779722446575761, + 0.0018068805802613497, + 0.0023261827882379293, + 0.0025809011422097683, + 0.0016170345479622483, + 0.004608790390193462, + 0.002559347776696086, + 0.005731356795877218, + 0.04892265051603317, + 0.004466603975743055, + 0.1038760393857956, + 0.04640057682991028, + 0.012607935816049576, + 0.023395363241434097, + 0.0046691615134477615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.026343639940023422 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187073335051537 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.014187073335051537, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187073335051537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06732413917779922, + 0.02908923476934433, + 0.2394355982542038, + 0.01667669042944908, + 0.019397877156734467, + 0.04371084272861481, + 0.06129375100135803, + 0.06696779280900955, + 0.01974141038954258, + 0.4363626539707184 + ], + "credit_map": [ + 0.005264165811240673, + 0.008947731927037239, + 0.007262823637574911, + 0.009707294404506683, + 0.00857077818363905, + 0.008120297454297543, + 0.010871820151805878, + 0.008579223416745663, + 0.0036467660684138536, + 0.0034546691458672285, + 0.003536500269547105, + 0.003424988593906164, + 0.003198038786649704, + 0.0035523767583072186, + 0.004912371747195721, + 0.0033635233994573355, + 0.017973419278860092, + 0.10045469552278519, + 0.008695350028574467, + 0.008350610733032227, + 0.026506129652261734, + 0.008737640455365181, + 0.057306572794914246, + 0.011411180719733238, + 0.0021396460942924023, + 0.002339625032618642, + 0.0017719358438625932, + 0.001600417192094028, + 0.002241771202534437, + 0.0021725825499743223, + 0.0018214599695056677, + 0.00258925324305892, + 0.0025713001377880573, + 0.0025984703097492456, + 0.001895772758871317, + 0.002764008240774274, + 0.0027526451740413904, + 0.002229155506938696, + 0.0023735910654067993, + 0.002212933963164687, + 0.004237317945808172, + 0.00484270416200161, + 0.005110390949994326, + 0.0067948102951049805, + 0.005115448031574488, + 0.004428957588970661, + 0.007047908380627632, + 0.006133305840194225, + 0.006311577279120684, + 0.007483107037842274, + 0.008372996002435684, + 0.009406262077391148, + 0.006939750164747238, + 0.005109356716275215, + 0.009632212109863758, + 0.008038493804633617, + 0.006795886438339949, + 0.00864832941442728, + 0.0080797728151083, + 0.0110193295404315, + 0.007293642032891512, + 0.006074911914765835, + 0.009371957741677761, + 0.009683958254754543, + 0.001627801451832056, + 0.0032381475903093815, + 0.002144506433978677, + 0.002017080085352063, + 0.0021771742030978203, + 0.0017609021160751581, + 0.004148934967815876, + 0.0026268644724041224, + 0.013345913030207157, + 0.13497251272201538, + 0.007863371632993221, + 0.17687688767910004, + 0.03560705855488777, + 0.019419211894273758, + 0.03886915743350983, + 0.009408572688698769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.014187073335051537 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017141900025308132 + ], + "opposite_logprob_tokens": [ + -6.37671422958374 + ], + "logprob_reward": -0.0017141900025308132, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017141900025308132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16216522455215454, + 0.05978039652109146, + 0.11146068572998047, + 0.026654250919818878, + 0.0684950202703476, + 0.07820963859558105, + 0.14011438190937042, + 0.023190250620245934, + 0.09223879873752594, + 0.23769134283065796 + ], + "credit_map": [ + 0.013640092685818672, + 0.016926320269703865, + 0.018514711409807205, + 0.022867029532790184, + 0.023449229076504707, + 0.019106434658169746, + 0.026271861046552658, + 0.021389545872807503, + 0.00517456978559494, + 0.0073396554216742516, + 0.005927381105720997, + 0.006699395831674337, + 0.00871707871556282, + 0.007165461778640747, + 0.011997952125966549, + 0.006758901756256819, + 0.00872553139925003, + 0.044827900826931, + 0.005483200773596764, + 0.006018099375069141, + 0.014622381888329983, + 0.00468183308839798, + 0.014171512797474861, + 0.01293022558093071, + 0.003786202287301421, + 0.003015626687556505, + 0.0027903462760150433, + 0.0037916696164757013, + 0.004117629490792751, + 0.003006202867254615, + 0.003438860410824418, + 0.0027077121194452047, + 0.010425519198179245, + 0.007702583447098732, + 0.006371036171913147, + 0.007306528277695179, + 0.00944380834698677, + 0.009268020279705524, + 0.009859043173491955, + 0.008118482306599617, + 0.006629898678511381, + 0.009829303249716759, + 0.006791458930820227, + 0.010665226727724075, + 0.010527023114264011, + 0.009864592924714088, + 0.01285658497363329, + 0.011045551858842373, + 0.008880972862243652, + 0.01652403734624386, + 0.015545238740742207, + 0.025294741615653038, + 0.017271609976887703, + 0.01479717530310154, + 0.020503416657447815, + 0.021297186613082886, + 0.00268167513422668, + 0.0031184155959635973, + 0.0031937286257743835, + 0.002908139256760478, + 0.0029260290320962667, + 0.002118104137480259, + 0.0031760900747030973, + 0.003068069228902459, + 0.005697021726518869, + 0.0220374446362257, + 0.005014119204133749, + 0.006132643204182386, + 0.009575972333550453, + 0.007587148807942867, + 0.030074195936322212, + 0.006120248697698116, + 0.0077141886577010155, + 0.06279446184635162, + 0.0050080278888344765, + 0.0944056287407875, + 0.02824520319700241, + 0.012263020500540733, + 0.02112395130097866, + 0.006136864423751831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0017141900025308132 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic close-up shot from a slightly overhead angle locks the viewer's gaze firmly on a cauldron of bubbling Sichuan hot pot. The vibrant, scalding broth is coated in a layer of shimmering, alluring red oil. Countless giant bubbles surge upward from the bottom, violently bursting on the surface, sending up tiny droplets of oil and releasing wisps of white steam. Freshly sliced, tender, tender beef rolls, freshly simmered in the broth, unfurl and tumble, their edges rapidly transforming from raw red to off-white. A few emerald green leaves, slightly wilting in the heat, join golden, saturated tofu and sliced shiitake mushrooms, bobbing and colliding with each surge of the broth, as if engaging in a passionate and chaotic dance. Warm lighting from above casts bright highlights on the glossy broth, imbuing the scene with a warm, rich, and earthy atmosphere, creating a mouth-watering and deeply impactful sensory experience.", + "seed": 27191, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_13/index_3387_seed_27191_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.009784638625569641, + "vlm_logprob": -0.009784638625569641, + "vlm_yes_no_margin": 5.425, + "vlm_token_credit": -0.009784638625569641, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00844467245042324 + ], + "opposite_logprob_tokens": [ + -5.758444786071777 + ], + "logprob_reward": -0.00844467245042324, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00844467245042324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18389633297920227, + 0.17140468955039978, + 0.016980858519673347, + 0.13810649514198303, + 0.05210971459746361, + 0.05464242771267891, + 0.05003627389669418, + 0.08422242850065231, + 0.21679091453552246, + 0.03180992230772972 + ], + "credit_map": [ + 0.0504542700946331, + 0.07095044106245041, + 0.009078112430870533, + 0.010498050600290298, + 0.01535853836685419, + 0.00805660616606474, + 0.011005314998328686, + 0.008494979701936245, + 0.033547479659318924, + 0.016905590891838074, + 0.010842153802514076, + 0.07272568345069885, + 0.008813773281872272, + 0.007627156563103199, + 0.012592574581503868, + 0.00835028663277626, + 0.0019684021826833487, + 0.0020989454351365566, + 0.0024183718487620354, + 0.0023264510091394186, + 0.002109008841216564, + 0.002478945069015026, + 0.0018731738673523068, + 0.0017075592186301947, + 0.04432642087340355, + 0.05695783346891403, + 0.007217298727482557, + 0.005567205138504505, + 0.007157319225370884, + 0.0037770343478769064, + 0.009644979611039162, + 0.003458374412730336, + 0.00860431045293808, + 0.006157021038234234, + 0.006084705702960491, + 0.007368186488747597, + 0.006738798692822456, + 0.005634389817714691, + 0.005987076088786125, + 0.005535222589969635, + 0.007176612503826618, + 0.008721528574824333, + 0.006174847017973661, + 0.004742263816297054, + 0.009529579430818558, + 0.008932319469749928, + 0.005849350709468126, + 0.003515924559906125, + 0.00576427299529314, + 0.005730501376092434, + 0.0062830811366438866, + 0.008243522606790066, + 0.0047799525782465935, + 0.0052542779594659805, + 0.00816874299198389, + 0.005811917595565319, + 0.00763850100338459, + 0.007435190957039595, + 0.011801570653915405, + 0.023251086473464966, + 0.009074467234313488, + 0.012924677692353725, + 0.007931704632937908, + 0.004165229387581348, + 0.022922201082110405, + 0.02407982386648655, + 0.025855083018541336, + 0.02196398563683033, + 0.033998627215623856, + 0.033958204090595245, + 0.029798364266753197, + 0.024214603006839752, + 0.0027940801810473204, + 0.003627707716077566, + 0.0033924237359315157, + 0.004848199430853128, + 0.005595339462161064, + 0.003933750092983246, + 0.005149906501173973, + 0.0024685144890099764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.00844467245042324 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017214493127539754 + ], + "opposite_logprob_tokens": [ + -6.376721382141113 + ], + "logprob_reward": -0.0017214493127539754, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017214493127539754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14157238602638245, + 0.14842480421066284, + 0.109184131026268, + 0.11992018669843674, + 0.032326892018318176, + 0.10894202440977097, + 0.03876007720828056, + 0.10458473116159439, + 0.153937429189682, + 0.04234732687473297 + ], + "credit_map": [ + 0.029937075451016426, + 0.03172924742102623, + 0.0151260020211339, + 0.010769599117338657, + 0.01614910364151001, + 0.013313357718288898, + 0.012824763543903828, + 0.011723232455551624, + 0.01806226372718811, + 0.02125876396894455, + 0.016846725717186928, + 0.016347091645002365, + 0.018728014081716537, + 0.0292904544621706, + 0.01389133557677269, + 0.014000161550939083, + 0.020608969032764435, + 0.04074752330780029, + 0.00638143066316843, + 0.005952685140073299, + 0.008180045522749424, + 0.009479858912527561, + 0.011682958342134953, + 0.006150663830339909, + 0.030976081266999245, + 0.040277935564517975, + 0.01098256092518568, + 0.009338774718344212, + 0.006118733901530504, + 0.004946095868945122, + 0.01080198772251606, + 0.006478012539446354, + 0.004983237013220787, + 0.004875125829130411, + 0.0062115807086229324, + 0.004378798883408308, + 0.003751804120838642, + 0.002944943029433489, + 0.002933244686573744, + 0.00224815565161407, + 0.013473762199282646, + 0.012985078617930412, + 0.014006891287863255, + 0.014695593155920506, + 0.01641465723514557, + 0.00867503322660923, + 0.02015683427453041, + 0.008534169755876064, + 0.005369542632251978, + 0.005811997689306736, + 0.006018171552568674, + 0.005642472300678492, + 0.004713025409728289, + 0.0032136866357177496, + 0.00531000504270196, + 0.0026811761781573296, + 0.009695661254227161, + 0.008929678238928318, + 0.016848647966980934, + 0.011322829872369766, + 0.01032954826951027, + 0.03483901917934418, + 0.007739597465842962, + 0.004879754967987537, + 0.013565645553171635, + 0.01570659503340721, + 0.020569488406181335, + 0.02312694489955902, + 0.024618439376354218, + 0.022753460332751274, + 0.019906915724277496, + 0.013689950108528137, + 0.0037203256506472826, + 0.005055398680269718, + 0.004372148774564266, + 0.007771920878440142, + 0.007053422275930643, + 0.00448048347607255, + 0.0056564295664429665, + 0.004237196873873472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0017214493127539754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797164723277092 + ], + "opposite_logprob_tokens": [ + -3.529797077178955 + ], + "logprob_reward": -0.029797164723277092, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797164723277092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3626200258731842, + 0.13689970970153809, + 0.10057373344898224, + 0.09094429016113281, + 0.0279233455657959, + 0.03601682186126709, + 0.029179269447922707, + 0.14054425060749054, + 0.03767595812678337, + 0.03762257844209671 + ], + "credit_map": [ + 0.0932912677526474, + 0.14921042323112488, + 0.02089095301926136, + 0.023732082918286324, + 0.028012214228510857, + 0.01253443956375122, + 0.02295907959342003, + 0.01198962889611721, + 0.01913846656680107, + 0.016321413218975067, + 0.013942801393568516, + 0.03734562546014786, + 0.011750899255275726, + 0.018538612872362137, + 0.01011096965521574, + 0.009750920347869396, + 0.016880445182323456, + 0.03678816184401512, + 0.005381226073950529, + 0.006330575793981552, + 0.0072314986027777195, + 0.011283954605460167, + 0.009961789473891258, + 0.006716090254485607, + 0.027511032298207283, + 0.03158112242817879, + 0.005186356604099274, + 0.0051778811030089855, + 0.00473782466724515, + 0.005976533517241478, + 0.006845504976809025, + 0.003928042948246002, + 0.004751946311444044, + 0.003063883865252137, + 0.0038492493331432343, + 0.003379413392394781, + 0.003031197702512145, + 0.0027253602165728807, + 0.003249537432566285, + 0.0038727596402168274, + 0.0055351280607283115, + 0.0053931414149701595, + 0.0038934065960347652, + 0.0037889908999204636, + 0.005451465956866741, + 0.003969667479395866, + 0.005253470968455076, + 0.0027315570041537285, + 0.0031131040304899216, + 0.003551941132172942, + 0.004096485674381256, + 0.003817151300609112, + 0.003108621807768941, + 0.0032390968408435583, + 0.005461478140205145, + 0.0027913928497582674, + 0.010796592570841312, + 0.011514359153807163, + 0.03481094166636467, + 0.011230558156967163, + 0.008006400428712368, + 0.05072180554270744, + 0.007843478582799435, + 0.005620123818516731, + 0.00278319139033556, + 0.005216511432081461, + 0.004715265706181526, + 0.005308747757226229, + 0.004831117112189531, + 0.004936382174491882, + 0.0044643268920481205, + 0.005420418456196785, + 0.0032589503098279238, + 0.005379792768508196, + 0.006438176613301039, + 0.0071204593405127525, + 0.004142274614423513, + 0.0038521368987858295, + 0.0038850533310323954, + 0.003545737825334072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.029797164723277092 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623034063726664 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.007623034063726664, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623034063726664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35083189606666565, + 0.2766403555870056, + 0.11186547577381134, + 0.0644213929772377, + 0.01891382411122322, + 0.02822727896273136, + 0.035230860114097595, + 0.04671263322234154, + 0.048263296484947205, + 0.018892992287874222 + ], + "credit_map": [ + 0.08817239850759506, + 0.14501436054706573, + 0.0217147059738636, + 0.021630536764860153, + 0.026434078812599182, + 0.012533416971564293, + 0.023083679378032684, + 0.012248723767697811, + 0.04994424059987068, + 0.030007006600499153, + 0.020352642983198166, + 0.10123702883720398, + 0.018329281359910965, + 0.0216459222137928, + 0.020041733980178833, + 0.01508247572928667, + 0.014972042292356491, + 0.029900124296545982, + 0.007509870920330286, + 0.007943741045892239, + 0.010449333116412163, + 0.02246091514825821, + 0.010986988432705402, + 0.007642464246600866, + 0.01841895282268524, + 0.025997593998908997, + 0.0047067939303815365, + 0.0025401217862963676, + 0.003809361020103097, + 0.00240866607055068, + 0.004825669806450605, + 0.0017142337746918201, + 0.0029124810826033354, + 0.0022147216368466616, + 0.00358545477502048, + 0.0028097748290747404, + 0.0020478295627981424, + 0.0018204208463430405, + 0.001921078423038125, + 0.0016020634211599827, + 0.004159779753535986, + 0.003508087247610092, + 0.004565793555229902, + 0.003114894963800907, + 0.004400039091706276, + 0.003720000386238098, + 0.003246913431212306, + 0.0015117722796276212, + 0.004455259535461664, + 0.005116453394293785, + 0.004847018513828516, + 0.0038425747770816088, + 0.004062028601765633, + 0.005122274626046419, + 0.004843974485993385, + 0.002941277576610446, + 0.0040079341270029545, + 0.004119346383959055, + 0.00594250438734889, + 0.00663069449365139, + 0.004648264031857252, + 0.012413511984050274, + 0.006134681403636932, + 0.00281569454818964, + 0.005198284052312374, + 0.006101373117417097, + 0.005263017490506172, + 0.0054661282338202, + 0.006334589794278145, + 0.007173762656748295, + 0.0065634348429739475, + 0.006162705831229687, + 0.0013885862426832318, + 0.002551429672166705, + 0.0020143918227404356, + 0.0036780755035579205, + 0.002164792502298951, + 0.0024868613108992577, + 0.002283029956743121, + 0.0023258260916918516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.007623034063726664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013368725776672363 + ], + "opposite_logprob_tokens": [ + -6.626337051391602 + ], + "logprob_reward": -0.0013368725776672363, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013368725776672363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31112056970596313, + 0.10106541216373444, + 0.08431630581617355, + 0.1303626298904419, + 0.0432717464864254, + 0.09362714737653732, + 0.04713290557265282, + 0.07002300024032593, + 0.08930095285177231, + 0.02977932058274746 + ], + "credit_map": [ + 0.0697735995054245, + 0.1285083144903183, + 0.018502969294786453, + 0.020646918565034866, + 0.025691745802760124, + 0.016567852348089218, + 0.02047794684767723, + 0.010951247066259384, + 0.016687646508216858, + 0.012083605863153934, + 0.010311201214790344, + 0.031167399138212204, + 0.007884464226663113, + 0.008467065170407295, + 0.007483608089387417, + 0.006980434060096741, + 0.013321694917976856, + 0.028676331043243408, + 0.005290328990668058, + 0.005048294551670551, + 0.006575825624167919, + 0.011359280906617641, + 0.008303460665047169, + 0.00574109610170126, + 0.04687327891588211, + 0.040054064244031906, + 0.013429009355604649, + 0.007717716041952372, + 0.005781684070825577, + 0.003462668275460601, + 0.009159970097243786, + 0.003884253092110157, + 0.007664795499294996, + 0.005177309270948172, + 0.006253756117075682, + 0.004393167328089476, + 0.0057410080917179585, + 0.00456339493393898, + 0.0053782169707119465, + 0.004100097808986902, + 0.02044074796140194, + 0.01071229949593544, + 0.015901505947113037, + 0.012154586613178253, + 0.011069915257394314, + 0.009372206404805183, + 0.008494982495903969, + 0.00548090785741806, + 0.006939319893717766, + 0.010118133388459682, + 0.005192816257476807, + 0.006929927971214056, + 0.00389517261646688, + 0.003930338658392429, + 0.006227262318134308, + 0.0038999412208795547, + 0.005077444948256016, + 0.006775071378797293, + 0.016519758850336075, + 0.012252844870090485, + 0.008649781346321106, + 0.013762295246124268, + 0.00415852852165699, + 0.002827283227816224, + 0.010059935972094536, + 0.012031766586005688, + 0.00980267021805048, + 0.008256476372480392, + 0.012729763053357601, + 0.013280230574309826, + 0.012565328739583492, + 0.010574787855148315, + 0.0031148605048656464, + 0.003865924896672368, + 0.003966528922319412, + 0.003729187184944749, + 0.004410646855831146, + 0.003149739233776927, + 0.004487667232751846, + 0.0030547671485692263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0013368725776672363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic close-up shot from a slightly overhead angle locks the viewer's gaze firmly on a cauldron of bubbling Sichuan hot pot. The vibrant, scalding broth is coated in a layer of shimmering, alluring red oil. Countless giant bubbles surge upward from the bottom, violently bursting on the surface, sending up tiny droplets of oil and releasing wisps of white steam. Freshly sliced, tender, tender beef rolls, freshly simmered in the broth, unfurl and tumble, their edges rapidly transforming from raw red to off-white. A few emerald green leaves, slightly wilting in the heat, join golden, saturated tofu and sliced shiitake mushrooms, bobbing and colliding with each surge of the broth, as if engaging in a passionate and chaotic dance. Warm lighting from above casts bright highlights on the glossy broth, imbuing the scene with a warm, rich, and earthy atmosphere, creating a mouth-watering and deeply impactful sensory experience.", + "seed": 27192, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_13/index_3387_seed_27192_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011094692209735513, + "vlm_logprob": -0.011094692209735513, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.011094692209735513, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00914066843688488 + ], + "opposite_logprob_tokens": [ + -5.759140491485596 + ], + "logprob_reward": -0.00914066843688488, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00914066843688488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22145597636699677, + 0.08179304003715515, + 0.06078440696001053, + 0.04859650507569313, + 0.1015249565243721, + 0.09884730726480484, + 0.0786718800663948, + 0.08738373965024948, + 0.11746852099895477, + 0.10347370058298111 + ], + "credit_map": [ + 0.020889226347208023, + 0.03644261136651039, + 0.03746756911277771, + 0.03370590880513191, + 0.03310706093907356, + 0.019621120765805244, + 0.020162416622042656, + 0.020060036331415176, + 0.0076856957748532295, + 0.01049215067178011, + 0.008636029437184334, + 0.01081011537462473, + 0.010002454742789268, + 0.006235061679035425, + 0.017145220190286636, + 0.010786302387714386, + 0.0063394783064723015, + 0.011263287626206875, + 0.0055619277991354465, + 0.006122439634054899, + 0.007798234932124615, + 0.008237517438828945, + 0.0077992333099246025, + 0.007662281859666109, + 0.006189664825797081, + 0.007872559130191803, + 0.004711670335382223, + 0.008230693638324738, + 0.00519905798137188, + 0.0053712064400315285, + 0.00495566613972187, + 0.0060659837909042835, + 0.011470017954707146, + 0.010476956143975258, + 0.00800132192671299, + 0.017447885125875473, + 0.011049313470721245, + 0.013085360638797283, + 0.016469813883304596, + 0.013524271547794342, + 0.007267283741384745, + 0.007736342027783394, + 0.015505900606513023, + 0.019656166434288025, + 0.008861823938786983, + 0.007337384857237339, + 0.02487275004386902, + 0.00760964211076498, + 0.007608714513480663, + 0.006390352733433247, + 0.01312602311372757, + 0.012636187486350536, + 0.008683986030519009, + 0.00930300448089838, + 0.011444320902228355, + 0.009479287080466747, + 0.00865933671593666, + 0.007612257264554501, + 0.013677914626896381, + 0.013786587864160538, + 0.012258549220860004, + 0.009074285626411438, + 0.011500595137476921, + 0.010814210399985313, + 0.018088432028889656, + 0.026406466960906982, + 0.022300396114587784, + 0.011861084960401058, + 0.01177878025919199, + 0.007704759482294321, + 0.00926669966429472, + 0.010061899200081825, + 0.014471837319433689, + 0.012671117670834064, + 0.014630393125116825, + 0.01576290838420391, + 0.012361297383904457, + 0.012442545965313911, + 0.011778973042964935, + 0.009354627691209316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.00914066843688488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015238827327266335 + ], + "opposite_logprob_tokens": [ + -6.501523971557617 + ], + "logprob_reward": -0.0015238827327266335, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015238827327266335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2170248180627823, + 0.10173413157463074, + 0.1270858347415924, + 0.07791174203157425, + 0.14196781814098358, + 0.04614148661494255, + 0.07404612749814987, + 0.10550853610038757, + 0.034483928233385086, + 0.07409564405679703 + ], + "credit_map": [ + 0.019390704110264778, + 0.03194786608219147, + 0.0353630967438221, + 0.03313545510172844, + 0.03516203537583351, + 0.02225540764629841, + 0.020229069516062737, + 0.019541170448064804, + 0.010176768526434898, + 0.013454057276248932, + 0.01431083120405674, + 0.017470654100179672, + 0.011405611410737038, + 0.008693581447005272, + 0.01460737083107233, + 0.011615253053605556, + 0.014068025164306164, + 0.015742840245366096, + 0.015171711333096027, + 0.013514237478375435, + 0.012991457246243954, + 0.011898231692612171, + 0.024354426190257072, + 0.01934490166604519, + 0.011938771232962608, + 0.007951688021421432, + 0.0069483136758208275, + 0.013825299218297005, + 0.006162924692034721, + 0.007748156785964966, + 0.013662545941770077, + 0.009674043394625187, + 0.01331139262765646, + 0.013917680829763412, + 0.009449356235563755, + 0.02098858170211315, + 0.010423889383673668, + 0.01856779120862484, + 0.03683505207300186, + 0.018474062904715538, + 0.0035139904357492924, + 0.003589269705116749, + 0.00572029035538435, + 0.008224858902394772, + 0.007542434614151716, + 0.003404740709811449, + 0.008123486302793026, + 0.00602241326123476, + 0.006521331612020731, + 0.006025041453540325, + 0.011597929522395134, + 0.011956633068621159, + 0.008021622896194458, + 0.009222034364938736, + 0.010860173963010311, + 0.009841358289122581, + 0.009088222868740559, + 0.00860279705375433, + 0.014116249047219753, + 0.014774063602089882, + 0.015615363605320454, + 0.013252373784780502, + 0.013622929342091084, + 0.016436530277132988, + 0.004225562326610088, + 0.005374941974878311, + 0.004933814983814955, + 0.006498660426586866, + 0.003395744366571307, + 0.0029476142954081297, + 0.003483915701508522, + 0.003623671131208539, + 0.010617698542773724, + 0.009202536195516586, + 0.0108030391857028, + 0.010273310355842113, + 0.007344032172113657, + 0.00883438903838396, + 0.010429433546960354, + 0.006591201759874821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0015238827327266335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695101737976074 + ], + "opposite_logprob_tokens": [ + -3.4086952209472656 + ], + "logprob_reward": -0.033695101737976074, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695101737976074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1375502496957779, + 0.0885043814778328, + 0.08411726355552673, + 0.1896866112947464, + 0.10457441210746765, + 0.06959463655948639, + 0.08461420983076096, + 0.057388853281736374, + 0.08562487363815308, + 0.09834454208612442 + ], + "credit_map": [ + 0.015967678278684616, + 0.026288708671927452, + 0.02027207426726818, + 0.015058365650475025, + 0.02079002745449543, + 0.012686401605606079, + 0.014647235162556171, + 0.011839760467410088, + 0.009876021184027195, + 0.012238280847668648, + 0.01396531704813242, + 0.01824376732110977, + 0.009584217332303524, + 0.006387929432094097, + 0.011007593013346195, + 0.00720125762745738, + 0.0077055334113538265, + 0.01771966926753521, + 0.007258506491780281, + 0.011662728153169155, + 0.010765578597784042, + 0.008585802279412746, + 0.011464258655905724, + 0.008955184370279312, + 0.018981115892529488, + 0.026277286931872368, + 0.01631585694849491, + 0.03299733251333237, + 0.01623360440135002, + 0.02676202729344368, + 0.025942834094166756, + 0.02617654949426651, + 0.010762250050902367, + 0.010074230842292309, + 0.006887235678732395, + 0.014847294427454472, + 0.008149581961333752, + 0.01679597608745098, + 0.020952429622411728, + 0.01610541343688965, + 0.004158188123255968, + 0.004311579279601574, + 0.010572107508778572, + 0.01627243682742119, + 0.008861720561981201, + 0.004253849387168884, + 0.0152584882453084, + 0.005906261503696442, + 0.007058521267026663, + 0.004051767289638519, + 0.01066541112959385, + 0.011156122200191021, + 0.010233836248517036, + 0.01456383801996708, + 0.012162754312157631, + 0.01472195889800787, + 0.0052353558130562305, + 0.005588555708527565, + 0.007594465743750334, + 0.007508950307965279, + 0.009197043254971504, + 0.0062703536823391914, + 0.00729683693498373, + 0.008697292767465115, + 0.009285833686590195, + 0.022839467972517014, + 0.021068183705210686, + 0.0077878497540950775, + 0.00827800389379263, + 0.004490199498832226, + 0.006262744311243296, + 0.005612599663436413, + 0.013110209256410599, + 0.010861752554774284, + 0.01582927256822586, + 0.0201089009642601, + 0.009757583029568195, + 0.010237474925816059, + 0.01184107270091772, + 0.006598278414458036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.033695101737976074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775268845260143 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775268845260143, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775268845260143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1802992820739746, + 0.13695302605628967, + 0.03814217075705528, + 0.05225029215216637, + 0.03246120736002922, + 0.08132102340459824, + 0.10495758801698685, + 0.024890080094337463, + 0.3106822073459625, + 0.03804317116737366 + ], + "credit_map": [ + 0.023732610046863556, + 0.03766334429383278, + 0.02416435070335865, + 0.018170857802033424, + 0.020964793860912323, + 0.018295032903552055, + 0.015630468726158142, + 0.02167780138552189, + 0.009365922771394253, + 0.01877651736140251, + 0.03243641182780266, + 0.01608242467045784, + 0.011784559115767479, + 0.006476155016571283, + 0.03155152499675751, + 0.010479501448571682, + 0.0028410833328962326, + 0.0073034754022955894, + 0.0039281887002289295, + 0.006108095869421959, + 0.004299567546695471, + 0.004681253340095282, + 0.005167676135897636, + 0.0038128243759274483, + 0.006325597409158945, + 0.007655190769582987, + 0.004551219288259745, + 0.008799242787063122, + 0.005453002639114857, + 0.004938527476042509, + 0.008581431582570076, + 0.005946073215454817, + 0.003355219727382064, + 0.003253296483308077, + 0.002242592629045248, + 0.004673728719353676, + 0.003088337602093816, + 0.003617675742134452, + 0.00777507433667779, + 0.004455283284187317, + 0.005195674020797014, + 0.0050363438203930855, + 0.010445917025208473, + 0.01775335893034935, + 0.008054313249886036, + 0.005727817770093679, + 0.02144644781947136, + 0.007661141455173492, + 0.009274215437471867, + 0.007953672669827938, + 0.01642899587750435, + 0.01536776777356863, + 0.012931964360177517, + 0.013919301331043243, + 0.015236491337418556, + 0.013845163397490978, + 0.0027486553881317377, + 0.002688149455934763, + 0.002769924933090806, + 0.005409170873463154, + 0.0030085730832070112, + 0.001990803051739931, + 0.003924149088561535, + 0.002350651891902089, + 0.028565824031829834, + 0.10979456454515457, + 0.08553238213062286, + 0.02109525538980961, + 0.02998758852481842, + 0.010082920081913471, + 0.014736606739461422, + 0.010887065902352333, + 0.0052149188704788685, + 0.0035341777838766575, + 0.006661469116806984, + 0.006640190724283457, + 0.004383151885122061, + 0.0044067734852433205, + 0.003966139629483223, + 0.0032363426871597767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.009775268845260143 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013385392958298326 + ], + "opposite_logprob_tokens": [ + -6.626338481903076 + ], + "logprob_reward": -0.0013385392958298326, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013385392958298326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1148885190486908, + 0.1044129878282547, + 0.09346896409988403, + 0.13807299733161926, + 0.13100875914096832, + 0.05892016366124153, + 0.03135331720113754, + 0.11526832729578018, + 0.06839403510093689, + 0.14421193301677704 + ], + "credit_map": [ + 0.01207973062992096, + 0.014257032424211502, + 0.011374767869710922, + 0.01922852173447609, + 0.0234078336507082, + 0.00997727271169424, + 0.014862339943647385, + 0.009701013565063477, + 0.011319316923618317, + 0.015155816450715065, + 0.01607847958803177, + 0.012601842172443867, + 0.011319540441036224, + 0.007584674283862114, + 0.019450295716524124, + 0.010903018526732922, + 0.00765767740085721, + 0.010158671997487545, + 0.008867121301591396, + 0.015276475809514523, + 0.009547819383442402, + 0.012977604754269123, + 0.016130786389112473, + 0.012852808460593224, + 0.018126310780644417, + 0.015004890970885754, + 0.013966483063995838, + 0.02129640430212021, + 0.017139675095677376, + 0.016455668956041336, + 0.01772969961166382, + 0.018353864550590515, + 0.01091456227004528, + 0.014673711732029915, + 0.0076610492542386055, + 0.01876700669527054, + 0.011484921909868717, + 0.017606163397431374, + 0.03622700273990631, + 0.013674340210855007, + 0.006042621564120054, + 0.005425318609923124, + 0.006115827243775129, + 0.008765283040702343, + 0.011091942898929119, + 0.004062699154019356, + 0.009264463558793068, + 0.008152005262672901, + 0.0031955381855368614, + 0.003311967244371772, + 0.004538334906101227, + 0.004243697505444288, + 0.00453593023121357, + 0.0032833486329764128, + 0.0051468852907419205, + 0.0030976145062595606, + 0.011494047939777374, + 0.011385275050997734, + 0.012926025316119194, + 0.01465483009815216, + 0.01902625896036625, + 0.011801489628851414, + 0.017520152032375336, + 0.016460249200463295, + 0.010139468125998974, + 0.012307791039347649, + 0.008636928163468838, + 0.011149958707392216, + 0.007918234914541245, + 0.006364727858453989, + 0.006122321356087923, + 0.005754601210355759, + 0.0194551944732666, + 0.016764601692557335, + 0.019124703481793404, + 0.021501434966921806, + 0.015932897105813026, + 0.020956747233867645, + 0.016582828015089035, + 0.01389352697879076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.0013385392958298326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic close-up shot from a slightly overhead angle locks the viewer's gaze firmly on a cauldron of bubbling Sichuan hot pot. The vibrant, scalding broth is coated in a layer of shimmering, alluring red oil. Countless giant bubbles surge upward from the bottom, violently bursting on the surface, sending up tiny droplets of oil and releasing wisps of white steam. Freshly sliced, tender, tender beef rolls, freshly simmered in the broth, unfurl and tumble, their edges rapidly transforming from raw red to off-white. A few emerald green leaves, slightly wilting in the heat, join golden, saturated tofu and sliced shiitake mushrooms, bobbing and colliding with each surge of the broth, as if engaging in a passionate and chaotic dance. Warm lighting from above casts bright highlights on the glossy broth, imbuing the scene with a warm, rich, and earthy atmosphere, creating a mouth-watering and deeply impactful sensory experience.", + "seed": 27193, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_13/index_3387_seed_27193_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01585900327190757, + "vlm_logprob": -0.01585900327190757, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.01585900327190757, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014371813274919987 + ], + "opposite_logprob_tokens": [ + -4.889371871948242 + ], + "logprob_reward": -0.014371813274919987, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.014371813274919987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22916486859321594, + 0.06180562451481819, + 0.05647628754377365, + 0.06540950387716293, + 0.19242557883262634, + 0.0355742946267128, + 0.05767384544014931, + 0.05953364074230194, + 0.0642201378941536, + 0.1777161806821823 + ], + "credit_map": [ + 0.028348850086331367, + 0.021421730518341064, + 0.03651929646730423, + 0.030048448592424393, + 0.034232594072818756, + 0.03362349793314934, + 0.024410933256149292, + 0.02055952325463295, + 0.007915129885077477, + 0.007915043272078037, + 0.010263168253004551, + 0.0071708946488797665, + 0.007836299017071724, + 0.007651646155864, + 0.008528907783329487, + 0.004524533171206713, + 0.00684400787577033, + 0.00483392970636487, + 0.010663356631994247, + 0.0054790969006717205, + 0.0073736379854381084, + 0.0073563274927437305, + 0.008753265254199505, + 0.005172666162252426, + 0.008570169098675251, + 0.005359964445233345, + 0.009566349908709526, + 0.006479366682469845, + 0.0068426490761339664, + 0.010110211558640003, + 0.009618869051337242, + 0.008861924521625042, + 0.010956311598420143, + 0.006416425108909607, + 0.008823477663099766, + 0.009836920537054539, + 0.06575025618076324, + 0.009794596582651138, + 0.07142438739538193, + 0.009423197247087955, + 0.005841306876391172, + 0.0021846918389201164, + 0.005245808511972427, + 0.004161417484283447, + 0.004863946232944727, + 0.0043387445621192455, + 0.004069486632943153, + 0.004868895281106234, + 0.006428358145058155, + 0.0045737638138234615, + 0.006671728100627661, + 0.007056301459670067, + 0.007312431000173092, + 0.008631804026663303, + 0.006639864295721054, + 0.010359598323702812, + 0.004661584738641977, + 0.00553797697648406, + 0.00517331762239337, + 0.009748327545821667, + 0.005811658222228289, + 0.007884364575147629, + 0.007888324558734894, + 0.01282808743417263, + 0.004768067039549351, + 0.006391850300133228, + 0.006459448952227831, + 0.011591630056500435, + 0.008984021842479706, + 0.006607642397284508, + 0.009166620671749115, + 0.010250856168568134, + 0.015549481846392155, + 0.02159411832690239, + 0.012595406733453274, + 0.017721213400363922, + 0.04799111187458038, + 0.023018766194581985, + 0.02446851320564747, + 0.014777570962905884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.014371813274919987 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019516960019245744 + ], + "opposite_logprob_tokens": [ + -6.251951694488525 + ], + "logprob_reward": -0.0019516960019245744, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019516960019245744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14668476581573486, + 0.18365620076656342, + 0.055684544146060944, + 0.07427742332220078, + 0.084544338285923, + 0.025062693282961845, + 0.04154377803206444, + 0.2145908921957016, + 0.04206697270274162, + 0.1318884640932083 + ], + "credit_map": [ + 0.015159295871853828, + 0.013863621279597282, + 0.030017586424946785, + 0.01734991930425167, + 0.023292262107133865, + 0.01854659430682659, + 0.014904391020536423, + 0.013551103882491589, + 0.018781905993819237, + 0.026191286742687225, + 0.01693193055689335, + 0.024459030479192734, + 0.03513685241341591, + 0.0252855084836483, + 0.024711143225431442, + 0.012158545665442944, + 0.007058574352413416, + 0.0050651500932872295, + 0.008917911909520626, + 0.006232604384422302, + 0.0069095357321202755, + 0.008166782557964325, + 0.006647832691669464, + 0.006686151027679443, + 0.010278276167809963, + 0.006915397010743618, + 0.010911676101386547, + 0.009520061314105988, + 0.008991746231913567, + 0.010797007009387016, + 0.009052875451743603, + 0.007810391951352358, + 0.006487693637609482, + 0.004113256465643644, + 0.008990958333015442, + 0.008103745058178902, + 0.020470289513468742, + 0.00512349558994174, + 0.02556423842906952, + 0.005690666846930981, + 0.0025675471406430006, + 0.0018736537313088775, + 0.0026739859022200108, + 0.004610160831362009, + 0.003273602109402418, + 0.0037487533409148455, + 0.0028397480491548777, + 0.0034752395004034042, + 0.004423850681632757, + 0.003056184621527791, + 0.005520200356841087, + 0.005532786715775728, + 0.0051999990828335285, + 0.00768906157463789, + 0.004666757304221392, + 0.005454936996102333, + 0.017291713505983353, + 0.013788912445306778, + 0.021198054775595665, + 0.027485908940434456, + 0.020084166899323463, + 0.026500076055526733, + 0.03312155604362488, + 0.05512049421668053, + 0.003984397277235985, + 0.0028599801007658243, + 0.0038470355793833733, + 0.006289766635745764, + 0.005297317169606686, + 0.004948008339852095, + 0.006376277189701796, + 0.008464189246296883, + 0.011571942828595638, + 0.01523085031658411, + 0.009701566770672798, + 0.014421042054891586, + 0.033870987594127655, + 0.014441007748246193, + 0.020810646936297417, + 0.011840414255857468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.0019516960019245744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307127743959427 + ], + "opposite_logprob_tokens": [ + -3.1680712699890137 + ], + "logprob_reward": -0.04307127743959427, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307127743959427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13160459697246552, + 0.0561826229095459, + 0.06115907430648804, + 0.03384247049689293, + 0.13018862903118134, + 0.07362446933984756, + 0.050037555396556854, + 0.14617538452148438, + 0.11828286945819855, + 0.19890227913856506 + ], + "credit_map": [ + 0.00969731155782938, + 0.018347345292568207, + 0.01941552199423313, + 0.021706219762563705, + 0.019726330414414406, + 0.01672731712460518, + 0.015141695737838745, + 0.010842850431799889, + 0.005109633319079876, + 0.006016905885189772, + 0.007502925116568804, + 0.007466247770935297, + 0.006658404134213924, + 0.007355127949267626, + 0.008763527497649193, + 0.007309847045689821, + 0.006826768163591623, + 0.005573146976530552, + 0.008686121553182602, + 0.005755129735916853, + 0.009019470773637295, + 0.009011908434331417, + 0.008972255513072014, + 0.007314274553209543, + 0.0037874924018979073, + 0.0026880940422415733, + 0.005681893322616816, + 0.003422204637899995, + 0.0036216266453266144, + 0.004903819877654314, + 0.004245356656610966, + 0.005491981748491526, + 0.008600748144090176, + 0.005383543204516172, + 0.009046509861946106, + 0.007266202475875616, + 0.040297213941812515, + 0.00673769460991025, + 0.04685819149017334, + 0.005998531822115183, + 0.008265524171292782, + 0.006430363282561302, + 0.010281686671078205, + 0.012001639232039452, + 0.007569152861833572, + 0.007449445780366659, + 0.011411424726247787, + 0.010215229354798794, + 0.005007921252399683, + 0.0032082567922770977, + 0.004117280710488558, + 0.006917343009263277, + 0.006092659197747707, + 0.011894170194864273, + 0.006429996807128191, + 0.006369927432388067, + 0.010031796991825104, + 0.008340422064065933, + 0.014765300787985325, + 0.01494293101131916, + 0.0154549740254879, + 0.017787454649806023, + 0.023839978501200676, + 0.041012536734342575, + 0.007575144991278648, + 0.012392756529152393, + 0.01120696309953928, + 0.017418988049030304, + 0.014263526536524296, + 0.015261751599609852, + 0.02095142751932144, + 0.019212311133742332, + 0.01930275931954384, + 0.03746868297457695, + 0.01776217482984066, + 0.027310870587825775, + 0.019593754783272743, + 0.02760489098727703, + 0.03079359419643879, + 0.019065558910369873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.04307127743959427 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184611573815346 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184611573815346, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184611573815346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08190838247537613, + 0.031296055763959885, + 0.020778141915798187, + 0.020725829526782036, + 0.23120759427547455, + 0.018626175820827484, + 0.020601516589522362, + 0.36261656880378723, + 0.14414824545383453, + 0.06809151917695999 + ], + "credit_map": [ + 0.007519463077187538, + 0.008646215312182903, + 0.012831303291022778, + 0.012813090346753597, + 0.01426214724779129, + 0.008593925274908543, + 0.009385676123201847, + 0.00785655528306961, + 0.003596249036490917, + 0.004689518827944994, + 0.0032333973795175552, + 0.004908462520688772, + 0.004209384322166443, + 0.00409943051636219, + 0.0041481913067400455, + 0.002411419991403818, + 0.0022874006535857916, + 0.0022261948324739933, + 0.0035588995087891817, + 0.002387029118835926, + 0.0023515187203884125, + 0.0026069488376379013, + 0.002599897561594844, + 0.0027602529153227806, + 0.0022254774812608957, + 0.0017995656235143542, + 0.003578907111659646, + 0.0026331234257668257, + 0.0023889753501862288, + 0.0024905626196414232, + 0.003230065107345581, + 0.002379151526838541, + 0.016259724274277687, + 0.00861342716962099, + 0.02121337316930294, + 0.014025661163032055, + 0.06745070219039917, + 0.014074183069169521, + 0.07571081817150116, + 0.013859689235687256, + 0.002460392890498042, + 0.0011215837439522147, + 0.002252232050523162, + 0.003297475865110755, + 0.0022865200880914927, + 0.0019170452142134309, + 0.0023964799474924803, + 0.002894445089623332, + 0.002156898844987154, + 0.001756213023327291, + 0.002693925052881241, + 0.0032423059456050396, + 0.002446040976792574, + 0.0030259587801992893, + 0.002229835372418165, + 0.0030503375455737114, + 0.02982426807284355, + 0.02144613489508629, + 0.03592411428689957, + 0.049047715961933136, + 0.03175890073180199, + 0.04807433485984802, + 0.05404960736632347, + 0.09249148517847061, + 0.011918123811483383, + 0.012658671475946903, + 0.015392601490020752, + 0.018967827782034874, + 0.02348928153514862, + 0.01512585673481226, + 0.017343595623970032, + 0.029252277687191963, + 0.0066879852674901485, + 0.011891665868461132, + 0.006184129510074854, + 0.008036784827709198, + 0.00936947949230671, + 0.007748291827738285, + 0.010635437443852425, + 0.0075377426110208035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.018184611573815346 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle focused on a cauldron of Sichuan hot pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth vibrant red with a layer of shimmering oil and countless giant bubbles bursting on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tender beef rolls visible unfurling and changing color from raw red to off-white in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green leaves, golden tofu, and sliced shiitake mushrooms visible bobbing in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with warm lighting from above casting bright highlights on the glossy broth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017156180692836642 + ], + "opposite_logprob_tokens": [ + -6.376715660095215 + ], + "logprob_reward": -0.0017156180692836642, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017156180692836642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23939357697963715, + 0.19162775576114655, + 0.045617107301950455, + 0.06902434676885605, + 0.04879041761159897, + 0.03973705694079399, + 0.0322946235537529, + 0.14363792538642883, + 0.08222593367099762, + 0.1076512336730957 + ], + "credit_map": [ + 0.026582665741443634, + 0.02846105396747589, + 0.03293829411268234, + 0.032447513192892075, + 0.026044484227895737, + 0.028348376974463463, + 0.03866772726178169, + 0.02590346150100231, + 0.01850147172808647, + 0.028459006920456886, + 0.024499045684933662, + 0.024583932012319565, + 0.02372145466506481, + 0.022038696333765984, + 0.030706042423844337, + 0.01911810040473938, + 0.006148836109787226, + 0.004525934811681509, + 0.008168711327016354, + 0.004455328918993473, + 0.007579741068184376, + 0.005910980980843306, + 0.004981447011232376, + 0.0038461245130747557, + 0.006842009257525206, + 0.0075886910781264305, + 0.009758143685758114, + 0.008377261459827423, + 0.008365228772163391, + 0.010524129495024681, + 0.008931081742048264, + 0.008637802675366402, + 0.004976051393896341, + 0.002409131731837988, + 0.008246403187513351, + 0.005231362767517567, + 0.00945227686315775, + 0.005646286997944117, + 0.007493576034903526, + 0.005335329100489616, + 0.003991513513028622, + 0.003593939356505871, + 0.0053957547061145306, + 0.0077578057534992695, + 0.003726257709786296, + 0.004003628622740507, + 0.005612976383417845, + 0.005655181128531694, + 0.0029307540971785784, + 0.0022559771314263344, + 0.003394812112674117, + 0.0038898878265172243, + 0.003534206422045827, + 0.006963539402931929, + 0.0046658506616950035, + 0.004659593570977449, + 0.009497536346316338, + 0.009088869206607342, + 0.011824057437479496, + 0.023134473711252213, + 0.013001312501728535, + 0.01882781647145748, + 0.02212618850171566, + 0.03613767772912979, + 0.00537109375, + 0.0074743530713021755, + 0.008029624819755554, + 0.012969396077096462, + 0.012650315649807453, + 0.00832682941108942, + 0.010006602853536606, + 0.017397722229361534, + 0.008397659286856651, + 0.01924082264304161, + 0.009389706887304783, + 0.015271781012415886, + 0.012064572423696518, + 0.016027245670557022, + 0.018618743866682053, + 0.008640710264444351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0017156180692836642 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_14/index_tensor([8459, 8459])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_14/index_tensor([8459, 8459])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f098fc5254c57b1829d0b1499e3aae05aa002ef1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_14/index_tensor([8459, 8459])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle shot captures an iconic coupe, its sleek metallic silver body accented by bold, geometric racing vinyl in a vibrant electric blue, executing a flawless, tight spin on rain-slicked, well-lit black asphalt. The entire scene is rendered in a vibrant urbane art style, defined by sharp outlines that imbue every element with a crisp graphic quality and a highly saturated color palette. As the car spins with controlled violence, its tires explode with dramatic sprays of water that capture the kaleidoscopic glow of holographic advertising, atomizing into a shimmering mist. The camera then begins to smoothly pull back and forth, seamlessly transitioning to a wide-angle shot, revealing a vast, multi-layered futuristic cityscape. Its densely packed buildings and glowing light paths are beautifully reflected in the mirror-like puddles covering the streets below, creating a stunning and vibrant cityscape.", + "seed": 9742, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_14/index_8459_seed_9742_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.042901539790909736, + "vlm_logprob": -0.042901539790909736, + "vlm_yes_no_margin": 4.024999952316284, + "vlm_token_credit": -0.042901539790909736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016155876219272614 + ], + "opposite_logprob_tokens": [ + -4.39115571975708 + ], + "logprob_reward": -0.016155876219272614, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016155876219272614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3252643346786499, + 0.09970807284116745, + 0.1360739767551422, + 0.07580038905143738, + 0.09029213339090347, + 0.10578852891921997, + 0.03786900267004967, + 0.038846611976623535, + 0.043013665825128555, + 0.047343265265226364 + ], + "credit_map": [ + 0.019640015438199043, + 0.019696425646543503, + 0.02576107159256935, + 0.02450588345527649, + 0.025264766067266464, + 0.030331503599882126, + 0.08871636539697647, + 0.0913483053445816, + 0.009819271042943, + 0.012292664498090744, + 0.015483947470784187, + 0.015815187245607376, + 0.011670744977891445, + 0.0076618255116045475, + 0.013625328429043293, + 0.013339100405573845, + 0.01701102964580059, + 0.016194844618439674, + 0.020411362871527672, + 0.01903272420167923, + 0.013234377838671207, + 0.014432759955525398, + 0.016946831718087196, + 0.018810037523508072, + 0.007180829998105764, + 0.008871311321854591, + 0.008057300001382828, + 0.010905910283327103, + 0.0078119379468262196, + 0.005553246010094881, + 0.017717281356453896, + 0.009702572599053383, + 0.015440910123288631, + 0.012703681364655495, + 0.012018549256026745, + 0.012388278730213642, + 0.012593647465109825, + 0.01044596079736948, + 0.008748751133680344, + 0.00595235638320446, + 0.02571825124323368, + 0.009658410213887691, + 0.027809441089630127, + 0.013083240948617458, + 0.009103209711611271, + 0.005731766577810049, + 0.008931580930948257, + 0.00575263099744916, + 0.0026502280961722136, + 0.005899905227124691, + 0.003092003520578146, + 0.0033999960869550705, + 0.005713604390621185, + 0.008532912470400333, + 0.00354422302916646, + 0.0050361319445073605, + 0.005427334923297167, + 0.004807232413440943, + 0.007562227547168732, + 0.004696803167462349, + 0.003620461095124483, + 0.0038402958307415247, + 0.004465480800718069, + 0.004426775500178337, + 0.004443655256181955, + 0.007117911707609892, + 0.0037639648653566837, + 0.006203353404998779, + 0.006606189999729395, + 0.005255323834717274, + 0.00507945753633976, + 0.004543807823210955, + 0.00421144487336278, + 0.005649934522807598, + 0.009367027319967747, + 0.006892194971442223, + 0.005154268350452185, + 0.00540276151150465, + 0.005533997435122728, + 0.005131636280566454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.016155876219272614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897930592298508 + ], + "opposite_logprob_tokens": [ + -2.578979253768921 + ], + "logprob_reward": -0.07897930592298508, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897930592298508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1268002688884735, + 0.10397007316350937, + 0.09305014461278915, + 0.05224520340561867, + 0.09110589325428009, + 0.10492587834596634, + 0.12739434838294983, + 0.157912015914917, + 0.03638729453086853, + 0.10620889812707901 + ], + "credit_map": [ + 0.008721638470888138, + 0.012116122990846634, + 0.012616876512765884, + 0.0098239965736866, + 0.011731421574950218, + 0.012036445550620556, + 0.024297181516885757, + 0.035456594079732895, + 0.011685349978506565, + 0.014291880652308464, + 0.01303159911185503, + 0.018720634281635284, + 0.010243801400065422, + 0.007056876085698605, + 0.01356520690023899, + 0.015374722890555859, + 0.016613829880952835, + 0.010200428776443005, + 0.016894932836294174, + 0.011429909616708755, + 0.00908589456230402, + 0.009045653976500034, + 0.008394281379878521, + 0.011385210789740086, + 0.005080939270555973, + 0.005577189847826958, + 0.007139044813811779, + 0.00946046132594347, + 0.004933337680995464, + 0.004538290668278933, + 0.009604777209460735, + 0.00591116352006793, + 0.016306346282362938, + 0.01591668650507927, + 0.013468622229993343, + 0.019168676808476448, + 0.006948054302483797, + 0.006263467483222485, + 0.007703071925789118, + 0.005330964457243681, + 0.021527202799916267, + 0.010644449852406979, + 0.021691853180527687, + 0.016968833282589912, + 0.0072418637573719025, + 0.00738377682864666, + 0.012272155843675137, + 0.007195739541202784, + 0.017069878056645393, + 0.017643164843320847, + 0.017627717927098274, + 0.015774298459291458, + 0.014551891945302486, + 0.015936432406306267, + 0.011600648052990437, + 0.01719031296670437, + 0.017757391557097435, + 0.023979712277650833, + 0.02522820420563221, + 0.01808650605380535, + 0.01244631502777338, + 0.01671968773007393, + 0.01854446344077587, + 0.025149740278720856, + 0.0035534522030502558, + 0.0061362371779978275, + 0.0055223447270691395, + 0.006052955519407988, + 0.004902068059891462, + 0.002778226975351572, + 0.0043488466180861, + 0.003093161853030324, + 0.014510979875922203, + 0.012035028077661991, + 0.01776955835521221, + 0.013255632482469082, + 0.015488997101783752, + 0.011227133683860302, + 0.01076270081102848, + 0.011158865876495838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.07897930592298508 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818484626710415 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.01818484626710415, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818484626710415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15354692935943604, + 0.05008533224463463, + 0.29068896174430847, + 0.028411585837602615, + 0.11690905690193176, + 0.036654725670814514, + 0.07301905006170273, + 0.12819337844848633, + 0.037305187433958054, + 0.08518582582473755 + ], + "credit_map": [ + 0.008164667524397373, + 0.007505467627197504, + 0.010287247598171234, + 0.011749116703867912, + 0.00902261771261692, + 0.012442357838153839, + 0.04194604232907295, + 0.052429407835006714, + 0.005424387287348509, + 0.007323754485696554, + 0.006432132795453072, + 0.008857552893459797, + 0.004934046883136034, + 0.0038797908928245306, + 0.006119146943092346, + 0.007114516105502844, + 0.06264682114124298, + 0.03925427421927452, + 0.04001326113939285, + 0.025357259437441826, + 0.02582404389977455, + 0.03252459317445755, + 0.026650739833712578, + 0.03841794654726982, + 0.0033731001894921064, + 0.002892571734264493, + 0.0026182043366134167, + 0.0039450908079743385, + 0.003174058860167861, + 0.003502913983538747, + 0.005588184576481581, + 0.0033174604177474976, + 0.0207561943680048, + 0.019386878237128258, + 0.017838746309280396, + 0.02221597544848919, + 0.010453527793288231, + 0.008626982569694519, + 0.010398322716355324, + 0.007232425268739462, + 0.006738136988133192, + 0.0032354402355849743, + 0.008094880729913712, + 0.004827407654374838, + 0.003060480346903205, + 0.003039733972400427, + 0.004381620325148106, + 0.0032770244870334864, + 0.010499165393412113, + 0.01022977102547884, + 0.010351007804274559, + 0.008516375906765461, + 0.008442550897598267, + 0.00995889212936163, + 0.006578525528311729, + 0.008442752063274384, + 0.015675421804189682, + 0.020920630544424057, + 0.024594975635409355, + 0.021915817633271217, + 0.010392655618488789, + 0.011251996271312237, + 0.011387527920305729, + 0.01205433625727892, + 0.003140433691442013, + 0.007493259850889444, + 0.005027256906032562, + 0.007729973178356886, + 0.003099812427535653, + 0.0034640466328710318, + 0.004328000824898481, + 0.00302239996381104, + 0.010392552241683006, + 0.009058823809027672, + 0.015574242919683456, + 0.012620124965906143, + 0.011191052384674549, + 0.008390901610255241, + 0.009659606032073498, + 0.008298511616885662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.01818484626710415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009195152088068426 + ], + "opposite_logprob_tokens": [ + -7.000919342041016 + ], + "logprob_reward": -0.0009195152088068426, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009195152088068426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13134092092514038, + 0.17558780312538147, + 0.1066480502486229, + 0.030064284801483154, + 0.19948826730251312, + 0.06290369480848312, + 0.0572822242975235, + 0.09856856614351273, + 0.07528016716241837, + 0.06283595412969589 + ], + "credit_map": [ + 0.008089408278465271, + 0.007240473758429289, + 0.013120565563440323, + 0.009838486090302467, + 0.008936886675655842, + 0.012685421854257584, + 0.030623558908700943, + 0.04080614075064659, + 0.021594803780317307, + 0.03580313175916672, + 0.01710660010576248, + 0.01614699698984623, + 0.022824186831712723, + 0.013290231116116047, + 0.02027488686144352, + 0.028546979650855064, + 0.021444736048579216, + 0.013876098208129406, + 0.01351558044552803, + 0.011204395443201065, + 0.010021412745118141, + 0.012485395185649395, + 0.009425870142877102, + 0.014674568548798561, + 0.0022761039435863495, + 0.0028168801218271255, + 0.0036221733316779137, + 0.00504764961078763, + 0.004655638709664345, + 0.0036256795283406973, + 0.004561120178550482, + 0.00345904054120183, + 0.03879808634519577, + 0.03324253484606743, + 0.03145381435751915, + 0.03131810203194618, + 0.019218232482671738, + 0.017625335603952408, + 0.016556309536099434, + 0.011275877244770527, + 0.012761116027832031, + 0.005904617719352245, + 0.014381366781890392, + 0.00981108658015728, + 0.0058691068552434444, + 0.0028648630250245333, + 0.006850857753306627, + 0.00446068262681365, + 0.009056982584297657, + 0.007435592822730541, + 0.007306216284632683, + 0.005426183808594942, + 0.007458625826984644, + 0.0079134376719594, + 0.006626033689826727, + 0.006059158593416214, + 0.008350406773388386, + 0.014738819561898708, + 0.01684361882507801, + 0.015385903418064117, + 0.010003538802266121, + 0.00947064720094204, + 0.016103971749544144, + 0.0076716686598956585, + 0.007884877733886242, + 0.009153425693511963, + 0.011058271862566471, + 0.01247598510235548, + 0.007901065982878208, + 0.00844855047762394, + 0.009743106551468372, + 0.008614890277385712, + 0.006753561086952686, + 0.006475562695413828, + 0.01153741218149662, + 0.00822904147207737, + 0.008691688999533653, + 0.007578116841614246, + 0.008467476814985275, + 0.005103099159896374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0009195152088068426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10026815533638 + ], + "opposite_logprob_tokens": [ + -2.3502681255340576 + ], + "logprob_reward": -0.10026815533638, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10026815533638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07260031253099442, + 0.04527386650443077, + 0.1029450073838234, + 0.06168688088655472, + 0.20011787116527557, + 0.08023322373628616, + 0.09396776556968689, + 0.21120604872703552, + 0.032982658594846725, + 0.09898635745048523 + ], + "credit_map": [ + 0.004278137348592281, + 0.007065259385854006, + 0.007830142974853516, + 0.007704025134444237, + 0.006044798996299505, + 0.007862251251935959, + 0.013030687347054482, + 0.018785003572702408, + 0.00556113151833415, + 0.005070428363978863, + 0.005766751244664192, + 0.006313619669526815, + 0.004822324030101299, + 0.0035468200221657753, + 0.006928338203579187, + 0.007264450658112764, + 0.02066277153789997, + 0.00834747590124607, + 0.01518677081912756, + 0.01530674658715725, + 0.009262236766517162, + 0.009360773488879204, + 0.009716026484966278, + 0.015102192759513855, + 0.005306169856339693, + 0.005115710664540529, + 0.006839077454060316, + 0.00792479608207941, + 0.010322344489395618, + 0.009163081645965576, + 0.010476856492459774, + 0.006538834422826767, + 0.04160696640610695, + 0.030093785375356674, + 0.030684376135468483, + 0.031461723148822784, + 0.018319809809327126, + 0.021798323839902878, + 0.014595318585634232, + 0.01155755203217268, + 0.013178152963519096, + 0.01087863277643919, + 0.011544693261384964, + 0.011748085729777813, + 0.011101734824478626, + 0.007695611100643873, + 0.0073913633823394775, + 0.006694948300719261, + 0.008278842084109783, + 0.019284935668110847, + 0.008635709062218666, + 0.007708418648689985, + 0.011095138266682625, + 0.01544212643057108, + 0.010295817628502846, + 0.013226764276623726, + 0.025342412292957306, + 0.051731616258621216, + 0.02957076206803322, + 0.033238932490348816, + 0.02109731175005436, + 0.017476696521043777, + 0.017888735979795456, + 0.014859572984278202, + 0.0029157367534935474, + 0.004536633845418692, + 0.004901976324617863, + 0.006207335274666548, + 0.005007525905966759, + 0.002746486570686102, + 0.0041159153915941715, + 0.002551042940467596, + 0.01114028412848711, + 0.010861508548259735, + 0.01726069487631321, + 0.013887518085539341, + 0.012091532349586487, + 0.011881611309945583, + 0.01250202301889658, + 0.009361173026263714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.10026815533638 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot captures an iconic coupe, its sleek metallic silver body accented by bold, geometric racing vinyl in a vibrant electric blue, executing a flawless, tight spin on rain-slicked, well-lit black asphalt. The entire scene is rendered in a vibrant urbane art style, defined by sharp outlines that imbue every element with a crisp graphic quality and a highly saturated color palette. As the car spins with controlled violence, its tires explode with dramatic sprays of water that capture the kaleidoscopic glow of holographic advertising, atomizing into a shimmering mist. The camera then begins to smoothly pull back and forth, seamlessly transitioning to a wide-angle shot, revealing a vast, multi-layered futuristic cityscape. Its densely packed buildings and glowing light paths are beautifully reflected in the mirror-like puddles covering the streets below, creating a stunning and vibrant cityscape.", + "seed": 9743, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_14/index_8459_seed_9743_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06988557651638985, + "vlm_logprob": -0.06988557651638985, + "vlm_yes_no_margin": 3.3999999523162843, + "vlm_token_credit": -0.06988557651638985, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02343667298555374 + ], + "opposite_logprob_tokens": [ + -4.023436546325684 + ], + "logprob_reward": -0.02343667298555374, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02343667298555374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17690499126911163, + 0.0846656933426857, + 0.09117740392684937, + 0.1173192635178566, + 0.12962467968463898, + 0.10736728459596634, + 0.04339021444320679, + 0.03469878435134888, + 0.046023424714803696, + 0.16882824897766113 + ], + "credit_map": [ + 0.020377468317747116, + 0.020538562908768654, + 0.02729046903550625, + 0.01980336755514145, + 0.02549276314675808, + 0.02242373861372471, + 0.016168123111128807, + 0.024810506030917168, + 0.01556174736469984, + 0.008979497477412224, + 0.01050089206546545, + 0.009547274559736252, + 0.011884128674864769, + 0.008813271299004555, + 0.00867948867380619, + 0.010699396021664143, + 0.015080285258591175, + 0.007223562337458134, + 0.012296995148062706, + 0.009001378901302814, + 0.008360292762517929, + 0.010369883850216866, + 0.016166167333722115, + 0.012678838334977627, + 0.016901491209864616, + 0.010685393586754799, + 0.0080955121666193, + 0.013204510323703289, + 0.019272392615675926, + 0.01059360709041357, + 0.014007236808538437, + 0.02455911971628666, + 0.025193361565470695, + 0.011048609390854836, + 0.014158607460558414, + 0.015182471834123135, + 0.006351795978844166, + 0.006612326018512249, + 0.03677631914615631, + 0.014301186427474022, + 0.014180799946188927, + 0.011851925402879715, + 0.015495015308260918, + 0.01874168962240219, + 0.013163154944777489, + 0.012094615027308464, + 0.011699700728058815, + 0.010140388272702694, + 0.0045453147031366825, + 0.005686802789568901, + 0.009771548211574554, + 0.005710994824767113, + 0.00326031306758523, + 0.004597092512995005, + 0.004289806820452213, + 0.005528339650481939, + 0.004150091204792261, + 0.004804352298378944, + 0.003467613598331809, + 0.005658097565174103, + 0.004279083106666803, + 0.003685155650600791, + 0.004026475362479687, + 0.0046279169619083405, + 0.005812182556837797, + 0.0043825688771903515, + 0.006331011652946472, + 0.005735006649047136, + 0.0046693128533661366, + 0.00416957912966609, + 0.008213061839342117, + 0.006710702553391457, + 0.008117751218378544, + 0.01008073054254055, + 0.009352589026093483, + 0.023802706971764565, + 0.01495737861841917, + 0.01388584729284048, + 0.05345717817544937, + 0.03517407923936844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.02343667298555374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866397753357887 + ], + "opposite_logprob_tokens": [ + -3.048664093017578 + ], + "logprob_reward": -0.04866397753357887, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866397753357887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09672892838716507, + 0.021796341985464096, + 0.2667267620563507, + 0.05296993628144264, + 0.1537207067012787, + 0.06140155345201492, + 0.06567411124706268, + 0.12375291436910629, + 0.08416934311389923, + 0.07305936515331268 + ], + "credit_map": [ + 0.01322331465780735, + 0.012669958174228668, + 0.011473669670522213, + 0.011500220745801926, + 0.012259124778211117, + 0.009816128760576248, + 0.008281624875962734, + 0.01750488393008709, + 0.0021571379620581865, + 0.001801528618671, + 0.002904048189520836, + 0.00213065417483449, + 0.0037352414801716805, + 0.002864864422008395, + 0.002679979195818305, + 0.0035228892229497433, + 0.0335104800760746, + 0.017318131402134895, + 0.025316564366221428, + 0.026443665847182274, + 0.028283867985010147, + 0.02595769613981247, + 0.07926136255264282, + 0.030635008588433266, + 0.0061795711517333984, + 0.006219911854714155, + 0.0050464803352952, + 0.006735115312039852, + 0.005530648864805698, + 0.008871368132531643, + 0.006963170133531094, + 0.007423667702823877, + 0.03601235896348953, + 0.012489586137235165, + 0.013070734217762947, + 0.01904442347586155, + 0.007008426822721958, + 0.011167775839567184, + 0.04285024106502533, + 0.012077153660356998, + 0.01033441349864006, + 0.006168005522340536, + 0.01279893983155489, + 0.006819272413849831, + 0.006748329848051071, + 0.005874269176274538, + 0.005644429475069046, + 0.007013891823589802, + 0.007583576254546642, + 0.007814290933310986, + 0.012261266820132732, + 0.01389961875975132, + 0.004976615309715271, + 0.005017292220145464, + 0.005682041402906179, + 0.008439409546554089, + 0.007596443872898817, + 0.014970359392464161, + 0.014324958436191082, + 0.0245976522564888, + 0.014689718373119831, + 0.015436563640832901, + 0.0160859152674675, + 0.01605130359530449, + 0.015897661447525024, + 0.010294578038156033, + 0.008025125600397587, + 0.00887439027428627, + 0.008828247897326946, + 0.012064000591635704, + 0.007597613148391247, + 0.012587727047502995, + 0.006876043975353241, + 0.004875604994595051, + 0.0038355395663529634, + 0.00810170266777277, + 0.008267969824373722, + 0.008692183531820774, + 0.0186732467263937, + 0.013737073168158531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.04866397753357887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193067327141762 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193067327141762, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193067327141762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11358164250850677, + 0.06534561514854431, + 0.116756871342659, + 0.053530771285295486, + 0.04633885622024536, + 0.0721140131354332, + 0.18201281130313873, + 0.10216950625181198, + 0.04769382253289223, + 0.20045606791973114 + ], + "credit_map": [ + 0.01067913044244051, + 0.01578863337635994, + 0.017118185758590698, + 0.011658698320388794, + 0.014307230710983276, + 0.012941967695951462, + 0.01025309693068266, + 0.020834704861044884, + 0.011545290239155293, + 0.006139859557151794, + 0.008508486673235893, + 0.007586863823235035, + 0.01070859283208847, + 0.005931790452450514, + 0.00759748974815011, + 0.007327237632125616, + 0.015160875394940376, + 0.008688417263329029, + 0.011814267374575138, + 0.011477134190499783, + 0.014036748558282852, + 0.011247050017118454, + 0.03167201206088066, + 0.0126603739336133, + 0.00560219120234251, + 0.00392512883991003, + 0.00767116155475378, + 0.004843559116125107, + 0.0076001775451004505, + 0.006639554630964994, + 0.008053682744503021, + 0.009195318445563316, + 0.004804643336683512, + 0.0054864133708179, + 0.004588793497532606, + 0.005073900334537029, + 0.0035787723027169704, + 0.00318999239243567, + 0.010533186607062817, + 0.009083157405257225, + 0.015052313916385174, + 0.007500205188989639, + 0.01340288482606411, + 0.01297715213149786, + 0.004743964411318302, + 0.006237835623323917, + 0.006284774746745825, + 0.005914884153753519, + 0.022183101624250412, + 0.029143625870347023, + 0.032753609120845795, + 0.027362458407878876, + 0.014805939048528671, + 0.016739604994654655, + 0.01932152546942234, + 0.019702941179275513, + 0.011433679610490799, + 0.01522411871701479, + 0.01162467710673809, + 0.01713641919195652, + 0.01450436282902956, + 0.010959777049720287, + 0.009438765235245228, + 0.011847712099552155, + 0.005598611664026976, + 0.006538706831634045, + 0.0037615266628563404, + 0.0060534849762916565, + 0.006033821497112513, + 0.006535734049975872, + 0.005885510705411434, + 0.007286428473889828, + 0.020532459020614624, + 0.015972884371876717, + 0.011461031623184681, + 0.024910584092140198, + 0.022741621360182762, + 0.019213879480957985, + 0.04943237453699112, + 0.036191247403621674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.014193067327141762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072618886828423 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072618886828423, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072618886828423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06520090997219086, + 0.0657869428396225, + 0.13094691932201385, + 0.059296172112226486, + 0.13924671709537506, + 0.09378980845212936, + 0.04276157543063164, + 0.06413114815950394, + 0.05814152956008911, + 0.2806982398033142 + ], + "credit_map": [ + 0.007588579319417477, + 0.007747879717499018, + 0.010446636937558651, + 0.007777459919452667, + 0.007913992740213871, + 0.007607916835695505, + 0.006067909765988588, + 0.010050534270703793, + 0.009537868201732635, + 0.0069945803843438625, + 0.007640954107046127, + 0.006752208806574345, + 0.012445196509361267, + 0.006687464192509651, + 0.007787882350385189, + 0.00794079527258873, + 0.018022233620285988, + 0.009529017843306065, + 0.013890892267227173, + 0.01591981016099453, + 0.013009347952902317, + 0.015433893539011478, + 0.0315437838435173, + 0.013597948476672173, + 0.008035442791879177, + 0.005354490131139755, + 0.0038883108645677567, + 0.006203959695994854, + 0.009232164360582829, + 0.005829442758113146, + 0.0067606838420033455, + 0.01399168185889721, + 0.019062979146838188, + 0.016635483130812645, + 0.013809005729854107, + 0.01710168831050396, + 0.006871802266687155, + 0.007758669555187225, + 0.03537633270025253, + 0.02263076975941658, + 0.012900661677122116, + 0.009748813696205616, + 0.013373201712965965, + 0.016690973192453384, + 0.011143765412271023, + 0.010325415059924126, + 0.010386930778622627, + 0.009220060892403126, + 0.0054712556302547455, + 0.0057125692255795, + 0.008787712082266808, + 0.007424750365316868, + 0.003898531198501587, + 0.0030689567793160677, + 0.004589271266013384, + 0.0038085300475358963, + 0.006828901823610067, + 0.009172635152935982, + 0.008336616680026054, + 0.013619979843497276, + 0.006751713342964649, + 0.006051869597285986, + 0.0067732445895671844, + 0.006596197374165058, + 0.00648507010191679, + 0.006286689080297947, + 0.006960074882954359, + 0.007456242572516203, + 0.007545380853116512, + 0.005217714235186577, + 0.00930245965719223, + 0.00888790376484394, + 0.013838738203048706, + 0.01532982848584652, + 0.015843451023101807, + 0.039562471210956573, + 0.02420804835855961, + 0.021347375586628914, + 0.08917637169361115, + 0.06139201670885086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.011072618886828423 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25206154584884644 + ], + "opposite_logprob_tokens": [ + -1.5020616054534912 + ], + "logprob_reward": -0.25206154584884644, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25206154584884644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07315151393413544, + 0.06470038741827011, + 0.2259853631258011, + 0.055884990841150284, + 0.05943477898836136, + 0.06663808226585388, + 0.08037881553173065, + 0.09021275490522385, + 0.11338508129119873, + 0.1702282577753067 + ], + "credit_map": [ + 0.008695357479155064, + 0.008910958655178547, + 0.009026569314301014, + 0.009612477384507656, + 0.01104777306318283, + 0.007885620929300785, + 0.006704216357320547, + 0.011268540285527706, + 0.008779752068221569, + 0.005966495256870985, + 0.005994467996060848, + 0.007090112194418907, + 0.011475716717541218, + 0.007363324984908104, + 0.008476445451378822, + 0.009554076008498669, + 0.02725156769156456, + 0.01572914607822895, + 0.01958506554365158, + 0.02245635911822319, + 0.024110985919833183, + 0.02283608727157116, + 0.06852028518915176, + 0.025495873764157295, + 0.007717409171164036, + 0.0067972964607179165, + 0.006287830416113138, + 0.006091733928769827, + 0.006194809917360544, + 0.005931990686804056, + 0.009753103367984295, + 0.007110819220542908, + 0.006576534826308489, + 0.0122330067679286, + 0.006020736414939165, + 0.006417096126824617, + 0.0031080960761755705, + 0.003842383623123169, + 0.012404249981045723, + 0.008832675404846668, + 0.010991417802870274, + 0.007109314203262329, + 0.011855186894536018, + 0.010489599779248238, + 0.005674147978425026, + 0.006468256004154682, + 0.0071777780540287495, + 0.00687238946557045, + 0.012399365194141865, + 0.013960185460746288, + 0.015053215436637402, + 0.013350505381822586, + 0.00612915214151144, + 0.005307693500071764, + 0.007439687382429838, + 0.0067390077747404575, + 0.00864453800022602, + 0.014968994073569775, + 0.009077928960323334, + 0.014553319662809372, + 0.01171280350536108, + 0.008991456590592861, + 0.011006495915353298, + 0.011257218196988106, + 0.01542348600924015, + 0.015573290176689625, + 0.012414982542395592, + 0.015573921613395214, + 0.013881690800189972, + 0.015372859314084053, + 0.012046429328620434, + 0.013098422437906265, + 0.022787243127822876, + 0.01940136030316353, + 0.009661014191806316, + 0.021457912400364876, + 0.027441255748271942, + 0.021542619913816452, + 0.024916432797908783, + 0.0230204239487648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.25206154584884644 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot captures an iconic coupe, its sleek metallic silver body accented by bold, geometric racing vinyl in a vibrant electric blue, executing a flawless, tight spin on rain-slicked, well-lit black asphalt. The entire scene is rendered in a vibrant urbane art style, defined by sharp outlines that imbue every element with a crisp graphic quality and a highly saturated color palette. As the car spins with controlled violence, its tires explode with dramatic sprays of water that capture the kaleidoscopic glow of holographic advertising, atomizing into a shimmering mist. The camera then begins to smoothly pull back and forth, seamlessly transitioning to a wide-angle shot, revealing a vast, multi-layered futuristic cityscape. Its densely packed buildings and glowing light paths are beautifully reflected in the mirror-like puddles covering the streets below, creating a stunning and vibrant cityscape.", + "seed": 9744, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_14/index_8459_seed_9744_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026919751695822926, + "vlm_logprob": -0.026919751695822926, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.026919751695822926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01882929354906082 + ], + "opposite_logprob_tokens": [ + -4.268829345703125 + ], + "logprob_reward": -0.01882929354906082, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01882929354906082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13894329965114594, + 0.09108045697212219, + 0.17088378965854645, + 0.12125248461961746, + 0.07830069959163666, + 0.10294724255800247, + 0.131951704621315, + 0.044952910393476486, + 0.05460783839225769, + 0.06507954746484756 + ], + "credit_map": [ + 0.01513691060245037, + 0.015939125791192055, + 0.013625151477754116, + 0.02473767288029194, + 0.014183909632265568, + 0.02434459514915943, + 0.016579793766140938, + 0.014396137557923794, + 0.010015386156737804, + 0.008600411005318165, + 0.009910060092806816, + 0.012726523913443089, + 0.014113410376012325, + 0.010072543285787106, + 0.013732125982642174, + 0.01190999150276184, + 0.0338096059858799, + 0.013432479463517666, + 0.011761567555367947, + 0.011831404641270638, + 0.05234681814908981, + 0.010973488911986351, + 0.014112494885921478, + 0.02261592075228691, + 0.018912415951490402, + 0.008028773590922356, + 0.05033325031399727, + 0.006845503114163876, + 0.004452327731996775, + 0.01570715382695198, + 0.008345318958163261, + 0.008627736940979958, + 0.013820294290781021, + 0.010103552602231503, + 0.009316101670265198, + 0.008779225870966911, + 0.012240500189363956, + 0.00613574031740427, + 0.008616252802312374, + 0.009289029985666275, + 0.014279112219810486, + 0.014376123435795307, + 0.01278819888830185, + 0.020926641300320625, + 0.00800279714167118, + 0.0070501998998224735, + 0.01042870618402958, + 0.015095461159944534, + 0.01761804334819317, + 0.015627039596438408, + 0.013802370056509972, + 0.01193959265947342, + 0.009322772733867168, + 0.03718807175755501, + 0.00833351444453001, + 0.0181202944368124, + 0.006845239549875259, + 0.006712971720844507, + 0.006796722766011953, + 0.006262943148612976, + 0.004154625348746777, + 0.0050813257694244385, + 0.0036349117290228605, + 0.00546417199075222, + 0.005260698962956667, + 0.0068097529001533985, + 0.007389148697257042, + 0.008137688040733337, + 0.00556735647842288, + 0.005032541695982218, + 0.009965525008738041, + 0.006445127539336681, + 0.006443082354962826, + 0.008174940012395382, + 0.008069557137787342, + 0.01033825147897005, + 0.00871382188051939, + 0.005640141200274229, + 0.011436625383794308, + 0.006263133138418198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.01882929354906082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496072769165039 + ], + "opposite_logprob_tokens": [ + -2.9299607276916504 + ], + "logprob_reward": -0.05496072769165039, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496072769165039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3677341043949127, + 0.033618196845054626, + 0.05187347158789635, + 0.13497234880924225, + 0.09063857793807983, + 0.03488842397928238, + 0.03492555022239685, + 0.13740774989128113, + 0.06389134377241135, + 0.05005023628473282 + ], + "credit_map": [ + 0.04025454819202423, + 0.044922471046447754, + 0.03113376908004284, + 0.05678124353289604, + 0.04459625110030174, + 0.06702479720115662, + 0.04185068607330322, + 0.04117034003138542, + 0.004923392552882433, + 0.003972182981669903, + 0.0037574893794953823, + 0.00407931674271822, + 0.004981446545571089, + 0.00398824829608202, + 0.0036793905310332775, + 0.00423673028126359, + 0.011085101403295994, + 0.00815938413143158, + 0.003746816422790289, + 0.0034442762844264507, + 0.0133018484339118, + 0.004861078690737486, + 0.0033605173230171204, + 0.003914446569979191, + 0.024438895285129547, + 0.010262544266879559, + 0.0609152689576149, + 0.007786902599036694, + 0.005171370692551136, + 0.009453007020056248, + 0.010990233160555363, + 0.005954128224402666, + 0.01098770834505558, + 0.011189745739102364, + 0.01288788951933384, + 0.010415432043373585, + 0.013735475018620491, + 0.010905968956649303, + 0.009132576175034046, + 0.011383782140910625, + 0.005057111848145723, + 0.005377993453294039, + 0.004654292948544025, + 0.002806077478453517, + 0.0038667439948767424, + 0.0032809593249112368, + 0.0028687501326203346, + 0.006976497825235128, + 0.0038197843823581934, + 0.004834567196667194, + 0.003753951285034418, + 0.003844072110950947, + 0.003492743941023946, + 0.00924750603735447, + 0.0025585417170077562, + 0.0033743809908628464, + 0.018737157806754112, + 0.023400161415338516, + 0.023107407614588737, + 0.011439954861998558, + 0.015303049236536026, + 0.024092795327305794, + 0.010729599744081497, + 0.010597622022032738, + 0.004947042092680931, + 0.0063344985246658325, + 0.008159119635820389, + 0.010340802371501923, + 0.0073682102374732494, + 0.006664985325187445, + 0.012263580225408077, + 0.007813101634383202, + 0.0069877589121460915, + 0.007015287410467863, + 0.006319764070212841, + 0.006591130513697863, + 0.006472073495388031, + 0.004839967470616102, + 0.007070868741720915, + 0.004753384739160538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.05496072769165039 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252134520560503 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252134520560503, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252134520560503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27733996510505676, + 0.07368602603673935, + 0.09197458624839783, + 0.16194097697734833, + 0.060790445655584335, + 0.03877625986933708, + 0.08026603609323502, + 0.12220580130815506, + 0.030052529647946358, + 0.06296734511852264 + ], + "credit_map": [ + 0.02529798448085785, + 0.032609447836875916, + 0.02242152951657772, + 0.039078932255506516, + 0.03997446224093437, + 0.0564437173306942, + 0.029141059145331383, + 0.03237282857298851, + 0.007998811081051826, + 0.007357587572187185, + 0.005930852144956589, + 0.006547077093273401, + 0.011147200129926205, + 0.011905889958143234, + 0.007273268420249224, + 0.015525338239967823, + 0.027506228536367416, + 0.008232069201767445, + 0.004072375595569611, + 0.005781812593340874, + 0.03109079971909523, + 0.004841290879994631, + 0.004719609394669533, + 0.0057304054498672485, + 0.03040982596576214, + 0.012474693357944489, + 0.07086549699306488, + 0.00968412309885025, + 0.007305529434233904, + 0.01440381444990635, + 0.009825307875871658, + 0.00697218906134367, + 0.008088778704404831, + 0.008753796108067036, + 0.009282109327614307, + 0.008119156584143639, + 0.00803541298955679, + 0.006240384187549353, + 0.005724584683775902, + 0.006546222139149904, + 0.01071010623127222, + 0.008240031078457832, + 0.004400982521474361, + 0.003006967715919018, + 0.004145638085901737, + 0.002838839776813984, + 0.0026118196547031403, + 0.002821873174980283, + 0.011171634308993816, + 0.0077476198785007, + 0.0076414914801716805, + 0.007249139714986086, + 0.005387895740568638, + 0.023976091295480728, + 0.005987416487187147, + 0.011104738339781761, + 0.017590027302503586, + 0.013990717008709908, + 0.02689974568784237, + 0.010760246776044369, + 0.012161966413259506, + 0.020131997764110565, + 0.009513302706182003, + 0.011157797649502754, + 0.006448670290410519, + 0.0043043396435678005, + 0.00428229384124279, + 0.003869849955663085, + 0.002304219175130129, + 0.004667075350880623, + 0.0022940202616155148, + 0.0018820618279278278, + 0.007058348972350359, + 0.010085605084896088, + 0.00713434349745512, + 0.009186708368360996, + 0.007357439957559109, + 0.00674362201243639, + 0.009271095506846905, + 0.006130182649940252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.005252134520560503 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006343498243950307 + ], + "opposite_logprob_tokens": [ + -7.37563419342041 + ], + "logprob_reward": -0.0006343498243950307, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006343498243950307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12232161313295364, + 0.044585999101400375, + 0.23701998591423035, + 0.1752764880657196, + 0.03233903646469116, + 0.07439576834440231, + 0.09362112730741501, + 0.05166328325867653, + 0.049139540642499924, + 0.11963719129562378 + ], + "credit_map": [ + 0.01088131032884121, + 0.014180871658027172, + 0.012608560733497143, + 0.0196247436106205, + 0.014601903967559338, + 0.022950857877731323, + 0.014047399163246155, + 0.013425960205495358, + 0.005192367359995842, + 0.004478249698877335, + 0.005364729091525078, + 0.005782490596175194, + 0.008149193599820137, + 0.004802763927727938, + 0.006311186123639345, + 0.004505023825913668, + 0.07190355658531189, + 0.01843785122036934, + 0.007886405102908611, + 0.010054582729935646, + 0.0883602425456047, + 0.013421377167105675, + 0.00990578904747963, + 0.017050186172127724, + 0.03457432612776756, + 0.013632064685225487, + 0.08004239201545715, + 0.009888451546430588, + 0.006808707024902105, + 0.0074629150331020355, + 0.015684867277741432, + 0.007182756904512644, + 0.004867999814450741, + 0.00437628198415041, + 0.005202597938477993, + 0.0039054343942552805, + 0.004785875324159861, + 0.002131859539076686, + 0.0029045913834124804, + 0.0041643972508609295, + 0.01557657029479742, + 0.012960828840732574, + 0.007745043840259314, + 0.007403355557471514, + 0.009152851998806, + 0.005589245352894068, + 0.0062716370448470116, + 0.009696238674223423, + 0.01227803435176611, + 0.011270214803516865, + 0.009525175206363201, + 0.00838987436145544, + 0.00709390128031373, + 0.027832133695483208, + 0.004889901261776686, + 0.012341895140707493, + 0.006372702773660421, + 0.007784613873809576, + 0.008291810750961304, + 0.004686538130044937, + 0.00421299459412694, + 0.008823278360068798, + 0.006710995454341173, + 0.004780346993356943, + 0.0063604018650949, + 0.008297269232571125, + 0.007231087423861027, + 0.006401007063686848, + 0.004628014285117388, + 0.005105665884912014, + 0.006224079988896847, + 0.004892014432698488, + 0.01715082675218582, + 0.014965696260333061, + 0.01345483586192131, + 0.016916293650865555, + 0.014842129312455654, + 0.011446299962699413, + 0.01951044611632824, + 0.011350665241479874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0006343498243950307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054922252893447876 + ], + "opposite_logprob_tokens": [ + -2.929922342300415 + ], + "logprob_reward": -0.054922252893447876, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054922252893447876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08884874731302261, + 0.13355320692062378, + 0.2403952181339264, + 0.15479162335395813, + 0.041143208742141724, + 0.07695828378200531, + 0.0671052411198616, + 0.046103328466415405, + 0.05539059638977051, + 0.09571058303117752 + ], + "credit_map": [ + 0.0075388383120298386, + 0.009761068038642406, + 0.007206323556602001, + 0.015454260632395744, + 0.01178255770355463, + 0.016247129067778587, + 0.010777618736028671, + 0.010080952197313309, + 0.01471676491200924, + 0.013805354945361614, + 0.01267369743436575, + 0.015892207622528076, + 0.022983402013778687, + 0.01836356893181801, + 0.012769698165357113, + 0.022348519414663315, + 0.07772915065288544, + 0.0189752709120512, + 0.007219188380986452, + 0.009102846495807171, + 0.08810034394264221, + 0.013700167648494244, + 0.010050734505057335, + 0.01551751047372818, + 0.027690501883625984, + 0.011531095951795578, + 0.059268537908792496, + 0.010935130529105663, + 0.008728341199457645, + 0.010774124413728714, + 0.017834795638918877, + 0.008029110729694366, + 0.00486723892390728, + 0.004932478070259094, + 0.005089036654680967, + 0.006242654286324978, + 0.004846907686442137, + 0.004687536507844925, + 0.0048997788690030575, + 0.0055775754153728485, + 0.012812705710530281, + 0.015389705076813698, + 0.008922173641622066, + 0.009551082737743855, + 0.008902688510715961, + 0.005935221444815397, + 0.006198469083756208, + 0.009246237576007843, + 0.0068870787508785725, + 0.00901549682021141, + 0.007441499270498753, + 0.01268824003636837, + 0.005433961283415556, + 0.009307023137807846, + 0.006458146497607231, + 0.009873793460428715, + 0.004370462149381638, + 0.006399576552212238, + 0.00691863801330328, + 0.005350856576114893, + 0.004425481893122196, + 0.005770753137767315, + 0.00548787834122777, + 0.007379684131592512, + 0.00597756914794445, + 0.006464200094342232, + 0.007060416508466005, + 0.0072377691976726055, + 0.006460668984800577, + 0.0073601179756224155, + 0.009216854348778725, + 0.005612996872514486, + 0.009229617193341255, + 0.015958474949002266, + 0.011303532868623734, + 0.012687166221439838, + 0.011956651695072651, + 0.009926949627697468, + 0.015029121190309525, + 0.00961906649172306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.054922252893447876 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle shot captures an iconic coupe, its sleek metallic silver body accented by bold, geometric racing vinyl in a vibrant electric blue, executing a flawless, tight spin on rain-slicked, well-lit black asphalt. The entire scene is rendered in a vibrant urbane art style, defined by sharp outlines that imbue every element with a crisp graphic quality and a highly saturated color palette. As the car spins with controlled violence, its tires explode with dramatic sprays of water that capture the kaleidoscopic glow of holographic advertising, atomizing into a shimmering mist. The camera then begins to smoothly pull back and forth, seamlessly transitioning to a wide-angle shot, revealing a vast, multi-layered futuristic cityscape. Its densely packed buildings and glowing light paths are beautifully reflected in the mirror-like puddles covering the streets below, creating a stunning and vibrant cityscape.", + "seed": 9745, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_14/index_8459_seed_9745_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02658347194083035, + "vlm_logprob": -0.02658347194083035, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.02658347194083035, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04400087520480156 + ], + "opposite_logprob_tokens": [ + -3.2940008640289307 + ], + "logprob_reward": -0.04400087520480156, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04400087520480156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12546616792678833, + 0.04118923097848892, + 0.15132978558540344, + 0.2091597318649292, + 0.22370576858520508, + 0.05513501912355423, + 0.051218900829553604, + 0.028853345662355423, + 0.07576518505811691, + 0.03817690908908844 + ], + "credit_map": [ + 0.005724191199988127, + 0.01012251153588295, + 0.007054852321743965, + 0.007931574247777462, + 0.02580602467060089, + 0.00856154877692461, + 0.023663712665438652, + 0.03660174086689949, + 0.0035792274866253138, + 0.00345498975366354, + 0.005977012682706118, + 0.006960391998291016, + 0.004567133262753487, + 0.006021067034453154, + 0.006574126426130533, + 0.004055279307067394, + 0.02055770345032215, + 0.007144635077565908, + 0.02021409384906292, + 0.01596810109913349, + 0.009106849320232868, + 0.006191643420606852, + 0.05415869876742363, + 0.017988057807087898, + 0.022793369367718697, + 0.0071188947185873985, + 0.09116928279399872, + 0.01940939575433731, + 0.0069196755066514015, + 0.007151077967137098, + 0.025369765236973763, + 0.029228271916508675, + 0.01711403764784336, + 0.02320433408021927, + 0.021503450348973274, + 0.03619490936398506, + 0.04305729642510414, + 0.033834636211395264, + 0.0294941533356905, + 0.01930292882025242, + 0.0022160662338137627, + 0.0033490865025669336, + 0.0024184465873986483, + 0.006111715454608202, + 0.003355303080752492, + 0.00423637917265296, + 0.015958184376358986, + 0.017489835619926453, + 0.005641453433781862, + 0.008302733302116394, + 0.007409214973449707, + 0.006655476056039333, + 0.009172958321869373, + 0.0029177535325288773, + 0.005785040091723204, + 0.005334264133125544, + 0.0025173283647745848, + 0.002687754575163126, + 0.006583128124475479, + 0.004586480557918549, + 0.002060686005279422, + 0.0024239986669272184, + 0.004197752568870783, + 0.003796215634793043, + 0.00536984670907259, + 0.009759511798620224, + 0.01400511059910059, + 0.011725863441824913, + 0.004598372615873814, + 0.005883100442588329, + 0.013760746456682682, + 0.010662637650966644, + 0.0063099986873567104, + 0.008589187636971474, + 0.003722516121342778, + 0.0033597040455788374, + 0.005219422746449709, + 0.0032069990411400795, + 0.00508988369256258, + 0.002679192926734686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.04400087520480156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594218745827675 + ], + "opposite_logprob_tokens": [ + -3.8955941200256348 + ], + "logprob_reward": -0.020594218745827675, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594218745827675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17852255702018738, + 0.030219538137316704, + 0.04346736520528793, + 0.1235695406794548, + 0.02252758853137493, + 0.3375293016433716, + 0.0348932221531868, + 0.08243338763713837, + 0.03624437376856804, + 0.11059313267469406 + ], + "credit_map": [ + 0.006897577550262213, + 0.00790318101644516, + 0.007186243776232004, + 0.011386582627892494, + 0.01850145123898983, + 0.011098814196884632, + 0.05560750141739845, + 0.059941209852695465, + 0.0025006879586726427, + 0.002497241832315922, + 0.004683223087340593, + 0.0037217645440250635, + 0.0038056590128690004, + 0.0031723070424050093, + 0.003819335252046585, + 0.006019320338964462, + 0.005470016971230507, + 0.0030278328340500593, + 0.0053950343281030655, + 0.004287471994757652, + 0.0026657944545149803, + 0.002202572301030159, + 0.01439706888049841, + 0.006021574139595032, + 0.014478144235908985, + 0.005932200234383345, + 0.054295364767313004, + 0.009188566356897354, + 0.0036771115846931934, + 0.004099759738892317, + 0.015637226402759552, + 0.016261160373687744, + 0.002136823022738099, + 0.002257188316434622, + 0.002821473404765129, + 0.0039905281737446785, + 0.0027991044335067272, + 0.0023133461363613605, + 0.0034794239327311516, + 0.0027297004126012325, + 0.008401231840252876, + 0.010381767526268959, + 0.018591376021504402, + 0.04788560792803764, + 0.013568391092121601, + 0.0197162926197052, + 0.1048898994922638, + 0.11409474164247513, + 0.0023673006799072027, + 0.005839049816131592, + 0.005132813006639481, + 0.0064972322434186935, + 0.00474963104352355, + 0.0021266441326588392, + 0.004209008999168873, + 0.003971540834754705, + 0.006325114518404007, + 0.005648695398122072, + 0.010759027674794197, + 0.010436723940074444, + 0.005997965112328529, + 0.012142729014158249, + 0.01228991150856018, + 0.018833227455615997, + 0.002666924148797989, + 0.003273740177974105, + 0.006318829953670502, + 0.006338512059301138, + 0.0024783790577203035, + 0.004493660759180784, + 0.006316543556749821, + 0.004357784986495972, + 0.014928730204701424, + 0.026659131050109863, + 0.011318211443722248, + 0.008826679550111294, + 0.01582399383187294, + 0.009354681707918644, + 0.018265744671225548, + 0.005415959749370813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.020594218745827675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782352484762669 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.009782352484762669, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782352484762669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16185633838176727, + 0.03489436209201813, + 0.16994687914848328, + 0.12861725687980652, + 0.025281470268964767, + 0.08498816937208176, + 0.062370263040065765, + 0.09272254258394241, + 0.055838339030742645, + 0.18348433077335358 + ], + "credit_map": [ + 0.00729962345212698, + 0.010043785907328129, + 0.008395585231482983, + 0.01159114670008421, + 0.0300469771027565, + 0.011520943604409695, + 0.037011049687862396, + 0.04594724252820015, + 0.005801861640065908, + 0.005587473511695862, + 0.0037895096465945244, + 0.004030527547001839, + 0.0033134538680315018, + 0.003262608079239726, + 0.004330133553594351, + 0.004778796341270208, + 0.022651389241218567, + 0.01240862999111414, + 0.021676145493984222, + 0.017294645309448242, + 0.010216182097792625, + 0.007131906226277351, + 0.06076173856854439, + 0.017806246876716614, + 0.014939269982278347, + 0.004683534149080515, + 0.05738702788949013, + 0.010795138776302338, + 0.003913043066859245, + 0.003950614016503096, + 0.01607607863843441, + 0.016872558742761612, + 0.002072657458484173, + 0.0024951796513050795, + 0.0029063576366752386, + 0.004186524078249931, + 0.0023878286592662334, + 0.003830934641882777, + 0.004640160594135523, + 0.002761828014627099, + 0.0036215060390532017, + 0.004670935217291117, + 0.0050046066753566265, + 0.017987575381994247, + 0.004925436805933714, + 0.006102071143686771, + 0.020966388285160065, + 0.02170964889228344, + 0.004699970129877329, + 0.010274393483996391, + 0.008123914711177349, + 0.007801681756973267, + 0.013956720009446144, + 0.003857254981994629, + 0.008055384270846844, + 0.005600943695753813, + 0.00628092885017395, + 0.00748672429472208, + 0.012741191312670708, + 0.013287585228681564, + 0.007531929295510054, + 0.008628062903881073, + 0.02111017517745495, + 0.015655945986509323, + 0.004083025734871626, + 0.006237095221877098, + 0.008560304529964924, + 0.006551331840455532, + 0.003970352001488209, + 0.0050275688990950584, + 0.014158732257783413, + 0.007249928079545498, + 0.018820060417056084, + 0.03238580748438835, + 0.02912610024213791, + 0.01761852391064167, + 0.02424810640513897, + 0.017819058150053024, + 0.026618391275405884, + 0.016848266124725342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.009782352484762669 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615630092099309 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.003615630092099309, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615630092099309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08766867965459824, + 0.02364412695169449, + 0.018952248618006706, + 0.315187007188797, + 0.019843842834234238, + 0.2770514488220215, + 0.08045672625303268, + 0.03128017857670784, + 0.07321692258119583, + 0.07269881665706635 + ], + "credit_map": [ + 0.005776106379926205, + 0.006701288744807243, + 0.004625377710908651, + 0.004957409575581551, + 0.024097779765725136, + 0.006387154571712017, + 0.012659478932619095, + 0.022464081645011902, + 0.002158841583877802, + 0.002632472198456526, + 0.003163394518196583, + 0.004250059369951487, + 0.002003770088776946, + 0.002864970825612545, + 0.003501837607473135, + 0.003068779595196247, + 0.001985123148187995, + 0.0025932721327990294, + 0.0033224455546587706, + 0.0022050361149013042, + 0.0017589052440598607, + 0.0018380176043137908, + 0.002870596246793866, + 0.002378851640969515, + 0.045911479741334915, + 0.015656521543860435, + 0.1312219649553299, + 0.025072352960705757, + 0.010225293226540089, + 0.011041448451578617, + 0.046067267656326294, + 0.02999066561460495, + 0.0016616733046248555, + 0.002205499680712819, + 0.001432829536497593, + 0.003583393758162856, + 0.0036863191053271294, + 0.00192848383449018, + 0.0033309874124825, + 0.0020146542228758335, + 0.00656084343791008, + 0.010135849006474018, + 0.011615075170993805, + 0.03508397191762924, + 0.010337118990719318, + 0.014750983566045761, + 0.08969626575708389, + 0.09887132048606873, + 0.00911303423345089, + 0.012976274825632572, + 0.010859973728656769, + 0.009569438174366951, + 0.014248126186430454, + 0.003856702707707882, + 0.010348674841225147, + 0.009484492242336273, + 0.0028760808054357767, + 0.003922759089618921, + 0.0037276733200997114, + 0.005263132508844137, + 0.0029363506473600864, + 0.003992977552115917, + 0.003997492603957653, + 0.004563706461340189, + 0.005048540886491537, + 0.011948510073125362, + 0.007088116370141506, + 0.008721120655536652, + 0.004159048665314913, + 0.004960270598530769, + 0.014075216837227345, + 0.017216091975569725, + 0.008930083364248276, + 0.016154909506440163, + 0.010241802781820297, + 0.006976154632866383, + 0.008054542355239391, + 0.006171016953885555, + 0.011057560332119465, + 0.005112743005156517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.003615630092099309 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an iconic coupe with a metallic silver body and electric blue racing vinyl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car executing a tight spin on rain-slicked black asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature water sprays from the tires reflecting holographic advertising lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and urbane with sharp outlines and a highly saturated color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pull back and forth to reveal a wide-angle view of a futuristic cityscape with glowing light paths?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05492428317666054 + ], + "opposite_logprob_tokens": [ + -2.929924249649048 + ], + "logprob_reward": -0.05492428317666054, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492428317666054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07201746851205826, + 0.04175786301493645, + 0.17313575744628906, + 0.10897251963615417, + 0.02804328314960003, + 0.16024231910705566, + 0.13894225656986237, + 0.06763360649347305, + 0.06917425245046616, + 0.14008063077926636 + ], + "credit_map": [ + 0.0047280676662921906, + 0.004398386459797621, + 0.003793545300140977, + 0.004294789861887693, + 0.012010972946882248, + 0.006244106218218803, + 0.015465893782675266, + 0.021081704646348953, + 0.0030796309001743793, + 0.0032132198102772236, + 0.008228294551372528, + 0.007744738832116127, + 0.004069728311151266, + 0.004530778154730797, + 0.006027996074408293, + 0.004863476846367121, + 0.02585522271692753, + 0.011660643853247166, + 0.021714508533477783, + 0.017602164298295975, + 0.01026364229619503, + 0.0066251144744455814, + 0.06038658693432808, + 0.01902787573635578, + 0.012216866947710514, + 0.005553515627980232, + 0.03236055746674538, + 0.007382543757557869, + 0.005058359354734421, + 0.005827940069139004, + 0.01248680055141449, + 0.028085939586162567, + 0.0027870002668350935, + 0.002765071112662554, + 0.0035533097106963396, + 0.003209134330973029, + 0.0028871875256299973, + 0.00475207669660449, + 0.005095166154205799, + 0.00299433758482337, + 0.005501764826476574, + 0.007368927355855703, + 0.007683814503252506, + 0.012447661720216274, + 0.005707824137061834, + 0.006542193703353405, + 0.0554075725376606, + 0.05958255007863045, + 0.012881317175924778, + 0.016563182696700096, + 0.017180373892188072, + 0.018304985016584396, + 0.027084955945611, + 0.012785892002284527, + 0.01655760034918785, + 0.017583956941962242, + 0.005416057538241148, + 0.009879578836262226, + 0.008578918874263763, + 0.009361589327454567, + 0.00650785630568862, + 0.0069629414938390255, + 0.012813868001103401, + 0.008112797513604164, + 0.005297558847814798, + 0.006911167874932289, + 0.011189043521881104, + 0.011111469939351082, + 0.0052708713337779045, + 0.01025562733411789, + 0.010810505598783493, + 0.008328007534146309, + 0.01870095543563366, + 0.0175112783908844, + 0.021634161472320557, + 0.017516769468784332, + 0.012911795638501644, + 0.013018404133617878, + 0.024501444771885872, + 0.014285819604992867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.05492428317666054 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_15/index_tensor([7559, 7559])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_15/index_tensor([7559, 7559])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a8291fdcdc1f041486be30c77b701ee6df960ad2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_15/index_tensor([7559, 7559])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion close-up focuses on a medical certificate, printed in black ink, gripped tightly by a pair of bony hands, their knuckles turning slightly white from the pressure. With a resounding, drawn-out, crisp tearing sound, the paper is decisively and violently split in two, then torn again, instantly shattering into countless fragments of varying sizes. The hands then release, and the camera tilts slightly upward, capturing the pieces of paper drifting through the air as if weightless. The warm, slanting afternoon sun streams through the window, creating a distinct Tyndall beam in the air, shimmering with tiny particles of dust. As each swirling, tumbling scrap of paper passes through the light, its jagged white edges momentarily illuminated, shimmering with a fleeting, soft glow. The background is a completely out-of-focus, soft-toned corner of the room. All visual focus is on these fragments fluttering like butterflies, falling in an elegant and slow posture. The whole picture is full of strong symbolic meaning, conveying a heavy yet relieved tranquil atmosphere in an almost silent dynamic.", + "seed": 13215, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_15/index_7559_seed_13215_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.39274647682905195, + "vlm_logprob": -0.39274647682905195, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.39274647682905195, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.805687665939331 + ], + "opposite_logprob_tokens": [ + -0.18068769574165344 + ], + "logprob_reward": -1.805687665939331, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.805687665939331, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1865784078836441, + 0.11285125464200974, + 0.04594526067376137, + 0.04460412263870239, + 0.14034290611743927, + 0.12822888791561127, + 0.10299819707870483, + 0.1460198312997818, + 0.06897132843732834, + 0.02345980890095234 + ], + "credit_map": [ + 0.02895127236843109, + 0.012182353995740414, + 0.030575450509786606, + 0.02014060504734516, + 0.01712612994015217, + 0.045357491821050644, + 0.014226609840989113, + 0.018018489703536034, + 0.02237909846007824, + 0.012755694799125195, + 0.017152080312371254, + 0.010138793848454952, + 0.015292408876121044, + 0.012626932002604008, + 0.01138249970972538, + 0.011123739182949066, + 0.006641042418777943, + 0.007199276704341173, + 0.0062043252401053905, + 0.006120431702584028, + 0.0040852162055671215, + 0.00504404678940773, + 0.004633493721485138, + 0.006017425563186407, + 0.00433265371248126, + 0.003966934513300657, + 0.006634729448705912, + 0.00973503664135933, + 0.002840470289811492, + 0.004759244155138731, + 0.0067097898572683334, + 0.005625262390822172, + 0.03310861811041832, + 0.01655811071395874, + 0.017217721790075302, + 0.01978854089975357, + 0.015117267146706581, + 0.017935048788785934, + 0.010247912257909775, + 0.010369689203798771, + 0.02327142469584942, + 0.02367427572607994, + 0.020098593086004257, + 0.014956281520426273, + 0.01742822490632534, + 0.012639747001230717, + 0.007094086613506079, + 0.009066246449947357, + 0.009367804042994976, + 0.01795361377298832, + 0.015078260563313961, + 0.009846581146121025, + 0.015280211344361305, + 0.015293573960661888, + 0.006649825721979141, + 0.013528326526284218, + 0.007788950577378273, + 0.012702546082437038, + 0.018681256100535393, + 0.026946233585476875, + 0.011690743267536163, + 0.01247165072709322, + 0.04283519461750984, + 0.0129032451659441, + 0.0033117285929620266, + 0.006562790367752314, + 0.007935424335300922, + 0.016678130254149437, + 0.006809397600591183, + 0.009129644371569157, + 0.01376915443688631, + 0.0047750468365848064, + 0.0024233649019151926, + 0.003191764932125807, + 0.0036457690875977278, + 0.0035504039842635393, + 0.00260135973803699, + 0.0020265961065888405, + 0.0031701643019914627, + 0.0028503823559731245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -1.805687665939331 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036412626504898 + ], + "opposite_logprob_tokens": [ + -2.3503642082214355 + ], + "logprob_reward": -0.10036412626504898, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036412626504898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20144346356391907, + 0.14187772572040558, + 0.03968179225921631, + 0.07609061151742935, + 0.05604225769639015, + 0.11925913393497467, + 0.10985250025987625, + 0.07866855710744858, + 0.13206364214420319, + 0.04502027481794357 + ], + "credit_map": [ + 0.019183900207281113, + 0.011163068003952503, + 0.03361980989575386, + 0.026194561272859573, + 0.014589488506317139, + 0.042980924248695374, + 0.024729344993829727, + 0.02898237481713295, + 0.02295699715614319, + 0.01324387826025486, + 0.021426528692245483, + 0.013292376883327961, + 0.01818375661969185, + 0.01724182814359665, + 0.015147233381867409, + 0.020385127514600754, + 0.007371753454208374, + 0.006694344338029623, + 0.0048163337633013725, + 0.0043753329664468765, + 0.002815345535054803, + 0.005579605232924223, + 0.0030847396701574326, + 0.004944339394569397, + 0.008705933578312397, + 0.006119146943092346, + 0.01042622048407793, + 0.013804519549012184, + 0.006231200881302357, + 0.010954286903142929, + 0.009662160649895668, + 0.010187141597270966, + 0.011353793554008007, + 0.005927038379013538, + 0.006449089385569096, + 0.006649353541433811, + 0.007334479596465826, + 0.0067114331759512424, + 0.005861111916601658, + 0.0057559534907341, + 0.01944643259048462, + 0.021821364760398865, + 0.015570552088320255, + 0.014983409084379673, + 0.018224410712718964, + 0.01305528823286295, + 0.007249349262565374, + 0.008908325806260109, + 0.009575669653713703, + 0.01844106987118721, + 0.01367778517305851, + 0.016384519636631012, + 0.013613458722829819, + 0.01464556809514761, + 0.00936872512102127, + 0.014145700260996819, + 0.005695651285350323, + 0.010015293024480343, + 0.012074485421180725, + 0.013615168631076813, + 0.0058232443407177925, + 0.006364079657942057, + 0.016721315681934357, + 0.008359326049685478, + 0.00859532505273819, + 0.018258417025208473, + 0.012269199825823307, + 0.022811951115727425, + 0.014650104567408562, + 0.015623897314071655, + 0.028299106284976006, + 0.011555643752217293, + 0.003347743535414338, + 0.004569461103528738, + 0.004326795693486929, + 0.005423401016741991, + 0.006739406380802393, + 0.007459597662091255, + 0.007885362952947617, + 0.005268508102744818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.10036412626504898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203879982233047 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.014203879982233047, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203879982233047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10173510760068893, + 0.16231928765773773, + 0.05179033428430557, + 0.05560548976063728, + 0.1532244086265564, + 0.08659777045249939, + 0.10677909851074219, + 0.14035558700561523, + 0.09307745099067688, + 0.04851549491286278 + ], + "credit_map": [ + 0.014808816835284233, + 0.008110281080007553, + 0.017113160341978073, + 0.010883592069149017, + 0.00901542417705059, + 0.016921035945415497, + 0.012989841401576996, + 0.011892947368323803, + 0.03301526606082916, + 0.015522489324212074, + 0.024694468826055527, + 0.014288771897554398, + 0.021045774221420288, + 0.019118070602416992, + 0.013331783935427666, + 0.02130265347659588, + 0.007761974353343248, + 0.009353778325021267, + 0.006571273785084486, + 0.008040269836783409, + 0.004501692019402981, + 0.00490591861307621, + 0.004644406959414482, + 0.006011022720485926, + 0.007066937163472176, + 0.005499609280377626, + 0.009399669244885445, + 0.010221283882856369, + 0.004258934408426285, + 0.004489443264901638, + 0.007958867587149143, + 0.006710744462907314, + 0.027181940153241158, + 0.015416387468576431, + 0.01689808815717697, + 0.02063656598329544, + 0.023887639865279198, + 0.019211789593100548, + 0.013982669450342655, + 0.016009310260415077, + 0.011536515317857265, + 0.009963403455913067, + 0.011180231347680092, + 0.0060949381440877914, + 0.017559541389346123, + 0.011587577871978283, + 0.005943035241216421, + 0.012732522562146187, + 0.008406782522797585, + 0.018121905624866486, + 0.01249952893704176, + 0.015730272978544235, + 0.01641635037958622, + 0.013984853401780128, + 0.00867948867380619, + 0.012939905747771263, + 0.0070921084843575954, + 0.01434299349784851, + 0.017659829929471016, + 0.02208910509943962, + 0.012573390267789364, + 0.012754755094647408, + 0.04091227427124977, + 0.012931126169860363, + 0.004325232934206724, + 0.00945355650037527, + 0.011532058008015156, + 0.02146797813475132, + 0.009307380765676498, + 0.01013402733951807, + 0.01815708726644516, + 0.008700125850737095, + 0.0034359970595687628, + 0.006419765297323465, + 0.007116949185729027, + 0.005406905896961689, + 0.006477023009210825, + 0.0051469579339027405, + 0.008898887783288956, + 0.0056130061857402325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.014203879982233047 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785067290067673 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.009785067290067673, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785067290067673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08617483824491501, + 0.09525029361248016, + 0.039569202810525894, + 0.16755104064941406, + 0.06294254213571548, + 0.2055189311504364, + 0.09925777465105057, + 0.1682530641555786, + 0.052636295557022095, + 0.022846007719635963 + ], + "credit_map": [ + 0.014218509197235107, + 0.00600028270855546, + 0.014002491720020771, + 0.01186784915626049, + 0.007043738383799791, + 0.017011575400829315, + 0.007492052856832743, + 0.008538343943655491, + 0.017272023484110832, + 0.010454704985022545, + 0.013809256255626678, + 0.008273991756141186, + 0.011112532578408718, + 0.01429577823728323, + 0.010836798697710037, + 0.009195207618176937, + 0.00782586820423603, + 0.004800373688340187, + 0.005724241025745869, + 0.004675772041082382, + 0.0027166029904037714, + 0.004104139748960733, + 0.004016389604657888, + 0.005705817136913538, + 0.020542297512292862, + 0.01798568107187748, + 0.02445203624665737, + 0.03233456611633301, + 0.01510990783572197, + 0.01214329432696104, + 0.022658007219433784, + 0.02232525125145912, + 0.015355399809777737, + 0.006909819785505533, + 0.00540615851059556, + 0.004120639991015196, + 0.00791396014392376, + 0.007641466334462166, + 0.00753118796274066, + 0.00806390680372715, + 0.03739257529377937, + 0.03264781832695007, + 0.03132334724068642, + 0.016654931008815765, + 0.031364936381578445, + 0.021436089649796486, + 0.01770678721368313, + 0.016992438584566116, + 0.010307208634912968, + 0.015743225812911987, + 0.012701880186796188, + 0.016835195943713188, + 0.014281794428825378, + 0.011322774924337864, + 0.006941689644008875, + 0.011124003678560257, + 0.010674352757632732, + 0.018310274928808212, + 0.030004601925611496, + 0.0450897291302681, + 0.01680869795382023, + 0.014491796493530273, + 0.022031724452972412, + 0.010841880924999714, + 0.0035792498383671045, + 0.006358956452459097, + 0.006544544827193022, + 0.010049786418676376, + 0.004810024052858353, + 0.006336340215057135, + 0.009440083056688309, + 0.005517312325537205, + 0.001787685789167881, + 0.002664310624822974, + 0.003359259804710746, + 0.0032794885337352753, + 0.0034385232720524073, + 0.0026230146177113056, + 0.002890184987336397, + 0.0028035398572683334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.009785067290067673 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336916446685791 + ], + "opposite_logprob_tokens": [ + -3.408691644668579 + ], + "logprob_reward": -0.0336916446685791, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336916446685791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10327182710170746, + 0.13958723843097687, + 0.04822653532028198, + 0.04471983015537262, + 0.17833548784255981, + 0.1114620715379715, + 0.10976072400808334, + 0.11864777654409409, + 0.09723661094903946, + 0.04875198006629944 + ], + "credit_map": [ + 0.014532739296555519, + 0.007437169551849365, + 0.01773039437830448, + 0.01045147329568863, + 0.009130217134952545, + 0.019436299800872803, + 0.010920759290456772, + 0.013632775284349918, + 0.021925192326307297, + 0.013096295297145844, + 0.021397680044174194, + 0.01265100110322237, + 0.018208548426628113, + 0.017371229827404022, + 0.016606556251645088, + 0.018330717459321022, + 0.006886743940412998, + 0.005666050594300032, + 0.006707944441586733, + 0.005753128323704004, + 0.003873970825225115, + 0.005621851421892643, + 0.005599969066679478, + 0.008116869255900383, + 0.005001698154956102, + 0.004787991289049387, + 0.011053048074245453, + 0.006493820808827877, + 0.002945270389318466, + 0.002748551545664668, + 0.006209731567651033, + 0.005479713901877403, + 0.037771061062812805, + 0.021878091618418694, + 0.02274899370968342, + 0.026350731030106544, + 0.021668002009391785, + 0.024017436429858208, + 0.012282103300094604, + 0.011619067750871181, + 0.02195746824145317, + 0.02280961535871029, + 0.01726851426064968, + 0.00980373751372099, + 0.014196190983057022, + 0.010319472290575504, + 0.0065488494001328945, + 0.008558214642107487, + 0.009144233539700508, + 0.015429632738232613, + 0.012443136423826218, + 0.015350949950516224, + 0.0162863340228796, + 0.015737051144242287, + 0.009616048075258732, + 0.015753332525491714, + 0.006646881345659494, + 0.011692085303366184, + 0.014227136969566345, + 0.02352049946784973, + 0.009778453037142754, + 0.01030035875737667, + 0.0319422222673893, + 0.010540120303630829, + 0.006692233961075544, + 0.010501476936042309, + 0.010510136373341084, + 0.022382114082574844, + 0.008883119560778141, + 0.010739323683083057, + 0.019614892080426216, + 0.007913300767540932, + 0.004233998246490955, + 0.006413976661860943, + 0.004874415695667267, + 0.007006716914474964, + 0.008430084213614464, + 0.007962042465806007, + 0.004430786240845919, + 0.005399955436587334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.0336916446685791 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion close-up focuses on a medical certificate, printed in black ink, gripped tightly by a pair of bony hands, their knuckles turning slightly white from the pressure. With a resounding, drawn-out, crisp tearing sound, the paper is decisively and violently split in two, then torn again, instantly shattering into countless fragments of varying sizes. The hands then release, and the camera tilts slightly upward, capturing the pieces of paper drifting through the air as if weightless. The warm, slanting afternoon sun streams through the window, creating a distinct Tyndall beam in the air, shimmering with tiny particles of dust. As each swirling, tumbling scrap of paper passes through the light, its jagged white edges momentarily illuminated, shimmering with a fleeting, soft glow. The background is a completely out-of-focus, soft-toned corner of the room. All visual focus is on these fragments fluttering like butterflies, falling in an elegant and slow posture. The whole picture is full of strong symbolic meaning, conveying a heavy yet relieved tranquil atmosphere in an almost silent dynamic.", + "seed": 13216, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_15/index_7559_seed_13216_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4878230206668377, + "vlm_logprob": -0.4878230206668377, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.4878230206668377, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.23868989944458 + ], + "opposite_logprob_tokens": [ + -0.11368986964225769 + ], + "logprob_reward": -2.23868989944458, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.23868989944458, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.25544723868370056, + 0.13243994116783142, + 0.13401801884174347, + 0.04159621149301529, + 0.06386343389749527, + 0.14436779916286469, + 0.030758067965507507, + 0.04260968789458275, + 0.10389521718025208, + 0.051004357635974884 + ], + "credit_map": [ + 0.014742097817361355, + 0.02647409960627556, + 0.029896778985857964, + 0.06786897033452988, + 0.016156600788235664, + 0.016920819878578186, + 0.04253160208463669, + 0.04085627943277359, + 0.013116824440658092, + 0.01791125163435936, + 0.014948045834898949, + 0.016992485150694847, + 0.02299526147544384, + 0.013674644753336906, + 0.017238376662135124, + 0.01556304655969143, + 0.01830081269145012, + 0.016014931723475456, + 0.018067829310894012, + 0.017713399603962898, + 0.015542564913630486, + 0.010231789201498032, + 0.017292210832238197, + 0.020854486152529716, + 0.005963616538792849, + 0.006013205740600824, + 0.006427839398384094, + 0.0047289179638028145, + 0.005293956492096186, + 0.004472117871046066, + 0.005175156984478235, + 0.0035213984083384275, + 0.009174155071377754, + 0.007536592893302441, + 0.008050594478845596, + 0.010747505351901054, + 0.007583294063806534, + 0.007961998693645, + 0.007922771386802197, + 0.004886525683104992, + 0.009504854679107666, + 0.010619405657052994, + 0.015605844557285309, + 0.0364244282245636, + 0.012799105606973171, + 0.01609966717660427, + 0.027366412803530693, + 0.015948083251714706, + 0.0023383810184895992, + 0.0015180815244093537, + 0.009557601064443588, + 0.0034943593200296164, + 0.002942170947790146, + 0.002781372983008623, + 0.004049510229378939, + 0.0040765912272036076, + 0.003480554325506091, + 0.005821912083774805, + 0.007395235355943441, + 0.005179314408451319, + 0.007224126718938351, + 0.004899176768958569, + 0.005146341398358345, + 0.0034630256704986095, + 0.006449687760323286, + 0.009343454614281654, + 0.009287985041737556, + 0.011894123628735542, + 0.01342377346009016, + 0.0297007467597723, + 0.011568065732717514, + 0.012227382510900497, + 0.0032016970217227936, + 0.01077504176646471, + 0.0036213723942637444, + 0.008145682513713837, + 0.008131565526127815, + 0.006498352158814669, + 0.007487154565751553, + 0.003143494948744774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -2.23868989944458 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07905884087085724 + ], + "opposite_logprob_tokens": [ + -2.579058885574341 + ], + "logprob_reward": -0.07905884087085724, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905884087085724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13734164834022522, + 0.23078088462352753, + 0.09310609102249146, + 0.04073804244399071, + 0.07894313335418701, + 0.1374758630990982, + 0.07080430537462234, + 0.059836722910404205, + 0.11216504126787186, + 0.038808248937129974 + ], + "credit_map": [ + 0.011777454987168312, + 0.01417582668364048, + 0.014605770818889141, + 0.02572592906653881, + 0.01620878092944622, + 0.010934957303106785, + 0.021501965820789337, + 0.022410959005355835, + 0.021795017644762993, + 0.029715824872255325, + 0.024355288594961166, + 0.029351456090807915, + 0.04240988940000534, + 0.027180271223187447, + 0.029005954042077065, + 0.026967182755470276, + 0.01237414125353098, + 0.010303471237421036, + 0.012735983356833458, + 0.010028127580881119, + 0.012016613967716694, + 0.008587026968598366, + 0.013531012460589409, + 0.013529717922210693, + 0.005380458664149046, + 0.006304978858679533, + 0.005598860327154398, + 0.004931697156280279, + 0.0057051475159823895, + 0.004641876555979252, + 0.0050302851013839245, + 0.0031447403598576784, + 0.0072048562578856945, + 0.009130378253757954, + 0.011839150451123714, + 0.01075945608317852, + 0.009679259732365608, + 0.01295285951346159, + 0.010864635929465294, + 0.0065125408582389355, + 0.008689679205417633, + 0.00953291542828083, + 0.014499270357191563, + 0.03543684631586075, + 0.012155751697719097, + 0.014801010489463806, + 0.027730152010917664, + 0.014630247838795185, + 0.00667427945882082, + 0.004110966343432665, + 0.02101607248187065, + 0.0073255086317658424, + 0.006224568001925945, + 0.005674609448760748, + 0.009878814220428467, + 0.009899485856294632, + 0.004933718591928482, + 0.0068153575994074345, + 0.01027693971991539, + 0.007012686226516962, + 0.010849467478692532, + 0.0065703545697033405, + 0.008758251555263996, + 0.004619942512363195, + 0.0060939970426261425, + 0.010928722098469734, + 0.010321319103240967, + 0.011728472076356411, + 0.015361055731773376, + 0.03203850984573364, + 0.012167192064225674, + 0.013525768183171749, + 0.0023561709094792604, + 0.004233669489622116, + 0.003487430978566408, + 0.0074788648635149, + 0.006667424459010363, + 0.005924312397837639, + 0.005494299810379744, + 0.0031660767272114754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.07905884087085724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901345938444138 + ], + "opposite_logprob_tokens": [ + -2.5790133476257324 + ], + "logprob_reward": -0.07901345938444138, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901345938444138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08812568336725235, + 0.0901588425040245, + 0.08132128417491913, + 0.09654978662729263, + 0.1756937950849533, + 0.10042425245046616, + 0.08834275603294373, + 0.05700930207967758, + 0.08344326913356781, + 0.138931006193161 + ], + "credit_map": [ + 0.009886812418699265, + 0.015780974179506302, + 0.010553155094385147, + 0.015535620041191578, + 0.007957861758768559, + 0.007028372026979923, + 0.012155836448073387, + 0.009227056056261063, + 0.007696095388382673, + 0.010149877518415451, + 0.012342696078121662, + 0.015972193330526352, + 0.011777251027524471, + 0.006933576427400112, + 0.013675832189619541, + 0.01161132100969553, + 0.012988079339265823, + 0.010383541695773602, + 0.010748934000730515, + 0.010273853316903114, + 0.009572713635861874, + 0.006783193908631802, + 0.010094905272126198, + 0.010476065799593925, + 0.011869886890053749, + 0.012864707969129086, + 0.019370660185813904, + 0.014561304822564125, + 0.009721986949443817, + 0.011319017969071865, + 0.00932349544018507, + 0.007518719416111708, + 0.013575389981269836, + 0.016674676910042763, + 0.0265883207321167, + 0.023183627054095268, + 0.01959431730210781, + 0.02837342396378517, + 0.03539362922310829, + 0.012310417369008064, + 0.0076914564706385136, + 0.009716084226965904, + 0.012401162646710873, + 0.021742690354585648, + 0.010190760716795921, + 0.009991386905312538, + 0.01602957583963871, + 0.012661140412092209, + 0.009570064954459667, + 0.0097624771296978, + 0.015138959512114525, + 0.013609739951789379, + 0.011351111344993114, + 0.007317779585719109, + 0.01129291020333767, + 0.010299715213477612, + 0.005985029973089695, + 0.008137035183608532, + 0.007985620759427547, + 0.009831043891608715, + 0.006369336973875761, + 0.004176340065896511, + 0.009713695384562016, + 0.004811202175915241, + 0.006543578580021858, + 0.008069109171628952, + 0.009061862714588642, + 0.014342325739562511, + 0.010774662718176842, + 0.017255011945962906, + 0.00881992094218731, + 0.008576801978051662, + 0.009173862636089325, + 0.027963727712631226, + 0.010232143104076385, + 0.0192299485206604, + 0.023809021338820457, + 0.021838732063770294, + 0.01537999790161848, + 0.011303575709462166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.07901345938444138 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649386465549469 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649386465549469, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649386465549469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15821473300457, + 0.08899249136447906, + 0.1194450631737709, + 0.10193322598934174, + 0.13220176100730896, + 0.13329049944877625, + 0.04391830042004585, + 0.044207748025655746, + 0.12310810387134552, + 0.05468810722231865 + ], + "credit_map": [ + 0.009215056896209717, + 0.015740200877189636, + 0.01870228536427021, + 0.04217033088207245, + 0.009258165024220943, + 0.01076442189514637, + 0.0273617934435606, + 0.025002464652061462, + 0.007480405271053314, + 0.009419519454240799, + 0.011626777239143848, + 0.015406259335577488, + 0.012639578431844711, + 0.007799353450536728, + 0.016271792352199554, + 0.008348798379302025, + 0.018653983250260353, + 0.013681993819773197, + 0.017519857734441757, + 0.015217596665024757, + 0.012054061517119408, + 0.009105019271373749, + 0.0163374375551939, + 0.01687510870397091, + 0.013769559562206268, + 0.01476011797785759, + 0.015271850861608982, + 0.014579933136701584, + 0.013784228824079037, + 0.011392990127205849, + 0.010044845752418041, + 0.008329693228006363, + 0.017679596319794655, + 0.01569192297756672, + 0.025241881608963013, + 0.01189159881323576, + 0.013078568503260612, + 0.0261402428150177, + 0.015910331159830093, + 0.006567603908479214, + 0.009577782824635506, + 0.009698250330984592, + 0.022338686510920525, + 0.026925580576062202, + 0.012956894002854824, + 0.016181230545043945, + 0.0235588476061821, + 0.012053228914737701, + 0.00367181608453393, + 0.0030237429309636354, + 0.011207526549696922, + 0.0054175700061023235, + 0.00490599125623703, + 0.0041222176514565945, + 0.006326969247311354, + 0.0052424645982682705, + 0.004362165462225676, + 0.004637280013412237, + 0.008029310032725334, + 0.00514239864423871, + 0.007318698335438967, + 0.004863400477916002, + 0.005785627756267786, + 0.004068867769092321, + 0.008165736682713032, + 0.011910416185855865, + 0.011300460435450077, + 0.014531251043081284, + 0.015636788681149483, + 0.034073878079652786, + 0.012881920672953129, + 0.014607646502554417, + 0.0031531204003840685, + 0.007795851677656174, + 0.0045150876976549625, + 0.007236779201775789, + 0.013710982166230679, + 0.009900101460516453, + 0.004526270553469658, + 0.003849911503493786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.008649386465549469 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337035171687603 + ], + "opposite_logprob_tokens": [ + -3.408703565597534 + ], + "logprob_reward": -0.0337035171687603, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337035171687603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08207066357135773, + 0.11181038618087769, + 0.0755217894911766, + 0.13572856783866882, + 0.12421321868896484, + 0.11614975333213806, + 0.09057998657226562, + 0.05783970281481743, + 0.14562880992889404, + 0.06045712158083916 + ], + "credit_map": [ + 0.008354578167200089, + 0.01578906737267971, + 0.008501472882926464, + 0.017414938658475876, + 0.005588390864431858, + 0.005897561088204384, + 0.011720695532858372, + 0.008803964592516422, + 0.009836745448410511, + 0.010905413888394833, + 0.016969110816717148, + 0.023714154958724976, + 0.011599473655223846, + 0.0091093173250556, + 0.01849418133497238, + 0.011181989684700966, + 0.009901996701955795, + 0.015863852575421333, + 0.01116261351853609, + 0.009663397446274757, + 0.008556005544960499, + 0.005736593622714281, + 0.007854558527469635, + 0.006782767828553915, + 0.008363829925656319, + 0.011554297991096973, + 0.028493735939264297, + 0.024694904685020447, + 0.014161155559122562, + 0.014569825492799282, + 0.023125432431697845, + 0.01076539047062397, + 0.009786296635866165, + 0.012886703945696354, + 0.01941663585603237, + 0.017932958900928497, + 0.014113251119852066, + 0.019653335213661194, + 0.020775796845555305, + 0.009648232720792294, + 0.007516840007156134, + 0.00826733186841011, + 0.01635696552693844, + 0.025570377707481384, + 0.01086064986884594, + 0.014193404465913773, + 0.02167258970439434, + 0.01171159278601408, + 0.007691617123782635, + 0.007719388231635094, + 0.025085538625717163, + 0.010839257389307022, + 0.010648653842508793, + 0.009971141815185547, + 0.011395265348255634, + 0.007229122798889875, + 0.004353884141892195, + 0.009656859561800957, + 0.008942252025008202, + 0.009316250681877136, + 0.008059559389948845, + 0.005248975474387407, + 0.00829083751887083, + 0.003971083089709282, + 0.0073181865736842155, + 0.014009177684783936, + 0.013427235186100006, + 0.015067089349031448, + 0.01997191086411476, + 0.04139995574951172, + 0.01668418198823929, + 0.017751086503267288, + 0.0038751482497900724, + 0.013120001181960106, + 0.004004128742963076, + 0.007917162030935287, + 0.009254395961761475, + 0.009850088506937027, + 0.007342006545513868, + 0.005094188265502453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.0337035171687603 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion close-up focuses on a medical certificate, printed in black ink, gripped tightly by a pair of bony hands, their knuckles turning slightly white from the pressure. With a resounding, drawn-out, crisp tearing sound, the paper is decisively and violently split in two, then torn again, instantly shattering into countless fragments of varying sizes. The hands then release, and the camera tilts slightly upward, capturing the pieces of paper drifting through the air as if weightless. The warm, slanting afternoon sun streams through the window, creating a distinct Tyndall beam in the air, shimmering with tiny particles of dust. As each swirling, tumbling scrap of paper passes through the light, its jagged white edges momentarily illuminated, shimmering with a fleeting, soft glow. The background is a completely out-of-focus, soft-toned corner of the room. All visual focus is on these fragments fluttering like butterflies, falling in an elegant and slow posture. The whole picture is full of strong symbolic meaning, conveying a heavy yet relieved tranquil atmosphere in an almost silent dynamic.", + "seed": 13217, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_15/index_7559_seed_13217_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.49068589247763156, + "vlm_logprob": -0.49068589247763156, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.49068589247763156, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.238658905029297 + ], + "opposite_logprob_tokens": [ + -0.11365880072116852 + ], + "logprob_reward": -2.238658905029297, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.238658905029297, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2508033514022827, + 0.03979500010609627, + 0.06882596015930176, + 0.04605802893638611, + 0.05832685902714729, + 0.11117877066135406, + 0.08264480531215668, + 0.08647365868091583, + 0.2085600197315216, + 0.04733350872993469 + ], + "credit_map": [ + 0.03491358086466789, + 0.017523379996418953, + 0.03931457921862602, + 0.02919944003224373, + 0.025866057723760605, + 0.03323135897517204, + 0.04193241149187088, + 0.02882254123687744, + 0.004774915985763073, + 0.005120528861880302, + 0.005607749801129103, + 0.004861314315348864, + 0.0046304394491016865, + 0.0028649105224758387, + 0.007219468243420124, + 0.004715670365840197, + 0.00795595720410347, + 0.006454548332840204, + 0.00838062446564436, + 0.010252625681459904, + 0.005387213546782732, + 0.009180651977658272, + 0.01017747726291418, + 0.011036857031285763, + 0.0038260933943092823, + 0.006561539135873318, + 0.007643645629286766, + 0.007704524323344231, + 0.005263295490294695, + 0.004573975224047899, + 0.006023781839758158, + 0.004461172502487898, + 0.00407332694157958, + 0.00659920321777463, + 0.008750388398766518, + 0.010045667178928852, + 0.005253538955003023, + 0.0073842997662723064, + 0.008928854018449783, + 0.007291576359421015, + 0.010174286551773548, + 0.016179196536540985, + 0.016832850873470306, + 0.021987129002809525, + 0.015238755382597446, + 0.014234478585422039, + 0.010100194253027439, + 0.006431886926293373, + 0.008029269985854626, + 0.011935992166399956, + 0.01074079517275095, + 0.011956095695495605, + 0.010351656936109066, + 0.006420184392482042, + 0.0126953748986125, + 0.010515435598790646, + 0.0073117283172905445, + 0.011091827414929867, + 0.009180384688079357, + 0.014131090603768826, + 0.00979780126363039, + 0.009005585685372353, + 0.011210652999579906, + 0.014744583517313004, + 0.01770879328250885, + 0.010043062269687653, + 0.07883283495903015, + 0.03867761790752411, + 0.012166919186711311, + 0.011260145343840122, + 0.02468828856945038, + 0.015182359144091606, + 0.004356334917247295, + 0.00869489274919033, + 0.00705002574250102, + 0.005961001385003328, + 0.008425241336226463, + 0.004641074221581221, + 0.004429791588336229, + 0.003775145160034299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -2.238658905029297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07011262327432632 + ], + "opposite_logprob_tokens": [ + -2.695112705230713 + ], + "logprob_reward": -0.07011262327432632, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07011262327432632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26695865392684937, + 0.051860082894563675, + 0.08460903912782669, + 0.04702151194214821, + 0.06230633333325386, + 0.14122022688388824, + 0.08922910690307617, + 0.0789792463183403, + 0.13589701056480408, + 0.04191874340176582 + ], + "credit_map": [ + 0.04222990572452545, + 0.023030854761600494, + 0.034964460879564285, + 0.02846454083919525, + 0.030670644715428352, + 0.035675227642059326, + 0.04085275158286095, + 0.0310702845454216, + 0.006514861714094877, + 0.0070906211622059345, + 0.006901358719915152, + 0.0070466590113937855, + 0.006283005233854055, + 0.004414270166307688, + 0.008672843687236309, + 0.004936466924846172, + 0.009484530426561832, + 0.008526192046701908, + 0.0101276570931077, + 0.01322128064930439, + 0.007113059051334858, + 0.010909594595432281, + 0.011942325159907341, + 0.013284399174153805, + 0.0038380585610866547, + 0.009455290623009205, + 0.01072046346962452, + 0.006139401812106371, + 0.004324681591242552, + 0.004272783640772104, + 0.00471307709813118, + 0.003557757707312703, + 0.0045734914019703865, + 0.006284201517701149, + 0.00833666417747736, + 0.009930023923516273, + 0.006304135080426931, + 0.009100010618567467, + 0.008633419871330261, + 0.009144393727183342, + 0.011723638512194157, + 0.02057274617254734, + 0.022722706198692322, + 0.030309058725833893, + 0.019417958334088326, + 0.016820333898067474, + 0.011362334713339806, + 0.008291443809866905, + 0.009538736194372177, + 0.014870616607367992, + 0.012237656861543655, + 0.013561957515776157, + 0.011918445117771626, + 0.005949808284640312, + 0.011043254286050797, + 0.010108635760843754, + 0.007155473344027996, + 0.010135353542864323, + 0.008145146071910858, + 0.012526360340416431, + 0.00890433695167303, + 0.008237087167799473, + 0.010227681137621403, + 0.013647809624671936, + 0.015332653187215328, + 0.016155574470758438, + 0.041270095854997635, + 0.021649988368153572, + 0.013363498263061047, + 0.005563108250498772, + 0.015143760479986668, + 0.007418352644890547, + 0.0041472879238426685, + 0.005567619111388922, + 0.006456025876104832, + 0.004963783547282219, + 0.006886121816933155, + 0.0045090545900166035, + 0.004523996729403734, + 0.0048648519441485405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.07011262327432632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006349414587021 + ], + "opposite_logprob_tokens": [ + -2.695063591003418 + ], + "logprob_reward": -0.07006349414587021, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006349414587021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13884268701076508, + 0.1026889756321907, + 0.041473645716905594, + 0.039164893329143524, + 0.04064018651843071, + 0.07345463335514069, + 0.06653980910778046, + 0.06851901113986969, + 0.3725644052028656, + 0.05611177161335945 + ], + "credit_map": [ + 0.0216110460460186, + 0.009390261024236679, + 0.018659546971321106, + 0.012619800865650177, + 0.021038245409727097, + 0.011305573396384716, + 0.032945774495601654, + 0.01127244159579277, + 0.012488188222050667, + 0.006944069638848305, + 0.015376690775156021, + 0.012669102288782597, + 0.010036091320216656, + 0.009088211692869663, + 0.020569337531924248, + 0.015517274849116802, + 0.0036781111266463995, + 0.0036499802954494953, + 0.005126622039824724, + 0.007979926653206348, + 0.004174103494733572, + 0.005040018819272518, + 0.006345090921968222, + 0.005479790270328522, + 0.002965451218187809, + 0.009369215928018093, + 0.009905573911964893, + 0.0039274548180401325, + 0.0027149871457368135, + 0.0028211395256221294, + 0.0037880537565797567, + 0.003673017956316471, + 0.005826552398502827, + 0.004767036065459251, + 0.006300789769738913, + 0.005396071821451187, + 0.004396651405841112, + 0.004409415647387505, + 0.004225499462336302, + 0.005318170879036188, + 0.0052431728690862656, + 0.007527949288487434, + 0.012392790988087654, + 0.016696440055966377, + 0.008040890097618103, + 0.009620850905776024, + 0.008158708922564983, + 0.005773829761892557, + 0.008376690559089184, + 0.009372935630381107, + 0.006185733247548342, + 0.00962916575372219, + 0.007115927524864674, + 0.005580671597272158, + 0.010207648389041424, + 0.010071031749248505, + 0.0055796424858272076, + 0.006662048399448395, + 0.0077789556235075, + 0.009099528193473816, + 0.009501184336841106, + 0.007466399576514959, + 0.007612450048327446, + 0.01481880247592926, + 0.03813003748655319, + 0.0240875706076622, + 0.13027068972587585, + 0.06560572236776352, + 0.023185981437563896, + 0.020564356818795204, + 0.04172326251864433, + 0.028996769338846207, + 0.004895207937806845, + 0.009025514125823975, + 0.008444380015134811, + 0.0075019970536231995, + 0.006157983560115099, + 0.0066721695475280285, + 0.0058863162994384766, + 0.007528204470872879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.07006349414587021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547949329018593 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012547949329018593, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547949329018593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15806126594543457, + 0.0876506045460701, + 0.06772901117801666, + 0.049934569746255875, + 0.14713531732559204, + 0.18630580604076385, + 0.09549764543771744, + 0.0683874636888504, + 0.04283330217003822, + 0.09646502882242203 + ], + "credit_map": [ + 0.020641766488552094, + 0.010385260917246342, + 0.024751637130975723, + 0.01884857378900051, + 0.016779372468590736, + 0.021508390083909035, + 0.02485661953687668, + 0.020289644598960876, + 0.010101990774273872, + 0.007976009510457516, + 0.011504940688610077, + 0.011503160931169987, + 0.008145703002810478, + 0.007262602914124727, + 0.01618810184299946, + 0.014968087896704674, + 0.008018522523343563, + 0.006692212074995041, + 0.008212652057409286, + 0.010172459296882153, + 0.006012930069118738, + 0.008787198923528194, + 0.00950255710631609, + 0.01033047679811716, + 0.005072334315627813, + 0.0100203026086092, + 0.01161952968686819, + 0.0063758776523172855, + 0.006009649019688368, + 0.00285312975756824, + 0.004711790941655636, + 0.0032719536684453487, + 0.015841590240597725, + 0.03469111770391464, + 0.03339120373129845, + 0.01991693675518036, + 0.01425324659794569, + 0.008904406800866127, + 0.012073916383087635, + 0.008062898181378841, + 0.01865324191749096, + 0.026865972205996513, + 0.026706738397479057, + 0.03180675581097603, + 0.02655431255698204, + 0.02646561712026596, + 0.018357759341597557, + 0.010895409621298313, + 0.007286118343472481, + 0.012123649008572102, + 0.012148182839155197, + 0.014290735125541687, + 0.010603792034089565, + 0.008079544641077518, + 0.016359509900212288, + 0.014606107026338577, + 0.00585266575217247, + 0.009489571675658226, + 0.006971444468945265, + 0.012226864695549011, + 0.007978679612278938, + 0.006856550928205252, + 0.008347872644662857, + 0.010663812980055809, + 0.0025552245788276196, + 0.0033772483002394438, + 0.01233983039855957, + 0.006474948022514582, + 0.005780134350061417, + 0.0022760096471756697, + 0.006513288710266352, + 0.0035166156012564898, + 0.009559244848787785, + 0.012284518219530582, + 0.011835633777081966, + 0.010161984711885452, + 0.020602112635970116, + 0.014894760213792324, + 0.010103157721459866, + 0.007023619022220373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.012547949329018593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062046490609645844 + ], + "opposite_logprob_tokens": [ + -2.812046527862549 + ], + "logprob_reward": -0.062046490609645844, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062046490609645844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15568913519382477, + 0.06824066489934921, + 0.057127438485622406, + 0.08533874154090881, + 0.07889450341463089, + 0.09733344614505768, + 0.09043846279382706, + 0.04181557893753052, + 0.23743824660778046, + 0.0876837819814682 + ], + "credit_map": [ + 0.019413840025663376, + 0.013654295355081558, + 0.023668980225920677, + 0.01782369613647461, + 0.015296121127903461, + 0.02146109566092491, + 0.021025029942393303, + 0.023346098139882088, + 0.0068104080855846405, + 0.008236618712544441, + 0.009551363065838814, + 0.007749881595373154, + 0.007127751596271992, + 0.005158476065844297, + 0.013337972573935986, + 0.010268203914165497, + 0.006745698396116495, + 0.0054926094599068165, + 0.0077091180719435215, + 0.011147002689540386, + 0.005528499372303486, + 0.006222277879714966, + 0.0083591528236866, + 0.005923082586377859, + 0.006081209052354097, + 0.020291823893785477, + 0.02255776710808277, + 0.010451218113303185, + 0.006973870098590851, + 0.006193169392645359, + 0.00740442331880331, + 0.005385268945246935, + 0.007462607230991125, + 0.01395491138100624, + 0.013765554875135422, + 0.014732681214809418, + 0.007199406158179045, + 0.0048981006257236, + 0.00933520495891571, + 0.007546043023467064, + 0.007242973428219557, + 0.010173194110393524, + 0.01744057796895504, + 0.021706372499465942, + 0.00995731633156538, + 0.01006108708679676, + 0.011068788357079029, + 0.009683139622211456, + 0.009335711598396301, + 0.013088294304907322, + 0.007991123013198376, + 0.01164272427558899, + 0.010896584950387478, + 0.007805540692061186, + 0.014909435994923115, + 0.014769056811928749, + 0.002913497621193528, + 0.004787955898791552, + 0.00549782020971179, + 0.006113147363066673, + 0.005140047520399094, + 0.005341947078704834, + 0.005040678661316633, + 0.006980489008128643, + 0.024036645889282227, + 0.044284164905548096, + 0.061053697019815445, + 0.03372785821557045, + 0.02376900613307953, + 0.01225694827735424, + 0.021715862676501274, + 0.01659407652914524, + 0.009305987507104874, + 0.009144505485892296, + 0.01233713235706091, + 0.008103796280920506, + 0.008349855430424213, + 0.013846198096871376, + 0.010031383484601974, + 0.016564933583140373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.062046490609645844 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion close-up focuses on a medical certificate, printed in black ink, gripped tightly by a pair of bony hands, their knuckles turning slightly white from the pressure. With a resounding, drawn-out, crisp tearing sound, the paper is decisively and violently split in two, then torn again, instantly shattering into countless fragments of varying sizes. The hands then release, and the camera tilts slightly upward, capturing the pieces of paper drifting through the air as if weightless. The warm, slanting afternoon sun streams through the window, creating a distinct Tyndall beam in the air, shimmering with tiny particles of dust. As each swirling, tumbling scrap of paper passes through the light, its jagged white edges momentarily illuminated, shimmering with a fleeting, soft glow. The background is a completely out-of-focus, soft-toned corner of the room. All visual focus is on these fragments fluttering like butterflies, falling in an elegant and slow posture. The whole picture is full of strong symbolic meaning, conveying a heavy yet relieved tranquil atmosphere in an almost silent dynamic.", + "seed": 13218, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_15/index_7559_seed_13218_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3473370783030987, + "vlm_logprob": -0.3473370783030987, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.3473370783030987, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6021007299423218 + ], + "opposite_logprob_tokens": [ + -0.22710071504116058 + ], + "logprob_reward": -1.6021007299423218, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6021007299423218, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09419218450784683, + 0.02661031484603882, + 0.04420572146773338, + 0.0813385471701622, + 0.14534036815166473, + 0.10840855538845062, + 0.3237466514110565, + 0.08153551816940308, + 0.05506357178092003, + 0.0395585298538208 + ], + "credit_map": [ + 0.011980358511209488, + 0.011849128641188145, + 0.006882338784635067, + 0.007980603724718094, + 0.017141634598374367, + 0.013739602640271187, + 0.012626045383512974, + 0.011992471292614937, + 0.0035652637016028166, + 0.0031963535584509373, + 0.0023321278858929873, + 0.0024881914723664522, + 0.003924105782061815, + 0.0033461928833276033, + 0.004766968544572592, + 0.0029911110177636147, + 0.006362673360854387, + 0.005409195087850094, + 0.003594723530113697, + 0.003851431654766202, + 0.00765151996165514, + 0.005818502511829138, + 0.006779584567993879, + 0.004738091491162777, + 0.011841367930173874, + 0.01123120915144682, + 0.006104941479861736, + 0.005817272234708071, + 0.010429584421217442, + 0.010105318389832973, + 0.012297961860895157, + 0.013510897755622864, + 0.039014704525470734, + 0.005309450905770063, + 0.04672665148973465, + 0.012188851833343506, + 0.011951527558267117, + 0.015034168027341366, + 0.00626986101269722, + 0.00884515792131424, + 0.008805959485471249, + 0.012625756673514843, + 0.01008106954395771, + 0.015319234691560268, + 0.0124037005007267, + 0.01620364747941494, + 0.019985660910606384, + 0.012983524240553379, + 0.020781010389328003, + 0.07577741146087646, + 0.12364635616540909, + 0.013663644902408123, + 0.0384615920484066, + 0.027409523725509644, + 0.011195938102900982, + 0.01281116995960474, + 0.006390135735273361, + 0.015591451898217201, + 0.009622148238122463, + 0.013907540589571, + 0.009606164880096912, + 0.0069910078309476376, + 0.01029874011874199, + 0.009128326550126076, + 0.005729242693632841, + 0.005421734880656004, + 0.011459887959063053, + 0.0060484400019049644, + 0.005042087286710739, + 0.007925390265882015, + 0.005058141890913248, + 0.008378646336495876, + 0.004329440649598837, + 0.0039514657109975815, + 0.007146394345909357, + 0.006323015782982111, + 0.002818373031914234, + 0.005817632656544447, + 0.004124369937926531, + 0.005047840531915426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -1.6021007299423218 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209757924079895 + ], + "opposite_logprob_tokens": [ + -2.8120975494384766 + ], + "logprob_reward": -0.06209757924079895, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209757924079895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10967343300580978, + 0.03643612191081047, + 0.0676710233092308, + 0.09799416363239288, + 0.061377134174108505, + 0.1026085913181305, + 0.3385950028896332, + 0.053732216358184814, + 0.09239790588617325, + 0.039514392614364624 + ], + "credit_map": [ + 0.014109437353909016, + 0.011908425949513912, + 0.007484416011720896, + 0.01315570529550314, + 0.02327980287373066, + 0.010953261516988277, + 0.019652927294373512, + 0.009129459038376808, + 0.006093996576964855, + 0.004140286706387997, + 0.003675668966025114, + 0.0030988543294370174, + 0.0036569247022271156, + 0.005255542695522308, + 0.005357867106795311, + 0.005156985484063625, + 0.01001759059727192, + 0.01047777570784092, + 0.0075706299394369125, + 0.005568492691963911, + 0.009398707188665867, + 0.007697233464568853, + 0.010583046823740005, + 0.006357548758387566, + 0.012460555881261826, + 0.015082906000316143, + 0.009327241219580173, + 0.008223081938922405, + 0.01489343773573637, + 0.009277984499931335, + 0.013698949478566647, + 0.015030015259981155, + 0.01592942886054516, + 0.00604321900755167, + 0.015081656165421009, + 0.006542873568832874, + 0.004358070436865091, + 0.006758790463209152, + 0.0033670340199023485, + 0.0032960684038698673, + 0.00827821809798479, + 0.009111827239394188, + 0.011210879310965538, + 0.014129847288131714, + 0.009372682310640812, + 0.015842530876398087, + 0.011436102911829948, + 0.023226505145430565, + 0.021453196182847023, + 0.03953315317630768, + 0.09628908336162567, + 0.014946547336876392, + 0.11252030730247498, + 0.029161136597394943, + 0.01203867420554161, + 0.012652908451855183, + 0.0045876698568463326, + 0.008085403591394424, + 0.006139735691249371, + 0.009447745978832245, + 0.0048282272182404995, + 0.005543885286897421, + 0.008085341192781925, + 0.007014211732894182, + 0.00682067358866334, + 0.011114031076431274, + 0.012928948737680912, + 0.016890764236450195, + 0.009777998551726341, + 0.008975554257631302, + 0.013832940720021725, + 0.01205699797719717, + 0.00480313366279006, + 0.005567426793277264, + 0.007542687468230724, + 0.005003963131457567, + 0.0031595868058502674, + 0.005396844353526831, + 0.004302268847823143, + 0.0037384836468845606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.06209757924079895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596671849489212 + ], + "opposite_logprob_tokens": [ + -3.895596742630005 + ], + "logprob_reward": -0.020596671849489212, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596671849489212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02796228602528572, + 0.02534599043428898, + 0.014016617089509964, + 0.049456194043159485, + 0.04732657968997955, + 0.10277018696069717, + 0.5388544201850891, + 0.08650606870651245, + 0.0760427862405777, + 0.0317188985645771 + ], + "credit_map": [ + 0.00319283246062696, + 0.0038373712450265884, + 0.002785546239465475, + 0.003731085918843746, + 0.004178510047495365, + 0.002909786533564329, + 0.0049160318449139595, + 0.002411122201010585, + 0.004588255658745766, + 0.0024490428622812033, + 0.0018737850477918983, + 0.001650091726332903, + 0.004669294226914644, + 0.0031951938290148973, + 0.004242862109094858, + 0.002677466254681349, + 0.002166559686884284, + 0.0017479043453931808, + 0.0012430151691660285, + 0.0015406699385493994, + 0.0016699127154424787, + 0.0012849377235397696, + 0.002919925842434168, + 0.001443691784515977, + 0.008375357836484909, + 0.010295328684151173, + 0.0045271096751093864, + 0.003701987210661173, + 0.006740784738212824, + 0.0051415227353572845, + 0.00559108704328537, + 0.005083012860268354, + 0.011169219389557838, + 0.0028817506972700357, + 0.014108535833656788, + 0.004584171809256077, + 0.0033597382716834545, + 0.00552003039047122, + 0.0025723809376358986, + 0.0031307521276175976, + 0.0059109157882630825, + 0.009059076197445393, + 0.012980423867702484, + 0.009032594040036201, + 0.010148412548005581, + 0.018386302515864372, + 0.008561513386666775, + 0.028690949082374573, + 0.02170957624912262, + 0.2150675356388092, + 0.13481086492538452, + 0.017855236306786537, + 0.11312678456306458, + 0.023462630808353424, + 0.004870832897722721, + 0.007950983941555023, + 0.007060408592224121, + 0.012413599528372288, + 0.010559618473052979, + 0.015808286145329475, + 0.007130960933864117, + 0.010751182213425636, + 0.011152464896440506, + 0.01162954792380333, + 0.005169651936739683, + 0.010515048168599606, + 0.007613308262079954, + 0.011493629775941372, + 0.012904551811516285, + 0.00929070170968771, + 0.011743342503905296, + 0.007312556263059378, + 0.004782889969646931, + 0.0033836488146334887, + 0.004846745636314154, + 0.003931940067559481, + 0.002864145440980792, + 0.0037085360381752253, + 0.003123517846688628, + 0.005077475216239691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.020596671849489212 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370732069015503 + ], + "opposite_logprob_tokens": [ + -3.4087073802948 + ], + "logprob_reward": -0.03370732069015503, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370732069015503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07169356197118759, + 0.014186599291861057, + 0.012943118810653687, + 0.02168315276503563, + 0.07380230724811554, + 0.07247497141361237, + 0.40878066420555115, + 0.15829862654209137, + 0.034189146012067795, + 0.13194790482521057 + ], + "credit_map": [ + 0.010393630713224411, + 0.0077087609097361565, + 0.004427253268659115, + 0.006287024822086096, + 0.015611798502504826, + 0.008501065894961357, + 0.012344223447144032, + 0.006419799756258726, + 0.0018339032540097833, + 0.0015853828517720103, + 0.0014323112554848194, + 0.00124258641153574, + 0.0022082189098000526, + 0.001645834418013692, + 0.002715275390073657, + 0.0015230854041874409, + 0.0015493861865252256, + 0.0015197108732536435, + 0.0014278360176831484, + 0.0017878368962556124, + 0.0017548379255458713, + 0.001367297489196062, + 0.0021134272683411837, + 0.0014227848732843995, + 0.002818252658471465, + 0.002937657991424203, + 0.002696794457733631, + 0.002063711639493704, + 0.004118440207093954, + 0.0026365239173173904, + 0.00215665390715003, + 0.002255118452012539, + 0.01877027377486229, + 0.0030105109326541424, + 0.02548435889184475, + 0.0066798957996070385, + 0.004415391944348812, + 0.006389806978404522, + 0.0029848271515220404, + 0.0060672336257994175, + 0.00411901855841279, + 0.007884345017373562, + 0.007067584432661533, + 0.008625415153801441, + 0.008515444584190845, + 0.010872564278542995, + 0.014502189122140408, + 0.01088840514421463, + 0.027727393433451653, + 0.042210906744003296, + 0.12204798310995102, + 0.01828511245548725, + 0.13761483132839203, + 0.03485320135951042, + 0.012104390189051628, + 0.01393680740147829, + 0.011560056358575821, + 0.02000184915959835, + 0.021970508620142937, + 0.027692191302776337, + 0.012846876867115498, + 0.020963523536920547, + 0.02344563417136669, + 0.019817976281046867, + 0.002905842149630189, + 0.003802971448749304, + 0.006704740226268768, + 0.006339086685329676, + 0.003234593663364649, + 0.002557247644290328, + 0.005983470939099789, + 0.00266119372099638, + 0.011628393083810806, + 0.012566260062158108, + 0.02551192231476307, + 0.018251750618219376, + 0.00984856765717268, + 0.016790593042969704, + 0.01822986640036106, + 0.019120555371046066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.03370732069015503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a medical certificate printed in black ink visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are bony hands gripping the paper with knuckles turning white from pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper torn into countless fragments of varying sizes that drift through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm slanting afternoon sun creating a distinct Tyndall beam with dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt slightly upward to capture the paper fragments drifting in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183089792728424 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183089792728424, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183089792728424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06716762483119965, + 0.013235999271273613, + 0.008298701606690884, + 0.013983825221657753, + 0.06406524777412415, + 0.04084736853837967, + 0.6882570385932922, + 0.05187011882662773, + 0.026089096441864967, + 0.026184964925050735 + ], + "credit_map": [ + 0.008107306435704231, + 0.007961457595229149, + 0.005671784747391939, + 0.008183028548955917, + 0.014980538748204708, + 0.006689746864140034, + 0.010391050949692726, + 0.0051827081479132175, + 0.0019786572083830833, + 0.0014059696113690734, + 0.0007448313408531249, + 0.0011055748909711838, + 0.0035404867958277464, + 0.00138789601624012, + 0.002065627835690975, + 0.0010069558629766107, + 0.0011601055739447474, + 0.001184353488497436, + 0.0007075433968566358, + 0.0007111106533557177, + 0.0011806996772065759, + 0.0009807082824409008, + 0.0015014589298516512, + 0.0008727216045372188, + 0.002208244288340211, + 0.0022031362168490887, + 0.0016764136962592602, + 0.0007607716252095997, + 0.0022991555742919445, + 0.0021810894832015038, + 0.00106618064455688, + 0.0015888337511569262, + 0.019482184201478958, + 0.0027263546362519264, + 0.021668659523129463, + 0.005468336399644613, + 0.004126050975173712, + 0.0046585919335484505, + 0.0020929197780787945, + 0.003842148231342435, + 0.0028959799092262983, + 0.004097866825759411, + 0.004374518990516663, + 0.0050238389521837234, + 0.004201996605843306, + 0.006249445956200361, + 0.00581618444994092, + 0.008187537081539631, + 0.0270856823772192, + 0.283809095621109, + 0.07961470633745193, + 0.01959981583058834, + 0.23427999019622803, + 0.02121482789516449, + 0.012481962330639362, + 0.010170971974730492, + 0.004330716095864773, + 0.009134150110185146, + 0.006183918099850416, + 0.007902334444224834, + 0.004805112723261118, + 0.007704266346991062, + 0.005879328120499849, + 0.0059302919544279575, + 0.002343216212466359, + 0.004724818281829357, + 0.0025967031251639128, + 0.0028961964417248964, + 0.004827046301215887, + 0.0025960863567888737, + 0.0032229966018348932, + 0.0028820312581956387, + 0.0030096308328211308, + 0.0022530509158968925, + 0.007075304631143808, + 0.0038468223065137863, + 0.0019097087206318974, + 0.0024421398993581533, + 0.003593186382204294, + 0.0020551192574203014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.018183089792728424 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_2/index_tensor([7369, 7369])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_2/index_tensor([7369, 7369])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..73cb59cc9ee6e3be57cce3141a48d32691d13f18 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_2/index_tensor([7369, 7369])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot, perched high above the sidewalk, follows a sleek, modern black-and-white police interceptor car as it speeds through rain-soaked city streets at night. A relentless downpour has slicked the asphalt to a black mirror, and the car's wide tires kick up a torrent of spray, barely maintaining traction as it hurtles forward. A discreet LED light bar atop the car's roof flickers silently, painting the wet street, the sleek car body, and the towering, dark skyscraper facades into a frenzied, pulsating flash of intense red and blue. The car's own powerful headlights carve a sharp white cone into the darkness, while the glow of distant neon signs and streetlights distorts and ripples across the wet surface and the puddles it carves, creating a palpable atmosphere imbued with a high-stakes sense of urgency and relentless pursuit.", + "seed": 31263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_2/index_7369_seed_31263_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19656678717583417, + "vlm_logprob": -0.19656678717583417, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.19656678717583417, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5772019624710083 + ], + "opposite_logprob_tokens": [ + -0.9522019624710083 + ], + "logprob_reward": -0.5772019624710083, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5772019624710083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10290790349245071, + 0.07638339698314667, + 0.4118737280368805, + 0.05833249166607857, + 0.16754120588302612, + 0.06962797790765762, + 0.04507697373628616, + 0.02833540551364422, + 0.01776283048093319, + 0.022158078849315643 + ], + "credit_map": [ + 0.017163008451461792, + 0.015539021231234074, + 0.01667611487209797, + 0.014780239202082157, + 0.008056212216615677, + 0.010581058450043201, + 0.009911561384797096, + 0.01020066998898983, + 0.010867743752896786, + 0.01541269849985838, + 0.010416459292173386, + 0.009035804308950901, + 0.00560724176466465, + 0.011349411681294441, + 0.005716435611248016, + 0.007977597415447235, + 0.04652348905801773, + 0.014036720618605614, + 0.020834097638726234, + 0.009811478666961193, + 0.031067771837115288, + 0.1626894474029541, + 0.013216600753366947, + 0.11369408667087555, + 0.008199653588235378, + 0.005937549751251936, + 0.010549213737249374, + 0.004670104011893272, + 0.00815279595553875, + 0.007611075881868601, + 0.00643001776188612, + 0.006782080978155136, + 0.011002588085830212, + 0.008894462138414383, + 0.014554079622030258, + 0.00939405057579279, + 0.05295872688293457, + 0.010998128913342953, + 0.04588603228330612, + 0.013853139244019985, + 0.008572600781917572, + 0.009156749583780766, + 0.010905878618359566, + 0.015201524831354618, + 0.0040645357221364975, + 0.006405712570995092, + 0.009790580719709396, + 0.005530384834855795, + 0.007263569161295891, + 0.00543167581781745, + 0.006036519538611174, + 0.00849339459091425, + 0.0035919207148253918, + 0.0050219399854540825, + 0.00508775282651186, + 0.004150196444243193, + 0.0059666140004992485, + 0.004398026969283819, + 0.003478806000202894, + 0.0028540357016026974, + 0.002952202456071973, + 0.0032552743796259165, + 0.0028110560961067677, + 0.002619385253638029, + 0.003903905861079693, + 0.0018431330099701881, + 0.002053149975836277, + 0.0023511999752372503, + 0.001817421056330204, + 0.0018092130776494741, + 0.0018865485908463597, + 0.002098257187753916, + 0.0037903611082583666, + 0.002137881936505437, + 0.002492590807378292, + 0.0016286132158711553, + 0.003565457183867693, + 0.003233417635783553, + 0.002147053135558963, + 0.0031627025455236435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.5772019624710083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008194923400879 + ], + "opposite_logprob_tokens": [ + -2.695081949234009 + ], + "logprob_reward": -0.07008194923400879, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008194923400879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08358863741159439, + 0.04769319295883179, + 0.1368628293275833, + 0.07678407430648804, + 0.040424566715955734, + 0.17127776145935059, + 0.07689525932073593, + 0.24464371800422668, + 0.06284292042255402, + 0.05898704007267952 + ], + "credit_map": [ + 0.009758575819432735, + 0.01611887477338314, + 0.01583430916070938, + 0.010335393249988556, + 0.008417841047048569, + 0.006744515150785446, + 0.008254592306911945, + 0.00812452845275402, + 0.007088611368089914, + 0.008188838139176369, + 0.008036096580326557, + 0.00371900899335742, + 0.0037272938061505556, + 0.007245976943522692, + 0.0038768972735852003, + 0.005810463801026344, + 0.01998605765402317, + 0.0040976982563734055, + 0.031363990157842636, + 0.015475373715162277, + 0.007585745304822922, + 0.015214364975690842, + 0.006171025335788727, + 0.03696856647729874, + 0.009671557694673538, + 0.006855701562017202, + 0.01703251525759697, + 0.009760548360645771, + 0.010428288020193577, + 0.006914281751960516, + 0.007440348621457815, + 0.008680825121700764, + 0.005886610131710768, + 0.004689032677561045, + 0.007830956019461155, + 0.003719240427017212, + 0.004148794338107109, + 0.0036742703523486853, + 0.005260340869426727, + 0.005215319339185953, + 0.019837990403175354, + 0.02202047035098076, + 0.03172536939382553, + 0.040694065392017365, + 0.010716760531067848, + 0.009720142930746078, + 0.02371990866959095, + 0.01284303143620491, + 0.013626360334455967, + 0.008031793870031834, + 0.01305412221699953, + 0.012772335670888424, + 0.005413615144789219, + 0.008382605388760567, + 0.009129834361374378, + 0.00648459792137146, + 0.06562317162752151, + 0.01858150027692318, + 0.045864105224609375, + 0.023977721109986305, + 0.014072000049054623, + 0.015623878687620163, + 0.025944167748093605, + 0.0349571667611599, + 0.009863565675914288, + 0.007454802747815847, + 0.007775321137160063, + 0.012164761312305927, + 0.004655944649130106, + 0.006764049641788006, + 0.007929899729788303, + 0.006234574131667614, + 0.010389095172286034, + 0.006853800266981125, + 0.008564962074160576, + 0.004190742503851652, + 0.007857647724449635, + 0.005821279250085354, + 0.004408712964504957, + 0.010900798253715038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.07008194923400879 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28139233589172363 + ], + "opposite_logprob_tokens": [ + -1.4063923358917236 + ], + "logprob_reward": -0.28139233589172363, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28139233589172363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08102772384881973, + 0.0658433586359024, + 0.20453056693077087, + 0.16010607779026031, + 0.0759497880935669, + 0.10111957788467407, + 0.06410962343215942, + 0.03909800201654434, + 0.08082662522792816, + 0.12738871574401855 + ], + "credit_map": [ + 0.01099375355988741, + 0.014645334333181381, + 0.012853492982685566, + 0.01374172605574131, + 0.006849478464573622, + 0.006921008694916964, + 0.007969245314598083, + 0.007053684443235397, + 0.009979895316064358, + 0.013156022876501083, + 0.007659293245524168, + 0.007544470485299826, + 0.005423417314887047, + 0.010161821730434895, + 0.0052178953774273396, + 0.006700537167489529, + 0.02306968905031681, + 0.007534058298915625, + 0.012258409522473812, + 0.006157738156616688, + 0.015400185249745846, + 0.079005166888237, + 0.006634479388594627, + 0.05447082966566086, + 0.01968725398182869, + 0.014320875518023968, + 0.026575619354844093, + 0.020798351615667343, + 0.021904131397604942, + 0.01697157882153988, + 0.0205422043800354, + 0.019306061789393425, + 0.007835118100047112, + 0.006214495748281479, + 0.00860962737351656, + 0.004767697770148516, + 0.009236263111233711, + 0.009795253165066242, + 0.020836053416132927, + 0.008655277080833912, + 0.012077061459422112, + 0.011566842906177044, + 0.012183308601379395, + 0.011501247994601727, + 0.009781298227608204, + 0.02179432474076748, + 0.013116366229951382, + 0.009099126793444157, + 0.011336592957377434, + 0.006104627624154091, + 0.005028136074542999, + 0.005359882954508066, + 0.005082510411739349, + 0.011808987706899643, + 0.009411259554326534, + 0.00997762382030487, + 0.005438988097012043, + 0.005319368094205856, + 0.0060034263879060745, + 0.004889247007668018, + 0.004182181786745787, + 0.004090957809239626, + 0.0054825772531330585, + 0.0036912530194967985, + 0.012865321710705757, + 0.0068494039587676525, + 0.011213429272174835, + 0.007164429873228073, + 0.016224179416894913, + 0.010077118873596191, + 0.007826662622392178, + 0.00860607624053955, + 0.021831514313817024, + 0.01016364898532629, + 0.022678464651107788, + 0.008715420961380005, + 0.02039395086467266, + 0.016037527471780777, + 0.009697065688669682, + 0.01787111908197403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.28139233589172363 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089949868619442 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089949868619442, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089949868619442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10526099056005478, + 0.04369747266173363, + 0.29808053374290466, + 0.12098123878240585, + 0.12416569143533707, + 0.06360617280006409, + 0.0468316450715065, + 0.12992948293685913, + 0.03382709622383118, + 0.03361967206001282 + ], + "credit_map": [ + 0.014121424406766891, + 0.013266230002045631, + 0.018996071070432663, + 0.020286519080400467, + 0.009127061814069748, + 0.009889036417007446, + 0.009454671293497086, + 0.01011996902525425, + 0.0062653073109686375, + 0.007412194274365902, + 0.006224425975233316, + 0.0033231840934604406, + 0.004671765957027674, + 0.006499877665191889, + 0.003965536132454872, + 0.0053351749666035175, + 0.03198162466287613, + 0.011251351796090603, + 0.03750452771782875, + 0.0183818768709898, + 0.02423408441245556, + 0.1090211495757103, + 0.009628395549952984, + 0.056077491492033005, + 0.017595462501049042, + 0.012646004557609558, + 0.018009241670370102, + 0.011990233324468136, + 0.018492022529244423, + 0.01178458146750927, + 0.014789629727602005, + 0.015674054622650146, + 0.010363456793129444, + 0.007950741797685623, + 0.012220249511301517, + 0.006284506060183048, + 0.04125114902853966, + 0.012563446536660194, + 0.022837527096271515, + 0.010694598779082298, + 0.011028053238987923, + 0.006550958845764399, + 0.009875470772385597, + 0.004018150269985199, + 0.008299187757074833, + 0.0106873270124197, + 0.008376993238925934, + 0.0047700293362140656, + 0.006989173591136932, + 0.005125990603119135, + 0.005889429710805416, + 0.006557057145982981, + 0.003427213989198208, + 0.006176729686558247, + 0.0062602004036307335, + 0.006405847612768412, + 0.034573301672935486, + 0.012187885120511055, + 0.01684863492846489, + 0.012076344341039658, + 0.01195619348436594, + 0.00916916225105524, + 0.014064191840589046, + 0.019053759053349495, + 0.006257845088839531, + 0.003552447771653533, + 0.006005863659083843, + 0.003881822107359767, + 0.002856701845303178, + 0.004120048135519028, + 0.0031742730643600225, + 0.003978092223405838, + 0.003044499084353447, + 0.003859204240143299, + 0.003983347676694393, + 0.0028938313480466604, + 0.005533876828849316, + 0.005464176647365093, + 0.0026350324042141438, + 0.0062057022005319595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.011089949868619442 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306773841381073 + ], + "opposite_logprob_tokens": [ + -3.168067693710327 + ], + "logprob_reward": -0.04306773841381073, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306773841381073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06037614867091179, + 0.07768193632364273, + 0.2955484986305237, + 0.16933777928352356, + 0.06021406874060631, + 0.11080768704414368, + 0.05988902226090431, + 0.04430220276117325, + 0.05166159197688103, + 0.07018112391233444 + ], + "credit_map": [ + 0.00983898714184761, + 0.007740695495158434, + 0.007242526393383741, + 0.007954204455018044, + 0.005827449262142181, + 0.007634224835783243, + 0.007931487634778023, + 0.006206569727510214, + 0.010105992667376995, + 0.01893109828233719, + 0.009973929263651371, + 0.005447701551020145, + 0.008104009553790092, + 0.013521450571715832, + 0.004255529493093491, + 0.007342217024415731, + 0.03195176646113396, + 0.010355659760534763, + 0.02165943756699562, + 0.010576414875686169, + 0.02248384803533554, + 0.11384414881467819, + 0.00904940813779831, + 0.07562779635190964, + 0.019583895802497864, + 0.014201820828020573, + 0.028415633365511894, + 0.017982378602027893, + 0.025422124192118645, + 0.014756889082491398, + 0.017680514603853226, + 0.03129451349377632, + 0.008020740002393723, + 0.007535693701356649, + 0.008120390586555004, + 0.0046217008493840694, + 0.009611387737095356, + 0.005260261241346598, + 0.010640116408467293, + 0.006403776817023754, + 0.014504935592412949, + 0.014647925272583961, + 0.01756219007074833, + 0.021250590682029724, + 0.006200320553034544, + 0.008219154551625252, + 0.01785561442375183, + 0.01056693959981203, + 0.010456344112753868, + 0.006776473950594664, + 0.006859768182039261, + 0.008476482704281807, + 0.0033670549746602774, + 0.005995724815875292, + 0.011007792316377163, + 0.0069493758492171764, + 0.009782745502889156, + 0.004243307281285524, + 0.00843344908207655, + 0.004087848123162985, + 0.0025805910117924213, + 0.003066843142732978, + 0.00521897291764617, + 0.006888441741466522, + 0.005376910325139761, + 0.004729094915091991, + 0.0079380227252841, + 0.007105869706720114, + 0.004988666158169508, + 0.006352430675178766, + 0.005432483274489641, + 0.009738106280565262, + 0.011245154775679111, + 0.005273601971566677, + 0.010177884250879288, + 0.0058210170827806, + 0.011478586122393608, + 0.009110389277338982, + 0.0059820725582540035, + 0.01109241507947445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.04306773841381073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, perched high above the sidewalk, follows a sleek, modern black-and-white police interceptor car as it speeds through rain-soaked city streets at night. A relentless downpour has slicked the asphalt to a black mirror, and the car's wide tires kick up a torrent of spray, barely maintaining traction as it hurtles forward. A discreet LED light bar atop the car's roof flickers silently, painting the wet street, the sleek car body, and the towering, dark skyscraper facades into a frenzied, pulsating flash of intense red and blue. The car's own powerful headlights carve a sharp white cone into the darkness, while the glow of distant neon signs and streetlights distorts and ripples across the wet surface and the puddles it carves, creating a palpable atmosphere imbued with a high-stakes sense of urgency and relentless pursuit.", + "seed": 31264, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_2/index_7369_seed_31264_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15496525913476944, + "vlm_logprob": -0.15496525913476944, + "vlm_yes_no_margin": 2.549999988079071, + "vlm_token_credit": -0.15496525913476944, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5314405560493469 + ], + "opposite_logprob_tokens": [ + -1.1564404964447021 + ], + "logprob_reward": -0.5314405560493469, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.5314405560493469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11660017818212509, + 0.06167919933795929, + 0.07742121070623398, + 0.10439138859510422, + 0.0859779641032219, + 0.03524323180317879, + 0.08529132604598999, + 0.0445663183927536, + 0.22022242844104767, + 0.1686067432165146 + ], + "credit_map": [ + 0.015162596479058266, + 0.017304958775639534, + 0.00920325517654419, + 0.00944394152611494, + 0.015576094388961792, + 0.015047307126224041, + 0.01626269891858101, + 0.01859932579100132, + 0.008266846649348736, + 0.007547779008746147, + 0.007801187690347433, + 0.006822013761848211, + 0.008531628176569939, + 0.008726054802536964, + 0.005900616757571697, + 0.008083073422312737, + 0.007972984574735165, + 0.011684014461934566, + 0.00866430439054966, + 0.008891928009688854, + 0.011779800988733768, + 0.008320086635649204, + 0.011033982038497925, + 0.009074111469089985, + 0.005499065853655338, + 0.007546443957835436, + 0.006931394338607788, + 0.00874126236885786, + 0.015843551605939865, + 0.008282233029603958, + 0.03667716309428215, + 0.014870275743305683, + 0.010408503003418446, + 0.006996283307671547, + 0.010094650089740753, + 0.006939152255654335, + 0.014108994044363499, + 0.010673009790480137, + 0.014020688831806183, + 0.012736676260828972, + 0.004295557737350464, + 0.006230267230421305, + 0.00308282021433115, + 0.00425340048968792, + 0.003761693136766553, + 0.005147093906998634, + 0.004323751665651798, + 0.004148645792156458, + 0.006133651826530695, + 0.010437210090458393, + 0.004489629529416561, + 0.005115867592394352, + 0.008155087940394878, + 0.025237901136279106, + 0.011850276030600071, + 0.013871702365577221, + 0.005034015979617834, + 0.004623011685907841, + 0.00578313646838069, + 0.004949916619807482, + 0.004379613324999809, + 0.004601280204951763, + 0.008816253393888474, + 0.006379088386893272, + 0.013049925677478313, + 0.010539183393120766, + 0.007729654666036367, + 0.007896888069808483, + 0.0766056701540947, + 0.012365520000457764, + 0.01488625817000866, + 0.07714933902025223, + 0.01619274914264679, + 0.06157161295413971, + 0.04978259652853012, + 0.012591744773089886, + 0.006538549903780222, + 0.0043421522714197636, + 0.010144731029868126, + 0.007442600559443235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.5314405560493469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12712228298187256 + ], + "opposite_logprob_tokens": [ + -2.127122402191162 + ], + "logprob_reward": -0.12712228298187256, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12712228298187256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12679429352283478, + 0.10622423142194748, + 0.0816967561841011, + 0.03618096932768822, + 0.0390772670507431, + 0.030528120696544647, + 0.1148335188627243, + 0.12686686217784882, + 0.148935467004776, + 0.18886245787143707 + ], + "credit_map": [ + 0.012259838171303272, + 0.010641908273100853, + 0.008816886693239212, + 0.00808693002909422, + 0.015655074268579483, + 0.035770200192928314, + 0.012159262783825397, + 0.023404190316796303, + 0.012428761459887028, + 0.012219257652759552, + 0.017603201791644096, + 0.014547898434102535, + 0.010266115888953209, + 0.012827776372432709, + 0.01149066723883152, + 0.014840543270111084, + 0.009815612807869911, + 0.008647807873785496, + 0.007446145173162222, + 0.006742538418620825, + 0.012831194326281548, + 0.01142931543290615, + 0.012279880233108997, + 0.012504265643656254, + 0.0046699801459908485, + 0.0031032131519168615, + 0.003138797590509057, + 0.003437124425545335, + 0.006465151906013489, + 0.005030438303947449, + 0.0054998924024403095, + 0.004836372099816799, + 0.005799878388643265, + 0.0032459900248795748, + 0.005083686672151089, + 0.00518910214304924, + 0.004601036198437214, + 0.004671449773013592, + 0.006665878929197788, + 0.003820243990048766, + 0.0027498749550431967, + 0.0027952990494668484, + 0.0025162845849990845, + 0.002059050602838397, + 0.0038566356524825096, + 0.00922435149550438, + 0.003524431260302663, + 0.0038021921645849943, + 0.01039697602391243, + 0.010588511824607849, + 0.014547388069331646, + 0.013655664399266243, + 0.010476142168045044, + 0.018840638920664787, + 0.019509999081492424, + 0.01681818813085556, + 0.02349805273115635, + 0.011089736595749855, + 0.016477959230542183, + 0.011032425798475742, + 0.014268259517848492, + 0.012778151780366898, + 0.019413774833083153, + 0.018308501690626144, + 0.008989281952381134, + 0.006589339580386877, + 0.006309991702437401, + 0.007044453639537096, + 0.050149694085121155, + 0.0084616057574749, + 0.011657791212201118, + 0.04973331093788147, + 0.02090819925069809, + 0.05918942391872406, + 0.03659709170460701, + 0.011716421693563461, + 0.011210662312805653, + 0.0089005958288908, + 0.017161840572953224, + 0.0231782253831625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.12712228298187256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812621906399727 + ], + "opposite_logprob_tokens": [ + -3.288126230239868 + ], + "logprob_reward": -0.03812621906399727, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812621906399727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21433404088020325, + 0.07816976308822632, + 0.05126689001917839, + 0.3549445569515228, + 0.049554493278265, + 0.07276985049247742, + 0.017678817734122276, + 0.0460691899061203, + 0.05761772394180298, + 0.057594720274209976 + ], + "credit_map": [ + 0.016970807686448097, + 0.018400277942419052, + 0.013606983236968517, + 0.022682616487145424, + 0.019487859681248665, + 0.04966970160603523, + 0.02746547758579254, + 0.046050313860177994, + 0.008314530365169048, + 0.007206829264760017, + 0.008822305127978325, + 0.0076994262635707855, + 0.009721015579998493, + 0.008876786567270756, + 0.013071967288851738, + 0.01445690542459488, + 0.004065834917128086, + 0.005295469891279936, + 0.004762173630297184, + 0.005002287216484547, + 0.007614327594637871, + 0.007661865558475256, + 0.00796495471149683, + 0.008899977430701256, + 0.020940404385328293, + 0.02324504964053631, + 0.02114209532737732, + 0.02649722807109356, + 0.052125558257102966, + 0.028821585699915886, + 0.12202363461256027, + 0.06014898791909218, + 0.008874934166669846, + 0.005944838747382164, + 0.008430836722254753, + 0.005296726711094379, + 0.00709445821121335, + 0.00444242125377059, + 0.00433975737541914, + 0.0051305219531059265, + 0.008199620991945267, + 0.009700600057840347, + 0.007706977892667055, + 0.008178488351404667, + 0.007139141671359539, + 0.01026573684066534, + 0.011174792423844337, + 0.010404491797089577, + 0.001646527205593884, + 0.00144615035969764, + 0.0016866109799593687, + 0.001304584788158536, + 0.001834568683989346, + 0.004440442193299532, + 0.002735296031460166, + 0.002584638074040413, + 0.007657169830054045, + 0.004427386913448572, + 0.006307921372354031, + 0.004395014140754938, + 0.0050125159323215485, + 0.0048005725257098675, + 0.007216769736260176, + 0.006251840852200985, + 0.0033918973058462143, + 0.002610391704365611, + 0.001768468995578587, + 0.0018539992161095142, + 0.020071908831596375, + 0.0031740216072648764, + 0.0038742010947316885, + 0.020872829481959343, + 0.005468116607517004, + 0.019638609141111374, + 0.015547974966466427, + 0.004635801538825035, + 0.0029148554895073175, + 0.0021393655333667994, + 0.00381386186927557, + 0.003436134196817875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.03812621906399727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087606549263 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.016087606549263, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087606549263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10328081250190735, + 0.11282212287187576, + 0.04921629652380943, + 0.11421278864145279, + 0.18114086985588074, + 0.0293926689773798, + 0.02651742473244667, + 0.05340936779975891, + 0.25230833888053894, + 0.07769932597875595 + ], + "credit_map": [ + 0.011325293220579624, + 0.008859710767865181, + 0.008734028786420822, + 0.007110176607966423, + 0.015149391256272793, + 0.025943027809262276, + 0.009450096637010574, + 0.016709085553884506, + 0.021522877737879753, + 0.014160718768835068, + 0.01057456061244011, + 0.01115184836089611, + 0.014470292255282402, + 0.014409857802093029, + 0.00997991394251585, + 0.016552051529288292, + 0.004836097825318575, + 0.007017877884209156, + 0.004772663582116365, + 0.004981913138180971, + 0.00582511629909277, + 0.005508725997060537, + 0.007229759823530912, + 0.009044134989380836, + 0.006206776015460491, + 0.009033884853124619, + 0.007605392951518297, + 0.009504832327365875, + 0.015052027069032192, + 0.010919737629592419, + 0.03619953989982605, + 0.01969059370458126, + 0.021899985149502754, + 0.02493428811430931, + 0.017926497384905815, + 0.01501227542757988, + 0.026577217504382133, + 0.027909573167562485, + 0.018382975831627846, + 0.028498057276010513, + 0.0031594375614076853, + 0.0034227315336465836, + 0.0037984801456332207, + 0.0034561119973659515, + 0.0026718745939433575, + 0.004652464762330055, + 0.0038659176789224148, + 0.004365650936961174, + 0.00243734591640532, + 0.003496668068692088, + 0.001942696631886065, + 0.002087595406919718, + 0.0024769289884716272, + 0.006301571149379015, + 0.0035179040860384703, + 0.004256713204085827, + 0.007067202590405941, + 0.004433602560311556, + 0.009094154462218285, + 0.005909498315304518, + 0.0064528645016252995, + 0.005444608628749847, + 0.00768069364130497, + 0.007326741702854633, + 0.01477657351642847, + 0.012999298050999641, + 0.01330109965056181, + 0.015465191565454006, + 0.0786857008934021, + 0.014863393269479275, + 0.019864585250616074, + 0.08235248178243637, + 0.008990973234176636, + 0.018604105338454247, + 0.010774502530694008, + 0.005967842880636454, + 0.005840504541993141, + 0.006111558992415667, + 0.009234155528247356, + 0.012175675481557846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.016087606549263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204963102936745 + ], + "opposite_logprob_tokens": [ + -2.812049627304077 + ], + "logprob_reward": -0.06204963102936745, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204963102936745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17411917448043823, + 0.13957250118255615, + 0.0508640855550766, + 0.18286126852035522, + 0.11939271539449692, + 0.06740565598011017, + 0.08861900120973587, + 0.03452856466174126, + 0.09599415957927704, + 0.04664283245801926 + ], + "credit_map": [ + 0.02662966214120388, + 0.02332470566034317, + 0.018133342266082764, + 0.01569393277168274, + 0.021903766319155693, + 0.02217305265367031, + 0.016601121053099632, + 0.029659593477845192, + 0.024047473445534706, + 0.021168749779462814, + 0.015275012701749802, + 0.01351605448871851, + 0.01552975457161665, + 0.013834699057042599, + 0.014093312434852123, + 0.02210744470357895, + 0.005533840507268906, + 0.0045480746775865555, + 0.00611512316390872, + 0.004417717456817627, + 0.005803661420941353, + 0.006894554942846298, + 0.00951262004673481, + 0.008038492873311043, + 0.014444850385189056, + 0.016049887984991074, + 0.01538133341819048, + 0.022164905443787575, + 0.021664123982191086, + 0.01567988656461239, + 0.05096646398305893, + 0.02650982141494751, + 0.016274800524115562, + 0.009824239648878574, + 0.014174380339682102, + 0.01140676997601986, + 0.02555614337325096, + 0.011403705924749374, + 0.012378092855215073, + 0.018374580889940262, + 0.007341468241065741, + 0.01064466405659914, + 0.005836643278598785, + 0.0061972495168447495, + 0.0072946310974657536, + 0.01073425356298685, + 0.008889099583029747, + 0.010467647574841976, + 0.008160123601555824, + 0.011315609328448772, + 0.01249619945883751, + 0.010065275244414806, + 0.008631754666566849, + 0.013610593974590302, + 0.0100952023640275, + 0.014244244433939457, + 0.006902774795889854, + 0.003154950914904475, + 0.003588163061067462, + 0.004249164368957281, + 0.003552912501618266, + 0.0033819114323705435, + 0.00449984148144722, + 0.005198848433792591, + 0.005914340727031231, + 0.006387941539287567, + 0.004799334332346916, + 0.005047176033258438, + 0.02864084579050541, + 0.006346149370074272, + 0.008659973740577698, + 0.030198389664292336, + 0.005437465384602547, + 0.008612365461885929, + 0.011485033668577671, + 0.006014759186655283, + 0.004172699525952339, + 0.002998994430527091, + 0.0036307864356786013, + 0.0042907241731882095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.06204963102936745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, perched high above the sidewalk, follows a sleek, modern black-and-white police interceptor car as it speeds through rain-soaked city streets at night. A relentless downpour has slicked the asphalt to a black mirror, and the car's wide tires kick up a torrent of spray, barely maintaining traction as it hurtles forward. A discreet LED light bar atop the car's roof flickers silently, painting the wet street, the sleek car body, and the towering, dark skyscraper facades into a frenzied, pulsating flash of intense red and blue. The car's own powerful headlights carve a sharp white cone into the darkness, while the glow of distant neon signs and streetlights distorts and ripples across the wet surface and the puddles it carves, creating a palpable atmosphere imbued with a high-stakes sense of urgency and relentless pursuit.", + "seed": 31265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_2/index_7369_seed_31265_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10644639478996396, + "vlm_logprob": -0.10644639478996396, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.10644639478996396, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34115874767303467 + ], + "opposite_logprob_tokens": [ + -1.5911587476730347 + ], + "logprob_reward": -0.34115874767303467, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.34115874767303467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2967871427536011, + 0.12905003130435944, + 0.057935044169425964, + 0.07514519989490509, + 0.09431567788124084, + 0.07371701300144196, + 0.06203942000865936, + 0.0762520432472229, + 0.09225735068321228, + 0.04250112175941467 + ], + "credit_map": [ + 0.021148383617401123, + 0.017880985513329506, + 0.019224487245082855, + 0.014247744344174862, + 0.01597738265991211, + 0.09764330834150314, + 0.01327996701002121, + 0.09738486260175705, + 0.012695474550127983, + 0.009901855140924454, + 0.014036229811608791, + 0.020849784836173058, + 0.013312557712197304, + 0.01617654785513878, + 0.02165372483432293, + 0.02042384073138237, + 0.003839933080598712, + 0.008088325150310993, + 0.006208914332091808, + 0.0069242483004927635, + 0.007285373285412788, + 0.004658191930502653, + 0.013225876726210117, + 0.0077041746117174625, + 0.006391196046024561, + 0.00870977621525526, + 0.007198461797088385, + 0.01100321114063263, + 0.009574508294463158, + 0.008180554024875164, + 0.011364791542291641, + 0.01272269245237112, + 0.013583725318312645, + 0.013764268718659878, + 0.011309299618005753, + 0.015238977037370205, + 0.01207431498914957, + 0.009378097020089626, + 0.010090045630931854, + 0.00887694675475359, + 0.004803778603672981, + 0.006655495148152113, + 0.008973369374871254, + 0.00988992489874363, + 0.008797180838882923, + 0.011177788488566875, + 0.0141491973772645, + 0.009270268492400646, + 0.003607321996241808, + 0.004180248826742172, + 0.004903937689960003, + 0.0050367312505841255, + 0.005372621584683657, + 0.0029572139028459787, + 0.021815378218889236, + 0.014165962114930153, + 0.007184905931353569, + 0.010665121488273144, + 0.00752310361713171, + 0.007938710041344166, + 0.01632094942033291, + 0.009437177330255508, + 0.008426477201282978, + 0.008755587972700596, + 0.004663461819291115, + 0.0091855488717556, + 0.008082601241767406, + 0.016611691564321518, + 0.01140788197517395, + 0.00886532198637724, + 0.01767030730843544, + 0.015770526602864265, + 0.0032193935476243496, + 0.0054462868720293045, + 0.004437804687768221, + 0.009892850182950497, + 0.0038872985169291496, + 0.0043952870182693005, + 0.005792513489723206, + 0.005429686047136784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.34115874767303467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12711922824382782 + ], + "opposite_logprob_tokens": [ + -2.127119302749634 + ], + "logprob_reward": -0.12711922824382782, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12711922824382782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09289476275444031, + 0.08620098978281021, + 0.08278626948595047, + 0.055534131824970245, + 0.09321184456348419, + 0.18831214308738708, + 0.12747789919376373, + 0.04833395406603813, + 0.1476534754037857, + 0.07759449630975723 + ], + "credit_map": [ + 0.010989606380462646, + 0.010533544234931469, + 0.011852157302200794, + 0.004302366636693478, + 0.008342573419213295, + 0.020686987787485123, + 0.0057313330471515656, + 0.020456193014979362, + 0.010123628191649914, + 0.009989893063902855, + 0.011579782702028751, + 0.009572158567607403, + 0.007712999824434519, + 0.009716620668768883, + 0.0138219790533185, + 0.013683927245438099, + 0.006330521311610937, + 0.015290352515876293, + 0.00958333071321249, + 0.009822563268244267, + 0.010822949931025505, + 0.00967468973249197, + 0.008484293706715107, + 0.01277756318449974, + 0.004673592746257782, + 0.006107713561505079, + 0.006257095839828253, + 0.010197076946496964, + 0.007027344312518835, + 0.007124318741261959, + 0.007933185435831547, + 0.006213800515979528, + 0.013408335857093334, + 0.012934820726513863, + 0.011967556551098824, + 0.013774135150015354, + 0.011625212617218494, + 0.008579508401453495, + 0.011911829002201557, + 0.009010445326566696, + 0.017563408240675926, + 0.016948789358139038, + 0.023717200383543968, + 0.030205465853214264, + 0.017691969871520996, + 0.01939506269991398, + 0.03585272282361984, + 0.02693752385675907, + 0.008378836326301098, + 0.009299074299633503, + 0.012846319004893303, + 0.020722201094031334, + 0.010281205177307129, + 0.008005483075976372, + 0.022746967151761055, + 0.03519781678915024, + 0.004891696851700544, + 0.005722670815885067, + 0.006756550632417202, + 0.005090971942991018, + 0.005618234630674124, + 0.006051396951079369, + 0.006505280267447233, + 0.007697154302150011, + 0.010294563136994839, + 0.017811298370361328, + 0.009791979566216469, + 0.016423102468252182, + 0.022911179810762405, + 0.013775944709777832, + 0.029364489018917084, + 0.027280913665890694, + 0.005539305042475462, + 0.0076971943490207195, + 0.005729166325181723, + 0.010682882741093636, + 0.00867678876966238, + 0.008097609505057335, + 0.021813564002513885, + 0.009357977658510208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.12711922824382782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764929736033082 + ], + "opposite_logprob_tokens": [ + -4.8826494216918945 + ], + "logprob_reward": -0.00764929736033082, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764929736033082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5412981510162354, + 0.07340806722640991, + 0.07287253439426422, + 0.02520304173231125, + 0.04654364287853241, + 0.0576976053416729, + 0.03567370027303696, + 0.040424223989248276, + 0.07607021182775497, + 0.03080877475440502 + ], + "credit_map": [ + 0.03236117959022522, + 0.05974205583333969, + 0.06005808338522911, + 0.07044734805822372, + 0.03592970594763756, + 0.10532747954130173, + 0.06401389092206955, + 0.11341840773820877, + 0.008364634588360786, + 0.007131133694201708, + 0.007785184308886528, + 0.012233762070536613, + 0.007296364288777113, + 0.00901709869503975, + 0.010235875844955444, + 0.011344019323587418, + 0.004521854221820831, + 0.011031574569642544, + 0.007511069532483816, + 0.01345585286617279, + 0.00916364323347807, + 0.008010431192815304, + 0.00940978154540062, + 0.009768334217369556, + 0.00186077319085598, + 0.0026350347325205803, + 0.0018243371741846204, + 0.004691514186561108, + 0.003529419656842947, + 0.0036355324555188417, + 0.004262812901288271, + 0.0027636189479380846, + 0.008304007351398468, + 0.006746706552803516, + 0.007596075069159269, + 0.00637425109744072, + 0.003963605035096407, + 0.0046742986887693405, + 0.005333932116627693, + 0.0035507692955434322, + 0.004844672977924347, + 0.005999369081109762, + 0.006319723557680845, + 0.012437397614121437, + 0.0038095014169812202, + 0.004342521540820599, + 0.011301975697278976, + 0.00864244345575571, + 0.003235703567042947, + 0.0022348249331116676, + 0.004140768200159073, + 0.003939627669751644, + 0.0031173331663012505, + 0.0025614926125854254, + 0.01025352068245411, + 0.006190432235598564, + 0.003846457228064537, + 0.0050413222052156925, + 0.003961581736803055, + 0.006322809495031834, + 0.007135235238820314, + 0.00548504339531064, + 0.00431672902777791, + 0.004315050318837166, + 0.008055202662944794, + 0.010605982504785061, + 0.0067804292775690556, + 0.00956572126597166, + 0.009008059278130531, + 0.009314578026533127, + 0.009831937029957771, + 0.012908314354717731, + 0.002808838617056608, + 0.0026517712976783514, + 0.005199900362640619, + 0.0063413092866539955, + 0.0028974241577088833, + 0.002153555629774928, + 0.004347601439803839, + 0.004408374894410372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.00764929736033082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018205448985099792 + ], + "opposite_logprob_tokens": [ + -4.018205642700195 + ], + "logprob_reward": -0.018205448985099792, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018205448985099792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3572092056274414, + 0.12946976721286774, + 0.09104250371456146, + 0.048936840146780014, + 0.05861213430762291, + 0.1126767247915268, + 0.04675276204943657, + 0.05593964830040932, + 0.07452822476625443, + 0.024832159280776978 + ], + "credit_map": [ + 0.03554491698741913, + 0.03631950914859772, + 0.031170189380645752, + 0.029068442061543465, + 0.02269289270043373, + 0.09014283120632172, + 0.023031577467918396, + 0.08923884481191635, + 0.00914923008531332, + 0.012392399832606316, + 0.011999117210507393, + 0.014224918559193611, + 0.01285967044532299, + 0.02652806043624878, + 0.018482619896531105, + 0.023833760991692543, + 0.007885986939072609, + 0.012222743593156338, + 0.013910303823649883, + 0.013855761848390102, + 0.008068747818470001, + 0.007761618588119745, + 0.016741406172513962, + 0.010595934465527534, + 0.004281825851649046, + 0.005894505884498358, + 0.004487228579819202, + 0.007695481181144714, + 0.005567724816501141, + 0.006284369621425867, + 0.007214012555778027, + 0.007511699106544256, + 0.0061933123506605625, + 0.007687950041145086, + 0.0079254861921072, + 0.008325870148837566, + 0.006539484485983849, + 0.007950670085847378, + 0.007401359733194113, + 0.0065880026668310165, + 0.011785642243921757, + 0.011329139582812786, + 0.012857096269726753, + 0.021258186548948288, + 0.009778720326721668, + 0.011645738035440445, + 0.01738579198718071, + 0.016636423766613007, + 0.0037800467107445, + 0.005401524715125561, + 0.006356149446219206, + 0.00960974209010601, + 0.005387815181165934, + 0.0021851048804819584, + 0.009950748644769192, + 0.004081638064235449, + 0.005978278815746307, + 0.0073299044743180275, + 0.005713052581995726, + 0.006478371098637581, + 0.00898810289800167, + 0.005204253364354372, + 0.006834402214735746, + 0.009413285180926323, + 0.007651870138943195, + 0.011936723254621029, + 0.005861381534487009, + 0.008752549067139626, + 0.009418499656021595, + 0.006378906778991222, + 0.012850767001509666, + 0.011677534319460392, + 0.0021691927686333656, + 0.003263333113864064, + 0.002405651146546006, + 0.004165315069258213, + 0.002610744209960103, + 0.002660444239154458, + 0.004422145429998636, + 0.0031353349331766367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.018205448985099792 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380992516875267 + ], + "opposite_logprob_tokens": [ + -3.2880992889404297 + ], + "logprob_reward": -0.0380992516875267, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380992516875267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.229377880692482, + 0.2084982842206955, + 0.06145670264959335, + 0.05274597182869911, + 0.1718861609697342, + 0.06511744111776352, + 0.0601239874958992, + 0.034215666353702545, + 0.08708500862121582, + 0.029492927715182304 + ], + "credit_map": [ + 0.01536934170871973, + 0.016445918008685112, + 0.013208997435867786, + 0.010550606064498425, + 0.013259232975542545, + 0.07531470060348511, + 0.011038747616112232, + 0.07419031113386154, + 0.015579688362777233, + 0.02100514993071556, + 0.02393166907131672, + 0.018112536519765854, + 0.01950427144765854, + 0.0467798076570034, + 0.025977540761232376, + 0.03760761767625809, + 0.0050001488998532295, + 0.008641164749860764, + 0.007047853898257017, + 0.007583758793771267, + 0.007390731479972601, + 0.006177171133458614, + 0.011540309526026249, + 0.008075564168393612, + 0.004157111048698425, + 0.005730470642447472, + 0.005339531693607569, + 0.00920884869992733, + 0.007167567033320665, + 0.004631940275430679, + 0.008370083756744862, + 0.00814041867852211, + 0.018675440922379494, + 0.02578299306333065, + 0.021419936791062355, + 0.027638761326670647, + 0.02418508008122444, + 0.02030492015182972, + 0.018639100715517998, + 0.015239926986396313, + 0.005359250586479902, + 0.00597692746669054, + 0.006067683920264244, + 0.013883726671338081, + 0.004639095161110163, + 0.005965617019683123, + 0.0132962791249156, + 0.009928862564265728, + 0.00541515601798892, + 0.00494346721097827, + 0.004699453711509705, + 0.005966214928776026, + 0.005721084773540497, + 0.0035204333253204823, + 0.020128510892391205, + 0.009729673154652119, + 0.002719690790399909, + 0.0034043097402900457, + 0.004589973017573357, + 0.0033421004191040993, + 0.004915099591016769, + 0.004434425849467516, + 0.005534203257411718, + 0.005275865085422993, + 0.011760805733501911, + 0.008635216392576694, + 0.007256579119712114, + 0.009870593436062336, + 0.009800637140870094, + 0.008251640014350414, + 0.016950378194451332, + 0.014559153467416763, + 0.0020621148869395256, + 0.0026033218018710613, + 0.002654341049492359, + 0.0058756121434271336, + 0.0033416070509701967, + 0.0033471896313130856, + 0.00601525092497468, + 0.0035934888292104006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0380992516875267 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, perched high above the sidewalk, follows a sleek, modern black-and-white police interceptor car as it speeds through rain-soaked city streets at night. A relentless downpour has slicked the asphalt to a black mirror, and the car's wide tires kick up a torrent of spray, barely maintaining traction as it hurtles forward. A discreet LED light bar atop the car's roof flickers silently, painting the wet street, the sleek car body, and the towering, dark skyscraper facades into a frenzied, pulsating flash of intense red and blue. The car's own powerful headlights carve a sharp white cone into the darkness, while the glow of distant neon signs and streetlights distorts and ripples across the wet surface and the puddles it carves, creating a palpable atmosphere imbued with a high-stakes sense of urgency and relentless pursuit.", + "seed": 31266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_2/index_7369_seed_31266_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10927189998328686, + "vlm_logprob": -0.10927189998328686, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.10927189998328686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3111312985420227 + ], + "opposite_logprob_tokens": [ + -1.936131238937378 + ], + "logprob_reward": -0.3111312985420227, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.3111312985420227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17033793032169342, + 0.13890700042247772, + 0.08058220893144608, + 0.11448481678962708, + 0.04493814706802368, + 0.02576737105846405, + 0.07519230246543884, + 0.20023910701274872, + 0.03543544188141823, + 0.11411574482917786 + ], + "credit_map": [ + 0.02693837694823742, + 0.02180095948278904, + 0.019168764352798462, + 0.010149803012609482, + 0.023674940690398216, + 0.03065980225801468, + 0.012733698822557926, + 0.025211574509739876, + 0.009808176197111607, + 0.00845828466117382, + 0.036701470613479614, + 0.006153909955173731, + 0.047943323850631714, + 0.011216975748538971, + 0.008733323775231838, + 0.00989152304828167, + 0.009713804349303246, + 0.009074073284864426, + 0.015860961750149727, + 0.010233582928776741, + 0.007007263135164976, + 0.010026824660599232, + 0.009314391762018204, + 0.009351297281682491, + 0.02494593895971775, + 0.017142150551080704, + 0.010333802551031113, + 0.017841648310422897, + 0.015201196074485779, + 0.009301146492362022, + 0.008868981152772903, + 0.010849938727915287, + 0.008151767775416374, + 0.0033706033136695623, + 0.00794400554150343, + 0.006232111249119043, + 0.0040198941715061665, + 0.002187645062804222, + 0.010275392793118954, + 0.0027567255310714245, + 0.0032558655366301537, + 0.0020975463557988405, + 0.0051478729583323, + 0.003416437190026045, + 0.0022953066509217024, + 0.003128602635115385, + 0.003104676026850939, + 0.0033210606779903173, + 0.007542167324572802, + 0.007343437988311052, + 0.005195098929107189, + 0.00938546285033226, + 0.014682917855679989, + 0.01142963394522667, + 0.010358107276260853, + 0.009255469776690006, + 0.011099273338913918, + 0.017864612862467766, + 0.017508331686258316, + 0.03798739239573479, + 0.018314572051167488, + 0.034582655876874924, + 0.045603375881910324, + 0.017278878018260002, + 0.003566860919818282, + 0.005104691255837679, + 0.00398276187479496, + 0.00575980544090271, + 0.0034780832938849926, + 0.004846231546252966, + 0.005623796489089727, + 0.0030732089653611183, + 0.011444712989032269, + 0.013773721642792225, + 0.00924905575811863, + 0.01502914633601904, + 0.012352976948022842, + 0.02230953238904476, + 0.012366654351353645, + 0.017589930444955826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.3111312985420227 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16043472290039062 + ], + "opposite_logprob_tokens": [ + -1.9104347229003906 + ], + "logprob_reward": -0.16043472290039062, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16043472290039062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13160432875156403, + 0.17667652666568756, + 0.03709322586655617, + 0.04555356502532959, + 0.04746561124920845, + 0.02948286011815071, + 0.1485193967819214, + 0.21706174314022064, + 0.03951498121023178, + 0.127027690410614 + ], + "credit_map": [ + 0.021561399102211, + 0.01547948271036148, + 0.02036416158080101, + 0.007855953648686409, + 0.01730768382549286, + 0.019897008314728737, + 0.01171835046261549, + 0.017420293763279915, + 0.013416478410363197, + 0.014174179174005985, + 0.04421577602624893, + 0.009549484588205814, + 0.051801647990942, + 0.014853200875222683, + 0.01433489192277193, + 0.014330889098346233, + 0.003261821810156107, + 0.005165844224393368, + 0.00522028049454093, + 0.0039512342773377895, + 0.004905498120933771, + 0.005797939840704203, + 0.004426027648150921, + 0.004364583175629377, + 0.010431252419948578, + 0.004398588556796312, + 0.0044484385289251804, + 0.007444914896041155, + 0.0047095404006540775, + 0.004377656150609255, + 0.004432498477399349, + 0.005310682114213705, + 0.006591194309294224, + 0.005883507896214724, + 0.010037547908723354, + 0.007358423434197903, + 0.004809274338185787, + 0.003015735652297735, + 0.006602851208299398, + 0.003167081391438842, + 0.0038475554902106524, + 0.002395915798842907, + 0.005366311874240637, + 0.0036324262619018555, + 0.002613561460748315, + 0.004000325221568346, + 0.0037387923803180456, + 0.0038879758212715387, + 0.015185440890491009, + 0.01454982627183199, + 0.011818524450063705, + 0.01808539591729641, + 0.027123762294650078, + 0.02218780294060707, + 0.018427548930048943, + 0.021141108125448227, + 0.012601474300026894, + 0.019716043025255203, + 0.018001804128289223, + 0.047797031700611115, + 0.016905980184674263, + 0.037620026618242264, + 0.04811842367053032, + 0.016300980001688004, + 0.0038231308571994305, + 0.003411425743252039, + 0.004139284137636423, + 0.00875236839056015, + 0.004177270457148552, + 0.005145443603396416, + 0.005643785931169987, + 0.004422274883836508, + 0.010504535399377346, + 0.012141428887844086, + 0.011247726157307625, + 0.013921819627285004, + 0.012172686867415905, + 0.02913033962249756, + 0.017761897295713425, + 0.02014726772904396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.16043472290039062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016094176098704338 + ], + "opposite_logprob_tokens": [ + -4.141094207763672 + ], + "logprob_reward": -0.016094176098704338, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016094176098704338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3631385862827301, + 0.06751902401447296, + 0.10539659112691879, + 0.1193145141005516, + 0.028432361781597137, + 0.06521617621183395, + 0.0548577755689621, + 0.13489311933517456, + 0.01868928037583828, + 0.042542532086372375 + ], + "credit_map": [ + 0.040763430297374725, + 0.02725537307560444, + 0.05287798494100571, + 0.019482798874378204, + 0.06904003024101257, + 0.06543266028165817, + 0.03040165826678276, + 0.05788464471697807, + 0.007322678342461586, + 0.005635241977870464, + 0.00868978537619114, + 0.006659017410129309, + 0.009851129725575447, + 0.00834561325609684, + 0.013759331777691841, + 0.00725622707977891, + 0.012421929277479649, + 0.014597326517105103, + 0.010593874379992485, + 0.012680171057581902, + 0.012674259021878242, + 0.014573757536709309, + 0.01173042505979538, + 0.016124850139021873, + 0.021465003490447998, + 0.018553853034973145, + 0.010347796604037285, + 0.017885152250528336, + 0.018078798428177834, + 0.010507871396839619, + 0.010946097783744335, + 0.01152995228767395, + 0.004372410476207733, + 0.0034327744506299496, + 0.006198801565915346, + 0.0033865943551063538, + 0.0021544815972447395, + 0.0018543321639299393, + 0.005132976919412613, + 0.0018999901367351413, + 0.005251327995210886, + 0.004794745706021786, + 0.009846304543316364, + 0.008284825831651688, + 0.004505648743361235, + 0.005517866462469101, + 0.013353079557418823, + 0.013662372715771198, + 0.006146719213575125, + 0.004453153349459171, + 0.0038657807745039463, + 0.004408902954310179, + 0.010562826879322529, + 0.010009546764194965, + 0.006809952203184366, + 0.008600895293056965, + 0.0115241389721632, + 0.014713633805513382, + 0.012409720569849014, + 0.025592811405658722, + 0.01030666846781969, + 0.022795528173446655, + 0.027847781777381897, + 0.009702833369374275, + 0.002666879678145051, + 0.002927028574049473, + 0.0012345616705715656, + 0.0026213410310447216, + 0.0014336335007101297, + 0.002466857898980379, + 0.0028780940920114517, + 0.0024608836974948645, + 0.00461059994995594, + 0.004576499108225107, + 0.003921234980225563, + 0.007218313869088888, + 0.00421857088804245, + 0.006488163955509663, + 0.004498734604567289, + 0.0070104137994349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.016094176098704338 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059982344508171 + ], + "opposite_logprob_tokens": [ + -3.895599842071533 + ], + "logprob_reward": -0.02059982344508171, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059982344508171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12093005329370499, + 0.08423825353384018, + 0.04005071893334389, + 0.11403926461935043, + 0.10323350131511688, + 0.11982066184282303, + 0.1319618672132492, + 0.17311346530914307, + 0.06641051173210144, + 0.0462016724050045 + ], + "credit_map": [ + 0.013690091669559479, + 0.011009260080754757, + 0.023286258801817894, + 0.008080714382231236, + 0.017414050176739693, + 0.018512725830078125, + 0.01356310024857521, + 0.01537385955452919, + 0.006917559541761875, + 0.004459203686565161, + 0.021311791613698006, + 0.004533485509455204, + 0.02726849354803562, + 0.008188012056052685, + 0.005340336821973324, + 0.006219369359314442, + 0.005411765072494745, + 0.006224179174751043, + 0.006604285910725594, + 0.00457728561013937, + 0.0038280077278614044, + 0.005852012429386377, + 0.0028302811551839113, + 0.004722902085632086, + 0.01451539434492588, + 0.02012641727924347, + 0.008505579084157944, + 0.017992591485381126, + 0.02162613905966282, + 0.01171203888952732, + 0.01018053013831377, + 0.009380573406815529, + 0.013829013332724571, + 0.015522864647209644, + 0.016632605344057083, + 0.014896461740136147, + 0.01211303286254406, + 0.0065843635238707066, + 0.015588197857141495, + 0.008066954091191292, + 0.011799032799899578, + 0.009001261554658413, + 0.017155326902866364, + 0.011947592720389366, + 0.010461275465786457, + 0.014239594340324402, + 0.02226163074374199, + 0.022954944521188736, + 0.012528068386018276, + 0.01313866302371025, + 0.010903092101216316, + 0.015452170744538307, + 0.027419313788414, + 0.019478127360343933, + 0.016750190407037735, + 0.016292251646518707, + 0.010167506523430347, + 0.01321779191493988, + 0.010187377221882343, + 0.02783372439444065, + 0.018656708300113678, + 0.03182070329785347, + 0.04713490605354309, + 0.014094728045165539, + 0.005596289876848459, + 0.0058579170145094395, + 0.008776628412306309, + 0.01198156364262104, + 0.0067320954985916615, + 0.008009213022887707, + 0.013779028318822384, + 0.005677772220224142, + 0.004711406771093607, + 0.00648094667121768, + 0.00451316824182868, + 0.007134930696338415, + 0.004561868496239185, + 0.005214334465563297, + 0.006597800180315971, + 0.0069872173480689526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.02059982344508171 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking up at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot that follows a black-and-white police interceptor car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving through rain-soaked city streets at night with a downpour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car have a flickering LED light bar on the roof emitting red and blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's headlights carving a sharp white cone of light into the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809947893023491 + ], + "opposite_logprob_tokens": [ + -3.288099527359009 + ], + "logprob_reward": -0.03809947893023491, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809947893023491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09075631946325302, + 0.08153978735208511, + 0.025633862242102623, + 0.09002276510000229, + 0.040085650980472565, + 0.07422937452793121, + 0.10587871819734573, + 0.3524826765060425, + 0.049698878079652786, + 0.08967206627130508 + ], + "credit_map": [ + 0.011313294991850853, + 0.008264860138297081, + 0.013681096956133842, + 0.005663394927978516, + 0.012210319750010967, + 0.01701609417796135, + 0.008017454296350479, + 0.014589792117476463, + 0.008818577975034714, + 0.006870574317872524, + 0.01626875065267086, + 0.007910650223493576, + 0.013983724638819695, + 0.008878616616129875, + 0.00984209030866623, + 0.008966786786913872, + 0.0025490624830126762, + 0.0031330271158367395, + 0.0031660329550504684, + 0.00260910508222878, + 0.0037250369787216187, + 0.00393420085310936, + 0.003338261041790247, + 0.003179132239893079, + 0.0158010832965374, + 0.012299076654016972, + 0.00893966481089592, + 0.017421254888176918, + 0.009616672992706299, + 0.008812678046524525, + 0.006779599003493786, + 0.010352732613682747, + 0.006841249763965607, + 0.0052012065425515175, + 0.004726060200482607, + 0.005671583116054535, + 0.004209940321743488, + 0.0020489441230893135, + 0.008722270838916302, + 0.0026643916498869658, + 0.004750883672386408, + 0.004090949427336454, + 0.0032184079755097628, + 0.006472354754805565, + 0.005738967563956976, + 0.007777331862598658, + 0.02082435041666031, + 0.02135612443089485, + 0.010398361831903458, + 0.00808256957679987, + 0.006409044843167067, + 0.010496113449335098, + 0.011211700737476349, + 0.025582393631339073, + 0.015320738777518272, + 0.018377793952822685, + 0.026520030573010445, + 0.0332406610250473, + 0.04257696121931076, + 0.07218058407306671, + 0.028127521276474, + 0.05458926782011986, + 0.06283199787139893, + 0.03241559490561485, + 0.0047896443866193295, + 0.005322562996298075, + 0.006923039443790913, + 0.008272320963442326, + 0.004416428040713072, + 0.006728373933583498, + 0.005750932265073061, + 0.0074955737218260765, + 0.005702431313693523, + 0.007250659167766571, + 0.005875797010958195, + 0.008940528146922588, + 0.012809768319129944, + 0.024764282628893852, + 0.011504283174872398, + 0.01282430812716484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.03809947893023491 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_3/index_tensor([2578, 2578])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_3/index_tensor([2578, 2578])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6d9d39945b1bb534fb4edbffbdf00b52b67970dc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_3/index_tensor([2578, 2578])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle, high-speed chase shot follows a young man on rollerblades as he engages a massive, fire-breathing dragon covered in obsidian scales. The young man's long, raven hair, tied in a high ponytail, whips wildly in the wind, his eyes piercing his steely face. He wears a sleeveless, bright red silk battle armor embroidered with golden flames, his bare arms muscular and defined. Beneath his feet, two rollerblades of bronze and flame spin at high speed, emitting dazzling sparks and allowing him to nimbly pivot and hover in the air. Just as the dragon coils around the serpentine's body, its maw opens, unleashing a devastating torrent of blazing orange-yellow flames. The young man reacts swiftly, leaning sideways and performing a breathtaking acrobatic maneuver to avoid the blaze. His immensely swift movements left a trail of faintly transparent afterimages of red clothing and golden flames in the twilight sky, a blend of deep purple and indigo, perfectly illustrating the ultimate sense of speed. The entire battlefield was situated atop a precipitous mountain range, with a mist-shrouded abyss below. The distant horizon burned with the afterglow of the setting sun, casting long, dramatic shadows over this epic confrontation. The air was filled with the tense atmosphere of scorching heat and clashing energy.", + "seed": 17560, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_3/index_2578_seed_17560_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10278991647064686, + "vlm_logprob": -0.10278991647064686, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.10278991647064686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11603289097547531 + ], + "opposite_logprob_tokens": [ + -2.241032838821411 + ], + "logprob_reward": -0.11603289097547531, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11603289097547531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19650280475616455, + 0.0870203897356987, + 0.05255494639277458, + 0.11413238197565079, + 0.11270235478878021, + 0.1654685139656067, + 0.07790149003267288, + 0.02927924506366253, + 0.10927577316761017, + 0.055162109434604645 + ], + "credit_map": [ + 0.022838067263364792, + 0.01977386325597763, + 0.024665864184498787, + 0.03128562495112419, + 0.019991284236311913, + 0.02200031653046608, + 0.01965661719441414, + 0.03629116714000702, + 0.008475648239254951, + 0.01044220756739378, + 0.009175212122499943, + 0.009501136839389801, + 0.017048180103302002, + 0.0077793169766664505, + 0.011911841109395027, + 0.01268684957176447, + 0.005221491679549217, + 0.00485047185793519, + 0.005546536762267351, + 0.006070781033486128, + 0.008070467039942741, + 0.0059813340194523335, + 0.009764746762812138, + 0.007049116771668196, + 0.010942391119897366, + 0.018678445369005203, + 0.008701271377503872, + 0.008919807150959969, + 0.015474445186555386, + 0.011012077331542969, + 0.01335323229432106, + 0.027050713077187538, + 0.011770885437726974, + 0.010739779099822044, + 0.0089922696352005, + 0.010262704454362392, + 0.013815522193908691, + 0.013012930750846863, + 0.018949760124087334, + 0.02515849471092224, + 0.01564875990152359, + 0.029656602069735527, + 0.015242199413478374, + 0.013134629465639591, + 0.024378575384616852, + 0.018197884783148766, + 0.023795267567038536, + 0.02541460283100605, + 0.009328971616923809, + 0.007614596746861935, + 0.015528319403529167, + 0.007633738685399294, + 0.007141055539250374, + 0.00811315793544054, + 0.007688784971833229, + 0.014852864667773247, + 0.004175444133579731, + 0.0038520386442542076, + 0.002801383612677455, + 0.0030551610980182886, + 0.004711453337222338, + 0.0037011962849646807, + 0.0035859476774930954, + 0.00339661817997694, + 0.01592034474015236, + 0.01413413230329752, + 0.011915035545825958, + 0.009525389410555363, + 0.01742498017847538, + 0.013118878938257694, + 0.009437786415219307, + 0.017799224704504013, + 0.004994395188987255, + 0.009883628226816654, + 0.0059460122138261795, + 0.00496545759961009, + 0.009000136516988277, + 0.005875865463167429, + 0.0064001441933214664, + 0.00809647049754858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.11603289097547531 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133079409599304 + ], + "opposite_logprob_tokens": [ + -1.4063308238983154 + ], + "logprob_reward": -0.28133079409599304, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133079409599304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29234039783477783, + 0.10953628271818161, + 0.047288209199905396, + 0.1282496452331543, + 0.09890054911375046, + 0.07975746691226959, + 0.06673313677310944, + 0.10628848522901535, + 0.029407670721411705, + 0.04149813950061798 + ], + "credit_map": [ + 0.03741178289055824, + 0.03222498670220375, + 0.02339990623295307, + 0.03564002737402916, + 0.04979242384433746, + 0.032945018261671066, + 0.045570291578769684, + 0.03535596281290054, + 0.01214565895497799, + 0.015047060325741768, + 0.01681831106543541, + 0.009391299448907375, + 0.016868116334080696, + 0.0124750267714262, + 0.01253066211938858, + 0.014260156080126762, + 0.006415648385882378, + 0.007047475315630436, + 0.005446561612188816, + 0.0053676944226026535, + 0.007333802990615368, + 0.004626820795238018, + 0.0058447448536753654, + 0.005205465946346521, + 0.01737453043460846, + 0.016712887212634087, + 0.01071122381836176, + 0.017581786960363388, + 0.013161038979887962, + 0.013800900429487228, + 0.01906636543571949, + 0.019840924069285393, + 0.007592099718749523, + 0.0120325842872262, + 0.011768965981900692, + 0.008333534933626652, + 0.017650118097662926, + 0.009577394463121891, + 0.017491204664111137, + 0.014454647898674011, + 0.009873677045106888, + 0.011860030703246593, + 0.008611176162958145, + 0.00596949877217412, + 0.010976895689964294, + 0.009114066138863564, + 0.011149827390909195, + 0.012202299199998379, + 0.005656613036990166, + 0.009539368562400341, + 0.008242171257734299, + 0.0048944600857794285, + 0.009475767612457275, + 0.0076115867123007774, + 0.007536064367741346, + 0.013777107931673527, + 0.014166321605443954, + 0.012974034994840622, + 0.008725930005311966, + 0.010468426160514355, + 0.01969723030924797, + 0.013499673455953598, + 0.012644214555621147, + 0.014112661592662334, + 0.002809937112033367, + 0.0031700844410806894, + 0.004271144978702068, + 0.003411957062780857, + 0.00395199004560709, + 0.0035739061422646046, + 0.0037366505712270737, + 0.004482002928853035, + 0.004286294803023338, + 0.007868481799960136, + 0.003679143264889717, + 0.00410673301666975, + 0.0059613133780658245, + 0.0046094381250441074, + 0.004871246870607138, + 0.006115491967648268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.28133079409599304 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032056272029877 + ], + "opposite_logprob_tokens": [ + -2.35032057762146 + ], + "logprob_reward": -0.10032056272029877, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032056272029877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29132673144340515, + 0.1261770874261856, + 0.06315824389457703, + 0.06709367036819458, + 0.14689238369464874, + 0.07961715757846832, + 0.07772724330425262, + 0.03254515677690506, + 0.033773504197597504, + 0.08168881386518478 + ], + "credit_map": [ + 0.03914529085159302, + 0.030313653871417046, + 0.039971549063920975, + 0.031721509993076324, + 0.04170968383550644, + 0.026901721954345703, + 0.030872521921992302, + 0.05069081857800484, + 0.012101458385586739, + 0.012209858745336533, + 0.016258539631962776, + 0.01227569580078125, + 0.020259292796254158, + 0.011983986012637615, + 0.01927962526679039, + 0.021808626130223274, + 0.0099501246586442, + 0.009233902208507061, + 0.007000304758548737, + 0.009068114683032036, + 0.008855928666889668, + 0.005113512743264437, + 0.007105575408786535, + 0.006830777507275343, + 0.006393119692802429, + 0.012953443452715874, + 0.008642980828881264, + 0.007153072394430637, + 0.007910146377980709, + 0.0058144512586295605, + 0.00981197226792574, + 0.00841448362916708, + 0.012482920661568642, + 0.015955476090312004, + 0.011297150515019894, + 0.014959953725337982, + 0.019989389926195145, + 0.014535315334796906, + 0.03598668426275253, + 0.021685490384697914, + 0.008003247901797295, + 0.013536334969103336, + 0.006581578869372606, + 0.008415534161031246, + 0.012476353906095028, + 0.009628468193113804, + 0.011688265018165112, + 0.009287380613386631, + 0.009682376869022846, + 0.010269751772284508, + 0.011794141493737698, + 0.008024577982723713, + 0.008583958260715008, + 0.0073380437679588795, + 0.01111410092562437, + 0.010920294560492039, + 0.004970215726643801, + 0.003336007008329034, + 0.002765652257949114, + 0.004669988993555307, + 0.005313916131854057, + 0.002451069187372923, + 0.004918020684272051, + 0.0041202884167432785, + 0.003840321209281683, + 0.0035967195872217417, + 0.006412432063370943, + 0.003874275367707014, + 0.0032279561273753643, + 0.004906405229121447, + 0.00395694375038147, + 0.003958451095968485, + 0.01161429937928915, + 0.010613097809255123, + 0.009363343007862568, + 0.0065049948170781136, + 0.010489227250218391, + 0.00665633287280798, + 0.010414847172796726, + 0.016032669693231583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.10032056272029877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628594525158405 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628594525158405, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628594525158405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22975753247737885, + 0.07165650278329849, + 0.08526496589183807, + 0.06841273605823517, + 0.06524623930454254, + 0.06603378057479858, + 0.08590587973594666, + 0.04909238964319229, + 0.039291586726903915, + 0.23933839797973633 + ], + "credit_map": [ + 0.023241456598043442, + 0.021452296525239944, + 0.03643094375729561, + 0.04167664423584938, + 0.03191808983683586, + 0.02580415830016136, + 0.02660767361521721, + 0.022626250982284546, + 0.00768706202507019, + 0.010354745201766491, + 0.011168926022946835, + 0.007423636969178915, + 0.010916478000581264, + 0.007414199877530336, + 0.00657265167683363, + 0.010118802078068256, + 0.013845515437424183, + 0.010487270541489124, + 0.01209108904004097, + 0.008777591399848461, + 0.009941084310412407, + 0.008810390718281269, + 0.008269589394330978, + 0.013042430393397808, + 0.006200413219630718, + 0.006699877791106701, + 0.006550947669893503, + 0.005311574786901474, + 0.008918434381484985, + 0.00740535045042634, + 0.008032030425965786, + 0.01929411105811596, + 0.006334095261991024, + 0.005929735489189625, + 0.005041010212153196, + 0.006224570795893669, + 0.008212873712182045, + 0.0070130666717886925, + 0.012035589665174484, + 0.014455297961831093, + 0.006023175083100796, + 0.008226273581385612, + 0.010875659994781017, + 0.004413478542119265, + 0.009239953942596912, + 0.009439361281692982, + 0.00871698372066021, + 0.009098894894123077, + 0.008173801004886627, + 0.014415399171411991, + 0.013373184949159622, + 0.008570722304284573, + 0.008944718167185783, + 0.010841568931937218, + 0.008721139281988144, + 0.012865345925092697, + 0.005252571776509285, + 0.00582532025873661, + 0.0036864466965198517, + 0.0060603381134569645, + 0.010150346904993057, + 0.004736843053251505, + 0.0073206182569265366, + 0.00605990132316947, + 0.005060408730059862, + 0.005295074079185724, + 0.006125690881162882, + 0.004139311611652374, + 0.004567581694573164, + 0.004666931927204132, + 0.0036751606967300177, + 0.005761428736150265, + 0.04183828830718994, + 0.05732575058937073, + 0.023581014946103096, + 0.014416255988180637, + 0.026310283690690994, + 0.03447745367884636, + 0.012837065383791924, + 0.028552286326885223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.007628594525158405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19159750640392303, + 0.13495352864265442, + 0.1219407245516777, + 0.15230002999305725, + 0.05556624382734299, + 0.04698652774095535, + 0.14160574972629547, + 0.05305681377649307, + 0.052113719284534454, + 0.04987914487719536 + ], + "credit_map": [ + 0.026536710560321808, + 0.017006775364279747, + 0.02863590605556965, + 0.023943841457366943, + 0.02054208144545555, + 0.022860294207930565, + 0.016577797010540962, + 0.03549409285187721, + 0.015148745849728584, + 0.016201229766011238, + 0.022084930911660194, + 0.014532055705785751, + 0.018246324732899666, + 0.01463512796908617, + 0.015398859977722168, + 0.01870625466108322, + 0.016846906393766403, + 0.015400927513837814, + 0.012711671181023121, + 0.013192196376621723, + 0.027355533093214035, + 0.010974274016916752, + 0.013464919291436672, + 0.011994286440312862, + 0.016651928424835205, + 0.013316255994141102, + 0.015568497590720654, + 0.01676032319664955, + 0.021367482841014862, + 0.015540399588644505, + 0.020911965519189835, + 0.03218318521976471, + 0.007936528883874416, + 0.007037971168756485, + 0.005837553180754185, + 0.005920757073909044, + 0.008069501258432865, + 0.005589364096522331, + 0.009038868360221386, + 0.006135699804872274, + 0.005461023189127445, + 0.005775435362011194, + 0.005168251693248749, + 0.003790854709222913, + 0.006833722349256277, + 0.004810110200196505, + 0.008107035420835018, + 0.007040091790258884, + 0.01538907177746296, + 0.021171312779188156, + 0.020570171996951103, + 0.013606298714876175, + 0.014486851170659065, + 0.01548495702445507, + 0.021399682387709618, + 0.019497400149703026, + 0.007606767117977142, + 0.005975182633846998, + 0.005730901379138231, + 0.005757705774158239, + 0.010795298963785172, + 0.004352221731096506, + 0.0065842908807098866, + 0.006254442501813173, + 0.006248390767723322, + 0.00592474639415741, + 0.006428142078220844, + 0.0072759599424898624, + 0.005838141776621342, + 0.0069421385414898396, + 0.007065000478178263, + 0.006391199771314859, + 0.006605525966733694, + 0.0069274576380848885, + 0.006295253988355398, + 0.004900592379271984, + 0.005427696742117405, + 0.0037657886277884245, + 0.007061683107167482, + 0.0088951475918293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.008636740036308765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, high-speed chase shot follows a young man on rollerblades as he engages a massive, fire-breathing dragon covered in obsidian scales. The young man's long, raven hair, tied in a high ponytail, whips wildly in the wind, his eyes piercing his steely face. He wears a sleeveless, bright red silk battle armor embroidered with golden flames, his bare arms muscular and defined. Beneath his feet, two rollerblades of bronze and flame spin at high speed, emitting dazzling sparks and allowing him to nimbly pivot and hover in the air. Just as the dragon coils around the serpentine's body, its maw opens, unleashing a devastating torrent of blazing orange-yellow flames. The young man reacts swiftly, leaning sideways and performing a breathtaking acrobatic maneuver to avoid the blaze. His immensely swift movements left a trail of faintly transparent afterimages of red clothing and golden flames in the twilight sky, a blend of deep purple and indigo, perfectly illustrating the ultimate sense of speed. The entire battlefield was situated atop a precipitous mountain range, with a mist-shrouded abyss below. The distant horizon burned with the afterglow of the setting sun, casting long, dramatic shadows over this epic confrontation. The air was filled with the tense atmosphere of scorching heat and clashing energy.", + "seed": 17561, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_3/index_2578_seed_17561_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.050905944779515266, + "vlm_logprob": -0.050905944779515266, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.050905944779515266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08067251741886139 + ], + "opposite_logprob_tokens": [ + -2.5806725025177 + ], + "logprob_reward": -0.08067251741886139, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08067251741886139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14149607717990875, + 0.19027327001094818, + 0.09824471175670624, + 0.07303397357463837, + 0.10097278654575348, + 0.07987333089113235, + 0.15439574420452118, + 0.03621075302362442, + 0.07477949559688568, + 0.05071986839175224 + ], + "credit_map": [ + 0.016546035185456276, + 0.016561849042773247, + 0.014794019982218742, + 0.017293205484747887, + 0.018867501989006996, + 0.01063209492713213, + 0.024881329387426376, + 0.021920034661889076, + 0.01694345474243164, + 0.040687572211027145, + 0.02299419976770878, + 0.0249598640948534, + 0.028245016932487488, + 0.02189427614212036, + 0.017057429999113083, + 0.017491469159722328, + 0.010307470336556435, + 0.011500361375510693, + 0.017848219722509384, + 0.014269031584262848, + 0.009326799772679806, + 0.010991394519805908, + 0.009950877167284489, + 0.01405055820941925, + 0.008063027635216713, + 0.009524167515337467, + 0.008632170036435127, + 0.009718186222016811, + 0.009048939682543278, + 0.013337953947484493, + 0.006678091362118721, + 0.008031442761421204, + 0.008999969810247421, + 0.00792766734957695, + 0.009360835887491703, + 0.008551464416086674, + 0.011174764484167099, + 0.01204850897192955, + 0.028469178825616837, + 0.014440391212701797, + 0.011231831274926662, + 0.013364477083086967, + 0.009255730547010899, + 0.01141687948256731, + 0.008736439980566502, + 0.008392110466957092, + 0.009447257034480572, + 0.008028602227568626, + 0.014942428097128868, + 0.014873020350933075, + 0.013724526390433311, + 0.022522253915667534, + 0.02015487663447857, + 0.03071032091975212, + 0.01596360094845295, + 0.02150472067296505, + 0.003792746923863888, + 0.005715003702789545, + 0.0037439160514622927, + 0.0049583096988499165, + 0.003370381658896804, + 0.006596718914806843, + 0.004281226079910994, + 0.003752450691536069, + 0.009018921293318272, + 0.009974407963454723, + 0.010906614363193512, + 0.014321643859148026, + 0.00748602207750082, + 0.006988633889704943, + 0.010101496241986752, + 0.005981755442917347, + 0.005390905309468508, + 0.004272105637937784, + 0.007507883477956057, + 0.007817050442099571, + 0.010128128342330456, + 0.005695770028978586, + 0.004864661954343319, + 0.005043358542025089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.08067251741886139 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293522268533707 + ], + "opposite_logprob_tokens": [ + -2.2379353046417236 + ], + "logprob_reward": -0.11293522268533707, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293522268533707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10623706132173538, + 0.11028523743152618, + 0.15393081307411194, + 0.2050248235464096, + 0.03634413331747055, + 0.07177207618951797, + 0.09978726506233215, + 0.050705697387456894, + 0.1363932341337204, + 0.029519611969590187 + ], + "credit_map": [ + 0.012102962471544743, + 0.015269366092979908, + 0.015368561260402203, + 0.02157197892665863, + 0.01275180745869875, + 0.00914495438337326, + 0.011103891767561436, + 0.008923552930355072, + 0.012469420209527016, + 0.019510982558131218, + 0.016814541071653366, + 0.015548360534012318, + 0.01648717373609543, + 0.010585128329694271, + 0.008284268900752068, + 0.010585363954305649, + 0.017475638538599014, + 0.01625264249742031, + 0.026068562641739845, + 0.019579293206334114, + 0.015476571395993233, + 0.031567636877298355, + 0.01154376845806837, + 0.015966705977916718, + 0.01571517065167427, + 0.0374784842133522, + 0.022576721385121346, + 0.02572210133075714, + 0.033371489495038986, + 0.02839585952460766, + 0.019625993445515633, + 0.022139012813568115, + 0.0030478225089609623, + 0.005159272346645594, + 0.0037336796522140503, + 0.004199513234198093, + 0.0052878521382808685, + 0.0037889075465500355, + 0.006411097943782806, + 0.0047159879468381405, + 0.0065992288291454315, + 0.01084296777844429, + 0.01223619095981121, + 0.008241276256740093, + 0.00861691776663065, + 0.00770543934777379, + 0.008358052000403404, + 0.00917200930416584, + 0.0114522073417902, + 0.010110430419445038, + 0.008644330315291882, + 0.012705273926258087, + 0.012258660979568958, + 0.020868901163339615, + 0.010196978226304054, + 0.013550489209592342, + 0.004882280249148607, + 0.007935740984976292, + 0.004892755765467882, + 0.007768387906253338, + 0.0052812048234045506, + 0.008719218894839287, + 0.005735073238611221, + 0.005491042044013739, + 0.013955045491456985, + 0.021281564608216286, + 0.01747765764594078, + 0.02197209745645523, + 0.013708099722862244, + 0.013879998587071896, + 0.019828103482723236, + 0.014290675520896912, + 0.002330541145056486, + 0.003467146772891283, + 0.003537080017849803, + 0.0038351244293153286, + 0.0056876735761761665, + 0.003351729130372405, + 0.004440706688910723, + 0.0028696120716631413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.11293522268533707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866068437695503 + ], + "opposite_logprob_tokens": [ + -3.0486607551574707 + ], + "logprob_reward": -0.04866068437695503, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866068437695503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10272571444511414, + 0.18971790373325348, + 0.1967523694038391, + 0.04652295261621475, + 0.16068127751350403, + 0.09238936752080917, + 0.07401302456855774, + 0.028933119028806686, + 0.06352972984313965, + 0.04473461955785751 + ], + "credit_map": [ + 0.014538918621838093, + 0.012385801412165165, + 0.013194327242672443, + 0.014765468426048756, + 0.009306381456553936, + 0.009990043006837368, + 0.01524475496262312, + 0.013300007209181786, + 0.01340128667652607, + 0.037440601736307144, + 0.024208281189203262, + 0.03103407844901085, + 0.02917032316327095, + 0.01907431334257126, + 0.018298722803592682, + 0.01709027774631977, + 0.019435079768300056, + 0.03149370476603508, + 0.02778705582022667, + 0.024833841249346733, + 0.020932428538799286, + 0.030891599133610725, + 0.015178564935922623, + 0.026200076565146446, + 0.005759428720921278, + 0.00694799330085516, + 0.004331407602876425, + 0.004925890825688839, + 0.008414073847234249, + 0.006082112435251474, + 0.004941781982779503, + 0.005120260175317526, + 0.017825331538915634, + 0.01067361980676651, + 0.014546073973178864, + 0.01521261315792799, + 0.023864997550845146, + 0.017010776326060295, + 0.0409327857196331, + 0.020615072920918465, + 0.008988761343061924, + 0.014487901702523232, + 0.014438756741583347, + 0.011500108987092972, + 0.00941300019621849, + 0.009526174515485764, + 0.011498508974909782, + 0.01253614854067564, + 0.0071786632761359215, + 0.007151630241423845, + 0.00664807902649045, + 0.010469548404216766, + 0.00953676924109459, + 0.015093432739377022, + 0.007739937398582697, + 0.010194960981607437, + 0.0030612200498580933, + 0.0038840528577566147, + 0.0027134795673191547, + 0.0043397098779678345, + 0.0027713957242667675, + 0.00566228199750185, + 0.0031397384591400623, + 0.0033612381666898727, + 0.006231324747204781, + 0.0075010135769844055, + 0.00889815203845501, + 0.013379547744989395, + 0.005932879168540239, + 0.006261568982154131, + 0.009260791353881359, + 0.006064448971301317, + 0.004044914618134499, + 0.003502858569845557, + 0.008998374454677105, + 0.010227988474071026, + 0.004658728837966919, + 0.0047200582921504974, + 0.004379503894597292, + 0.0042021917179226875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.04866068437695503 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636845551431179 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636845551431179, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636845551431179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0870191752910614, + 0.1310068964958191, + 0.0750041976571083, + 0.08682160824537277, + 0.041133299469947815, + 0.13419266045093536, + 0.07711578905582428, + 0.07424697279930115, + 0.16319556534290314, + 0.13026389479637146 + ], + "credit_map": [ + 0.014907361008226871, + 0.011158949695527554, + 0.013989150524139404, + 0.008457179181277752, + 0.010251062922179699, + 0.014773200266063213, + 0.007753328885883093, + 0.005728931166231632, + 0.01499467808753252, + 0.01413801684975624, + 0.019515469670295715, + 0.02628544345498085, + 0.007807945366948843, + 0.015047043561935425, + 0.011615283787250519, + 0.021602993831038475, + 0.008615609258413315, + 0.009244900196790695, + 0.012609625235199928, + 0.009870380163192749, + 0.0075762621127069, + 0.012240946292877197, + 0.005855404306203127, + 0.008991063572466373, + 0.006321070250123739, + 0.009484957903623581, + 0.01224400196224451, + 0.009683151729404926, + 0.008726922795176506, + 0.02477039210498333, + 0.0060716974548995495, + 0.009519405663013458, + 0.0028977880720049143, + 0.004804091528058052, + 0.0035784330684691668, + 0.004713860340416431, + 0.0051008425652980804, + 0.003984710667282343, + 0.010188248939812183, + 0.0058653224259614944, + 0.017581142485141754, + 0.01920521818101406, + 0.013405226171016693, + 0.013325277715921402, + 0.02651839703321457, + 0.017244862392544746, + 0.011599048972129822, + 0.015313475392758846, + 0.008408709429204464, + 0.007940257899463177, + 0.006749930791556835, + 0.0104970782995224, + 0.009049362502992153, + 0.013834742829203606, + 0.008745073340833187, + 0.011890629306435585, + 0.009855295531451702, + 0.010010012425482273, + 0.006785605102777481, + 0.009627732448279858, + 0.007486875168979168, + 0.011448615230619907, + 0.007825186476111412, + 0.011207642033696175, + 0.019842151552438736, + 0.025097938254475594, + 0.020410459488630295, + 0.02598540112376213, + 0.017129292711615562, + 0.016004974022507668, + 0.022469744086265564, + 0.01625559851527214, + 0.013593425042927265, + 0.010687886737287045, + 0.015444040298461914, + 0.015541675500571728, + 0.021160181611776352, + 0.019869975745677948, + 0.017687957733869553, + 0.01627875119447708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.004636845551431179 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624453864991665 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624453864991665, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624453864991665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380005180835724, + 0.14662353694438934, + 0.06026601046323776, + 0.05594105273485184, + 0.06672623753547668, + 0.12552709877490997, + 0.10985211282968521, + 0.06491543352603912, + 0.20326921343803406, + 0.02887873724102974 + ], + "credit_map": [ + 0.019771505147218704, + 0.015056909993290901, + 0.013981766067445278, + 0.02722437120974064, + 0.01422835886478424, + 0.019063202664256096, + 0.015515628270804882, + 0.01315879262983799, + 0.015116491355001926, + 0.020042231306433678, + 0.026296187192201614, + 0.027983754873275757, + 0.01079290360212326, + 0.01630566082894802, + 0.009257263503968716, + 0.020829053595662117, + 0.006830183789134026, + 0.005676836706697941, + 0.011381345801055431, + 0.007889128290116787, + 0.005420065019279718, + 0.012064318172633648, + 0.004992696922272444, + 0.006011440418660641, + 0.005776963196694851, + 0.01020801905542612, + 0.006744640879333019, + 0.005201748572289944, + 0.009274683892726898, + 0.007155937142670155, + 0.005838262848556042, + 0.005740799475461245, + 0.0058211591094732285, + 0.007463755086064339, + 0.011654027737677097, + 0.007636394817382097, + 0.00707643385976553, + 0.007331415545195341, + 0.01144768949598074, + 0.00829536747187376, + 0.013394973240792751, + 0.015461726114153862, + 0.019971035420894623, + 0.018877128139138222, + 0.013409742154181004, + 0.013140963390469551, + 0.016662856563925743, + 0.014608674682676792, + 0.01140996441245079, + 0.011492560617625713, + 0.00934626441448927, + 0.016256388276815414, + 0.013970906846225262, + 0.020231982693076134, + 0.011816945858299732, + 0.015327109955251217, + 0.009741801768541336, + 0.006739091128110886, + 0.005749312695115805, + 0.007524511311203241, + 0.006743375211954117, + 0.012762443162500858, + 0.007600043900310993, + 0.008054856210947037, + 0.019181353971362114, + 0.029111893847584724, + 0.027037745341658592, + 0.036687981337308884, + 0.019570516422390938, + 0.01901056617498398, + 0.030337031930685043, + 0.022332148626446724, + 0.0023525224532932043, + 0.0021254948806017637, + 0.006318974308669567, + 0.007465218659490347, + 0.0031591125298291445, + 0.0029300295282155275, + 0.0020993223879486322, + 0.0024280655197799206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.007624453864991665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, high-speed chase shot follows a young man on rollerblades as he engages a massive, fire-breathing dragon covered in obsidian scales. The young man's long, raven hair, tied in a high ponytail, whips wildly in the wind, his eyes piercing his steely face. He wears a sleeveless, bright red silk battle armor embroidered with golden flames, his bare arms muscular and defined. Beneath his feet, two rollerblades of bronze and flame spin at high speed, emitting dazzling sparks and allowing him to nimbly pivot and hover in the air. Just as the dragon coils around the serpentine's body, its maw opens, unleashing a devastating torrent of blazing orange-yellow flames. The young man reacts swiftly, leaning sideways and performing a breathtaking acrobatic maneuver to avoid the blaze. His immensely swift movements left a trail of faintly transparent afterimages of red clothing and golden flames in the twilight sky, a blend of deep purple and indigo, perfectly illustrating the ultimate sense of speed. The entire battlefield was situated atop a precipitous mountain range, with a mist-shrouded abyss below. The distant horizon burned with the afterglow of the setting sun, casting long, dramatic shadows over this epic confrontation. The air was filled with the tense atmosphere of scorching heat and clashing energy.", + "seed": 17562, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_3/index_2578_seed_17562_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07724160961806774, + "vlm_logprob": -0.07724160961806774, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.07724160961806774, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10201390087604523 + ], + "opposite_logprob_tokens": [ + -2.3520138263702393 + ], + "logprob_reward": -0.10201390087604523, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10201390087604523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03404809907078743, + 0.06522537022829056, + 0.05908884108066559, + 0.0594402514398098, + 0.6001362800598145, + 0.020200563594698906, + 0.01681390590965748, + 0.05829915776848793, + 0.0368841178715229, + 0.04986339807510376 + ], + "credit_map": [ + 0.0036258769687265158, + 0.0034330752678215504, + 0.005005902610719204, + 0.0048662093468010426, + 0.0030761766247451305, + 0.004869724158197641, + 0.0050203269347548485, + 0.004150808323174715, + 0.008738317526876926, + 0.008096253499388695, + 0.008447633124887943, + 0.00493492791429162, + 0.011692422442138195, + 0.009002485312521458, + 0.0048618800938129425, + 0.009451445192098618, + 0.009025166742503643, + 0.00608337065204978, + 0.009246399626135826, + 0.005165889859199524, + 0.004984376486390829, + 0.010647106915712357, + 0.0045090531930327415, + 0.009427479468286037, + 0.006610597483813763, + 0.008111647330224514, + 0.006904140580445528, + 0.006790748797357082, + 0.009274409152567387, + 0.007559174671769142, + 0.006237886380404234, + 0.007951642386615276, + 0.22848893702030182, + 0.24965599179267883, + 0.019834894686937332, + 0.023859800770878792, + 0.024609815329313278, + 0.014451952651143074, + 0.024103745818138123, + 0.015131169930100441, + 0.0027513927780091763, + 0.002381773665547371, + 0.00424807658419013, + 0.0024914778769016266, + 0.0021051454823464155, + 0.002095674630254507, + 0.002120960270985961, + 0.0020060616079717875, + 0.0025699855759739876, + 0.00197941018268466, + 0.003801919287070632, + 0.0021836368832737207, + 0.0021139022428542376, + 0.001440594787709415, + 0.0011831928277388215, + 0.001541264122352004, + 0.007446626666933298, + 0.007716013118624687, + 0.0072621568106114864, + 0.012115861289203167, + 0.008199160918593407, + 0.004711233079433441, + 0.0050983307883143425, + 0.005749773234128952, + 0.004622923210263252, + 0.005113229621201754, + 0.00728283915668726, + 0.003391033038496971, + 0.006468360777944326, + 0.004506650380790234, + 0.002562248846516013, + 0.0029368316754698753, + 0.005568308290094137, + 0.006466791499406099, + 0.004201630130410194, + 0.00797636155039072, + 0.006252750288695097, + 0.007747393101453781, + 0.0056882258504629135, + 0.0059619382955133915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.10201390087604523 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559168934822083 + ], + "opposite_logprob_tokens": [ + -1.6005916595458984 + ], + "logprob_reward": -0.22559168934822083, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559168934822083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09492964297533035, + 0.11074712872505188, + 0.0807073637843132, + 0.13207080960273743, + 0.28681492805480957, + 0.023820456117391586, + 0.042797017842531204, + 0.13306181132793427, + 0.04901033639907837, + 0.04604049026966095 + ], + "credit_map": [ + 0.009850750677287579, + 0.008777935057878494, + 0.01583258993923664, + 0.014125295914709568, + 0.011507224291563034, + 0.009152458980679512, + 0.012928048148751259, + 0.012755339965224266, + 0.01974222995340824, + 0.0074351467192173, + 0.014075623825192451, + 0.008036337792873383, + 0.029252681881189346, + 0.007561414502561092, + 0.010579884983599186, + 0.01406380720436573, + 0.0167652927339077, + 0.010918975807726383, + 0.00850062444806099, + 0.006510445382446051, + 0.007871557027101517, + 0.01134375762194395, + 0.005029648542404175, + 0.013767065480351448, + 0.015020708553493023, + 0.01875573955476284, + 0.01261516846716404, + 0.013763288967311382, + 0.02450006827712059, + 0.017408542335033417, + 0.014444884844124317, + 0.015562421642243862, + 0.11043374985456467, + 0.12110121548175812, + 0.010910727083683014, + 0.00957325380295515, + 0.012464617379009724, + 0.00718032056465745, + 0.007138965185731649, + 0.008012098260223866, + 0.0031254144851118326, + 0.003556241048499942, + 0.0033509384375065565, + 0.002716947579756379, + 0.002495266031473875, + 0.0024394227657467127, + 0.002649450907483697, + 0.0034867743961513042, + 0.006720604840666056, + 0.005021166522055864, + 0.00849133264273405, + 0.006281799171119928, + 0.005703159142285585, + 0.003413696074858308, + 0.0040177879855036736, + 0.0031474721617996693, + 0.018200282007455826, + 0.01856285147368908, + 0.017909185960888863, + 0.028497949242591858, + 0.01499897800385952, + 0.011923959478735924, + 0.012104494497179985, + 0.010864109732210636, + 0.005287345964461565, + 0.008009976707398891, + 0.008082748390734196, + 0.0069697038270533085, + 0.005646424368023872, + 0.006146220024675131, + 0.0040469844825565815, + 0.004820933565497398, + 0.006094027776271105, + 0.006670114118605852, + 0.007722463924437761, + 0.006473504938185215, + 0.005676880478858948, + 0.004871147684752941, + 0.004755507223308086, + 0.003776846919208765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.22559168934822083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048674650490283966 + ], + "opposite_logprob_tokens": [ + -3.0486745834350586 + ], + "logprob_reward": -0.048674650490283966, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048674650490283966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034868642687797546, + 0.0704692080616951, + 0.058402061462402344, + 0.04825352877378464, + 0.6139188408851624, + 0.016427449882030487, + 0.06142983213067055, + 0.04455173388123512, + 0.00964427087455988, + 0.04203440248966217 + ], + "credit_map": [ + 0.004158471245318651, + 0.0025257528759539127, + 0.006014535203576088, + 0.004816283471882343, + 0.00408134562894702, + 0.00478788698092103, + 0.004184762481600046, + 0.004299605265259743, + 0.018528539687395096, + 0.005010613240301609, + 0.0035679189022630453, + 0.003610305255278945, + 0.028845058754086494, + 0.004500329494476318, + 0.003411461366340518, + 0.0029949836898595095, + 0.007518613710999489, + 0.004846459720283747, + 0.006476821377873421, + 0.006367725320160389, + 0.007463496644049883, + 0.007270526606589556, + 0.008727353066205978, + 0.009731066413223743, + 0.005008616019040346, + 0.005495442543178797, + 0.007459884975105524, + 0.008200529031455517, + 0.005778762046247721, + 0.005696175619959831, + 0.005363129079341888, + 0.005250989459455013, + 0.2401360422372818, + 0.26432543992996216, + 0.018882719799876213, + 0.020799420773983, + 0.026485728099942207, + 0.015161864459514618, + 0.014119317755103111, + 0.01400830876082182, + 0.0016365668270736933, + 0.0018024761229753494, + 0.0024867476895451546, + 0.0025263160932809114, + 0.0014437440549954772, + 0.0022553347516804934, + 0.0019796660635620356, + 0.0022965981625020504, + 0.00842939130961895, + 0.009291366674005985, + 0.012569578364491463, + 0.006579986773431301, + 0.008096707984805107, + 0.0066573782823979855, + 0.004342945292592049, + 0.005462472327053547, + 0.0037287732120603323, + 0.004924466833472252, + 0.004771897103637457, + 0.00841459259390831, + 0.0074721574783325195, + 0.0048620617017149925, + 0.004792396444827318, + 0.005585388746112585, + 0.0011928214225918055, + 0.0010800310410559177, + 0.0017554732039570808, + 0.0018186401575803757, + 0.0013118917122483253, + 0.0010406285291537642, + 0.0007280469289980829, + 0.0007167382282204926, + 0.004917462822049856, + 0.005762618500739336, + 0.007827582769095898, + 0.004788432270288467, + 0.006227067206054926, + 0.005004621110856533, + 0.004203709773719311, + 0.003302906872704625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.048674650490283966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191736526787281 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003191736526787281, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191736526787281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06484774500131607, + 0.14969168603420258, + 0.03688611090183258, + 0.04257112368941307, + 0.4184475243091583, + 0.09066291153430939, + 0.02198917604982853, + 0.10621848702430725, + 0.03186250850558281, + 0.036822717636823654 + ], + "credit_map": [ + 0.008647611364722252, + 0.005339071620255709, + 0.011001486331224442, + 0.007692497223615646, + 0.0075178551487624645, + 0.00893170852214098, + 0.007395623251795769, + 0.008321896195411682, + 0.03963053226470947, + 0.010784200392663479, + 0.009903271682560444, + 0.00597706763073802, + 0.060480449348688126, + 0.010100731626152992, + 0.0064722211100161076, + 0.006343216169625521, + 0.006066636182367802, + 0.004591487348079681, + 0.004958481062203646, + 0.00328643131069839, + 0.003217654535546899, + 0.005393811035901308, + 0.003163460874930024, + 0.006208151578903198, + 0.004257974214851856, + 0.004723129794001579, + 0.005259175319224596, + 0.0050074257887899876, + 0.00754095334559679, + 0.006673689931631088, + 0.004448874853551388, + 0.0046599023044109344, + 0.16056951880455017, + 0.17694102227687836, + 0.017218170687556267, + 0.017970971763134003, + 0.01765356957912445, + 0.00963473692536354, + 0.00899414625018835, + 0.009465442039072514, + 0.012446296401321888, + 0.009904524311423302, + 0.018560605123639107, + 0.009594963863492012, + 0.008111218921840191, + 0.009190555661916733, + 0.009648135863244534, + 0.013206611387431622, + 0.0028202617540955544, + 0.00251992279663682, + 0.0031243355479091406, + 0.0018805431900545955, + 0.005149980075657368, + 0.002544994466006756, + 0.001601656200364232, + 0.0023474835325032473, + 0.014209371991455555, + 0.01537332683801651, + 0.014257406815886497, + 0.022669440135359764, + 0.015322290360927582, + 0.0072235423140227795, + 0.008898339234292507, + 0.008264773525297642, + 0.003076830878853798, + 0.0037854481488466263, + 0.005181360058486462, + 0.005465988535434008, + 0.004093198571354151, + 0.0033352263271808624, + 0.0031097957398742437, + 0.0038146632723510265, + 0.004895415622740984, + 0.005020655691623688, + 0.0035462561063468456, + 0.006327456794679165, + 0.004203287418931723, + 0.004393416456878185, + 0.0039993650279939175, + 0.00443686731159687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.003191736526787281 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736070849001408 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736070849001408, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736070849001408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0833124965429306, + 0.22328492999076843, + 0.08677686750888824, + 0.03268785402178764, + 0.1137247234582901, + 0.0640922412276268, + 0.11037827283143997, + 0.175128772854805, + 0.028453614562749863, + 0.08216021955013275 + ], + "credit_map": [ + 0.008772465400397778, + 0.009167180396616459, + 0.012541511096060276, + 0.008540663868188858, + 0.009240616112947464, + 0.014008661732077599, + 0.010443110018968582, + 0.010598287917673588, + 0.052481308579444885, + 0.01614067330956459, + 0.01425244938582182, + 0.011149371042847633, + 0.08134059607982635, + 0.017492348328232765, + 0.013286923989653587, + 0.017141276970505714, + 0.01911255717277527, + 0.009811905212700367, + 0.007188780698925257, + 0.006854387931525707, + 0.008984426036477089, + 0.0063719600439071655, + 0.009567727334797382, + 0.018885118886828423, + 0.004772807937115431, + 0.006004386581480503, + 0.003977368585765362, + 0.0033259561751037836, + 0.005025277845561504, + 0.002976513933390379, + 0.0037039448507130146, + 0.0029016027692705393, + 0.038237325847148895, + 0.04213246330618858, + 0.005445305723696947, + 0.005019414704293013, + 0.006206131540238857, + 0.0055964915081858635, + 0.006243747193366289, + 0.0048438492231070995, + 0.005583038087934256, + 0.008080181665718555, + 0.008859945461153984, + 0.009478757157921791, + 0.006883146241307259, + 0.010186701081693172, + 0.007109412457793951, + 0.007911057211458683, + 0.011927482672035694, + 0.011953416280448437, + 0.01319082174450159, + 0.018154025077819824, + 0.014782989397644997, + 0.012467638589441776, + 0.008984304033219814, + 0.018917595967650414, + 0.023739617317914963, + 0.0252130925655365, + 0.02278277464210987, + 0.03424989432096481, + 0.021617455407977104, + 0.015778440982103348, + 0.017056003212928772, + 0.01469149999320507, + 0.002030601492151618, + 0.0020336236339062452, + 0.007911186665296555, + 0.005152950529009104, + 0.002768783364444971, + 0.002425543265417218, + 0.0026470620650798082, + 0.0034838630817830563, + 0.008141850121319294, + 0.008156068623065948, + 0.01597471535205841, + 0.012275305576622486, + 0.008584333583712578, + 0.008541903458535671, + 0.009409578517079353, + 0.011076460592448711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.006736070849001408 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, high-speed chase shot follows a young man on rollerblades as he engages a massive, fire-breathing dragon covered in obsidian scales. The young man's long, raven hair, tied in a high ponytail, whips wildly in the wind, his eyes piercing his steely face. He wears a sleeveless, bright red silk battle armor embroidered with golden flames, his bare arms muscular and defined. Beneath his feet, two rollerblades of bronze and flame spin at high speed, emitting dazzling sparks and allowing him to nimbly pivot and hover in the air. Just as the dragon coils around the serpentine's body, its maw opens, unleashing a devastating torrent of blazing orange-yellow flames. The young man reacts swiftly, leaning sideways and performing a breathtaking acrobatic maneuver to avoid the blaze. His immensely swift movements left a trail of faintly transparent afterimages of red clothing and golden flames in the twilight sky, a blend of deep purple and indigo, perfectly illustrating the ultimate sense of speed. The entire battlefield was situated atop a precipitous mountain range, with a mist-shrouded abyss below. The distant horizon burned with the afterglow of the setting sun, casting long, dramatic shadows over this epic confrontation. The air was filled with the tense atmosphere of scorching heat and clashing energy.", + "seed": 17563, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_3/index_2578_seed_17563_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08213819898664951, + "vlm_logprob": -0.08213819898664951, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.08213819898664951, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11440514773130417 + ], + "opposite_logprob_tokens": [ + -2.2394051551818848 + ], + "logprob_reward": -0.11440514773130417, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11440514773130417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22000660002231598, + 0.11197204142808914, + 0.08865451067686081, + 0.04381413012742996, + 0.042307015508413315, + 0.11022836714982986, + 0.1687902957201004, + 0.05195516720414162, + 0.05491901561617851, + 0.10735281556844711 + ], + "credit_map": [ + 0.01978171616792679, + 0.022089151665568352, + 0.03751334920525551, + 0.03555719926953316, + 0.027669399976730347, + 0.019176863133907318, + 0.03763652220368385, + 0.020582400262355804, + 0.009026485495269299, + 0.011813323013484478, + 0.014469739980995655, + 0.014414569362998009, + 0.01736016944050789, + 0.015871373936533928, + 0.010494587011635303, + 0.018521787598729134, + 0.006828685291111469, + 0.012067089788615704, + 0.01426432654261589, + 0.012393360957503319, + 0.010108410380780697, + 0.009103258140385151, + 0.013141118921339512, + 0.010748257860541344, + 0.0037365180905908346, + 0.006129023618996143, + 0.005648535210639238, + 0.006496841087937355, + 0.006346232257783413, + 0.004914425779134035, + 0.005680967587977648, + 0.004861587658524513, + 0.0037468222435563803, + 0.0057547749020159245, + 0.0052468664944171906, + 0.005039554554969072, + 0.006149562541395426, + 0.005728950258344412, + 0.005606817547231913, + 0.005033665802329779, + 0.013146031647920609, + 0.012210733257234097, + 0.017542432993650436, + 0.01379457674920559, + 0.015872152522206306, + 0.014204396866261959, + 0.00923904124647379, + 0.014218991622328758, + 0.01854303479194641, + 0.01697426103055477, + 0.027317427098751068, + 0.014704161323606968, + 0.03211796656250954, + 0.0175640769302845, + 0.015204507857561111, + 0.02636486105620861, + 0.008204841986298561, + 0.0056741926819086075, + 0.006485642399638891, + 0.00441421614959836, + 0.006837461609393358, + 0.005193352233618498, + 0.0058306059800088406, + 0.009314851835370064, + 0.006682030390948057, + 0.007839694619178772, + 0.006558449473232031, + 0.006589463446289301, + 0.00510367238894105, + 0.005901042837649584, + 0.00810927338898182, + 0.008135391399264336, + 0.009812144562602043, + 0.014360026456415653, + 0.014259173534810543, + 0.014753838069736958, + 0.01306610181927681, + 0.01457905676215887, + 0.015104172751307487, + 0.011418298818171024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.11440514773130417 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559253871440887 + ], + "opposite_logprob_tokens": [ + -1.6005924940109253 + ], + "logprob_reward": -0.22559253871440887, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559253871440887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15390364825725555, + 0.07822312414646149, + 0.09555687755346298, + 0.05280917137861252, + 0.04042651876807213, + 0.058093152940273285, + 0.30819860100746155, + 0.06948778033256531, + 0.06593479216098785, + 0.07736635208129883 + ], + "credit_map": [ + 0.02176686003804207, + 0.013672507368028164, + 0.02263372763991356, + 0.02316201850771904, + 0.019968688488006592, + 0.013978777453303337, + 0.02440347522497177, + 0.014317589811980724, + 0.010267947800457478, + 0.0059543089009821415, + 0.011847265064716339, + 0.005866489373147488, + 0.007673002313822508, + 0.011150870472192764, + 0.006448773667216301, + 0.01901446469128132, + 0.0064213150180876255, + 0.009611361660063267, + 0.020097874104976654, + 0.009333156980574131, + 0.012267934158444405, + 0.010563394986093044, + 0.012082653120160103, + 0.015179183334112167, + 0.004299269989132881, + 0.006666058674454689, + 0.006352628581225872, + 0.006258707027882338, + 0.009624824859201908, + 0.007719125598669052, + 0.005121523514389992, + 0.0067670345306396484, + 0.004184018354862928, + 0.005411474499851465, + 0.002822421956807375, + 0.003760299412533641, + 0.006872081663459539, + 0.008368401788175106, + 0.004531121347099543, + 0.004476703237742186, + 0.00863686390221119, + 0.0041648950427770615, + 0.005867613013833761, + 0.004671510774642229, + 0.008564637042582035, + 0.010358474217355251, + 0.004868183750659227, + 0.010960971936583519, + 0.03807632252573967, + 0.03503933921456337, + 0.05948004871606827, + 0.02845102548599243, + 0.04930080845952034, + 0.032935455441474915, + 0.02846483699977398, + 0.03645078092813492, + 0.012488305568695068, + 0.011725828982889652, + 0.007061393931508064, + 0.006256150547415018, + 0.00676006218418479, + 0.00725385919213295, + 0.007868359796702862, + 0.01007381733506918, + 0.008605096489191055, + 0.009421112015843391, + 0.007239759434014559, + 0.008617592975497246, + 0.006868402939289808, + 0.00538540119305253, + 0.009316395036876202, + 0.010481025092303753, + 0.0103622917085886, + 0.012690678238868713, + 0.011243675835430622, + 0.013003812171518803, + 0.00650783721357584, + 0.0075782351195812225, + 0.008600625209510326, + 0.007379199378192425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.22559253871440887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496829003095627 + ], + "opposite_logprob_tokens": [ + -2.9299683570861816 + ], + "logprob_reward": -0.05496829003095627, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496829003095627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16517561674118042, + 0.14432765543460846, + 0.1368108093738556, + 0.07472406327724457, + 0.04570460692048073, + 0.03848014026880264, + 0.12350060790777206, + 0.037402451038360596, + 0.11546436697244644, + 0.11840969324111938 + ], + "credit_map": [ + 0.015924789011478424, + 0.018769213929772377, + 0.027620412409305573, + 0.02035687305033207, + 0.021702447906136513, + 0.020211651921272278, + 0.02663949690759182, + 0.013950735330581665, + 0.011604491621255875, + 0.01729493960738182, + 0.0220348984003067, + 0.009956562891602516, + 0.027190249413251877, + 0.02049332670867443, + 0.01738501712679863, + 0.018368177115917206, + 0.009498191997408867, + 0.015445292927324772, + 0.025698184967041016, + 0.01793963462114334, + 0.015331468544900417, + 0.014560791663825512, + 0.023683514446020126, + 0.014653750695288181, + 0.005988025572150946, + 0.009727034717798233, + 0.013070637360215187, + 0.011195571161806583, + 0.009223034605383873, + 0.00829826109111309, + 0.01029561460018158, + 0.0069258809089660645, + 0.004113182425498962, + 0.0060245017521083355, + 0.004596393089741468, + 0.005322635173797607, + 0.006248375866562128, + 0.00717031117528677, + 0.0061368318274617195, + 0.006092372350394726, + 0.005288338754326105, + 0.003376653417944908, + 0.004985007923096418, + 0.004592112731188536, + 0.004928877577185631, + 0.005173470359295607, + 0.0034458129666745663, + 0.00668986514210701, + 0.012441040016710758, + 0.016354816034436226, + 0.0155265424400568, + 0.009264148771762848, + 0.021817343309521675, + 0.014501336961984634, + 0.012011362239718437, + 0.021584011614322662, + 0.006332641001790762, + 0.006032232660800219, + 0.002992534078657627, + 0.003947087097913027, + 0.005872123874723911, + 0.0042929272167384624, + 0.0036039066035300493, + 0.004328999202698469, + 0.013389983214437962, + 0.02135315351188183, + 0.013020187616348267, + 0.012566904537379742, + 0.009715001098811626, + 0.011131525039672852, + 0.01453483197838068, + 0.019752783700823784, + 0.016333600506186485, + 0.011427365243434906, + 0.017878489568829536, + 0.017540132626891136, + 0.011479653418064117, + 0.012773774564266205, + 0.018994558602571487, + 0.011982114985585213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.05496829003095627 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946207791566849 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946207791566849, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946207791566849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15357297658920288, + 0.16455353796482086, + 0.12378604710102081, + 0.044105153530836105, + 0.043331969529390335, + 0.054216742515563965, + 0.09316808730363846, + 0.24909554421901703, + 0.0332171805202961, + 0.04095272719860077 + ], + "credit_map": [ + 0.018290724605321884, + 0.01513410359621048, + 0.026405317708849907, + 0.020981887355446815, + 0.019397642463445663, + 0.014338801614940166, + 0.02269737794995308, + 0.016327133402228355, + 0.01891699805855751, + 0.012412527576088905, + 0.02412874810397625, + 0.01255184505134821, + 0.019117265939712524, + 0.022413646802306175, + 0.017232060432434082, + 0.03778044506907463, + 0.010420495644211769, + 0.013290982693433762, + 0.025810034945607185, + 0.01412473525851965, + 0.011854044161736965, + 0.012816853821277618, + 0.020330941304564476, + 0.015137960202991962, + 0.004139927215874195, + 0.0056116217747330666, + 0.007227432448416948, + 0.005269035696983337, + 0.00392164709046483, + 0.005016562063246965, + 0.00611118646338582, + 0.0068077389150857925, + 0.006734426598995924, + 0.003662425559014082, + 0.006679953541606665, + 0.00519167585298419, + 0.005561468191444874, + 0.006296053528785706, + 0.004314005374908447, + 0.0048919618129730225, + 0.005561858415603638, + 0.005181425251066685, + 0.004997165873646736, + 0.006247090641409159, + 0.007650068495422602, + 0.007530867122113705, + 0.005242896266281605, + 0.011805370450019836, + 0.010122725740075111, + 0.01151863019913435, + 0.01196527574211359, + 0.009151351638138294, + 0.01760057918727398, + 0.010309644043445587, + 0.009729133918881416, + 0.012770742177963257, + 0.03550860285758972, + 0.03198149800300598, + 0.0370032899081707, + 0.0195627398788929, + 0.030843960121273994, + 0.025485487654805183, + 0.019572565332055092, + 0.04913739114999771, + 0.0033157458528876305, + 0.0045830425806343555, + 0.003505726344883442, + 0.005193198565393686, + 0.0035987677983939648, + 0.00328971934504807, + 0.004744910169392824, + 0.004986068233847618, + 0.005046180449426174, + 0.005328151397407055, + 0.006362196058034897, + 0.006551960948854685, + 0.0037464944180101156, + 0.00562657043337822, + 0.004742294084280729, + 0.0035488770809024572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.005946207791566849 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man wearing sleeveless bright red silk battle armor with golden flame embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the antagonist a massive dragon covered in obsidian scales that is breathing fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the young man's rollerblades bronze-colored and emitting sparks while he performs acrobatic maneuvers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment a precipitous mountain range with a mist-shrouded abyss below and a twilight sky of deep purple and indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low-angle, high-speed perspective following the young man?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778810665011406 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778810665011406, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778810665011406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26189231872558594, + 0.05854957178235054, + 0.05143582448363304, + 0.09354032576084137, + 0.10963264852762222, + 0.04895932972431183, + 0.10340861976146698, + 0.06950438767671585, + 0.12816181778907776, + 0.07491518557071686 + ], + "credit_map": [ + 0.027097390964627266, + 0.01904708705842495, + 0.03812680393457413, + 0.0493970662355423, + 0.0341789647936821, + 0.020148299634456635, + 0.053841814398765564, + 0.020054906606674194, + 0.005795953795313835, + 0.003958472050726414, + 0.010190569795668125, + 0.006142898462712765, + 0.0073456005193293095, + 0.007158720400184393, + 0.0076783630065619946, + 0.010278990492224693, + 0.003530242247506976, + 0.006009427364915609, + 0.011011463589966297, + 0.005921133328229189, + 0.006204705219715834, + 0.005115918815135956, + 0.006117046345025301, + 0.007525887805968523, + 0.009818620048463345, + 0.00994421448558569, + 0.016544746235013008, + 0.010007725097239017, + 0.008448074571788311, + 0.012867625802755356, + 0.008444185368716717, + 0.01746514067053795, + 0.014010382816195488, + 0.011693611741065979, + 0.013446316123008728, + 0.008617880754172802, + 0.01468429435044527, + 0.022564826533198357, + 0.011893772520124912, + 0.01272156834602356, + 0.004745409823954105, + 0.003703443333506584, + 0.004887823946774006, + 0.005013573914766312, + 0.00934887770563364, + 0.00918011274188757, + 0.0040443455800414085, + 0.008035742677748203, + 0.010846068151295185, + 0.00962647795677185, + 0.008818570524454117, + 0.010240473784506321, + 0.024839108809828758, + 0.013482535257935524, + 0.01085173524916172, + 0.014703648164868355, + 0.00939001515507698, + 0.0103457011282444, + 0.008396287448704243, + 0.007132925558835268, + 0.007819343358278275, + 0.007102426141500473, + 0.007238945458084345, + 0.012078741565346718, + 0.013471993617713451, + 0.016377490013837814, + 0.013482778333127499, + 0.017637236043810844, + 0.02174082025885582, + 0.00940733216702938, + 0.020079337060451508, + 0.015964820981025696, + 0.008712821640074253, + 0.011421870440244675, + 0.010197902098298073, + 0.007394962012767792, + 0.006758473813533783, + 0.01567954756319523, + 0.007959981448948383, + 0.006789628881961107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.009778810665011406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_4/index_tensor([2553, 2553])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_4/index_tensor([2553, 2553])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8243bec2ee4697f64d226eba0e30dec54ca3f6c9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_4/index_tensor([2553, 2553])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "subsequent low-angle, close-up tracking shot focuses on a chubby, orange cat gliding uncontrollably across a polished, dark hardwood floor. Its fluffy orange tabby fur glows warmly in the soft natural light streaming in from the side windows. Its emerald green eyes widen in pure surprise at the sudden gliding motion, its ears tilted to the sides in a comical, airplane-like formation, and its mouth parted slightly, as if in a silent exclamation of surprise. The camera precisely captures its fully extended front paws, their pink pads scraping against the floor in a futile attempt to regain a sliver of grip, only to continue its gliding motion. As its body hurtles forward, almost flat on the ground, the smooth floor texture stretches across the frame into a dynamic blur. The blurred legs and feet of furniture in the background flash by, further heightening the sense of speed. The scene is imbued with warm tones and a sudden, comical sense of panic, perfectly capturing this endearing moment brimming with life.", + "seed": 32620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_4/index_2553_seed_32620_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13928498616442084, + "vlm_logprob": -0.13928498616442084, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.13928498616442084, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008868236094713211 + ], + "opposite_logprob_tokens": [ + -5.258868217468262 + ], + "logprob_reward": -0.008868236094713211, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008868236094713211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2643180191516876, + 0.09582311660051346, + 0.13301712274551392, + 0.04183419421315193, + 0.0434085913002491, + 0.03766656294465065, + 0.06809131056070328, + 0.11852705478668213, + 0.1549988090991974, + 0.04231520742177963 + ], + "credit_map": [ + 0.02684769593179226, + 0.037356119602918625, + 0.053761329501867294, + 0.024361131712794304, + 0.0374816432595253, + 0.03359859064221382, + 0.028506409376859665, + 0.022405078634619713, + 0.00588818546384573, + 0.012578321620821953, + 0.011168555356562138, + 0.01136483158916235, + 0.022864146158099174, + 0.00968577153980732, + 0.009695108979940414, + 0.01257819589227438, + 0.009986897930502892, + 0.013497350737452507, + 0.01821819692850113, + 0.014544459991157055, + 0.029287289828062057, + 0.013064454309642315, + 0.020646877586841583, + 0.013771599158644676, + 0.009411519393324852, + 0.003360175061970949, + 0.007117057219147682, + 0.003918501548469067, + 0.004080302082002163, + 0.0033518807031214237, + 0.0046889036893844604, + 0.0059058512561023235, + 0.004617834463715553, + 0.0027510132640600204, + 0.008392213843762875, + 0.004889452829957008, + 0.007710989098995924, + 0.0046737161464989185, + 0.005452933721244335, + 0.004920434206724167, + 0.003099684137851, + 0.002072338480502367, + 0.007215915247797966, + 0.0051802657544612885, + 0.006056813057512045, + 0.0070391506887972355, + 0.00276082381606102, + 0.004241570830345154, + 0.005559283774346113, + 0.009170562028884888, + 0.00564240850508213, + 0.005627322010695934, + 0.01764627918601036, + 0.008421827107667923, + 0.0059325555339455605, + 0.010091079398989677, + 0.012604229152202606, + 0.01215748954564333, + 0.016131293028593063, + 0.009678698144853115, + 0.02585834451019764, + 0.008331053890287876, + 0.012256244197487831, + 0.021509705111384392, + 0.008747396990656853, + 0.009078997187316418, + 0.04572049528360367, + 0.042613107711076736, + 0.014607397839426994, + 0.013781928457319736, + 0.008870347402989864, + 0.011579138226807117, + 0.006826016586273909, + 0.0028981484938412905, + 0.008687146008014679, + 0.00649813748896122, + 0.007070272229611874, + 0.0027468593325465918, + 0.0034991581924259663, + 0.004089468624442816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.008868236094713211 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.632763147354126 + ], + "opposite_logprob_tokens": [ + -0.757763147354126 + ], + "logprob_reward": -0.632763147354126, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.632763147354126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2082829624414444, + 0.0819038525223732, + 0.15526923537254333, + 0.09462860226631165, + 0.1352766454219818, + 0.05257878080010414, + 0.0963144600391388, + 0.04533463343977928, + 0.062466464936733246, + 0.06794442236423492 + ], + "credit_map": [ + 0.02473580278456211, + 0.033453624695539474, + 0.02676977775990963, + 0.01931719481945038, + 0.036566175520420074, + 0.024046149104833603, + 0.024081436917185783, + 0.01931280829012394, + 0.007761961780488491, + 0.006453678011894226, + 0.01164441742002964, + 0.00938647985458374, + 0.008607125841081142, + 0.010029896162450314, + 0.019515879452228546, + 0.008504413068294525, + 0.011993840336799622, + 0.019323986023664474, + 0.020377829670906067, + 0.019916675984859467, + 0.032132744789123535, + 0.01614721305668354, + 0.020650627091526985, + 0.014726300723850727, + 0.023062720894813538, + 0.009320049546658993, + 0.01154768094420433, + 0.012205233797430992, + 0.009398025460541248, + 0.007759003899991512, + 0.010253551416099072, + 0.011082337237894535, + 0.011842948384582996, + 0.009195004589855671, + 0.02343812957406044, + 0.014659921638667583, + 0.021670661866664886, + 0.011749782599508762, + 0.015538769774138927, + 0.02718142606317997, + 0.004678075201809406, + 0.004215128719806671, + 0.00801970437169075, + 0.007523103151470423, + 0.00778178172186017, + 0.008561166934669018, + 0.0045544966123998165, + 0.007245320826768875, + 0.007389722857624292, + 0.0063034771010279655, + 0.01027866080403328, + 0.008560613729059696, + 0.00914735347032547, + 0.025911035016179085, + 0.005292212590575218, + 0.023431388661265373, + 0.005288265645503998, + 0.004760146141052246, + 0.005522836931049824, + 0.004711412359029055, + 0.010549302212893963, + 0.0068754334934055805, + 0.004506261087954044, + 0.0031209709122776985, + 0.0046415249817073345, + 0.0042145405896008015, + 0.022589126601815224, + 0.01024419255554676, + 0.006289190147072077, + 0.004553753416985273, + 0.004692032467573881, + 0.005242104176431894, + 0.01086486503481865, + 0.005936937406659126, + 0.012309356592595577, + 0.013441129587590694, + 0.007849645800888538, + 0.0040890867821872234, + 0.006644751876592636, + 0.006808650214225054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.632763147354126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307573288679123 + ], + "opposite_logprob_tokens": [ + -3.1680757999420166 + ], + "logprob_reward": -0.04307573288679123, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307573288679123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27255505323410034, + 0.08859701454639435, + 0.09619398415088654, + 0.08575711399316788, + 0.060099873691797256, + 0.0362609438598156, + 0.04937151074409485, + 0.05673857033252716, + 0.2330794483423233, + 0.02134651131927967 + ], + "credit_map": [ + 0.02646888606250286, + 0.035052306950092316, + 0.0600757971405983, + 0.026391368359327316, + 0.036660343408584595, + 0.032562240958213806, + 0.02647240087389946, + 0.02887170948088169, + 0.005114135332405567, + 0.010558178648352623, + 0.018073001876473427, + 0.007351974491029978, + 0.01465577818453312, + 0.0065200128592550755, + 0.016372764483094215, + 0.009951155632734299, + 0.0074535240419209, + 0.010131900198757648, + 0.012404697947204113, + 0.008807338774204254, + 0.020069748163223267, + 0.009147501550614834, + 0.01582430675625801, + 0.012354962527751923, + 0.011705707758665085, + 0.006959485821425915, + 0.010228886269032955, + 0.008644101209938526, + 0.014821075834333897, + 0.0067928992211818695, + 0.01580204628407955, + 0.01080290973186493, + 0.005025051534175873, + 0.0039534056559205055, + 0.006628038827329874, + 0.006752383895218372, + 0.01044671144336462, + 0.007465472910553217, + 0.008246265351772308, + 0.011582539416849613, + 0.003089176258072257, + 0.003289199201390147, + 0.006460356526076794, + 0.005606357008218765, + 0.0034968259278684855, + 0.008472706191241741, + 0.0020237574353814125, + 0.0038225629832595587, + 0.0042502968572080135, + 0.007110796868801117, + 0.004928804002702236, + 0.004549716599285603, + 0.013767807744443417, + 0.005709415767341852, + 0.005020837299525738, + 0.004033831879496574, + 0.007005305029451847, + 0.004445372149348259, + 0.009025011211633682, + 0.005337945651262999, + 0.011753280647099018, + 0.008707446977496147, + 0.006299776956439018, + 0.004164428450167179, + 0.013067550957202911, + 0.013418012298643589, + 0.05958307534456253, + 0.07756355404853821, + 0.014513678848743439, + 0.02034517005085945, + 0.01258776430040598, + 0.022000627592206, + 0.0025598315987735987, + 0.002424700651317835, + 0.003396811196580529, + 0.0030228414107114077, + 0.0035334965214133263, + 0.0014668414369225502, + 0.002181369112804532, + 0.0027606175281107426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.04307573288679123 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040924143977463245 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.0040924143977463245, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040924143977463245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2510869801044464, + 0.03876231983304024, + 0.08421497792005539, + 0.05506659299135208, + 0.13273465633392334, + 0.15780828893184662, + 0.02964661456644535, + 0.10790053755044937, + 0.10665597766637802, + 0.036123014986515045 + ], + "credit_map": [ + 0.026422403752803802, + 0.029568007215857506, + 0.059650719165802, + 0.036858633160591125, + 0.023338444530963898, + 0.030478697270154953, + 0.02270759828388691, + 0.022062482312321663, + 0.003182015847414732, + 0.003236358519643545, + 0.006121746730059385, + 0.004499032162129879, + 0.0042537422850728035, + 0.004654016345739365, + 0.008430072106420994, + 0.0043853349052369595, + 0.00638930406421423, + 0.008111965842545033, + 0.01123606413602829, + 0.008137480355799198, + 0.01962154358625412, + 0.007909750565886497, + 0.013896338641643524, + 0.008912532590329647, + 0.007708215620368719, + 0.0048828343860805035, + 0.005417699459940195, + 0.006391210947185755, + 0.009706765413284302, + 0.005057185422629118, + 0.009924679063260555, + 0.005977999418973923, + 0.00965141225606203, + 0.008754721842706203, + 0.014028619043529034, + 0.011125774122774601, + 0.01871173270046711, + 0.01632177270948887, + 0.017615966498851776, + 0.03652466461062431, + 0.01239939033985138, + 0.012561740353703499, + 0.02688336744904518, + 0.027673810720443726, + 0.016626371070742607, + 0.038023799657821655, + 0.009806375950574875, + 0.013833447359502316, + 0.0024836414959281683, + 0.0025518203619867563, + 0.002371489768847823, + 0.0026802602224051952, + 0.003998673055320978, + 0.006789823528379202, + 0.00225247023627162, + 0.006518435664474964, + 0.015418855473399162, + 0.006728654261678457, + 0.009809504263103008, + 0.010356300510466099, + 0.01290854811668396, + 0.02464824542403221, + 0.00830751471221447, + 0.019722918048501015, + 0.006798835005611181, + 0.010725404135882854, + 0.012806534767150879, + 0.020545391365885735, + 0.023088842630386353, + 0.00918818823993206, + 0.009917312301695347, + 0.013585470616817474, + 0.003853181377053261, + 0.0038136541843414307, + 0.005922249518334866, + 0.0052079735323786736, + 0.0060615865513682365, + 0.002537365071475506, + 0.004298134241253138, + 0.004428867716342211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0040924143977463245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625400088727474 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625400088727474, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625400088727474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29946163296699524, + 0.07239078730344772, + 0.07786622643470764, + 0.06635928899049759, + 0.045383334159851074, + 0.06941559165716171, + 0.0872901901602745, + 0.06940897554159164, + 0.12454409897327423, + 0.08787988871335983 + ], + "credit_map": [ + 0.031669121235609055, + 0.04866514354944229, + 0.05318307876586914, + 0.030757324770092964, + 0.0432451069355011, + 0.035058364272117615, + 0.03182327374815941, + 0.025060219690203667, + 0.00556424492970109, + 0.0071503412909805775, + 0.009816660545766354, + 0.009801628068089485, + 0.009259643964469433, + 0.007483240216970444, + 0.014107380993664265, + 0.009207646362483501, + 0.005343336146324873, + 0.007365212310105562, + 0.011001070961356163, + 0.010268619284033775, + 0.018010780215263367, + 0.0070746238343417645, + 0.010642393492162228, + 0.008160189725458622, + 0.013367361389100552, + 0.005201051477342844, + 0.010727016255259514, + 0.007459986489266157, + 0.0068896557204425335, + 0.006283923052251339, + 0.008041753433644772, + 0.008388536050915718, + 0.005201546009629965, + 0.004265994299203157, + 0.006693451665341854, + 0.005563630256801844, + 0.006659915205091238, + 0.004290867131203413, + 0.0056691840291023254, + 0.007038742769509554, + 0.006008354481309652, + 0.0069207302294671535, + 0.014240305870771408, + 0.009483733214437962, + 0.009649625048041344, + 0.009261306375265121, + 0.005673971958458424, + 0.008177564479410648, + 0.0052883983589708805, + 0.006543278228491545, + 0.006638754159212112, + 0.005582685116678476, + 0.011388829909265041, + 0.023815371096134186, + 0.0053021227940917015, + 0.022730752825737, + 0.008863733150064945, + 0.004598098807036877, + 0.007159295491874218, + 0.005195312201976776, + 0.007380705326795578, + 0.017646975815296173, + 0.004708365071564913, + 0.013856490142643452, + 0.007161654997617006, + 0.005474831443279982, + 0.039253927767276764, + 0.04730679839849472, + 0.006210530176758766, + 0.00464472733438015, + 0.006897675339132547, + 0.007593953516334295, + 0.008706019259989262, + 0.008598760701715946, + 0.016623612493276596, + 0.016356751322746277, + 0.012567326426506042, + 0.005815817043185234, + 0.008772412315011024, + 0.010439185425639153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.007625400088727474 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, close-up tracking shot focuses on a chubby, orange cat gliding uncontrollably across a polished, dark hardwood floor. Its fluffy orange tabby fur glows warmly in the soft natural light streaming in from the side windows. Its emerald green eyes widen in pure surprise at the sudden gliding motion, its ears tilted to the sides in a comical, airplane-like formation, and its mouth parted slightly, as if in a silent exclamation of surprise. The camera precisely captures its fully extended front paws, their pink pads scraping against the floor in a futile attempt to regain a sliver of grip, only to continue its gliding motion. As its body hurtles forward, almost flat on the ground, the smooth floor texture stretches across the frame into a dynamic blur. The blurred legs and feet of furniture in the background flash by, further heightening the sense of speed. The scene is imbued with warm tones and a sudden, comical sense of panic, perfectly capturing this endearing moment brimming with life.", + "seed": 32621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_4/index_2553_seed_32621_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1941458542831242, + "vlm_logprob": -0.1941458542831242, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.1941458542831242, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0047646318562328815 + ], + "opposite_logprob_tokens": [ + -6.254764556884766 + ], + "logprob_reward": -0.0047646318562328815, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0047646318562328815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18694326281547546, + 0.0853450670838356, + 0.1070631891489029, + 0.08230119198560715, + 0.08961798995733261, + 0.05935627594590187, + 0.17880366742610931, + 0.057078953832387924, + 0.09295836091041565, + 0.06053214520215988 + ], + "credit_map": [ + 0.013735983520746231, + 0.01028942596167326, + 0.02706080861389637, + 0.024114908650517464, + 0.013020964339375496, + 0.03390990570187569, + 0.020005719736218452, + 0.044805534183979034, + 0.008871546946465969, + 0.012905078008770943, + 0.010020007379353046, + 0.009987263940274715, + 0.014567910693585873, + 0.009647712111473083, + 0.011005828157067299, + 0.008339716121554375, + 0.014077379368245602, + 0.013192417100071907, + 0.01276556495577097, + 0.012452569790184498, + 0.010016056708991528, + 0.020728768780827522, + 0.012299769558012486, + 0.011530659161508083, + 0.005941066425293684, + 0.012029364705085754, + 0.011811193078756332, + 0.011237841099500656, + 0.008363214321434498, + 0.006194728426635265, + 0.011032299138605595, + 0.015691477805376053, + 0.010746886022388935, + 0.009999381378293037, + 0.026389731094241142, + 0.01389523409307003, + 0.007001038640737534, + 0.005546304397284985, + 0.00869003776460886, + 0.007349367719143629, + 0.006795304827392101, + 0.00541275180876255, + 0.007219955325126648, + 0.014703003689646721, + 0.006485681980848312, + 0.00473645469173789, + 0.0072500454261898994, + 0.0067530726082623005, + 0.013653570786118507, + 0.02419331669807434, + 0.01621139794588089, + 0.018057256937026978, + 0.032255154103040695, + 0.023928040638566017, + 0.023112529888749123, + 0.027392378076910973, + 0.004185111727565527, + 0.006696416065096855, + 0.006167478859424591, + 0.011728417128324509, + 0.00616008136421442, + 0.007031405810266733, + 0.007825690321624279, + 0.007284347899258137, + 0.010689706541597843, + 0.006237375549972057, + 0.010209159925580025, + 0.020285096019506454, + 0.011093918234109879, + 0.009978314861655235, + 0.01059947069734335, + 0.013865313492715359, + 0.00576336495578289, + 0.005742530804127455, + 0.00785155687481165, + 0.00817040354013443, + 0.01066743303090334, + 0.008267746306955814, + 0.007810650859028101, + 0.006258456036448479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.0047646318562328815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8982789516448975 + ], + "opposite_logprob_tokens": [ + -0.5232789516448975 + ], + "logprob_reward": -0.8982789516448975, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8982789516448975, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.170134499669075, + 0.07514355331659317, + 0.11978571861982346, + 0.18226897716522217, + 0.06698915362358093, + 0.047816574573516846, + 0.05559598654508591, + 0.03604394942522049, + 0.11134310066699982, + 0.13487857580184937 + ], + "credit_map": [ + 0.01585429348051548, + 0.010060323402285576, + 0.03536887466907501, + 0.018513590097427368, + 0.010479185730218887, + 0.02111469767987728, + 0.0170921441167593, + 0.04165137559175491, + 0.007925670593976974, + 0.01286984235048294, + 0.006256322376430035, + 0.007489679381251335, + 0.014784504659473896, + 0.008579827845096588, + 0.009542842395603657, + 0.007694851607084274, + 0.013580749742686749, + 0.014890347607433796, + 0.012730991467833519, + 0.011848388239741325, + 0.01017843745648861, + 0.028342576697468758, + 0.01510544028133154, + 0.013108768500387669, + 0.015877148136496544, + 0.016910675913095474, + 0.01799659989774227, + 0.01729475148022175, + 0.02724609524011612, + 0.013375785201787949, + 0.04059206321835518, + 0.032975826412439346, + 0.007658532354980707, + 0.008618094027042389, + 0.019389517605304718, + 0.008219478651881218, + 0.006413060240447521, + 0.004123665392398834, + 0.005760272964835167, + 0.006806525867432356, + 0.006756681948900223, + 0.004569884855300188, + 0.006766165141016245, + 0.010830688290297985, + 0.004556834697723389, + 0.004918241873383522, + 0.004097389988601208, + 0.005320683587342501, + 0.0039000282995402813, + 0.005087353754788637, + 0.006725345738232136, + 0.010197426192462444, + 0.006670673377811909, + 0.006253178231418133, + 0.007497256156057119, + 0.009264719672501087, + 0.0026524162385612726, + 0.004546307493001223, + 0.004305343143641949, + 0.006977403070777655, + 0.0047312104143202305, + 0.003701393958181143, + 0.004115669522434473, + 0.005014202557504177, + 0.011315692216157913, + 0.009877271950244904, + 0.010624079965054989, + 0.01473416481167078, + 0.01676470786333084, + 0.021079255267977715, + 0.01310321781784296, + 0.013844690285623074, + 0.013772179372608662, + 0.025680208578705788, + 0.01735728047788143, + 0.015268392860889435, + 0.018069686368107796, + 0.015783077105879784, + 0.015480930916965008, + 0.013466810807585716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.8982789516448975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033718038350343704 + ], + "opposite_logprob_tokens": [ + -3.4087181091308594 + ], + "logprob_reward": -0.033718038350343704, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033718038350343704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4287837743759155, + 0.042821288108825684, + 0.06033996492624283, + 0.15940774977207184, + 0.1020214632153511, + 0.08962464332580566, + 0.013397956266999245, + 0.01919035241007805, + 0.05838940292596817, + 0.026023419573903084 + ], + "credit_map": [ + 0.02819746546447277, + 0.024885855615139008, + 0.04312235116958618, + 0.044808581471443176, + 0.02443852461874485, + 0.06279578059911728, + 0.056246280670166016, + 0.14428894221782684, + 0.0042300112545490265, + 0.005154167767614126, + 0.004311589524149895, + 0.006511113606393337, + 0.0068796006962656975, + 0.0050642522983253, + 0.005496727302670479, + 0.005173826590180397, + 0.007177839055657387, + 0.006794026121497154, + 0.012850763276219368, + 0.008009014651179314, + 0.006111949682235718, + 0.005406036972999573, + 0.008639328181743622, + 0.005351006984710693, + 0.013547045178711414, + 0.014480603858828545, + 0.027786128222942352, + 0.012063687667250633, + 0.01956585980951786, + 0.011745359748601913, + 0.038218364119529724, + 0.022000711411237717, + 0.0114229004830122, + 0.01234889030456543, + 0.028667476028203964, + 0.009606707841157913, + 0.012789781205356121, + 0.00921397190541029, + 0.010725006461143494, + 0.007246735505759716, + 0.00972059741616249, + 0.012200450524687767, + 0.01329868659377098, + 0.009077581577003002, + 0.013094529509544373, + 0.007749895099550486, + 0.013231240212917328, + 0.011251660995185375, + 0.0012802458368241787, + 0.0014921874972060323, + 0.001694984850473702, + 0.002549918135628104, + 0.0014097604434937239, + 0.001288432627916336, + 0.0015536483842879534, + 0.0021287789568305016, + 0.00209605460986495, + 0.0021520841401070356, + 0.0020772996358573437, + 0.00289487955160439, + 0.0025296066887676716, + 0.0022319387644529343, + 0.002298472449183464, + 0.0029100142419338226, + 0.008724736049771309, + 0.00473546888679266, + 0.005746134091168642, + 0.005030602216720581, + 0.006689003668725491, + 0.014685525558888912, + 0.004457636270672083, + 0.008320297114551067, + 0.0027816512156277895, + 0.0023208209313452244, + 0.0024149473756551743, + 0.004854846280068159, + 0.0038133196067065, + 0.0031504863873124123, + 0.003784389467909932, + 0.0029029573779553175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.033718038350343704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762551836669445 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.00762551836669445, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762551836669445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1589941680431366, + 0.08545107394456863, + 0.13012027740478516, + 0.1413392871618271, + 0.07117979228496552, + 0.06214561685919762, + 0.05994170159101486, + 0.06200994923710823, + 0.19371475279331207, + 0.03510339930653572 + ], + "credit_map": [ + 0.01250192429870367, + 0.015933366492390633, + 0.027501579374074936, + 0.020846428349614143, + 0.018919730558991432, + 0.019435398280620575, + 0.017898807302117348, + 0.02595693990588188, + 0.006821527145802975, + 0.014210711233317852, + 0.00679633067920804, + 0.00816376879811287, + 0.01854700595140457, + 0.0095713771879673, + 0.011984718032181263, + 0.0093556372448802, + 0.01052164752036333, + 0.013572383671998978, + 0.030442772433161736, + 0.015773437917232513, + 0.012564282864332199, + 0.021267300471663475, + 0.015583374537527561, + 0.01039507519453764, + 0.01055990532040596, + 0.013459128327667713, + 0.027544433251023293, + 0.01212095282971859, + 0.013107172213494778, + 0.012337896972894669, + 0.027035366743803024, + 0.02517443522810936, + 0.007413629908114672, + 0.006578235421329737, + 0.019433636218309402, + 0.010868511162698269, + 0.007715696934610605, + 0.004820387344807386, + 0.008044772781431675, + 0.006304924841970205, + 0.008915913291275501, + 0.008727123960852623, + 0.00851401872932911, + 0.00950698647648096, + 0.006129552610218525, + 0.00518027413636446, + 0.005364525131881237, + 0.009807220660150051, + 0.005746085662394762, + 0.0061541106551885605, + 0.007810965180397034, + 0.012226466089487076, + 0.007432845421135426, + 0.0060019902884960175, + 0.006916039157658815, + 0.007653198670595884, + 0.003867390099912882, + 0.005859903525561094, + 0.01094102580100298, + 0.007488248869776726, + 0.009486110880970955, + 0.0052411118522286415, + 0.007373543456196785, + 0.01175261102616787, + 0.024708272889256477, + 0.01993686519563198, + 0.01758834719657898, + 0.02001221850514412, + 0.03386852517724037, + 0.033984895795583725, + 0.02237255685031414, + 0.02124306745827198, + 0.00448573799803853, + 0.0035815814044326544, + 0.003962729126214981, + 0.003995189908891916, + 0.00569463986903429, + 0.005085085518658161, + 0.0038025358226150274, + 0.004495898727327585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.00762551836669445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342131197452545 + ], + "opposite_logprob_tokens": [ + -3.6513421535491943 + ], + "logprob_reward": -0.026342131197452545, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342131197452545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17374786734580994, + 0.09835927188396454, + 0.0692722499370575, + 0.1647857427597046, + 0.030090924352407455, + 0.038931455463171005, + 0.08436764031648636, + 0.04851571470499039, + 0.21285893023014069, + 0.07907021045684814 + ], + "credit_map": [ + 0.013974123634397984, + 0.012658076360821724, + 0.02913784794509411, + 0.01940545253455639, + 0.016263753175735474, + 0.019024066627025604, + 0.020709844306111336, + 0.04257470369338989, + 0.008006027899682522, + 0.01793510839343071, + 0.006921570282429457, + 0.011213988997042179, + 0.021396061405539513, + 0.010604165494441986, + 0.012264271266758442, + 0.010018075816333294, + 0.006412101443856955, + 0.010394719429314137, + 0.007961645722389221, + 0.0079395966604352, + 0.006971324793994427, + 0.013829628936946392, + 0.009057990275323391, + 0.0067052459344267845, + 0.01382433157414198, + 0.015315555967390537, + 0.016008855774998665, + 0.017573952674865723, + 0.024457216262817383, + 0.012730023823678493, + 0.03457131236791611, + 0.030304495245218277, + 0.0031664082780480385, + 0.0033163900952786207, + 0.007227124646306038, + 0.0036688384134322405, + 0.003357244422659278, + 0.0024780267849564552, + 0.0032561919651925564, + 0.0036207016091793776, + 0.005789404269307852, + 0.0049022892490029335, + 0.004859829321503639, + 0.006104981526732445, + 0.003924482502043247, + 0.00453087268397212, + 0.0037779216654598713, + 0.0050416733138263226, + 0.007611438166350126, + 0.012885392643511295, + 0.008718211203813553, + 0.014842372387647629, + 0.008689194917678833, + 0.00870377104729414, + 0.010079185478389263, + 0.012838070280849934, + 0.0033999814186245203, + 0.005323548801243305, + 0.007830828428268433, + 0.007749336306005716, + 0.007147543597966433, + 0.004308957606554031, + 0.005645350553095341, + 0.007110167294740677, + 0.030409367755055428, + 0.01932220160961151, + 0.0194644033908844, + 0.02666424959897995, + 0.03244154527783394, + 0.03766326606273651, + 0.02259269542992115, + 0.02430121973156929, + 0.008337512612342834, + 0.015880584716796875, + 0.00991606991738081, + 0.007542554289102554, + 0.011083903722465038, + 0.00858950987458229, + 0.009207719005644321, + 0.00851235631853342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.026342131197452545 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, close-up tracking shot focuses on a chubby, orange cat gliding uncontrollably across a polished, dark hardwood floor. Its fluffy orange tabby fur glows warmly in the soft natural light streaming in from the side windows. Its emerald green eyes widen in pure surprise at the sudden gliding motion, its ears tilted to the sides in a comical, airplane-like formation, and its mouth parted slightly, as if in a silent exclamation of surprise. The camera precisely captures its fully extended front paws, their pink pads scraping against the floor in a futile attempt to regain a sliver of grip, only to continue its gliding motion. As its body hurtles forward, almost flat on the ground, the smooth floor texture stretches across the frame into a dynamic blur. The blurred legs and feet of furniture in the background flash by, further heightening the sense of speed. The scene is imbued with warm tones and a sudden, comical sense of panic, perfectly capturing this endearing moment brimming with life.", + "seed": 32622, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_4/index_2553_seed_32622_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08798744603991508, + "vlm_logprob": -0.08798744603991508, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.08798744603991508, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006918877363204956 + ], + "opposite_logprob_tokens": [ + -5.881918907165527 + ], + "logprob_reward": -0.006918877363204956, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006918877363204956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13000419735908508, + 0.06080593168735504, + 0.1267240047454834, + 0.058933038264513016, + 0.08037212491035461, + 0.0512237474322319, + 0.0811961367726326, + 0.10216282308101654, + 0.12231658399105072, + 0.18626143038272858 + ], + "credit_map": [ + 0.009834909811615944, + 0.013654164038598537, + 0.014058269560337067, + 0.022999949753284454, + 0.021732276305556297, + 0.010762723162770271, + 0.014229238033294678, + 0.02273266203701496, + 0.004272212274372578, + 0.006407072767615318, + 0.00979722011834383, + 0.008366036228835583, + 0.007998043671250343, + 0.0072578382678329945, + 0.0078494343906641, + 0.008858075365424156, + 0.0075034452602267265, + 0.016588056460022926, + 0.012407825328409672, + 0.018825557082891464, + 0.020816586911678314, + 0.01704087108373642, + 0.015536504797637463, + 0.018005158752202988, + 0.004844596143811941, + 0.005506616551429033, + 0.003632600884884596, + 0.006555801723152399, + 0.013760482892394066, + 0.006198138929903507, + 0.009546739980578423, + 0.008888063952326775, + 0.004866456147283316, + 0.0066382307559251785, + 0.01439305953681469, + 0.008155609481036663, + 0.012745780870318413, + 0.014068990014493465, + 0.010684659704566002, + 0.008819337002933025, + 0.0054437341168522835, + 0.0054757618345320225, + 0.00551799638196826, + 0.007860609330236912, + 0.0071165855042636395, + 0.004606474190950394, + 0.0068462444469332695, + 0.008356341160833836, + 0.007603245787322521, + 0.009589350782334805, + 0.010301992297172546, + 0.01602408103644848, + 0.009592199698090553, + 0.007343496661633253, + 0.012106667272746563, + 0.008635101839900017, + 0.0117200231179595, + 0.010713867843151093, + 0.01251577865332365, + 0.012793506495654583, + 0.01642887108027935, + 0.009148046374320984, + 0.011431594379246235, + 0.017411136999726295, + 0.008113235235214233, + 0.006622808985412121, + 0.01665586419403553, + 0.01509232260286808, + 0.03758439049124718, + 0.01795254275202751, + 0.011117172427475452, + 0.009178247302770615, + 0.009705731645226479, + 0.013244055211544037, + 0.00806029886007309, + 0.040510255843400955, + 0.04730972647666931, + 0.029324989765882492, + 0.022264093160629272, + 0.01584228128194809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.006918877363204956 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38704702258110046 + ], + "opposite_logprob_tokens": [ + -1.1370470523834229 + ], + "logprob_reward": -0.38704702258110046, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38704702258110046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0635039433836937, + 0.06020902097225189, + 0.09480046480894089, + 0.07432318478822708, + 0.04046155884861946, + 0.1360723227262497, + 0.09124855697154999, + 0.06834530085325241, + 0.09265510737895966, + 0.27838051319122314 + ], + "credit_map": [ + 0.006275470834225416, + 0.00844422448426485, + 0.0077415769919753075, + 0.010462689213454723, + 0.005286317318677902, + 0.006315730046480894, + 0.005821292754262686, + 0.013156644068658352, + 0.006910006515681744, + 0.005114305764436722, + 0.011914363130927086, + 0.007882476784288883, + 0.0040098028257489204, + 0.008671451359987259, + 0.0060025048442184925, + 0.00970410741865635, + 0.00919287372380495, + 0.016588596627116203, + 0.008650047704577446, + 0.015711428597569466, + 0.009138233959674835, + 0.0090587567538023, + 0.011139844544231892, + 0.01532068382948637, + 0.011657420545816422, + 0.010831576772034168, + 0.009081545285880566, + 0.007401165086776018, + 0.011343895457684994, + 0.007938599213957787, + 0.009208868257701397, + 0.006860113237053156, + 0.0026619916316121817, + 0.003819873556494713, + 0.005249303299933672, + 0.0065527246333658695, + 0.006852824240922928, + 0.0056776488199830055, + 0.005057694390416145, + 0.004589498043060303, + 0.010547512210905552, + 0.010990535840392113, + 0.01621747761964798, + 0.024705372750759125, + 0.0169037077575922, + 0.01607740856707096, + 0.017162170261144638, + 0.02346814237535, + 0.00960138626396656, + 0.013962984085083008, + 0.008320720866322517, + 0.015584920533001423, + 0.01050669513642788, + 0.008171437308192253, + 0.014620685949921608, + 0.01047972496598959, + 0.006654960103332996, + 0.00893025565892458, + 0.005830899812281132, + 0.009790697135031223, + 0.018997352570295334, + 0.004585267975926399, + 0.00796512607485056, + 0.0055907475762069225, + 0.006226150784641504, + 0.0037056587170809507, + 0.019139761105179787, + 0.009582259692251682, + 0.03529468923807144, + 0.007555813062936068, + 0.0059223235584795475, + 0.005228453781455755, + 0.014525769278407097, + 0.01634669303894043, + 0.013176496140658855, + 0.06012744456529617, + 0.07646016031503677, + 0.04161617532372475, + 0.03296326845884323, + 0.02316451258957386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.38704702258110046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033711932599544525 + ], + "opposite_logprob_tokens": [ + -3.4087119102478027 + ], + "logprob_reward": -0.033711932599544525, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033711932599544525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17124523222446442, + 0.13178277015686035, + 0.12453651428222656, + 0.06071125343441963, + 0.17415255308151245, + 0.08420710265636444, + 0.04904991015791893, + 0.06782012432813644, + 0.055333733558654785, + 0.08116080611944199 + ], + "credit_map": [ + 0.01574770361185074, + 0.016904374584555626, + 0.019062060862779617, + 0.03554511442780495, + 0.028225289657711983, + 0.011941562406718731, + 0.021613316610455513, + 0.02220579795539379, + 0.010171926580369473, + 0.02599530853331089, + 0.028240147978067398, + 0.01107549574226141, + 0.014342567883431911, + 0.01278773508965969, + 0.01183572132140398, + 0.017333859577775, + 0.008073454722762108, + 0.01631913147866726, + 0.012157163582742214, + 0.017869189381599426, + 0.021206418052315712, + 0.012469949200749397, + 0.017464755102992058, + 0.018976442515850067, + 0.004081999883055687, + 0.004733323119580746, + 0.004042021930217743, + 0.006812899373471737, + 0.01229175552725792, + 0.012878273613750935, + 0.009496327489614487, + 0.006374644581228495, + 0.010197276249527931, + 0.012928848154842854, + 0.02255384251475334, + 0.016986727714538574, + 0.03534034267067909, + 0.027333449572324753, + 0.02717980556190014, + 0.021632255986332893, + 0.005938755813986063, + 0.0060683549381792545, + 0.009969730861485004, + 0.015793176367878914, + 0.011524428613483906, + 0.009259354323148727, + 0.011489425785839558, + 0.01416387315839529, + 0.003664890769869089, + 0.003960796166211367, + 0.007243137340992689, + 0.009966353885829449, + 0.005666404031217098, + 0.005169333424419165, + 0.00674661248922348, + 0.0066323792561888695, + 0.0063898456282913685, + 0.009711437858641148, + 0.006478673778474331, + 0.010814296081662178, + 0.013028216548264027, + 0.0040678163059055805, + 0.006798832677304745, + 0.01053099799901247, + 0.004838204476982355, + 0.004254416562616825, + 0.007006066385656595, + 0.007221968844532967, + 0.010188402608036995, + 0.010257869027554989, + 0.006336839869618416, + 0.005229963455349207, + 0.008550656028091908, + 0.006071988493204117, + 0.008747998625040054, + 0.013484706170856953, + 0.01644267328083515, + 0.009097659960389137, + 0.01010989397764206, + 0.008655227720737457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.033711932599544525 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633760545402765 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004633760545402765, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633760545402765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12791210412979126, + 0.13451246917247772, + 0.08438339829444885, + 0.07484851777553558, + 0.0709025040268898, + 0.10897213220596313, + 0.07272901386022568, + 0.0841175988316536, + 0.08419987559318542, + 0.15742243826389313 + ], + "credit_map": [ + 0.014362191781401634, + 0.01259158831089735, + 0.01542453933507204, + 0.02888437733054161, + 0.010413003154098988, + 0.009268217720091343, + 0.010659252293407917, + 0.02630893513560295, + 0.011187639087438583, + 0.02314269356429577, + 0.022237854078412056, + 0.020319180563092232, + 0.018386300653219223, + 0.013053013011813164, + 0.013784576207399368, + 0.012401207350194454, + 0.007600697688758373, + 0.010885600000619888, + 0.007396405562758446, + 0.016835205256938934, + 0.009087596088647842, + 0.009733607061207294, + 0.008879521861672401, + 0.013964764773845673, + 0.004023248795419931, + 0.006643784232437611, + 0.005765244364738464, + 0.009113387204706669, + 0.015282634645700455, + 0.011523501016199589, + 0.012510623782873154, + 0.009986094199120998, + 0.007332052569836378, + 0.010612767189741135, + 0.008834603242576122, + 0.00992963369935751, + 0.009895273484289646, + 0.007968389429152012, + 0.008534451015293598, + 0.007795331533998251, + 0.007894324138760567, + 0.009500282816588879, + 0.010687713511288166, + 0.021843653172254562, + 0.014571855776011944, + 0.011498292908072472, + 0.014631349593400955, + 0.018344657495617867, + 0.004335673060268164, + 0.00686879875138402, + 0.009666848927736282, + 0.009330181404948235, + 0.010248834267258644, + 0.012311660684645176, + 0.007511535193771124, + 0.012455478310585022, + 0.009817625395953655, + 0.009939946234226227, + 0.008596113882958889, + 0.011466564610600471, + 0.016282150521874428, + 0.00837999302893877, + 0.01001513097435236, + 0.009620074182748795, + 0.005615534260869026, + 0.004423191770911217, + 0.013922742567956448, + 0.010302742011845112, + 0.031188583001494408, + 0.007166428491473198, + 0.005880805663764477, + 0.005699852481484413, + 0.009204253554344177, + 0.009202204644680023, + 0.010202483274042606, + 0.034593842923641205, + 0.044283125549554825, + 0.02079213410615921, + 0.01616080477833748, + 0.012983591295778751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.004633760545402765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625637110322714 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625637110322714, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625637110322714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0897231251001358, + 0.07665607333183289, + 0.05052490532398224, + 0.052933406084775925, + 0.07827664166688919, + 0.17437998950481415, + 0.05450437217950821, + 0.04971744120121002, + 0.05375886708498001, + 0.31952518224716187 + ], + "credit_map": [ + 0.008015763014554977, + 0.007642444688826799, + 0.011562418192625046, + 0.02085021324455738, + 0.007636501453816891, + 0.005415070336312056, + 0.006809117738157511, + 0.021791590377688408, + 0.006196996662765741, + 0.008559701964259148, + 0.014508731663227081, + 0.01401740126311779, + 0.007851059548556805, + 0.009468359872698784, + 0.007099473848938942, + 0.008954344317317009, + 0.004253926686942577, + 0.00682655768468976, + 0.005381590686738491, + 0.007700716145336628, + 0.0061105601489543915, + 0.006323203910142183, + 0.005271503236144781, + 0.008656840771436691, + 0.00526862358674407, + 0.006699416320770979, + 0.003966035321354866, + 0.0049784742295742035, + 0.011400799266994, + 0.0063361600041389465, + 0.008597959764301777, + 0.005685934331268072, + 0.008930242620408535, + 0.011276759207248688, + 0.013612539507448673, + 0.01006253995001316, + 0.0075029004365205765, + 0.012146607041358948, + 0.006799641996622086, + 0.00794539786875248, + 0.00843011774122715, + 0.008656730875372887, + 0.01772462949156761, + 0.016246305778622627, + 0.011316991411149502, + 0.042714864015579224, + 0.01230879221111536, + 0.05698155611753464, + 0.003959497902542353, + 0.007705868687480688, + 0.0055513158440589905, + 0.011045826599001884, + 0.006148246582597494, + 0.005682005546987057, + 0.008188808336853981, + 0.006222797092050314, + 0.005691188387572765, + 0.004407298285514116, + 0.006382177118211985, + 0.005897659342736006, + 0.007725652307271957, + 0.007912451401352882, + 0.004351602867245674, + 0.0073494077660143375, + 0.005114844534546137, + 0.0027677216567099094, + 0.00945159513503313, + 0.006102374289184809, + 0.0190731193870306, + 0.00426261592656374, + 0.0033311080187559128, + 0.00365548487752676, + 0.01644742488861084, + 0.021156568080186844, + 0.011810851283371449, + 0.06429493427276611, + 0.08247054368257523, + 0.049604371190071106, + 0.04291978105902672, + 0.030820704996585846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.007625637110322714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, close-up tracking shot focuses on a chubby, orange cat gliding uncontrollably across a polished, dark hardwood floor. Its fluffy orange tabby fur glows warmly in the soft natural light streaming in from the side windows. Its emerald green eyes widen in pure surprise at the sudden gliding motion, its ears tilted to the sides in a comical, airplane-like formation, and its mouth parted slightly, as if in a silent exclamation of surprise. The camera precisely captures its fully extended front paws, their pink pads scraping against the floor in a futile attempt to regain a sliver of grip, only to continue its gliding motion. As its body hurtles forward, almost flat on the ground, the smooth floor texture stretches across the frame into a dynamic blur. The blurred legs and feet of furniture in the background flash by, further heightening the sense of speed. The scene is imbued with warm tones and a sudden, comical sense of panic, perfectly capturing this endearing moment brimming with life.", + "seed": 32623, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_4/index_2553_seed_32623_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0787951460108161, + "vlm_logprob": -0.0787951460108161, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.0787951460108161, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011244144290685654 + ], + "opposite_logprob_tokens": [ + -4.886244297027588 + ], + "logprob_reward": -0.011244144290685654, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011244144290685654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14100128412246704, + 0.10286161303520203, + 0.046763766556978226, + 0.09856592863798141, + 0.13009384274482727, + 0.03936968371272087, + 0.10654740035533905, + 0.0875946506857872, + 0.12528951466083527, + 0.12191226333379745 + ], + "credit_map": [ + 0.016179563477635384, + 0.024524901062250137, + 0.012839575298130512, + 0.017885897308588028, + 0.018300902098417282, + 0.014493749476969242, + 0.012786314822733402, + 0.023990394547581673, + 0.007502325344830751, + 0.007730467244982719, + 0.023427540436387062, + 0.009594427421689034, + 0.01280864980071783, + 0.024826519191265106, + 0.009948630817234516, + 0.0070230597630143166, + 0.007524334359914064, + 0.006376350298523903, + 0.007728366181254387, + 0.005036734044551849, + 0.006858875975012779, + 0.0029040679801255465, + 0.005666639190167189, + 0.004668398294597864, + 0.00839163362979889, + 0.012384410947561264, + 0.016834571957588196, + 0.012347733601927757, + 0.009416578337550163, + 0.010515851899981499, + 0.012437011115252972, + 0.016238145530223846, + 0.011108492501080036, + 0.021341193467378616, + 0.013100226409733295, + 0.01884397678077221, + 0.022355658933520317, + 0.00802659336477518, + 0.019421715289354324, + 0.01589600183069706, + 0.006021429318934679, + 0.005989253520965576, + 0.004894870333373547, + 0.0041318475268781185, + 0.007564425002783537, + 0.0030792932957410812, + 0.00421061459928751, + 0.0034779568668454885, + 0.01608877442777157, + 0.008469142951071262, + 0.011468234471976757, + 0.013713385909795761, + 0.01794356293976307, + 0.007242577616125345, + 0.013978580944240093, + 0.017643149942159653, + 0.007135408464819193, + 0.00990211870521307, + 0.013663619756698608, + 0.014155329205095768, + 0.016117574647068977, + 0.00507313059642911, + 0.012440573424100876, + 0.009106896817684174, + 0.008181964047253132, + 0.009786263108253479, + 0.03202385455369949, + 0.017797833308577538, + 0.01995578408241272, + 0.009443446062505245, + 0.01637151837348938, + 0.011728869751095772, + 0.010430028662085533, + 0.02330935001373291, + 0.010490589775145054, + 0.01610635779798031, + 0.018601106479763985, + 0.011852710507810116, + 0.014552724547684193, + 0.01656939834356308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.011244144290685654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34861063957214355 + ], + "opposite_logprob_tokens": [ + -1.2236106395721436 + ], + "logprob_reward": -0.34861063957214355, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34861063957214355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07772266864776611, + 0.16817259788513184, + 0.09441591799259186, + 0.13675768673419952, + 0.02396315708756447, + 0.017975393682718277, + 0.0948798805475235, + 0.09205754101276398, + 0.05773070454597473, + 0.23632444441318512 + ], + "credit_map": [ + 0.008724444545805454, + 0.009650496765971184, + 0.011154421605169773, + 0.010563422925770283, + 0.011374592781066895, + 0.006750172469764948, + 0.00896766409277916, + 0.010537467896938324, + 0.028681954368948936, + 0.012707888148725033, + 0.04389725998044014, + 0.029346119612455368, + 0.011774887330830097, + 0.017498353496193886, + 0.014465953223407269, + 0.009800185449421406, + 0.015776244923472404, + 0.014551810920238495, + 0.012721486389636993, + 0.010815689340233803, + 0.014171076938509941, + 0.00487404502928257, + 0.013689246959984303, + 0.007816320285201073, + 0.014692723751068115, + 0.018346088007092476, + 0.021244866773486137, + 0.0159552413970232, + 0.01630689576268196, + 0.019654422998428345, + 0.011723225004971027, + 0.018834222108125687, + 0.002289342461153865, + 0.003208791371434927, + 0.0028130793944001198, + 0.0023808206897228956, + 0.0038729894440621138, + 0.002571258693933487, + 0.0033450722694396973, + 0.0034818043932318687, + 0.003607366932556033, + 0.002306718612089753, + 0.0018079110886901617, + 0.0021670677233487368, + 0.002921855775639415, + 0.0015864152228459716, + 0.0016380748711526394, + 0.0019399826414883137, + 0.013205092400312424, + 0.008347256109118462, + 0.009393705986440182, + 0.008790507912635803, + 0.01860196702182293, + 0.006566848140209913, + 0.013411929830908775, + 0.016562582924962044, + 0.008550663478672504, + 0.010874524712562561, + 0.01067723985761404, + 0.012274466454982758, + 0.014688397757709026, + 0.006157243624329567, + 0.01654154621064663, + 0.01229346077889204, + 0.007612396962940693, + 0.004913268610835075, + 0.007866451516747475, + 0.007372692227363586, + 0.0060150292702019215, + 0.009385664016008377, + 0.00677278870716691, + 0.007792415097355843, + 0.014249931089580059, + 0.023557620123028755, + 0.07028040289878845, + 0.02278793416917324, + 0.023211859166622162, + 0.013206363655626774, + 0.021240929141640663, + 0.04778940975666046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.34861063957214355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294946178793907 + ], + "opposite_logprob_tokens": [ + -3.7732949256896973 + ], + "logprob_reward": -0.023294946178793907, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294946178793907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16938148438930511, + 0.16852311789989471, + 0.034931737929582596, + 0.07207030057907104, + 0.03641624003648758, + 0.039414964616298676, + 0.07684721052646637, + 0.1920223981142044, + 0.12081625312566757, + 0.08957630395889282 + ], + "credit_map": [ + 0.018285533413290977, + 0.027271749451756477, + 0.01891644112765789, + 0.019093932583928108, + 0.023675797507166862, + 0.013062555342912674, + 0.020867984741926193, + 0.02820749208331108, + 0.015124578960239887, + 0.022291461005806923, + 0.0411086343228817, + 0.01972772739827633, + 0.01842588558793068, + 0.02309308387339115, + 0.01968468725681305, + 0.009067061357200146, + 0.004551880992949009, + 0.004151880741119385, + 0.004617699887603521, + 0.005171599797904491, + 0.006246043834835291, + 0.001936434768140316, + 0.005217219237238169, + 0.0030389775056391954, + 0.006062028463929892, + 0.010207249782979488, + 0.010302189737558365, + 0.009744742885231972, + 0.008105329237878323, + 0.006033965852111578, + 0.009872600436210632, + 0.011742199771106243, + 0.0036928802728652954, + 0.0064170160330832005, + 0.004522103350609541, + 0.00408145459368825, + 0.007368284743279219, + 0.0030620235484093428, + 0.003882306395098567, + 0.0033901722636073828, + 0.005631951615214348, + 0.004239530302584171, + 0.0061178067699074745, + 0.003872805507853627, + 0.006293036509305239, + 0.003041899763047695, + 0.005856093019247055, + 0.0043618399649858475, + 0.01072798203676939, + 0.005890895612537861, + 0.007768784649670124, + 0.009829571470618248, + 0.013317517936229706, + 0.0056563992984592915, + 0.009051688015460968, + 0.014604372903704643, + 0.015627549961209297, + 0.02452917769551277, + 0.02187955006957054, + 0.023722955957055092, + 0.03679211810231209, + 0.010279537178575993, + 0.03581608086824417, + 0.02337542548775673, + 0.012755698524415493, + 0.012952283956110477, + 0.01638156920671463, + 0.014074921607971191, + 0.025198176503181458, + 0.009885587729513645, + 0.017710845917463303, + 0.011857167817652225, + 0.006077853962779045, + 0.006110691465437412, + 0.03234746307134628, + 0.007574555929750204, + 0.004873654805123806, + 0.004395959433168173, + 0.006454088259488344, + 0.02174203284084797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.023294946178793907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093126393854618 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093126393854618, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093126393854618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10004284977912903, + 0.230148047208786, + 0.09596787393093109, + 0.08225776255130768, + 0.04924558848142624, + 0.028502369299530983, + 0.14107030630111694, + 0.11990034580230713, + 0.029229607433080673, + 0.12363533675670624 + ], + "credit_map": [ + 0.012079650536179543, + 0.012339157983660698, + 0.012060103937983513, + 0.010164901614189148, + 0.016398126259446144, + 0.012616679072380066, + 0.011754621751606464, + 0.012629596516489983, + 0.04088905453681946, + 0.015347677282989025, + 0.05595330148935318, + 0.03694780543446541, + 0.01731569692492485, + 0.029822515323758125, + 0.020479291677474976, + 0.01339268684387207, + 0.015872713178396225, + 0.012990603223443031, + 0.017396802082657814, + 0.011777141131460667, + 0.01311588380485773, + 0.004950822796672583, + 0.010978621430695057, + 0.008885275572538376, + 0.009400630369782448, + 0.010765444487333298, + 0.011271324940025806, + 0.008561198599636555, + 0.010646789334714413, + 0.013726554811000824, + 0.007217358332127333, + 0.010668450966477394, + 0.005361509975045919, + 0.010650888085365295, + 0.00552580039948225, + 0.004465130157768726, + 0.011734561063349247, + 0.00221550022251904, + 0.0059011466801166534, + 0.0033910435158759356, + 0.00330505077727139, + 0.004744885955005884, + 0.003087373450398445, + 0.003336094319820404, + 0.004531323444098234, + 0.0028901773039251566, + 0.003476299112662673, + 0.003131162142381072, + 0.02003547176718712, + 0.01245085895061493, + 0.014066552743315697, + 0.013227500021457672, + 0.027112895622849464, + 0.009903067722916603, + 0.020098090171813965, + 0.02417585998773575, + 0.02040989324450493, + 0.013465034775435925, + 0.011789166368544102, + 0.015011167153716087, + 0.014895735308527946, + 0.013457035645842552, + 0.016891606152057648, + 0.013980698771774769, + 0.0023099377285689116, + 0.002587586874142289, + 0.004374724347144365, + 0.003552932757884264, + 0.004330283496528864, + 0.003669848432764411, + 0.004625650588423014, + 0.0037786439061164856, + 0.008282791823148727, + 0.006546023767441511, + 0.04612521827220917, + 0.010331129655241966, + 0.005908160470426083, + 0.008068825118243694, + 0.008397270925343037, + 0.02997591346502304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.004093126393854618 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat gliding uncontrollably across a polished dark hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the cat's emerald green eyes wide and its ears tilted to the sides in an airplane-like formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle in a close-up tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the background furniture legs and feet visible as blurred shapes indicating speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732873618602753 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732873618602753, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732873618602753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06181250512599945, + 0.0949033796787262, + 0.04115959256887436, + 0.0733795240521431, + 0.0218708124011755, + 0.03366483002901077, + 0.07797084003686905, + 0.11074492335319519, + 0.03293168544769287, + 0.4515618681907654 + ], + "credit_map": [ + 0.007764044217765331, + 0.007106900680810213, + 0.008764141239225864, + 0.006739187519997358, + 0.010463502258062363, + 0.005034467205405235, + 0.006914876867085695, + 0.009025385603308678, + 0.015751522034406662, + 0.007369066122919321, + 0.0141261862590909, + 0.014635179191827774, + 0.00821499153971672, + 0.01781075820326805, + 0.01057289820164442, + 0.006422773934900761, + 0.009786801412701607, + 0.00734275346621871, + 0.006948579102754593, + 0.005015305243432522, + 0.004177730530500412, + 0.001511764945462346, + 0.0029164464212954044, + 0.003460211679339409, + 0.00828340370208025, + 0.009629899635910988, + 0.009260188788175583, + 0.0087935421615839, + 0.008731629699468613, + 0.011039153672754765, + 0.0071447486989200115, + 0.010496950708329678, + 0.0022368337959051132, + 0.0033140378072857857, + 0.0020670800004154444, + 0.0024141829926520586, + 0.003810873720794916, + 0.002788445446640253, + 0.002446574391797185, + 0.0027927840128540993, + 0.006164306774735451, + 0.004486455116420984, + 0.004245517309755087, + 0.004296950530260801, + 0.004536849446594715, + 0.003147607436403632, + 0.0028087988030165434, + 0.003978345077484846, + 0.011839556507766247, + 0.006454817019402981, + 0.007662795949727297, + 0.008445856161415577, + 0.013934079557657242, + 0.0056465426459908485, + 0.00946130696684122, + 0.014525889419019222, + 0.01471586525440216, + 0.012496969662606716, + 0.011938069947063923, + 0.015674656257033348, + 0.014930913224816322, + 0.008919375017285347, + 0.019697781652212143, + 0.01237129233777523, + 0.002315465360879898, + 0.004422595724463463, + 0.0047597261145710945, + 0.004653162322938442, + 0.00507226400077343, + 0.002208041027188301, + 0.005901068449020386, + 0.0035993640776723623, + 0.04007498919963837, + 0.08728151768445969, + 0.08711349219083786, + 0.04267357289791107, + 0.052660416811704636, + 0.03226181119680405, + 0.03429364040493965, + 0.07520244270563126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.006732873618602753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_5/index_tensor([8218, 8218])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_5/index_tensor([8218, 8218])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..805ac52cf1466ac02a58afae1bed5a1defe34baf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_5/index_tensor([8218, 8218])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, long shot, shot from a slightly upward angle, first focuses on a weathered beacon tower on the ancient Great Wall. Four soldiers clad in dark iron plate armor are at the height of their triumphant ecstasy. Their armor, composed of thousands of plates, is etched with scratches and dents from battle, its leather straps rustling in the strong wind. The bright red tassels of their iron helmets dance wildly in the biting mountain breeze. Their weathered faces, stained with sweat and dust, are now filled with undisguised, heartfelt joy, their mouths open in a silent roar. Steadily and powerfully, they raise their spears and ring-handled swords high into the sky, the sharp metal tips gleaming blindingly in the golden afternoon sun that pierces the clouds. As the camera steadily and deliberately pulls back, the image gradually reveals the beacon towers built of massive gray stones beneath their feet, and the entire Great Wall, winding and undulating like a giant dragon, disappearing into the vast mountains in the distance. Thick clouds swirl across the sky, and beams of light pouring through the torn rifts cast vast, shifting shadows across the mountain ranges. The entire scene, a stark palette of gray, brown, and iron black, is infused with vitality by the touch of sunlight and the red of the helmet tassels, creating a sense of epic grandeur, primal power, and the powerful emotion of a hard-won victory in a harsh environment.", + "seed": 33471, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_5/index_8218_seed_33471_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1950676292181015, + "vlm_logprob": -0.1950676292181015, + "vlm_yes_no_margin": 2.1499999880790712, + "vlm_token_credit": -0.1950676292181015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4342113137245178 + ], + "opposite_logprob_tokens": [ + -1.059211254119873 + ], + "logprob_reward": -0.4342113137245178, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4342113137245178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17199264466762543, + 0.11391206085681915, + 0.08182670176029205, + 0.09703227877616882, + 0.13548453152179718, + 0.056491557508707047, + 0.05710302293300629, + 0.08116399496793747, + 0.1633732169866562, + 0.041619908064603806 + ], + "credit_map": [ + 0.02428276278078556, + 0.03205260634422302, + 0.02292233519256115, + 0.026219697669148445, + 0.01582820527255535, + 0.019885653629899025, + 0.014452449046075344, + 0.016348950564861298, + 0.009794133715331554, + 0.01003606989979744, + 0.02313908562064171, + 0.020851420238614082, + 0.00867411121726036, + 0.007653194013983011, + 0.013335146941244602, + 0.020428907126188278, + 0.011625483632087708, + 0.012797405943274498, + 0.012025869451463223, + 0.009139083325862885, + 0.009455309249460697, + 0.007544215302914381, + 0.00992915965616703, + 0.009310180321335793, + 0.007673195097595453, + 0.014307714067399502, + 0.011707567609846592, + 0.016318554058670998, + 0.014904703944921494, + 0.01265659835189581, + 0.0074070896953344345, + 0.012056862004101276, + 0.013668989762663841, + 0.010035688057541847, + 0.026347417384386063, + 0.021062318235635757, + 0.013203993439674377, + 0.015408718027174473, + 0.014589043334126472, + 0.021168364211916924, + 0.005842171609401703, + 0.004531992599368095, + 0.008614891208708286, + 0.008890817873179913, + 0.007779478561133146, + 0.00906309112906456, + 0.004537966102361679, + 0.007231154479086399, + 0.00520308455452323, + 0.004770240280777216, + 0.013993460685014725, + 0.009389378130435944, + 0.004515344742685556, + 0.0050690365023911, + 0.005933905020356178, + 0.008228582330048084, + 0.021378861740231514, + 0.008489453233778477, + 0.012867857702076435, + 0.00860702432692051, + 0.0064244819805026054, + 0.00501718744635582, + 0.006747134029865265, + 0.01163200382143259, + 0.01694709062576294, + 0.05591036379337311, + 0.012982853688299656, + 0.010763943195343018, + 0.03004675731062889, + 0.009983265772461891, + 0.007289208471775055, + 0.019449736922979355, + 0.011279535479843616, + 0.0037471153773367405, + 0.003212571609765291, + 0.003940868191421032, + 0.007640708237886429, + 0.003439289052039385, + 0.002495028078556061, + 0.005864795297384262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.4342113137245178 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283104240894318 + ], + "opposite_logprob_tokens": [ + -2.0178310871124268 + ], + "logprob_reward": -0.14283104240894318, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283104240894318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20528794825077057, + 0.08171803504228592, + 0.1351839154958725, + 0.10252879559993744, + 0.166202574968338, + 0.056181129068136215, + 0.045448094606399536, + 0.05579878017306328, + 0.05625488609075546, + 0.09539587795734406 + ], + "credit_map": [ + 0.025600126013159752, + 0.03651145473122597, + 0.02527388371527195, + 0.029418502002954483, + 0.019336525350809097, + 0.02521177940070629, + 0.026073461398482323, + 0.01786220259964466, + 0.0069340551272034645, + 0.006802450865507126, + 0.008647841401398182, + 0.010679705068469048, + 0.008389039896428585, + 0.011870147660374641, + 0.01082431897521019, + 0.01757047325372696, + 0.019599610939621925, + 0.01674075610935688, + 0.0215636994689703, + 0.015212027356028557, + 0.015596483834087849, + 0.013453053310513496, + 0.016499890014529228, + 0.016518378630280495, + 0.009317321702837944, + 0.012145157903432846, + 0.011550645343959332, + 0.014010476879775524, + 0.014719181694090366, + 0.013314411044120789, + 0.011144709773361683, + 0.01632688380777836, + 0.012169895693659782, + 0.016780776903033257, + 0.01704861968755722, + 0.019966760650277138, + 0.036093562841415405, + 0.03230082616209984, + 0.014508659020066261, + 0.017333464697003365, + 0.012677476741373539, + 0.005732773337513208, + 0.008279665373265743, + 0.0071363565512001514, + 0.005536728538572788, + 0.00504565704613924, + 0.003858019132167101, + 0.00791444443166256, + 0.0037483605556190014, + 0.003451949218288064, + 0.01049252413213253, + 0.007323601748794317, + 0.004322992637753487, + 0.0042334143072366714, + 0.004987596999853849, + 0.006887651514261961, + 0.006163908634334803, + 0.006847964134067297, + 0.010830223560333252, + 0.009327427484095097, + 0.004108136519789696, + 0.003865996841341257, + 0.0062479679472744465, + 0.008407148532569408, + 0.006553032901138067, + 0.016911571845412254, + 0.005424552131444216, + 0.004461764823645353, + 0.009788084775209427, + 0.003552059642970562, + 0.0033565890043973923, + 0.006207222584635019, + 0.03121970035135746, + 0.008293447084724903, + 0.004250315949320793, + 0.005322773475199938, + 0.025998704135417938, + 0.0053660208359360695, + 0.003362139454111457, + 0.011582771316170692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.14283104240894318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330426312983036 + ], + "opposite_logprob_tokens": [ + -3.7733042240142822 + ], + "logprob_reward": -0.02330426312983036, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330426312983036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1623150259256363, + 0.07661895453929901, + 0.15009945631027222, + 0.1305842101573944, + 0.23715312778949738, + 0.05888243764638901, + 0.05454946681857109, + 0.03884509950876236, + 0.0371728390455246, + 0.053779419511556625 + ], + "credit_map": [ + 0.02534584328532219, + 0.04588048905134201, + 0.017522484064102173, + 0.019908055663108826, + 0.013720924966037273, + 0.015683112666010857, + 0.013957473449409008, + 0.010296640917658806, + 0.006722867488861084, + 0.0057343472726643085, + 0.0117284394800663, + 0.011352275498211384, + 0.009791608899831772, + 0.008127093315124512, + 0.011443359777331352, + 0.011718960478901863, + 0.017562057822942734, + 0.01958473026752472, + 0.022238530218601227, + 0.013767745345830917, + 0.019957730546593666, + 0.014598273672163486, + 0.02403114177286625, + 0.018359245732426643, + 0.012417296878993511, + 0.014262991957366467, + 0.017368625849485397, + 0.02661317214369774, + 0.015638509765267372, + 0.01492997631430626, + 0.012598938308656216, + 0.016754690557718277, + 0.03145955130457878, + 0.024415068328380585, + 0.031392429023981094, + 0.04466358199715614, + 0.024224627763032913, + 0.031872138381004333, + 0.01679360494017601, + 0.03233210742473602, + 0.009568407200276852, + 0.006760741118341684, + 0.007394628133624792, + 0.009100498631596565, + 0.0059174965135753155, + 0.008508832193911076, + 0.004705105442553759, + 0.006926730740815401, + 0.006693169940263033, + 0.0036436154041439295, + 0.010156448930501938, + 0.008768884465098381, + 0.004465540871024132, + 0.00833008624613285, + 0.004966427572071552, + 0.007525292225182056, + 0.003963317722082138, + 0.006126004736870527, + 0.003980689682066441, + 0.006375249475240707, + 0.003541387151926756, + 0.005057371687144041, + 0.005456977523863316, + 0.0043441010639071465, + 0.0035678218118846416, + 0.005895585287362337, + 0.00677173025906086, + 0.003953245934098959, + 0.005293773952871561, + 0.0038010473363101482, + 0.0035615747328847647, + 0.004328061826527119, + 0.005998601671308279, + 0.007807701826095581, + 0.005088907666504383, + 0.006226623430848122, + 0.009843701496720314, + 0.004042393993586302, + 0.003619804745540023, + 0.011151686310768127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.02330426312983036 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350608095526695 + ], + "opposite_logprob_tokens": [ + -3.651350498199463 + ], + "logprob_reward": -0.026350608095526695, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350608095526695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11029835045337677, + 0.07863656431436539, + 0.4121781289577484, + 0.06067528948187828, + 0.07144583761692047, + 0.03186211735010147, + 0.0447402186691761, + 0.15434923768043518, + 0.018481556326150894, + 0.017332660034298897 + ], + "credit_map": [ + 0.014734562486410141, + 0.0165295097976923, + 0.017731690779328346, + 0.015870800241827965, + 0.009197799488902092, + 0.008269957266747952, + 0.012281432747840881, + 0.015682602301239967, + 0.006231740117073059, + 0.006034460384398699, + 0.010751314461231232, + 0.01245425920933485, + 0.006772617809474468, + 0.007859745062887669, + 0.01067810133099556, + 0.017854323610663414, + 0.052371881902217865, + 0.08093180507421494, + 0.06438253074884415, + 0.03819252550601959, + 0.04945677891373634, + 0.03562444448471069, + 0.04624428600072861, + 0.04497389495372772, + 0.005577002186328173, + 0.006914193741977215, + 0.0074869622476398945, + 0.007951179519295692, + 0.009848583489656448, + 0.007362791337072849, + 0.006745822727680206, + 0.008788755163550377, + 0.005924566648900509, + 0.005950199905782938, + 0.006712878122925758, + 0.008851169608533382, + 0.014681771397590637, + 0.013561926782131195, + 0.005853625945746899, + 0.00990970153361559, + 0.006893153302371502, + 0.0033828075975179672, + 0.004249352030456066, + 0.003749071853235364, + 0.004731542430818081, + 0.0032033519819378853, + 0.0020319344475865364, + 0.003620903706178069, + 0.006751418113708496, + 0.004308713134378195, + 0.007525634951889515, + 0.005728063639253378, + 0.004238314460963011, + 0.007155613042414188, + 0.0028138039633631706, + 0.00621865876019001, + 0.014585105702280998, + 0.01181239914149046, + 0.01438358798623085, + 0.018929673358798027, + 0.01540396362543106, + 0.038747966289520264, + 0.018646661192178726, + 0.02183988131582737, + 0.002237513894215226, + 0.00292748655192554, + 0.002545521594583988, + 0.0024631705600768328, + 0.002339094178751111, + 0.0024782950058579445, + 0.0016051928978413343, + 0.001885280362330377, + 0.0025226185098290443, + 0.0037792243529111147, + 0.0018376677762717009, + 0.0018757423385977745, + 0.0023754623252898455, + 0.001971747260540724, + 0.0011487778974696994, + 0.0018214192241430283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.026350608095526695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34864091873168945 + ], + "opposite_logprob_tokens": [ + -1.2236409187316895 + ], + "logprob_reward": -0.34864091873168945, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34864091873168945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13932214677333832, + 0.06037936359643936, + 0.168205127120018, + 0.09766847640275955, + 0.09162230044603348, + 0.06412915885448456, + 0.046749286353588104, + 0.036119550466537476, + 0.21585039794445038, + 0.07995418459177017 + ], + "credit_map": [ + 0.015915673226118088, + 0.023157767951488495, + 0.024954669177532196, + 0.023708971217274666, + 0.012960677966475487, + 0.009719517081975937, + 0.016599414870142937, + 0.012305443175137043, + 0.006190244574099779, + 0.004644356202334166, + 0.00633088918402791, + 0.008365883491933346, + 0.007032476365566254, + 0.007362966425716877, + 0.009430969133973122, + 0.011021573096513748, + 0.024885300546884537, + 0.025419460609555244, + 0.02510962262749672, + 0.017376085743308067, + 0.020769894123077393, + 0.015165001153945923, + 0.019272470846772194, + 0.020207276567816734, + 0.009793837554752827, + 0.01066291518509388, + 0.012215755879878998, + 0.013185210525989532, + 0.014572721906006336, + 0.012598267756402493, + 0.01196668017655611, + 0.012673084624111652, + 0.010547907091677189, + 0.007370727136731148, + 0.014060840010643005, + 0.013877912424504757, + 0.011389506980776787, + 0.013797336257994175, + 0.00913200806826353, + 0.011446056887507439, + 0.00797396618872881, + 0.005837427452206612, + 0.00919455848634243, + 0.009883464314043522, + 0.008317310363054276, + 0.011350936256349087, + 0.004547954071313143, + 0.007023535668849945, + 0.004971369635313749, + 0.0030588344670832157, + 0.008061191998422146, + 0.00617040554061532, + 0.004674723371863365, + 0.00992992427200079, + 0.004007438663393259, + 0.0058753942139446735, + 0.005763451103121042, + 0.004426657222211361, + 0.005550785921514034, + 0.004433327354490757, + 0.0027084548491984606, + 0.004500857088714838, + 0.003798001678660512, + 0.004938012920320034, + 0.023032093420624733, + 0.07877510040998459, + 0.01367300096899271, + 0.011907530948519707, + 0.04289314150810242, + 0.011025246232748032, + 0.008861353620886803, + 0.025682907551527023, + 0.006533915642648935, + 0.012993891723453999, + 0.007565448526293039, + 0.008870062418282032, + 0.01872229017317295, + 0.005499640014022589, + 0.004949595779180527, + 0.014819329604506493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.34864091873168945 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, long shot, shot from a slightly upward angle, first focuses on a weathered beacon tower on the ancient Great Wall. Four soldiers clad in dark iron plate armor are at the height of their triumphant ecstasy. Their armor, composed of thousands of plates, is etched with scratches and dents from battle, its leather straps rustling in the strong wind. The bright red tassels of their iron helmets dance wildly in the biting mountain breeze. Their weathered faces, stained with sweat and dust, are now filled with undisguised, heartfelt joy, their mouths open in a silent roar. Steadily and powerfully, they raise their spears and ring-handled swords high into the sky, the sharp metal tips gleaming blindingly in the golden afternoon sun that pierces the clouds. As the camera steadily and deliberately pulls back, the image gradually reveals the beacon towers built of massive gray stones beneath their feet, and the entire Great Wall, winding and undulating like a giant dragon, disappearing into the vast mountains in the distance. Thick clouds swirl across the sky, and beams of light pouring through the torn rifts cast vast, shifting shadows across the mountain ranges. The entire scene, a stark palette of gray, brown, and iron black, is infused with vitality by the touch of sunlight and the red of the helmet tassels, creating a sense of epic grandeur, primal power, and the powerful emotion of a hard-won victory in a harsh environment.", + "seed": 33472, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_5/index_8218_seed_33472_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14899626076221467, + "vlm_logprob": -0.14899626076221467, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.14899626076221467, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26459982991218567 + ], + "opposite_logprob_tokens": [ + -1.5145998001098633 + ], + "logprob_reward": -0.26459982991218567, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.26459982991218567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14698755741119385, + 0.2452167570590973, + 0.16242606937885284, + 0.0785413384437561, + 0.0584128312766552, + 0.08144848793745041, + 0.08313722163438797, + 0.06200998276472092, + 0.021102475002408028, + 0.060717321932315826 + ], + "credit_map": [ + 0.015048351138830185, + 0.017262164503335953, + 0.01949789933860302, + 0.021413272246718407, + 0.011733833700418472, + 0.013259273022413254, + 0.024505697190761566, + 0.024267062544822693, + 0.024128518998622894, + 0.01677505485713482, + 0.035218458622694016, + 0.02961716055870056, + 0.023711472749710083, + 0.02295607142150402, + 0.03761577606201172, + 0.055194247514009476, + 0.01285619754344225, + 0.019614817574620247, + 0.01671672984957695, + 0.01991204172372818, + 0.013608968816697598, + 0.025008060038089752, + 0.03493931517004967, + 0.019769933074712753, + 0.007526060566306114, + 0.0112193888053298, + 0.009590153582394123, + 0.009921664372086525, + 0.010462376289069653, + 0.007950793951749802, + 0.011553741991519928, + 0.010317156091332436, + 0.007811737712472677, + 0.008428318426012993, + 0.008305697701871395, + 0.006629326846450567, + 0.007827655412256718, + 0.0055892933160066605, + 0.007137634325772524, + 0.0066831656731665134, + 0.012398738414049149, + 0.02116725966334343, + 0.012938722036778927, + 0.005261867772787809, + 0.007930871099233627, + 0.007161829154938459, + 0.006636160425841808, + 0.007953047752380371, + 0.011960246600210667, + 0.010840126313269138, + 0.016738414764404297, + 0.004026282578706741, + 0.010908856056630611, + 0.012056421488523483, + 0.006195622030645609, + 0.01041125226765871, + 0.01029428094625473, + 0.00812759529799223, + 0.009087979793548584, + 0.003634444670751691, + 0.005212063435465097, + 0.012035000137984753, + 0.005161659326404333, + 0.008456961251795292, + 0.002687720349058509, + 0.002841775305569172, + 0.0037628833670169115, + 0.0018406379967927933, + 0.001632891595363617, + 0.0037061909679323435, + 0.0017906263237819076, + 0.0028397466521710157, + 0.008671486750245094, + 0.006025040987879038, + 0.0067215790040791035, + 0.0057708825916051865, + 0.007325138431042433, + 0.010929436422884464, + 0.007165643852204084, + 0.008108112029731274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.26459982991218567 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706905603408813 + ], + "opposite_logprob_tokens": [ + -2.1270689964294434 + ], + "logprob_reward": -0.12706905603408813, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706905603408813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10290277749300003, + 0.1880253702402115, + 0.20810309052467346, + 0.052488598972558975, + 0.033272042870521545, + 0.07011523842811584, + 0.08069956302642822, + 0.03557637706398964, + 0.1267538219690323, + 0.1020631492137909 + ], + "credit_map": [ + 0.011034679599106312, + 0.00660628080368042, + 0.016510916873812675, + 0.01129844319075346, + 0.008531367406249046, + 0.009764760732650757, + 0.014062725938856602, + 0.025093592703342438, + 0.018997205421328545, + 0.011014046147465706, + 0.027067190036177635, + 0.023111911490559578, + 0.018768534064292908, + 0.016850873827934265, + 0.027625467628240585, + 0.04459012299776077, + 0.012540476396679878, + 0.01122343074530363, + 0.01813574507832527, + 0.017869379371404648, + 0.015972938388586044, + 0.05291348323225975, + 0.05880851671099663, + 0.020639104768633842, + 0.004431379958987236, + 0.0064203510992228985, + 0.0055152419954538345, + 0.007436459884047508, + 0.006610303185880184, + 0.004829994402825832, + 0.008917584083974361, + 0.008327278308570385, + 0.004212573636323214, + 0.0052666086703538895, + 0.004563651513308287, + 0.00408880366012454, + 0.0030483382288366556, + 0.003445325419306755, + 0.004586691036820412, + 0.004060046281665564, + 0.011775103397667408, + 0.017690695822238922, + 0.009178223088383675, + 0.0053444793447852135, + 0.006935413461178541, + 0.006371055264025927, + 0.005119113251566887, + 0.007701152935624123, + 0.011419673450291157, + 0.009227419272065163, + 0.013820664957165718, + 0.004191004671156406, + 0.009813574142754078, + 0.013989199884235859, + 0.007738426327705383, + 0.010499597527086735, + 0.0058099995367228985, + 0.0034098594915121794, + 0.007083036936819553, + 0.0029510522726923227, + 0.003449863288551569, + 0.004595920443534851, + 0.003953852225095034, + 0.004322787281125784, + 0.01563967764377594, + 0.014871769584715366, + 0.016634440049529076, + 0.010789685882627964, + 0.015346990898251534, + 0.021284429356455803, + 0.013084354810416698, + 0.019102461636066437, + 0.01588168367743492, + 0.009089166298508644, + 0.008894997648894787, + 0.008528967387974262, + 0.013010026887059212, + 0.018750004470348358, + 0.013834848068654537, + 0.014073450118303299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.12706905603408813 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867850989103317 + ], + "opposite_logprob_tokens": [ + -3.048678398132324 + ], + "logprob_reward": -0.04867850989103317, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867850989103317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1690797209739685, + 0.0859263688325882, + 0.04775677248835564, + 0.07551976293325424, + 0.05562986060976982, + 0.16863267123699188, + 0.10991790890693665, + 0.07508521527051926, + 0.07889571040868759, + 0.13355599343776703 + ], + "credit_map": [ + 0.020740775391459465, + 0.014712752774357796, + 0.021115025505423546, + 0.020700328052043915, + 0.021214544773101807, + 0.016305027529597282, + 0.021455205976963043, + 0.03283606842160225, + 0.006295326631516218, + 0.006595608778297901, + 0.011946634389460087, + 0.01224484108388424, + 0.00734814302995801, + 0.007569518871605396, + 0.012046819552779198, + 0.021879475563764572, + 0.0035442637745290995, + 0.004027905408293009, + 0.004975028336048126, + 0.00705412682145834, + 0.006545100826770067, + 0.007444566115736961, + 0.009375913999974728, + 0.004789869301021099, + 0.00692642480134964, + 0.011594711802899837, + 0.0051948861218988895, + 0.009297285228967667, + 0.00930401124060154, + 0.00742326071485877, + 0.011111649684607983, + 0.01466753613203764, + 0.008777600713074207, + 0.011084974743425846, + 0.008507262915372849, + 0.00536130229011178, + 0.005234796088188887, + 0.005303862038999796, + 0.004938553087413311, + 0.0064215040765702724, + 0.022348854690790176, + 0.034017164260149, + 0.021713480353355408, + 0.011194054037332535, + 0.02260447107255459, + 0.02827497571706772, + 0.012351008132100105, + 0.016128651797771454, + 0.015447025187313557, + 0.01597941853106022, + 0.016478823497891426, + 0.005135409999638796, + 0.013755224645137787, + 0.01774495840072632, + 0.01042787916958332, + 0.014949165284633636, + 0.013130296021699905, + 0.010633126832544804, + 0.008059829473495483, + 0.0057502673007547855, + 0.006787167862057686, + 0.015292047522962093, + 0.0062070623971521854, + 0.009225419722497463, + 0.009511817246675491, + 0.009595963172614574, + 0.009991185739636421, + 0.00718321418389678, + 0.01021731086075306, + 0.009811591356992722, + 0.010193683207035065, + 0.012390943244099617, + 0.029452795162796974, + 0.01625349558889866, + 0.012288645841181278, + 0.009445915929973125, + 0.011492406949400902, + 0.02558821439743042, + 0.011555744335055351, + 0.01747877150774002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.04867850989103317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294828832149506 + ], + "opposite_logprob_tokens": [ + -3.7732949256896973 + ], + "logprob_reward": -0.023294828832149506, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294828832149506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21879427134990692, + 0.0764903575181961, + 0.1686229705810547, + 0.08416619151830673, + 0.20008103549480438, + 0.041287802159786224, + 0.06946668773889542, + 0.04633162543177605, + 0.028443988412618637, + 0.06631509214639664 + ], + "credit_map": [ + 0.01688714139163494, + 0.02254425920546055, + 0.024567920714616776, + 0.02716864086687565, + 0.03881557658314705, + 0.028170593082904816, + 0.03686922788619995, + 0.023770906031131744, + 0.00693650683388114, + 0.004808369092643261, + 0.011593934148550034, + 0.007610028143972158, + 0.006399174220860004, + 0.0063693043775856495, + 0.01260191947221756, + 0.02017112448811531, + 0.010404874570667744, + 0.013410796411335468, + 0.014533176086843014, + 0.020110128447413445, + 0.018224351108074188, + 0.037475936114788055, + 0.03909952938556671, + 0.01536418404430151, + 0.00918885599821806, + 0.015319129452109337, + 0.006637812126427889, + 0.009868036955595016, + 0.00839102454483509, + 0.007536367047578096, + 0.014600004069507122, + 0.012624964118003845, + 0.0295561533421278, + 0.024432089179754257, + 0.026395434513688087, + 0.024969832971692085, + 0.018460161983966827, + 0.02584254927933216, + 0.02395366132259369, + 0.026471154764294624, + 0.004418292082846165, + 0.006134097930043936, + 0.007074812427163124, + 0.002487191464751959, + 0.005511004012078047, + 0.0038788141682744026, + 0.006761087570339441, + 0.005022503901273012, + 0.0090818852186203, + 0.010478329844772816, + 0.011453124694526196, + 0.003786670044064522, + 0.009740469045937061, + 0.007614523638039827, + 0.009933055378496647, + 0.007378630340099335, + 0.007438506931066513, + 0.007544404827058315, + 0.004675308242440224, + 0.003051226260140538, + 0.0065917386673390865, + 0.008681242354214191, + 0.003681620117276907, + 0.004667576402425766, + 0.003385094925761223, + 0.004059068392962217, + 0.0038399468176066875, + 0.0028846561908721924, + 0.0035800582263618708, + 0.004275431390851736, + 0.0027990874368697405, + 0.003620644798502326, + 0.00838431902229786, + 0.008636001497507095, + 0.006567865610122681, + 0.005941238719969988, + 0.008432207629084587, + 0.011337864212691784, + 0.008779183030128479, + 0.008236411027610302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.023294828832149506 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813390791416168 + ], + "opposite_logprob_tokens": [ + -1.4063390493392944 + ], + "logprob_reward": -0.2813390791416168, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813390791416168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12998849153518677, + 0.0927153155207634, + 0.0886945128440857, + 0.07043153047561646, + 0.044994279742240906, + 0.1349804401397705, + 0.2173401117324829, + 0.054397597908973694, + 0.05279181897640228, + 0.11366588622331619 + ], + "credit_map": [ + 0.02358691580593586, + 0.014043227769434452, + 0.015635523945093155, + 0.01379501260817051, + 0.01256544329226017, + 0.013508052565157413, + 0.01778237894177437, + 0.01907194033265114, + 0.007896753028035164, + 0.007104908581823111, + 0.014953525736927986, + 0.009821707382798195, + 0.008297647349536419, + 0.007969502359628677, + 0.014251524582505226, + 0.022419746965169907, + 0.007670590654015541, + 0.006846138276159763, + 0.010365194641053677, + 0.01094291266053915, + 0.010016856715083122, + 0.017604388296604156, + 0.017223380506038666, + 0.008025044575333595, + 0.006195527967065573, + 0.006764030549675226, + 0.009266646578907967, + 0.011040124110877514, + 0.005653706844896078, + 0.005473643075674772, + 0.01508599054068327, + 0.010951860807836056, + 0.005770078860223293, + 0.004590020049363375, + 0.00827850867062807, + 0.005244559142738581, + 0.004171273205429316, + 0.004295992199331522, + 0.0060286796651780605, + 0.006615168880671263, + 0.018350377678871155, + 0.02729671262204647, + 0.017396636307239532, + 0.007656018249690533, + 0.012655488215386868, + 0.0172470360994339, + 0.018207238987088203, + 0.016170933842658997, + 0.026834610849618912, + 0.025301070883870125, + 0.04505906626582146, + 0.011300239711999893, + 0.028248818591237068, + 0.030146675184369087, + 0.023865582421422005, + 0.026584049686789513, + 0.011224191635847092, + 0.006952185183763504, + 0.0069158500991761684, + 0.00398962153121829, + 0.004563984461128712, + 0.008836761116981506, + 0.005466867238283157, + 0.006448138039559126, + 0.0064539010636508465, + 0.0067198327742516994, + 0.006657908204942942, + 0.003679362591356039, + 0.005337740294635296, + 0.010714189149439335, + 0.0050360397435724735, + 0.008192845620214939, + 0.018039273098111153, + 0.011024907231330872, + 0.010128961876034737, + 0.011427327059209347, + 0.010100157000124454, + 0.02446570061147213, + 0.01353056263178587, + 0.014948990195989609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.2813390791416168 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, long shot, shot from a slightly upward angle, first focuses on a weathered beacon tower on the ancient Great Wall. Four soldiers clad in dark iron plate armor are at the height of their triumphant ecstasy. Their armor, composed of thousands of plates, is etched with scratches and dents from battle, its leather straps rustling in the strong wind. The bright red tassels of their iron helmets dance wildly in the biting mountain breeze. Their weathered faces, stained with sweat and dust, are now filled with undisguised, heartfelt joy, their mouths open in a silent roar. Steadily and powerfully, they raise their spears and ring-handled swords high into the sky, the sharp metal tips gleaming blindingly in the golden afternoon sun that pierces the clouds. As the camera steadily and deliberately pulls back, the image gradually reveals the beacon towers built of massive gray stones beneath their feet, and the entire Great Wall, winding and undulating like a giant dragon, disappearing into the vast mountains in the distance. Thick clouds swirl across the sky, and beams of light pouring through the torn rifts cast vast, shifting shadows across the mountain ranges. The entire scene, a stark palette of gray, brown, and iron black, is infused with vitality by the touch of sunlight and the red of the helmet tassels, creating a sense of epic grandeur, primal power, and the powerful emotion of a hard-won victory in a harsh environment.", + "seed": 33473, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_5/index_8218_seed_33473_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05234237872064114, + "vlm_logprob": -0.05234237872064114, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.05234237872064114, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043286148458719254 + ], + "opposite_logprob_tokens": [ + -3.293286085128784 + ], + "logprob_reward": -0.043286148458719254, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.043286148458719254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10448283702135086, + 0.2916708290576935, + 0.024684324860572815, + 0.04140503332018852, + 0.0588885135948658, + 0.06663800030946732, + 0.12137803435325623, + 0.05789594352245331, + 0.06804005056619644, + 0.16491644084453583 + ], + "credit_map": [ + 0.0072135538794100285, + 0.014433084055781364, + 0.013747524470090866, + 0.015986064448952675, + 0.012247782200574875, + 0.013042797334492207, + 0.014425359666347504, + 0.013386677019298077, + 0.0121537484228611, + 0.009542570449411869, + 0.0420602448284626, + 0.010605748742818832, + 0.011436455883085728, + 0.06940906494855881, + 0.10526779294013977, + 0.031195199117064476, + 0.003940691705793142, + 0.001960612600669265, + 0.002845140639692545, + 0.0026197179686278105, + 0.003269117558375001, + 0.0035256827250123024, + 0.003681925358250737, + 0.002841437701135874, + 0.004124232567846775, + 0.00236520916223526, + 0.005511135794222355, + 0.0029754596762359142, + 0.007027949206531048, + 0.005415671970695257, + 0.006937817670404911, + 0.007047557272017002, + 0.009145794436335564, + 0.0037749113980680704, + 0.006081453524529934, + 0.005759637337177992, + 0.008577429689466953, + 0.008885716088116169, + 0.0049140434712171555, + 0.011749526485800743, + 0.009748438373208046, + 0.007769490592181683, + 0.0074925608932971954, + 0.007648312021046877, + 0.007815128192305565, + 0.009532646276056767, + 0.008997409604489803, + 0.007634010631591082, + 0.03652261570096016, + 0.010422341525554657, + 0.007447205483913422, + 0.01332872174680233, + 0.00779198482632637, + 0.03227785974740982, + 0.007714624051004648, + 0.005872684996575117, + 0.01136634312570095, + 0.005258761812001467, + 0.006735717877745628, + 0.004878608975559473, + 0.007569881621748209, + 0.014876596629619598, + 0.003622600808739662, + 0.0035874343011528254, + 0.01349848322570324, + 0.004583144094794989, + 0.0058737704530358315, + 0.0037646328564733267, + 0.011373926885426044, + 0.021148066967725754, + 0.0036415879148989916, + 0.004156433045864105, + 0.045653946697711945, + 0.014950629323720932, + 0.007810624316334724, + 0.005744523368775845, + 0.013478538952767849, + 0.06439100950956345, + 0.006251432467252016, + 0.006635741796344519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.043286148458719254 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904938399791718 + ], + "opposite_logprob_tokens": [ + -2.4640493392944336 + ], + "logprob_reward": -0.08904938399791718, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904938399791718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0750594437122345, + 0.36825430393218994, + 0.020016612485051155, + 0.03496231511235237, + 0.07588310539722443, + 0.06298366189002991, + 0.16513368487358093, + 0.03716176375746727, + 0.027290012687444687, + 0.13325507938861847 + ], + "credit_map": [ + 0.006014151033014059, + 0.012688940390944481, + 0.008104325272142887, + 0.007933652959764004, + 0.00889200996607542, + 0.010874918662011623, + 0.010372556746006012, + 0.010178885422647, + 0.02274516597390175, + 0.010590865276753902, + 0.05671083554625511, + 0.01427238155156374, + 0.014363163150846958, + 0.08235909044742584, + 0.12849298119544983, + 0.03871981054544449, + 0.0020040590316057205, + 0.002267054282128811, + 0.0030476965475827456, + 0.002006886526942253, + 0.0024448935873806477, + 0.0025803539901971817, + 0.003023704746738076, + 0.0026419630739837885, + 0.003017802955582738, + 0.002702548401430249, + 0.00503262085840106, + 0.0024474177043884993, + 0.006871543824672699, + 0.004522053990513086, + 0.004502398893237114, + 0.005865928251296282, + 0.010094513185322285, + 0.005884520709514618, + 0.008595920167863369, + 0.00746119162067771, + 0.011933946050703526, + 0.009188718162477016, + 0.009753498248755932, + 0.01297079212963581, + 0.009777434170246124, + 0.007090001832693815, + 0.007443099282681942, + 0.007863292470574379, + 0.006313185207545757, + 0.00848938338458538, + 0.008571555837988853, + 0.007435712032020092, + 0.05695093050599098, + 0.012767093256115913, + 0.006099608261138201, + 0.00897534191608429, + 0.00967421755194664, + 0.057179972529411316, + 0.007277967873960733, + 0.0062085664831101894, + 0.005070789251476526, + 0.00468313368037343, + 0.003490985371172428, + 0.004217948764562607, + 0.006670975126326084, + 0.004379086662083864, + 0.004398950841277838, + 0.004249893594533205, + 0.004472355358302593, + 0.0018900737632066011, + 0.0026419260539114475, + 0.0015114540001377463, + 0.0044686258770525455, + 0.007382587064057589, + 0.0019369876245036721, + 0.002986003179103136, + 0.04273677617311478, + 0.011458728462457657, + 0.00468722078949213, + 0.00315516022965312, + 0.007470122538506985, + 0.05470220744609833, + 0.003949722275137901, + 0.0050951456651091576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.08904938399791718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020597022026777267 + ], + "opposite_logprob_tokens": [ + -3.895596981048584 + ], + "logprob_reward": -0.020597022026777267, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597022026777267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11385247856378555, + 0.1277429759502411, + 0.02883068099617958, + 0.06528519093990326, + 0.14018835127353668, + 0.0730968490242958, + 0.1071711927652359, + 0.03523765876889229, + 0.048473287373781204, + 0.26012134552001953 + ], + "credit_map": [ + 0.007213856093585491, + 0.01447453536093235, + 0.014729488641023636, + 0.020926211029291153, + 0.01462615467607975, + 0.008350551128387451, + 0.021411558613181114, + 0.012120124883949757, + 0.006838307250291109, + 0.00403082650154829, + 0.02603517845273018, + 0.0047189220786094666, + 0.007247402332723141, + 0.012499489821493626, + 0.05018577724695206, + 0.01618707925081253, + 0.0031754181254655123, + 0.0018532999092712998, + 0.0036731937434524298, + 0.002423505298793316, + 0.00377485784702003, + 0.0034855566918849945, + 0.004019305109977722, + 0.006425545550882816, + 0.005559387616813183, + 0.004539798013865948, + 0.00858917273581028, + 0.004655007738620043, + 0.012806915678083897, + 0.008879239670932293, + 0.009861416183412075, + 0.010394249111413956, + 0.0225839801132679, + 0.01151921134442091, + 0.021492695435881615, + 0.014568882994353771, + 0.015134818851947784, + 0.024772366508841515, + 0.013843569904565811, + 0.016272831708192825, + 0.01314728707075119, + 0.00862372387200594, + 0.008180927485227585, + 0.008042575791478157, + 0.0067905946634709835, + 0.01104261726140976, + 0.00876451563090086, + 0.00850460585206747, + 0.03921494632959366, + 0.007895419374108315, + 0.0028829367365688086, + 0.0032596474047750235, + 0.005084894597530365, + 0.042182374745607376, + 0.003134678117930889, + 0.0035162889398634434, + 0.004694650415331125, + 0.004639832768589258, + 0.0038545560091733932, + 0.004031026735901833, + 0.005957134999334812, + 0.0048307511024177074, + 0.003988022916018963, + 0.0032416845206171274, + 0.009799163788557053, + 0.0034342268481850624, + 0.0049692788161337376, + 0.0029780680779367685, + 0.007588494569063187, + 0.013934122398495674, + 0.002545357448980212, + 0.00322457542642951, + 0.08482836931943893, + 0.024293111637234688, + 0.006847166456282139, + 0.004542703274637461, + 0.015125571750104427, + 0.10781966894865036, + 0.007428696379065514, + 0.009236069396138191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.020597022026777267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808735474944115 + ], + "opposite_logprob_tokens": [ + -3.529808759689331 + ], + "logprob_reward": -0.029808735474944115, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808735474944115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16774101555347443, + 0.10817641764879227, + 0.04759906604886055, + 0.2809894382953644, + 0.050246670842170715, + 0.08398882299661636, + 0.033387213945388794, + 0.046951912343502045, + 0.03310435637831688, + 0.14781512320041656 + ], + "credit_map": [ + 0.013783760368824005, + 0.019702088087797165, + 0.022903544828295708, + 0.030176149681210518, + 0.02036406844854355, + 0.012603994458913803, + 0.028971197083592415, + 0.019236214458942413, + 0.021485472097992897, + 0.009232887998223305, + 0.011894547380506992, + 0.011961810290813446, + 0.010981792584061623, + 0.02679968811571598, + 0.00933589693158865, + 0.006484312936663628, + 0.004049389623105526, + 0.0033392361365258694, + 0.005826946347951889, + 0.0036222150083631277, + 0.007092206738889217, + 0.007210932206362486, + 0.006486646365374327, + 0.0099714957177639, + 0.022055113688111305, + 0.01856653206050396, + 0.03192649781703949, + 0.016683034598827362, + 0.06781361252069473, + 0.040608037263154984, + 0.04989941418170929, + 0.03343719244003296, + 0.009609156288206577, + 0.004169002175331116, + 0.0054582576267421246, + 0.005200364161282778, + 0.006935845594853163, + 0.005759914871305227, + 0.005203375592827797, + 0.007910754531621933, + 0.01225602999329567, + 0.008775806054472923, + 0.009433815255761147, + 0.010208385065197945, + 0.009129021316766739, + 0.0106612928211689, + 0.012650970369577408, + 0.010873506776988506, + 0.0059554120525717735, + 0.004477035254240036, + 0.0031998255290091038, + 0.004994136281311512, + 0.003986896947026253, + 0.0031115980818867683, + 0.004492339212447405, + 0.0031699698884040117, + 0.011366503313183784, + 0.0041127861477434635, + 0.0027142935432493687, + 0.002394392853602767, + 0.006083052139729261, + 0.013992023654282093, + 0.0028430118691176176, + 0.0034458450973033905, + 0.0036364709958434105, + 0.002122163539752364, + 0.004871687851846218, + 0.004882102832198143, + 0.005320910830050707, + 0.0031850673258304596, + 0.0036818459630012512, + 0.0054041058756411076, + 0.04613647609949112, + 0.014000150375068188, + 0.004854162689298391, + 0.004009542986750603, + 0.008663259446620941, + 0.0597035177052021, + 0.004995673429220915, + 0.005452333949506283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.029808735474944115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897060364484787 + ], + "opposite_logprob_tokens": [ + -2.5789706707000732 + ], + "logprob_reward": -0.07897060364484787, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897060364484787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09224415570497513, + 0.2189519703388214, + 0.014757318422198296, + 0.019185204058885574, + 0.07355046272277832, + 0.04758313298225403, + 0.1677919328212738, + 0.023682674393057823, + 0.08309762179851532, + 0.2591554820537567 + ], + "credit_map": [ + 0.00669182650744915, + 0.014934766106307507, + 0.013575880788266659, + 0.011373275890946388, + 0.009735871106386185, + 0.013762004673480988, + 0.014412330463528633, + 0.007758201099932194, + 0.008091887459158897, + 0.006565325893461704, + 0.0352645069360733, + 0.007977643981575966, + 0.010345459915697575, + 0.042949847877025604, + 0.08261125534772873, + 0.025146065279841423, + 0.0016711895586922765, + 0.0016187402652576566, + 0.0016406774520874023, + 0.0015503522008657455, + 0.001987708732485771, + 0.0025811053346842527, + 0.0020346397068351507, + 0.0016729059861972928, + 0.0018522432073950768, + 0.0016585207777097821, + 0.0024167951196432114, + 0.0016953566810116172, + 0.0033129157964140177, + 0.002519141649827361, + 0.0032437043264508247, + 0.002486528130248189, + 0.009710643440485, + 0.0055258008651435375, + 0.010242295451462269, + 0.009230383671820164, + 0.011984949931502342, + 0.014632610604166985, + 0.005578102543950081, + 0.006645679473876953, + 0.004978954792022705, + 0.0035532540641725063, + 0.006956202443689108, + 0.005502241663634777, + 0.005155766382813454, + 0.004857203457504511, + 0.009606446139514446, + 0.00697306776419282, + 0.06345061957836151, + 0.012846534140408039, + 0.004495894070714712, + 0.004066438414156437, + 0.008147167973220348, + 0.06578481942415237, + 0.0033744005486369133, + 0.005626062396913767, + 0.004463008604943752, + 0.0019374748226255178, + 0.002509571146219969, + 0.0017024939879775047, + 0.003313876921311021, + 0.006641467101871967, + 0.0017152945511043072, + 0.0013994879554957151, + 0.018638622015714645, + 0.0066824001260101795, + 0.007799485232681036, + 0.004974632058292627, + 0.011815470643341541, + 0.024328049272298813, + 0.004268020857125521, + 0.004590949509292841, + 0.07302527129650116, + 0.027682848274707794, + 0.008793877437710762, + 0.007327709347009659, + 0.01829148828983307, + 0.10624779760837555, + 0.007431726437062025, + 0.010354766622185707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.07897060364484787 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, long shot, shot from a slightly upward angle, first focuses on a weathered beacon tower on the ancient Great Wall. Four soldiers clad in dark iron plate armor are at the height of their triumphant ecstasy. Their armor, composed of thousands of plates, is etched with scratches and dents from battle, its leather straps rustling in the strong wind. The bright red tassels of their iron helmets dance wildly in the biting mountain breeze. Their weathered faces, stained with sweat and dust, are now filled with undisguised, heartfelt joy, their mouths open in a silent roar. Steadily and powerfully, they raise their spears and ring-handled swords high into the sky, the sharp metal tips gleaming blindingly in the golden afternoon sun that pierces the clouds. As the camera steadily and deliberately pulls back, the image gradually reveals the beacon towers built of massive gray stones beneath their feet, and the entire Great Wall, winding and undulating like a giant dragon, disappearing into the vast mountains in the distance. Thick clouds swirl across the sky, and beams of light pouring through the torn rifts cast vast, shifting shadows across the mountain ranges. The entire scene, a stark palette of gray, brown, and iron black, is infused with vitality by the touch of sunlight and the red of the helmet tassels, creating a sense of epic grandeur, primal power, and the powerful emotion of a hard-won victory in a harsh environment.", + "seed": 33474, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_5/index_8218_seed_33474_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04466493222862482, + "vlm_logprob": -0.04466493222862482, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.04466493222862482, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06394071131944656 + ], + "opposite_logprob_tokens": [ + -2.8139407634735107 + ], + "logprob_reward": -0.06394071131944656, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06394071131944656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09948863089084625, + 0.051718637347221375, + 0.04054313153028488, + 0.05526207759976387, + 0.030845899134874344, + 0.04663289338350296, + 0.056778792291879654, + 0.12255395203828812, + 0.3624829351902008, + 0.13369303941726685 + ], + "credit_map": [ + 0.01157188881188631, + 0.010307303629815578, + 0.013921890407800674, + 0.0124899260699749, + 0.006966412533074617, + 0.012461700476706028, + 0.013744303956627846, + 0.018025202676653862, + 0.006047633942216635, + 0.0036009917967021465, + 0.011486141011118889, + 0.004775505978614092, + 0.005286321043968201, + 0.006279774475842714, + 0.008214321918785572, + 0.006027949508279562, + 0.005436690989881754, + 0.004574075806885958, + 0.00761206541210413, + 0.004042352084070444, + 0.004575933795422316, + 0.004693107679486275, + 0.004101917147636414, + 0.005506991408765316, + 0.006333111319690943, + 0.004968265537172556, + 0.009634836576879025, + 0.00457413587719202, + 0.006214182358235121, + 0.006083986256271601, + 0.010419473983347416, + 0.007034084293991327, + 0.004309188574552536, + 0.001621048548258841, + 0.009742670692503452, + 0.004801914561539888, + 0.0020421037916094065, + 0.004667128901928663, + 0.001490843016654253, + 0.0021710011642426252, + 0.008227940648794174, + 0.00521585438400507, + 0.003992144018411636, + 0.004359932150691748, + 0.005388719495385885, + 0.00749016460031271, + 0.006288310047239065, + 0.005669828038662672, + 0.00923381932079792, + 0.007112485822290182, + 0.007155272178351879, + 0.008560985326766968, + 0.00659368559718132, + 0.006529552396386862, + 0.0062810154631733894, + 0.005311974789947271, + 0.021148569881916046, + 0.019200975075364113, + 0.013181084766983986, + 0.00746429106220603, + 0.027475018054246902, + 0.017419936135411263, + 0.00821095984429121, + 0.008453120477497578, + 0.024782124906778336, + 0.1269715577363968, + 0.017397427931427956, + 0.011011716909706593, + 0.056356675922870636, + 0.09764242172241211, + 0.013839393854141235, + 0.01448164414614439, + 0.014673351310193539, + 0.020362209528684616, + 0.01575649157166481, + 0.011293389834463596, + 0.02789868414402008, + 0.02022731304168701, + 0.010992336086928844, + 0.012489265762269497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.06394071131944656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294256895780563 + ], + "opposite_logprob_tokens": [ + -2.2379424571990967 + ], + "logprob_reward": -0.11294256895780563, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294256895780563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10161095857620239, + 0.04475141316652298, + 0.07675986737012863, + 0.03813495114445686, + 0.09786448627710342, + 0.08181998878717422, + 0.05627590790390968, + 0.14950542151927948, + 0.2776702344417572, + 0.07560683786869049 + ], + "credit_map": [ + 0.014834894798696041, + 0.012506944127380848, + 0.018261468037962914, + 0.014823256991803646, + 0.006033743266016245, + 0.010222879238426685, + 0.010638500563800335, + 0.014289271086454391, + 0.004516342654824257, + 0.0030804933048784733, + 0.010455968789756298, + 0.005222220439463854, + 0.005436014849692583, + 0.004796973895281553, + 0.005712293088436127, + 0.005531106144189835, + 0.009021877311170101, + 0.007979028858244419, + 0.013127836398780346, + 0.007971950806677341, + 0.00711438525468111, + 0.009943430311977863, + 0.008758598938584328, + 0.012842764146625996, + 0.004431469365954399, + 0.003942045383155346, + 0.006518597714602947, + 0.004089840687811375, + 0.004098282195627689, + 0.005167119204998016, + 0.005841277074068785, + 0.00404632231220603, + 0.014729867689311504, + 0.006977969780564308, + 0.03126073628664017, + 0.022853048518300056, + 0.004104017745703459, + 0.008299569599330425, + 0.0037143791560083628, + 0.00592489680275321, + 0.009368452243506908, + 0.008540370501577854, + 0.006784444209188223, + 0.011037256568670273, + 0.00855433102697134, + 0.0149764996021986, + 0.01389696728438139, + 0.008661665953695774, + 0.011566626839339733, + 0.005425062961876392, + 0.00728760939091444, + 0.0075600226409733295, + 0.006391672417521477, + 0.0060824439860880375, + 0.006782485172152519, + 0.005179983098059893, + 0.020064953714609146, + 0.015199168585240841, + 0.02567046321928501, + 0.011804942041635513, + 0.023338062688708305, + 0.01825706660747528, + 0.018840664997696877, + 0.01633010245859623, + 0.018411245197057724, + 0.09511372447013855, + 0.014549466781318188, + 0.00924958847463131, + 0.03984619304537773, + 0.08074506372213364, + 0.007549924310296774, + 0.012205013073980808, + 0.010336095467209816, + 0.011758165434002876, + 0.010665926150977612, + 0.00883775856345892, + 0.008463454432785511, + 0.00982232391834259, + 0.008920365944504738, + 0.00680273724719882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.11294256895780563 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596904680132866 + ], + "opposite_logprob_tokens": [ + -3.895596981048584 + ], + "logprob_reward": -0.020596904680132866, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596904680132866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08933887630701065, + 0.03795218840241432, + 0.041304223239421844, + 0.02949579618871212, + 0.035036519169807434, + 0.03055071271955967, + 0.12171369791030884, + 0.1052820235490799, + 0.46881455183029175, + 0.040511369705200195 + ], + "credit_map": [ + 0.014817661605775356, + 0.01045210100710392, + 0.011650016531348228, + 0.013243911787867546, + 0.00717922393232584, + 0.012693485245108604, + 0.009202277287840843, + 0.010100197978317738, + 0.004037685226649046, + 0.00324810529127717, + 0.009679040871560574, + 0.003285648999735713, + 0.003907571546733379, + 0.005297133233398199, + 0.003963898867368698, + 0.0045331064611673355, + 0.004508133511990309, + 0.006024927832186222, + 0.007301832549273968, + 0.00492951599881053, + 0.004221511539071798, + 0.0057497527450323105, + 0.004429026041179895, + 0.004139520227909088, + 0.003734221449121833, + 0.00304793962277472, + 0.006339835003018379, + 0.002813753904774785, + 0.003291763598099351, + 0.0039033975917845964, + 0.003234951524063945, + 0.0031299334950745106, + 0.002976498333737254, + 0.002257685409858823, + 0.006159856915473938, + 0.006867182441055775, + 0.0031952867284417152, + 0.007242418825626373, + 0.0029498832300305367, + 0.003387707518413663, + 0.004200926516205072, + 0.0026194388046860695, + 0.003085883567109704, + 0.002838134067133069, + 0.004244111943989992, + 0.006743935868144035, + 0.0029994123615324497, + 0.003818869823589921, + 0.025687076151371002, + 0.014522678218781948, + 0.014350863173604012, + 0.01786966435611248, + 0.015669705346226692, + 0.01417904905974865, + 0.010593945160508156, + 0.008840718306601048, + 0.014271210879087448, + 0.016444645822048187, + 0.010753700509667397, + 0.008939595893025398, + 0.023204155266284943, + 0.011012377217411995, + 0.010634268634021282, + 0.010022074915468693, + 0.0323089174926281, + 0.1504383236169815, + 0.026316286996006966, + 0.01533596869558096, + 0.06932511925697327, + 0.1401684582233429, + 0.014648607932031155, + 0.020272891968488693, + 0.005393874365836382, + 0.006341918371617794, + 0.004942338913679123, + 0.004543926101177931, + 0.004547857213765383, + 0.0059577953070402145, + 0.004989793989807367, + 0.003793865442276001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.020596904680132866 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066374257206917 + ], + "opposite_logprob_tokens": [ + -4.141066551208496 + ], + "logprob_reward": -0.016066374257206917, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066374257206917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09001374989748001, + 0.019001124426722527, + 0.05816579610109329, + 0.08394331485033035, + 0.15574614703655243, + 0.03946671262383461, + 0.08120439946651459, + 0.08350305259227753, + 0.3629412353038788, + 0.026014501228928566 + ], + "credit_map": [ + 0.01251295953989029, + 0.010397406294941902, + 0.010984235443174839, + 0.014189784415066242, + 0.007180477492511272, + 0.011664362624287605, + 0.010422187857329845, + 0.012662334367632866, + 0.002450364176183939, + 0.0017688559601083398, + 0.004366903565824032, + 0.0015193329891189933, + 0.00183304853271693, + 0.0025460636243224144, + 0.002254064893350005, + 0.0022624912671744823, + 0.006673246156424284, + 0.005026801954954863, + 0.009238449856638908, + 0.006088916212320328, + 0.007950912229716778, + 0.0077800452709198, + 0.008515464141964912, + 0.006891964934766293, + 0.009827998466789722, + 0.0084989657625556, + 0.02114242874085903, + 0.00980721227824688, + 0.007460054010152817, + 0.010141524486243725, + 0.008656417019665241, + 0.008408717811107635, + 0.030024534091353416, + 0.008928172290325165, + 0.053448908030986786, + 0.02754569612443447, + 0.007133224047720432, + 0.014836033806204796, + 0.005978253670036793, + 0.007851325906813145, + 0.003975715953856707, + 0.004990953486412764, + 0.003940715454518795, + 0.005010920576751232, + 0.004370779264718294, + 0.007354209665209055, + 0.00613750983029604, + 0.003685907693579793, + 0.021253760904073715, + 0.011139900423586369, + 0.009640309028327465, + 0.008004617877304554, + 0.008455582894384861, + 0.00898299552500248, + 0.008019079454243183, + 0.0057081542909145355, + 0.01195854414254427, + 0.013566924259066582, + 0.00718995276838541, + 0.006974115967750549, + 0.02040044218301773, + 0.008645388297736645, + 0.007303211838006973, + 0.007464476861059666, + 0.02536233700811863, + 0.11446981132030487, + 0.020945727825164795, + 0.012889409437775612, + 0.05365821346640587, + 0.1062699481844902, + 0.012623623944818974, + 0.01672215573489666, + 0.0029424482490867376, + 0.0029368374962359667, + 0.00417284807190299, + 0.0031906855292618275, + 0.003177357604727149, + 0.00337069109082222, + 0.003382961731404066, + 0.0028406712226569653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.016066374257206917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the ancient Great Wall featuring a weathered beacon tower made of massive gray stones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are four soldiers visible wearing dark iron plate armor composed of thousands of plates with scratches and dents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers raising spears and ring-handled swords high into the sky with sharp metal tips gleaming in the sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' iron helmets have bright red tassels that are dancing wildly in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera steadily pull back to reveal the winding Great Wall disappearing into the vast mountains in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778101928532124 + ], + "opposite_logprob_tokens": [ + -4.634778022766113 + ], + "logprob_reward": -0.009778101928532124, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778101928532124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0648193284869194, + 0.059158485382795334, + 0.05285494774580002, + 0.03149798884987831, + 0.03619905188679695, + 0.07027000188827515, + 0.06776256114244461, + 0.10729125142097473, + 0.3225424587726593, + 0.1876038908958435 + ], + "credit_map": [ + 0.0072109331376850605, + 0.003614633111283183, + 0.013249761424958706, + 0.0085784075781703, + 0.004616766236722469, + 0.008209893479943275, + 0.007502502761781216, + 0.011836430057883263, + 0.006594041828066111, + 0.004791438113898039, + 0.014265035279095173, + 0.005993686616420746, + 0.005483405664563179, + 0.005668289959430695, + 0.00781545415520668, + 0.008547132834792137, + 0.005838062148541212, + 0.0044080778025090694, + 0.008845283649861813, + 0.006840586196631193, + 0.007553865667432547, + 0.007145446725189686, + 0.005440170411020517, + 0.006783452350646257, + 0.0037092475686222315, + 0.0038014964666217566, + 0.004410761874169111, + 0.004896607715636492, + 0.0031894221901893616, + 0.003994655329734087, + 0.003667400451377034, + 0.0038283944595605135, + 0.0030986620113253593, + 0.00305721047334373, + 0.007382805924862623, + 0.004912879317998886, + 0.0058405096642673016, + 0.005590735003352165, + 0.0032122256234288216, + 0.003104023402556777, + 0.008813869208097458, + 0.008042518980801105, + 0.00558538269251585, + 0.007574783638119698, + 0.009241979569196701, + 0.012027273885905743, + 0.011993453837931156, + 0.006990738213062286, + 0.013712886720895767, + 0.007558428682386875, + 0.00832726713269949, + 0.011501248925924301, + 0.0069956728257238865, + 0.007066905032843351, + 0.0073345499113202095, + 0.005265598651021719, + 0.01877174712717533, + 0.019425861537456512, + 0.008371786214411259, + 0.008415075950324535, + 0.019648799672722816, + 0.013126189820468426, + 0.009535500779747963, + 0.009996290318667889, + 0.02817791886627674, + 0.060938965529203415, + 0.025608263909816742, + 0.02438924089074135, + 0.056117963045835495, + 0.07058016955852509, + 0.033157020807266235, + 0.023572901263833046, + 0.01914442889392376, + 0.024579063057899475, + 0.020048052072525024, + 0.017041588202118874, + 0.038897860795259476, + 0.02953292615711689, + 0.021843185648322105, + 0.016516797244548798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.009778101928532124 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_6/index_tensor([9033, 9033])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_6/index_tensor([9033, 9033])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cbc2756b10cb7c44b9fdeed9ed0e6efcd39bd729 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_6/index_tensor([9033, 9033])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a black-and-white world of thick film grain and high-contrast lighting, a rapidly rotating, circling camera closely follows a black cat as it performs an astonishing feat on the parallel bars. The cat's body is pitch black, its fur smooth as satin, its silver highlights reflecting off the single, harsh spotlight cast from above. The definition of its muscles is clearly visible with every thrust. Its eyes, narrowed slightly, gleam with a cunning and cold light, as if enjoying this dark performance all its own. With its front paws firmly grasping the cold metal bars, it smoothly executes a series of complex swings and twists with a gesture that blends the grace of a ballerina with the strength of a top athlete. Its body carves perfect arcs through the air, its long tail acting as a precise rudder, maintaining its incredible balance. The camera's rapid rotation blurs the background, focusing entirely on the cat's sinisterly beautiful movements. Deep shadows dance wildly on the ground as it moves, creating a surreal atmosphere that is both dizzying and slightly unsettling.", + "seed": 29279, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_6/index_9033_seed_29279_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017858849745243786, + "vlm_logprob": -0.017858849745243786, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.017858849745243786, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007129466626793146 + ], + "opposite_logprob_tokens": [ + -6.007129669189453 + ], + "logprob_reward": -0.007129466626793146, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.007129466626793146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14765481650829315, + 0.13877341151237488, + 0.055529624223709106, + 0.0318506583571434, + 0.0490141287446022, + 0.10789318382740021, + 0.21512039005756378, + 0.11245761066675186, + 0.03638697415590286, + 0.10531923174858093 + ], + "credit_map": [ + 0.014279918745160103, + 0.015705473721027374, + 0.02315925806760788, + 0.019969258457422256, + 0.017427481710910797, + 0.021078525111079216, + 0.020657306537032127, + 0.015377595089375973, + 0.01831277646124363, + 0.02889721840620041, + 0.02228040248155594, + 0.007137003354728222, + 0.025362670421600342, + 0.01757386513054371, + 0.008110548369586468, + 0.011098933406174183, + 0.0076935505494475365, + 0.010787686333060265, + 0.00493074394762516, + 0.005214016418904066, + 0.007705471478402615, + 0.006847378797829151, + 0.005906845908612013, + 0.0064439293928444386, + 0.003932449501007795, + 0.003244782565161586, + 0.004321925342082977, + 0.004123480059206486, + 0.0032759588211774826, + 0.0047693713568151, + 0.003315404523164034, + 0.004867284558713436, + 0.006489356979727745, + 0.008362927474081516, + 0.005339916329830885, + 0.00534470658749342, + 0.005331792403012514, + 0.005933685228228569, + 0.004379103425890207, + 0.007832640781998634, + 0.01707991026341915, + 0.0130758136510849, + 0.006605719216167927, + 0.010057920590043068, + 0.016725249588489532, + 0.014495682902634144, + 0.008062887005507946, + 0.02179000712931156, + 0.03505992889404297, + 0.06522328406572342, + 0.01564330793917179, + 0.007298819255083799, + 0.037615783512592316, + 0.018911248072981834, + 0.01632336713373661, + 0.019044650718569756, + 0.012269821017980576, + 0.023105790838599205, + 0.009553455747663975, + 0.01285459753125906, + 0.021240759640932083, + 0.010799931362271309, + 0.007468490395694971, + 0.015164771117269993, + 0.0026745451614260674, + 0.0042009977623820305, + 0.002909665461629629, + 0.0034668019507080317, + 0.004144378937780857, + 0.008125776425004005, + 0.0038902678061276674, + 0.0069745383225381374, + 0.011970669031143188, + 0.013680793344974518, + 0.009595843032002449, + 0.014842981472611427, + 0.016548316925764084, + 0.01801140420138836, + 0.008127938024699688, + 0.01254128385335207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.007129466626793146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647613227367401 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647613227367401, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647613227367401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14287149906158447, + 0.12233790755271912, + 0.14270424842834473, + 0.05189589783549309, + 0.10966009646654129, + 0.0854499340057373, + 0.11480653285980225, + 0.10716630518436432, + 0.06682877987623215, + 0.05627874657511711 + ], + "credit_map": [ + 0.01335163414478302, + 0.01295450795441866, + 0.019889770075678825, + 0.018653573468327522, + 0.017872152850031853, + 0.023372098803520203, + 0.022023264318704605, + 0.014754518866539001, + 0.012787992134690285, + 0.028276890516281128, + 0.01293592993170023, + 0.0038440555799752474, + 0.03295082598924637, + 0.01527401152998209, + 0.005243710242211819, + 0.01102450955659151, + 0.022604865953326225, + 0.018005648627877235, + 0.01141916774213314, + 0.008851885795593262, + 0.017952853813767433, + 0.037965692579746246, + 0.014138956554234028, + 0.011765187606215477, + 0.005147136282175779, + 0.004429755732417107, + 0.005288710352033377, + 0.005613882560282946, + 0.005096084903925657, + 0.010656137950718403, + 0.006049790419638157, + 0.009614404290914536, + 0.012783697806298733, + 0.014357533305883408, + 0.01396979857236147, + 0.010568701662123203, + 0.01675150729715824, + 0.01536789070814848, + 0.010053589008748531, + 0.015807386487722397, + 0.013708511367440224, + 0.013677232898771763, + 0.007050812244415283, + 0.008565014228224754, + 0.011102835647761822, + 0.012871288694441319, + 0.009124630130827427, + 0.009349624626338482, + 0.018787527456879616, + 0.02183425985276699, + 0.01215472910553217, + 0.006528041325509548, + 0.017146019265055656, + 0.012020258232951164, + 0.008930321782827377, + 0.017405392602086067, + 0.009982939809560776, + 0.02011885493993759, + 0.00867459736764431, + 0.01071230135858059, + 0.019713472574949265, + 0.011196832172572613, + 0.007203905377537012, + 0.01956341229379177, + 0.006455274764448404, + 0.010502282530069351, + 0.006313755642622709, + 0.006213358603417873, + 0.008257611654698849, + 0.01373149361461401, + 0.006524847354739904, + 0.008830166421830654, + 0.007519119884818792, + 0.008187717758119106, + 0.0051938267424702644, + 0.00872455071657896, + 0.008879967033863068, + 0.0077524008229374886, + 0.004780948162078857, + 0.005240222439169884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.008647613227367401 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978388637304306 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.00978388637304306, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978388637304306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13383451104164124, + 0.06039011478424072, + 0.05869225785136223, + 0.07421711087226868, + 0.06259820610284805, + 0.08384620398283005, + 0.12016832083463669, + 0.14448854327201843, + 0.03708309680223465, + 0.22468163073062897 + ], + "credit_map": [ + 0.013978750444948673, + 0.012176127173006535, + 0.02074369601905346, + 0.013897771015763283, + 0.01719862036406994, + 0.021280618384480476, + 0.022050641477108, + 0.01250828430056572, + 0.008833915926516056, + 0.008203716017305851, + 0.006295026279985905, + 0.0025917773600667715, + 0.00899742916226387, + 0.014702733606100082, + 0.004328131675720215, + 0.00643738079816103, + 0.008150114677846432, + 0.010779806412756443, + 0.005059374030679464, + 0.005222567822784185, + 0.005386388394981623, + 0.012665980495512486, + 0.0056318906135857105, + 0.005796133540570736, + 0.007441061083227396, + 0.010217320173978806, + 0.007160883396863937, + 0.009120997972786427, + 0.013452211394906044, + 0.010235229507088661, + 0.006228828802704811, + 0.010360583662986755, + 0.0076933810487389565, + 0.00829041562974453, + 0.0072968085296452045, + 0.00680454121902585, + 0.00669997138902545, + 0.011107786558568478, + 0.006007648538798094, + 0.00869765505194664, + 0.010984020307660103, + 0.007881234399974346, + 0.006930558942258358, + 0.008053066208958626, + 0.008350388146936893, + 0.015927208587527275, + 0.013161104172468185, + 0.012558628804981709, + 0.017098689451813698, + 0.023952282965183258, + 0.014352604746818542, + 0.0061127315275371075, + 0.017156731337308884, + 0.011108881793916225, + 0.012275386601686478, + 0.018111007288098335, + 0.015645897015929222, + 0.03076603077352047, + 0.009904969483613968, + 0.013867226429283619, + 0.03141259029507637, + 0.01499610859900713, + 0.00969617534428835, + 0.01819954626262188, + 0.00411227997392416, + 0.005771148484200239, + 0.004658011719584465, + 0.003840029938146472, + 0.004906391724944115, + 0.004063376225531101, + 0.0054053072817623615, + 0.0043265484273433685, + 0.02213747613132, + 0.02842901274561882, + 0.01890072599053383, + 0.03859005868434906, + 0.03832491114735603, + 0.03537638112902641, + 0.01841617561876774, + 0.02450687624514103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.00978388637304306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202206760644913 + ], + "opposite_logprob_tokens": [ + -2.8120219707489014 + ], + "logprob_reward": -0.06202206760644913, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202206760644913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07309725135564804, + 0.05148068815469742, + 0.05383949726819992, + 0.08691944926977158, + 0.09042012691497803, + 0.07272196561098099, + 0.08916482329368591, + 0.2067038118839264, + 0.056848958134651184, + 0.2188034951686859 + ], + "credit_map": [ + 0.008831348270177841, + 0.00839169230312109, + 0.011844725348055363, + 0.01011348981410265, + 0.009547230787575245, + 0.009141335263848305, + 0.006536676548421383, + 0.008690744638442993, + 0.007973579689860344, + 0.006598498206585646, + 0.0065479339100420475, + 0.003684749361127615, + 0.007545785047113895, + 0.008695431984961033, + 0.003971862606704235, + 0.0064628408290445805, + 0.007280416786670685, + 0.011528197675943375, + 0.0071963160298764706, + 0.0036291752476245165, + 0.005749458447098732, + 0.008414186537265778, + 0.005690635181963444, + 0.00435110367834568, + 0.007276699412614107, + 0.007836018688976765, + 0.010256253182888031, + 0.010953124612569809, + 0.014052344486117363, + 0.017122088000178337, + 0.0071159303188323975, + 0.01230697613209486, + 0.010986682027578354, + 0.012864569202065468, + 0.010365886613726616, + 0.007741585839539766, + 0.01401657983660698, + 0.012346294708549976, + 0.007870960980653763, + 0.014227555133402348, + 0.010606899857521057, + 0.006854009348899126, + 0.006392038427293301, + 0.00612581567838788, + 0.008309335447847843, + 0.010578612796962261, + 0.0097929323092103, + 0.014062311500310898, + 0.01790257915854454, + 0.022755105048418045, + 0.006263484712690115, + 0.004136267118155956, + 0.011470209807157516, + 0.007817983627319336, + 0.00877750851213932, + 0.010041673667728901, + 0.018671616911888123, + 0.039968252182006836, + 0.014886366203427315, + 0.018114518374204636, + 0.04175080731511116, + 0.02552011050283909, + 0.016212793067097664, + 0.03157931938767433, + 0.005706725642085075, + 0.006852858699858189, + 0.005901450756937265, + 0.004874482750892639, + 0.005991247482597828, + 0.008223616518080235, + 0.00647258386015892, + 0.012825985439121723, + 0.0249717328697443, + 0.024746417999267578, + 0.018447069451212883, + 0.024502234533429146, + 0.02347813919186592, + 0.04351835697889328, + 0.021364975720643997, + 0.03777454048395157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.06202206760644913 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001711214892566204 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.001711214892566204, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001711214892566204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13977302610874176, + 0.06037365272641182, + 0.09800692647695541, + 0.06891576945781708, + 0.061322275549173355, + 0.16722680628299713, + 0.08797156065702438, + 0.18581490218639374, + 0.04153532162308693, + 0.08905968815088272 + ], + "credit_map": [ + 0.013160050846636295, + 0.014867412857711315, + 0.029310693964362144, + 0.017880886793136597, + 0.022061113268136978, + 0.01716425083577633, + 0.008786598220467567, + 0.01654203236103058, + 0.009207070805132389, + 0.01056353747844696, + 0.009644235484302044, + 0.0037610873114317656, + 0.009102126583456993, + 0.007023724727332592, + 0.004251299891620874, + 0.006820575334131718, + 0.01702697202563286, + 0.028763895854353905, + 0.009944652207195759, + 0.007644543424248695, + 0.008988032117486, + 0.013006017543375492, + 0.006889673415571451, + 0.005743145477026701, + 0.008097033016383648, + 0.006555015221238136, + 0.006093164440244436, + 0.007083253003656864, + 0.008453872986137867, + 0.015620437450706959, + 0.005828026682138443, + 0.011184969916939735, + 0.007120317313820124, + 0.0066835288889706135, + 0.005517131648957729, + 0.006582391448318958, + 0.009488056413829327, + 0.0111765768378973, + 0.005700212437659502, + 0.009054065681993961, + 0.022690756246447563, + 0.020412353798747063, + 0.013824240304529667, + 0.015457632951438427, + 0.02226937748491764, + 0.03025157004594803, + 0.017942124977707863, + 0.024378756061196327, + 0.016059288755059242, + 0.021425150334835052, + 0.005777677986770868, + 0.004968231078237295, + 0.014323444105684757, + 0.0068937987089157104, + 0.007533910218626261, + 0.010990059934556484, + 0.019334586337208748, + 0.04137987270951271, + 0.013174516148865223, + 0.017675623297691345, + 0.041222017258405685, + 0.017418142408132553, + 0.012677248567342758, + 0.02293291874229908, + 0.0036611128598451614, + 0.004418283700942993, + 0.00344463880173862, + 0.004023382905870676, + 0.004239864647388458, + 0.005780908744782209, + 0.004666951950639486, + 0.011300180107355118, + 0.010904155671596527, + 0.010841033421456814, + 0.008964060805737972, + 0.014913142658770084, + 0.013564692810177803, + 0.012824655510485172, + 0.006079587619751692, + 0.010968374088406563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.001711214892566204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a black-and-white world of thick film grain and high-contrast lighting, a rapidly rotating, circling camera closely follows a black cat as it performs an astonishing feat on the parallel bars. The cat's body is pitch black, its fur smooth as satin, its silver highlights reflecting off the single, harsh spotlight cast from above. The definition of its muscles is clearly visible with every thrust. Its eyes, narrowed slightly, gleam with a cunning and cold light, as if enjoying this dark performance all its own. With its front paws firmly grasping the cold metal bars, it smoothly executes a series of complex swings and twists with a gesture that blends the grace of a ballerina with the strength of a top athlete. Its body carves perfect arcs through the air, its long tail acting as a precise rudder, maintaining its incredible balance. The camera's rapid rotation blurs the background, focusing entirely on the cat's sinisterly beautiful movements. Deep shadows dance wildly on the ground as it moves, creating a surreal atmosphere that is both dizzying and slightly unsettling.", + "seed": 29280, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_6/index_9033_seed_29280_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025341124879196285, + "vlm_logprob": -0.025341124879196285, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.025341124879196285, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005812883377075195 + ], + "opposite_logprob_tokens": [ + -6.380812644958496 + ], + "logprob_reward": -0.005812883377075195, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.005812883377075195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2273087501525879, + 0.10277749598026276, + 0.11505868285894394, + 0.0624363012611866, + 0.09748674184083939, + 0.09184470772743225, + 0.06220213323831558, + 0.07736460864543915, + 0.08751863241195679, + 0.07600198686122894 + ], + "credit_map": [ + 0.029128769412636757, + 0.051853545010089874, + 0.02088274247944355, + 0.014087515883147717, + 0.02867930196225643, + 0.02609081007540226, + 0.017124459147453308, + 0.039461616426706314, + 0.010913284495472908, + 0.026726847514510155, + 0.005359012167900801, + 0.005614210851490498, + 0.008741665631532669, + 0.02579631097614765, + 0.010150745511054993, + 0.00947541743516922, + 0.010322902351617813, + 0.013443746604025364, + 0.006499134004116058, + 0.009416202083230019, + 0.03049411065876484, + 0.021368158981204033, + 0.01480138674378395, + 0.008713035844266415, + 0.004884791094809771, + 0.006397940684109926, + 0.007971641607582569, + 0.01068920362740755, + 0.008216473273932934, + 0.01115751639008522, + 0.005798626225441694, + 0.00732010742649436, + 0.015421326272189617, + 0.015962427482008934, + 0.01596246287226677, + 0.010164269246160984, + 0.012740779668092728, + 0.008492921479046345, + 0.008247379213571548, + 0.010495178401470184, + 0.01527587603777647, + 0.010877450928092003, + 0.01616283692419529, + 0.013211299665272236, + 0.010789127089083195, + 0.007261903490871191, + 0.00997213739901781, + 0.008294078521430492, + 0.010066845454275608, + 0.007297122851014137, + 0.005551811307668686, + 0.008650069124996662, + 0.00607670983299613, + 0.008102280087769032, + 0.0076998937875032425, + 0.008757397532463074, + 0.012650354765355587, + 0.0072876145131886005, + 0.008190405555069447, + 0.008520156145095825, + 0.0076002818532288074, + 0.01245369203388691, + 0.012262262403964996, + 0.008399838581681252, + 0.010563485324382782, + 0.012323426082730293, + 0.007188532967120409, + 0.013058989308774471, + 0.013487536460161209, + 0.012902571819722652, + 0.008060784079134464, + 0.009933307766914368, + 0.006986916530877352, + 0.009696781635284424, + 0.004521125927567482, + 0.004623109940439463, + 0.02274630405008793, + 0.011506212875247002, + 0.0065102968364953995, + 0.009411242790520191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.005812883377075195 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764006981626153 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.00764006981626153, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764006981626153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17397423088550568, + 0.03289709612727165, + 0.12847784161567688, + 0.11753291636705399, + 0.03525640442967415, + 0.2439124435186386, + 0.04156864807009697, + 0.06951945275068283, + 0.07193762063980103, + 0.08492328971624374 + ], + "credit_map": [ + 0.021464955061674118, + 0.02233075723052025, + 0.014755569398403168, + 0.01367929857224226, + 0.015860408544540405, + 0.02423267439007759, + 0.014279783703386784, + 0.0473707877099514, + 0.004562198184430599, + 0.005004273261874914, + 0.004512725863605738, + 0.003759191371500492, + 0.0035729750525206327, + 0.004195923916995525, + 0.0034865173511207104, + 0.0038032899610698223, + 0.011128440499305725, + 0.013592010363936424, + 0.006837609224021435, + 0.007577989250421524, + 0.042228814214468, + 0.02825901098549366, + 0.01058528758585453, + 0.008268688805401325, + 0.013950248248875141, + 0.018327219411730766, + 0.01723029650747776, + 0.011075574904680252, + 0.016923386603593826, + 0.020284656435251236, + 0.005785757675766945, + 0.013955782167613506, + 0.004966465290635824, + 0.005892282351851463, + 0.006146745290607214, + 0.003115769475698471, + 0.005606095306575298, + 0.0028008774388581514, + 0.003217147197574377, + 0.003511021612212062, + 0.04733175039291382, + 0.038404401391744614, + 0.04614412039518356, + 0.017864294350147247, + 0.03874295949935913, + 0.018988633528351784, + 0.020218465477228165, + 0.016217822209000587, + 0.0056637926027178764, + 0.005164196714758873, + 0.005775486119091511, + 0.0033443754073232412, + 0.004543596878647804, + 0.00790430512279272, + 0.004168060142546892, + 0.005004833918064833, + 0.012312663719058037, + 0.010296761989593506, + 0.004908510018140078, + 0.008831486105918884, + 0.007361646741628647, + 0.010781136341392994, + 0.0098732253536582, + 0.0051540229469537735, + 0.008716355077922344, + 0.006254906766116619, + 0.006553913466632366, + 0.009422616101801395, + 0.007778815925121307, + 0.009373726323246956, + 0.011212112382054329, + 0.012625169940292835, + 0.008081886917352676, + 0.00907683465629816, + 0.005960012320429087, + 0.005882600788027048, + 0.02007053792476654, + 0.011336023919284344, + 0.010358069092035294, + 0.014157326892018318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.00764006981626153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783060289919376 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783060289919376, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783060289919376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3689996600151062, + 0.05700014531612396, + 0.12867216765880585, + 0.06445030122995377, + 0.0676996186375618, + 0.07715228199958801, + 0.04271754249930382, + 0.06912371516227722, + 0.08628934621810913, + 0.03789525851607323 + ], + "credit_map": [ + 0.038509976118803024, + 0.07607761025428772, + 0.015447039157152176, + 0.02043318562209606, + 0.04046027362346649, + 0.05758993327617645, + 0.022464953362941742, + 0.09801667928695679, + 0.007603132165968418, + 0.010656683705747128, + 0.0035523187834769487, + 0.005162008572369814, + 0.0067791324108839035, + 0.008301369845867157, + 0.007008139975368977, + 0.00793735682964325, + 0.009783899411559105, + 0.018007799983024597, + 0.009172998368740082, + 0.009369590319693089, + 0.03591516613960266, + 0.023985285311937332, + 0.011345995590090752, + 0.01109143067151308, + 0.0057013025507330894, + 0.009145654737949371, + 0.009126896969974041, + 0.00809010211378336, + 0.010504714213311672, + 0.011156247928738594, + 0.003902322379872203, + 0.0068230596370995045, + 0.016378479078412056, + 0.018894951790571213, + 0.008636318147182465, + 0.004358798265457153, + 0.005600457079708576, + 0.003474421799182892, + 0.004142792895436287, + 0.0062133981846272945, + 0.014453014358878136, + 0.008039426989853382, + 0.01191726978868246, + 0.010239475406706333, + 0.00678740069270134, + 0.007392313797026873, + 0.009170019067823887, + 0.009153362363576889, + 0.005571487359702587, + 0.005244449246674776, + 0.005258297082036734, + 0.005979897920042276, + 0.004451926331967115, + 0.00584746990352869, + 0.00493656937032938, + 0.005427447147667408, + 0.009010755456984043, + 0.010106301866471767, + 0.007220122963190079, + 0.010136492550373077, + 0.005969622638076544, + 0.009414740838110447, + 0.012519914656877518, + 0.004745763260871172, + 0.010277698747813702, + 0.010309448465704918, + 0.007463885936886072, + 0.013641164638102055, + 0.006510526407510042, + 0.009691024199128151, + 0.013481694273650646, + 0.01491390448063612, + 0.00498500419780612, + 0.004490338731557131, + 0.0036696004681289196, + 0.0026459055952727795, + 0.00640427740290761, + 0.005204943474382162, + 0.004738058429211378, + 0.005757130682468414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.009783060289919376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027894377708435 + ], + "opposite_logprob_tokens": [ + -2.350278854370117 + ], + "logprob_reward": -0.10027894377708435, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027894377708435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3119228482246399, + 0.11460007727146149, + 0.14870722591876984, + 0.04738406091928482, + 0.0233218502253294, + 0.15484966337680817, + 0.06412496417760849, + 0.020749233663082123, + 0.034488365054130554, + 0.07985170930624008 + ], + "credit_map": [ + 0.027170896530151367, + 0.07051035761833191, + 0.02147478424012661, + 0.014663023874163628, + 0.032070644199848175, + 0.04333743825554848, + 0.01821824535727501, + 0.08447745442390442, + 0.01206082385033369, + 0.023435860872268677, + 0.006964307278394699, + 0.006536220200359821, + 0.013250971212983131, + 0.028050914406776428, + 0.013688894920051098, + 0.010612077079713345, + 0.012590729631483555, + 0.016896851360797882, + 0.008899468928575516, + 0.011835191398859024, + 0.04468437656760216, + 0.030440911650657654, + 0.01188359409570694, + 0.011476102285087109, + 0.0037706356961280107, + 0.008680593222379684, + 0.005707734730094671, + 0.005629718769341707, + 0.008412269875407219, + 0.007903642021119595, + 0.0030699612107127905, + 0.004209504462778568, + 0.003761498723179102, + 0.0036182748153805733, + 0.0034994485322386026, + 0.0032554124481976032, + 0.002902862150222063, + 0.0014253932749852538, + 0.002304171910509467, + 0.002554788952693343, + 0.03315136209130287, + 0.029600756242871284, + 0.022524237632751465, + 0.011785008013248444, + 0.026791011914610863, + 0.011331114917993546, + 0.009974871762096882, + 0.009691299870610237, + 0.008959045633673668, + 0.007952003739774227, + 0.006003752816468477, + 0.007538692094385624, + 0.005879635456949472, + 0.0073248594999313354, + 0.00804990530014038, + 0.01241707056760788, + 0.0025918642058968544, + 0.0029281217139214277, + 0.002035925630480051, + 0.0026824488304555416, + 0.003370396327227354, + 0.002722105011343956, + 0.0030590256210416555, + 0.0013593473704531789, + 0.004905343055725098, + 0.004441482946276665, + 0.0026414315216243267, + 0.004658661317080259, + 0.003831783076748252, + 0.004623166285455227, + 0.005170136224478483, + 0.0042163603939116, + 0.007307198829948902, + 0.011712570674717426, + 0.00381158827804029, + 0.005189821124076843, + 0.02115277759730816, + 0.010812434367835522, + 0.008392476476728916, + 0.01147284172475338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.10027894377708435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190667135640979 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.003190667135640979, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190667135640979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2820264995098114, + 0.2830910086631775, + 0.08892518281936646, + 0.03999611735343933, + 0.0326722115278244, + 0.09508416801691055, + 0.06614606082439423, + 0.023604217916727066, + 0.02593374252319336, + 0.06252080202102661 + ], + "credit_map": [ + 0.027747489511966705, + 0.061820127069950104, + 0.014191338792443275, + 0.01189669594168663, + 0.021650198847055435, + 0.04323580861091614, + 0.016247732564806938, + 0.08523708581924438, + 0.030147375538945198, + 0.06747506558895111, + 0.01534417737275362, + 0.01565530151128769, + 0.024638064205646515, + 0.0701962560415268, + 0.02607758529484272, + 0.033557165414094925, + 0.010744381695985794, + 0.00911454576998949, + 0.007179444655776024, + 0.006734059192240238, + 0.01909734681248665, + 0.015314181335270405, + 0.01396394893527031, + 0.00677727023139596, + 0.0033394484780728817, + 0.005722430534660816, + 0.005740462336689234, + 0.005953352432698011, + 0.006591794081032276, + 0.005957223474979401, + 0.002254250692203641, + 0.004437155090272427, + 0.006557576358318329, + 0.007019806653261185, + 0.0035573875065892935, + 0.0029907708521932364, + 0.0038697165437042713, + 0.002650086535140872, + 0.0026695383712649345, + 0.0033573268447071314, + 0.02067231945693493, + 0.0172169990837574, + 0.01589222252368927, + 0.007289971690624952, + 0.0148988151922822, + 0.006307541858404875, + 0.005814103409647942, + 0.006992196198552847, + 0.00833609513938427, + 0.013653391972184181, + 0.005351785104721785, + 0.005480803083628416, + 0.009070205502212048, + 0.010362218134105206, + 0.006124918814748526, + 0.007766647730022669, + 0.003015553578734398, + 0.0029344926588237286, + 0.002457453403621912, + 0.002970821922644973, + 0.003549396526068449, + 0.003544958308339119, + 0.0033668060787022114, + 0.0017647346248850226, + 0.003744647605344653, + 0.003308485960587859, + 0.0026359015610069036, + 0.0026331369299441576, + 0.002531979698687792, + 0.00432015722617507, + 0.0027721631340682507, + 0.003987269476056099, + 0.005365589167922735, + 0.007508896756917238, + 0.004130609333515167, + 0.0036876858212053776, + 0.014133482240140438, + 0.007940252311527729, + 0.008056065067648888, + 0.011698217131197453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.003190667135640979 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a black-and-white world of thick film grain and high-contrast lighting, a rapidly rotating, circling camera closely follows a black cat as it performs an astonishing feat on the parallel bars. The cat's body is pitch black, its fur smooth as satin, its silver highlights reflecting off the single, harsh spotlight cast from above. The definition of its muscles is clearly visible with every thrust. Its eyes, narrowed slightly, gleam with a cunning and cold light, as if enjoying this dark performance all its own. With its front paws firmly grasping the cold metal bars, it smoothly executes a series of complex swings and twists with a gesture that blends the grace of a ballerina with the strength of a top athlete. Its body carves perfect arcs through the air, its long tail acting as a precise rudder, maintaining its incredible balance. The camera's rapid rotation blurs the background, focusing entirely on the cat's sinisterly beautiful movements. Deep shadows dance wildly on the ground as it moves, creating a surreal atmosphere that is both dizzying and slightly unsettling.", + "seed": 29281, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_6/index_9033_seed_29281_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02415339248254895, + "vlm_logprob": -0.02415339248254895, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.02415339248254895, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006445097737014294 + ], + "opposite_logprob_tokens": [ + -5.881444931030273 + ], + "logprob_reward": -0.006445097737014294, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006445097737014294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1831929087638855, + 0.10993091762065887, + 0.07889552414417267, + 0.07531513273715973, + 0.12434955686330795, + 0.10687988251447678, + 0.07754120975732803, + 0.044760093092918396, + 0.13256463408470154, + 0.06657014042139053 + ], + "credit_map": [ + 0.027923721820116043, + 0.017859874293208122, + 0.01813347265124321, + 0.018501993268728256, + 0.021097151562571526, + 0.027917835861444473, + 0.017112065106630325, + 0.03464679419994354, + 0.013835037127137184, + 0.010019072331488132, + 0.01594691164791584, + 0.012791305780410767, + 0.010259447619318962, + 0.013464754447340965, + 0.012656531296670437, + 0.020957857370376587, + 0.010906798765063286, + 0.009354199282824993, + 0.012881958857178688, + 0.009012512862682343, + 0.0099589042365551, + 0.00879112258553505, + 0.0063148909248411655, + 0.01167512871325016, + 0.01787351444363594, + 0.0071853455156087875, + 0.00835242960602045, + 0.0059107830747962, + 0.007375181186944246, + 0.009146028198301792, + 0.007083903066813946, + 0.01238794345408678, + 0.01468681450933218, + 0.02001786231994629, + 0.012708904221653938, + 0.01071852631866932, + 0.015017838217318058, + 0.019266974180936813, + 0.01589886285364628, + 0.016033772379159927, + 0.012552523985505104, + 0.01761675253510475, + 0.010893472470343113, + 0.008179248310625553, + 0.020937271416187286, + 0.0161509420722723, + 0.011674022302031517, + 0.008875643834471703, + 0.010795483365654945, + 0.010687659494578838, + 0.00901934877038002, + 0.007559900637716055, + 0.016468077898025513, + 0.009540289640426636, + 0.007069312036037445, + 0.006401130463927984, + 0.005582180805504322, + 0.0050136623904109, + 0.00499073788523674, + 0.005600349511951208, + 0.00510215712711215, + 0.00557959359139204, + 0.005056906957179308, + 0.007834500633180141, + 0.014658392407000065, + 0.027390403673052788, + 0.01485166884958744, + 0.011312384158372879, + 0.018630482256412506, + 0.015672806650400162, + 0.016865625977516174, + 0.013182866387069225, + 0.008036738261580467, + 0.010687895119190216, + 0.009204600006341934, + 0.007709656376391649, + 0.009200884960591793, + 0.007215222343802452, + 0.008082574233412743, + 0.006432562600821257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.006445097737014294 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608150638639927 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.01608150638639927, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608150638639927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13339704275131226, + 0.1180611103773117, + 0.07462392747402191, + 0.03597022965550423, + 0.06249123439192772, + 0.12113530188798904, + 0.02235955372452736, + 0.12951979041099548, + 0.2509653568267822, + 0.051476385444402695 + ], + "credit_map": [ + 0.020664580166339874, + 0.011453982442617416, + 0.015118852257728577, + 0.013952802866697311, + 0.014558553695678711, + 0.01893216371536255, + 0.012443632818758488, + 0.0262724868953228, + 0.016888437792658806, + 0.006695882882922888, + 0.014468912966549397, + 0.010327008552849293, + 0.013114092871546745, + 0.016328468918800354, + 0.01831144280731678, + 0.021926870569586754, + 0.010459251701831818, + 0.010672143660485744, + 0.008431490510702133, + 0.010717668570578098, + 0.010163445957005024, + 0.0076216296292841434, + 0.009329070337116718, + 0.007229230832308531, + 0.006722670514136553, + 0.0024917430710047483, + 0.004620070103555918, + 0.0028657224029302597, + 0.0024859944824129343, + 0.00408881576731801, + 0.00336713925935328, + 0.00932807382196188, + 0.007703656330704689, + 0.0067085749469697475, + 0.008567944169044495, + 0.007307893130928278, + 0.005791048519313335, + 0.008226616308093071, + 0.010197383351624012, + 0.00798811949789524, + 0.014146908186376095, + 0.0194883830845356, + 0.011536470614373684, + 0.010015620850026608, + 0.020868316292762756, + 0.01841447688639164, + 0.013928467407822609, + 0.012736663222312927, + 0.002670455025509, + 0.0031867518555372953, + 0.003724335692822933, + 0.0019925890956074, + 0.0027712699957191944, + 0.0027117538265883923, + 0.0022146785631775856, + 0.0030877203680574894, + 0.017208650708198547, + 0.014025215059518814, + 0.00949750654399395, + 0.019822318106889725, + 0.021107405424118042, + 0.016044938936829567, + 0.01267560850828886, + 0.019138149917125702, + 0.028139928355813026, + 0.0497102364897728, + 0.032511405646800995, + 0.018805956467986107, + 0.030968202278017998, + 0.033195678144693375, + 0.027461476624011993, + 0.030172470957040787, + 0.006123379338532686, + 0.008566359058022499, + 0.0066968114115297794, + 0.005260340869426727, + 0.007067240308970213, + 0.0059935408644378185, + 0.0076372274197638035, + 0.0041314889676868916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.01608150638639927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741517689079046 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741517689079046, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741517689079046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09663465619087219, + 0.12350481748580933, + 0.10471390187740326, + 0.1379644274711609, + 0.0834067314863205, + 0.10787874460220337, + 0.04017973691225052, + 0.15046951174736023, + 0.10878729075193405, + 0.04646020755171776 + ], + "credit_map": [ + 0.014019708149135113, + 0.00823050457984209, + 0.011055588722229004, + 0.01071001123636961, + 0.007823496125638485, + 0.01790185645222664, + 0.008256567642092705, + 0.01863691583275795, + 0.01647130772471428, + 0.008602021262049675, + 0.020230116322636604, + 0.014833603985607624, + 0.007580731064081192, + 0.014480693265795708, + 0.019487276673316956, + 0.021819062530994415, + 0.013394162990152836, + 0.013531185686588287, + 0.018641630187630653, + 0.010326548479497433, + 0.014025372453033924, + 0.012816355563700199, + 0.008260734379291534, + 0.01371790561825037, + 0.020592283457517624, + 0.018199114128947258, + 0.013807902112603188, + 0.021552523598074913, + 0.02174665778875351, + 0.015113419853150845, + 0.014963238500058651, + 0.011989284306764603, + 0.011367364786565304, + 0.007839889265596867, + 0.011422711424529552, + 0.010550807230174541, + 0.0070333704352378845, + 0.011989842168986797, + 0.012585083022713661, + 0.010617656633257866, + 0.013844512403011322, + 0.01186153944581747, + 0.008479863405227661, + 0.011196915991604328, + 0.01673353835940361, + 0.018738524988293648, + 0.01416421215981245, + 0.012859629467129707, + 0.005102075636386871, + 0.005100277252495289, + 0.004490571096539497, + 0.0035437610931694508, + 0.009298940189182758, + 0.005418891552835703, + 0.0043409732170403, + 0.002884248271584511, + 0.020796846598386765, + 0.026636067777872086, + 0.01913173496723175, + 0.018798379227519035, + 0.017196713015437126, + 0.01758577674627304, + 0.01329672522842884, + 0.01702725701034069, + 0.01185974944382906, + 0.02204202301800251, + 0.013836550526320934, + 0.0076421708799898624, + 0.014404370449483395, + 0.013865359127521515, + 0.01266816258430481, + 0.01246890239417553, + 0.00615726551041007, + 0.007600321434438229, + 0.00430268282070756, + 0.004912219941616058, + 0.008682784624397755, + 0.005096183158457279, + 0.006565392017364502, + 0.003143356880173087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.006741517689079046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901165425777435 + ], + "opposite_logprob_tokens": [ + -2.4640116691589355 + ], + "logprob_reward": -0.08901165425777435, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901165425777435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12262111157178879, + 0.1534419059753418, + 0.15713872015476227, + 0.07635290175676346, + 0.18623493611812592, + 0.06071384996175766, + 0.045986514538526535, + 0.063024140894413, + 0.07162722945213318, + 0.06285867094993591 + ], + "credit_map": [ + 0.017073845490813255, + 0.012668661773204803, + 0.01304388977587223, + 0.013317085802555084, + 0.01564355567097664, + 0.015341815538704395, + 0.013039003126323223, + 0.022493261843919754, + 0.022572100162506104, + 0.009477085433900356, + 0.020939264446496964, + 0.014391215518116951, + 0.014906405471265316, + 0.022493697702884674, + 0.019365541636943817, + 0.02929658629000187, + 0.023300891742110252, + 0.017020760104060173, + 0.0216488279402256, + 0.019255325198173523, + 0.020702311769127846, + 0.01883416809141636, + 0.016041481867432594, + 0.020334957167506218, + 0.01235991157591343, + 0.01071221474558115, + 0.006370775867253542, + 0.013471201062202454, + 0.01052586454898119, + 0.008474006317555904, + 0.008141416124999523, + 0.006297507788985968, + 0.02581457607448101, + 0.015021780505776405, + 0.024759020656347275, + 0.023717978969216347, + 0.012989870272576809, + 0.02568504773080349, + 0.03224337473511696, + 0.026003288105130196, + 0.007162801921367645, + 0.007917801849544048, + 0.005105298012495041, + 0.004250355064868927, + 0.010644655674695969, + 0.012241763062775135, + 0.00657234713435173, + 0.006818827707320452, + 0.005832980386912823, + 0.006286941934376955, + 0.00520396139472723, + 0.004011505749076605, + 0.011082411743700504, + 0.0058782706037163734, + 0.0037402764428406954, + 0.003950168844312429, + 0.010063454508781433, + 0.01157432608306408, + 0.007714050821959972, + 0.0066205584444105625, + 0.006986074149608612, + 0.0077288090251386166, + 0.00692133791744709, + 0.005415528081357479, + 0.008353297598659992, + 0.012521285563707352, + 0.009304284118115902, + 0.004836921114474535, + 0.007614294067025185, + 0.008930540643632412, + 0.011679831892251968, + 0.00838677678257227, + 0.008412460796535015, + 0.010332637466490269, + 0.007973967120051384, + 0.007108376361429691, + 0.008536336943507195, + 0.006996069569140673, + 0.007912877947092056, + 0.005585944280028343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.08901165425777435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024871863424777985 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024871863424777985, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024871863424777985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07056992501020432, + 0.17494003474712372, + 0.07266142964363098, + 0.12277060747146606, + 0.3088665306568146, + 0.05189540982246399, + 0.02704986371099949, + 0.043902162462472916, + 0.045561015605926514, + 0.0817829966545105 + ], + "credit_map": [ + 0.008119655773043633, + 0.007501899264752865, + 0.006397820543497801, + 0.00999850407242775, + 0.007726550567895174, + 0.006680939346551895, + 0.011388083919882774, + 0.01275647897273302, + 0.02326056733727455, + 0.011249753646552563, + 0.02188042551279068, + 0.019155068323016167, + 0.018903806805610657, + 0.02237386628985405, + 0.024168817326426506, + 0.03394772484898567, + 0.010914934799075127, + 0.008810877799987793, + 0.009791078977286816, + 0.010387586429715157, + 0.00826080609112978, + 0.007650379557162523, + 0.006683871615678072, + 0.01016189157962799, + 0.02009587548673153, + 0.015310177579522133, + 0.015228252857923508, + 0.016384702175855637, + 0.015937265008687973, + 0.013786237686872482, + 0.010469772852957249, + 0.015558329410851002, + 0.04129718244075775, + 0.032822053879499435, + 0.03678421676158905, + 0.03225447237491608, + 0.023629548028111458, + 0.04774156212806702, + 0.05245994031429291, + 0.04187754541635513, + 0.006334367673844099, + 0.007622881326824427, + 0.004481166135519743, + 0.00401254603639245, + 0.009201142005622387, + 0.009634170681238174, + 0.005417953245341778, + 0.005191183183342218, + 0.0034715975634753704, + 0.003008629661053419, + 0.0031552985310554504, + 0.002296434948220849, + 0.0049059102311730385, + 0.0030296947807073593, + 0.0025974761229008436, + 0.004584823735058308, + 0.0055410717613995075, + 0.008137383498251438, + 0.004461964126676321, + 0.004768667742609978, + 0.006156249903142452, + 0.005187541712075472, + 0.005860964301973581, + 0.003788315923884511, + 0.004850011318922043, + 0.009540845640003681, + 0.005591558758169413, + 0.0037081006448715925, + 0.004562443122267723, + 0.005754040088504553, + 0.0066480208188295364, + 0.004905999172478914, + 0.0100879967212677, + 0.012773888185620308, + 0.009927562437951565, + 0.007927240803837776, + 0.013618183322250843, + 0.009707323275506496, + 0.011395405977964401, + 0.006345393601804972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.0024871863424777985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a black-and-white world of thick film grain and high-contrast lighting, a rapidly rotating, circling camera closely follows a black cat as it performs an astonishing feat on the parallel bars. The cat's body is pitch black, its fur smooth as satin, its silver highlights reflecting off the single, harsh spotlight cast from above. The definition of its muscles is clearly visible with every thrust. Its eyes, narrowed slightly, gleam with a cunning and cold light, as if enjoying this dark performance all its own. With its front paws firmly grasping the cold metal bars, it smoothly executes a series of complex swings and twists with a gesture that blends the grace of a ballerina with the strength of a top athlete. Its body carves perfect arcs through the air, its long tail acting as a precise rudder, maintaining its incredible balance. The camera's rapid rotation blurs the background, focusing entirely on the cat's sinisterly beautiful movements. Deep shadows dance wildly on the ground as it moves, creating a surreal atmosphere that is both dizzying and slightly unsettling.", + "seed": 29282, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_6/index_9033_seed_29282_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013763820938766002, + "vlm_logprob": -0.013763820938766002, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.013763820938766002, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004983622580766678 + ], + "opposite_logprob_tokens": [ + -6.004983425140381 + ], + "logprob_reward": -0.004983622580766678, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.004983622580766678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17497701942920685, + 0.0527302548289299, + 0.0576210655272007, + 0.1191701889038086, + 0.1405085027217865, + 0.08871684968471527, + 0.08004456013441086, + 0.04835517704486847, + 0.1470528244972229, + 0.09082362055778503 + ], + "credit_map": [ + 0.02162882126867771, + 0.02113185077905655, + 0.02507851831614971, + 0.032898060977458954, + 0.023795461282134056, + 0.013253462500870228, + 0.026448769494891167, + 0.0107420664280653, + 0.006844086106866598, + 0.005173330195248127, + 0.007254765834659338, + 0.005636271089315414, + 0.010455837473273277, + 0.005462910048663616, + 0.005488562397658825, + 0.006414484232664108, + 0.005620064213871956, + 0.007108879275619984, + 0.00847415067255497, + 0.007230210583657026, + 0.01065074373036623, + 0.0065730963833630085, + 0.008234329521656036, + 0.0037295836955308914, + 0.01484181359410286, + 0.01120518334209919, + 0.021236147731542587, + 0.009820591658353806, + 0.018451223149895668, + 0.018247144296765327, + 0.01826515980064869, + 0.007102916017174721, + 0.012292321771383286, + 0.035231441259384155, + 0.018812207505106926, + 0.010044274851679802, + 0.03255973383784294, + 0.012620290741324425, + 0.00879722274839878, + 0.010151000693440437, + 0.012524807825684547, + 0.009908493608236313, + 0.01622934080660343, + 0.008379345759749413, + 0.00881576631218195, + 0.010787038132548332, + 0.011418242007493973, + 0.010653812438249588, + 0.008500376716256142, + 0.013905659317970276, + 0.007274496369063854, + 0.007087424863129854, + 0.0165864247828722, + 0.01054705772548914, + 0.008773840963840485, + 0.007369277998805046, + 0.004883493296802044, + 0.005399297457188368, + 0.006282759830355644, + 0.005567571613937616, + 0.0052860029973089695, + 0.005521143786609173, + 0.006102111656218767, + 0.009312798269093037, + 0.020286740735173225, + 0.020688004791736603, + 0.021294984966516495, + 0.016890691593289375, + 0.019782066345214844, + 0.012824918143451214, + 0.02003287523984909, + 0.015252531506121159, + 0.008945553563535213, + 0.020353225991129875, + 0.008894795551896095, + 0.007619265001267195, + 0.009855424053966999, + 0.014873247593641281, + 0.010440596379339695, + 0.009841508232057095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.004983622580766678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420564204454422 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.01420564204454422, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420564204454422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18503284454345703, + 0.18792353570461273, + 0.034693554043769836, + 0.08448521792888641, + 0.06754797697067261, + 0.05763855576515198, + 0.06435530632734299, + 0.062466952949762344, + 0.12462837249040604, + 0.13122773170471191 + ], + "credit_map": [ + 0.017976872622966766, + 0.016969630494713783, + 0.02886459231376648, + 0.02875751070678234, + 0.021284371614456177, + 0.02436239644885063, + 0.02973480336368084, + 0.01708265393972397, + 0.01353995781391859, + 0.021490884944796562, + 0.020900418981909752, + 0.027803974226117134, + 0.03184903785586357, + 0.02835765853524208, + 0.025497127324342728, + 0.01848446950316429, + 0.0032062800601124763, + 0.004181039985269308, + 0.0032878045458346605, + 0.006028047297149897, + 0.005434808321297169, + 0.0036306048277765512, + 0.00622965395450592, + 0.002695314818993211, + 0.009123357012867928, + 0.00835124496370554, + 0.010345447808504105, + 0.00866372138261795, + 0.013142628595232964, + 0.020271124318242073, + 0.00995574053376913, + 0.0046319542452692986, + 0.005224904976785183, + 0.011721164919435978, + 0.008059865795075893, + 0.006001515779644251, + 0.017437592148780823, + 0.009082251228392124, + 0.005269842222332954, + 0.004750838037580252, + 0.00865386612713337, + 0.005068088881671429, + 0.009845651686191559, + 0.005700675304979086, + 0.006023326888680458, + 0.008959985338151455, + 0.006204377394169569, + 0.007182585075497627, + 0.008436139672994614, + 0.011391655541956425, + 0.005784633103758097, + 0.005874848924577236, + 0.013155505061149597, + 0.006995456293225288, + 0.007586692925542593, + 0.005130374804139137, + 0.008159984834492207, + 0.01036118995398283, + 0.007292808964848518, + 0.007660082075744867, + 0.006181453354656696, + 0.00695396400988102, + 0.006101284641772509, + 0.009756184183061123, + 0.013487234711647034, + 0.0179080031812191, + 0.012465562671422958, + 0.018016818910837173, + 0.019810721278190613, + 0.011485756374895573, + 0.019043006002902985, + 0.012411271221935749, + 0.012163540348410606, + 0.02281073108315468, + 0.011086724698543549, + 0.013561611995100975, + 0.014895301312208176, + 0.020947197452187538, + 0.021434763446450233, + 0.01432785764336586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.01420564204454422 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098943900316954 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.004098943900316954, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098943900316954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.182489812374115, + 0.12563058733940125, + 0.05633013695478439, + 0.048346903175115585, + 0.1613757163286209, + 0.058073755353689194, + 0.10970517992973328, + 0.06429517269134521, + 0.0809558853507042, + 0.11279686540365219 + ], + "credit_map": [ + 0.028329353779554367, + 0.018564974889159203, + 0.022236356511712074, + 0.025623783469200134, + 0.022789152339100838, + 0.01805107854306698, + 0.03170472010970116, + 0.015190400183200836, + 0.014128158800303936, + 0.011469148099422455, + 0.014938220381736755, + 0.013547061011195183, + 0.025537928566336632, + 0.01482357457280159, + 0.01537152100354433, + 0.01581498049199581, + 0.005933323875069618, + 0.0074490513652563095, + 0.005811590701341629, + 0.007291346322745085, + 0.008535635657608509, + 0.007275679148733616, + 0.007994908839464188, + 0.006038601044565439, + 0.004106063395738602, + 0.006068985443562269, + 0.006298527121543884, + 0.005787026137113571, + 0.006396793294698, + 0.007331983186304569, + 0.007225566543638706, + 0.005131957121193409, + 0.015246002934873104, + 0.04676828533411026, + 0.013355952687561512, + 0.008896137587726116, + 0.03785516694188118, + 0.019878322258591652, + 0.0103744613006711, + 0.00900138821452856, + 0.007188189309090376, + 0.00616359943524003, + 0.008627419359982014, + 0.006823921576142311, + 0.0059537701308727264, + 0.010204387828707695, + 0.006273696199059486, + 0.0068387738429009914, + 0.011324035003781319, + 0.016836978495121002, + 0.012329990975558758, + 0.011019681580364704, + 0.015430085361003876, + 0.012194650247693062, + 0.016951700672507286, + 0.013618053868412971, + 0.008255968801677227, + 0.00915098749101162, + 0.009056435897946358, + 0.011332575231790543, + 0.005389079917222261, + 0.0065692756325006485, + 0.0069716558791697025, + 0.0075691998936235905, + 0.010119566693902016, + 0.011474661529064178, + 0.00802649836987257, + 0.009417662397027016, + 0.012963845394551754, + 0.009176485240459442, + 0.00960668083280325, + 0.010170486755669117, + 0.01279613096266985, + 0.025760570541024208, + 0.009500653482973576, + 0.010556583292782307, + 0.012849893420934677, + 0.019217241555452347, + 0.012046042829751968, + 0.010069750249385834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.004098943900316954 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304501786828041 + ], + "opposite_logprob_tokens": [ + -3.1680450439453125 + ], + "logprob_reward": -0.04304501786828041, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304501786828041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08938514441251755, + 0.1625838577747345, + 0.039542704820632935, + 0.04496900737285614, + 0.258574515581131, + 0.058921944350004196, + 0.08425673842430115, + 0.11912921816110611, + 0.03732753545045853, + 0.105309396982193 + ], + "credit_map": [ + 0.012271924875676632, + 0.014411570504307747, + 0.013762082904577255, + 0.014254089444875717, + 0.01038719154894352, + 0.006101708859205246, + 0.011988100595772266, + 0.006208473816514015, + 0.013618086464703083, + 0.012258240021765232, + 0.0251504797488451, + 0.013157187961041927, + 0.03697311133146286, + 0.02011905051767826, + 0.021162863820791245, + 0.020144835114479065, + 0.0039467113092541695, + 0.0048643141053617, + 0.004005821887403727, + 0.004943477921187878, + 0.006813365966081619, + 0.005141452420502901, + 0.0057769292034208775, + 0.004050629213452339, + 0.003720535896718502, + 0.004926815629005432, + 0.006729749497026205, + 0.004487552214413881, + 0.009807547554373741, + 0.006333895027637482, + 0.005992549937218428, + 0.0029703625477850437, + 0.0273478701710701, + 0.05511964485049248, + 0.027681101113557816, + 0.0210607647895813, + 0.03636695444583893, + 0.052812427282333374, + 0.023929886519908905, + 0.014255852438509464, + 0.005106829572468996, + 0.006173394620418549, + 0.008242510259151459, + 0.0076063042506575584, + 0.0052367872558534145, + 0.009410071186721325, + 0.008225355297327042, + 0.008920691907405853, + 0.00848242174834013, + 0.012935866601765156, + 0.008516086265444756, + 0.008959813974797726, + 0.013449409045279026, + 0.012889769859611988, + 0.011689674109220505, + 0.007333704270422459, + 0.01919412985444069, + 0.016886398196220398, + 0.018514368683099747, + 0.01937319152057171, + 0.010189867578446865, + 0.01385855209082365, + 0.01180644053965807, + 0.009306266903877258, + 0.004822017624974251, + 0.007150988094508648, + 0.0037223733961582184, + 0.00439062062650919, + 0.004953030031174421, + 0.003915062639862299, + 0.004988271743059158, + 0.003385169431567192, + 0.013932649046182632, + 0.020776743069291115, + 0.009085471741855145, + 0.01093438733369112, + 0.014459515921771526, + 0.010172151029109955, + 0.015183242969214916, + 0.010765226557850838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.04304501786828041 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with thick film grain and high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a black cat with smooth fur and silver highlights performing on parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat execute complex swings and twists while maintaining balance with its tail acting as a rudder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera rapidly rotating and circling to closely follow the cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred due to the camera's rapid rotation while a single harsh spotlight shines from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002485878299921751 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.002485878299921751, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002485878299921751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08809079229831696, + 0.14310842752456665, + 0.0644284337759018, + 0.10430175065994263, + 0.2509990334510803, + 0.046336960047483444, + 0.050073932856321335, + 0.07994148135185242, + 0.03732116147875786, + 0.1353980153799057 + ], + "credit_map": [ + 0.010592442005872726, + 0.013902639038860798, + 0.010006102733314037, + 0.012456034310162067, + 0.009991743601858616, + 0.010595401749014854, + 0.012601662427186966, + 0.007944771088659763, + 0.012132971547544003, + 0.01171329990029335, + 0.02317281812429428, + 0.013836417347192764, + 0.030946487560868263, + 0.018613316118717194, + 0.016008710488677025, + 0.01668441854417324, + 0.006810228805989027, + 0.007093584164977074, + 0.007715095765888691, + 0.007271877955645323, + 0.009128025732934475, + 0.00795720610767603, + 0.009256766177713871, + 0.00919564813375473, + 0.010580944828689098, + 0.0077773029915988445, + 0.014778682962059975, + 0.012030759826302528, + 0.016045695170760155, + 0.016030706465244293, + 0.020419254899024963, + 0.006638405378907919, + 0.02716626226902008, + 0.06896010041236877, + 0.029118774458765984, + 0.01676175557076931, + 0.04646232724189758, + 0.03280101343989372, + 0.019117318093776703, + 0.01061146892607212, + 0.005319408606737852, + 0.004386316053569317, + 0.009001736529171467, + 0.004246002994477749, + 0.0037659804802387953, + 0.006589359138160944, + 0.004806098993867636, + 0.008222060278058052, + 0.005301583092659712, + 0.007164797745645046, + 0.00558463903144002, + 0.005192000884562731, + 0.009116639383137226, + 0.0053557162173092365, + 0.0076136402785778046, + 0.004744914826005697, + 0.006999853532761335, + 0.008860412053763866, + 0.007636132184416056, + 0.008225657977163792, + 0.00951316487044096, + 0.015115492045879364, + 0.0090795261785388, + 0.014511238783597946, + 0.004535638261586428, + 0.004596991930156946, + 0.005264529958367348, + 0.0051135472021996975, + 0.004197047557681799, + 0.003618111601099372, + 0.0058159735053777695, + 0.004179320763796568, + 0.01769605651497841, + 0.01476906519383192, + 0.01870041713118553, + 0.020730702206492424, + 0.016582660377025604, + 0.012371049262583256, + 0.021381884813308716, + 0.013166178949177265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.002485878299921751 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_tensor([1435, 1435])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_tensor([1435, 1435])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e4e4a75d3cac2e175e06428e2a0b0bee73bc0173 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_tensor([1435, 1435])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "super-slow-motion overhead shot, perfectly still, focuses on a massive, vintage Baroque-style crystal chandelier suspended in the center of the majestic opera house's dome. This magnificent fixture, composed of layered brass frames and hundreds of intricately carved crystal pendants, radiates a warm, dazzling golden light, casting intricate shadows on the empty red velvet seats below. Suddenly, with a powerful, silent tremor, the entire chandelier begins to shake violently, the crystal pendants clashing with each other, producing an increasingly rapid and crisp tinkling sound. As the tremor reaches its peak, the chandelier's main metal frame suddenly bends and breaks painfully under the unbearable stress, leading to a catastrophic structural collapse. Countless crystal fragments and glass beads exploded violently from the center outward, then poured down like a deadly yet beautiful flash rain. In slow motion, each irregular fragment slowly rolled and rotated in the air, its sharp cross-section capturing and madly refracting the last flicker of light before the chandelier went out, drawing short and sharp arcs of cold light against the gradually darkening background. The whole picture was full of sudden destructiveness and a heart-pounding tragic beauty.", + "seed": 11862, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_1435_seed_11862_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5458802745677531, + "vlm_logprob": -0.5458802745677531, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.5458802745677531, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011377571150660515 + ], + "opposite_logprob_tokens": [ + -5.386377334594727 + ], + "logprob_reward": -0.011377571150660515, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.011377571150660515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1785171777009964, + 0.057268764823675156, + 0.06106627732515335, + 0.05842207372188568, + 0.15776367485523224, + 0.054732345044612885, + 0.07441132515668869, + 0.053012337535619736, + 0.14557114243507385, + 0.15923485159873962 + ], + "credit_map": [ + 0.026378849521279335, + 0.024469275027513504, + 0.017534011974930763, + 0.020895859226584435, + 0.025116000324487686, + 0.01668878272175789, + 0.02623545378446579, + 0.021198928356170654, + 0.008991342037916183, + 0.008898123167455196, + 0.005454602185636759, + 0.0062804738990962505, + 0.00793642271310091, + 0.007582552265375853, + 0.0066329860128462315, + 0.005492259282618761, + 0.011325661092996597, + 0.008337760344147682, + 0.006645523477345705, + 0.005111915059387684, + 0.007901952601969242, + 0.01067543774843216, + 0.005436855833977461, + 0.005631173495203257, + 0.006342420354485512, + 0.007612987421452999, + 0.004975140560418367, + 0.007375427521765232, + 0.008904741145670414, + 0.008439863100647926, + 0.00831974670290947, + 0.006451746448874474, + 0.024298029020428658, + 0.02369757927954197, + 0.011897331103682518, + 0.019282065331935883, + 0.0240168496966362, + 0.023303763940930367, + 0.019366592168807983, + 0.011901481077075005, + 0.009299918077886105, + 0.006428492721170187, + 0.006207692436873913, + 0.005418851505964994, + 0.008406965993344784, + 0.008507943712174892, + 0.005150382872670889, + 0.005312097724527121, + 0.008188959211111069, + 0.008678088895976543, + 0.00818321667611599, + 0.010765986517071724, + 0.011221123859286308, + 0.008870751596987247, + 0.01025407575070858, + 0.008249125443398952, + 0.005732976831495762, + 0.0052598887123167515, + 0.0066147055476903915, + 0.009518892504274845, + 0.006256519351154566, + 0.006452068220824003, + 0.0063774739392101765, + 0.006799814756959677, + 0.02326621115207672, + 0.01564800925552845, + 0.028401145711541176, + 0.015464209951460361, + 0.01246277429163456, + 0.023374753072857857, + 0.012079962529242039, + 0.014874078333377838, + 0.026715727522969246, + 0.018238607794046402, + 0.026654794812202454, + 0.015180026181042194, + 0.015994112938642502, + 0.01985374465584755, + 0.01270754262804985, + 0.02389029785990715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.011377571150660515 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003640929702669382 + ], + "opposite_logprob_tokens": [ + -5.628641128540039 + ], + "logprob_reward": -0.003640929702669382, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003640929702669382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1472068727016449, + 0.05994982272386551, + 0.044084396213293076, + 0.0848400890827179, + 0.10268006473779678, + 0.1745031327009201, + 0.13880497217178345, + 0.09666070342063904, + 0.050760943442583084, + 0.10050902515649796 + ], + "credit_map": [ + 0.015766160562634468, + 0.017591306939721107, + 0.019305139780044556, + 0.02188648097217083, + 0.024195300415158272, + 0.011706439778208733, + 0.017196185886859894, + 0.01955985277891159, + 0.005551142152398825, + 0.008363007567822933, + 0.006630868650972843, + 0.008234231732785702, + 0.009189647622406483, + 0.0068582771345973015, + 0.008487764745950699, + 0.006634887773543596, + 0.006157442461699247, + 0.006704619154334068, + 0.004910077899694443, + 0.004520083777606487, + 0.005587979219853878, + 0.0065971557050943375, + 0.005529471207410097, + 0.004077569581568241, + 0.012800458818674088, + 0.01175373513251543, + 0.0074134706519544125, + 0.00937487930059433, + 0.013405599631369114, + 0.011659730225801468, + 0.011060848832130432, + 0.007371368817985058, + 0.01859874464571476, + 0.010637996718287468, + 0.010959671810269356, + 0.013919535093009472, + 0.016819709911942482, + 0.01287356298416853, + 0.010231364518404007, + 0.008639479987323284, + 0.02838587947189808, + 0.021087173372507095, + 0.022303657606244087, + 0.024108629673719406, + 0.025447657331824303, + 0.021255746483802795, + 0.018354633823037148, + 0.013559749349951744, + 0.017375130206346512, + 0.018461275845766068, + 0.014944396913051605, + 0.01773308403789997, + 0.019498245790600777, + 0.018238360062241554, + 0.01939684897661209, + 0.013157615438103676, + 0.011431054212152958, + 0.010684600099921227, + 0.012770493514835835, + 0.014459928497672081, + 0.01202110480517149, + 0.011535687372088432, + 0.013020938262343407, + 0.01073689665645361, + 0.008233005180954933, + 0.004699036478996277, + 0.006563667673617601, + 0.006206585559993982, + 0.005987682845443487, + 0.009061580523848534, + 0.00452380022034049, + 0.005485589616000652, + 0.019854214042425156, + 0.01048737671226263, + 0.014043799601495266, + 0.00863097608089447, + 0.0077696447260677814, + 0.012914515100419521, + 0.009523834101855755, + 0.01728466898202896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.003640929702669382 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.2379064559936523 + ], + "opposite_logprob_tokens": [ + -0.1129065752029419 + ], + "logprob_reward": -2.2379064559936523, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.2379064559936523, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11793693900108337, + 0.061356063932180405, + 0.06475728750228882, + 0.04580552130937576, + 0.25466108322143555, + 0.1693647801876068, + 0.14364071190357208, + 0.05125489458441734, + 0.029712092131376266, + 0.06151062250137329 + ], + "credit_map": [ + 0.014194771647453308, + 0.01593312807381153, + 0.0176345556974411, + 0.015244701877236366, + 0.016111597418785095, + 0.012856293469667435, + 0.014442509040236473, + 0.011519386433064938, + 0.007415330968797207, + 0.008231146261096, + 0.007357041351497173, + 0.00790427252650261, + 0.008296014741063118, + 0.007859254255890846, + 0.007960350252687931, + 0.006332656368613243, + 0.011040663346648216, + 0.006846636068075895, + 0.007274118717759848, + 0.0066306632943451405, + 0.0082398671656847, + 0.008290279656648636, + 0.009424187242984772, + 0.007010866422206163, + 0.006866776384413242, + 0.005849860608577728, + 0.004254227038472891, + 0.0056016999296844006, + 0.00579261127859354, + 0.007492891978472471, + 0.005939176771789789, + 0.0040082745254039764, + 0.05618925020098686, + 0.03560252487659454, + 0.025942925363779068, + 0.02446029707789421, + 0.031039254739880562, + 0.03371970355510712, + 0.02733921818435192, + 0.0203679408878088, + 0.01991833932697773, + 0.022360607981681824, + 0.019746404141187668, + 0.02307005040347576, + 0.027160750702023506, + 0.02157890796661377, + 0.02242632769048214, + 0.013103388249874115, + 0.017935506999492645, + 0.018457695841789246, + 0.014522436074912548, + 0.019590698182582855, + 0.02008483000099659, + 0.024248624220490456, + 0.01727754808962345, + 0.011523379944264889, + 0.006850971840322018, + 0.005769117269665003, + 0.008561619557440281, + 0.006455837748944759, + 0.006617036648094654, + 0.005542452447116375, + 0.005714390426874161, + 0.005743469577282667, + 0.0038096678908914328, + 0.0032089001033455133, + 0.0032148079480975866, + 0.004735175985842943, + 0.0034058711025863886, + 0.0047013224102556705, + 0.0027516542468219995, + 0.003884692909196019, + 0.008501152507960796, + 0.005845409817993641, + 0.005763066466897726, + 0.008538995869457722, + 0.007164824288338423, + 0.007651856169104576, + 0.0067612300626933575, + 0.011284091509878635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -2.2379064559936523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08919037878513336 + ], + "opposite_logprob_tokens": [ + -2.4641904830932617 + ], + "logprob_reward": -0.08919037878513336, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08919037878513336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10358887165784836, + 0.05776922032237053, + 0.053171250969171524, + 0.09765242785215378, + 0.1410723775625229, + 0.1500510722398758, + 0.22088459134101868, + 0.06792505085468292, + 0.04836991801857948, + 0.05951516330242157 + ], + "credit_map": [ + 0.013611223548650742, + 0.0139383003115654, + 0.012280361726880074, + 0.0096468236297369, + 0.018508581444621086, + 0.011771395802497864, + 0.013904690742492676, + 0.00992748886346817, + 0.005870444234460592, + 0.007419003173708916, + 0.004607465583831072, + 0.006884223781526089, + 0.011931386776268482, + 0.008664068765938282, + 0.007674988359212875, + 0.004717640578746796, + 0.009846233762800694, + 0.0056500742211937904, + 0.006621828768402338, + 0.005234463606029749, + 0.0061590238474309444, + 0.005685893353074789, + 0.007415760774165392, + 0.006557977758347988, + 0.015878941863775253, + 0.013375995680689812, + 0.007827537134289742, + 0.01179707981646061, + 0.014408978633582592, + 0.012923184782266617, + 0.015061083249747753, + 0.006379628088325262, + 0.029674101620912552, + 0.01750093698501587, + 0.015371608547866344, + 0.014749468304216862, + 0.01632663980126381, + 0.019664455205202103, + 0.015462854877114296, + 0.012322310358285904, + 0.017795389518141747, + 0.019199809059500694, + 0.017910471186041832, + 0.019237570464611053, + 0.02333870716392994, + 0.021593661978840828, + 0.01978580839931965, + 0.011189655400812626, + 0.027565566822886467, + 0.03015626035630703, + 0.0222149845212698, + 0.03054860793054104, + 0.030566638335585594, + 0.0353352390229702, + 0.02753211371600628, + 0.01696518063545227, + 0.008648384362459183, + 0.007434845436364412, + 0.01158098317682743, + 0.00772424042224884, + 0.008026672527194023, + 0.006136185955256224, + 0.00947263091802597, + 0.008901108987629414, + 0.007078894879668951, + 0.005222178529947996, + 0.0059555829502642155, + 0.00653063552454114, + 0.006287838332355022, + 0.0068464120849967, + 0.005094004329293966, + 0.005354373715817928, + 0.008529528975486755, + 0.006578063126653433, + 0.006633064243942499, + 0.007330234162509441, + 0.006941129919141531, + 0.007873467169702053, + 0.006205337587743998, + 0.00942433811724186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.08919037878513336 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3872860372066498 + ], + "opposite_logprob_tokens": [ + -1.1372860670089722 + ], + "logprob_reward": -0.3872860372066498, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3872860372066498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2292511761188507, + 0.08420083671808243, + 0.051619868725538254, + 0.06506519019603729, + 0.14474505186080933, + 0.07570420950651169, + 0.05140922963619232, + 0.060960665345191956, + 0.059189245104789734, + 0.17785444855690002 + ], + "credit_map": [ + 0.0279851071536541, + 0.02921486087143421, + 0.03406582400202751, + 0.02474251203238964, + 0.032503508031368256, + 0.020839499309659004, + 0.026957659050822258, + 0.03294222056865692, + 0.011211390607059002, + 0.01261894591152668, + 0.010647740215063095, + 0.009381826035678387, + 0.010571793653070927, + 0.009340612217783928, + 0.009619268588721752, + 0.010809273459017277, + 0.011664455756545067, + 0.00595293752849102, + 0.005685667973011732, + 0.005235630087554455, + 0.005702710244804621, + 0.005452046170830727, + 0.0047109369188547134, + 0.007215491030365229, + 0.00937145296484232, + 0.00964498519897461, + 0.005749027244746685, + 0.008516558445990086, + 0.009790932759642601, + 0.0072816042229533195, + 0.010282134637236595, + 0.004428497515618801, + 0.015396765433251858, + 0.017329709604382515, + 0.016009602695703506, + 0.022008473053574562, + 0.028855526819825172, + 0.015205765143036842, + 0.018077855929732323, + 0.01186137180775404, + 0.012727338820695877, + 0.009108537808060646, + 0.007947398349642754, + 0.009692653082311153, + 0.011078700423240662, + 0.00898660533130169, + 0.0077468836680054665, + 0.00841609388589859, + 0.006928070914000273, + 0.0053567783907055855, + 0.006618168670684099, + 0.00792099628597498, + 0.008255885913968086, + 0.006293661426752806, + 0.005485399626195431, + 0.004550270736217499, + 0.008189979009330273, + 0.007223919034004211, + 0.005176587961614132, + 0.006321069318801165, + 0.0077258446253836155, + 0.008806890808045864, + 0.009064809419214725, + 0.008451568894088268, + 0.0073877242393791676, + 0.0072853416204452515, + 0.01061480026692152, + 0.008602400310337543, + 0.00670568086206913, + 0.007445396389812231, + 0.006003588903695345, + 0.005144314840435982, + 0.03191344812512398, + 0.02004694566130638, + 0.0323948934674263, + 0.01556449756026268, + 0.019214315339922905, + 0.020813651382923126, + 0.014268247410655022, + 0.023638470098376274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.3872860372066498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "super-slow-motion overhead shot, perfectly still, focuses on a massive, vintage Baroque-style crystal chandelier suspended in the center of the majestic opera house's dome. This magnificent fixture, composed of layered brass frames and hundreds of intricately carved crystal pendants, radiates a warm, dazzling golden light, casting intricate shadows on the empty red velvet seats below. Suddenly, with a powerful, silent tremor, the entire chandelier begins to shake violently, the crystal pendants clashing with each other, producing an increasingly rapid and crisp tinkling sound. As the tremor reaches its peak, the chandelier's main metal frame suddenly bends and breaks painfully under the unbearable stress, leading to a catastrophic structural collapse. Countless crystal fragments and glass beads exploded violently from the center outward, then poured down like a deadly yet beautiful flash rain. In slow motion, each irregular fragment slowly rolled and rotated in the air, its sharp cross-section capturing and madly refracting the last flicker of light before the chandelier went out, drawing short and sharp arcs of cold light against the gradually darkening background. The whole picture was full of sudden destructiveness and a heart-pounding tragic beauty.", + "seed": 11863, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_1435_seed_11863_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5858501006849111, + "vlm_logprob": -0.5858501006849111, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.5858501006849111, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021172692999243736 + ], + "opposite_logprob_tokens": [ + -4.771172523498535 + ], + "logprob_reward": -0.021172692999243736, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.021172692999243736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17051075398921967, + 0.32235589623451233, + 0.057012852281332016, + 0.038111016154289246, + 0.04820027947425842, + 0.10391288250684738, + 0.08056017756462097, + 0.029263028874993324, + 0.0723157450556755, + 0.0777573212981224 + ], + "credit_map": [ + 0.0198162030428648, + 0.01441932562738657, + 0.01214083842933178, + 0.016472840681672096, + 0.0362611748278141, + 0.019245587289333344, + 0.03587868809700012, + 0.016276095062494278, + 0.022428609430789948, + 0.034326810389757156, + 0.016729524359107018, + 0.03303101658821106, + 0.03129802271723747, + 0.07791659981012344, + 0.08712546527385712, + 0.019499853253364563, + 0.006460533477365971, + 0.004314945545047522, + 0.014938648790121078, + 0.005975759122520685, + 0.006081875413656235, + 0.0055579692125320435, + 0.009146519005298615, + 0.004536601714789867, + 0.004863357171416283, + 0.004701623693108559, + 0.0037424657493829727, + 0.005188298411667347, + 0.003969169221818447, + 0.006795979104936123, + 0.005837984383106232, + 0.0030121381860226393, + 0.006085684988647699, + 0.006717551965266466, + 0.004833521321415901, + 0.004257449880242348, + 0.0050553614273667336, + 0.00873855222016573, + 0.004915312398225069, + 0.007596845272928476, + 0.00846132542937994, + 0.019466349855065346, + 0.012530294246971607, + 0.011163895949721336, + 0.014971341006457806, + 0.012181670404970646, + 0.010839574970304966, + 0.014298433437943459, + 0.008191033266484737, + 0.010085714980959892, + 0.013915221206843853, + 0.009380107745528221, + 0.010903243906795979, + 0.008554080501198769, + 0.007494261488318443, + 0.012036511674523354, + 0.003269175998866558, + 0.0036991634406149387, + 0.0030918174888938665, + 0.0034623125102370977, + 0.0033378242515027523, + 0.0036282541695982218, + 0.004574912134557962, + 0.004199567716568708, + 0.007771623786538839, + 0.007262173108756542, + 0.01013251394033432, + 0.009012502618134022, + 0.006858892738819122, + 0.013473082333803177, + 0.008961906656622887, + 0.008843052200973034, + 0.012204962782561779, + 0.008739233016967773, + 0.012000465765595436, + 0.007659806404262781, + 0.007660032715648413, + 0.011107495985925198, + 0.009657664224505424, + 0.008727654814720154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.021172692999243736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005992185790091753 + ], + "opposite_logprob_tokens": [ + -5.130992412567139 + ], + "logprob_reward": -0.005992185790091753, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005992185790091753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13863925635814667, + 0.11246931552886963, + 0.07400863617658615, + 0.0751538947224617, + 0.0355437770485878, + 0.08191311359405518, + 0.12915170192718506, + 0.04673060402274132, + 0.254423052072525, + 0.05196666345000267 + ], + "credit_map": [ + 0.017595212906599045, + 0.01764061115682125, + 0.014017629437148571, + 0.02231840416789055, + 0.013596873730421066, + 0.024599207565188408, + 0.015293988399207592, + 0.013577340170741081, + 0.008723214268684387, + 0.007416918408125639, + 0.006934423930943012, + 0.013353615067899227, + 0.0070870015770196915, + 0.0294937901198864, + 0.03243681788444519, + 0.007023535203188658, + 0.007573325652629137, + 0.009261604398488998, + 0.00999918021261692, + 0.008834874257445335, + 0.011126862838864326, + 0.011313335038721561, + 0.008623812347650528, + 0.007275642827153206, + 0.0065191397443413734, + 0.01900293678045273, + 0.006192032713443041, + 0.008165236562490463, + 0.011959218420088291, + 0.009103831835091114, + 0.007404486648738384, + 0.006807012017816305, + 0.0032857421319931746, + 0.006175902206450701, + 0.006100893020629883, + 0.003885924117639661, + 0.0036967890337109566, + 0.005077695939689875, + 0.0036157602444291115, + 0.00370506732724607, + 0.008006257936358452, + 0.01045742817223072, + 0.0058073499239981174, + 0.005742271896451712, + 0.018367040902376175, + 0.019421124830842018, + 0.005435911472886801, + 0.008675727061927319, + 0.015765760093927383, + 0.019519567489624023, + 0.016871625557541847, + 0.014664731919765472, + 0.016726695001125336, + 0.017712604254484177, + 0.009279555641114712, + 0.018611159175634384, + 0.005469715688377619, + 0.0055182636715471745, + 0.00447737704962492, + 0.0051819137297570705, + 0.005127253942191601, + 0.0061081391759216785, + 0.007322800811380148, + 0.007525140885263681, + 0.023584116250276566, + 0.03710131719708443, + 0.03668251261115074, + 0.03469103202223778, + 0.027952121570706367, + 0.02694096975028515, + 0.04486221820116043, + 0.022608766332268715, + 0.008288462646305561, + 0.007494544610381126, + 0.007835146971046925, + 0.006528774276375771, + 0.005057525355368853, + 0.006801136769354343, + 0.0044882046058773994, + 0.005472866352647543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.005992185790091753 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.4640207290649414 + ], + "opposite_logprob_tokens": [ + -0.08902081102132797 + ], + "logprob_reward": -2.4640207290649414, + "yes_no_margin_reward": -2.375, + "token_credit_reward": -2.4640207290649414, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.043626703321933746, + 0.17642749845981598, + 0.019934969022870064, + 0.032240916043519974, + 0.061076246201992035, + 0.035092249512672424, + 0.0744762271642685, + 0.046928536146879196, + 0.44103923439979553, + 0.0691574364900589 + ], + "credit_map": [ + 0.003995291888713837, + 0.003983804956078529, + 0.004533147439360619, + 0.0060422723181545734, + 0.007569894660264254, + 0.005306586157530546, + 0.008011342026293278, + 0.004184362478554249, + 0.013561750762164593, + 0.010008553974330425, + 0.007750422693789005, + 0.01870826818048954, + 0.012245013378560543, + 0.049840401858091354, + 0.05342242121696472, + 0.010890663601458073, + 0.001832142355851829, + 0.0021566832438111305, + 0.002909064758569002, + 0.003402607748284936, + 0.0021635815501213074, + 0.0026312677655369043, + 0.002917887642979622, + 0.001921735005453229, + 0.0029789546970278025, + 0.005939059425145388, + 0.003202308202162385, + 0.0033654742874205112, + 0.003695670049637556, + 0.005341933574527502, + 0.0037843671161681414, + 0.0039331489242613316, + 0.006949628237634897, + 0.010325438342988491, + 0.0062725660391151905, + 0.00605440279468894, + 0.007004816550761461, + 0.007587239611893892, + 0.00650391960516572, + 0.010378236882388592, + 0.0031028662342578173, + 0.006152941379696131, + 0.002641125116497278, + 0.0045739952474832535, + 0.004237876273691654, + 0.004391499795019627, + 0.004226746968924999, + 0.00576519500464201, + 0.010373838245868683, + 0.00793098472058773, + 0.012965008616447449, + 0.008842089213430882, + 0.0075012813322246075, + 0.010051056742668152, + 0.006729208864271641, + 0.010082758963108063, + 0.005995407234877348, + 0.00719862524420023, + 0.004278284031897783, + 0.00640144944190979, + 0.005889467895030975, + 0.005076996516436338, + 0.006029628682881594, + 0.0060586752369999886, + 0.034741733223199844, + 0.058722592890262604, + 0.04720216616988182, + 0.06105131655931473, + 0.0490545891225338, + 0.06426748633384705, + 0.07228133082389832, + 0.05371801555156708, + 0.011954030022025108, + 0.009573951363563538, + 0.008646707981824875, + 0.007895030081272125, + 0.007166462950408459, + 0.009347736835479736, + 0.0073114726692438126, + 0.007262038998305798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -2.4640207290649414 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08921655267477036 + ], + "opposite_logprob_tokens": [ + -2.464216470718384 + ], + "logprob_reward": -0.08921655267477036, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08921655267477036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09523797035217285, + 0.038646500557661057, + 0.023500382900238037, + 0.05992206558585167, + 0.0918324664235115, + 0.07306843250989914, + 0.07878056913614273, + 0.05278466269373894, + 0.3845064043998718, + 0.10172052681446075 + ], + "credit_map": [ + 0.010456307791173458, + 0.007907546125352383, + 0.006149308290332556, + 0.007926436141133308, + 0.0207983311265707, + 0.00845345575362444, + 0.02498670108616352, + 0.008559889160096645, + 0.003631538012996316, + 0.006737635470926762, + 0.002814328297972679, + 0.00473112752661109, + 0.004978101700544357, + 0.007022255565971136, + 0.005955386906862259, + 0.0027761294040828943, + 0.0024470144417136908, + 0.003337829140946269, + 0.003169087925925851, + 0.0034478604793548584, + 0.0032221744768321514, + 0.002456382615491748, + 0.003626454621553421, + 0.0017935804789885879, + 0.006887971889227629, + 0.010914996266365051, + 0.007400129921734333, + 0.007197362836450338, + 0.006410290952771902, + 0.007235424127429724, + 0.0066577536053955555, + 0.007218140177428722, + 0.011740815825760365, + 0.015399133786559105, + 0.008513207547366619, + 0.008787459693849087, + 0.009971519000828266, + 0.010830960236489773, + 0.009725190699100494, + 0.01686418056488037, + 0.005920498166233301, + 0.011972678825259209, + 0.005140206776559353, + 0.006425766274333, + 0.01649641804397106, + 0.012511739507317543, + 0.007097579538822174, + 0.007503554690629244, + 0.011268648318946362, + 0.0097226956859231, + 0.012828131206333637, + 0.009231735952198505, + 0.008818610571324825, + 0.010000224225223064, + 0.006638572085648775, + 0.010271954350173473, + 0.005984599702060223, + 0.0064400313422083855, + 0.006419312674552202, + 0.005505929235368967, + 0.00478345574811101, + 0.008989148773252964, + 0.006242603994905949, + 0.008419584482908249, + 0.02939884178340435, + 0.05294371396303177, + 0.0396554172039032, + 0.053717974573373795, + 0.042414333671331406, + 0.055276013910770416, + 0.06442005932331085, + 0.04668007045984268, + 0.018174856901168823, + 0.01266456488519907, + 0.014247080311179161, + 0.011218104511499405, + 0.01066180132329464, + 0.014705641195178032, + 0.009303121827542782, + 0.010745365172624588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.08921655267477036 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3488483428955078 + ], + "opposite_logprob_tokens": [ + -1.2238483428955078 + ], + "logprob_reward": -0.3488483428955078, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3488483428955078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11969941854476929, + 0.40040311217308044, + 0.023762885481119156, + 0.053020332008600235, + 0.11462879925966263, + 0.09298011660575867, + 0.03195088729262352, + 0.05492876470088959, + 0.08307003974914551, + 0.025555681437253952 + ], + "credit_map": [ + 0.010158434510231018, + 0.009357579983770847, + 0.00935711245983839, + 0.01030308473855257, + 0.020411334931850433, + 0.014368019066751003, + 0.03452189266681671, + 0.011221960186958313, + 0.03226901963353157, + 0.031584832817316055, + 0.02197430655360222, + 0.039632320404052734, + 0.0304645337164402, + 0.11121997982263565, + 0.10845815390348434, + 0.024799922481179237, + 0.001948573044501245, + 0.0037319129332900047, + 0.0030696571338921785, + 0.003983609843999147, + 0.0030187591910362244, + 0.0025773870293051004, + 0.0027304962277412415, + 0.002702490659430623, + 0.004736163187772036, + 0.01398500893265009, + 0.005125087685883045, + 0.005537090823054314, + 0.007808174937963486, + 0.006006921175867319, + 0.005998499225825071, + 0.003823388135060668, + 0.01453919056802988, + 0.017892315983772278, + 0.012665674090385437, + 0.010596059262752533, + 0.011742094531655312, + 0.014333963394165039, + 0.010778430849313736, + 0.02208106778562069, + 0.008099540136754513, + 0.01646447740495205, + 0.006530660204589367, + 0.007508392445743084, + 0.02124006301164627, + 0.016952194273471832, + 0.0087932413443923, + 0.007391550578176975, + 0.0030762723181396723, + 0.004225895740091801, + 0.004658878315240145, + 0.005568738095462322, + 0.0036800308153033257, + 0.0034424045588821173, + 0.0036588944494724274, + 0.0036397732328623533, + 0.00651921983808279, + 0.006576041225343943, + 0.005303492303937674, + 0.006937212310731411, + 0.006041920743882656, + 0.006770459935069084, + 0.008302323520183563, + 0.008478092029690742, + 0.006284733302891254, + 0.014029634185135365, + 0.008299783803522587, + 0.012901348993182182, + 0.009661365300416946, + 0.007065615151077509, + 0.01678810827434063, + 0.008039452135562897, + 0.0031877770088613033, + 0.004360468592494726, + 0.0036720663774758577, + 0.002743449527770281, + 0.0026458522770553827, + 0.002928665140643716, + 0.0028764260932803154, + 0.0031409766525030136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.3488483428955078 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "super-slow-motion overhead shot, perfectly still, focuses on a massive, vintage Baroque-style crystal chandelier suspended in the center of the majestic opera house's dome. This magnificent fixture, composed of layered brass frames and hundreds of intricately carved crystal pendants, radiates a warm, dazzling golden light, casting intricate shadows on the empty red velvet seats below. Suddenly, with a powerful, silent tremor, the entire chandelier begins to shake violently, the crystal pendants clashing with each other, producing an increasingly rapid and crisp tinkling sound. As the tremor reaches its peak, the chandelier's main metal frame suddenly bends and breaks painfully under the unbearable stress, leading to a catastrophic structural collapse. Countless crystal fragments and glass beads exploded violently from the center outward, then poured down like a deadly yet beautiful flash rain. In slow motion, each irregular fragment slowly rolled and rotated in the air, its sharp cross-section capturing and madly refracting the last flicker of light before the chandelier went out, drawing short and sharp arcs of cold light against the gradually darkening background. The whole picture was full of sudden destructiveness and a heart-pounding tragic beauty.", + "seed": 11864, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_1435_seed_11864_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.540522447321564, + "vlm_logprob": -0.540522447321564, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.540522447321564, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017801983281970024 + ], + "opposite_logprob_tokens": [ + -5.017801761627197 + ], + "logprob_reward": -0.017801983281970024, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.017801983281970024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4554523527622223, + 0.015095951966941357, + 0.016106590628623962, + 0.05295274779200554, + 0.021558981388807297, + 0.09522823989391327, + 0.16031324863433838, + 0.13025899231433868, + 0.04551059380173683, + 0.007522377651184797 + ], + "credit_map": [ + 0.02407483197748661, + 0.0277115311473608, + 0.1447088122367859, + 0.016645099967718124, + 0.02162184752523899, + 0.17547528445720673, + 0.013394836336374283, + 0.03182007744908333, + 0.0013574117328971624, + 0.002274214755743742, + 0.0013412279076874256, + 0.0013722479343414307, + 0.0016291610663756728, + 0.003274662420153618, + 0.0015632391441613436, + 0.00228378688916564, + 0.0020151534117758274, + 0.0016424605855718255, + 0.0024681943468749523, + 0.0014744922518730164, + 0.001900830422528088, + 0.00228738971054554, + 0.001395871862769127, + 0.002922199899330735, + 0.0053283353336155415, + 0.004752123262733221, + 0.005687551572918892, + 0.003122579539194703, + 0.008457699790596962, + 0.013516715727746487, + 0.0035663126036524773, + 0.008521427400410175, + 0.0030114329420030117, + 0.0028177229687571526, + 0.0024705652613192797, + 0.0018274316098541021, + 0.0027141578029841185, + 0.003024749457836151, + 0.003246454754844308, + 0.002446467522531748, + 0.005268062464892864, + 0.010142354294657707, + 0.021249398589134216, + 0.005473514087498188, + 0.009983754716813564, + 0.010566938668489456, + 0.005435411352664232, + 0.027108801528811455, + 0.009751381352543831, + 0.00882395263761282, + 0.057844378054142, + 0.006796581670641899, + 0.010827915742993355, + 0.045163221657276154, + 0.0033291575964540243, + 0.017776653170585632, + 0.017394378781318665, + 0.018093280494213104, + 0.015472478233277798, + 0.008611194789409637, + 0.025896796956658363, + 0.018553148955106735, + 0.010907236486673355, + 0.015330483205616474, + 0.006489609833806753, + 0.00561037752777338, + 0.007063062395900488, + 0.004475768655538559, + 0.007738481741398573, + 0.004757240414619446, + 0.003467201953753829, + 0.005908850580453873, + 0.0007587378495372832, + 0.001240671263076365, + 0.0011465353891253471, + 0.0007353695691563189, + 0.0013816060964018106, + 0.0008067252347245812, + 0.0006270438898354769, + 0.0008256884175352752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.017801983281970024 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004130760673433542 + ], + "opposite_logprob_tokens": [ + -5.504130840301514 + ], + "logprob_reward": -0.004130760673433542, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004130760673433542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4610317647457123, + 0.035062920302152634, + 0.01403027679771185, + 0.03472280502319336, + 0.03819712996482849, + 0.042772211134433746, + 0.18948902189731598, + 0.12848010659217834, + 0.04162955656647682, + 0.014584241434931755 + ], + "credit_map": [ + 0.02293221838772297, + 0.031358446925878525, + 0.14571884274482727, + 0.017269879579544067, + 0.02355395443737507, + 0.17529715597629547, + 0.013523281551897526, + 0.03137795627117157, + 0.0032385506201535463, + 0.004162913653999567, + 0.0034478234592825174, + 0.004634634125977755, + 0.0037467768415808678, + 0.0063475496135652065, + 0.005258488468825817, + 0.004226176533848047, + 0.002151109278202057, + 0.0016758247511461377, + 0.002220096532255411, + 0.0015265726251527667, + 0.0013363537145778537, + 0.0013080249773338437, + 0.0016671974444761872, + 0.0021450959611684084, + 0.004431226756423712, + 0.003684003371745348, + 0.0028938285540789366, + 0.002364308573305607, + 0.0047756703570485115, + 0.008300310000777245, + 0.0024775839410722256, + 0.0057958695106208324, + 0.002664300613105297, + 0.004622242879122496, + 0.004524131305515766, + 0.0025996402837336063, + 0.007967116311192513, + 0.004232484381645918, + 0.003098112531006336, + 0.008489096537232399, + 0.002905856352299452, + 0.005779007915407419, + 0.009649382904171944, + 0.0027335253544151783, + 0.005034984089434147, + 0.003662603907287121, + 0.002950763562694192, + 0.010056084021925926, + 0.010975640267133713, + 0.02769239991903305, + 0.04231448099017143, + 0.006063955370336771, + 0.04687745124101639, + 0.016218151897192, + 0.0053316932171583176, + 0.0340152271091938, + 0.015746718272566795, + 0.018286483362317085, + 0.014778558164834976, + 0.008763881400227547, + 0.027033543214201927, + 0.01904679834842682, + 0.010233848355710506, + 0.014590267091989517, + 0.006544942501932383, + 0.0049400730058550835, + 0.006876900792121887, + 0.004248146433383226, + 0.0061300587840378284, + 0.004017496015876532, + 0.0038059670478105545, + 0.00506596639752388, + 0.0014849224826321006, + 0.0024510386865586042, + 0.0020151410717517138, + 0.0016934557352215052, + 0.0016911489656195045, + 0.0021319910883903503, + 0.0012034456012770534, + 0.0019130967557430267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.004130760673433542 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.1270415782928467 + ], + "opposite_logprob_tokens": [ + -0.1270415484905243 + ], + "logprob_reward": -2.1270415782928467, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.1270415782928467, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.33452773094177246, + 0.02377748303115368, + 0.017367754131555557, + 0.07595933228731155, + 0.09614487737417221, + 0.11826115101575851, + 0.14377275109291077, + 0.15637342631816864, + 0.018844975158572197, + 0.014970485121011734 + ], + "credit_map": [ + 0.01948225311934948, + 0.014198177494108677, + 0.11729367077350616, + 0.01463998481631279, + 0.015093616209924221, + 0.11905872821807861, + 0.01046676468104124, + 0.024294573813676834, + 0.0020471238531172276, + 0.0025812142994254827, + 0.0018903209129348397, + 0.0018454758683219552, + 0.002776410663500428, + 0.0058953831903636456, + 0.002708075800910592, + 0.004033476579934359, + 0.0018836968811228871, + 0.0023963325656950474, + 0.0024177555460482836, + 0.0019982096273452044, + 0.001623177551664412, + 0.0022664328571408987, + 0.0018460178980603814, + 0.002936131088063121, + 0.008038761094212532, + 0.007142509799450636, + 0.006843109615147114, + 0.0041022635996341705, + 0.012032327242195606, + 0.020218906924128532, + 0.0054306816309690475, + 0.012150775641202927, + 0.00640465272590518, + 0.013761916197836399, + 0.011202076449990273, + 0.006092877592891455, + 0.01985860988497734, + 0.007828288711607456, + 0.007432969752699137, + 0.02356349676847458, + 0.006781804375350475, + 0.0081429872661829, + 0.028906814754009247, + 0.006623756140470505, + 0.012283803895115852, + 0.01192000973969698, + 0.006259807851165533, + 0.03734215721487999, + 0.00687422463670373, + 0.019264785572886467, + 0.007867030799388885, + 0.004288843367248774, + 0.03480222448706627, + 0.04282882437109947, + 0.004549473989754915, + 0.023297343403100967, + 0.019227957352995872, + 0.026263900101184845, + 0.01578409969806671, + 0.011976064182817936, + 0.03206991031765938, + 0.020815979689359665, + 0.012744009494781494, + 0.017491498962044716, + 0.002226912183687091, + 0.0027726583648473024, + 0.002693758113309741, + 0.0015888722846284509, + 0.003201513085514307, + 0.002519087865948677, + 0.0016567467246204615, + 0.002185426652431488, + 0.0016171210445463657, + 0.002175608417019248, + 0.0019427090883255005, + 0.001822256832383573, + 0.0018998326268047094, + 0.0019234400242567062, + 0.001538751763291657, + 0.0020507650915533304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -2.1270415782928467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07912855595350266 + ], + "opposite_logprob_tokens": [ + -2.5791285037994385 + ], + "logprob_reward": -0.07912855595350266, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07912855595350266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4537959098815918, + 0.022157369181513786, + 0.011377763003110886, + 0.05315827578306198, + 0.0766415074467659, + 0.0549161396920681, + 0.20549003779888153, + 0.06034662947058678, + 0.040036868304014206, + 0.02207946963608265 + ], + "credit_map": [ + 0.025211049243807793, + 0.025244716554880142, + 0.15026438236236572, + 0.01772020198404789, + 0.021708227694034576, + 0.1695866584777832, + 0.013285745866596699, + 0.030774928629398346, + 0.0026124820578843355, + 0.0029310230165719986, + 0.0021031545475125313, + 0.0026464867405593395, + 0.0023798360489308834, + 0.0039092605002224445, + 0.0027092737145721912, + 0.0028658544179052114, + 0.001337782945483923, + 0.001335008768364787, + 0.001414275960996747, + 0.001059647765941918, + 0.0014727420639246702, + 0.0015978042501956224, + 0.0013112492160871625, + 0.0018492521485313773, + 0.005384800024330616, + 0.006228983402252197, + 0.0053564319387078285, + 0.003317291149869561, + 0.008472592569887638, + 0.012345804832875729, + 0.003818585304543376, + 0.00823378749191761, + 0.0053426967933773994, + 0.011346164159476757, + 0.00997846107929945, + 0.006005310919135809, + 0.012332665733993053, + 0.006348338443785906, + 0.0070664044469594955, + 0.01822146773338318, + 0.007254498545080423, + 0.006171697285026312, + 0.009466676041483879, + 0.00422716373577714, + 0.005792451091110706, + 0.005284214857965708, + 0.004001516383141279, + 0.01271792035549879, + 0.00994132924824953, + 0.019091608002781868, + 0.029588529840111732, + 0.006593761034309864, + 0.036882366985082626, + 0.07124049216508865, + 0.005692838225513697, + 0.026459114626049995, + 0.007484104949980974, + 0.010396350175142288, + 0.0068817525170743465, + 0.0050589460879564285, + 0.009765620343387127, + 0.00721314549446106, + 0.005978910252451897, + 0.0075677973218262196, + 0.006292516831308603, + 0.004657506011426449, + 0.0068111298605799675, + 0.0038908387068659067, + 0.006097754463553429, + 0.004061727784574032, + 0.0031685936264693737, + 0.005056801717728376, + 0.0027863343711942434, + 0.003211633302271366, + 0.0028145217802375555, + 0.0021116849966347218, + 0.003518275683745742, + 0.002497768262401223, + 0.001775840064510703, + 0.003363411407917738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.07912855595350266 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4745093584060669 + ], + "opposite_logprob_tokens": [ + -0.9745093584060669 + ], + "logprob_reward": -0.4745093584060669, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4745093584060669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3986697196960449, + 0.03726737201213837, + 0.016232427209615707, + 0.07646725326776505, + 0.09846928715705872, + 0.03910941258072853, + 0.23996193706989288, + 0.04473177343606949, + 0.03125710412859917, + 0.017833730205893517 + ], + "credit_map": [ + 0.02041260525584221, + 0.014983047731220722, + 0.1296895444393158, + 0.015647830441594124, + 0.018666882067918777, + 0.1601058542728424, + 0.010851932689547539, + 0.028312014415860176, + 0.003893779357895255, + 0.004502038937062025, + 0.004666806664317846, + 0.002938046818599105, + 0.003956626169383526, + 0.008670474402606487, + 0.00353625719435513, + 0.00510333850979805, + 0.0014937376836314797, + 0.0029082319233566523, + 0.002211470389738679, + 0.0012841575080528855, + 0.0020686304196715355, + 0.0020632727537304163, + 0.001670093392021954, + 0.002532829763367772, + 0.00800139270722866, + 0.0069320122711360455, + 0.008546669036149979, + 0.0056517762131989, + 0.009663411416113377, + 0.01830124668776989, + 0.006604294292628765, + 0.012766445055603981, + 0.008845537900924683, + 0.021015621721744537, + 0.010623586364090443, + 0.009599735029041767, + 0.01563509926199913, + 0.008835882879793644, + 0.01143921259790659, + 0.012474605813622475, + 0.0031824687030166388, + 0.004655945580452681, + 0.009152015671133995, + 0.00238850899040699, + 0.003373346757143736, + 0.004102789796888828, + 0.0028900092002004385, + 0.009364325553178787, + 0.012393929995596409, + 0.021640393882989883, + 0.03479529172182083, + 0.006229684222489595, + 0.04377055540680885, + 0.08685561269521713, + 0.0067399595864117146, + 0.027536502107977867, + 0.005360949318856001, + 0.005809336435049772, + 0.005648582708090544, + 0.002994578331708908, + 0.008480384014546871, + 0.007347999140620232, + 0.003302070777863264, + 0.005787866190075874, + 0.005121380090713501, + 0.0034453938715159893, + 0.005562055855989456, + 0.0026556667871773243, + 0.003680572845041752, + 0.002942855004221201, + 0.0031461783219128847, + 0.0047030020505189896, + 0.00190097582526505, + 0.0024267311673611403, + 0.0028169883880764246, + 0.002424759091809392, + 0.0020283800549805164, + 0.002259632106870413, + 0.001605070661753416, + 0.0023711910471320152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.4745093584060669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "super-slow-motion overhead shot, perfectly still, focuses on a massive, vintage Baroque-style crystal chandelier suspended in the center of the majestic opera house's dome. This magnificent fixture, composed of layered brass frames and hundreds of intricately carved crystal pendants, radiates a warm, dazzling golden light, casting intricate shadows on the empty red velvet seats below. Suddenly, with a powerful, silent tremor, the entire chandelier begins to shake violently, the crystal pendants clashing with each other, producing an increasingly rapid and crisp tinkling sound. As the tremor reaches its peak, the chandelier's main metal frame suddenly bends and breaks painfully under the unbearable stress, leading to a catastrophic structural collapse. Countless crystal fragments and glass beads exploded violently from the center outward, then poured down like a deadly yet beautiful flash rain. In slow motion, each irregular fragment slowly rolled and rotated in the air, its sharp cross-section capturing and madly refracting the last flicker of light before the chandelier went out, drawing short and sharp arcs of cold light against the gradually darkening background. The whole picture was full of sudden destructiveness and a heart-pounding tragic beauty.", + "seed": 11865, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_7/index_1435_seed_11865_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1525509651750326, + "vlm_logprob": -0.1525509651750326, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.1525509651750326, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015031756833195686 + ], + "opposite_logprob_tokens": [ + -5.265031814575195 + ], + "logprob_reward": -0.015031756833195686, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.015031756833195686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07749250531196594, + 0.04488217085599899, + 0.19181011617183685, + 0.0414264015853405, + 0.14451923966407776, + 0.06322918087244034, + 0.10981956869363785, + 0.10497146099805832, + 0.06573648750782013, + 0.1561128795146942 + ], + "credit_map": [ + 0.011969834566116333, + 0.010821937583386898, + 0.0125096645206213, + 0.00719634210690856, + 0.0087188221514225, + 0.008936872705817223, + 0.006753586698323488, + 0.010585452429950237, + 0.007897073403000832, + 0.005535770673304796, + 0.005410583224147558, + 0.0038935064803808928, + 0.005114677827805281, + 0.004818506073206663, + 0.003349190577864647, + 0.008862870745360851, + 0.023203160613775253, + 0.018512651324272156, + 0.01721273921430111, + 0.0099704684689641, + 0.02936907857656479, + 0.06801628321409225, + 0.009210562333464622, + 0.016315193846821785, + 0.002841302892193198, + 0.01525355875492096, + 0.0018106993520632386, + 0.0029421921353787184, + 0.011276026256382465, + 0.0018606747034937143, + 0.0033272949513047934, + 0.002114657312631607, + 0.01196100004017353, + 0.011692332103848457, + 0.01316206343472004, + 0.015322870574891567, + 0.014809933491051197, + 0.01901138760149479, + 0.02925809845328331, + 0.02930157072842121, + 0.008400758728384972, + 0.009655880741775036, + 0.006356977857649326, + 0.010799474082887173, + 0.007230062037706375, + 0.007357122376561165, + 0.007165152579545975, + 0.00626375712454319, + 0.016524480655789375, + 0.015475669875741005, + 0.01054913830012083, + 0.012497683055698872, + 0.010669195093214512, + 0.011828792281448841, + 0.018671996891498566, + 0.013602619990706444, + 0.012243729084730148, + 0.009315251372754574, + 0.013618731871247292, + 0.01412796601653099, + 0.011214003898203373, + 0.010627951472997665, + 0.01709161326289177, + 0.0167322289198637, + 0.004787604324519634, + 0.007283997256308794, + 0.008235990069806576, + 0.011563577689230442, + 0.006688677705824375, + 0.006359932012856007, + 0.013443868607282639, + 0.0073728482238948345, + 0.04802059009671211, + 0.010756039060652256, + 0.018471991643309593, + 0.009803314693272114, + 0.008767840452492237, + 0.011231336742639542, + 0.03856847062706947, + 0.010493302717804909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.015031756833195686 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006759041920304298 + ], + "opposite_logprob_tokens": [ + -5.006759166717529 + ], + "logprob_reward": -0.006759041920304298, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006759041920304298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08313841372728348, + 0.03358389437198639, + 0.09967304021120071, + 0.13629387319087982, + 0.09742601215839386, + 0.1141577661037445, + 0.14848490059375763, + 0.0793209820985794, + 0.08197397738695145, + 0.12594714760780334 + ], + "credit_map": [ + 0.00874499510973692, + 0.010444295592606068, + 0.00887164007872343, + 0.007449670694768429, + 0.01278039999306202, + 0.0077098822221159935, + 0.009869244880974293, + 0.017268281430006027, + 0.005681377835571766, + 0.005111362785100937, + 0.004370720125734806, + 0.0039518363773822784, + 0.0036628427915275097, + 0.004102194681763649, + 0.003735610283911228, + 0.002967948792502284, + 0.011978588066995144, + 0.010927367024123669, + 0.00986296683549881, + 0.006165944505482912, + 0.01424152310937643, + 0.03248116746544838, + 0.005674024112522602, + 0.008341464214026928, + 0.006640072446316481, + 0.047433264553546906, + 0.008226989768445492, + 0.00487670348957181, + 0.04301198944449425, + 0.008943628519773483, + 0.0067458306439220905, + 0.010415399447083473, + 0.009206397458910942, + 0.010798615403473377, + 0.009537706151604652, + 0.012600313872098923, + 0.011986235156655312, + 0.01316213421523571, + 0.012498593889176846, + 0.017636017873883247, + 0.01243543066084385, + 0.01926536113023758, + 0.012747997418045998, + 0.019039716571569443, + 0.01653856411576271, + 0.01167520321905613, + 0.0132892532274127, + 0.009166245348751545, + 0.027409719303250313, + 0.024437064304947853, + 0.011720177717506886, + 0.017695138230919838, + 0.01665155030786991, + 0.017374718561768532, + 0.022361161187291145, + 0.01083536259829998, + 0.009393662214279175, + 0.009503578767180443, + 0.00938076339662075, + 0.009512309916317463, + 0.008791882544755936, + 0.0076112160459160805, + 0.011454726569354534, + 0.013672851026058197, + 0.006388217676430941, + 0.009900947101414204, + 0.008675348944962025, + 0.024145573377609253, + 0.007603907026350498, + 0.00507302675396204, + 0.012034770101308823, + 0.008152193389832973, + 0.011181619018316269, + 0.0075002978555858135, + 0.009339870885014534, + 0.01205350924283266, + 0.008523674681782722, + 0.026541773229837418, + 0.039574623107910156, + 0.01123177632689476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.006759041920304298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233055353164673 + ], + "opposite_logprob_tokens": [ + -0.8983055353164673 + ], + "logprob_reward": -0.5233055353164673, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233055353164673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09238173067569733, + 0.023220760747790337, + 0.05715430900454521, + 0.15824981033802032, + 0.08806067705154419, + 0.22172215580940247, + 0.12359369546175003, + 0.03594225272536278, + 0.07844069600105286, + 0.12123387306928635 + ], + "credit_map": [ + 0.011055490002036095, + 0.012041511945426464, + 0.009465877898037434, + 0.004971057642251253, + 0.01715232990682125, + 0.006815474480390549, + 0.005340585019439459, + 0.025539405643939972, + 0.0023848672863096, + 0.00428665429353714, + 0.002203466137871146, + 0.002728428691625595, + 0.004418367985635996, + 0.0022145125549286604, + 0.0021431033965200186, + 0.0028413617983460426, + 0.008110318332910538, + 0.0055334619246423244, + 0.005453342106193304, + 0.0022256639786064625, + 0.008274969644844532, + 0.01935284025967121, + 0.003067633369937539, + 0.005136080551892519, + 0.007192573510110378, + 0.037439703941345215, + 0.015045400708913803, + 0.006271170452237129, + 0.04277358204126358, + 0.00790537241846323, + 0.008937872014939785, + 0.03268413618206978, + 0.009154753759503365, + 0.009612016379833221, + 0.009101975709199905, + 0.012119222432374954, + 0.011671771295368671, + 0.012472887523472309, + 0.011520305648446083, + 0.012407741509377956, + 0.019421830773353577, + 0.036050908267498016, + 0.02415783703327179, + 0.042691394686698914, + 0.031205613166093826, + 0.020577536895871162, + 0.03034699708223343, + 0.017270023003220558, + 0.01592435874044895, + 0.015772853046655655, + 0.012414381839334965, + 0.014362181536853313, + 0.014484865590929985, + 0.01256695855408907, + 0.026755494996905327, + 0.011312611401081085, + 0.0038332371041178703, + 0.0035178312100470066, + 0.0036747066769748926, + 0.005631338804960251, + 0.003984643146395683, + 0.00372303812764585, + 0.006521881558001041, + 0.00505557656288147, + 0.004641201812773943, + 0.014077320694923401, + 0.009414903819561005, + 0.011280626989901066, + 0.009450296871364117, + 0.007081498857587576, + 0.015350913628935814, + 0.007143935654312372, + 0.007754498161375523, + 0.0071427421644330025, + 0.024040037766098976, + 0.008069539442658424, + 0.008134400472044945, + 0.016364509239792824, + 0.041845425963401794, + 0.007882720790803432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.5233055353164673 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607188768684864 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.01607188768684864, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607188768684864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05564302206039429, + 0.040498558431863785, + 0.05189232900738716, + 0.28624293208122253, + 0.08271586149930954, + 0.16126753389835358, + 0.16219255328178406, + 0.04668968915939331, + 0.05001072213053703, + 0.0628468319773674 + ], + "credit_map": [ + 0.007307776249945164, + 0.007545227184891701, + 0.006255940068513155, + 0.004531954415142536, + 0.009657616727054119, + 0.005289608146995306, + 0.004392451606690884, + 0.010662444867193699, + 0.00411399407312274, + 0.005640794523060322, + 0.0038056059274822474, + 0.0031767513137310743, + 0.007839593105018139, + 0.003044870449230075, + 0.003166220150887966, + 0.009710727259516716, + 0.005986945237964392, + 0.004783474840223789, + 0.004492845386266708, + 0.0022148583084344864, + 0.006280852481722832, + 0.018000461161136627, + 0.0023696888238191605, + 0.007763202302157879, + 0.012233817018568516, + 0.08195324242115021, + 0.0220661498606205, + 0.010893689468502998, + 0.08278080075979233, + 0.014625981450080872, + 0.016718439757823944, + 0.0449707992374897, + 0.008478649891912937, + 0.008912974037230015, + 0.008916917257010937, + 0.011732565239071846, + 0.011121351271867752, + 0.012064510025084019, + 0.011124133132398129, + 0.010364753194153309, + 0.01652827113866806, + 0.025560341775417328, + 0.018566010519862175, + 0.027848143130540848, + 0.02110045962035656, + 0.016967743635177612, + 0.021608475595712662, + 0.01308810617774725, + 0.028255382552742958, + 0.024294456467032433, + 0.015290348790585995, + 0.018751323223114014, + 0.018303273245692253, + 0.017128627747297287, + 0.026427745819091797, + 0.01374139729887247, + 0.005309398751705885, + 0.005198356695473194, + 0.004479197785258293, + 0.006640061270445585, + 0.005196715239435434, + 0.004703553859144449, + 0.006721298210322857, + 0.008441107347607613, + 0.0029438831843435764, + 0.00864570215344429, + 0.006681350991129875, + 0.006256416440010071, + 0.005667183082550764, + 0.00424389261752367, + 0.010808654129505157, + 0.004763638135045767, + 0.006662546191364527, + 0.003805021056905389, + 0.014864620752632618, + 0.0041702124290168285, + 0.004440322518348694, + 0.007140922360122204, + 0.01771865226328373, + 0.004044537898153067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.01607188768684864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an overhead view of a massive vintage Baroque-style crystal chandelier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chandelier composed of layered brass frames and hundreds of intricately carved crystal pendants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chandelier shake violently, with crystal pendants clashing and the main metal frame bending and breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless crystal fragments explode outward and pour down like rain in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in super-slow motion with a perfectly still camera?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158660411834717 + ], + "opposite_logprob_tokens": [ + -1.7015866041183472 + ], + "logprob_reward": -0.20158660411834717, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158660411834717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.144592747092247, + 0.06575103849172592, + 0.057638466358184814, + 0.14389930665493011, + 0.07443540543317795, + 0.15665476024150848, + 0.056754592806100845, + 0.07572489976882935, + 0.12956364452838898, + 0.09498507529497147 + ], + "credit_map": [ + 0.023144982755184174, + 0.017933286726474762, + 0.01681496575474739, + 0.013997551053762436, + 0.020922182127833366, + 0.01449463702738285, + 0.01255189348012209, + 0.024733256548643112, + 0.006878644227981567, + 0.008469493128359318, + 0.007015539798885584, + 0.0040606181137263775, + 0.01610342040657997, + 0.004613928962498903, + 0.0038378930184990168, + 0.014771508984267712, + 0.006616115570068359, + 0.007293212227523327, + 0.008025646209716797, + 0.006257233675569296, + 0.008230600506067276, + 0.010602666065096855, + 0.004902504850178957, + 0.005710487719625235, + 0.007003382313996553, + 0.04264495521783829, + 0.009373152628540993, + 0.008240625262260437, + 0.041305966675281525, + 0.007547830697149038, + 0.011844419874250889, + 0.015938982367515564, + 0.006007771007716656, + 0.005982707254588604, + 0.007388770580291748, + 0.00818457081913948, + 0.00680480245500803, + 0.00629763538017869, + 0.019924238324165344, + 0.013844906352460384, + 0.014665433205664158, + 0.025080598890781403, + 0.01393414381891489, + 0.030871696770191193, + 0.020477887243032455, + 0.015023651532828808, + 0.023737454786896706, + 0.01286389771848917, + 0.005684603936970234, + 0.005739317741245031, + 0.006045210640877485, + 0.00683607067912817, + 0.006385353393852711, + 0.005765076261013746, + 0.013235832564532757, + 0.007063128985464573, + 0.006553403567522764, + 0.00871468335390091, + 0.009313317947089672, + 0.012505792081356049, + 0.009554042480885983, + 0.00760778971016407, + 0.010504861362278461, + 0.010971015319228172, + 0.00865287147462368, + 0.012146025896072388, + 0.013716038316488266, + 0.04555026441812515, + 0.012829028069972992, + 0.007657521869987249, + 0.019654234871268272, + 0.009357660077512264, + 0.03922126069664955, + 0.007101702503859997, + 0.012484895065426826, + 0.0038993323687464, + 0.00797184556722641, + 0.014048571698367596, + 0.005383390933275223, + 0.004874089732766151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.20158660411834717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_8/index_tensor([1534, 1534])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_8/index_tensor([1534, 1534])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bd185b212645ecdc4027d34479939e70650b92b3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_8/index_tensor([1534, 1534])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot from a horizontal angle reveals a gloomy environment with virtually no external light. A strange, feathery creature slowly and gracefully unfurls from a deep, almost black sand. Composed of countless delicate, translucent filaments, the creature forms an elegant structure resembling the feathers of a fern. Its entire surface is evenly distributed with tiny bioluminescent points emitting a steady, cool white glow, providing the sole source of light in the darkness. At the moment of full extension, the entire organism suddenly and violently contracts inward, as if in a silent, energetic howl. Simultaneously, all the light points on its body transform from a steady shimmer into a frantic, high-frequency pulse, with drastic changes in intensity and rhythm. The stark contrast of black and white creates a whimsical atmosphere that is both eerie and vibrant, imbued with a sense of mystery and awe.", + "seed": 7618, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_8/index_1534_seed_7618_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08932881746441126, + "vlm_logprob": -0.08932881746441126, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.08932881746441126, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03778787702322006 + ], + "opposite_logprob_tokens": [ + -3.412787914276123 + ], + "logprob_reward": -0.03778787702322006, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03778787702322006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14707452058792114, + 0.12084295600652695, + 0.07981771975755692, + 0.08197123557329178, + 0.08195272088050842, + 0.08236328512430191, + 0.15637630224227905, + 0.03742113709449768, + 0.12291238456964493, + 0.08926776796579361 + ], + "credit_map": [ + 0.016012927517294884, + 0.014627647586166859, + 0.02046126499772072, + 0.023912834003567696, + 0.011331726796925068, + 0.019219154492020607, + 0.027019981294870377, + 0.014488987624645233, + 0.007669771555811167, + 0.012239567935466766, + 0.020174797624349594, + 0.020682139322161674, + 0.019527962431311607, + 0.011305509135127068, + 0.018548978492617607, + 0.010694220662117004, + 0.0071089728735387325, + 0.00705705676227808, + 0.01268914993852377, + 0.013947834260761738, + 0.007151218131184578, + 0.007175442297011614, + 0.014289950020611286, + 0.010398092679679394, + 0.013557025231420994, + 0.008143316023051739, + 0.009935353882610798, + 0.009912075474858284, + 0.008009197190403938, + 0.012208785861730576, + 0.010983468033373356, + 0.009222017601132393, + 0.014571529813110828, + 0.006157714407891035, + 0.012000258080661297, + 0.008648136630654335, + 0.009486890397965908, + 0.011829379014670849, + 0.009192905388772488, + 0.010065902955830097, + 0.012829055078327656, + 0.0097225122153759, + 0.008010078221559525, + 0.012563953176140785, + 0.007126639597117901, + 0.010620719753205776, + 0.011958683840930462, + 0.009531642310321331, + 0.01743469201028347, + 0.016430623829364777, + 0.014551006257534027, + 0.02711685188114643, + 0.011297507211565971, + 0.0213167704641819, + 0.027576245367527008, + 0.02065260149538517, + 0.0033733807504177094, + 0.004597761668264866, + 0.0037721265107393265, + 0.007416300009936094, + 0.0031937796156853437, + 0.0038117049261927605, + 0.00649569695815444, + 0.004760386422276497, + 0.016179218888282776, + 0.012678665108978748, + 0.010608219541609287, + 0.016526952385902405, + 0.014911843463778496, + 0.019211433827877045, + 0.018820947036147118, + 0.01397511176764965, + 0.007549244910478592, + 0.008860968053340912, + 0.008691530674695969, + 0.0156366266310215, + 0.01163316611200571, + 0.010132910683751106, + 0.014630472287535667, + 0.012132843025028706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.03778787702322006 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748621817678213 + ], + "opposite_logprob_tokens": [ + -5.006748676300049 + ], + "logprob_reward": -0.006748621817678213, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748621817678213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15299807488918304, + 0.13997918367385864, + 0.3092919588088989, + 0.15883173048496246, + 0.0686635971069336, + 0.027384327724575996, + 0.05496569350361824, + 0.026954129338264465, + 0.02807822823524475, + 0.032853078097105026 + ], + "credit_map": [ + 0.01299649104475975, + 0.014405624940991402, + 0.029214242473244667, + 0.024616029113531113, + 0.01486801728606224, + 0.014268673956394196, + 0.020713623613119125, + 0.021915368735790253, + 0.009289655834436417, + 0.010236931033432484, + 0.027167683467268944, + 0.02390860579907894, + 0.01634504459798336, + 0.013417051173746586, + 0.021376628428697586, + 0.018237581476569176, + 0.03218746557831764, + 0.026166917756199837, + 0.038911741226911545, + 0.05332449451088905, + 0.027117377147078514, + 0.03426234424114227, + 0.06155533716082573, + 0.03576626628637314, + 0.031705331057310104, + 0.009486416354775429, + 0.015291005373001099, + 0.014841943979263306, + 0.015591982752084732, + 0.033819910138845444, + 0.017566900700330734, + 0.02052823267877102, + 0.015265922993421555, + 0.004861088935285807, + 0.007165135815739632, + 0.005228868220001459, + 0.005718969739973545, + 0.017167560756206512, + 0.006279986817389727, + 0.006976062897592783, + 0.0050097364000976086, + 0.0030996769201010466, + 0.002478444715961814, + 0.0030301527585834265, + 0.0029387117829173803, + 0.005451089236885309, + 0.002754516201093793, + 0.0026220008730888367, + 0.005901790689677, + 0.006392424926161766, + 0.005337191745638847, + 0.009844469837844372, + 0.004073191434144974, + 0.0068219671957194805, + 0.009371181949973106, + 0.007223477587103844, + 0.005143142305314541, + 0.003025384619832039, + 0.002080395817756653, + 0.0027070126961916685, + 0.0022853414993733168, + 0.006042275112122297, + 0.002911839634180069, + 0.0027587369550019503, + 0.0047044879756867886, + 0.00276722339913249, + 0.002822409849613905, + 0.003019908908754587, + 0.003041369142010808, + 0.005647256039083004, + 0.0033131747040897608, + 0.0027623972855508327, + 0.003271394642069936, + 0.0033060009591281414, + 0.0033246607054024935, + 0.005372470710426569, + 0.002748757367953658, + 0.0037923483178019524, + 0.005380913615226746, + 0.00565652921795845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.006748621817678213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978530291467905 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.00978530291467905, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978530291467905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18237996101379395, + 0.14912308752536774, + 0.10301892459392548, + 0.09001307189464569, + 0.06870396435260773, + 0.051442574709653854, + 0.255588173866272, + 0.02611406333744526, + 0.03241877630352974, + 0.04119737446308136 + ], + "credit_map": [ + 0.01895616203546524, + 0.01700877957046032, + 0.03243763744831085, + 0.03226852789521217, + 0.015865709632635117, + 0.02123512141406536, + 0.025347240269184113, + 0.01926078461110592, + 0.010179605334997177, + 0.016147607937455177, + 0.019491082057356834, + 0.024565299972891808, + 0.024013405665755272, + 0.014931426383554935, + 0.025014230981469154, + 0.014780434779822826, + 0.009382566437125206, + 0.01063485350459814, + 0.019131701439619064, + 0.01960051991045475, + 0.008313632570207119, + 0.010028763674199581, + 0.015402759425342083, + 0.010524129495024681, + 0.010631807148456573, + 0.009350991807878017, + 0.008260891772806644, + 0.010251596570014954, + 0.01453119982033968, + 0.010940821841359138, + 0.015636775642633438, + 0.010408986359834671, + 0.011805380694568157, + 0.005046059377491474, + 0.008947843685746193, + 0.006681754253804684, + 0.006356194149702787, + 0.01017671450972557, + 0.009304052218794823, + 0.010385963134467602, + 0.009015548974275589, + 0.005770728923380375, + 0.00600768718868494, + 0.003636502893641591, + 0.005404390394687653, + 0.008865855634212494, + 0.005743581801652908, + 0.00699828052893281, + 0.026564795523881912, + 0.028322815895080566, + 0.02344224601984024, + 0.045355334877967834, + 0.01912134326994419, + 0.03178824484348297, + 0.047246191650629044, + 0.03374722599983215, + 0.004447526764124632, + 0.0029302218463271856, + 0.002461756346747279, + 0.0037288577295839787, + 0.002059254329651594, + 0.004531308077275753, + 0.003230629488825798, + 0.002724509220570326, + 0.005344905890524387, + 0.0030945097096264362, + 0.003510154550895095, + 0.0033234059810638428, + 0.003920108545571566, + 0.006207222584635019, + 0.0035262759774923325, + 0.003492193529382348, + 0.0034914028365164995, + 0.004784711170941591, + 0.0034488236997276545, + 0.007212775759398937, + 0.005792482756078243, + 0.004104010295122862, + 0.006743585225194693, + 0.00561958272010088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.00978530291467905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525747099891305 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.00525747099891305, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525747099891305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10893946141004562, + 0.16884779930114746, + 0.14385652542114258, + 0.09481590986251831, + 0.06842949241399765, + 0.092739999294281, + 0.17871442437171936, + 0.040572065860033035, + 0.04732118919491768, + 0.055763185024261475 + ], + "credit_map": [ + 0.007767930161207914, + 0.012283997610211372, + 0.019795475527644157, + 0.01779833249747753, + 0.010407190769910812, + 0.010269658640027046, + 0.01795186661183834, + 0.012665008194744587, + 0.011458305642008781, + 0.010459094308316708, + 0.024683916941285133, + 0.034390341490507126, + 0.011225788854062557, + 0.02333999052643776, + 0.03244009613990784, + 0.02085026726126671, + 0.012051121331751347, + 0.012204739265143871, + 0.02121807634830475, + 0.024716030806303024, + 0.01224651001393795, + 0.013274616561830044, + 0.030019737780094147, + 0.018125690519809723, + 0.011665597558021545, + 0.010414151474833488, + 0.008213594555854797, + 0.010050160810351372, + 0.01596486009657383, + 0.01177703682333231, + 0.015317851677536964, + 0.011412655003368855, + 0.01063977275043726, + 0.005942246876657009, + 0.010267173871397972, + 0.007815481163561344, + 0.008299736306071281, + 0.007523453328758478, + 0.009118903428316116, + 0.008822727017104626, + 0.012296738103032112, + 0.013170408084988594, + 0.011740939691662788, + 0.009273438714444637, + 0.012178458273410797, + 0.010796988382935524, + 0.012032720260322094, + 0.011250311508774757, + 0.019544463604688644, + 0.01816808618605137, + 0.017756562680006027, + 0.03104015812277794, + 0.013610228896141052, + 0.022471081465482712, + 0.032121442258358, + 0.02400238811969757, + 0.005793512333184481, + 0.004620559513568878, + 0.004251215141266584, + 0.004402871709316969, + 0.004095917567610741, + 0.0072756665758788586, + 0.004524465184658766, + 0.005607859697192907, + 0.004363847430795431, + 0.00500325346365571, + 0.004397199954837561, + 0.008353603072464466, + 0.00417833449319005, + 0.006417779251933098, + 0.008769597858190536, + 0.005837574135512114, + 0.004781996365636587, + 0.005353009328246117, + 0.005062234587967396, + 0.009107332676649094, + 0.006553308572620153, + 0.00660964148119092, + 0.008928781375288963, + 0.009366879239678383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.00525747099891305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870648145675659 + ], + "opposite_logprob_tokens": [ + -1.137064814567566 + ], + "logprob_reward": -0.3870648145675659, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870648145675659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1323389858007431, + 0.16924577951431274, + 0.13223892450332642, + 0.10179471224546432, + 0.06098748371005058, + 0.0840272605419159, + 0.08906768262386322, + 0.05854126811027527, + 0.12634502351284027, + 0.0454128161072731 + ], + "credit_map": [ + 0.013804769143462181, + 0.01229173969477415, + 0.01812584325671196, + 0.027669968083500862, + 0.011624119244515896, + 0.017137328162789345, + 0.019019102677702904, + 0.0126661267131567, + 0.01134044025093317, + 0.020209286361932755, + 0.02152278460562229, + 0.031547676771879196, + 0.02952035330235958, + 0.017441682517528534, + 0.021922005340456963, + 0.015741555020213127, + 0.009770232252776623, + 0.008073905482888222, + 0.023507727310061455, + 0.025800319388508797, + 0.010956074111163616, + 0.017787877470254898, + 0.02156156301498413, + 0.014781228266656399, + 0.019749432802200317, + 0.00801433902233839, + 0.010202151723206043, + 0.007639552466571331, + 0.011535209603607655, + 0.018496667966246605, + 0.013248578645288944, + 0.01290877815335989, + 0.006564363371580839, + 0.00726290512830019, + 0.005672459956258535, + 0.009103749878704548, + 0.008319612592458725, + 0.00763414753600955, + 0.0102233262732625, + 0.006206915248185396, + 0.012351665645837784, + 0.009282706305384636, + 0.011721879243850708, + 0.00908185075968504, + 0.00923224538564682, + 0.010622642934322357, + 0.011042030528187752, + 0.010692239739000797, + 0.011934087611734867, + 0.008582918904721737, + 0.009250884875655174, + 0.014495527371764183, + 0.006763769313693047, + 0.01196697074919939, + 0.013070850633084774, + 0.013002674095332623, + 0.006670934613794088, + 0.009616701863706112, + 0.006068079266697168, + 0.00909173209220171, + 0.006224160548299551, + 0.00796581618487835, + 0.007386883720755577, + 0.0055169593542814255, + 0.021568752825260162, + 0.012609537690877914, + 0.010539541020989418, + 0.011877874843776226, + 0.014612065628170967, + 0.030816249549388885, + 0.014059726148843765, + 0.01026128325611353, + 0.0054055810905992985, + 0.004078510217368603, + 0.003144108457490802, + 0.008311769925057888, + 0.005144086200743914, + 0.007235973607748747, + 0.006891895085573196, + 0.005200892686843872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.3870648145675659 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot from a horizontal angle reveals a gloomy environment with virtually no external light. A strange, feathery creature slowly and gracefully unfurls from a deep, almost black sand. Composed of countless delicate, translucent filaments, the creature forms an elegant structure resembling the feathers of a fern. Its entire surface is evenly distributed with tiny bioluminescent points emitting a steady, cool white glow, providing the sole source of light in the darkness. At the moment of full extension, the entire organism suddenly and violently contracts inward, as if in a silent, energetic howl. Simultaneously, all the light points on its body transform from a steady shimmer into a frantic, high-frequency pulse, with drastic changes in intensity and rhythm. The stark contrast of black and white creates a whimsical atmosphere that is both eerie and vibrant, imbued with a sense of mystery and awe.", + "seed": 7619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_8/index_1534_seed_7619_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08905481630936266, + "vlm_logprob": -0.08905481630936266, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.08905481630936266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1878926008939743 + ], + "opposite_logprob_tokens": [ + -1.8128925561904907 + ], + "logprob_reward": -0.1878926008939743, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1878926008939743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1988430917263031, + 0.3217684030532837, + 0.039948366582393646, + 0.16253390908241272, + 0.08339167386293411, + 0.03409731015563011, + 0.05663089454174042, + 0.02574325166642666, + 0.04847439005970955, + 0.02856874279677868 + ], + "credit_map": [ + 0.028658896684646606, + 0.015026291832327843, + 0.021714331582188606, + 0.021908404305577278, + 0.0180966779589653, + 0.042979393154382706, + 0.017987174913287163, + 0.0324719175696373, + 0.027985213324427605, + 0.09240974485874176, + 0.01631106063723564, + 0.016197960823774338, + 0.03668433800339699, + 0.028087962418794632, + 0.0187870804220438, + 0.08530504256486893, + 0.004433372989296913, + 0.005697161424905062, + 0.00548920501023531, + 0.006178822834044695, + 0.003409761004149914, + 0.004007869400084019, + 0.0033709302078932524, + 0.007361243478953838, + 0.0159751009196043, + 0.027647757902741432, + 0.014839783310890198, + 0.012016414664685726, + 0.045865584164857864, + 0.017451195046305656, + 0.013982092961668968, + 0.014755982905626297, + 0.007541503291577101, + 0.016489185392856598, + 0.005096721462905407, + 0.004743623547255993, + 0.02906900830566883, + 0.0056453365832567215, + 0.006460343953222036, + 0.008345944806933403, + 0.0029373173601925373, + 0.00508100027218461, + 0.0027746015693992376, + 0.0027524337638169527, + 0.008725790306925774, + 0.0029990419279783964, + 0.003263466991484165, + 0.0055636572651565075, + 0.0060228146612644196, + 0.008971089497208595, + 0.005852122791111469, + 0.004371530842036009, + 0.0125203812494874, + 0.006580411456525326, + 0.005534052848815918, + 0.0067784893326461315, + 0.0025038975290954113, + 0.0032046933192759752, + 0.0024836098309606314, + 0.006297122221440077, + 0.003174097742885351, + 0.0020996113307774067, + 0.003142008325085044, + 0.002838212065398693, + 0.005066707730293274, + 0.008961914107203484, + 0.004205610137432814, + 0.005580581724643707, + 0.010457135736942291, + 0.004555338993668556, + 0.005797174759209156, + 0.003849924774840474, + 0.0013748644851148129, + 0.0024613002315163612, + 0.006647206377238035, + 0.008477849885821342, + 0.0016341618029400706, + 0.0013791018864139915, + 0.00412532314658165, + 0.002468934515491128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.1878926008939743 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086198389530182 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.016086198389530182, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086198389530182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3719542324542999, + 0.15339498221874237, + 0.012723165564239025, + 0.15437749028205872, + 0.2201087772846222, + 0.005477935075759888, + 0.01383821852505207, + 0.010075511410832405, + 0.012993115931749344, + 0.04505658522248268 + ], + "credit_map": [ + 0.11444760113954544, + 0.013900259509682655, + 0.013525846414268017, + 0.015154055319726467, + 0.02411741204559803, + 0.13977956771850586, + 0.01283544022589922, + 0.03819403797388077, + 0.010122394189238548, + 0.058091096580028534, + 0.0048014456406235695, + 0.006692501250654459, + 0.02294745109975338, + 0.00924769975244999, + 0.010853298008441925, + 0.03063909336924553, + 0.001385124633088708, + 0.002387621672824025, + 0.0012532295659184456, + 0.0012803307035937905, + 0.0016191760078072548, + 0.0011652343673631549, + 0.001473244745284319, + 0.002159204101189971, + 0.012992823496460915, + 0.014795608818531036, + 0.017850687727332115, + 0.01997249200940132, + 0.036177679896354675, + 0.019016100093722343, + 0.01779436133801937, + 0.015777720138430595, + 0.014840339310467243, + 0.04519791528582573, + 0.007691747974604368, + 0.009334482252597809, + 0.08315276354551315, + 0.011766293086111546, + 0.012803268618881702, + 0.03532195836305618, + 0.0007225269800983369, + 0.0008476161165162921, + 0.0007099039503373206, + 0.0005617839633487165, + 0.0006953318370506167, + 0.0007037683972157538, + 0.0006218946073204279, + 0.0006151092238724232, + 0.0016132899327203631, + 0.001859978074207902, + 0.0016069180564954877, + 0.0012923784088343382, + 0.002417482202872634, + 0.0021203537471592426, + 0.0012465404579415917, + 0.0016812782268971205, + 0.0009785061702132225, + 0.001480903709307313, + 0.0016757113626226783, + 0.0009734206832945347, + 0.0014193353708833456, + 0.0009209187119267881, + 0.001580134965479374, + 0.0010465800296515226, + 0.001345977303571999, + 0.0021019624546170235, + 0.0012203434016555548, + 0.0017125816084444523, + 0.0016048618126660585, + 0.00173549959436059, + 0.0018776260549202561, + 0.0013942637015134096, + 0.0016493486473336816, + 0.0024449292104691267, + 0.005571005865931511, + 0.013555795885622501, + 0.0025597314815968275, + 0.0013593384064733982, + 0.01482069306075573, + 0.003095741383731365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.016086198389530182 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081226170063019 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081226170063019, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081226170063019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22441436350345612, + 0.20195166766643524, + 0.041201185435056686, + 0.06409648805856705, + 0.16147984564304352, + 0.030240265652537346, + 0.06178177148103714, + 0.046797093003988266, + 0.05296642333269119, + 0.11507091671228409 + ], + "credit_map": [ + 0.06796098500490189, + 0.007511885836720467, + 0.008641674183309078, + 0.006266621872782707, + 0.013340710662305355, + 0.0887540951371193, + 0.005969063378870487, + 0.025969337671995163, + 0.015166633762419224, + 0.06579656898975372, + 0.007354197092354298, + 0.007808716036379337, + 0.026412487030029297, + 0.014509578235447407, + 0.014077350497245789, + 0.05082613229751587, + 0.003547684522345662, + 0.009530073031783104, + 0.004630549810826778, + 0.003069942817091942, + 0.006258153822273016, + 0.004293962847441435, + 0.0033950016368180513, + 0.006475815083831549, + 0.005098293535411358, + 0.007589456159621477, + 0.0069625964388251305, + 0.006759720854461193, + 0.01604553312063217, + 0.009302406571805477, + 0.006747367326170206, + 0.0055911182425916195, + 0.010742210783064365, + 0.03336849808692932, + 0.006557046435773373, + 0.0076340725645422935, + 0.06106496974825859, + 0.008552130311727524, + 0.00894610770046711, + 0.024614812806248665, + 0.0032696749549359083, + 0.00452757254242897, + 0.0035629523918032646, + 0.0021901463624089956, + 0.006637027021497488, + 0.003835320705547929, + 0.002837995532900095, + 0.0033795766066759825, + 0.007277695927768946, + 0.009241253137588501, + 0.006915634032338858, + 0.00523796770721674, + 0.011539648286998272, + 0.007402298040688038, + 0.0059147728607058525, + 0.008252501487731934, + 0.004981884732842445, + 0.003986966330558062, + 0.005804417654871941, + 0.010268256068229675, + 0.004377477336674929, + 0.003340156050398946, + 0.00461842305958271, + 0.009419513866305351, + 0.005341286305338144, + 0.006794567685574293, + 0.0065583279356360435, + 0.008910628035664558, + 0.009591436013579369, + 0.005176505073904991, + 0.006596507970243692, + 0.003997165709733963, + 0.008403751067817211, + 0.011712600477039814, + 0.02195107564330101, + 0.031623635441064835, + 0.008275717496871948, + 0.005498131737112999, + 0.019376324489712715, + 0.008229685015976429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.011081226170063019 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645389039069414 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645389039069414, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645389039069414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6327493786811829, + 0.059036076068878174, + 0.029017962515354156, + 0.09443903714418411, + 0.08645281940698624, + 0.012878570705652237, + 0.009152916260063648, + 0.009170269593596458, + 0.03474953770637512, + 0.03235337510704994 + ], + "credit_map": [ + 0.1956315040588379, + 0.020235316827893257, + 0.019912725314497948, + 0.017231155186891556, + 0.0382949598133564, + 0.2571103870868683, + 0.014352704398334026, + 0.0699806734919548, + 0.003619450842961669, + 0.01964103803038597, + 0.004005889408290386, + 0.0032052251044660807, + 0.008626650087535381, + 0.004292868077754974, + 0.004110909998416901, + 0.011534048244357109, + 0.0026603485457599163, + 0.0061018881388008595, + 0.0028708812315016985, + 0.0026497940998524427, + 0.002788497367873788, + 0.003064139047637582, + 0.0022119253408163786, + 0.006670489441603422, + 0.0092698959633708, + 0.014775269664824009, + 0.00797310657799244, + 0.005012300331145525, + 0.028303787112236023, + 0.011560271494090557, + 0.009218106046319008, + 0.008326304145157337, + 0.008940514177083969, + 0.013075435534119606, + 0.006876467261463404, + 0.005597731098532677, + 0.025732452049851418, + 0.006437689531594515, + 0.008313759230077267, + 0.011478775180876255, + 0.0017089156899601221, + 0.00173906737472862, + 0.0015273335156962276, + 0.0010881595080718398, + 0.0016178993973881006, + 0.0019707721658051014, + 0.0014271740801632404, + 0.001799248973838985, + 0.0008701195474714041, + 0.0011527275200933218, + 0.0011736521264538169, + 0.0008430466987192631, + 0.0015580608742311597, + 0.0011742034694179893, + 0.0009788918541744351, + 0.001402214402332902, + 0.0009965781355276704, + 0.0011971627827733755, + 0.000982065568678081, + 0.0015010065399110317, + 0.001366142532788217, + 0.0006900010048411787, + 0.001209734007716179, + 0.0012275787303224206, + 0.003631262108683586, + 0.004631913732737303, + 0.003996559884399176, + 0.004995655734091997, + 0.006395417265594006, + 0.003595601534470916, + 0.004591679200530052, + 0.0029114489443600178, + 0.001675027422606945, + 0.0030997975263744593, + 0.004999786149710417, + 0.0069149392656981945, + 0.0023325379006564617, + 0.0015850329073145986, + 0.005772887263447046, + 0.005973366554826498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.004645389039069414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556866705417633 + ], + "opposite_logprob_tokens": [ + -1.6005686521530151 + ], + "logprob_reward": -0.22556866705417633, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556866705417633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49988752603530884, + 0.11713922768831253, + 0.023020612075924873, + 0.049305737018585205, + 0.049772463738918304, + 0.010679065249860287, + 0.027601877227425575, + 0.0590328611433506, + 0.07106372714042664, + 0.09249687939882278 + ], + "credit_map": [ + 0.15395882725715637, + 0.0158756785094738, + 0.01585141196846962, + 0.014331071637570858, + 0.028706805780529976, + 0.20404662191867828, + 0.010892892256379128, + 0.056224193423986435, + 0.007378803566098213, + 0.04374459385871887, + 0.0031191499438136816, + 0.005270030349493027, + 0.017736196517944336, + 0.007005384657531977, + 0.008210371248424053, + 0.02467469684779644, + 0.002431172411888838, + 0.004273315891623497, + 0.003164528636261821, + 0.00204061740078032, + 0.002436915412545204, + 0.002545745810493827, + 0.0030928112100809813, + 0.0030355043709278107, + 0.005192657932639122, + 0.0074834758415818214, + 0.005860281642526388, + 0.003296295180916786, + 0.01180682610720396, + 0.006091069430112839, + 0.004431590437889099, + 0.005143539514392614, + 0.005529746413230896, + 0.006094750016927719, + 0.0038848405238240957, + 0.0028955130837857723, + 0.015516066923737526, + 0.004030653275549412, + 0.005426387768238783, + 0.006394504103809595, + 0.0012873229570686817, + 0.00139663927257061, + 0.001296570640988648, + 0.0009555742726661265, + 0.0014850753359496593, + 0.001583239296451211, + 0.0013943294761702418, + 0.0012803138233721256, + 0.003060624934732914, + 0.00310297473333776, + 0.0036935906391590834, + 0.0025005978532135487, + 0.004203481134027243, + 0.0031650292221456766, + 0.003346511395648122, + 0.004529066849499941, + 0.005249070003628731, + 0.006564343813806772, + 0.005387091543525457, + 0.0121003994718194, + 0.010772954672574997, + 0.004373561125248671, + 0.0069859568029642105, + 0.007599485106766224, + 0.006208722945302725, + 0.0094675337895751, + 0.009404096752405167, + 0.014783550053834915, + 0.01107625663280487, + 0.005083535332232714, + 0.00790791679173708, + 0.007132109720259905, + 0.006334207020699978, + 0.01159917563199997, + 0.021678704768419266, + 0.011359397321939468, + 0.008898427709937096, + 0.005313723348081112, + 0.010713771916925907, + 0.016599474474787712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.22556866705417633 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot from a horizontal angle reveals a gloomy environment with virtually no external light. A strange, feathery creature slowly and gracefully unfurls from a deep, almost black sand. Composed of countless delicate, translucent filaments, the creature forms an elegant structure resembling the feathers of a fern. Its entire surface is evenly distributed with tiny bioluminescent points emitting a steady, cool white glow, providing the sole source of light in the darkness. At the moment of full extension, the entire organism suddenly and violently contracts inward, as if in a silent, energetic howl. Simultaneously, all the light points on its body transform from a steady shimmer into a frantic, high-frequency pulse, with drastic changes in intensity and rhythm. The stark contrast of black and white creates a whimsical atmosphere that is both eerie and vibrant, imbued with a sense of mystery and awe.", + "seed": 7620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_8/index_1534_seed_7620_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10048783342354, + "vlm_logprob": -0.10048783342354, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.10048783342354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3909914493560791 + ], + "opposite_logprob_tokens": [ + -1.140991449356079 + ], + "logprob_reward": -0.3909914493560791, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3909914493560791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036727700382471085, + 0.019076496362686157, + 0.4477822482585907, + 0.015588155016303062, + 0.15618422627449036, + 0.028065523132681847, + 0.03912383317947388, + 0.1981734037399292, + 0.03259535878896713, + 0.026683073490858078 + ], + "credit_map": [ + 0.004130158573389053, + 0.005339285358786583, + 0.003171629272401333, + 0.002730740001425147, + 0.0028584664687514305, + 0.011408288963139057, + 0.002229532925412059, + 0.0048595997504889965, + 0.0022768396884202957, + 0.0018580557079985738, + 0.0018171824049204588, + 0.002421214012429118, + 0.0022665655706077814, + 0.0033092747908085585, + 0.0017561070853844285, + 0.003371256636455655, + 0.035787638276815414, + 0.025628723204135895, + 0.026295559480786324, + 0.0063815852627158165, + 0.1063484475016594, + 0.1892164945602417, + 0.012254834175109863, + 0.04586895927786827, + 0.0015479744179174304, + 0.002152038970962167, + 0.0019118840573355556, + 0.0016617679502815008, + 0.0022226106375455856, + 0.002262200228869915, + 0.0014647984644398093, + 0.00236487970687449, + 0.009678823873400688, + 0.0075618792325258255, + 0.04137646034359932, + 0.005963000003248453, + 0.0073094540275633335, + 0.05838644132018089, + 0.006051487289369106, + 0.01985667273402214, + 0.0036278918851166964, + 0.002155926311388612, + 0.004769176710397005, + 0.002729297848418355, + 0.0029020849615335464, + 0.006719466298818588, + 0.0020454281475394964, + 0.00311625050380826, + 0.005449057091027498, + 0.002513565355911851, + 0.007339246571063995, + 0.0038695293478667736, + 0.002775944536551833, + 0.005175309255719185, + 0.0036223733332008123, + 0.008378805592656136, + 0.019936896860599518, + 0.0034303630236536264, + 0.07915503531694412, + 0.008625984191894531, + 0.00501555809751153, + 0.04280532896518707, + 0.006421214900910854, + 0.03278302773833275, + 0.004197040572762489, + 0.0023618170525878668, + 0.005952106788754463, + 0.004846809897571802, + 0.002554754028096795, + 0.004317830316722393, + 0.004041472915560007, + 0.0043235281482338905, + 0.002091789385303855, + 0.0016023588832467794, + 0.007291657850146294, + 0.0027192982379347086, + 0.0012581178452819586, + 0.007200682535767555, + 0.002522199647501111, + 0.0019969698041677475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.3909914493560791 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097969900816679 + ], + "opposite_logprob_tokens": [ + -4.634797096252441 + ], + "logprob_reward": -0.0097969900816679, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097969900816679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032358523458242416, + 0.009040450677275658, + 0.6297162771224976, + 0.010300148278474808, + 0.05206596851348877, + 0.030550576746463776, + 0.024768706411123276, + 0.11411066353321075, + 0.04097165912389755, + 0.05611696094274521 + ], + "credit_map": [ + 0.00404551113024354, + 0.0029021541122347116, + 0.0036660577170550823, + 0.0029629014898091555, + 0.003707433817908168, + 0.00826255977153778, + 0.002778946654871106, + 0.004032960627228022, + 0.0013255380326882005, + 0.001000917167402804, + 0.0010410447139292955, + 0.0011730869300663471, + 0.0017709553940221667, + 0.0008801885996945202, + 0.0007031848654150963, + 0.0011455350322648883, + 0.051819443702697754, + 0.03619775176048279, + 0.03763065114617348, + 0.010060865432024002, + 0.14861059188842773, + 0.2646629214286804, + 0.017667384818196297, + 0.06306669861078262, + 0.0011520162224769592, + 0.0015712748281657696, + 0.0012496946146711707, + 0.0008824362885206938, + 0.0012655529426410794, + 0.0014360147761180997, + 0.000983065227046609, + 0.00176009361166507, + 0.0029095944482833147, + 0.002147563500329852, + 0.019550981000065804, + 0.003244127845391631, + 0.0022022766061127186, + 0.013260087929666042, + 0.0024353135377168655, + 0.00631602481007576, + 0.003526776097714901, + 0.0026774650905281305, + 0.005477530881762505, + 0.0025251894257962704, + 0.0033083599992096424, + 0.006753440015017986, + 0.002201470546424389, + 0.0040803453885018826, + 0.0028889887034893036, + 0.002942851046100259, + 0.004234608728438616, + 0.0029629033524543047, + 0.003552017267793417, + 0.0024797348305583, + 0.002495398046448827, + 0.0032122046686708927, + 0.01214357279241085, + 0.002090110909193754, + 0.04607025533914566, + 0.005461032036691904, + 0.0025310111232101917, + 0.022100845351815224, + 0.0037463961634784937, + 0.019967442378401756, + 0.0065034059807658195, + 0.0019506120588630438, + 0.008987005800008774, + 0.004013041500002146, + 0.001761112012900412, + 0.008151931688189507, + 0.003489891067147255, + 0.006114658899605274, + 0.0063681770116090775, + 0.002896253252401948, + 0.008914664387702942, + 0.007464507129043341, + 0.003208668204024434, + 0.009782799519598484, + 0.008308596909046173, + 0.009173291735351086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.0097969900816679 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789198637008667 + ], + "opposite_logprob_tokens": [ + -4.634788990020752 + ], + "logprob_reward": -0.009789198637008667, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789198637008667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027431797236204147, + 0.009598752483725548, + 0.47947269678115845, + 0.013329416513442993, + 0.040782101452350616, + 0.037843361496925354, + 0.11656524240970612, + 0.16170547902584076, + 0.06447618454694748, + 0.048794981092214584 + ], + "credit_map": [ + 0.003122172085568309, + 0.002201386960223317, + 0.00389288947917521, + 0.0017872375901788473, + 0.0023054650519043207, + 0.010550426319241524, + 0.0013512081932276487, + 0.002221011556684971, + 0.0011541425483301282, + 0.0012326620053499937, + 0.0009699981310404837, + 0.0011111990315839648, + 0.001176005112938583, + 0.001857271301560104, + 0.0009976942092180252, + 0.0010997796198353171, + 0.038752876222133636, + 0.02813686616718769, + 0.02866358868777752, + 0.006889409385621548, + 0.11413639038801193, + 0.2022659331560135, + 0.012591357342898846, + 0.04803627356886864, + 0.0013218306703493, + 0.0019969469867646694, + 0.0015209977282211185, + 0.0020391708239912987, + 0.0019765484612435102, + 0.001347506302408874, + 0.0018175403820350766, + 0.001308875042013824, + 0.0026532593183219433, + 0.0018371688202023506, + 0.013860227540135384, + 0.0025522562209516764, + 0.0018877543043345213, + 0.002415319439023733, + 0.0026420888025313616, + 0.012934025377035141, + 0.003843451151624322, + 0.002511040773242712, + 0.007973607629537582, + 0.0032013726886361837, + 0.0026637164410203695, + 0.004474857822060585, + 0.00355151598341763, + 0.009623797610402107, + 0.01350422017276287, + 0.011755099520087242, + 0.02262294664978981, + 0.010639295913279057, + 0.015252500772476196, + 0.013431185856461525, + 0.011820020154118538, + 0.017539963126182556, + 0.014496926218271255, + 0.004069566261023283, + 0.06185721606016159, + 0.007893047295510769, + 0.006258344277739525, + 0.04207097366452217, + 0.005793075542896986, + 0.01926633156836033, + 0.008994660340249538, + 0.005459675099700689, + 0.014068901538848877, + 0.007284540683031082, + 0.004130280576646328, + 0.00796041265130043, + 0.006366622634232044, + 0.010211094282567501, + 0.004706386476755142, + 0.0017888437723740935, + 0.018142947927117348, + 0.0036589992232620716, + 0.0018948721699416637, + 0.009951447136700153, + 0.0023659979924559593, + 0.006285486742854118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.009789198637008667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028255560901015997 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028255560901015997, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028255560901015997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05014805495738983, + 0.020555052906274796, + 0.45159319043159485, + 0.02225344255566597, + 0.10680871456861496, + 0.06374836713075638, + 0.05335722491145134, + 0.08288592100143433, + 0.03806763142347336, + 0.11058233678340912 + ], + "credit_map": [ + 0.008355963975191116, + 0.006143801845610142, + 0.004312549252063036, + 0.004290601704269648, + 0.0032500936649739742, + 0.01484670490026474, + 0.004313168115913868, + 0.00463517801836133, + 0.0019137008348479867, + 0.002452173503115773, + 0.0015331044560298324, + 0.002632065676152706, + 0.0031261725816875696, + 0.004289533011615276, + 0.002283915877342224, + 0.00232438906095922, + 0.03509160876274109, + 0.028204303234815598, + 0.02490248531103134, + 0.006441748235374689, + 0.1106824278831482, + 0.18842536211013794, + 0.012678309343755245, + 0.045166973024606705, + 0.0022299359552562237, + 0.002988838590681553, + 0.003216600976884365, + 0.0030523822642862797, + 0.0030622510239481926, + 0.0023137421812862158, + 0.002942739287391305, + 0.0024469548370689154, + 0.00674033910036087, + 0.006238602567464113, + 0.01403653435409069, + 0.004944018088281155, + 0.006275419145822525, + 0.040080081671476364, + 0.006193249486386776, + 0.0223004799336195, + 0.00635281577706337, + 0.0029860367067158222, + 0.012119553983211517, + 0.0033063278533518314, + 0.003069053404033184, + 0.01173867005854845, + 0.0031694441568106413, + 0.021006468683481216, + 0.005431216675788164, + 0.0060523697175085545, + 0.00768493814393878, + 0.005967693403363228, + 0.008149241097271442, + 0.006492766551673412, + 0.0060404641553759575, + 0.007538538426160812, + 0.005882111378014088, + 0.0029103041160851717, + 0.025830242782831192, + 0.003910745959728956, + 0.0037393379025161266, + 0.030516352504491806, + 0.0037088433746248484, + 0.006387987174093723, + 0.004810404032468796, + 0.0017786325188353658, + 0.00827676523476839, + 0.003984665963798761, + 0.0018269101856276393, + 0.007708173245191574, + 0.00462261401116848, + 0.005059469491243362, + 0.008987000212073326, + 0.0021291046869009733, + 0.04353084787726402, + 0.006610088050365448, + 0.0026164387818425894, + 0.03425988182425499, + 0.00400558253750205, + 0.008443406783044338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.0028255560901015997 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903597295284271 + ], + "opposite_logprob_tokens": [ + -2.464035987854004 + ], + "logprob_reward": -0.08903597295284271, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903597295284271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025933459401130676, + 0.020104652270674706, + 0.5341469049453735, + 0.02935388870537281, + 0.1890355497598648, + 0.058871954679489136, + 0.02840040996670723, + 0.03661681339144707, + 0.031565796583890915, + 0.045970529317855835 + ], + "credit_map": [ + 0.004227814264595509, + 0.0028071317356079817, + 0.003255632473155856, + 0.002312536584213376, + 0.001959871733561158, + 0.006255853921175003, + 0.00221935473382473, + 0.00289526442065835, + 0.0017824573442339897, + 0.0019522352376952767, + 0.0025481225457042456, + 0.0028370062354952097, + 0.0024253891315311193, + 0.004108296241611242, + 0.0021235684398561716, + 0.002327576046809554, + 0.04258976876735687, + 0.0317293256521225, + 0.03128731623291969, + 0.00792434997856617, + 0.12844620645046234, + 0.22433698177337646, + 0.01415521465241909, + 0.0536777600646019, + 0.0031041447073221207, + 0.0031510856933891773, + 0.004159459378570318, + 0.003607323160395026, + 0.0031205490231513977, + 0.00517208082601428, + 0.002863266272470355, + 0.004175978247076273, + 0.016129346564412117, + 0.006598100531846285, + 0.06903237849473953, + 0.007254856172949076, + 0.0046611749567091465, + 0.01739894412457943, + 0.0050299083814024925, + 0.06293085217475891, + 0.005201264750212431, + 0.0033921641297638416, + 0.010207107290625572, + 0.0036412745248526335, + 0.003985778894275427, + 0.00957783404737711, + 0.004717097617685795, + 0.01814943552017212, + 0.0030829510651528835, + 0.0031681249383836985, + 0.004725477192550898, + 0.0030212129931896925, + 0.003999926149845123, + 0.003667339449748397, + 0.0033443323336541653, + 0.003391045378521085, + 0.004924908746033907, + 0.0024424451403319836, + 0.009671550244092941, + 0.0035381035413593054, + 0.0022894376888871193, + 0.0037207119166851044, + 0.0028346527833491564, + 0.007195002865046263, + 0.004194260109215975, + 0.002236833330243826, + 0.005821498576551676, + 0.004094227217137814, + 0.002450608881190419, + 0.004059219267219305, + 0.004784349352121353, + 0.003924799617379904, + 0.0058089978992938995, + 0.0025207807775586843, + 0.011811630800366402, + 0.004473804961889982, + 0.0026045276317745447, + 0.004653123673051596, + 0.0035763459745794535, + 0.010521317832171917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.08903597295284271 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot from a horizontal angle reveals a gloomy environment with virtually no external light. A strange, feathery creature slowly and gracefully unfurls from a deep, almost black sand. Composed of countless delicate, translucent filaments, the creature forms an elegant structure resembling the feathers of a fern. Its entire surface is evenly distributed with tiny bioluminescent points emitting a steady, cool white glow, providing the sole source of light in the darkness. At the moment of full extension, the entire organism suddenly and violently contracts inward, as if in a silent, energetic howl. Simultaneously, all the light points on its body transform from a steady shimmer into a frantic, high-frequency pulse, with drastic changes in intensity and rhythm. The stark contrast of black and white creates a whimsical atmosphere that is both eerie and vibrant, imbued with a sense of mystery and awe.", + "seed": 7621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_8/index_1534_seed_7621_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04030070542357862, + "vlm_logprob": -0.04030070542357862, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.04030070542357862, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15321393311023712 + ], + "opposite_logprob_tokens": [ + -2.0282139778137207 + ], + "logprob_reward": -0.15321393311023712, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.15321393311023712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29096218943595886, + 0.07231413573026657, + 0.05593949183821678, + 0.07188817113637924, + 0.07545037567615509, + 0.16863344609737396, + 0.09570024907588959, + 0.10439024865627289, + 0.042407769709825516, + 0.022313915193080902 + ], + "credit_map": [ + 0.01825723797082901, + 0.020513420924544334, + 0.046095214784145355, + 0.028605302795767784, + 0.019782206043601036, + 0.055655766278505325, + 0.05301201343536377, + 0.04904104396700859, + 0.005215656477957964, + 0.008290525525808334, + 0.007578718941658735, + 0.010438147932291031, + 0.005856937263160944, + 0.011204246431589127, + 0.012165660038590431, + 0.011564240790903568, + 0.003980443812906742, + 0.007037206087261438, + 0.007255615200847387, + 0.007227569352835417, + 0.006180366035550833, + 0.006612097378820181, + 0.009284485131502151, + 0.008361713029444218, + 0.00689941318705678, + 0.007104723248630762, + 0.009845522232353687, + 0.007753433659672737, + 0.006419527344405651, + 0.008276065811514854, + 0.012598997913300991, + 0.012990490533411503, + 0.007152758538722992, + 0.007866035215556622, + 0.008401254191994667, + 0.007456056773662567, + 0.006288257893174887, + 0.013541312888264656, + 0.012512297369539738, + 0.012232406996190548, + 0.014001691713929176, + 0.02495199628174305, + 0.020505165681242943, + 0.02279289811849594, + 0.014890393242239952, + 0.027888353914022446, + 0.02125357836484909, + 0.02234938181936741, + 0.009373859502375126, + 0.0165165513753891, + 0.013063427060842514, + 0.013765825890004635, + 0.0070138187147676945, + 0.007391931489109993, + 0.011055818758904934, + 0.017519015818834305, + 0.00810658186674118, + 0.024829652160406113, + 0.010429484769701958, + 0.020636387169361115, + 0.011727000586688519, + 0.007109815254807472, + 0.012019224464893341, + 0.009532098658382893, + 0.004990628454834223, + 0.0067201764322817326, + 0.0040414477698504925, + 0.007020886521786451, + 0.0033880474511533976, + 0.003673554165288806, + 0.005601765587925911, + 0.006971264258027077, + 0.0032168489415198565, + 0.0021914788521826267, + 0.0024504568427801132, + 0.002822980983182788, + 0.00189744564704597, + 0.0035103620029985905, + 0.003070166101679206, + 0.0031541779171675444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.15321393311023712 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090893298387527 + ], + "opposite_logprob_tokens": [ + -4.5110907554626465 + ], + "logprob_reward": -0.011090893298387527, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090893298387527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39855486154556274, + 0.09408872574567795, + 0.08131435513496399, + 0.08289442211389542, + 0.03585479408502579, + 0.06750339269638062, + 0.10535511374473572, + 0.06423244625329971, + 0.05215216055512428, + 0.01804967224597931 + ], + "credit_map": [ + 0.025744851678609848, + 0.03887917101383209, + 0.07005349546670914, + 0.05167866125702858, + 0.027021322399377823, + 0.04953397437930107, + 0.07046463340520859, + 0.0651787668466568, + 0.006387047935277224, + 0.009075326845049858, + 0.017152652144432068, + 0.013072551228106022, + 0.007617925759404898, + 0.010897473432123661, + 0.015307513065636158, + 0.014578239060938358, + 0.00623973598703742, + 0.009963908232748508, + 0.013503698632121086, + 0.010980418883264065, + 0.008373385295271873, + 0.01137138158082962, + 0.00962899997830391, + 0.011252840980887413, + 0.005387124605476856, + 0.012445076368749142, + 0.011983445845544338, + 0.011844202876091003, + 0.007308226078748703, + 0.011052120476961136, + 0.010675798170268536, + 0.012198437005281448, + 0.0034121668431907892, + 0.0048263599164783955, + 0.004972390364855528, + 0.0037540572229772806, + 0.003339095739647746, + 0.005678392946720123, + 0.004023168236017227, + 0.005849165841937065, + 0.007051351945847273, + 0.011323543265461922, + 0.007116740569472313, + 0.010161942802369595, + 0.006375443655997515, + 0.009767880663275719, + 0.007164661306887865, + 0.008541835471987724, + 0.009055634960532188, + 0.01620226539671421, + 0.011718538589775562, + 0.014850446954369545, + 0.008009896613657475, + 0.010242321528494358, + 0.016816023737192154, + 0.01845998503267765, + 0.007401951123028994, + 0.011620400473475456, + 0.006277715787291527, + 0.009414038620889187, + 0.006394952069967985, + 0.005672936327755451, + 0.010597244836390018, + 0.006853216327726841, + 0.007030868902802467, + 0.01013876125216484, + 0.005110176280140877, + 0.0071130055002868176, + 0.005357156042009592, + 0.0061423094011843204, + 0.004669482819736004, + 0.006590405944734812, + 0.0026501179672777653, + 0.0023551734630018473, + 0.0017412283923476934, + 0.0020119070541113615, + 0.0015624138759449124, + 0.003203745698556304, + 0.002063749125227332, + 0.002461338648572564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.011090893298387527 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084762401878834 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084762401878834, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084762401878834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21768763661384583, + 0.09477901458740234, + 0.05734894797205925, + 0.07319118827581406, + 0.04279639571905136, + 0.25563469529151917, + 0.07324160635471344, + 0.06475277990102768, + 0.08630333840847015, + 0.034264326095581055 + ], + "credit_map": [ + 0.016524896025657654, + 0.02510041370987892, + 0.02467338554561138, + 0.03573428839445114, + 0.01997678354382515, + 0.026574507355690002, + 0.03790793940424919, + 0.031195422634482384, + 0.009237241931259632, + 0.009849111549556255, + 0.010784212499856949, + 0.008977766148746014, + 0.008655217476189137, + 0.01744876429438591, + 0.014497916214168072, + 0.015328786335885525, + 0.006107245106250048, + 0.006650245748460293, + 0.010836438275873661, + 0.006813521031290293, + 0.004991305526345968, + 0.0061263395473361015, + 0.0070261843502521515, + 0.008797668851912022, + 0.005494609009474516, + 0.011774315498769283, + 0.008694038726389408, + 0.011589997448027134, + 0.006268523167818785, + 0.0070413281209766865, + 0.011915034614503384, + 0.0104133365675807, + 0.0026256542187184095, + 0.007694641128182411, + 0.004793718922883272, + 0.006026684772223234, + 0.004246527794748545, + 0.0040130168199539185, + 0.007933307439088821, + 0.005462846718728542, + 0.024203116074204445, + 0.0495443269610405, + 0.024499360471963882, + 0.03429296240210533, + 0.029034197330474854, + 0.03332221135497093, + 0.027497155591845512, + 0.033241353929042816, + 0.006817986257374287, + 0.009359536692500114, + 0.008278772234916687, + 0.012015851214528084, + 0.005555482115596533, + 0.00560382753610611, + 0.011356079950928688, + 0.014254077337682247, + 0.007026171777397394, + 0.014776291325688362, + 0.005114557221531868, + 0.012891973368823528, + 0.005154068116098642, + 0.00553244398906827, + 0.006954038515686989, + 0.007303230930119753, + 0.010632496327161789, + 0.018139375373721123, + 0.008746487088501453, + 0.014634194783866405, + 0.008116237819194794, + 0.009648885577917099, + 0.007553355302661657, + 0.008832301013171673, + 0.0033170590177178383, + 0.0034524230286478996, + 0.0026169721968472004, + 0.0027481361757963896, + 0.0019583466928452253, + 0.008710341528058052, + 0.002772537525743246, + 0.008688509464263916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.011084762401878834 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002823535120114684 + ], + "opposite_logprob_tokens": [ + -5.877823352813721 + ], + "logprob_reward": -0.002823535120114684, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002823535120114684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12642717361450195, + 0.09334868937730789, + 0.058243874460458755, + 0.15574142336845398, + 0.0942157506942749, + 0.10490137338638306, + 0.13030578196048737, + 0.11282607167959213, + 0.0965123102068901, + 0.027477482333779335 + ], + "credit_map": [ + 0.010111546143889427, + 0.01110891718417406, + 0.018286973237991333, + 0.019835295155644417, + 0.009250858798623085, + 0.014664110727608204, + 0.023482242599129677, + 0.019687242805957794, + 0.00555147510021925, + 0.012001898139715195, + 0.012294563464820385, + 0.014670701697468758, + 0.008191497065126896, + 0.013700584881007671, + 0.011561423540115356, + 0.015376551076769829, + 0.005184022709727287, + 0.006178481038659811, + 0.0070228674449026585, + 0.008121617138385773, + 0.005813505966216326, + 0.006827958859503269, + 0.009730767458677292, + 0.009364661760628223, + 0.016136236488819122, + 0.01384325698018074, + 0.02467990294098854, + 0.015920959413051605, + 0.013079223223030567, + 0.021765371784567833, + 0.025611847639083862, + 0.02470463328063488, + 0.005856595002114773, + 0.016551226377487183, + 0.010424889624118805, + 0.01141894981265068, + 0.010845680721104145, + 0.009678451344370842, + 0.016614370048046112, + 0.012825597077608109, + 0.009884346276521683, + 0.013316990807652473, + 0.013051035813987255, + 0.01872306317090988, + 0.010051026940345764, + 0.013747579418122768, + 0.012923741713166237, + 0.013203602284193039, + 0.00932229869067669, + 0.0172526054084301, + 0.013227395713329315, + 0.024337520822882652, + 0.012112773954868317, + 0.0128252524882555, + 0.02688753791153431, + 0.014340408146381378, + 0.009624766185879707, + 0.02373085729777813, + 0.008718104101717472, + 0.023136090487241745, + 0.01120277214795351, + 0.009270761162042618, + 0.015545617789030075, + 0.011597110889852047, + 0.01243410911411047, + 0.020795857533812523, + 0.009506461210548878, + 0.015196958556771278, + 0.009328300133347511, + 0.010846516117453575, + 0.008180554956197739, + 0.010223563760519028, + 0.003748095128685236, + 0.0027344448026269674, + 0.0030345774721354246, + 0.0026009255088865757, + 0.002722138538956642, + 0.004699795041233301, + 0.0021770764142274857, + 0.005760431755334139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.002823535120114684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment gloomy with virtually no external light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a strange, feathery creature composed of translucent filaments unfurl from deep black sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature's surface emit a steady cool white bioluminescent glow that serves as the sole light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature suddenly contract inward while its light points transform into a frantic, high-frequency pulse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290403187274933 + ], + "opposite_logprob_tokens": [ + -3.7732903957366943 + ], + "logprob_reward": -0.023290403187274933, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290403187274933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1363087296485901, + 0.06897851824760437, + 0.0779080018401146, + 0.07169316709041595, + 0.09822580218315125, + 0.13038939237594604, + 0.1517021507024765, + 0.11421898752450943, + 0.12496975064277649, + 0.025605492293834686 + ], + "credit_map": [ + 0.009075225330889225, + 0.022982941940426826, + 0.018501022830605507, + 0.017611078917980194, + 0.01395375281572342, + 0.01539516169577837, + 0.018771803006529808, + 0.020017748698592186, + 0.004874564707279205, + 0.011715753935277462, + 0.00915658101439476, + 0.01095658540725708, + 0.0058200787752866745, + 0.007232045289129019, + 0.008935515768826008, + 0.010287394747138023, + 0.007091618608683348, + 0.011562624014914036, + 0.010756067000329494, + 0.010327886790037155, + 0.006403835024684668, + 0.008119010366499424, + 0.011315634474158287, + 0.012331326492130756, + 0.00650404067710042, + 0.011483476497232914, + 0.008319987915456295, + 0.011186075396835804, + 0.006213839165866375, + 0.00701398029923439, + 0.010722204111516476, + 0.010249563492834568, + 0.008617888204753399, + 0.013530637137591839, + 0.012664131820201874, + 0.014083595015108585, + 0.00969434529542923, + 0.009977346286177635, + 0.014912115409970284, + 0.01474574115127325, + 0.012318500317633152, + 0.02348673716187477, + 0.014538818039000034, + 0.018663572147488594, + 0.013744333758950233, + 0.01590951532125473, + 0.015569386072456837, + 0.01615852303802967, + 0.012479796074330807, + 0.01741982437670231, + 0.01870860904455185, + 0.023629389703273773, + 0.016624828800559044, + 0.012448879890143871, + 0.027755575254559517, + 0.02263525128364563, + 0.009629782289266586, + 0.028053566813468933, + 0.007775224279612303, + 0.02314986288547516, + 0.00937945768237114, + 0.009327782317996025, + 0.014416920952498913, + 0.012486395426094532, + 0.014540845528244972, + 0.02604244463145733, + 0.011432022787630558, + 0.021569231525063515, + 0.011527354829013348, + 0.013444053940474987, + 0.012848732993006706, + 0.013565060682594776, + 0.002821862231940031, + 0.002853844780474901, + 0.0021438198164105415, + 0.0029683909378945827, + 0.0016248057363554835, + 0.005885773804038763, + 0.0028160803485661745, + 0.004490915220230818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.023290403187274933 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_9/index_tensor([3801, 3801])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_9/index_tensor([3801, 3801])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..857a9b7071cd397448ae768b47c02d87972c0b9c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_9/index_tensor([3801, 3801])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, sweeping 360-degree shot, shot from a slightly upward angle, focuses on an eagle perched atop a precarious cliff. The golden late afternoon sun illuminates the jagged, weathered gray rock face. Below, a vast, bottomless canyon stretches, and distant mountains loom through the mist. The eagle's plumage is dark brown, almost black, and its crest on the back of its neck gleams golden-brown in the light. Its powerful yellow talons grip the rock, while its sharp amber eyes warily survey its territory. As the camera steadily rotates around it, the eagle suddenly unfurls its massive wings, each feather visible, its wingspan nearly obscuring a corner of the sky. With its head held high and its sharp, hooked yellow beak extended, it lets out a piercing cry that echoes through the empty canyon. The sunlight outlines a dazzling golden edge on the dark feathers of the eagle, and also makes the rough texture of the rock become three-dimensional and distinct. The whole picture, in warm tones, through slow and continuous circular motion, fully displays the majestic posture of this bird of prey and the magnificent natural scenery around it, creating an epic feeling of grandeur, aloofness and primitive vitality.", + "seed": 3228, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_9/index_3801_seed_3228_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3095624649897218, + "vlm_logprob": -0.3095624649897218, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.3095624649897218, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010120858438313007 + ], + "opposite_logprob_tokens": [ + -4.885120868682861 + ], + "logprob_reward": -0.010120858438313007, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010120858438313007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15530063211917877, + 0.06339532881975174, + 0.035644643008708954, + 0.07336100190877914, + 0.09561402350664139, + 0.0572284460067749, + 0.28101304173469543, + 0.12036552280187607, + 0.03871334344148636, + 0.07936405390501022 + ], + "credit_map": [ + 0.015665778890252113, + 0.019050506874918938, + 0.02255888655781746, + 0.021067705005407333, + 0.01406815368682146, + 0.028276534751057625, + 0.014706375077366829, + 0.01990668848156929, + 0.009297141805291176, + 0.0060927970334887505, + 0.009816846810281277, + 0.005950918886810541, + 0.0073033482767641544, + 0.011594699695706367, + 0.006055662874132395, + 0.007283915765583515, + 0.007601647637784481, + 0.0033211419358849525, + 0.00617650430649519, + 0.0031323805451393127, + 0.0037992699071764946, + 0.005298216827213764, + 0.003717708168551326, + 0.00259776902385056, + 0.008515716530382633, + 0.00977344997227192, + 0.010917273350059986, + 0.007723354268819094, + 0.007054867222905159, + 0.011189145967364311, + 0.009464125148952007, + 0.008723068051040173, + 0.010732121765613556, + 0.007683625444769859, + 0.014109889976680279, + 0.014094335958361626, + 0.01085349265486002, + 0.011144960299134254, + 0.013623169623315334, + 0.013372426852583885, + 0.01218134444206953, + 0.008225404657423496, + 0.007023966871201992, + 0.00633775582537055, + 0.004968264605849981, + 0.008094422519207, + 0.004702950362116098, + 0.0056943343952298164, + 0.057695262134075165, + 0.019443122670054436, + 0.07506886124610901, + 0.032305922359228134, + 0.018127350136637688, + 0.030696364119648933, + 0.01912868581712246, + 0.028547462075948715, + 0.017855204641819, + 0.013077198527753353, + 0.012659565545618534, + 0.012100707739591599, + 0.023433223366737366, + 0.014828605577349663, + 0.012712927535176277, + 0.013698081485927105, + 0.00543056707829237, + 0.006263407412916422, + 0.004435989074409008, + 0.0051576835103333, + 0.004314670339226723, + 0.003270963206887245, + 0.003291563829407096, + 0.006548501085489988, + 0.009368149563670158, + 0.009030715562403202, + 0.008751343935728073, + 0.010508572682738304, + 0.013523043133318424, + 0.009329233318567276, + 0.008335837163031101, + 0.010517163202166557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.010120858438313007 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609957218170166 + ], + "opposite_logprob_tokens": [ + -4.141099452972412 + ], + "logprob_reward": -0.01609957218170166, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609957218170166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15659041702747345, + 0.14252053201198578, + 0.08285870403051376, + 0.09414268285036087, + 0.043888188898563385, + 0.07516232132911682, + 0.0692928358912468, + 0.07556598633527756, + 0.10461483150720596, + 0.15536347031593323 + ], + "credit_map": [ + 0.01388117577880621, + 0.02302698977291584, + 0.024789534509181976, + 0.017548510804772377, + 0.019640764221549034, + 0.022656947374343872, + 0.017658062279224396, + 0.01738842763006687, + 0.015574352815747261, + 0.017853721976280212, + 0.020908594131469727, + 0.007395523134618998, + 0.025641372427344322, + 0.030361153185367584, + 0.013084704987704754, + 0.011701098643243313, + 0.01796148717403412, + 0.009673640131950378, + 0.014169381931424141, + 0.00615807855501771, + 0.009384578093886375, + 0.007432312238961458, + 0.008728304877877235, + 0.00935092382133007, + 0.01290830411016941, + 0.0105354068800807, + 0.013536385260522366, + 0.009558519348502159, + 0.008287597447633743, + 0.012178671546280384, + 0.017204763367772102, + 0.009933029301464558, + 0.00630147522315383, + 0.004663093015551567, + 0.006567862816154957, + 0.005745972041040659, + 0.004334726836532354, + 0.006016128230839968, + 0.0046376679092645645, + 0.005621262360364199, + 0.01067401934415102, + 0.00858169887214899, + 0.008194824680685997, + 0.013471568934619427, + 0.008497118018567562, + 0.013501529581844807, + 0.006096131168305874, + 0.006145431660115719, + 0.008102970197796822, + 0.007172416429966688, + 0.007791281212121248, + 0.009990940801799297, + 0.010063041932880878, + 0.0077612679451704025, + 0.01058666780591011, + 0.007824253290891647, + 0.011105972342193127, + 0.011105244979262352, + 0.006577229127287865, + 0.007286198902875185, + 0.015671607106924057, + 0.007281928788870573, + 0.00827262457460165, + 0.008265174925327301, + 0.014915660955011845, + 0.01337706483900547, + 0.020792951807379723, + 0.016780409961938858, + 0.011570756323635578, + 0.010566013865172863, + 0.006654699333012104, + 0.0099572679027915, + 0.017210664227604866, + 0.025525931268930435, + 0.01622620038688183, + 0.016799556091427803, + 0.02087024226784706, + 0.02090679667890072, + 0.016020959243178368, + 0.021803127601742744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01609957218170166 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979085173457861 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.00979085173457861, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979085173457861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1259121298789978, + 0.06025675684213638, + 0.031045781448483467, + 0.10406451672315598, + 0.015680158510804176, + 0.03254856541752815, + 0.07754351198673248, + 0.28050529956817627, + 0.057585883885622025, + 0.21485741436481476 + ], + "credit_map": [ + 0.010005493648350239, + 0.01597617194056511, + 0.018206484615802765, + 0.016652295365929604, + 0.015340595506131649, + 0.0209911298006773, + 0.014601134695112705, + 0.014138822443783283, + 0.007393068168312311, + 0.0047305854968726635, + 0.009719724766910076, + 0.005436666775494814, + 0.005951295606791973, + 0.014086749404668808, + 0.006042391061782837, + 0.006896278355270624, + 0.005849324632436037, + 0.0031742991413921118, + 0.005410062614828348, + 0.0028594601899385452, + 0.0039472440257668495, + 0.004523657262325287, + 0.002988983178511262, + 0.0022927522659301758, + 0.013015476986765862, + 0.012404794804751873, + 0.016627700999379158, + 0.01057177223265171, + 0.010461771860718727, + 0.015038584358990192, + 0.014187783002853394, + 0.011756631545722485, + 0.002269408665597439, + 0.0018417491810396314, + 0.002242548391222954, + 0.0013993275351822376, + 0.0015194083098322153, + 0.0029546150472015142, + 0.0016865513753145933, + 0.0017665496561676264, + 0.004554049111902714, + 0.003917671274393797, + 0.004287044983357191, + 0.006332647055387497, + 0.0040100920014083385, + 0.0033797952346503735, + 0.00368209695443511, + 0.0023851667065173388, + 0.019144469872117043, + 0.005326560698449612, + 0.025497183203697205, + 0.00598619831725955, + 0.004073144402354956, + 0.008386150002479553, + 0.004798963665962219, + 0.004330847412347794, + 0.03425755351781845, + 0.045010168105363846, + 0.03218693286180496, + 0.030973033979535103, + 0.04151134565472603, + 0.03488342836499214, + 0.03251936659216881, + 0.029163489118218422, + 0.0065201944671571255, + 0.0054526920430362225, + 0.008328301832079887, + 0.009344113059341908, + 0.0059608700685203075, + 0.005613388959318399, + 0.007944864220917225, + 0.008421460166573524, + 0.02551041916012764, + 0.025138573721051216, + 0.024583054706454277, + 0.030084827914834023, + 0.03610986843705177, + 0.0283505879342556, + 0.023679222911596298, + 0.021400876343250275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.00979085173457861 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.502018690109253 + ], + "opposite_logprob_tokens": [ + -0.25201863050460815 + ], + "logprob_reward": -1.502018690109253, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.502018690109253, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1567162424325943, + 0.09244661033153534, + 0.04267019033432007, + 0.11572816222906113, + 0.028071656823158264, + 0.09786432981491089, + 0.06324882805347443, + 0.0992976650595665, + 0.09519324451684952, + 0.2087630182504654 + ], + "credit_map": [ + 0.016424257308244705, + 0.01385671179741621, + 0.022167876362800598, + 0.015679005533456802, + 0.01990974321961403, + 0.032061200588941574, + 0.01595558226108551, + 0.020661871880292892, + 0.00867162924259901, + 0.008895640261471272, + 0.01420067623257637, + 0.007662577088922262, + 0.012384261935949326, + 0.021030519157648087, + 0.009849006310105324, + 0.0097523033618927, + 0.006715270224958658, + 0.005986510775983334, + 0.006815784145146608, + 0.005610273685306311, + 0.004457472823560238, + 0.003704370930790901, + 0.006143698003143072, + 0.0032368069514632225, + 0.012260125018656254, + 0.01607409119606018, + 0.01644832082092762, + 0.01241592038422823, + 0.012562994845211506, + 0.012962967157363892, + 0.017063962295651436, + 0.015939781442284584, + 0.00413120910525322, + 0.003671499202027917, + 0.0038013230077922344, + 0.0023623413871973753, + 0.002387535059824586, + 0.006446501240134239, + 0.0023226281628012657, + 0.0029486198909580708, + 0.016828134655952454, + 0.008989816531538963, + 0.01214098371565342, + 0.020660534501075745, + 0.008469182066619396, + 0.014577290043234825, + 0.008153436705470085, + 0.008044948801398277, + 0.00897843949496746, + 0.006566711701452732, + 0.012100470252335072, + 0.006221368908882141, + 0.00765733327716589, + 0.008466437458992004, + 0.008546177297830582, + 0.004711884073913097, + 0.024509426206350327, + 0.010910194367170334, + 0.012263663113117218, + 0.01100308820605278, + 0.01194065436720848, + 0.011935148388147354, + 0.010022541508078575, + 0.006712953094393015, + 0.012806338258087635, + 0.012608814053237438, + 0.017241276800632477, + 0.014398756437003613, + 0.010440931655466557, + 0.007446324452757835, + 0.010818501003086567, + 0.009432298131287098, + 0.02624175325036049, + 0.03022751957178116, + 0.026658063754439354, + 0.027377445250749588, + 0.03332275152206421, + 0.0259567704051733, + 0.019947659224271774, + 0.019031057134270668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -1.502018690109253 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782352484762669 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.009782352484762669, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782352484762669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3407558798789978, + 0.11850333213806152, + 0.08039547502994537, + 0.04066426679491997, + 0.044411275535821915, + 0.0328247994184494, + 0.04722141847014427, + 0.1073940247297287, + 0.06115081533789635, + 0.1266787201166153 + ], + "credit_map": [ + 0.030922509729862213, + 0.03158276528120041, + 0.059528451412916183, + 0.02630644664168358, + 0.036563314497470856, + 0.08145464956760406, + 0.02710590325295925, + 0.0472918376326561, + 0.0162991750985384, + 0.008789360523223877, + 0.021224036812782288, + 0.007940882816910744, + 0.008420504629611969, + 0.03624868020415306, + 0.0070786252617836, + 0.012502061203122139, + 0.022916225716471672, + 0.010707673616707325, + 0.013155142776668072, + 0.0056359004229307175, + 0.006840775720775127, + 0.011636647395789623, + 0.005240104626864195, + 0.004263000097125769, + 0.004590184893459082, + 0.004756601061671972, + 0.008084460161626339, + 0.0036541607696563005, + 0.004031324293464422, + 0.007078215014189482, + 0.0034002282191067934, + 0.00506909005343914, + 0.006468720734119415, + 0.004140711389482021, + 0.0068313018418848515, + 0.004547180607914925, + 0.004905000329017639, + 0.00632905401289463, + 0.00531240738928318, + 0.0058768982999026775, + 0.00599918095394969, + 0.002618323313072324, + 0.004438037518411875, + 0.004691049922257662, + 0.0033646472729742527, + 0.005070885177701712, + 0.003961703274399042, + 0.0026809703558683395, + 0.009096602909266949, + 0.007575893308967352, + 0.006449003703892231, + 0.003438298823311925, + 0.006403493229299784, + 0.006184460129588842, + 0.004205521661788225, + 0.0038681463338434696, + 0.02437528595328331, + 0.012645972892642021, + 0.014488555490970612, + 0.010214883834123611, + 0.012874174863100052, + 0.0134709682315588, + 0.0096849724650383, + 0.00963920634239912, + 0.007237363141030073, + 0.008712160401046276, + 0.007793313357979059, + 0.011046933010220528, + 0.005587655119597912, + 0.00473760487511754, + 0.006241328082978725, + 0.009794458746910095, + 0.01575382985174656, + 0.014764913357794285, + 0.014972411096096039, + 0.017192417755723, + 0.02288052998483181, + 0.015734246000647545, + 0.012707985937595367, + 0.012672379612922668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.009782352484762669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "sweeping, sweeping 360-degree shot, shot from a slightly upward angle, focuses on an eagle perched atop a precarious cliff. The golden late afternoon sun illuminates the jagged, weathered gray rock face. Below, a vast, bottomless canyon stretches, and distant mountains loom through the mist. The eagle's plumage is dark brown, almost black, and its crest on the back of its neck gleams golden-brown in the light. Its powerful yellow talons grip the rock, while its sharp amber eyes warily survey its territory. As the camera steadily rotates around it, the eagle suddenly unfurls its massive wings, each feather visible, its wingspan nearly obscuring a corner of the sky. With its head held high and its sharp, hooked yellow beak extended, it lets out a piercing cry that echoes through the empty canyon. The sunlight outlines a dazzling golden edge on the dark feathers of the eagle, and also makes the rough texture of the rock become three-dimensional and distinct. The whole picture, in warm tones, through slow and continuous circular motion, fully displays the majestic posture of this bird of prey and the magnificent natural scenery around it, creating an epic feeling of grandeur, aloofness and primitive vitality.", + "seed": 3229, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_9/index_3801_seed_3229_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.40656396858394145, + "vlm_logprob": -0.40656396858394145, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.40656396858394145, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0595005601644516 + ], + "opposite_logprob_tokens": [ + -2.9345004558563232 + ], + "logprob_reward": -0.0595005601644516, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0595005601644516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11614573746919632, + 0.07767128944396973, + 0.15152090787887573, + 0.2120635062456131, + 0.04932895302772522, + 0.06415844708681107, + 0.2140270471572876, + 0.053807664662599564, + 0.02542226016521454, + 0.035854145884513855 + ], + "credit_map": [ + 0.011529614217579365, + 0.010282731615006924, + 0.01263176929205656, + 0.015090075321495533, + 0.0179781224578619, + 0.013663429766893387, + 0.013347228057682514, + 0.021622775122523308, + 0.007048366125673056, + 0.006962737999856472, + 0.00714032631367445, + 0.005327689461410046, + 0.006401788908988237, + 0.01818474754691124, + 0.010611175559461117, + 0.015994466841220856, + 0.017977258190512657, + 0.015861600637435913, + 0.019707517698407173, + 0.014582673087716103, + 0.015795066952705383, + 0.029676996171474457, + 0.0159723162651062, + 0.021947501227259636, + 0.020641610026359558, + 0.026255473494529724, + 0.03136884421110153, + 0.03448792174458504, + 0.03210679814219475, + 0.029554998502135277, + 0.017816392704844475, + 0.019831495359539986, + 0.003909602295607328, + 0.004349257331341505, + 0.007202991284430027, + 0.0046342601999640465, + 0.008786980994045734, + 0.0061209434643387794, + 0.006892284844070673, + 0.0074326395988464355, + 0.005494667217135429, + 0.007933140732347965, + 0.011110612191259861, + 0.005727862939238548, + 0.009121960960328579, + 0.0076244245283305645, + 0.006772716995328665, + 0.010373069904744625, + 0.01861056312918663, + 0.022152043879032135, + 0.029541781172156334, + 0.02190251834690571, + 0.027402611449360847, + 0.024871014058589935, + 0.03472268581390381, + 0.03482385352253914, + 0.003147512674331665, + 0.004136094823479652, + 0.006648566108196974, + 0.006418332923203707, + 0.011045225895941257, + 0.00869051180779934, + 0.006491278763860464, + 0.007230149116367102, + 0.0012913341633975506, + 0.0013630155008286238, + 0.004679908510297537, + 0.006660813000053167, + 0.002329054055735469, + 0.0025757180992513895, + 0.0028141611255705357, + 0.0037082605995237827, + 0.0038174984510987997, + 0.00198788708075881, + 0.005855206400156021, + 0.00445707980543375, + 0.0021996626164764166, + 0.004119097255170345, + 0.0031451452523469925, + 0.010272568091750145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.0595005601644516 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011101503856480122 + ], + "opposite_logprob_tokens": [ + -4.511101722717285 + ], + "logprob_reward": -0.011101503856480122, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011101503856480122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18439659476280212, + 0.08059678226709366, + 0.06946992129087448, + 0.10328102111816406, + 0.061653513461351395, + 0.031251464039087296, + 0.16738781332969666, + 0.08961482346057892, + 0.03788076341152191, + 0.1744672954082489 + ], + "credit_map": [ + 0.034648895263671875, + 0.020159780979156494, + 0.016765179112553596, + 0.01915893703699112, + 0.019458366557955742, + 0.03167019784450531, + 0.01396858599036932, + 0.0285666361451149, + 0.00732355797663331, + 0.008818432688713074, + 0.007878840900957584, + 0.00909420382231474, + 0.019682735204696655, + 0.010391752235591412, + 0.007168709300458431, + 0.010238549672067165, + 0.007680507376790047, + 0.0086282379925251, + 0.01864263229072094, + 0.00561729958280921, + 0.005534186959266663, + 0.009756823070347309, + 0.004704103339463472, + 0.008906135335564613, + 0.008497143164277077, + 0.017618844285607338, + 0.013101025484502316, + 0.013354099355638027, + 0.010433075949549675, + 0.016679100692272186, + 0.011569724418222904, + 0.012028010562062263, + 0.005020568147301674, + 0.005435767583549023, + 0.010440043173730373, + 0.004234141670167446, + 0.00618131086230278, + 0.008272379636764526, + 0.01073553878813982, + 0.011333764530718327, + 0.002476477064192295, + 0.0030574966222047806, + 0.006647053640335798, + 0.0026684983167797327, + 0.002857435029000044, + 0.004162851255387068, + 0.0035179401747882366, + 0.005863710772246122, + 0.013910266570746899, + 0.02807317115366459, + 0.01871637813746929, + 0.012457441538572311, + 0.030553624033927917, + 0.018154766410589218, + 0.0187322199344635, + 0.026789939031004906, + 0.004940672777593136, + 0.007575302850455046, + 0.011045037768781185, + 0.009297839365899563, + 0.012981616891920567, + 0.015054816380143166, + 0.0095361964777112, + 0.019183341413736343, + 0.002156480448320508, + 0.0028238974045962095, + 0.006046634633094072, + 0.007125223986804485, + 0.0037662885151803493, + 0.004056334495544434, + 0.0043885475024580956, + 0.0075173573568463326, + 0.018224384635686874, + 0.010471267625689507, + 0.024029234424233437, + 0.016079528257250786, + 0.009997093118727207, + 0.03025166504085064, + 0.014655590988695621, + 0.050758540630340576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.011101503856480122 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25208839774131775 + ], + "opposite_logprob_tokens": [ + -1.5020884275436401 + ], + "logprob_reward": -0.25208839774131775, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25208839774131775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11585193872451782, + 0.062343254685401917, + 0.07358241081237793, + 0.12705937027931213, + 0.047006797045469284, + 0.05587530881166458, + 0.34317299723625183, + 0.1004863902926445, + 0.01717321015894413, + 0.057448260486125946 + ], + "credit_map": [ + 0.027511270716786385, + 0.011115258559584618, + 0.010744092054665089, + 0.009381938725709915, + 0.01655278354883194, + 0.02527102641761303, + 0.005894296802580357, + 0.009381287731230259, + 0.01180195715278387, + 0.007533492986112833, + 0.006889846175909042, + 0.005721640773117542, + 0.006081791594624519, + 0.012673587538301945, + 0.004442701116204262, + 0.007198243401944637, + 0.009898127987980843, + 0.00963528547435999, + 0.014402985572814941, + 0.0069680847227573395, + 0.005661272443830967, + 0.009058993309736252, + 0.005944147240370512, + 0.012013517320156097, + 0.012011785991489887, + 0.021525327116250992, + 0.017780274152755737, + 0.016490021720528603, + 0.018009379506111145, + 0.01740984059870243, + 0.00980117917060852, + 0.01403158251196146, + 0.004117786418646574, + 0.005439565982669592, + 0.007011567708104849, + 0.004208221100270748, + 0.007303027901798487, + 0.005704471841454506, + 0.007108274381607771, + 0.006113884039223194, + 0.004297242965549231, + 0.007948199287056923, + 0.007720572873950005, + 0.005183883011341095, + 0.0071021756157279015, + 0.006643605884164572, + 0.005821005441248417, + 0.011158633045852184, + 0.036689162254333496, + 0.055196069180965424, + 0.031056923791766167, + 0.026257304474711418, + 0.07039902359247208, + 0.03733213245868683, + 0.03445184975862503, + 0.05179056152701378, + 0.004996090661734343, + 0.006774940062314272, + 0.017290784046053886, + 0.012707591988146305, + 0.013852475211024284, + 0.012145201675593853, + 0.010791081935167313, + 0.02192823402583599, + 0.0010962765663862228, + 0.0015218380140140653, + 0.002728068735450506, + 0.003288273699581623, + 0.00197256775572896, + 0.0017191363731399179, + 0.002275649458169937, + 0.0025714028161019087, + 0.00504140043631196, + 0.002804876770824194, + 0.0067984857596457005, + 0.004180999007076025, + 0.004509999882429838, + 0.014930486679077148, + 0.004756968934088945, + 0.014425051398575306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.25208839774131775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7014895677566528 + ], + "opposite_logprob_tokens": [ + -0.20148955285549164 + ], + "logprob_reward": -1.7014895677566528, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7014895677566528, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12664775550365448, + 0.06651856005191803, + 0.12433014810085297, + 0.14695999026298523, + 0.06604969501495361, + 0.02334889955818653, + 0.28926923871040344, + 0.055184002965688705, + 0.02723170444369316, + 0.07445991039276123 + ], + "credit_map": [ + 0.02533208765089512, + 0.016410456970334053, + 0.010614688508212566, + 0.006217566784471273, + 0.010798524133861065, + 0.025280144065618515, + 0.011143156327307224, + 0.020851152017712593, + 0.008258840069174767, + 0.009051631204783916, + 0.00759872468188405, + 0.006147483829408884, + 0.008648082613945007, + 0.013169584795832634, + 0.006886540446430445, + 0.006757679395377636, + 0.015087217092514038, + 0.012986832298338413, + 0.026130614802241325, + 0.012309537269175053, + 0.008289864286780357, + 0.018587978556752205, + 0.008787744678556919, + 0.022150374948978424, + 0.013666818849742413, + 0.022111205384135246, + 0.023288585245609283, + 0.017718249931931496, + 0.020131373777985573, + 0.019637735560536385, + 0.012354232370853424, + 0.018051810562610626, + 0.0060926200821995735, + 0.007125207222998142, + 0.010766956955194473, + 0.006511279847472906, + 0.009080917574465275, + 0.007311269640922546, + 0.011288481764495373, + 0.007872974500060081, + 0.0021042253356426954, + 0.002200234215706587, + 0.004523961804807186, + 0.0024436162784695625, + 0.002133508212864399, + 0.0031964946538209915, + 0.002225745003670454, + 0.004521115683019161, + 0.018754247575998306, + 0.041413143277168274, + 0.03312024474143982, + 0.02546328864991665, + 0.05361182987689972, + 0.037019453942775726, + 0.030505837872624397, + 0.04938118904829025, + 0.0024666546378284693, + 0.0034516078885644674, + 0.010822155512869358, + 0.005839762277901173, + 0.005886424332857132, + 0.008335296995937824, + 0.005748956464231014, + 0.01263314951211214, + 0.0014153377851471305, + 0.0014464244013652205, + 0.0050905197858810425, + 0.0049303313717246056, + 0.0021789181046187878, + 0.002524841809645295, + 0.00476471520960331, + 0.004880619700998068, + 0.0062895268201828, + 0.00425270851701498, + 0.011284484528005123, + 0.009662026539444923, + 0.006201574113219976, + 0.013003304600715637, + 0.006725439801812172, + 0.017040856182575226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -1.7014895677566528 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639813400804996 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639813400804996, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639813400804996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2832830846309662, + 0.0940360277891159, + 0.14176546037197113, + 0.06902510672807693, + 0.03294935077428818, + 0.020080752670764923, + 0.14248992502689362, + 0.15448422729969025, + 0.018853016197681427, + 0.04303307831287384 + ], + "credit_map": [ + 0.0736321359872818, + 0.028081059455871582, + 0.024357978254556656, + 0.015117548406124115, + 0.023292964324355125, + 0.07107075303792953, + 0.013484720140695572, + 0.03424590826034546, + 0.013383490964770317, + 0.012751530855894089, + 0.012910937890410423, + 0.0076916394755244255, + 0.00834599882364273, + 0.019008925184607506, + 0.005230350885540247, + 0.014713144861161709, + 0.020468594506382942, + 0.015918802469968796, + 0.03418576344847679, + 0.008736273273825645, + 0.006566510070115328, + 0.029938263818621635, + 0.008418410085141659, + 0.017532844096422195, + 0.007309013977646828, + 0.010174657218158245, + 0.010814042761921883, + 0.0069788191467523575, + 0.007896681316196918, + 0.00981401838362217, + 0.007690023630857468, + 0.008347839117050171, + 0.003586881561204791, + 0.0035819467157125473, + 0.005442029796540737, + 0.0029947077855467796, + 0.004087480716407299, + 0.00414324551820755, + 0.004047835245728493, + 0.0050652180798351765, + 0.0024953163228929043, + 0.0019068947294726968, + 0.00279372395016253, + 0.001975715858861804, + 0.0015843667788431048, + 0.0026057660579681396, + 0.003313459688797593, + 0.003405507653951645, + 0.010093459859490395, + 0.020146436989307404, + 0.015499534085392952, + 0.01287970319390297, + 0.02861599624156952, + 0.01596011035144329, + 0.016724590212106705, + 0.02257007174193859, + 0.008553553372621536, + 0.009168166667222977, + 0.01798904873430729, + 0.018216170370578766, + 0.010856325738132, + 0.0395972765982151, + 0.01405397243797779, + 0.03604969382286072, + 0.0011699690949171782, + 0.0011217397404834628, + 0.002789184218272567, + 0.00358183984644711, + 0.002349165501073003, + 0.0017561663407832384, + 0.002364477841183543, + 0.0037204716354608536, + 0.003630915191024542, + 0.0025357988197356462, + 0.005958146415650845, + 0.0030699397902935743, + 0.0030322533566504717, + 0.008007903583347797, + 0.0037686426658183336, + 0.013029474765062332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.008639813400804996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "sweeping, sweeping 360-degree shot, shot from a slightly upward angle, focuses on an eagle perched atop a precarious cliff. The golden late afternoon sun illuminates the jagged, weathered gray rock face. Below, a vast, bottomless canyon stretches, and distant mountains loom through the mist. The eagle's plumage is dark brown, almost black, and its crest on the back of its neck gleams golden-brown in the light. Its powerful yellow talons grip the rock, while its sharp amber eyes warily survey its territory. As the camera steadily rotates around it, the eagle suddenly unfurls its massive wings, each feather visible, its wingspan nearly obscuring a corner of the sky. With its head held high and its sharp, hooked yellow beak extended, it lets out a piercing cry that echoes through the empty canyon. The sunlight outlines a dazzling golden edge on the dark feathers of the eagle, and also makes the rough texture of the rock become three-dimensional and distinct. The whole picture, in warm tones, through slow and continuous circular motion, fully displays the majestic posture of this bird of prey and the magnificent natural scenery around it, creating an epic feeling of grandeur, aloofness and primitive vitality.", + "seed": 3230, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_9/index_3801_seed_3230_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2305540181696415, + "vlm_logprob": -0.2305540181696415, + "vlm_yes_no_margin": 2.9750000119209288, + "vlm_token_credit": -0.2305540181696415, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019902996718883514 + ], + "opposite_logprob_tokens": [ + -4.019903182983398 + ], + "logprob_reward": -0.019902996718883514, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019902996718883514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07338552176952362, + 0.06418877094984055, + 0.14025607705116272, + 0.11268065124750137, + 0.05914550647139549, + 0.10302794724702835, + 0.07609157264232635, + 0.1411818116903305, + 0.10754550993442535, + 0.12249668687582016 + ], + "credit_map": [ + 0.013014636002480984, + 0.013479883782565594, + 0.008530492894351482, + 0.009176508523523808, + 0.00891739223152399, + 0.009739318862557411, + 0.005972474347800016, + 0.004554809536784887, + 0.008347157388925552, + 0.007252129260450602, + 0.010243909433484077, + 0.007152402773499489, + 0.008115487173199654, + 0.011971225962042809, + 0.00643563736230135, + 0.004670817404985428, + 0.02110718935728073, + 0.018657315522432327, + 0.03096829168498516, + 0.013680973090231419, + 0.01653899997472763, + 0.012966780923306942, + 0.017759263515472412, + 0.00857724528759718, + 0.016602421179413795, + 0.02300182916224003, + 0.0183439739048481, + 0.013559876941144466, + 0.011297126300632954, + 0.011175411753356457, + 0.01021015178412199, + 0.008489852771162987, + 0.012492883019149303, + 0.010185208171606064, + 0.011875269003212452, + 0.00786916259676218, + 0.004413163755089045, + 0.0033739826176315546, + 0.005516663659363985, + 0.0034191682934761047, + 0.013107852078974247, + 0.013097594492137432, + 0.016948046162724495, + 0.013266786932945251, + 0.01464220229536295, + 0.01632566936314106, + 0.010132350958883762, + 0.005507438909262419, + 0.013351699337363243, + 0.00720312399789691, + 0.009181046858429909, + 0.010433110408484936, + 0.014880452305078506, + 0.007414740510284901, + 0.007646132260560989, + 0.005981264170259237, + 0.01608947291970253, + 0.023623505607247353, + 0.013114950619637966, + 0.012702947482466698, + 0.01935085467994213, + 0.03165359050035477, + 0.01222267746925354, + 0.0124238021671772, + 0.011343104764819145, + 0.007830921560525894, + 0.012119190767407417, + 0.016782449558377266, + 0.006972128059715033, + 0.016018256545066833, + 0.01835155487060547, + 0.01812789961695671, + 0.01993168704211712, + 0.011737124063074589, + 0.02162415161728859, + 0.017917754128575325, + 0.011498968116939068, + 0.017171544954180717, + 0.009642494842410088, + 0.01297294907271862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.019902996718883514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061407081782818 + ], + "opposite_logprob_tokens": [ + -3.8956141471862793 + ], + "logprob_reward": -0.02061407081782818, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061407081782818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2199862152338028, + 0.17761653661727905, + 0.07860910147428513, + 0.04626672342419624, + 0.09357576817274094, + 0.07138117402791977, + 0.04575558006763458, + 0.09763751924037933, + 0.07221071422100067, + 0.09696070849895477 + ], + "credit_map": [ + 0.02805253118276596, + 0.04516126215457916, + 0.02397244982421398, + 0.03053220734000206, + 0.023416075855493546, + 0.030256709083914757, + 0.021497981622815132, + 0.017096975818276405, + 0.025668788701295853, + 0.016568180173635483, + 0.04769570752978325, + 0.016822388395667076, + 0.016757572069764137, + 0.03121599555015564, + 0.013233178295195103, + 0.00965473148971796, + 0.0193942841142416, + 0.006485047284513712, + 0.01668824441730976, + 0.006517563946545124, + 0.008861755952239037, + 0.008064442314207554, + 0.007901538163423538, + 0.0046962229534983635, + 0.006160636432468891, + 0.008560944348573685, + 0.007246953435242176, + 0.0069170184433460236, + 0.004642744082957506, + 0.004965172149240971, + 0.004671163856983185, + 0.0031020869500935078, + 0.016767684370279312, + 0.010946241207420826, + 0.020126333460211754, + 0.011487527750432491, + 0.008504987694323063, + 0.009343955665826797, + 0.008061151951551437, + 0.008337879553437233, + 0.009976480156183243, + 0.007546538952738047, + 0.014277669601142406, + 0.01285682711750269, + 0.009083911776542664, + 0.007381064351648092, + 0.005717364605516195, + 0.004541316535323858, + 0.007396124303340912, + 0.005201548337936401, + 0.006754656787961721, + 0.006075878161936998, + 0.006655273027718067, + 0.006852154154330492, + 0.0032603899016976357, + 0.0035595514345914125, + 0.00906334351748228, + 0.01037064753472805, + 0.020003175362944603, + 0.010770543478429317, + 0.008736314252018929, + 0.022210124880075455, + 0.008891548030078411, + 0.007591824978590012, + 0.010156487114727497, + 0.00516174640506506, + 0.011551761999726295, + 0.012103554792702198, + 0.006257002707570791, + 0.007788420654833317, + 0.011211658827960491, + 0.007980083115398884, + 0.014386190101504326, + 0.012116975151002407, + 0.016660356894135475, + 0.011336490511894226, + 0.008954070508480072, + 0.01172499917447567, + 0.011077210307121277, + 0.010704422369599342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02061407081782818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865625500679016 + ], + "opposite_logprob_tokens": [ + -3.0486562252044678 + ], + "logprob_reward": -0.04865625500679016, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865625500679016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10832938551902771, + 0.2086815983057022, + 0.10930947214365005, + 0.14608658850193024, + 0.08956768363714218, + 0.06803860515356064, + 0.0915398821234703, + 0.10429622232913971, + 0.042973775416612625, + 0.031176753342151642 + ], + "credit_map": [ + 0.01613905280828476, + 0.013397588394582272, + 0.010229324921965599, + 0.020637379959225655, + 0.011493156664073467, + 0.013963359408080578, + 0.011850484646856785, + 0.01061903964728117, + 0.025195959955453873, + 0.023597735911607742, + 0.04054843634366989, + 0.019573412835597992, + 0.026040436699986458, + 0.0318770557641983, + 0.02549298293888569, + 0.016355568543076515, + 0.01520767342299223, + 0.017606772482395172, + 0.01790716126561165, + 0.011108867824077606, + 0.014930791221559048, + 0.011007952503859997, + 0.01337029691785574, + 0.008169947192072868, + 0.02329518459737301, + 0.021255185827612877, + 0.028746938332915306, + 0.01641164906322956, + 0.015027266927063465, + 0.01650293543934822, + 0.013456596061587334, + 0.011390835046768188, + 0.012575927190482616, + 0.012135914526879787, + 0.013955430127680302, + 0.012231587432324886, + 0.010827833786606789, + 0.010869864374399185, + 0.009489134885370731, + 0.007481995038688183, + 0.009150903671979904, + 0.008872347883880138, + 0.013420118018984795, + 0.010451947338879108, + 0.00822571199387312, + 0.009143774397671223, + 0.005722469184547663, + 0.00305133406072855, + 0.02052079513669014, + 0.00882040522992611, + 0.012973576784133911, + 0.010397721081972122, + 0.015486177057027817, + 0.00910616759210825, + 0.008333107456564903, + 0.005901934113353491, + 0.010751551017165184, + 0.01149902306497097, + 0.018915709108114243, + 0.012944883666932583, + 0.010816721245646477, + 0.022533701732754707, + 0.009359534829854965, + 0.007475097198039293, + 0.0053814430721104145, + 0.006274391897022724, + 0.004389119800180197, + 0.005163901951164007, + 0.004135345574468374, + 0.005839226301759481, + 0.006064955610781908, + 0.005725390277802944, + 0.004776515532284975, + 0.0040382277220487595, + 0.004878217354416847, + 0.0033149404916912317, + 0.003223754931241274, + 0.004329432267695665, + 0.003438555169850588, + 0.003177109407261014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.04865625500679016 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538148880004883 + ], + "opposite_logprob_tokens": [ + -0.42881494760513306 + ], + "logprob_reward": -1.0538148880004883, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0538148880004883, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17727957665920258, + 0.0899493396282196, + 0.24084614217281342, + 0.05439358204603195, + 0.07415870577096939, + 0.036162592470645905, + 0.06136761233210564, + 0.1684531718492508, + 0.05461039021611214, + 0.0427788570523262 + ], + "credit_map": [ + 0.02840881049633026, + 0.026725826784968376, + 0.021554188802838326, + 0.030286595225334167, + 0.023611726239323616, + 0.020351409912109375, + 0.01610511727631092, + 0.010235900059342384, + 0.013227204792201519, + 0.011037071235477924, + 0.017230218276381493, + 0.010514402762055397, + 0.010490274988114834, + 0.011424264870584011, + 0.00893463846296072, + 0.00709126191213727, + 0.06701023876667023, + 0.02926616184413433, + 0.06615442782640457, + 0.013443906791508198, + 0.021746698766946793, + 0.0207721795886755, + 0.012776227667927742, + 0.009676306508481503, + 0.007794208358973265, + 0.008856753818690777, + 0.009418883360922337, + 0.008207116276025772, + 0.005691081751137972, + 0.005117295775562525, + 0.0049372948706150055, + 0.004370949696749449, + 0.012680688872933388, + 0.011172918602824211, + 0.015891805291175842, + 0.00834366399794817, + 0.008900641463696957, + 0.006232660263776779, + 0.005405803211033344, + 0.0055305240675807, + 0.004588306415826082, + 0.003979007713496685, + 0.005450454540550709, + 0.005505091976374388, + 0.004846314899623394, + 0.005580702796578407, + 0.004283906891942024, + 0.0019288056064397097, + 0.00854438915848732, + 0.010955843143165112, + 0.008853648789227009, + 0.009108034893870354, + 0.008531407453119755, + 0.005804495885968208, + 0.00543162040412426, + 0.0041381726041436195, + 0.018278367817401886, + 0.0229930616915226, + 0.020083943381905556, + 0.0174417681992054, + 0.021181778982281685, + 0.03499751538038254, + 0.020473582670092583, + 0.013003146275877953, + 0.007231408264487982, + 0.008759346790611744, + 0.005483098328113556, + 0.00720997154712677, + 0.005644730757921934, + 0.00886967871338129, + 0.0064535499550402164, + 0.0049586063250899315, + 0.00712223444133997, + 0.004181806463748217, + 0.006764902733266354, + 0.01024144422262907, + 0.00351545843295753, + 0.004192143213003874, + 0.0037157097831368446, + 0.003045156830921769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -1.0538148880004883 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781880304217339 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781880304217339, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781880304217339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.225871279835701, + 0.049603160470724106, + 0.19756722450256348, + 0.05891023948788643, + 0.04400986433029175, + 0.03546040132641792, + 0.14732617139816284, + 0.14599131047725677, + 0.04734490439295769, + 0.04791540279984474 + ], + "credit_map": [ + 0.025849128141999245, + 0.04406391829252243, + 0.02552505023777485, + 0.041479598730802536, + 0.025671184062957764, + 0.0275091715157032, + 0.02234109863638878, + 0.013432142324745655, + 0.008403144776821136, + 0.007612780202180147, + 0.009035559371113777, + 0.004719727206975222, + 0.004468961618840694, + 0.006053222343325615, + 0.005449121817946434, + 0.00386064313352108, + 0.03692108392715454, + 0.027302857488393784, + 0.051523562520742416, + 0.013468229211866856, + 0.022505521774291992, + 0.02144823968410492, + 0.012794199399650097, + 0.011603537946939468, + 0.008395952172577381, + 0.00790251325815916, + 0.011864044703543186, + 0.007646224461495876, + 0.007284952327609062, + 0.005600483622401953, + 0.0046094791032373905, + 0.005606597289443016, + 0.006675452459603548, + 0.006627976428717375, + 0.009313768707215786, + 0.0060127172619104385, + 0.004010621923953295, + 0.004592043347656727, + 0.003736679209396243, + 0.0030406038276851177, + 0.005107048898935318, + 0.003551182569935918, + 0.006047613453119993, + 0.005206969100981951, + 0.005914749577641487, + 0.004193530417978764, + 0.0032803392969071865, + 0.002158966613933444, + 0.036686889827251434, + 0.015816912055015564, + 0.021941572427749634, + 0.01904057338833809, + 0.023580633103847504, + 0.011220416985452175, + 0.011401906609535217, + 0.007637254893779755, + 0.016692491248250008, + 0.021424980834126472, + 0.017192402854561806, + 0.011802117340266705, + 0.016885118559002876, + 0.03643800690770149, + 0.014161495491862297, + 0.011394699104130268, + 0.007164497394114733, + 0.0052427505142986774, + 0.006483552977442741, + 0.009481020271778107, + 0.004438156262040138, + 0.005943837575614452, + 0.004430405329912901, + 0.004160682205110788, + 0.0093081621453166, + 0.004031280055642128, + 0.013831586576998234, + 0.009102055802941322, + 0.002694279421120882, + 0.0033534590620547533, + 0.003629673970863223, + 0.0019649025052785873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.009781880304217339 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "sweeping, sweeping 360-degree shot, shot from a slightly upward angle, focuses on an eagle perched atop a precarious cliff. The golden late afternoon sun illuminates the jagged, weathered gray rock face. Below, a vast, bottomless canyon stretches, and distant mountains loom through the mist. The eagle's plumage is dark brown, almost black, and its crest on the back of its neck gleams golden-brown in the light. Its powerful yellow talons grip the rock, while its sharp amber eyes warily survey its territory. As the camera steadily rotates around it, the eagle suddenly unfurls its massive wings, each feather visible, its wingspan nearly obscuring a corner of the sky. With its head held high and its sharp, hooked yellow beak extended, it lets out a piercing cry that echoes through the empty canyon. The sunlight outlines a dazzling golden edge on the dark feathers of the eagle, and also makes the rough texture of the rock become three-dimensional and distinct. The whole picture, in warm tones, through slow and continuous circular motion, fully displays the majestic posture of this bird of prey and the magnificent natural scenery around it, creating an epic feeling of grandeur, aloofness and primitive vitality.", + "seed": 3231, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000010/group_9/index_3801_seed_3231_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14128630524501204, + "vlm_logprob": -0.14128630524501204, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.14128630524501204, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018566157668828964 + ], + "opposite_logprob_tokens": [ + -4.143566131591797 + ], + "logprob_reward": -0.018566157668828964, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018566157668828964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03871817886829376, + 0.07157639414072037, + 0.27240604162216187, + 0.03472210094332695, + 0.16168899834156036, + 0.016518566757440567, + 0.014990992844104767, + 0.012936359271407127, + 0.2501131296157837, + 0.12632925808429718 + ], + "credit_map": [ + 0.00469833193346858, + 0.006298671010881662, + 0.005365713499486446, + 0.0037296866066753864, + 0.0028595016337931156, + 0.005852015223354101, + 0.002757155569270253, + 0.007157104555517435, + 0.010372700169682503, + 0.014060045592486858, + 0.00841041374951601, + 0.005019540898501873, + 0.00558199267834425, + 0.018148234114050865, + 0.003996622748672962, + 0.005986850243061781, + 0.07014699280261993, + 0.0708831176161766, + 0.034518394619226456, + 0.017134586349129677, + 0.01570737361907959, + 0.04160295054316521, + 0.011965672485530376, + 0.010446922853589058, + 0.0034869201481342316, + 0.00427585281431675, + 0.00569863012060523, + 0.004886616952717304, + 0.00319272605702281, + 0.0061490293592214584, + 0.003249005414545536, + 0.003783316584303975, + 0.022154372185468674, + 0.012398852035403252, + 0.039422690868377686, + 0.004201916046440601, + 0.0030902992002665997, + 0.06965813785791397, + 0.004953423049300909, + 0.005809304770082235, + 0.001766058849170804, + 0.0019830302335321903, + 0.002694793278351426, + 0.002109204651787877, + 0.001945726340636611, + 0.00238432246260345, + 0.0016603850526735187, + 0.0019750462379306555, + 0.0010733974631875753, + 0.002436312148347497, + 0.002013847464695573, + 0.0015154461143538356, + 0.0015846365131437778, + 0.0028150449506938457, + 0.001499488833360374, + 0.0020528181921690702, + 0.001696962513960898, + 0.00259214174002409, + 0.0019076085882261395, + 0.001240959856659174, + 0.0013413239503279328, + 0.0014667471405118704, + 0.0014614961110055447, + 0.0012291192542761564, + 0.02604340948164463, + 0.019618907943367958, + 0.040027253329753876, + 0.019192509353160858, + 0.02722179889678955, + 0.035735834389925, + 0.019823450595140457, + 0.06244996190071106, + 0.019571524113416672, + 0.015231306664645672, + 0.018128296360373497, + 0.010024791583418846, + 0.019150981679558754, + 0.011980506591498852, + 0.013984186574816704, + 0.018257655203342438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.018566157668828964 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822757162153721 + ], + "opposite_logprob_tokens": [ + -4.018227577209473 + ], + "logprob_reward": -0.01822757162153721, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822757162153721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05301986262202263, + 0.13409681618213654, + 0.05496370419859886, + 0.10643359273672104, + 0.2554493248462677, + 0.022626163437962532, + 0.14310356974601746, + 0.030651791021227837, + 0.06968262046575546, + 0.12997247278690338 + ], + "credit_map": [ + 0.006478749681264162, + 0.009747833013534546, + 0.006774465087801218, + 0.004891143646091223, + 0.0054186987690627575, + 0.009043948724865913, + 0.0039197541773319244, + 0.006745263934135437, + 0.019927533343434334, + 0.027400044724345207, + 0.01711510680615902, + 0.012510803528130054, + 0.00933423638343811, + 0.02876320853829384, + 0.008763279765844345, + 0.010282600298523903, + 0.005812942981719971, + 0.006542588118463755, + 0.0082142548635602, + 0.0033886926248669624, + 0.004812674596905708, + 0.015527511946856976, + 0.003819367615506053, + 0.006845673080533743, + 0.01602492667734623, + 0.013305569998919964, + 0.015164929442107677, + 0.01089938823133707, + 0.013791187666356564, + 0.021606553345918655, + 0.0076794871129095554, + 0.007961555384099483, + 0.05696936696767807, + 0.01706416718661785, + 0.041159361600875854, + 0.005863504484295845, + 0.007248034700751305, + 0.11319257318973541, + 0.00677006971091032, + 0.007182256318628788, + 0.0017443555407226086, + 0.002296315971761942, + 0.002656526630744338, + 0.003273305017501116, + 0.0030753249302506447, + 0.0034686110448092222, + 0.003190812189131975, + 0.0029209137428551912, + 0.012240279465913773, + 0.01940336637198925, + 0.02451745979487896, + 0.011005125939846039, + 0.011495670303702354, + 0.03855089470744133, + 0.009321680292487144, + 0.016569094732403755, + 0.0043023088946938515, + 0.005236300174146891, + 0.004775629378855228, + 0.003163200104609132, + 0.002956117270514369, + 0.003603170393034816, + 0.002978866221383214, + 0.003636198351159692, + 0.0071071479469537735, + 0.0046980613842606544, + 0.011261978186666965, + 0.006205372512340546, + 0.006456600036472082, + 0.007395431865006685, + 0.005854093935340643, + 0.020703930407762527, + 0.015822017565369606, + 0.014811527915298939, + 0.023696107789874077, + 0.010798290371894836, + 0.016362264752388, + 0.013895242474973202, + 0.015601174905896187, + 0.018985841423273087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.01822757162153721 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370651230216026 + ], + "opposite_logprob_tokens": [ + -3.4087064266204834 + ], + "logprob_reward": -0.03370651230216026, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370651230216026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03887151554226875, + 0.045702382922172546, + 0.13083456456661224, + 0.03250718116760254, + 0.18689970672130585, + 0.03589697182178497, + 0.08390691131353378, + 0.017581988126039505, + 0.2768099009990692, + 0.15098881721496582 + ], + "credit_map": [ + 0.0038909022696316242, + 0.005723905749619007, + 0.005385236814618111, + 0.0045102378353476524, + 0.003338742535561323, + 0.006129442248493433, + 0.003513074479997158, + 0.006379979196935892, + 0.007026453036814928, + 0.01146608404815197, + 0.005547087639570236, + 0.003203712636604905, + 0.0031416569836437702, + 0.009060155600309372, + 0.0026961707044392824, + 0.003561068791896105, + 0.036559801548719406, + 0.036013517528772354, + 0.016901597380638123, + 0.008483811281621456, + 0.007571727968752384, + 0.015184137970209122, + 0.005758250597864389, + 0.004361734725534916, + 0.0033678689505904913, + 0.003194759599864483, + 0.005070798099040985, + 0.003399932524189353, + 0.00388324330560863, + 0.008082368411123753, + 0.002535433741286397, + 0.002972780726850033, + 0.04102875292301178, + 0.012281154282391071, + 0.032232701778411865, + 0.003595510497689247, + 0.00408635288476944, + 0.08510255068540573, + 0.003656521439552307, + 0.004916179459542036, + 0.0028327363543212414, + 0.0038329772651195526, + 0.0037601811345666647, + 0.005878538824617863, + 0.005250276066362858, + 0.005688969977200031, + 0.00447947159409523, + 0.004173824563622475, + 0.005387279205024242, + 0.011097418144345284, + 0.01102541759610176, + 0.008702291175723076, + 0.009424019604921341, + 0.01633892022073269, + 0.011600113473832607, + 0.010331460274755955, + 0.0028179893270134926, + 0.0012900022557005286, + 0.002532021375373006, + 0.0014621560694649816, + 0.001500646467320621, + 0.002421630546450615, + 0.0027487515471875668, + 0.002808792982250452, + 0.03018326312303543, + 0.02241262048482895, + 0.04310949891805649, + 0.020833492279052734, + 0.02792368270456791, + 0.043049536645412445, + 0.021927716210484505, + 0.06737010926008224, + 0.01961197517812252, + 0.0151680251583457, + 0.02574457786977291, + 0.010916649363934994, + 0.019630467519164085, + 0.017509274184703827, + 0.01760142855346203, + 0.024806438013911247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.03370651230216026 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.63274085521698 + ], + "opposite_logprob_tokens": [ + -0.75774085521698 + ], + "logprob_reward": -0.63274085521698, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.63274085521698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04052899405360222, + 0.026373447850346565, + 0.0830991342663765, + 0.045048873871564865, + 0.36300450563430786, + 0.02971004880964756, + 0.3130871057510376, + 0.028840987011790276, + 0.03419370576739311, + 0.036113202571868896 + ], + "credit_map": [ + 0.004496279638260603, + 0.007310230750590563, + 0.005937742069363594, + 0.0037188054993748665, + 0.002487302990630269, + 0.0043537793681025505, + 0.0031229981686919928, + 0.00910185743123293, + 0.0034322261344641447, + 0.0045035709626972675, + 0.0034239469096064568, + 0.0021921528968960047, + 0.002482551848515868, + 0.005961657501757145, + 0.0016632262850180268, + 0.0027141147293150425, + 0.022997045889496803, + 0.023561112582683563, + 0.010429846122860909, + 0.005362342111766338, + 0.005198427475988865, + 0.008458245545625687, + 0.0036870520561933517, + 0.003405066207051277, + 0.005254873540252447, + 0.005381040275096893, + 0.006054166704416275, + 0.004384069237858057, + 0.006042949389666319, + 0.011003941297531128, + 0.0032223444432020187, + 0.0037054899148643017, + 0.047562263906002045, + 0.027099084109067917, + 0.09267326444387436, + 0.009192219004034996, + 0.006327828858047724, + 0.1583470106124878, + 0.010774852707982063, + 0.01102799829095602, + 0.003102790331467986, + 0.003673020750284195, + 0.0037668521981686354, + 0.0035885442048311234, + 0.0030679828487336636, + 0.005210452247411013, + 0.003194307442754507, + 0.0041060978546738625, + 0.02700604312121868, + 0.043700624257326126, + 0.050626736134290695, + 0.023269226774573326, + 0.026918774470686913, + 0.08709369599819183, + 0.017315838485956192, + 0.03715618699789047, + 0.004839487839490175, + 0.004875681363046169, + 0.003975683823227882, + 0.002920826431363821, + 0.002858392195776105, + 0.0029126321896910667, + 0.002663973718881607, + 0.0037943064235150814, + 0.0034142720978707075, + 0.0028206834103912115, + 0.004652871284633875, + 0.003110695630311966, + 0.003899801755324006, + 0.0038829604163765907, + 0.0050787897780537605, + 0.007333629764616489, + 0.005078176502138376, + 0.003365198615938425, + 0.006768084596842527, + 0.0029938146471977234, + 0.0032453236635774374, + 0.004636415746062994, + 0.0035897409543395042, + 0.006436447612941265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.63274085521698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an eagle perched on a cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage dark brown or black with a golden-brown crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady 360-degree rotation around the eagle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast canyon with distant mountains visible through mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190429415553808 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.003190429415553808, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190429415553808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06325378268957138, + 0.22063486278057098, + 0.03163718059659004, + 0.033067021518945694, + 0.47064390778541565, + 0.028647491708397865, + 0.07038255780935287, + 0.02066832408308983, + 0.025552038103342056, + 0.03551284968852997 + ], + "credit_map": [ + 0.007259025238454342, + 0.013244664296507835, + 0.008822323754429817, + 0.006096257362514734, + 0.005470641423016787, + 0.01032997202128172, + 0.004560387227684259, + 0.007470507174730301, + 0.03606202453374863, + 0.04883452132344246, + 0.027214907109737396, + 0.016095293685793877, + 0.015130537562072277, + 0.04832102358341217, + 0.01257331296801567, + 0.01640324853360653, + 0.004270477220416069, + 0.004338119179010391, + 0.004809388890862465, + 0.0019052246352657676, + 0.002429818734526634, + 0.009089306928217411, + 0.00202327873557806, + 0.0027715677861124277, + 0.003745263908058405, + 0.004475794266909361, + 0.005255299154669046, + 0.003611702937632799, + 0.0034057199954986572, + 0.0069009605795145035, + 0.002249276963993907, + 0.003423003014177084, + 0.050946835428476334, + 0.03691380098462105, + 0.12790943682193756, + 0.01340730581432581, + 0.008149223402142525, + 0.19935694336891174, + 0.016610993072390556, + 0.01734933629631996, + 0.0032687117345631123, + 0.0035529721062630415, + 0.0036035897210240364, + 0.0046858456917107105, + 0.003237892873585224, + 0.004506936762481928, + 0.0029754014685750008, + 0.002816140651702881, + 0.005302655976265669, + 0.00985664688050747, + 0.008566825650632381, + 0.005224414635449648, + 0.005302833393216133, + 0.02258836105465889, + 0.005132750142365694, + 0.008408070541918278, + 0.0031522808130830526, + 0.0023631718941032887, + 0.003103091148659587, + 0.002043704502284527, + 0.0019333937671035528, + 0.003127814969047904, + 0.0016731198411434889, + 0.0032717438880354166, + 0.002968043554574251, + 0.002579589607194066, + 0.003818673314526677, + 0.0025682009290903807, + 0.001979715656489134, + 0.004040352534502745, + 0.0026578791439533234, + 0.004939584527164698, + 0.003924175165593624, + 0.0031216463539749384, + 0.005138732027262449, + 0.0031889919191598892, + 0.004574708174914122, + 0.004309753887355328, + 0.006313140969723463, + 0.004941698629409075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.003190429415553808 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_0/index_tensor([5451, 5451])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_0/index_tensor([5451, 5451])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8469116cd802f5b2d0ae5fcd65ea214ecbaa40ff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_0/index_tensor([5451, 5451])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely slow pull-focus shot begins, initially focusing precisely on a sunflower in the foreground with an extremely shallow depth of field. It's a perfect flower in full bloom, its broad, unfurling golden petals velvety in the soft morning light, each petal delicately textured at the edges. At the center, a deep brown disk is dotted with tiny seeds arranged in a Fibonacci spiral. A drop of glistening morning dew hangs from the tip of one petal, reflecting the blurred background. The background is now a dreamy bokeh of soft yellows, hazy greens, and the pale blue of the distant sky, brimming with light and color reminiscent of an Impressionist painting. As the lens's focus begins to smoothly and steadily shift back, the sharp, clear sunflower in the foreground gradually loses its outline, softening and falling out of focus, while the blurred blocks of color in the background begin to coalesce. Finally, the entire vast sunflower field is clearly displayed before the eyes. Thousands of sunflowers form a golden wave, their faces all facing the direction of the rising sun. The green stems and broad leaves form a rich sense of layering in the sea of flowers. The distant horizon connects with the blue sky. The entire picture transitions from a focus on microscopic beauty to admiration for the grand natural landscape, creating a tranquil, magnificent and vibrant visual atmosphere.", + "seed": 5374, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_0/index_5451_seed_5374_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.033586347149685024, + "vlm_logprob": -0.033586347149685024, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.033586347149685024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003465125570073724 + ], + "opposite_logprob_tokens": [ + -6.503465175628662 + ], + "logprob_reward": -0.003465125570073724, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.003465125570073724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.191039577126503, + 0.306571900844574, + 0.047397855669260025, + 0.07988674193620682, + 0.020667733624577522, + 0.013717143796384335, + 0.13570396602153778, + 0.04886344075202942, + 0.1438191533088684, + 0.012332489714026451 + ], + "credit_map": [ + 0.020408013835549355, + 0.021483847871422768, + 0.030152028426527977, + 0.01655488647520542, + 0.019175510853528976, + 0.038788363337516785, + 0.016871804371476173, + 0.027605101466178894, + 0.030130166560411453, + 0.10259822756052017, + 0.03403029963374138, + 0.037324946373701096, + 0.020091302692890167, + 0.016942406073212624, + 0.02886989153921604, + 0.03658464923501015, + 0.0034415321424603462, + 0.01565600372850895, + 0.0025937205646187067, + 0.009357107803225517, + 0.003357364796102047, + 0.0034555583260953426, + 0.004792331717908382, + 0.004744230769574642, + 0.013835684396326542, + 0.007724208291620016, + 0.00897241197526455, + 0.023909373208880424, + 0.005741661414504051, + 0.005579340271651745, + 0.007322251796722412, + 0.006801800802350044, + 0.0035998623352497816, + 0.002127581974491477, + 0.0029010975267738104, + 0.002464447170495987, + 0.0025372160598635674, + 0.0018911045044660568, + 0.0024978078436106443, + 0.002648614812642336, + 0.002168519888073206, + 0.0012640905333682895, + 0.002187756821513176, + 0.0017996380338445306, + 0.0021757143549621105, + 0.0014398215571418405, + 0.0012409612536430359, + 0.0014406403061002493, + 0.013318118639290333, + 0.004948448855429888, + 0.03708113729953766, + 0.017108840867877007, + 0.005788553971797228, + 0.0043314783833920956, + 0.03981081396341324, + 0.013316559605300426, + 0.0032162268180400133, + 0.002728306455537677, + 0.011938422918319702, + 0.006404982414096594, + 0.004387872759252787, + 0.0028343265876173973, + 0.012745662592351437, + 0.004607632290571928, + 0.00773822795599699, + 0.011368422769010067, + 0.020181061699986458, + 0.017748303711414337, + 0.027816949412226677, + 0.03273436054587364, + 0.01099441759288311, + 0.01523739006370306, + 0.0013450153637677431, + 0.0011122203432023525, + 0.0017466482240706682, + 0.001932143117301166, + 0.001338460948318243, + 0.0013374501140788198, + 0.002018947387114167, + 0.0015016027027741075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.003465125570073724 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018250280991196632 + ], + "opposite_logprob_tokens": [ + -4.018250465393066 + ], + "logprob_reward": -0.018250280991196632, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018250280991196632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22080183029174805, + 0.5263755321502686, + 0.04056936502456665, + 0.02922876924276352, + 0.023492900654673576, + 0.01875017024576664, + 0.020056599751114845, + 0.0366436243057251, + 0.07058961689472198, + 0.013491583056747913 + ], + "credit_map": [ + 0.01829676702618599, + 0.04131205752491951, + 0.030507579445838928, + 0.01931111328303814, + 0.021176131442189217, + 0.03803880140185356, + 0.02420150302350521, + 0.027957873418927193, + 0.045509543269872665, + 0.1893184334039688, + 0.04836178198456764, + 0.06387675553560257, + 0.036500465124845505, + 0.0301706250756979, + 0.040688954293727875, + 0.07194896787405014, + 0.0058371638879179955, + 0.0071592433378100395, + 0.005423196591436863, + 0.0042468649335205555, + 0.0054382928647100925, + 0.005455032456666231, + 0.0034466832876205444, + 0.0035628899931907654, + 0.005927998572587967, + 0.002875575330108404, + 0.003082870040088892, + 0.008449369110167027, + 0.0019915031734853983, + 0.0019378057913854718, + 0.0029199353884905577, + 0.002043710555881262, + 0.002990433946251869, + 0.002549346536397934, + 0.0032429045531898737, + 0.006530799902975559, + 0.0013391528045758605, + 0.0016319460701197386, + 0.0028690777253359556, + 0.0023392399307340384, + 0.0018957445863634348, + 0.0016555247129872441, + 0.0030287180561572313, + 0.004108520224690437, + 0.0013514107558876276, + 0.0020303926430642605, + 0.002364633372053504, + 0.0023152269423007965, + 0.0015766212018206716, + 0.0013090081047266722, + 0.0025389259681105614, + 0.004737405106425285, + 0.0009896476985886693, + 0.0008089686161838472, + 0.006470454391092062, + 0.0016255690716207027, + 0.0027224395889788866, + 0.0017263384070247412, + 0.005047002341598272, + 0.00813098531216383, + 0.0019125593826174736, + 0.0021046637557446957, + 0.012256043031811714, + 0.0027435901574790478, + 0.004174747969955206, + 0.004212938714772463, + 0.022685861214995384, + 0.009204215370118618, + 0.006839156616479158, + 0.007461795583367348, + 0.009766360744833946, + 0.0062445420771837234, + 0.0012386911548674107, + 0.001405311282724142, + 0.001981621142476797, + 0.0024708574637770653, + 0.0018503795145079494, + 0.0011004707776010036, + 0.0020989777985960245, + 0.0013452739221975207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.018250280991196632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032995045185089 + ], + "opposite_logprob_tokens": [ + -2.350329875946045 + ], + "logprob_reward": -0.10032995045185089, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032995045185089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30749839544296265, + 0.16341142356395721, + 0.19027027487754822, + 0.07719556242227554, + 0.034968484193086624, + 0.036940865218639374, + 0.06606703996658325, + 0.03906236216425896, + 0.06401720643043518, + 0.020568441599607468 + ], + "credit_map": [ + 0.024252992123365402, + 0.09953640401363373, + 0.03602875769138336, + 0.030209539458155632, + 0.03173171728849411, + 0.022886991500854492, + 0.03142179548740387, + 0.0314301997423172, + 0.012381265871226788, + 0.05230772867798805, + 0.013423563912510872, + 0.019656868651509285, + 0.013901842758059502, + 0.01385059766471386, + 0.011074136942625046, + 0.026815403252840042, + 0.015616019256412983, + 0.06190710514783859, + 0.01136303972452879, + 0.03733619675040245, + 0.01484584342688322, + 0.019275743514299393, + 0.014016184955835342, + 0.015910137444734573, + 0.009200746193528175, + 0.008788318373262882, + 0.007838606834411621, + 0.010289965197443962, + 0.011198141612112522, + 0.007944018580019474, + 0.01156006008386612, + 0.010375699028372765, + 0.0031297216191887856, + 0.003474342869594693, + 0.0039430046454072, + 0.007028911728411913, + 0.002026633359491825, + 0.0026385136879980564, + 0.007099573966115713, + 0.005627777427434921, + 0.00437844917178154, + 0.003967824392020702, + 0.006650819443166256, + 0.006951768416911364, + 0.002884424990043044, + 0.0037092063575983047, + 0.004428584594279528, + 0.003969782497733831, + 0.006991856265813112, + 0.0037760408595204353, + 0.0194704569876194, + 0.0073590753600001335, + 0.003444722620770335, + 0.0026492709293961525, + 0.015372112393379211, + 0.007003505248576403, + 0.00516129843890667, + 0.0026783240027725697, + 0.009090631268918514, + 0.004209588281810284, + 0.0024829150643199682, + 0.002878751838579774, + 0.008959075435996056, + 0.0036017769016325474, + 0.0037298521492630243, + 0.004228275269269943, + 0.01211728248745203, + 0.0051553635857999325, + 0.011088086292147636, + 0.013858986087143421, + 0.007464200723916292, + 0.006375152617692947, + 0.0021180652547627687, + 0.002112551359459758, + 0.0024568785447627306, + 0.00308542069979012, + 0.0018208574037998915, + 0.0035402236972004175, + 0.0026911459863185883, + 0.002743297955021262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.10032995045185089 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078105196356773 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.016078105196356773, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078105196356773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23876531422138214, + 0.057146910578012466, + 0.09858088195323944, + 0.18359428644180298, + 0.07088618725538254, + 0.05781814455986023, + 0.05695386230945587, + 0.04396947845816612, + 0.1681816577911377, + 0.02410331927239895 + ], + "credit_map": [ + 0.02199230156838894, + 0.05927615240216255, + 0.030180754140019417, + 0.021198753267526627, + 0.02295476756989956, + 0.03205600753426552, + 0.027027009055018425, + 0.024079563096165657, + 0.004345973953604698, + 0.015464412979781628, + 0.0076852054335176945, + 0.008287296630442142, + 0.0036996870767325163, + 0.0039010178297758102, + 0.008545774035155773, + 0.0052175456658005714, + 0.0080112898722291, + 0.03104275092482567, + 0.008812155574560165, + 0.015467804856598377, + 0.010070749558508396, + 0.012670089490711689, + 0.005507707130163908, + 0.0069983419962227345, + 0.03431941568851471, + 0.017890561372041702, + 0.018688945099711418, + 0.060783449560403824, + 0.011799125000834465, + 0.012268551625311375, + 0.01347049418836832, + 0.014373742043972015, + 0.008927540853619576, + 0.007593189831823111, + 0.008349339477717876, + 0.010065403766930103, + 0.007787768729031086, + 0.010698424652218819, + 0.00869867391884327, + 0.008765840902924538, + 0.008652491495013237, + 0.007776289712637663, + 0.008166703395545483, + 0.007979738526046276, + 0.00919504277408123, + 0.00535906245931983, + 0.005808974150568247, + 0.004879839718341827, + 0.004809721373021603, + 0.0021175118163228035, + 0.011404742486774921, + 0.010131064802408218, + 0.0033895547967404127, + 0.002755049616098404, + 0.01677950657904148, + 0.00556670967489481, + 0.0027727538254112005, + 0.004041068721562624, + 0.010644561611115932, + 0.004886326380074024, + 0.0045403060503304005, + 0.00302490103058517, + 0.009468009695410728, + 0.004591550678014755, + 0.009101655334234238, + 0.022739103063941002, + 0.0373648963868618, + 0.025404350832104683, + 0.030811846256256104, + 0.013576206751167774, + 0.015702024102211, + 0.01348157785832882, + 0.0028579600621014833, + 0.0017728953389450908, + 0.0035036259796470404, + 0.002910217270255089, + 0.0026206383481621742, + 0.004004208836704493, + 0.004161198623478413, + 0.002272574696689844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.016078105196356773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808273538947105 + ], + "opposite_logprob_tokens": [ + -3.529808282852173 + ], + "logprob_reward": -0.029808273538947105, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808273538947105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2873638868331909, + 0.07757188379764557, + 0.03389594703912735, + 0.041716866195201874, + 0.04077070951461792, + 0.031156154349446297, + 0.0632808655500412, + 0.2294122278690338, + 0.1718749701976776, + 0.02295641601085663 + ], + "credit_map": [ + 0.03156358003616333, + 0.08847315609455109, + 0.03211308270692825, + 0.027060044929385185, + 0.02961793914437294, + 0.021611077710986137, + 0.028813382610678673, + 0.028111634775996208, + 0.006939226295799017, + 0.021332457661628723, + 0.005222815554589033, + 0.008131456561386585, + 0.00690721208229661, + 0.007186507806181908, + 0.007137749809771776, + 0.014714467339217663, + 0.004067263565957546, + 0.004616585094481707, + 0.0038051027804613113, + 0.0033201007172465324, + 0.004614442121237516, + 0.0052892123349010944, + 0.004133606795221567, + 0.004049636423587799, + 0.004694425500929356, + 0.0046368553303182125, + 0.005081054754555225, + 0.008071839809417725, + 0.004707466345280409, + 0.00389673188328743, + 0.005009882152080536, + 0.005618618335574865, + 0.008342980407178402, + 0.003839037613943219, + 0.005648002494126558, + 0.005291115026921034, + 0.0028847060166299343, + 0.0038531494792550802, + 0.005103664007037878, + 0.005808057729154825, + 0.0051911561749875546, + 0.003597184084355831, + 0.004685546737164259, + 0.005502516869455576, + 0.0024486121255904436, + 0.0030502204317599535, + 0.003424060298129916, + 0.0032568613532930613, + 0.007351510692387819, + 0.004831642843782902, + 0.008571983315050602, + 0.011338639073073864, + 0.0036769560538232327, + 0.0032300788443535566, + 0.01745973341166973, + 0.006820326205343008, + 0.009882026351988316, + 0.009433737024664879, + 0.014042070135474205, + 0.06966942548751831, + 0.011837980709969997, + 0.011137409135699272, + 0.08826690167188644, + 0.015142709016799927, + 0.01009121723473072, + 0.011152154766023159, + 0.025396136566996574, + 0.01657208986580372, + 0.03440656512975693, + 0.03954705968499184, + 0.01542892400175333, + 0.019280843436717987, + 0.001623769523575902, + 0.0016467543318867683, + 0.002500326605513692, + 0.0032548371236771345, + 0.0035689647775143385, + 0.004670772701501846, + 0.003167587099596858, + 0.0025234066415578127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.029808273538947105 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely slow pull-focus shot begins, initially focusing precisely on a sunflower in the foreground with an extremely shallow depth of field. It's a perfect flower in full bloom, its broad, unfurling golden petals velvety in the soft morning light, each petal delicately textured at the edges. At the center, a deep brown disk is dotted with tiny seeds arranged in a Fibonacci spiral. A drop of glistening morning dew hangs from the tip of one petal, reflecting the blurred background. The background is now a dreamy bokeh of soft yellows, hazy greens, and the pale blue of the distant sky, brimming with light and color reminiscent of an Impressionist painting. As the lens's focus begins to smoothly and steadily shift back, the sharp, clear sunflower in the foreground gradually loses its outline, softening and falling out of focus, while the blurred blocks of color in the background begin to coalesce. Finally, the entire vast sunflower field is clearly displayed before the eyes. Thousands of sunflowers form a golden wave, their faces all facing the direction of the rising sun. The green stems and broad leaves form a rich sense of layering in the sea of flowers. The distant horizon connects with the blue sky. The entire picture transitions from a focus on microscopic beauty to admiration for the grand natural landscape, creating a tranquil, magnificent and vibrant visual atmosphere.", + "seed": 5375, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_0/index_5451_seed_5375_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.031366859562695026, + "vlm_logprob": -0.031366859562695026, + "vlm_yes_no_margin": 4.599999904632568, + "vlm_token_credit": -0.031366859562695026, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0020573418587446213 + ], + "opposite_logprob_tokens": [ + -8.002057075500488 + ], + "logprob_reward": -0.0020573418587446213, + "yes_no_margin_reward": 7.999999523162842, + "token_credit_reward": -0.0020573418587446213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12353887408971786, + 0.4438631236553192, + 0.15818580985069275, + 0.0588504821062088, + 0.040157251060009, + 0.04764846712350845, + 0.0325046107172966, + 0.0345580093562603, + 0.021366441622376442, + 0.03932695835828781 + ], + "credit_map": [ + 0.0077523114159703255, + 0.007677693851292133, + 0.0344720259308815, + 0.016809118911623955, + 0.007842420600354671, + 0.03231305629014969, + 0.009530398063361645, + 0.007141856476664543, + 0.036929044872522354, + 0.08792264014482498, + 0.12924595177173615, + 0.08063472807407379, + 0.026820307597517967, + 0.030857041478157043, + 0.03116181492805481, + 0.020291585475206375, + 0.02028614841401577, + 0.017703069373965263, + 0.018914157524704933, + 0.021612562239170074, + 0.01365978829562664, + 0.021264299750328064, + 0.027736850082874298, + 0.017008939757943153, + 0.005963329691439867, + 0.01143637578934431, + 0.006430135108530521, + 0.008108559064567089, + 0.004882661625742912, + 0.008781255222856998, + 0.0063685523346066475, + 0.006879616063088179, + 0.0028354169335216284, + 0.004691082052886486, + 0.007786201313138008, + 0.005269394256174564, + 0.0031372569501399994, + 0.006016512401401997, + 0.004955927841365337, + 0.0054654586128890514, + 0.0030857298988848925, + 0.0036299433559179306, + 0.007010172586888075, + 0.004819035995751619, + 0.002395043848082423, + 0.0028232086915522814, + 0.015990497544407845, + 0.007894831709563732, + 0.0036032088100910187, + 0.0033558872528374195, + 0.003373038722202182, + 0.0047922683879733086, + 0.002964884275570512, + 0.002691039117053151, + 0.00676875002682209, + 0.004955535288900137, + 0.004250876605510712, + 0.0025580863002687693, + 0.0062276278622448444, + 0.005436336621642113, + 0.0025673790369182825, + 0.0026562053244560957, + 0.005302196368575096, + 0.0055593024007976055, + 0.002650683745741844, + 0.0026208360213786364, + 0.003202500520274043, + 0.002667153486981988, + 0.0025753641966730356, + 0.0017061237012967467, + 0.0026430687867105007, + 0.003300710814073682, + 0.005309565924108028, + 0.0026996999513357878, + 0.004752609878778458, + 0.005422254092991352, + 0.0029259216971695423, + 0.0018153693526983261, + 0.011323133483529091, + 0.00507840421050787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0020573418587446213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012574084103107452 + ], + "opposite_logprob_tokens": [ + -4.387574195861816 + ], + "logprob_reward": -0.012574084103107452, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012574084103107452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1055179089307785, + 0.6496113538742065, + 0.03492068126797676, + 0.07387858629226685, + 0.015394385904073715, + 0.06032602861523628, + 0.018620528280735016, + 0.015918582677841187, + 0.011639958247542381, + 0.014172004535794258 + ], + "credit_map": [ + 0.007880686782300472, + 0.009523874148726463, + 0.033408086746931076, + 0.015038198791444302, + 0.008070039562880993, + 0.012057532556355, + 0.009700302965939045, + 0.009839195758104324, + 0.07386727631092072, + 0.2620016038417816, + 0.07306399196386337, + 0.04578197002410889, + 0.07676047831773758, + 0.06572176516056061, + 0.033351145684719086, + 0.019063135609030724, + 0.004188536200672388, + 0.0033260469790548086, + 0.004466507118195295, + 0.004170352127403021, + 0.003023941768333316, + 0.006706612650305033, + 0.0041845375671982765, + 0.004854149185121059, + 0.005535022355616093, + 0.008231575600802898, + 0.008385044522583485, + 0.021324846893548965, + 0.004926450550556183, + 0.0065039354376494884, + 0.004567866213619709, + 0.014403843320906162, + 0.0011670378735288978, + 0.0016488570254296064, + 0.0031954252626746893, + 0.002085550222545862, + 0.0013177964137867093, + 0.001775495707988739, + 0.0023020023945719004, + 0.0019022229826077819, + 0.002736383117735386, + 0.001919344998896122, + 0.022968342527747154, + 0.004430185537785292, + 0.0019347129855304956, + 0.0029441192746162415, + 0.005803869571536779, + 0.01758906990289688, + 0.0022498571779578924, + 0.0012322966940701008, + 0.0021089264191687107, + 0.0027363246772438288, + 0.0013031166745349765, + 0.00127730134408921, + 0.003441848326474428, + 0.004270857200026512, + 0.0022618600632995367, + 0.0008884877897799015, + 0.0024071340449154377, + 0.002674770075827837, + 0.0008612595847807825, + 0.000947149412240833, + 0.003075335407629609, + 0.0028025880455970764, + 0.0018457622500136495, + 0.0011303164064884186, + 0.0018342406256124377, + 0.001423420151695609, + 0.0010462678037583828, + 0.0008930829353630543, + 0.0019293638179078698, + 0.0015375047223642468, + 0.0022274660877883434, + 0.0011348309926688671, + 0.0018982469337061048, + 0.0019060203339904547, + 0.001291635213419795, + 0.0008258198504336178, + 0.0032308166846632957, + 0.0016571695450693369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.012574084103107452 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293905228376389 + ], + "opposite_logprob_tokens": [ + -2.2379391193389893 + ], + "logprob_reward": -0.11293905228376389, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293905228376389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18539035320281982, + 0.5295697450637817, + 0.11762602627277374, + 0.044135887175798416, + 0.05176546797156334, + 0.020394589751958847, + 0.015148788690567017, + 0.010457759723067284, + 0.009462342597544193, + 0.016049040481448174 + ], + "credit_map": [ + 0.011811343021690845, + 0.010839538648724556, + 0.0463305301964283, + 0.0243154838681221, + 0.012960098683834076, + 0.0559057742357254, + 0.01376277208328247, + 0.009464802220463753, + 0.04257868602871895, + 0.11563852429389954, + 0.14947831630706787, + 0.0876302570104599, + 0.034926146268844604, + 0.03901608660817146, + 0.03720666095614433, + 0.023095062002539635, + 0.016151996329426765, + 0.01402421947568655, + 0.014111530035734177, + 0.01670789159834385, + 0.010014815255999565, + 0.015119890682399273, + 0.01940508745610714, + 0.012090601027011871, + 0.005358655471354723, + 0.005226363893598318, + 0.005762571003288031, + 0.010009344667196274, + 0.002923445776104927, + 0.004477858077734709, + 0.005080146715044975, + 0.005297502037137747, + 0.003684357972815633, + 0.006586391478776932, + 0.006489863619208336, + 0.006857589818537235, + 0.0054107289761304855, + 0.009363541379570961, + 0.006394583731889725, + 0.0069784135557711124, + 0.0022328062914311886, + 0.0009962364565581083, + 0.002024741843342781, + 0.0013607749715447426, + 0.0009805939625948668, + 0.0013054521987214684, + 0.005433134734630585, + 0.006060849875211716, + 0.001736538833938539, + 0.001331509556621313, + 0.0022067835088819265, + 0.002308144001290202, + 0.0016344166360795498, + 0.0011519452091306448, + 0.0018766821594908834, + 0.002902768086642027, + 0.0015123033663257957, + 0.0008693434647284448, + 0.0015332924667745829, + 0.0014389952411875129, + 0.0013048115652054548, + 0.0008973198710009456, + 0.0014383831294253469, + 0.0014633104437962174, + 0.0014003138057887554, + 0.0007616957300342619, + 0.0015154544962570071, + 0.0016269293846562505, + 0.0007593248155899346, + 0.0007031408604234457, + 0.0014090873301029205, + 0.0012863967567682266, + 0.0019929336849600077, + 0.0013933012960478663, + 0.0019155240152031183, + 0.0022128985729068518, + 0.0014511216431856155, + 0.0009843219304457307, + 0.00412383209913969, + 0.0019751060754060745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.11293905228376389 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110772168263793 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.0110772168263793, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110772168263793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18491536378860474, + 0.12836532294750214, + 0.17404909431934357, + 0.15078110992908478, + 0.06927438080310822, + 0.16022498905658722, + 0.048342015594244, + 0.03047131560742855, + 0.01997232437133789, + 0.03360407426953316 + ], + "credit_map": [ + 0.010848535224795341, + 0.019989795982837677, + 0.048262860625982285, + 0.028253566473722458, + 0.013780470937490463, + 0.0343819335103035, + 0.016699293628335, + 0.012698892503976822, + 0.011633853428065777, + 0.031689565628767014, + 0.02762591280043125, + 0.019825367256999016, + 0.009586644358932972, + 0.011374123394489288, + 0.009955774992704391, + 0.00667409785091877, + 0.02262287214398384, + 0.0202200748026371, + 0.02266886830329895, + 0.025924645364284515, + 0.0126746641471982, + 0.02515318989753723, + 0.026541661471128464, + 0.01824311539530754, + 0.012536856345832348, + 0.038373928517103195, + 0.014538745395839214, + 0.023249177262187004, + 0.015581964515149593, + 0.022570423781871796, + 0.01392369158565998, + 0.010006320662796497, + 0.007912056520581245, + 0.007948830723762512, + 0.010867494158446789, + 0.010626155883073807, + 0.006033886224031448, + 0.008832466788589954, + 0.008037114515900612, + 0.009016375988721848, + 0.0074982959777116776, + 0.004550351295620203, + 0.050614140927791595, + 0.011263036169111729, + 0.004591108299791813, + 0.006312408950179815, + 0.015446314588189125, + 0.059949327260255814, + 0.010150355286896229, + 0.005570080131292343, + 0.0061498056165874004, + 0.006207473110407591, + 0.0035418933257460594, + 0.003703043796122074, + 0.0067644184455275536, + 0.006254944950342178, + 0.003325005294755101, + 0.002802580129355192, + 0.003306192345917225, + 0.004171520005911589, + 0.005278393160551786, + 0.0023745896760374308, + 0.005581728182733059, + 0.003631305182352662, + 0.004648089874535799, + 0.0017754127038642764, + 0.002835177583619952, + 0.0028034751303493977, + 0.0017156510148197412, + 0.0015775178326293826, + 0.0024137087166309357, + 0.0022032917477190495, + 0.004632560070604086, + 0.00285966950468719, + 0.004208809696137905, + 0.00446348125115037, + 0.0032316772267222404, + 0.0020208884961903095, + 0.008328701369464397, + 0.003858286654576659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0110772168263793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186602741479874 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186602741479874, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186602741479874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10846652090549469, + 0.38386788964271545, + 0.04188993200659752, + 0.22620758414268494, + 0.03910093754529953, + 0.04274455085396767, + 0.05557823181152344, + 0.03221069276332855, + 0.024329030886292458, + 0.04560462757945061 + ], + "credit_map": [ + 0.007681269198656082, + 0.006639513187110424, + 0.029218360781669617, + 0.010709117166697979, + 0.00769864022731781, + 0.033512409776449203, + 0.0071436031721532345, + 0.005863613449037075, + 0.030654776841402054, + 0.08045703172683716, + 0.1097664013504982, + 0.06654544174671173, + 0.02471759356558323, + 0.02732190676033497, + 0.027251102030277252, + 0.017153630033135414, + 0.006279775407165289, + 0.0044521670788526535, + 0.004975415766239166, + 0.004776119254529476, + 0.0036648716777563095, + 0.005958095658570528, + 0.006662154570221901, + 0.005121331196278334, + 0.017696909606456757, + 0.027798211202025414, + 0.035768501460552216, + 0.06243642419576645, + 0.01284561026841402, + 0.02300841547548771, + 0.013435904867947102, + 0.03321761637926102, + 0.0032917240168899298, + 0.003594285110011697, + 0.005454078316688538, + 0.0055862669833004475, + 0.004537180997431278, + 0.005527036730200052, + 0.0051207635551691055, + 0.005989603698253632, + 0.0037599357310682535, + 0.0028974898159503937, + 0.011077942326664925, + 0.004777838476002216, + 0.0028452433180063963, + 0.002138101030141115, + 0.0029413565061986446, + 0.01230664923787117, + 0.009809560142457485, + 0.0040331026539206505, + 0.0059149679727852345, + 0.007320881355553865, + 0.005230187438428402, + 0.003187075722962618, + 0.01030512060970068, + 0.009777337312698364, + 0.003621606854721904, + 0.0021979708690196276, + 0.005261758808046579, + 0.005201642867177725, + 0.0032158803660422564, + 0.0018907799385488033, + 0.005043712444603443, + 0.0057773408479988575, + 0.0021040283609181643, + 0.0023743482306599617, + 0.0035063850227743387, + 0.0038911167066544294, + 0.0020587309263646603, + 0.001650615013204515, + 0.004904329776763916, + 0.003839476965367794, + 0.004479998257011175, + 0.002944162581115961, + 0.005221976898610592, + 0.008249963633716106, + 0.0027611874975264072, + 0.002003681380301714, + 0.014298617839813232, + 0.005645041354000568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.018186602741479874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely slow pull-focus shot begins, initially focusing precisely on a sunflower in the foreground with an extremely shallow depth of field. It's a perfect flower in full bloom, its broad, unfurling golden petals velvety in the soft morning light, each petal delicately textured at the edges. At the center, a deep brown disk is dotted with tiny seeds arranged in a Fibonacci spiral. A drop of glistening morning dew hangs from the tip of one petal, reflecting the blurred background. The background is now a dreamy bokeh of soft yellows, hazy greens, and the pale blue of the distant sky, brimming with light and color reminiscent of an Impressionist painting. As the lens's focus begins to smoothly and steadily shift back, the sharp, clear sunflower in the foreground gradually loses its outline, softening and falling out of focus, while the blurred blocks of color in the background begin to coalesce. Finally, the entire vast sunflower field is clearly displayed before the eyes. Thousands of sunflowers form a golden wave, their faces all facing the direction of the rising sun. The green stems and broad leaves form a rich sense of layering in the sea of flowers. The distant horizon connects with the blue sky. The entire picture transitions from a focus on microscopic beauty to admiration for the grand natural landscape, creating a tranquil, magnificent and vibrant visual atmosphere.", + "seed": 5376, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_0/index_5451_seed_5376_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05783239626325667, + "vlm_logprob": -0.05783239626325667, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.05783239626325667, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025889708194881678 + ], + "opposite_logprob_tokens": [ + -7.877588748931885 + ], + "logprob_reward": -0.0025889708194881678, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0025889708194881678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05302969738841057, + 0.15667687356472015, + 0.0441322848200798, + 0.05224224179983139, + 0.02881597727537155, + 0.05375697836279869, + 0.04137145355343819, + 0.027600498870015144, + 0.044493332505226135, + 0.49788063764572144 + ], + "credit_map": [ + 0.00687780836597085, + 0.003976122941821814, + 0.017674773931503296, + 0.004132834263145924, + 0.005448866169899702, + 0.004094752483069897, + 0.003333890112116933, + 0.007490651216357946, + 0.015625877305865288, + 0.05593778192996979, + 0.01211579144001007, + 0.007904238998889923, + 0.012255754321813583, + 0.00870752613991499, + 0.009577659890055656, + 0.03455225005745888, + 0.005978938192129135, + 0.004515702836215496, + 0.00655697425827384, + 0.00315248966217041, + 0.004296512342989445, + 0.008654464967548847, + 0.004540637601166964, + 0.006436570547521114, + 0.008498743176460266, + 0.007353512104600668, + 0.00870204996317625, + 0.004081239923834801, + 0.004142229910939932, + 0.0069530862383544445, + 0.0041006398387253284, + 0.008410744369029999, + 0.005204332061111927, + 0.003005434526130557, + 0.0033001962583512068, + 0.0024778989609330893, + 0.0029712372925132513, + 0.00260570808313787, + 0.0061667682603001595, + 0.003084404394030571, + 0.0029348607640713453, + 0.018285933881998062, + 0.01799129694700241, + 0.003800581442192197, + 0.0025785332545638084, + 0.001637447508983314, + 0.004001914989203215, + 0.002526417374610901, + 0.004614759236574173, + 0.0062909964472055435, + 0.008698779158294201, + 0.009953673928976059, + 0.0030185922514647245, + 0.0027769850566983223, + 0.0025225423742085695, + 0.003495129058137536, + 0.0047422731295228004, + 0.002219210844486952, + 0.0038082909304648638, + 0.004829173441976309, + 0.002558404579758644, + 0.002243144903331995, + 0.00393642857670784, + 0.0032635745592415333, + 0.005068106576800346, + 0.0038263117894530296, + 0.007344373967498541, + 0.004251915495842695, + 0.004622481297701597, + 0.002722380915656686, + 0.004764216486364603, + 0.011893549002707005, + 0.03561808168888092, + 0.09656959772109985, + 0.060072772204875946, + 0.07201399654150009, + 0.035605572164058685, + 0.07912524044513702, + 0.031085016205906868, + 0.08779039233922958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.0025889708194881678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014242897741496563 + ], + "opposite_logprob_tokens": [ + -4.264243125915527 + ], + "logprob_reward": -0.014242897741496563, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014242897741496563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04905074089765549, + 0.06368578970432281, + 0.01662791520357132, + 0.054921627044677734, + 0.6080573201179504, + 0.0845344066619873, + 0.047533947974443436, + 0.017608487978577614, + 0.0341181643307209, + 0.023861626163125038 + ], + "credit_map": [ + 0.006212274543941021, + 0.0028925444930791855, + 0.007087800186127424, + 0.0026199501007795334, + 0.008322758600115776, + 0.006835694424808025, + 0.004615744575858116, + 0.01046398002654314, + 0.005985582713037729, + 0.007666515186429024, + 0.007636894006282091, + 0.00510991457849741, + 0.008133304305374622, + 0.009474135935306549, + 0.006338316947221756, + 0.013341128826141357, + 0.002374574076384306, + 0.0021178608294576406, + 0.0019194153137505054, + 0.0011355163296684623, + 0.0018674333114176989, + 0.0037843864411115646, + 0.0015088632935658097, + 0.0019198671216145158, + 0.009913311339914799, + 0.006506319623440504, + 0.009424162097275257, + 0.003232619957998395, + 0.004518840461969376, + 0.007629807572811842, + 0.0037121714558452368, + 0.009984401054680347, + 0.029862329363822937, + 0.03948119655251503, + 0.027105268090963364, + 0.09447472542524338, + 0.05525081977248192, + 0.02205962873995304, + 0.27486521005630493, + 0.06495818495750427, + 0.004150473512709141, + 0.03045819327235222, + 0.029786575585603714, + 0.005395740270614624, + 0.0038035211618989706, + 0.0033130054362118244, + 0.0036957720294594765, + 0.003931129816919565, + 0.003965939860790968, + 0.007054117973893881, + 0.01156568806618452, + 0.013517038896679878, + 0.0026850623544305563, + 0.002635406097397208, + 0.0027927865739911795, + 0.0033179125748574734, + 0.0035710809752345085, + 0.0021742030512541533, + 0.0025756964460015297, + 0.0024650453124195337, + 0.002130059525370598, + 0.0010845959186553955, + 0.0019776395056396723, + 0.0016301677096635103, + 0.006942181382328272, + 0.0025495211593806744, + 0.006507907062768936, + 0.0032788196112960577, + 0.0015403697034344077, + 0.0022342177107930183, + 0.002270692028105259, + 0.008794459514319897, + 0.002877109916880727, + 0.00402358453720808, + 0.003366952994838357, + 0.0030886335298419, + 0.0017386259278282523, + 0.0029717227444052696, + 0.0016449225367978215, + 0.004150077700614929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.014242897741496563 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558435797691345 + ], + "opposite_logprob_tokens": [ + -1.6005843877792358 + ], + "logprob_reward": -0.22558435797691345, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558435797691345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10342994332313538, + 0.06047753989696503, + 0.06515368819236755, + 0.04052071273326874, + 0.4592389166355133, + 0.14992092549800873, + 0.02962239272892475, + 0.0457710437476635, + 0.015320111066102982, + 0.030544757843017578 + ], + "credit_map": [ + 0.010877116583287716, + 0.008183819241821766, + 0.025792470201849937, + 0.007048708386719227, + 0.012543341144919395, + 0.013749980367720127, + 0.006856546271592379, + 0.01837795600295067, + 0.0062204524874687195, + 0.018758678808808327, + 0.010687661357223988, + 0.003776564495638013, + 0.004627315327525139, + 0.006063438951969147, + 0.00379374623298645, + 0.006549679674208164, + 0.008865359239280224, + 0.006694470066577196, + 0.008487392216920853, + 0.004493966698646545, + 0.006294380873441696, + 0.014915978536009789, + 0.006770569831132889, + 0.008631561882793903, + 0.006346371024847031, + 0.0046318876557052135, + 0.007088056765496731, + 0.0025979899801313877, + 0.003409275785088539, + 0.005698434542864561, + 0.0029865035321563482, + 0.007762191817164421, + 0.02054237388074398, + 0.029159322381019592, + 0.01745631732046604, + 0.069979228079319, + 0.04270626977086067, + 0.01593259908258915, + 0.20942404866218567, + 0.05403869226574898, + 0.006086564157158136, + 0.05382189899682999, + 0.06089207902550697, + 0.008890127763152122, + 0.005193987861275673, + 0.003471981268376112, + 0.007057690992951393, + 0.004506582859903574, + 0.002822048030793667, + 0.00434029009193182, + 0.006373824086040258, + 0.007028484717011452, + 0.0017391643486917019, + 0.0016702599823474884, + 0.0021253882441669703, + 0.0035229292698204517, + 0.004186366684734821, + 0.00568639999255538, + 0.007960159331560135, + 0.007270906586199999, + 0.0077469986863434315, + 0.0027571995742619038, + 0.005470483098179102, + 0.004692522343248129, + 0.0010708378395065665, + 0.0012230562279000878, + 0.0029603238217532635, + 0.0026571250054985285, + 0.0012948537478223443, + 0.0009184273658320308, + 0.0017348896944895387, + 0.003460596315562725, + 0.004712637048214674, + 0.004798576235771179, + 0.0038475068286061287, + 0.004175765905529261, + 0.0023379286285489798, + 0.004329531453549862, + 0.001982548274099827, + 0.004360260907560587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.22558435797691345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645012974739075 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008645012974739075, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645012974739075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030438922345638275, + 0.07195128500461578, + 0.01894577220082283, + 0.06688045710325241, + 0.3783433437347412, + 0.06363068521022797, + 0.06550059467554092, + 0.066016785800457, + 0.03366696462035179, + 0.2046252191066742 + ], + "credit_map": [ + 0.0037408003117889166, + 0.0029517842922359705, + 0.007405225653201342, + 0.002340389182791114, + 0.0034786819014698267, + 0.0029836436733603477, + 0.002179930917918682, + 0.005358463153243065, + 0.008752578869462013, + 0.01634855568408966, + 0.010661562904715538, + 0.0035431061405688524, + 0.005410970188677311, + 0.006602213252335787, + 0.00371592678129673, + 0.016916362568736076, + 0.0022390824742615223, + 0.002071832539513707, + 0.0034743803553283215, + 0.002029560972005129, + 0.0017193673411384225, + 0.002789052901789546, + 0.0018063668394461274, + 0.0028161273803561926, + 0.011996193788945675, + 0.009763042442500591, + 0.009777240455150604, + 0.005871464032679796, + 0.005885165650397539, + 0.00588288763538003, + 0.006534133106470108, + 0.01117032952606678, + 0.01919223740696907, + 0.025440087541937828, + 0.0164658036082983, + 0.05901859328150749, + 0.0337059460580349, + 0.014013277366757393, + 0.17114131152629852, + 0.03936607390642166, + 0.0035978234373033047, + 0.02058551460504532, + 0.019913824275135994, + 0.0036544152535498142, + 0.0034724732395261526, + 0.004041658248752356, + 0.003457678249105811, + 0.004907292779535055, + 0.007548190653324127, + 0.008330714888870716, + 0.010237828828394413, + 0.007609354332089424, + 0.010180137120187283, + 0.00786159373819828, + 0.004922613967210054, + 0.008810156024992466, + 0.007627510000020266, + 0.007898950949311256, + 0.010086272843182087, + 0.014402472414076328, + 0.006285077892243862, + 0.004967641085386276, + 0.0072495946660637856, + 0.007499262224882841, + 0.005083846393972635, + 0.0031761829741299152, + 0.0049758413806557655, + 0.005872226320207119, + 0.002818198874592781, + 0.0020117193926125765, + 0.005255009979009628, + 0.004473936278373003, + 0.013663462363183498, + 0.040015265345573425, + 0.024998893961310387, + 0.030112283304333687, + 0.013930955901741982, + 0.032290685921907425, + 0.012403542175889015, + 0.0372101254761219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.008645012974739075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810074180364609 + ], + "opposite_logprob_tokens": [ + -3.2881007194519043 + ], + "logprob_reward": -0.03810074180364609, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810074180364609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016370302066206932, + 0.03757938742637634, + 0.024858230724930763, + 0.014267852529883385, + 0.30122658610343933, + 0.0358661450445652, + 0.031097007915377617, + 0.06518065929412842, + 0.09452929347753525, + 0.37902459502220154 + ], + "credit_map": [ + 0.0033128981012851, + 0.001088631572201848, + 0.0022976635955274105, + 0.0009435749379917979, + 0.001194033888168633, + 0.002015601610764861, + 0.0011880932142958045, + 0.0043298047967255116, + 0.00254248408600688, + 0.01046160701662302, + 0.003115120343863964, + 0.0019387133652344346, + 0.002647555200383067, + 0.010468197055161, + 0.002198943169787526, + 0.004206766840070486, + 0.0031248501036316156, + 0.0028028362430632114, + 0.0035715557169169188, + 0.0020694760605692863, + 0.0029369418043643236, + 0.004942023195326328, + 0.0025672351475805044, + 0.002843310823664069, + 0.002297233557328582, + 0.0016486616805195808, + 0.002471372252330184, + 0.0012897291453555226, + 0.0009880784200504422, + 0.002021102700382471, + 0.0012880220310762525, + 0.002263653092086315, + 0.01645527221262455, + 0.019941797479987144, + 0.014702988788485527, + 0.04744492098689079, + 0.026619572192430496, + 0.01160765066742897, + 0.13456112146377563, + 0.02989327348768711, + 0.0024747326970100403, + 0.011240249499678612, + 0.011840208433568478, + 0.002324344590306282, + 0.001717729726806283, + 0.0014946549199521542, + 0.002543642884120345, + 0.002230584155768156, + 0.003513491014018655, + 0.003322253469377756, + 0.00484742596745491, + 0.005773794837296009, + 0.002066778251901269, + 0.0033190895337611437, + 0.0029283317271620035, + 0.00532584497705102, + 0.01551064569503069, + 0.009006486274302006, + 0.008283369243144989, + 0.00867476500570774, + 0.0071096536703407764, + 0.004611602984368801, + 0.006205613259226084, + 0.005778519902378321, + 0.015628855675458908, + 0.008402644656598568, + 0.015510899014770985, + 0.011916679330170155, + 0.0061180745251476765, + 0.005344068631529808, + 0.011166519485414028, + 0.020441556349396706, + 0.026101835072040558, + 0.07543883472681046, + 0.045425895601511, + 0.0549597442150116, + 0.025914406403899193, + 0.06060181185603142, + 0.022860942408442497, + 0.06772114336490631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.03810074180364609 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely slow pull-focus shot begins, initially focusing precisely on a sunflower in the foreground with an extremely shallow depth of field. It's a perfect flower in full bloom, its broad, unfurling golden petals velvety in the soft morning light, each petal delicately textured at the edges. At the center, a deep brown disk is dotted with tiny seeds arranged in a Fibonacci spiral. A drop of glistening morning dew hangs from the tip of one petal, reflecting the blurred background. The background is now a dreamy bokeh of soft yellows, hazy greens, and the pale blue of the distant sky, brimming with light and color reminiscent of an Impressionist painting. As the lens's focus begins to smoothly and steadily shift back, the sharp, clear sunflower in the foreground gradually loses its outline, softening and falling out of focus, while the blurred blocks of color in the background begin to coalesce. Finally, the entire vast sunflower field is clearly displayed before the eyes. Thousands of sunflowers form a golden wave, their faces all facing the direction of the rising sun. The green stems and broad leaves form a rich sense of layering in the sea of flowers. The distant horizon connects with the blue sky. The entire picture transitions from a focus on microscopic beauty to admiration for the grand natural landscape, creating a tranquil, magnificent and vibrant visual atmosphere.", + "seed": 5377, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_0/index_5451_seed_5377_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024815944489091635, + "vlm_logprob": -0.024815944489091635, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.024815944489091635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003400735557079315 + ], + "opposite_logprob_tokens": [ + -8.503400802612305 + ], + "logprob_reward": -0.003400735557079315, + "yes_no_margin_reward": 8.5, + "token_credit_reward": -0.003400735557079315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03806816786527634, + 0.17869076132774353, + 0.028517691418528557, + 0.07237699627876282, + 0.07884584367275238, + 0.10073432326316833, + 0.06973041594028473, + 0.04887726530432701, + 0.3334493637084961, + 0.050709184259176254 + ], + "credit_map": [ + 0.004698076285421848, + 0.004374682903289795, + 0.005943493451923132, + 0.0047563957050442696, + 0.004206954501569271, + 0.0033480587881058455, + 0.0040496294386684895, + 0.006690880283713341, + 0.0048224180936813354, + 0.006846176460385323, + 0.011147705838084221, + 0.008265387266874313, + 0.008657886646687984, + 0.008085647597908974, + 0.06606250256299973, + 0.06480305641889572, + 0.003421676578000188, + 0.0025922288186848164, + 0.0029852455481886864, + 0.0023529091849923134, + 0.003542401595041156, + 0.0046750218607485294, + 0.0026986172888427973, + 0.006249594502151012, + 0.007421545684337616, + 0.011346301063895226, + 0.009018897078931332, + 0.00733636599034071, + 0.007607026491314173, + 0.0043713669292628765, + 0.010764123871922493, + 0.014511379413306713, + 0.015719082206487656, + 0.006143118720501661, + 0.01770191267132759, + 0.007923095487058163, + 0.005062102805823088, + 0.005367258097976446, + 0.010601813904941082, + 0.01032746210694313, + 0.009900061413645744, + 0.0041186194866895676, + 0.030389904975891113, + 0.01212610024958849, + 0.004194417968392372, + 0.004434001166373491, + 0.01266444195061922, + 0.022906776517629623, + 0.01227850653231144, + 0.0037703446578234434, + 0.01317392848432064, + 0.0065290555357933044, + 0.005312756635248661, + 0.0034850656520575285, + 0.009388021193444729, + 0.015792742371559143, + 0.004621730651706457, + 0.002285753609612584, + 0.012750674039125443, + 0.003169916570186615, + 0.003324930090457201, + 0.005505193490535021, + 0.0053754569962620735, + 0.011843616142868996, + 0.018393145874142647, + 0.010019557550549507, + 0.11231561750173569, + 0.015008939430117607, + 0.008225073106586933, + 0.015852708369493484, + 0.026552405208349228, + 0.1270819455385208, + 0.002392389578744769, + 0.0025006860960274935, + 0.017388347536325455, + 0.003478980390354991, + 0.0020868394058197737, + 0.0018686404218897223, + 0.01297610905021429, + 0.008017194457352161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.003400735557079315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009819536469876766 + ], + "opposite_logprob_tokens": [ + -4.634819507598877 + ], + "logprob_reward": -0.009819536469876766, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009819536469876766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27730998396873474, + 0.27876797318458557, + 0.018419088795781136, + 0.030484924092888832, + 0.03451039269566536, + 0.038769785314798355, + 0.08606641739606857, + 0.05535493046045303, + 0.1464996039867401, + 0.03381684049963951 + ], + "credit_map": [ + 0.029153848066926003, + 0.03737334534525871, + 0.04828687012195587, + 0.02753896825015545, + 0.027677295729517937, + 0.02944256365299225, + 0.030981620773673058, + 0.04685547575354576, + 0.009822502732276917, + 0.010640686377882957, + 0.014896037988364697, + 0.013033642433583736, + 0.014037120155990124, + 0.013867020606994629, + 0.10147953033447266, + 0.10099142789840698, + 0.0021988400258123875, + 0.002000760054215789, + 0.002438727766275406, + 0.0017170669743791223, + 0.001823291415348649, + 0.0018370908219367266, + 0.002659277059137821, + 0.003744034096598625, + 0.003943160641938448, + 0.00548618845641613, + 0.003894941648468375, + 0.003801066428422928, + 0.0026149728801101446, + 0.0027282992377877235, + 0.004131045192480087, + 0.003885249374434352, + 0.003619031049311161, + 0.0038425009697675705, + 0.005217592231929302, + 0.0033827966544777155, + 0.003820652374997735, + 0.0028493269346654415, + 0.0052695125341415405, + 0.006508977618068457, + 0.0039002213161438704, + 0.0024445559829473495, + 0.00960117019712925, + 0.004996476229280233, + 0.0018494181567803025, + 0.0019317694241181016, + 0.005369785241782665, + 0.008676386438310146, + 0.008143316954374313, + 0.003119814209640026, + 0.019400600343942642, + 0.00505299586802721, + 0.0053518773056566715, + 0.006303868256509304, + 0.011721412651240826, + 0.02697252854704857, + 0.00546500226482749, + 0.0024550140369683504, + 0.013146650046110153, + 0.0034786406904459, + 0.00580954272300005, + 0.008724072948098183, + 0.006536759436130524, + 0.009739246219396591, + 0.008009937591850758, + 0.005332267377525568, + 0.040908899158239365, + 0.006333062890917063, + 0.0039715697057545185, + 0.00833950750529766, + 0.016379479318857193, + 0.05722486972808838, + 0.0024896860122680664, + 0.001604210352525115, + 0.007597158197313547, + 0.0017566279275342822, + 0.001851238077506423, + 0.0020210177171975374, + 0.01084485836327076, + 0.0056520444341003895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.009819536469876766 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903542906045914 + ], + "opposite_logprob_tokens": [ + -2.4640355110168457 + ], + "logprob_reward": -0.08903542906045914, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903542906045914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18531668186187744, + 0.21433193981647491, + 0.10221540927886963, + 0.1951073855161667, + 0.13595502078533173, + 0.04486384242773056, + 0.025401314720511436, + 0.02919926866889, + 0.03716597333550453, + 0.030443178489804268 + ], + "credit_map": [ + 0.018141353502869606, + 0.02540970966219902, + 0.033488690853118896, + 0.020761093124747276, + 0.017499413341283798, + 0.01635325886309147, + 0.02183317206799984, + 0.03182997554540634, + 0.007121149450540543, + 0.008797494694590569, + 0.015002021566033363, + 0.010248559527099133, + 0.010286093689501286, + 0.015590967610478401, + 0.07435458898544312, + 0.07293105870485306, + 0.008119530975818634, + 0.0073049599304795265, + 0.010098868981003761, + 0.008043580688536167, + 0.007133266888558865, + 0.012237686663866043, + 0.02595863863825798, + 0.023318879306316376, + 0.016448497772216797, + 0.022867536172270775, + 0.023447373881936073, + 0.017656318843364716, + 0.02305263839662075, + 0.01761016435921192, + 0.03152649849653244, + 0.04249836504459381, + 0.019505873322486877, + 0.010409905575215816, + 0.017154039815068245, + 0.019013280048966408, + 0.011997018940746784, + 0.008677510544657707, + 0.022178545594215393, + 0.0270188357681036, + 0.0048683625645935535, + 0.001765055232681334, + 0.013567890040576458, + 0.00484426086768508, + 0.0017817418556660414, + 0.0021340532694011927, + 0.006202294025570154, + 0.00970018282532692, + 0.0040686363354325294, + 0.0011791075812652707, + 0.00491937855258584, + 0.002062680898234248, + 0.002128619933500886, + 0.0021568001247942448, + 0.0030298931524157524, + 0.00585619593039155, + 0.0024617272429168224, + 0.001039199880324304, + 0.009215888567268848, + 0.0015856276731938124, + 0.001546557410620153, + 0.0023037088103592396, + 0.003017340088263154, + 0.008029220625758171, + 0.001994143472984433, + 0.0020018168725073338, + 0.0062364451587200165, + 0.0017718849703669548, + 0.002344910055398941, + 0.004740630276501179, + 0.0060576521791517735, + 0.012018490582704544, + 0.0018047640332952142, + 0.0019923876971006393, + 0.003291880479082465, + 0.002288301708176732, + 0.0023231692612171173, + 0.0022790124639868736, + 0.007309325505048037, + 0.009154337458312511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.08903542906045914 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630014326423407 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007630014326423407, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630014326423407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07141277194023132, + 0.3113947808742523, + 0.03455866500735283, + 0.1124449297785759, + 0.0653456449508667, + 0.029949937015771866, + 0.1710728406906128, + 0.1322777420282364, + 0.044683538377285004, + 0.026859169825911522 + ], + "credit_map": [ + 0.007072844542562962, + 0.010176623240113258, + 0.01040208525955677, + 0.007785613648593426, + 0.007146618328988552, + 0.00949840433895588, + 0.007894660346210003, + 0.011435925029218197, + 0.008988180197775364, + 0.01228682417422533, + 0.01825683005154133, + 0.015009548515081406, + 0.014686474576592445, + 0.012660464271903038, + 0.11542241275310516, + 0.11408405005931854, + 0.0027377023361623287, + 0.0018662960501387715, + 0.0037743982393294573, + 0.002710688393563032, + 0.002216739347204566, + 0.0022393600083887577, + 0.011282619088888168, + 0.0077308607287704945, + 0.012132622301578522, + 0.019081251695752144, + 0.012880240567028522, + 0.011584705673158169, + 0.012189396657049656, + 0.008215612731873989, + 0.016959672793745995, + 0.019401423633098602, + 0.006632007658481598, + 0.0054031796753406525, + 0.007947788573801517, + 0.008425239473581314, + 0.007337419781833887, + 0.005413517355918884, + 0.010647296905517578, + 0.013539195992052555, + 0.0028877221047878265, + 0.0026260893791913986, + 0.004972358699887991, + 0.00413915840908885, + 0.0027736590709537268, + 0.0027049435302615166, + 0.005633972585201263, + 0.004212032072246075, + 0.017400357872247696, + 0.009166182018816471, + 0.032564692199230194, + 0.011900600045919418, + 0.012990793213248253, + 0.018813632428646088, + 0.023554448038339615, + 0.044682133942842484, + 0.013227906078100204, + 0.00742035498842597, + 0.028097135946154594, + 0.006779126822948456, + 0.017065437510609627, + 0.02451237291097641, + 0.014520042575895786, + 0.020655367523431778, + 0.003063753480091691, + 0.0021161457989364862, + 0.014489787630736828, + 0.002378592500463128, + 0.0018090787343680859, + 0.0037134536541998386, + 0.0049534947611391544, + 0.01215923298150301, + 0.002005057642236352, + 0.0022727635223418474, + 0.003195850644260645, + 0.0017395088216289878, + 0.0017695907736197114, + 0.002723964862525463, + 0.007549123372882605, + 0.005603311583399773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.007630014326423407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sunflower shown in full bloom with golden petals and a deep brown center containing a visible spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of glistening dew visible hanging from the tip of one of the sunflower's petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a smooth pull-focus shot that shifts focus from the foreground sunflower to the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end by revealing a vast field of thousands of sunflowers facing the rising sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194007031619549 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014194007031619549, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194007031619549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058221228420734406, + 0.2464732676744461, + 0.014668967574834824, + 0.033098623156547546, + 0.02114001475274563, + 0.1456354856491089, + 0.0688110738992691, + 0.2188733071088791, + 0.13629886507987976, + 0.0567791610956192 + ], + "credit_map": [ + 0.006631257012486458, + 0.007917962968349457, + 0.010730409994721413, + 0.005932965315878391, + 0.0054683685302734375, + 0.004882322158664465, + 0.006822926923632622, + 0.00983501598238945, + 0.007043336518108845, + 0.009808006696403027, + 0.015726277604699135, + 0.011712748557329178, + 0.011464258655905724, + 0.01036124862730503, + 0.09049218893051147, + 0.08986521512269974, + 0.0013483985094353557, + 0.0011175026884302497, + 0.0014601120492443442, + 0.0013520264765247703, + 0.0014513814821839333, + 0.0031489115208387375, + 0.0016772941453382373, + 0.003113341284915805, + 0.0033441558480262756, + 0.004833494778722525, + 0.003978627268224955, + 0.0035283968318253756, + 0.0033672440331429243, + 0.0024975070264190435, + 0.004883308429270983, + 0.006665888242423534, + 0.0036081820726394653, + 0.0016414289129897952, + 0.004161066375672817, + 0.0023719873279333115, + 0.0015077909920364618, + 0.002248754957690835, + 0.0026848679408431053, + 0.0029159386176615953, + 0.014988794922828674, + 0.0068418714217841625, + 0.0440276637673378, + 0.015871457755565643, + 0.005632001906633377, + 0.006425795610994101, + 0.019501276314258575, + 0.03234662860631943, + 0.009342223405838013, + 0.0032646069303154945, + 0.013201843947172165, + 0.0050961049273610115, + 0.0056731076911091805, + 0.005429571960121393, + 0.00974216777831316, + 0.01706145517528057, + 0.01999044045805931, + 0.011201868765056133, + 0.052864573895931244, + 0.010799792595207691, + 0.02422800101339817, + 0.035313501954078674, + 0.023201927542686462, + 0.04127321392297745, + 0.007213195785880089, + 0.0031365163158625364, + 0.046501755714416504, + 0.006218737456947565, + 0.0028304492589086294, + 0.005462965928018093, + 0.01129195187240839, + 0.05364328995347023, + 0.004386509768664837, + 0.005364051554352045, + 0.00928483810275793, + 0.003258721437305212, + 0.004346965812146664, + 0.006961789447814226, + 0.009154785424470901, + 0.014021508395671844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.014194007031619549 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_1/index_tensor([3744, 3744])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_1/index_tensor([3744, 3744])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6f02778a716d16509837940d2886432bbf1dbc4b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_1/index_tensor([3744, 3744])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, fluid, circular dolly shot begins, swirling at eye level around a vibrant cooking scene in a cozy, warmly lit kitchen, all rendered in the vibrant, captivating style of Pixar animation. At the centerpiece, a sturdy, amber-colored glass bottle of craft beer stands proudly, its surface rimmed with cool condensation. Its label is a playful, purely graphic design, depicting rolling hills and sunlight, saturated with saturated oranges and greens. Nearby, a cherry-red, enameled pot simmers gently on a gas stove, a thick, bubbling tomato sauce topped with distinctly diced onions and sweet bean paste, releasing a subtle burst of steam. The bottle and pan rest on a polished, light-colored wooden countertop, artfully stacked with an array of remarkably fresh ingredients: plump, glossy red tomatoes, a whole head of garlic with papery skin, and a sprig of bright green basil, all seemingly impossibly perfect. As the camera continues to swirl gracefully, an unseen hand, rendered in soft, stylized features, briefly enters the frame, stirring the sauce steadily with a wooden spoon in fluid, deliberate movements. The entire scene is bathed in a soft golden glow from overhead chandeliers, casting long, soft shadows that make the colors pop and create a cheerful and warm family atmosphere that celebrates the simple joy of preparing a delicious meal.", + "seed": 22921, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_1/index_3744_seed_22921_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0643165297806263, + "vlm_logprob": -0.0643165297806263, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.0643165297806263, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10191486030817032 + ], + "opposite_logprob_tokens": [ + -2.601914882659912 + ], + "logprob_reward": -0.10191486030817032, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.10191486030817032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05800054594874382, + 0.1168874204158783, + 0.08236521482467651, + 0.08184633404016495, + 0.061551760882139206, + 0.11818108707666397, + 0.13059864938259125, + 0.1166529431939125, + 0.0634494200348854, + 0.17046663165092468 + ], + "credit_map": [ + 0.007081582676619291, + 0.011235862970352173, + 0.004624511115252972, + 0.007270275615155697, + 0.007596455514431, + 0.00787610374391079, + 0.0061082118190824986, + 0.0062075424939394, + 0.009616376832127571, + 0.03525333106517792, + 0.009612247347831726, + 0.013912705704569817, + 0.014956693165004253, + 0.01648046262562275, + 0.00964671652764082, + 0.0074088857509195805, + 0.006620224565267563, + 0.011875258758664131, + 0.009368132799863815, + 0.010389831848442554, + 0.015345405787229538, + 0.011620718985795975, + 0.008480729535222054, + 0.008664911612868309, + 0.010790791362524033, + 0.010860058479011059, + 0.006576399318873882, + 0.008387279696762562, + 0.01671215519309044, + 0.012916159816086292, + 0.009290843270719051, + 0.006312649231404066, + 0.011338809505105019, + 0.008057433180510998, + 0.003657531226053834, + 0.007027829997241497, + 0.008045798167586327, + 0.008122161030769348, + 0.0053357528522610664, + 0.009966445155441761, + 0.012894040904939175, + 0.016568487510085106, + 0.010287362150847912, + 0.010183128528296947, + 0.02896548993885517, + 0.01382508222013712, + 0.013074187561869621, + 0.01238331664353609, + 0.010722296312451363, + 0.016821788623929024, + 0.013909251429140568, + 0.01211349107325077, + 0.025522829964756966, + 0.016173267737030983, + 0.02131967805325985, + 0.014016037806868553, + 0.011413217522203922, + 0.031039902940392494, + 0.007408075034618378, + 0.009320165030658245, + 0.02706786058843136, + 0.013098030351102352, + 0.008552544750273228, + 0.008753144182264805, + 0.0073992591351270676, + 0.009058735333383083, + 0.006821069400757551, + 0.0087081054225564, + 0.0069634695537388325, + 0.010520917363464832, + 0.008697387762367725, + 0.005280479323118925, + 0.018535031005740166, + 0.01607361063361168, + 0.014832992106676102, + 0.055838312953710556, + 0.016347842290997505, + 0.008091578260064125, + 0.028030499815940857, + 0.012716762721538544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.10191486030817032 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813975676894188 + ], + "opposite_logprob_tokens": [ + -3.288139820098877 + ], + "logprob_reward": -0.03813975676894188, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813975676894188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07773035019636154, + 0.023589078336954117, + 0.04405910521745682, + 0.08657462149858475, + 0.24083098769187927, + 0.04011266678571701, + 0.02681533806025982, + 0.3156460225582123, + 0.029218826442956924, + 0.11542303115129471 + ], + "credit_map": [ + 0.00958595797419548, + 0.01106469240039587, + 0.008040782064199448, + 0.0067842393182218075, + 0.011873018927872181, + 0.011281484737992287, + 0.012064368464052677, + 0.0070358081720769405, + 0.002405652077868581, + 0.004223262425512075, + 0.002547943266108632, + 0.002619182225316763, + 0.0030820544343441725, + 0.0031094190198928118, + 0.003727253759279847, + 0.001874312525615096, + 0.0028247074224054813, + 0.005137807223945856, + 0.006590140983462334, + 0.004923993721604347, + 0.00769397895783186, + 0.0062041860073804855, + 0.0056664296425879, + 0.005017859395593405, + 0.011513074859976768, + 0.008507601916790009, + 0.005453667603433132, + 0.009903918951749802, + 0.01855020970106125, + 0.014589178375899792, + 0.009998125024139881, + 0.008058836683630943, + 0.022731957957148552, + 0.041895199567079544, + 0.014043847098946571, + 0.021899472922086716, + 0.05426650121808052, + 0.033628933131694794, + 0.029225902631878853, + 0.02313915826380253, + 0.005129626952111721, + 0.005331134889274836, + 0.0058791483752429485, + 0.005999226123094559, + 0.005537808872759342, + 0.003880650270730257, + 0.00417014816775918, + 0.004184923600405455, + 0.0024435764644294977, + 0.00413349037989974, + 0.002440348733216524, + 0.0019075848395004869, + 0.005045290570706129, + 0.0030372554901987314, + 0.0041321623139083385, + 0.003675628686323762, + 0.026219896972179413, + 0.08440092206001282, + 0.018798107281327248, + 0.019454875960946083, + 0.07535198330879211, + 0.03896285966038704, + 0.0246847253292799, + 0.02777264267206192, + 0.004016298335045576, + 0.0036255151499062777, + 0.0029433080926537514, + 0.0026256025303155184, + 0.00313454563729465, + 0.004588007926940918, + 0.004428049549460411, + 0.0038574994541704655, + 0.014635652303695679, + 0.011331361718475819, + 0.011335695162415504, + 0.03588838130235672, + 0.009576654061675072, + 0.004553947132080793, + 0.018994523212313652, + 0.009106816723942757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.03813975676894188 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308246821165085 + ], + "opposite_logprob_tokens": [ + -3.1680824756622314 + ], + "logprob_reward": -0.04308246821165085, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308246821165085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07929717749357224, + 0.10086279362440109, + 0.07789250463247299, + 0.09833598136901855, + 0.20160290598869324, + 0.08383913338184357, + 0.05642394348978996, + 0.1784336119890213, + 0.034640226513147354, + 0.0886717215180397 + ], + "credit_map": [ + 0.011847501620650291, + 0.010001865215599537, + 0.007977204397320747, + 0.007023789454251528, + 0.010328500531613827, + 0.012232502922415733, + 0.013036639429628849, + 0.006849169731140137, + 0.00939644780009985, + 0.02120448835194111, + 0.010421688668429852, + 0.008831472136080265, + 0.014087609946727753, + 0.01371042151004076, + 0.016701210290193558, + 0.006509451661258936, + 0.006280103698372841, + 0.010593063198029995, + 0.009222714230418205, + 0.007407839410007, + 0.01214613951742649, + 0.011863972060382366, + 0.01238743495196104, + 0.007991240359842777, + 0.012567557394504547, + 0.008801927790045738, + 0.005951947532594204, + 0.011438566260039806, + 0.02046896517276764, + 0.016698915511369705, + 0.011833060532808304, + 0.01057504490017891, + 0.029596934095025063, + 0.03971008211374283, + 0.013723177835345268, + 0.014301368966698647, + 0.03889939561486244, + 0.029120653867721558, + 0.019414613023400307, + 0.01683669164776802, + 0.012079943902790546, + 0.015217244625091553, + 0.006451871711760759, + 0.008777111768722534, + 0.013332223519682884, + 0.009218466468155384, + 0.008939693681895733, + 0.009822584688663483, + 0.00496713537722826, + 0.006883631926029921, + 0.00860245805233717, + 0.005007500760257244, + 0.010301044210791588, + 0.00507014337927103, + 0.008824664168059826, + 0.006767367944121361, + 0.019400762394070625, + 0.03888824209570885, + 0.02430248074233532, + 0.0196445994079113, + 0.03302476555109024, + 0.012792414054274559, + 0.01917557790875435, + 0.011204767040908337, + 0.00379829085431993, + 0.004799126647412777, + 0.004892265889793634, + 0.0058136992156505585, + 0.0033085080794990063, + 0.0030829343013465405, + 0.005989633500576019, + 0.0029557710513472557, + 0.013694856315851212, + 0.009777942672371864, + 0.008218914270401001, + 0.019162792712450027, + 0.009753914549946785, + 0.00687070470303297, + 0.011445361189544201, + 0.009747236035764217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.04308246821165085 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035744309425354 + ], + "opposite_logprob_tokens": [ + -2.3503575325012207 + ], + "logprob_reward": -0.10035744309425354, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035744309425354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030086608603596687, + 0.040095504373311996, + 0.02190842479467392, + 0.04693056270480156, + 0.05576228350400925, + 0.1265135556459427, + 0.057757746428251266, + 0.1796426773071289, + 0.031484607607126236, + 0.4098180830478668 + ], + "credit_map": [ + 0.0027486735489219427, + 0.005641669500619173, + 0.0024362648837268353, + 0.0034968070685863495, + 0.0036437835078686476, + 0.004006090108305216, + 0.004956588614732027, + 0.0031567299738526344, + 0.004849638324230909, + 0.008669232949614525, + 0.0037755537778139114, + 0.003217503661289811, + 0.0072954995557665825, + 0.005275919567793608, + 0.003994107246398926, + 0.003018044400960207, + 0.002177025657147169, + 0.003435298800468445, + 0.003034881316125393, + 0.0022485805675387383, + 0.003273272654041648, + 0.002898440696299076, + 0.0029033171012997627, + 0.0019376041600480676, + 0.005777362268418074, + 0.0052691856399178505, + 0.0024311242159456015, + 0.005555536597967148, + 0.010615219362080097, + 0.007741695735603571, + 0.005196398589760065, + 0.004344038665294647, + 0.005246077198535204, + 0.008203490637242794, + 0.003432482248172164, + 0.006427912972867489, + 0.011481169611215591, + 0.007346938364207745, + 0.007755621802061796, + 0.0058685895055532455, + 0.019515102729201317, + 0.027268221601843834, + 0.01307759340852499, + 0.011092289350926876, + 0.023446790874004364, + 0.014385220594704151, + 0.007853497751057148, + 0.009874815121293068, + 0.004770117346197367, + 0.005560395307838917, + 0.008429143577814102, + 0.007916255854070187, + 0.00920359417796135, + 0.005279081407934427, + 0.01125947292894125, + 0.00533968023955822, + 0.030913839116692543, + 0.03996578976511955, + 0.01435431931167841, + 0.017320696264505386, + 0.032008785754442215, + 0.015540705062448978, + 0.01889561116695404, + 0.010642913170158863, + 0.0037113099824637175, + 0.0034749654587358236, + 0.003843666287139058, + 0.0035014126915484667, + 0.003364990232512355, + 0.005584015976637602, + 0.004664664622396231, + 0.003339582122862339, + 0.048077721148729324, + 0.04582454636693001, + 0.03856223449110985, + 0.09033655375242233, + 0.05302905663847923, + 0.029441500082612038, + 0.05909372493624687, + 0.04545273631811142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.10035744309425354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380881205201149 + ], + "opposite_logprob_tokens": [ + -3.288088083267212 + ], + "logprob_reward": -0.0380881205201149, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380881205201149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37895646691322327, + 0.0744754895567894, + 0.060734204947948456, + 0.0737236738204956, + 0.05401588976383209, + 0.0815034955739975, + 0.08127809315919876, + 0.03347108140587807, + 0.02185644581913948, + 0.1399850696325302 + ], + "credit_map": [ + 0.037649959325790405, + 0.0543738529086113, + 0.03755989670753479, + 0.04066646099090576, + 0.04674236848950386, + 0.05433187633752823, + 0.0583491250872612, + 0.0492829792201519, + 0.008422316052019596, + 0.0190077293664217, + 0.0057504503056406975, + 0.008693800307810307, + 0.009909795597195625, + 0.010994015261530876, + 0.006782539188861847, + 0.004914845805615187, + 0.0042639742605388165, + 0.0067701032385230064, + 0.005045661702752113, + 0.00698826415464282, + 0.010402417741715908, + 0.00893619004637003, + 0.012466656975448132, + 0.005860942415893078, + 0.009550848975777626, + 0.007778518367558718, + 0.004733718000352383, + 0.009529177099466324, + 0.016341716051101685, + 0.012535697780549526, + 0.008128696121275425, + 0.005125310737639666, + 0.00743755092844367, + 0.011189606972038746, + 0.003939530812203884, + 0.005641786381602287, + 0.010453948751091957, + 0.007627897430211306, + 0.0032339063473045826, + 0.0044916667975485325, + 0.008943991735577583, + 0.01438735332340002, + 0.007132148370146751, + 0.007623969577252865, + 0.013694997876882553, + 0.008839206770062447, + 0.010794044472277164, + 0.010087788105010986, + 0.007090841885656118, + 0.007516619749367237, + 0.01006886176764965, + 0.009112849831581116, + 0.014365321956574917, + 0.009041236713528633, + 0.014839443378150463, + 0.00924292765557766, + 0.004161914810538292, + 0.004311894532293081, + 0.004664179868996143, + 0.0045076459646224976, + 0.0038036294281482697, + 0.004368459805846214, + 0.005372733809053898, + 0.0022806262131780386, + 0.0023918321821838617, + 0.002516825683414936, + 0.0034976571332663298, + 0.0026417404878884554, + 0.0021499504800885916, + 0.0030053798109292984, + 0.0035686774645000696, + 0.002084384672343731, + 0.01579243876039982, + 0.016508441418409348, + 0.016172466799616814, + 0.02327161841094494, + 0.02058839611709118, + 0.010346529074013233, + 0.019996456801891327, + 0.017308741807937622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0380881205201149 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, fluid, circular dolly shot begins, swirling at eye level around a vibrant cooking scene in a cozy, warmly lit kitchen, all rendered in the vibrant, captivating style of Pixar animation. At the centerpiece, a sturdy, amber-colored glass bottle of craft beer stands proudly, its surface rimmed with cool condensation. Its label is a playful, purely graphic design, depicting rolling hills and sunlight, saturated with saturated oranges and greens. Nearby, a cherry-red, enameled pot simmers gently on a gas stove, a thick, bubbling tomato sauce topped with distinctly diced onions and sweet bean paste, releasing a subtle burst of steam. The bottle and pan rest on a polished, light-colored wooden countertop, artfully stacked with an array of remarkably fresh ingredients: plump, glossy red tomatoes, a whole head of garlic with papery skin, and a sprig of bright green basil, all seemingly impossibly perfect. As the camera continues to swirl gracefully, an unseen hand, rendered in soft, stylized features, briefly enters the frame, stirring the sauce steadily with a wooden spoon in fluid, deliberate movements. The entire scene is bathed in a soft golden glow from overhead chandeliers, casting long, soft shadows that make the colors pop and create a cheerful and warm family atmosphere that celebrates the simple joy of preparing a delicious meal.", + "seed": 22922, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_1/index_3744_seed_22922_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06870580408722163, + "vlm_logprob": -0.06870580408722163, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.06870580408722163, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10174420475959778 + ], + "opposite_logprob_tokens": [ + -2.6017441749572754 + ], + "logprob_reward": -0.10174420475959778, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.10174420475959778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05570515617728233, + 0.04874930530786514, + 0.07463525235652924, + 0.2599342167377472, + 0.048681147396564484, + 0.13579215109348297, + 0.05668860301375389, + 0.07846227288246155, + 0.20049989223480225, + 0.04085201397538185 + ], + "credit_map": [ + 0.009911934845149517, + 0.008745262399315834, + 0.005815865006297827, + 0.004641516134142876, + 0.007955314591526985, + 0.006930023431777954, + 0.006368071306496859, + 0.005337169859558344, + 0.004833195824176073, + 0.012464145198464394, + 0.004536155611276627, + 0.0030944054014980793, + 0.009014618583023548, + 0.006174120586365461, + 0.004388768225908279, + 0.004243890754878521, + 0.005976354703307152, + 0.00843726098537445, + 0.01190781220793724, + 0.008718283846974373, + 0.008525356650352478, + 0.006270367186516523, + 0.01499338261783123, + 0.009806433692574501, + 0.012321424670517445, + 0.097620889544487, + 0.005156305152922869, + 0.007614776026457548, + 0.022119488567113876, + 0.10023631900548935, + 0.006342445965856314, + 0.008522561751306057, + 0.008876212872564793, + 0.011367574334144592, + 0.006197252310812473, + 0.004650934133678675, + 0.006734611466526985, + 0.005030219908803701, + 0.0027464712038636208, + 0.0030778723303228617, + 0.011793043464422226, + 0.00496290996670723, + 0.034159060567617416, + 0.017992980778217316, + 0.007830310612916946, + 0.04029352217912674, + 0.011283925734460354, + 0.007476406637579203, + 0.004098694305866957, + 0.0051045785658061504, + 0.01034359447658062, + 0.012671277858316898, + 0.005063746124505997, + 0.006059519946575165, + 0.007560786325484514, + 0.005786404013633728, + 0.0065115573816001415, + 0.00963534228503704, + 0.015045546926558018, + 0.010730430483818054, + 0.009423164650797844, + 0.01056006271392107, + 0.009578208439052105, + 0.006977963726967573, + 0.006415863987058401, + 0.007311227265745401, + 0.009349118918180466, + 0.08164697885513306, + 0.0066085211001336575, + 0.0065477220341563225, + 0.07642792165279388, + 0.0061925435438752174, + 0.0029578404501080513, + 0.00699927844107151, + 0.004037266131490469, + 0.00719589414075017, + 0.004229011945426464, + 0.005081371404230595, + 0.005297959316521883, + 0.00505339028313756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.10174420475959778 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813505172729492 + ], + "opposite_logprob_tokens": [ + -3.288135051727295 + ], + "logprob_reward": -0.03813505172729492, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813505172729492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24116694927215576, + 0.05123382806777954, + 0.02709595486521721, + 0.20615746080875397, + 0.025115089491009712, + 0.2748669385910034, + 0.016135646030306816, + 0.053496118634939194, + 0.08279410749673843, + 0.021937889978289604 + ], + "credit_map": [ + 0.023755107074975967, + 0.03210514411330223, + 0.013768100179731846, + 0.01509914267808199, + 0.05059712007641792, + 0.028258593752980232, + 0.05992916598916054, + 0.01765458472073078, + 0.0061486694030463696, + 0.01030200906097889, + 0.005795743782073259, + 0.004967858549207449, + 0.007587227504700422, + 0.006760310847312212, + 0.00441332021728158, + 0.005258691962808371, + 0.0027629684191197157, + 0.0030278265476226807, + 0.0027789867017418146, + 0.006799135822802782, + 0.0036612283438444138, + 0.0029081874527037144, + 0.002463175915181637, + 0.002694444265216589, + 0.009192156605422497, + 0.07921858876943588, + 0.0046400632709264755, + 0.007031132932752371, + 0.018783334642648697, + 0.07531151175498962, + 0.005104183219373226, + 0.006876485887914896, + 0.00255518383346498, + 0.004089868161827326, + 0.0025637284852564335, + 0.0025847721844911575, + 0.0027509559877216816, + 0.0032815204467624426, + 0.0038519534282386303, + 0.0034371051006019115, + 0.026576198637485504, + 0.006452200468629599, + 0.0710473284125328, + 0.029819516465067863, + 0.016457295045256615, + 0.09485651552677155, + 0.016695978119969368, + 0.012961911037564278, + 0.0014877563808113337, + 0.002026086440309882, + 0.002556274179369211, + 0.00196647085249424, + 0.0019336671102792025, + 0.0021451131906360388, + 0.0021455897949635983, + 0.0018746877321973443, + 0.004513507708907127, + 0.005437060724943876, + 0.009863168001174927, + 0.008193260990083218, + 0.006424569059163332, + 0.004434209782630205, + 0.009394672699272633, + 0.005235666409134865, + 0.0026951890904456377, + 0.004428261891007423, + 0.004534170962870121, + 0.030661994591355324, + 0.002958711702376604, + 0.0032947545405477285, + 0.03063328191637993, + 0.0035877374466508627, + 0.0021875270176678896, + 0.003959375899285078, + 0.0017757016466930509, + 0.00313685555011034, + 0.0029415495228022337, + 0.003244217252358794, + 0.0023640969302505255, + 0.002328564878553152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.03813505172729492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081154346466064 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.016081154346466064, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081154346466064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09840551018714905, + 0.030315961688756943, + 0.030757589265704155, + 0.23347830772399902, + 0.06445557624101639, + 0.17444393038749695, + 0.0643971636891365, + 0.11039059609174728, + 0.15964823961257935, + 0.03370707109570503 + ], + "credit_map": [ + 0.013503163121640682, + 0.01938518136739731, + 0.005468638148158789, + 0.00835740938782692, + 0.01764560304582119, + 0.01319380383938551, + 0.012766210362315178, + 0.008085500448942184, + 0.002977123251184821, + 0.006940622813999653, + 0.0035676234401762486, + 0.0022255254443734884, + 0.005825369618833065, + 0.004494902677834034, + 0.0021030723582953215, + 0.0021817234810441732, + 0.0027069603092968464, + 0.003348926780745387, + 0.0034686089493334293, + 0.004137290176004171, + 0.0038684625178575516, + 0.0034788991324603558, + 0.0061749909073114395, + 0.0035734502598643303, + 0.010593188926577568, + 0.08778592944145203, + 0.006025393959134817, + 0.008615599013864994, + 0.026819627732038498, + 0.07969949394464493, + 0.006603123154491186, + 0.0073359510861337185, + 0.008469457738101482, + 0.012691229581832886, + 0.00940268486738205, + 0.006800306495279074, + 0.0075422246009111404, + 0.005793035961687565, + 0.006614646874368191, + 0.007141986396163702, + 0.01760733500123024, + 0.004374597687274218, + 0.045891497284173965, + 0.018546776846051216, + 0.010267759673297405, + 0.06064850091934204, + 0.00928004551678896, + 0.007827423512935638, + 0.004152233712375164, + 0.005802833009511232, + 0.01186604704707861, + 0.011775429360568523, + 0.006973553448915482, + 0.006339339539408684, + 0.012142134830355644, + 0.005345591809600592, + 0.00797684770077467, + 0.01117182057350874, + 0.015676289796829224, + 0.018768129870295525, + 0.016894584521651268, + 0.012070686556398869, + 0.019208140671253204, + 0.00862409733235836, + 0.002582571469247341, + 0.004702848847955465, + 0.0054360926151275635, + 0.06952117383480072, + 0.004534523468464613, + 0.002132250927388668, + 0.06685348600149155, + 0.0038852880243211985, + 0.003219034057110548, + 0.005827188491821289, + 0.001978848362341523, + 0.005814779549837112, + 0.0045120143331587315, + 0.0051674433052539825, + 0.0038007902912795544, + 0.003386972937732935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.016081154346466064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994391918182373 + ], + "opposite_logprob_tokens": [ + -1.8049439191818237 + ], + "logprob_reward": -0.17994391918182373, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994391918182373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10413350909948349, + 0.024531524628400803, + 0.02429831214249134, + 0.25347191095352173, + 0.16895848512649536, + 0.17913256585597992, + 0.04154011234641075, + 0.1187165379524231, + 0.05677236244082451, + 0.02844470925629139 + ], + "credit_map": [ + 0.01921837404370308, + 0.02309403195977211, + 0.00692263338714838, + 0.008253062143921852, + 0.01235982310026884, + 0.012580029666423798, + 0.011922975070774555, + 0.009782576002180576, + 0.0017517188098281622, + 0.005772502161562443, + 0.0028831649105995893, + 0.002307586371898651, + 0.004554832354187965, + 0.003545523853972554, + 0.0020477576181292534, + 0.0016684380825608969, + 0.0024581397883594036, + 0.0027114979457110167, + 0.0028318988624960184, + 0.004134971648454666, + 0.003116646083071828, + 0.002602766500785947, + 0.0037392948288470507, + 0.0027030964847654104, + 0.010437114164233208, + 0.09713693708181381, + 0.005660505034029484, + 0.007413263898342848, + 0.024254031479358673, + 0.09439096599817276, + 0.006732359062880278, + 0.007446751929819584, + 0.025056950747966766, + 0.03592449799180031, + 0.024668056517839432, + 0.01792220212519169, + 0.021588023751974106, + 0.01498938538134098, + 0.01079647522419691, + 0.018012892454862595, + 0.01613602787256241, + 0.004570289980620146, + 0.04829201102256775, + 0.02252177894115448, + 0.010776002891361713, + 0.057082753628492355, + 0.010824954137206078, + 0.008928751572966576, + 0.0027439172845333815, + 0.003983198199421167, + 0.006107535678893328, + 0.00610526604577899, + 0.008361538872122765, + 0.005184726789593697, + 0.005503410939127207, + 0.0035505187697708607, + 0.009292518720030785, + 0.011265534907579422, + 0.020658696070313454, + 0.0184401236474514, + 0.021991046145558357, + 0.013595130294561386, + 0.01484206598252058, + 0.008631423115730286, + 0.0024384509306401014, + 0.0035170845221728086, + 0.0034237306099385023, + 0.020460478961467743, + 0.003097591456025839, + 0.002153424546122551, + 0.01970982365310192, + 0.001971771940588951, + 0.0024773452896624804, + 0.0047508710995316505, + 0.002954282099381089, + 0.004782326519489288, + 0.0030450259801000357, + 0.0037765142042189837, + 0.0033852136693894863, + 0.0032731296960264444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.17994391918182373 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624690420925617 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624690420925617, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624690420925617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10503888130187988, + 0.03626212105154991, + 0.08514087647199631, + 0.5506129264831543, + 0.02987401932477951, + 0.055340297520160675, + 0.020511671900749207, + 0.029008427634835243, + 0.07566814124584198, + 0.012542545795440674 + ], + "credit_map": [ + 0.013255948200821877, + 0.018610138446092606, + 0.007585698273032904, + 0.008307443000376225, + 0.01941719651222229, + 0.012214120477437973, + 0.01623266004025936, + 0.009415668435394764, + 0.0036244031507521868, + 0.006045524496585131, + 0.0033525119069963694, + 0.003676589345559478, + 0.0053446339443326, + 0.0075021847151219845, + 0.0023425272665917873, + 0.004373745061457157, + 0.0072898417711257935, + 0.01287737488746643, + 0.010082066990435123, + 0.012540720403194427, + 0.013501504436135292, + 0.00973878800868988, + 0.009552239440381527, + 0.009558344259858131, + 0.02717794105410576, + 0.21166177093982697, + 0.011981905438005924, + 0.015549772419035435, + 0.05082738399505615, + 0.20296625792980194, + 0.012033171020448208, + 0.01841476373374462, + 0.002877888735383749, + 0.004585564136505127, + 0.0022971229627728462, + 0.0020618485286831856, + 0.0036128258798271418, + 0.006296079605817795, + 0.003237975062802434, + 0.004904714412987232, + 0.004805282689630985, + 0.004374181851744652, + 0.009295078925788403, + 0.008564184419810772, + 0.005902913399040699, + 0.010992246679961681, + 0.0058199502527713776, + 0.00558646023273468, + 0.0018576948205009103, + 0.0028958937618881464, + 0.0036392470356076956, + 0.002779810456559062, + 0.0020519024692475796, + 0.0019151411252096295, + 0.0034108376130461693, + 0.0019611434545367956, + 0.0022726943716406822, + 0.0035423091612756252, + 0.004082869738340378, + 0.003930427599698305, + 0.00515238381922245, + 0.002418105024844408, + 0.005339286755770445, + 0.0022703499998897314, + 0.0035538466181606054, + 0.005505397450178862, + 0.0033517328556627035, + 0.026471171528100967, + 0.004618991166353226, + 0.00478050671517849, + 0.02472650445997715, + 0.00265998812392354, + 0.0009484966285526752, + 0.002167073544114828, + 0.0012484333710744977, + 0.0019085478270426393, + 0.0012895468389615417, + 0.001291847787797451, + 0.002204416086897254, + 0.001484184293076396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.007624690420925617 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, fluid, circular dolly shot begins, swirling at eye level around a vibrant cooking scene in a cozy, warmly lit kitchen, all rendered in the vibrant, captivating style of Pixar animation. At the centerpiece, a sturdy, amber-colored glass bottle of craft beer stands proudly, its surface rimmed with cool condensation. Its label is a playful, purely graphic design, depicting rolling hills and sunlight, saturated with saturated oranges and greens. Nearby, a cherry-red, enameled pot simmers gently on a gas stove, a thick, bubbling tomato sauce topped with distinctly diced onions and sweet bean paste, releasing a subtle burst of steam. The bottle and pan rest on a polished, light-colored wooden countertop, artfully stacked with an array of remarkably fresh ingredients: plump, glossy red tomatoes, a whole head of garlic with papery skin, and a sprig of bright green basil, all seemingly impossibly perfect. As the camera continues to swirl gracefully, an unseen hand, rendered in soft, stylized features, briefly enters the frame, stirring the sauce steadily with a wooden spoon in fluid, deliberate movements. The entire scene is bathed in a soft golden glow from overhead chandeliers, casting long, soft shadows that make the colors pop and create a cheerful and warm family atmosphere that celebrates the simple joy of preparing a delicious meal.", + "seed": 22923, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_1/index_3744_seed_22923_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06064502131193876, + "vlm_logprob": -0.06064502131193876, + "vlm_yes_no_margin": 3.2749999523162843, + "vlm_token_credit": -0.06064502131193876, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12313870340585709 + ], + "opposite_logprob_tokens": [ + -2.373138666152954 + ], + "logprob_reward": -0.12313870340585709, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12313870340585709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08743437379598618, + 0.08194462954998016, + 0.14999903738498688, + 0.05750966817140579, + 0.09238504618406296, + 0.16005036234855652, + 0.11278768628835678, + 0.15967141091823578, + 0.04394358769059181, + 0.054274190217256546 + ], + "credit_map": [ + 0.017922261729836464, + 0.015429976396262646, + 0.01353567186743021, + 0.009344212710857391, + 0.004776793997734785, + 0.015018131583929062, + 0.005150240380316973, + 0.006257086526602507, + 0.007170911878347397, + 0.006872023455798626, + 0.011467237025499344, + 0.012442631646990776, + 0.010840080678462982, + 0.015846945345401764, + 0.00875658169388771, + 0.008548224344849586, + 0.019393298774957657, + 0.02307475358247757, + 0.021107029169797897, + 0.020620746538043022, + 0.01937880553305149, + 0.025581298395991325, + 0.01000426895916462, + 0.010838830843567848, + 0.007897171191871166, + 0.0062581803649663925, + 0.00632683327421546, + 0.009856409393250942, + 0.006970565766096115, + 0.00653063552454114, + 0.009838061407208443, + 0.003831808688119054, + 0.01666877791285515, + 0.010304754599928856, + 0.006791805848479271, + 0.009205855429172516, + 0.006534721702337265, + 0.027075495570898056, + 0.00861947238445282, + 0.007184159476310015, + 0.02009880729019642, + 0.008648738265037537, + 0.05760221928358078, + 0.04910295829176903, + 0.00508407736197114, + 0.0054765427485108376, + 0.009153160266578197, + 0.004883856512606144, + 0.01283947192132473, + 0.011702804826200008, + 0.011094740591943264, + 0.01987440325319767, + 0.012864276766777039, + 0.009660073556005955, + 0.027224667370319366, + 0.007527240086346865, + 0.015760671347379684, + 0.01615489087998867, + 0.017079027369618416, + 0.02617114596068859, + 0.02079782821238041, + 0.015125829726457596, + 0.029975049197673798, + 0.018606984987854958, + 0.004193836357444525, + 0.006456648465245962, + 0.00437201140448451, + 0.0067832497879862785, + 0.00579387042671442, + 0.0058156587183475494, + 0.00600050063803792, + 0.00452780956402421, + 0.004324737936258316, + 0.00470815971493721, + 0.006794396787881851, + 0.011542223393917084, + 0.008978938683867455, + 0.0053879390470683575, + 0.008715158328413963, + 0.0038226356264203787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.12313870340585709 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020623063668608665 + ], + "opposite_logprob_tokens": [ + -3.895622968673706 + ], + "logprob_reward": -0.020623063668608665, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020623063668608665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2519584894180298, + 0.06341998279094696, + 0.03797666355967522, + 0.03381721302866936, + 0.022244824096560478, + 0.042375270277261734, + 0.217991903424263, + 0.24889880418777466, + 0.05683392658829689, + 0.02448289841413498 + ], + "credit_map": [ + 0.043626561760902405, + 0.029407396912574768, + 0.027519160881638527, + 0.020041828975081444, + 0.033714182674884796, + 0.036316681653261185, + 0.028359748423099518, + 0.03297293931245804, + 0.006764563266187906, + 0.0067250472493469715, + 0.006855190731585026, + 0.007082376163452864, + 0.011014348827302456, + 0.010652042925357819, + 0.007999876514077187, + 0.0063265361823141575, + 0.003156738355755806, + 0.004070538096129894, + 0.006268548313528299, + 0.0054789655841887, + 0.006584180518984795, + 0.006266437470912933, + 0.0033185083884745836, + 0.0028327470645308495, + 0.0049496665596961975, + 0.004634595010429621, + 0.0037454471457749605, + 0.005124200601130724, + 0.004397958051413298, + 0.004172641783952713, + 0.004453521221876144, + 0.002339184982702136, + 0.0025007545482367277, + 0.0029257768765091896, + 0.003969860263168812, + 0.0033636947628110647, + 0.002600280335173011, + 0.0027878056280314922, + 0.0022054770961403847, + 0.001891175750643015, + 0.004508011508733034, + 0.004281220026314259, + 0.007558944635093212, + 0.007829451933503151, + 0.003869428997859359, + 0.004807721823453903, + 0.004676405806094408, + 0.004844088573008776, + 0.014150670729577541, + 0.00924587156623602, + 0.011552647687494755, + 0.06924512982368469, + 0.015404738485813141, + 0.009124997071921825, + 0.08039721846580505, + 0.0088706249371171, + 0.02691037394106388, + 0.024875253438949585, + 0.028745120391249657, + 0.03793267160654068, + 0.03201965615153313, + 0.02767607383430004, + 0.046672020107507706, + 0.024067658931016922, + 0.0052872635424137115, + 0.006138855591416359, + 0.007226998917758465, + 0.008314820937812328, + 0.008514217101037502, + 0.007501596584916115, + 0.008491287007927895, + 0.005358894355595112, + 0.0025358167476952076, + 0.0027337672654539347, + 0.0027388930320739746, + 0.004768311977386475, + 0.0027506717015057802, + 0.0025475118309259415, + 0.004349434748291969, + 0.0020584927406162024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.020623063668608665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197722733020782 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.018197722733020782, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018197722733020782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05232458561658859, + 0.02338743582367897, + 0.04085075855255127, + 0.02616778388619423, + 0.05160687491297722, + 0.2512245774269104, + 0.22050347924232483, + 0.25078946352005005, + 0.06107686832547188, + 0.02206810563802719 + ], + "credit_map": [ + 0.010816800408065319, + 0.009338093921542168, + 0.007476234342902899, + 0.0053220512345433235, + 0.004651600960642099, + 0.008589211851358414, + 0.002824817318469286, + 0.0033057809341698885, + 0.00243271398358047, + 0.0021900988649576902, + 0.0022188422735780478, + 0.004177286755293608, + 0.0027441647835075855, + 0.00325645855627954, + 0.0034777713008224964, + 0.0028901027981191874, + 0.0049094813875854015, + 0.006665239110589027, + 0.005989127792418003, + 0.006182973273098469, + 0.005367089528590441, + 0.005981938447803259, + 0.002787408186122775, + 0.002967504784464836, + 0.0033375381026417017, + 0.002896730788052082, + 0.002407422289252281, + 0.003190249903127551, + 0.004599632695317268, + 0.0030574738048017025, + 0.004017420578747988, + 0.0026613175868988037, + 0.007476951461285353, + 0.007532982155680656, + 0.005051887594163418, + 0.006457632407546043, + 0.004722451791167259, + 0.009665788151323795, + 0.005256892181932926, + 0.005442292895168066, + 0.017579641193151474, + 0.01676025427877903, + 0.07943863421678543, + 0.08583997935056686, + 0.008745160885155201, + 0.009420470334589481, + 0.018951116129755974, + 0.01448933407664299, + 0.01711808703839779, + 0.009497832506895065, + 0.012751749716699123, + 0.06938222795724869, + 0.010469945147633553, + 0.008339578285813332, + 0.08482104539871216, + 0.008123049512505531, + 0.026924340054392815, + 0.027270624414086342, + 0.027845721691846848, + 0.030413519591093063, + 0.038677770644426346, + 0.03287816047668457, + 0.03626355901360512, + 0.030515797436237335, + 0.00613936735317111, + 0.006225167773663998, + 0.006590843666344881, + 0.008045557886362076, + 0.010004409588873386, + 0.010808823630213737, + 0.009013615548610687, + 0.004249086603522301, + 0.0021600753534585238, + 0.002112692454829812, + 0.0021303517278283834, + 0.004237915854901075, + 0.003523311810567975, + 0.00267780851572752, + 0.003475522855296731, + 0.0017504296265542507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.018197722733020782 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707513570785522 + ], + "opposite_logprob_tokens": [ + -2.1270751953125 + ], + "logprob_reward": -0.12707513570785522, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707513570785522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039711352437734604, + 0.03230541944503784, + 0.020921414718031883, + 0.02349625900387764, + 0.035467728972435, + 0.03082219697535038, + 0.1446927785873413, + 0.6217703223228455, + 0.02503924071788788, + 0.025773322209715843 + ], + "credit_map": [ + 0.007807161659002304, + 0.007567923981696367, + 0.007649510633200407, + 0.004063572734594345, + 0.0024358611553907394, + 0.006697688717395067, + 0.0016662750858813524, + 0.0018233557930216193, + 0.002474643522873521, + 0.0030258370097726583, + 0.00551956333220005, + 0.00496452022343874, + 0.0038606508169323206, + 0.006937092635780573, + 0.0031131331343203783, + 0.002409973181784153, + 0.002768568927422166, + 0.004141976125538349, + 0.003384120063856244, + 0.002750478917732835, + 0.0020729328971356153, + 0.0032273749820888042, + 0.0012116855941712856, + 0.0013642758131027222, + 0.0028710654005408287, + 0.002922444138675928, + 0.002083214232698083, + 0.0031921137124300003, + 0.00470262486487627, + 0.0027839101385325193, + 0.003162122331559658, + 0.001778762205503881, + 0.003782191313803196, + 0.0028525246307253838, + 0.009953651577234268, + 0.008790750056505203, + 0.0017881928943097591, + 0.0028288953471928835, + 0.0027611693367362022, + 0.002710350789129734, + 0.003886014921590686, + 0.003555999370291829, + 0.004477147478610277, + 0.006317880004644394, + 0.0027740083169192076, + 0.0034322214778512716, + 0.003526827786117792, + 0.0028520922642201185, + 0.023323774337768555, + 0.008842600509524345, + 0.010625855065882206, + 0.032385725528001785, + 0.009601877070963383, + 0.006509168539196253, + 0.04807257652282715, + 0.005331183783710003, + 0.07018239796161652, + 0.06283452361822128, + 0.053919460624456406, + 0.10581144690513611, + 0.08057685941457748, + 0.06345458328723907, + 0.13749676942825317, + 0.04749421402812004, + 0.002273841528221965, + 0.0017953321803361177, + 0.0028156861662864685, + 0.004144599195569754, + 0.003015066497027874, + 0.0038146558217704296, + 0.005024642217904329, + 0.0021554133854806423, + 0.0027844568248838186, + 0.00328320381231606, + 0.00263347290456295, + 0.002851394470781088, + 0.00649412302300334, + 0.003060638904571533, + 0.003039162140339613, + 0.001626867218874395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.12707513570785522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190481044352055 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190481044352055, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190481044352055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2592163681983948, + 0.06270460039377213, + 0.07605201750993729, + 0.04731548950076103, + 0.11519376933574677, + 0.14165431261062622, + 0.14327850937843323, + 0.06511857360601425, + 0.0427064448595047, + 0.04675989970564842 + ], + "credit_map": [ + 0.04199464991688728, + 0.02525479532778263, + 0.029279228299856186, + 0.020124338567256927, + 0.03300764039158821, + 0.04034871608018875, + 0.027704445645213127, + 0.04150255024433136, + 0.004891896620392799, + 0.0045294854789972305, + 0.007066364400088787, + 0.010794395580887794, + 0.007002216763794422, + 0.009734761901199818, + 0.007730463519692421, + 0.010955024510622025, + 0.015532897785305977, + 0.012851064093410969, + 0.0093929348513484, + 0.007419274654239416, + 0.0102652283385396, + 0.012129639275372028, + 0.0038836458697915077, + 0.00457733403891325, + 0.006568843498826027, + 0.006461878772825003, + 0.004723075311630964, + 0.00753991212695837, + 0.007953220047056675, + 0.005615369416773319, + 0.00455849152058363, + 0.0038946978747844696, + 0.012051485478878021, + 0.008333981037139893, + 0.026233019307255745, + 0.013992393389344215, + 0.00993863120675087, + 0.02904328890144825, + 0.008120095357298851, + 0.0074808732606470585, + 0.020150739699602127, + 0.009279826655983925, + 0.044852662831544876, + 0.036760762333869934, + 0.007138141896575689, + 0.008106925524771214, + 0.010462271980941296, + 0.004902989137917757, + 0.01928621344268322, + 0.013407446444034576, + 0.01169507671147585, + 0.021086351945996284, + 0.024491747841238976, + 0.014243463054299355, + 0.028683114796876907, + 0.010385082103312016, + 0.00884927250444889, + 0.008887813426554203, + 0.006503797136247158, + 0.007472706492990255, + 0.008032860234379768, + 0.009750607423484325, + 0.010097634978592396, + 0.0055238851346075535, + 0.005231904331594706, + 0.004293439909815788, + 0.005023124627768993, + 0.0061760712414979935, + 0.004972198512405157, + 0.0053304643370211124, + 0.006418723147362471, + 0.005260516423732042, + 0.004682695493102074, + 0.005884403828531504, + 0.004850413650274277, + 0.00742452684789896, + 0.009526376612484455, + 0.005032501649111509, + 0.0064056189730763435, + 0.0029533628839999437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.014190481044352055 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, fluid, circular dolly shot begins, swirling at eye level around a vibrant cooking scene in a cozy, warmly lit kitchen, all rendered in the vibrant, captivating style of Pixar animation. At the centerpiece, a sturdy, amber-colored glass bottle of craft beer stands proudly, its surface rimmed with cool condensation. Its label is a playful, purely graphic design, depicting rolling hills and sunlight, saturated with saturated oranges and greens. Nearby, a cherry-red, enameled pot simmers gently on a gas stove, a thick, bubbling tomato sauce topped with distinctly diced onions and sweet bean paste, releasing a subtle burst of steam. The bottle and pan rest on a polished, light-colored wooden countertop, artfully stacked with an array of remarkably fresh ingredients: plump, glossy red tomatoes, a whole head of garlic with papery skin, and a sprig of bright green basil, all seemingly impossibly perfect. As the camera continues to swirl gracefully, an unseen hand, rendered in soft, stylized features, briefly enters the frame, stirring the sauce steadily with a wooden spoon in fluid, deliberate movements. The entire scene is bathed in a soft golden glow from overhead chandeliers, casting long, soft shadows that make the colors pop and create a cheerful and warm family atmosphere that celebrates the simple joy of preparing a delicious meal.", + "seed": 22924, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_1/index_3744_seed_22924_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05741950329393149, + "vlm_logprob": -0.05741950329393149, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.05741950329393149, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10490276664495468 + ], + "opposite_logprob_tokens": [ + -2.604902744293213 + ], + "logprob_reward": -0.10490276664495468, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.10490276664495468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028739964589476585, + 0.041232019662857056, + 0.6539355516433716, + 0.07214242964982986, + 0.011624310165643692, + 0.022335726767778397, + 0.04807368665933609, + 0.08762087672948837, + 0.012457692995667458, + 0.021837709471583366 + ], + "credit_map": [ + 0.0024114151019603014, + 0.0034429810475558043, + 0.0028838429134339094, + 0.0035939854569733143, + 0.004662262741476297, + 0.00481528602540493, + 0.0033142976462841034, + 0.0036158941220492125, + 0.003205471206456423, + 0.0051571885123848915, + 0.002803487004712224, + 0.003677199361845851, + 0.0035141147673130035, + 0.005423772148787975, + 0.008014080114662647, + 0.009436706081032753, + 0.012591768987476826, + 0.00816395040601492, + 0.011616626754403114, + 0.11345064640045166, + 0.022189849987626076, + 0.014485523104667664, + 0.2776130437850952, + 0.1938241422176361, + 0.0024205250665545464, + 0.004573381505906582, + 0.01903894916176796, + 0.02986132726073265, + 0.0015065480256453156, + 0.0018011096399277449, + 0.005042725242674351, + 0.007897856645286083, + 0.0010098208440467715, + 0.0014843873213976622, + 0.0023953162599354982, + 0.0013185909483581781, + 0.001511729322373867, + 0.001219940255396068, + 0.0011883650440722704, + 0.0014961595879867673, + 0.002043656539171934, + 0.003614764427766204, + 0.0034932398702949286, + 0.0026160082779824734, + 0.0019387000938877463, + 0.0018236346077173948, + 0.0034954044967889786, + 0.0033103181049227715, + 0.00583674618974328, + 0.008138824254274368, + 0.00736202672123909, + 0.004339015576988459, + 0.006497346796095371, + 0.00486344238743186, + 0.007944919168949127, + 0.003091363003477454, + 0.004105226136744022, + 0.002714881906285882, + 0.03470521792769432, + 0.011021235957741737, + 0.0038892526645213366, + 0.00431498559191823, + 0.02389269880950451, + 0.0029773772694170475, + 0.001213034032844007, + 0.0021069177892059088, + 0.001538899727165699, + 0.0019403855549171567, + 0.0014866228448227048, + 0.0019598582293838263, + 0.0013543058885261416, + 0.0008576684631407261, + 0.001891668769530952, + 0.0035751331597566605, + 0.002860501641407609, + 0.004366020672023296, + 0.0020561141427606344, + 0.003095305059105158, + 0.0026340933982282877, + 0.0013588722795248032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.10490276664495468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700686126947403 + ], + "opposite_logprob_tokens": [ + -2.695068597793579 + ], + "logprob_reward": -0.0700686126947403, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700686126947403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06804807484149933, + 0.07816722244024277, + 0.5140365362167358, + 0.12289231270551682, + 0.014627030119299889, + 0.03970780223608017, + 0.0626809373497963, + 0.03582939878106117, + 0.04005901888012886, + 0.023951711133122444 + ], + "credit_map": [ + 0.009878148324787617, + 0.0093411635607481, + 0.005760061554610729, + 0.009164074435830116, + 0.005444584414362907, + 0.00760577991604805, + 0.011119633913040161, + 0.009734620340168476, + 0.007632879540324211, + 0.004033791832625866, + 0.00796184316277504, + 0.012033553794026375, + 0.005895202048122883, + 0.007064193021506071, + 0.017918407917022705, + 0.01562735065817833, + 0.011743374168872833, + 0.006984648294746876, + 0.009267440997064114, + 0.08691661804914474, + 0.017492571845650673, + 0.012559321708977222, + 0.215866357088089, + 0.15320618450641632, + 0.003988384269177914, + 0.005505678243935108, + 0.02103229984641075, + 0.050323985517024994, + 0.003172736382111907, + 0.004548511002212763, + 0.010979180224239826, + 0.02334153652191162, + 0.0013892039423808455, + 0.0020773319993168116, + 0.002488055033609271, + 0.0018803890561684966, + 0.0020166656468063593, + 0.0013899257173761725, + 0.0016457939054816961, + 0.0017396630719304085, + 0.0035531159956008196, + 0.007431981153786182, + 0.005169988144189119, + 0.004096955060958862, + 0.008435170166194439, + 0.0032188319601118565, + 0.003928143996745348, + 0.003873615525662899, + 0.005191131494939327, + 0.007374250795692205, + 0.01578054018318653, + 0.005004886072129011, + 0.006146710366010666, + 0.004700943827629089, + 0.015201051719486713, + 0.003281415207311511, + 0.0018718848004937172, + 0.0009003402083180845, + 0.014335581101477146, + 0.004943312611430883, + 0.0015041654696688056, + 0.0014909981982782483, + 0.009656845591962337, + 0.0011262658517807722, + 0.004022060893476009, + 0.00614752434194088, + 0.005536485463380814, + 0.0056377905420959, + 0.007676663808524609, + 0.004145679995417595, + 0.004492583684623241, + 0.0024002252612262964, + 0.0024346173740923405, + 0.003630983177572489, + 0.002719361800700426, + 0.003691664896905422, + 0.0025229842867702246, + 0.0028177767526358366, + 0.00270572560839355, + 0.0034285944420844316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.0700686126947403 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547007761895657 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012547007761895657, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547007761895657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03504564240574837, + 0.034086935222148895, + 0.43419477343559265, + 0.060340896248817444, + 0.03711564093828201, + 0.045320119708776474, + 0.10243840515613556, + 0.09516691416501999, + 0.06736385822296143, + 0.08892673999071121 + ], + "credit_map": [ + 0.0038460749201476574, + 0.004774428904056549, + 0.004031146876513958, + 0.00444443617016077, + 0.004123147111386061, + 0.005129828583449125, + 0.00409907940775156, + 0.0045975022949278355, + 0.003809206420555711, + 0.0035987335722893476, + 0.003188252681866288, + 0.003923460841178894, + 0.003064068965613842, + 0.005463954992592335, + 0.005081808660179377, + 0.005957453045994043, + 0.009328937157988548, + 0.005578914657235146, + 0.00853018555790186, + 0.0722506195306778, + 0.0151058379560709, + 0.010578787885606289, + 0.18169046938419342, + 0.13113105297088623, + 0.002967016538605094, + 0.008884099312126637, + 0.013743438757956028, + 0.02035427838563919, + 0.002196670277044177, + 0.0025538024492561817, + 0.004530949052423239, + 0.005110647063702345, + 0.003306726925075054, + 0.004712691064924002, + 0.007864945568144321, + 0.0048230295069515705, + 0.004888105671852827, + 0.004070148803293705, + 0.003950343932956457, + 0.003499650862067938, + 0.002763556083664298, + 0.006334812380373478, + 0.008129715919494629, + 0.005482801236212254, + 0.004500307142734528, + 0.0036241442430764437, + 0.007239655591547489, + 0.007245130371302366, + 0.01259347703307867, + 0.015830378979444504, + 0.010656505823135376, + 0.010169748216867447, + 0.019780678674578667, + 0.015788478776812553, + 0.012272220104932785, + 0.0053469291888177395, + 0.0038262081798166037, + 0.0020208253990858793, + 0.0410616360604763, + 0.014875510707497597, + 0.0026889455039054155, + 0.002510529477149248, + 0.026001421734690666, + 0.0021818417590111494, + 0.006237117573618889, + 0.011679919436573982, + 0.0075387959368526936, + 0.0073918853886425495, + 0.01196126639842987, + 0.01106640137732029, + 0.007093230728060007, + 0.004395247437059879, + 0.007590155582875013, + 0.017348404973745346, + 0.010062847286462784, + 0.018150364980101585, + 0.009277069009840488, + 0.00893549993634224, + 0.010447092354297638, + 0.0071153161115944386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.012547007761895657 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900651544332504 + ], + "opposite_logprob_tokens": [ + -2.5790064334869385 + ], + "logprob_reward": -0.07900651544332504, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900651544332504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01045177597552538, + 0.009690907783806324, + 0.49772801995277405, + 0.02924521081149578, + 0.018830519169569016, + 0.021764537319540977, + 0.3280457556247711, + 0.0030868311878293753, + 0.005410654470324516, + 0.07574575394392014 + ], + "credit_map": [ + 0.001004772144369781, + 0.001096921507269144, + 0.000992397777736187, + 0.0010901137720793486, + 0.002028045943006873, + 0.0014675125712528825, + 0.0014987303875386715, + 0.001273282920010388, + 0.0006976150907576084, + 0.0009034592076204717, + 0.0018542796606197953, + 0.0011726886732503772, + 0.001028532860800624, + 0.001093525206670165, + 0.0015512030804529786, + 0.0013896049931645393, + 0.01093281526118517, + 0.006593487691134214, + 0.008955448865890503, + 0.08422307670116425, + 0.016794847324490547, + 0.011686000041663647, + 0.20986193418502808, + 0.14868046343326569, + 0.0007116761989891529, + 0.0025657769292593002, + 0.00741825345903635, + 0.011792690493166447, + 0.0005166769842617214, + 0.0008314599399454892, + 0.0017357327742502093, + 0.003672945313155651, + 0.002482334617525339, + 0.0024783469270914793, + 0.00324949249625206, + 0.0021140575408935547, + 0.0026383050717413425, + 0.002776936860755086, + 0.0016421063337475061, + 0.0014489421155303717, + 0.0015286258421838284, + 0.003675028681755066, + 0.003143427660688758, + 0.0026577506214380264, + 0.0038736716378480196, + 0.0019131337758153677, + 0.002886389847844839, + 0.0020865106489509344, + 0.04080996289849281, + 0.051950596272945404, + 0.04684102162718773, + 0.03418699651956558, + 0.056421179324388504, + 0.042618509382009506, + 0.0374128632247448, + 0.017804645001888275, + 0.0005003645783290267, + 0.00035880511859431863, + 0.00039037963142618537, + 0.00032541374093852937, + 0.0004933932214044034, + 0.0005197952268645167, + 0.00031096243765205145, + 0.00018771772738546133, + 0.0006742000114172697, + 0.0008665650384500623, + 0.0007888340041972697, + 0.0006980569451116025, + 0.0006994444411247969, + 0.0008115539094433188, + 0.0005312652210704982, + 0.00034073530696332455, + 0.0077516911551356316, + 0.01966015063226223, + 0.00789799913764, + 0.011817657388746738, + 0.009319703094661236, + 0.005721905268728733, + 0.007818414829671383, + 0.005758239887654781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 9 + ], + "selected_reward": -0.07900651544332504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a slow, fluid, circular dolly shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sturdy, amber-colored glass bottle of craft beer with condensation on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cherry-red, enameled pot simmering on a gas stove with bubbling tomato sauce, diced onions, and sweet bean paste?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an unseen hand enter the frame to stir the sauce with a wooden spoon in fluid, deliberate movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are fresh ingredients including plump red tomatoes, a whole head of garlic, and a sprig of bright green basil visible on the countertop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572613924741745 + ], + "opposite_logprob_tokens": [ + -3.8955726623535156 + ], + "logprob_reward": -0.020572613924741745, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572613924741745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28279080986976624, + 0.12349945306777954, + 0.06773488223552704, + 0.1518602818250656, + 0.04274923354387283, + 0.04627818614244461, + 0.12473800033330917, + 0.05900437384843826, + 0.03363873064517975, + 0.06770607084035873 + ], + "credit_map": [ + 0.03308012709021568, + 0.03167407214641571, + 0.029938342049717903, + 0.037562839686870575, + 0.03702140972018242, + 0.029146982356905937, + 0.04842498153448105, + 0.03594207763671875, + 0.01207910105586052, + 0.008331415243446827, + 0.008352219127118587, + 0.012136653997004032, + 0.008782057091593742, + 0.015568527393043041, + 0.02897316962480545, + 0.029276316985487938, + 0.0036743308883160353, + 0.003378507914021611, + 0.0036914139054715633, + 0.007814761251211166, + 0.005176398903131485, + 0.004591029137372971, + 0.016945406794548035, + 0.022463027387857437, + 0.0032753264531493187, + 0.006564888637512922, + 0.03541316092014313, + 0.06556835770606995, + 0.002150316722691059, + 0.0037771675270050764, + 0.010887411423027515, + 0.024223649874329567, + 0.005166243761777878, + 0.00571201927959919, + 0.008257229812443256, + 0.005241705570369959, + 0.004382419865578413, + 0.0033062563743442297, + 0.005437380168586969, + 0.005245976150035858, + 0.0025749769993126392, + 0.004176797810941935, + 0.008949974551796913, + 0.0074605378322303295, + 0.006129579618573189, + 0.003428648691624403, + 0.008647441864013672, + 0.004910231102257967, + 0.013951358385384083, + 0.01776704378426075, + 0.020920030772686005, + 0.01464688777923584, + 0.019257815554738045, + 0.015433665364980698, + 0.016386909410357475, + 0.0063742902129888535, + 0.002698963973671198, + 0.0018220694037154317, + 0.023096250370144844, + 0.0054135313257575035, + 0.0026028547435998917, + 0.002473351312801242, + 0.018779831007122993, + 0.0021175246220082045, + 0.0034059928730130196, + 0.003633132902905345, + 0.008380159735679626, + 0.003745290217921138, + 0.00391529081389308, + 0.002839125692844391, + 0.006013880949467421, + 0.0017058581579476595, + 0.006840009707957506, + 0.015062200836837292, + 0.0067959558218717575, + 0.010532000102102757, + 0.00846073403954506, + 0.007667472120374441, + 0.007193049881607294, + 0.005154645070433617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.020572613924741745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_10/index_tensor([2254, 2254])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_10/index_tensor([2254, 2254])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9b2acd7b7bc2ffafda349a9a858e649c081963a5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_10/index_tensor([2254, 2254])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a professor in his 50s, with thinning salt-and-pepper hair and wire-rimmed glasses, standing before a large, dark green blackboard in a traditional university lecture hall. He wears a classic brown tweed blazer with worn leather patches at the elbows, over a crisp light blue button-down shirt. His body is slightly turned away from the camera as he focuses intently on the blackboard, which is already partially covered with complex diagrams and mathematical symbols. With his right hand, he steadily and deliberately slides a small piece of white chalk across the surface, meticulously forming a new physics equation using Greek letters and symbols, each firm stroke releasing a faint burst of chalk dust. In the softly blurred foreground, the heads of the students are clearly visible, all hunched over their wooden tables, their hands moving in sync as they copy the professor's work into their notebooks. The scene is bathed in the even, practical glow of overhead lighting, creating an atmosphere of focused, tranquil, and attentive intelligence.", + "seed": 2397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_10/index_2254_seed_2397_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.030890599451959135, + "vlm_logprob": -0.030890599451959135, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.030890599451959135, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.050841476768255234 + ], + "opposite_logprob_tokens": [ + -3.0508415699005127 + ], + "logprob_reward": -0.050841476768255234, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.050841476768255234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1424504667520523, + 0.10616347938776016, + 0.09595510363578796, + 0.06017504632472992, + 0.08460172265768051, + 0.04773008078336716, + 0.3197120726108551, + 0.03933832421898842, + 0.05199931561946869, + 0.051874443888664246 + ], + "credit_map": [ + 0.02403363399207592, + 0.02214212156832218, + 0.01341519970446825, + 0.02532319724559784, + 0.01855473592877388, + 0.018367890268564224, + 0.011529367417097092, + 0.0090843066573143, + 0.016578979790210724, + 0.018311791121959686, + 0.011111237108707428, + 0.02275850810110569, + 0.011433431878685951, + 0.010925338603556156, + 0.008488615974783897, + 0.006555563770234585, + 0.01960751600563526, + 0.014633744955062866, + 0.011695909313857555, + 0.018552076071500778, + 0.009549339301884174, + 0.00991519819945097, + 0.006886069662868977, + 0.005115237552672625, + 0.009555223397910595, + 0.010481408797204494, + 0.004737445153295994, + 0.017216090112924576, + 0.004975092597305775, + 0.005384834948927164, + 0.004857593681663275, + 0.0029673511162400246, + 0.014353017322719097, + 0.019715871661901474, + 0.007292506750673056, + 0.016647368669509888, + 0.008007337339222431, + 0.008282291702926159, + 0.006250742822885513, + 0.004052574746310711, + 0.008474670350551605, + 0.010191896930336952, + 0.007256249897181988, + 0.003730050753802061, + 0.005433168262243271, + 0.005334715358912945, + 0.0033803945407271385, + 0.003928928170353174, + 0.0695793479681015, + 0.03020532801747322, + 0.03614543378353119, + 0.11573867499828339, + 0.011452239006757736, + 0.022124094888567924, + 0.022686423733830452, + 0.011780505999922752, + 0.005215085111558437, + 0.008307741023600101, + 0.004913329146802425, + 0.003996724728494883, + 0.005525958724319935, + 0.004184611141681671, + 0.00363274198025465, + 0.003562123281881213, + 0.008089018985629082, + 0.010497591458261013, + 0.0044615608640015125, + 0.00846580695360899, + 0.005721167661249638, + 0.007004644256085157, + 0.003767448477447033, + 0.00399206904694438, + 0.008184844627976418, + 0.008631276898086071, + 0.004578528460115194, + 0.008639398962259293, + 0.005775136407464743, + 0.008508686907589436, + 0.004605651833117008, + 0.0029509160667657852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.050841476768255234 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371585160493851 + ], + "opposite_logprob_tokens": [ + -3.4087159633636475 + ], + "logprob_reward": -0.03371585160493851, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371585160493851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14630816876888275, + 0.12516289949417114, + 0.13061287999153137, + 0.101750947535038, + 0.10295362770557404, + 0.03162689879536629, + 0.11623828113079071, + 0.07503880560398102, + 0.0651373416185379, + 0.10517022013664246 + ], + "credit_map": [ + 0.016567327082157135, + 0.02131122723221779, + 0.01804628036916256, + 0.03337479010224342, + 0.016229553148150444, + 0.016035398468375206, + 0.013455820269882679, + 0.011287782341241837, + 0.016654880717396736, + 0.01827714592218399, + 0.013189449906349182, + 0.028722159564495087, + 0.015207190997898579, + 0.013304528780281544, + 0.009949267841875553, + 0.009858274832367897, + 0.02549157850444317, + 0.018490280956029892, + 0.018614904955029488, + 0.02654116414487362, + 0.011335017159581184, + 0.012370246462523937, + 0.010400170460343361, + 0.00736951781436801, + 0.01729126274585724, + 0.01802745833992958, + 0.007753800135105848, + 0.02804449386894703, + 0.009082240983843803, + 0.00891320500522852, + 0.007890836335718632, + 0.004747647792100906, + 0.01974959671497345, + 0.021114129573106766, + 0.014602337032556534, + 0.011225632391870022, + 0.010595181956887245, + 0.010631266050040722, + 0.007334961090236902, + 0.007700522430241108, + 0.005507879890501499, + 0.005279682110995054, + 0.004267116077244282, + 0.00481117470189929, + 0.0034732078202068806, + 0.003577915718778968, + 0.0028067692182958126, + 0.001903153257444501, + 0.019329961389303207, + 0.012646985240280628, + 0.01451854221522808, + 0.039021823555231094, + 0.007265453692525625, + 0.009404686279594898, + 0.009380672127008438, + 0.004670146387070417, + 0.011582775041460991, + 0.013922945596277714, + 0.009180312976241112, + 0.009397527202963829, + 0.010893029160797596, + 0.00731044402346015, + 0.006258048582822084, + 0.006493718829005957, + 0.010234085842967033, + 0.013641861267387867, + 0.005155382677912712, + 0.009710727259516716, + 0.007611692883074284, + 0.009634830988943577, + 0.004761195741593838, + 0.004387565888464451, + 0.014691120013594627, + 0.018423618748784065, + 0.010301542468369007, + 0.020561274141073227, + 0.009729411453008652, + 0.011133525520563126, + 0.009301639162003994, + 0.011028085835278034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03371585160493851 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639576844871044 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639576844871044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639576844871044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2753317654132843, + 0.08196596056222916, + 0.10065053403377533, + 0.08410133421421051, + 0.036032311618328094, + 0.05612265691161156, + 0.11599339544773102, + 0.13016900420188904, + 0.0352129191160202, + 0.08442012965679169 + ], + "credit_map": [ + 0.03497457876801491, + 0.03250528499484062, + 0.04220137745141983, + 0.04260649532079697, + 0.04352091625332832, + 0.036118630319833755, + 0.024484578520059586, + 0.018919875845313072, + 0.010521914809942245, + 0.011801052838563919, + 0.009515892714262009, + 0.016554316505789757, + 0.011760917492210865, + 0.009465295821428299, + 0.006622671149671078, + 0.005723896902054548, + 0.016604699194431305, + 0.015416078269481659, + 0.009902464225888252, + 0.0244925357401371, + 0.010006056167185307, + 0.010314768180251122, + 0.007957948371767998, + 0.005955986212939024, + 0.013226063922047615, + 0.013616436161100864, + 0.008605742827057838, + 0.022043505683541298, + 0.007398359477519989, + 0.008205761201679707, + 0.006964344531297684, + 0.004041118547320366, + 0.005801553372293711, + 0.005579313728958368, + 0.004293003585189581, + 0.005214511416852474, + 0.004210525192320347, + 0.004764911253005266, + 0.0038212111685425043, + 0.00234727980569005, + 0.007355577778071165, + 0.01067270990461111, + 0.009958089329302311, + 0.005260589066892862, + 0.007191166747361422, + 0.007152285426855087, + 0.005639494862407446, + 0.0028927410021424294, + 0.016629571095108986, + 0.012052170932292938, + 0.015099517069756985, + 0.03250211849808693, + 0.009155391715466976, + 0.010429175570607185, + 0.010979685932397842, + 0.009145772084593773, + 0.017436083406209946, + 0.02069680206477642, + 0.014768269844353199, + 0.01962372288107872, + 0.0144405048340559, + 0.01582573540508747, + 0.013841741718351841, + 0.013536141254007816, + 0.004256414715200663, + 0.005088360980153084, + 0.004451355896890163, + 0.005855690222233534, + 0.0034893243573606014, + 0.003911883570253849, + 0.004212378989905119, + 0.003947511315345764, + 0.011594586074352264, + 0.008301101624965668, + 0.011335117742419243, + 0.010092215612530708, + 0.00806595478206873, + 0.012328685261309147, + 0.012045497074723244, + 0.010656976141035557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.008639576844871044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307721555233002 + ], + "opposite_logprob_tokens": [ + -3.168077230453491 + ], + "logprob_reward": -0.04307721555233002, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307721555233002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1672157496213913, + 0.06007161736488342, + 0.05500747635960579, + 0.07360132038593292, + 0.03086031600832939, + 0.019480982795357704, + 0.4846719801425934, + 0.04477565735578537, + 0.030357221141457558, + 0.03395763412117958 + ], + "credit_map": [ + 0.01629280298948288, + 0.026348186656832695, + 0.019392523914575577, + 0.04026252031326294, + 0.020216163247823715, + 0.01799100637435913, + 0.01775510422885418, + 0.008957436308264732, + 0.006043252535164356, + 0.00951097160577774, + 0.00717049278318882, + 0.014194595627486706, + 0.008519448339939117, + 0.005943295080214739, + 0.004845560062676668, + 0.0038440090138465166, + 0.007661766838282347, + 0.0072614885866642, + 0.007719896268099546, + 0.008046390488743782, + 0.008046485483646393, + 0.007064979523420334, + 0.005790624301880598, + 0.00341584999114275, + 0.007912000641226768, + 0.009875534102320671, + 0.006995178759098053, + 0.024629075080156326, + 0.006068449933081865, + 0.006725440267473459, + 0.006681564263999462, + 0.00471408199518919, + 0.005253483075648546, + 0.00657140975818038, + 0.0039954716339707375, + 0.00509609654545784, + 0.002853495068848133, + 0.003186456859111786, + 0.0023015292827039957, + 0.001602374599315226, + 0.0023634228855371475, + 0.0044170874170959, + 0.002415435155853629, + 0.0028912231791764498, + 0.002503537805750966, + 0.002360703656449914, + 0.0015381413977593184, + 0.0009914323454722762, + 0.10513404756784439, + 0.0465129092335701, + 0.05442415922880173, + 0.17772550880908966, + 0.016802340745925903, + 0.03243221715092659, + 0.033614423125982285, + 0.018026432022452354, + 0.006090342067182064, + 0.008106046356260777, + 0.004568201024085283, + 0.006056251935660839, + 0.00620758393779397, + 0.004863039590418339, + 0.004323045257478952, + 0.004561153240501881, + 0.004906306974589825, + 0.003466979833319783, + 0.004175156354904175, + 0.0036453825887292624, + 0.0032198275439441204, + 0.0037861948367208242, + 0.0036988290958106518, + 0.003458548104390502, + 0.005394134670495987, + 0.005549415946006775, + 0.003654763801023364, + 0.004739577881991863, + 0.003976487088948488, + 0.005023004952818155, + 0.003284272039309144, + 0.0023359840270131826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.04307721555233002 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178876489400864 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178876489400864, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178876489400864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10053831338882446, + 0.08677779883146286, + 0.14440886676311493, + 0.03062485158443451, + 0.04065065085887909, + 0.026451414451003075, + 0.4726112484931946, + 0.03101661428809166, + 0.03622639179229736, + 0.030693823471665382 + ], + "credit_map": [ + 0.012655496597290039, + 0.01294063962996006, + 0.015450602397322655, + 0.014326930977404118, + 0.013823124580085278, + 0.011235298588871956, + 0.010766279883682728, + 0.009339936077594757, + 0.013153261505067348, + 0.014176126569509506, + 0.010808659717440605, + 0.02068066969513893, + 0.007929153740406036, + 0.008526023477315903, + 0.006861421745270491, + 0.004642486572265625, + 0.022712115198373795, + 0.02097228728234768, + 0.021263914182782173, + 0.02191728912293911, + 0.019925100728869438, + 0.014798329211771488, + 0.011961274780333042, + 0.010858564637601376, + 0.0045026252046227455, + 0.003749711439013481, + 0.005801233928650618, + 0.004359859507530928, + 0.0034316990058869123, + 0.003157559549435973, + 0.0029951755423098803, + 0.0026269855443388224, + 0.007405261509120464, + 0.008832777850329876, + 0.005523460451513529, + 0.006174265872687101, + 0.0035756982397288084, + 0.0039564804174005985, + 0.0029409313574433327, + 0.002241777488961816, + 0.003354325657710433, + 0.003981658723205328, + 0.004191187210381031, + 0.004301807843148708, + 0.0028231858741492033, + 0.0023485906422138214, + 0.0030117162968963385, + 0.002438941737636924, + 0.10543988645076752, + 0.04807252436876297, + 0.05072402209043503, + 0.1615218073129654, + 0.021232757717370987, + 0.03467198833823204, + 0.0328458733856678, + 0.018102368339896202, + 0.00445569958537817, + 0.003843219019472599, + 0.004148818552494049, + 0.00396487582474947, + 0.0033219167962670326, + 0.00400247098878026, + 0.003512867260724306, + 0.003766748122870922, + 0.004781687632203102, + 0.005660623777657747, + 0.004612776916474104, + 0.006197981536388397, + 0.0040368372574448586, + 0.004128876142203808, + 0.004222639836370945, + 0.0025849712546914816, + 0.004710123408585787, + 0.005820733495056629, + 0.003051617881283164, + 0.00426884088665247, + 0.0036759586073458195, + 0.004638828802853823, + 0.0025117010809481144, + 0.00201601954177022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.018178876489400864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a professor in his 50s, with thinning salt-and-pepper hair and wire-rimmed glasses, standing before a large, dark green blackboard in a traditional university lecture hall. He wears a classic brown tweed blazer with worn leather patches at the elbows, over a crisp light blue button-down shirt. His body is slightly turned away from the camera as he focuses intently on the blackboard, which is already partially covered with complex diagrams and mathematical symbols. With his right hand, he steadily and deliberately slides a small piece of white chalk across the surface, meticulously forming a new physics equation using Greek letters and symbols, each firm stroke releasing a faint burst of chalk dust. In the softly blurred foreground, the heads of the students are clearly visible, all hunched over their wooden tables, their hands moving in sync as they copy the professor's work into their notebooks. The scene is bathed in the even, practical glow of overhead lighting, creating an atmosphere of focused, tranquil, and attentive intelligence.", + "seed": 2398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_10/index_2254_seed_2398_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019397181831300257, + "vlm_logprob": -0.019397181831300257, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.019397181831300257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028851602226495743 + ], + "opposite_logprob_tokens": [ + -3.6538515090942383 + ], + "logprob_reward": -0.028851602226495743, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028851602226495743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25967803597450256, + 0.12476658821105957, + 0.07450057566165924, + 0.03853221610188484, + 0.13590261340141296, + 0.08375201374292374, + 0.17190256714820862, + 0.046142324805259705, + 0.03021330200135708, + 0.03460976108908653 + ], + "credit_map": [ + 0.028586462140083313, + 0.031305085867643356, + 0.014150413684546947, + 0.015603804029524326, + 0.04014822840690613, + 0.03868686035275459, + 0.032491348683834076, + 0.05870581418275833, + 0.013543998822569847, + 0.012052910402417183, + 0.006960645318031311, + 0.007147382479161024, + 0.01706448569893837, + 0.019311711192131042, + 0.014658301137387753, + 0.034027136862277985, + 0.0091347461566329, + 0.007079970091581345, + 0.004270929377526045, + 0.0047054109163582325, + 0.011362671852111816, + 0.011249944567680359, + 0.009194368496537209, + 0.01750253327190876, + 0.003341380273923278, + 0.004335608799010515, + 0.002447880571708083, + 0.0025851207319647074, + 0.0061977035365998745, + 0.005941214971244335, + 0.006245704833418131, + 0.007437601685523987, + 0.015087943524122238, + 0.01713171973824501, + 0.00858211424201727, + 0.010706906206905842, + 0.015381665900349617, + 0.02132209949195385, + 0.013408484868705273, + 0.03428167477250099, + 0.009536482393741608, + 0.010910207405686378, + 0.005986712872982025, + 0.00538709806278348, + 0.017873985692858696, + 0.012886377051472664, + 0.014184115454554558, + 0.006987027358263731, + 0.017551353201270103, + 0.012033911421895027, + 0.015592490322887897, + 0.009790212847292423, + 0.01918873004615307, + 0.033965352922677994, + 0.020191462710499763, + 0.0435890331864357, + 0.006340604275465012, + 0.004694676026701927, + 0.004710668232291937, + 0.003689451841637492, + 0.005427627824246883, + 0.008765259757637978, + 0.005832289345562458, + 0.00668173935264349, + 0.003465992398560047, + 0.005346940830349922, + 0.00321630435064435, + 0.002481495961546898, + 0.0035696967970579863, + 0.005362404976040125, + 0.0026678156573325396, + 0.004102648701518774, + 0.0038822696078568697, + 0.0042617470026016235, + 0.002609502524137497, + 0.0024711955338716507, + 0.005136249586939812, + 0.004965328145772219, + 0.003916794899851084, + 0.007366671692579985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.028851602226495743 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302748799324036 + ], + "opposite_logprob_tokens": [ + -3.7733027935028076 + ], + "logprob_reward": -0.023302748799324036, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302748799324036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27202627062797546, + 0.16473844647407532, + 0.08636868745088577, + 0.06725969165563583, + 0.17873743176460266, + 0.04448704421520233, + 0.08045728504657745, + 0.04523056745529175, + 0.027401797473430634, + 0.03329278901219368 + ], + "credit_map": [ + 0.027893852442502975, + 0.03374519199132919, + 0.02593512088060379, + 0.028461620211601257, + 0.029131729155778885, + 0.03542667254805565, + 0.03594428673386574, + 0.05548780784010887, + 0.017371032387018204, + 0.017962271347641945, + 0.01570812612771988, + 0.016149885952472687, + 0.019190557301044464, + 0.02216402068734169, + 0.020518368110060692, + 0.03567419946193695, + 0.00923865009099245, + 0.008077796548604965, + 0.004956481978297234, + 0.005180333741009235, + 0.018266992643475533, + 0.012730357237160206, + 0.012847041711211205, + 0.015071037225425243, + 0.006652345880866051, + 0.008167263120412827, + 0.004775151144713163, + 0.0042632208205759525, + 0.016374453902244568, + 0.010707532986998558, + 0.010521947406232357, + 0.005797786172479391, + 0.022631360217928886, + 0.022784775123000145, + 0.01480095461010933, + 0.01492940541356802, + 0.02160712331533432, + 0.02489451691508293, + 0.020868446677923203, + 0.036220867186784744, + 0.003290628083050251, + 0.004694673232734203, + 0.0028832873795181513, + 0.0030440997797995806, + 0.01108185201883316, + 0.006645564455538988, + 0.007386094890534878, + 0.0054608481004834175, + 0.004597017075866461, + 0.006253459956496954, + 0.003548235399648547, + 0.0035366166848689318, + 0.0074110934510827065, + 0.02156711556017399, + 0.007721875794231892, + 0.025821886956691742, + 0.005044379737228155, + 0.0044645881280303, + 0.00424934970214963, + 0.003182540647685528, + 0.00727143045514822, + 0.00809252168983221, + 0.006296772044152021, + 0.006628991104662418, + 0.002311468357220292, + 0.003664683783426881, + 0.0025160075165331364, + 0.002972506685182452, + 0.0037854108959436417, + 0.0032624711748212576, + 0.004573310259729624, + 0.004315940663218498, + 0.003377939574420452, + 0.004141008947044611, + 0.002094713971018791, + 0.0025304551236331463, + 0.00589084904640913, + 0.004774613305926323, + 0.005885140970349312, + 0.004598069936037064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.023302748799324036 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254387855529785 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254387855529785, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254387855529785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32085931301116943, + 0.1996500939130783, + 0.06339067965745926, + 0.0944301187992096, + 0.1005621999502182, + 0.03715871274471283, + 0.06244996190071106, + 0.049039576202631, + 0.018953926861286163, + 0.053505346179008484 + ], + "credit_map": [ + 0.03624839335680008, + 0.05017905682325363, + 0.01713155396282673, + 0.025504516437649727, + 0.033766601234674454, + 0.05063129588961601, + 0.030657729133963585, + 0.07674018293619156, + 0.025016481056809425, + 0.024344785138964653, + 0.022112492471933365, + 0.017841974273324013, + 0.021489886566996574, + 0.025903182104229927, + 0.023451922461390495, + 0.03948938101530075, + 0.006186898332089186, + 0.008883355185389519, + 0.0033259729389101267, + 0.004074010066688061, + 0.01099623367190361, + 0.010182655416429043, + 0.0058016106486320496, + 0.013939947821199894, + 0.01182086393237114, + 0.011011921800673008, + 0.007249744143337011, + 0.007758717983961105, + 0.008041408844292164, + 0.013447979465126991, + 0.010274248197674751, + 0.024825237691402435, + 0.012440829537808895, + 0.013721592724323273, + 0.00729667954146862, + 0.00865986105054617, + 0.010431420058012009, + 0.015709146857261658, + 0.008114819414913654, + 0.024187859147787094, + 0.0040971520356833935, + 0.004850026685744524, + 0.0038530079182237387, + 0.0024904636666178703, + 0.007539159618318081, + 0.005553591530770063, + 0.004469577223062515, + 0.0043057347647845745, + 0.0032573381904512644, + 0.00786744337528944, + 0.002936300588771701, + 0.0028945370577275753, + 0.005209054332226515, + 0.018741030246019363, + 0.0050040422938764095, + 0.016540218144655228, + 0.007265808992087841, + 0.005863578524440527, + 0.005420253146439791, + 0.004054966848343611, + 0.006434142589569092, + 0.005932922940701246, + 0.006160160526633263, + 0.007907749153673649, + 0.0020716991275548935, + 0.003564127953723073, + 0.0018541559111326933, + 0.0019355207914486527, + 0.0017670433735474944, + 0.0033642135094851255, + 0.0019380735466256738, + 0.002459093928337097, + 0.005938434507697821, + 0.007531235925853252, + 0.0049559990875422955, + 0.004553235601633787, + 0.007085687946528196, + 0.008439334109425545, + 0.00610022759065032, + 0.008901195600628853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005254387855529785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980225533246994 + ], + "opposite_logprob_tokens": [ + -3.5298023223876953 + ], + "logprob_reward": -0.02980225533246994, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980225533246994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31320083141326904, + 0.11623785644769669, + 0.10153504461050034, + 0.12375154346227646, + 0.14146003127098083, + 0.018489744514226913, + 0.06858963519334793, + 0.02561856247484684, + 0.058663878589868546, + 0.03245292976498604 + ], + "credit_map": [ + 0.043889112770557404, + 0.03272160142660141, + 0.033571507781744, + 0.027455786243081093, + 0.05045502632856369, + 0.043205875903367996, + 0.05480686575174332, + 0.027095025405287743, + 0.0074195461347699165, + 0.012252440676093102, + 0.01128250453621149, + 0.006755207199603319, + 0.03160719573497772, + 0.01902679353952408, + 0.016269274055957794, + 0.011624886654317379, + 0.008802264928817749, + 0.012215431779623032, + 0.006759257987141609, + 0.0049027372151613235, + 0.026778584346175194, + 0.015704134479165077, + 0.01249806210398674, + 0.013874555006623268, + 0.013961517252027988, + 0.011515950784087181, + 0.008003384806215763, + 0.008669677190482616, + 0.018362216651439667, + 0.01789148524403572, + 0.017073463648557663, + 0.028273846954107285, + 0.019474292173981667, + 0.014961536042392254, + 0.014669755473732948, + 0.010060136206448078, + 0.01779746264219284, + 0.018998360261321068, + 0.014989613555371761, + 0.030508873984217644, + 0.002417647745460272, + 0.002601094776764512, + 0.001245154649950564, + 0.0014564919983968139, + 0.002292404416948557, + 0.002729164669290185, + 0.002510192571207881, + 0.0032375906594097614, + 0.0031613705214112997, + 0.005847733002156019, + 0.002547794720157981, + 0.001982686808332801, + 0.004645693581551313, + 0.02295883372426033, + 0.0048202089965343475, + 0.022625302895903587, + 0.0027167699299752712, + 0.0028875796124339104, + 0.002418030984699726, + 0.0017146228346973658, + 0.005206183064728975, + 0.003728603944182396, + 0.0029731786344200373, + 0.003973590210080147, + 0.005504624452441931, + 0.005480136256664991, + 0.003973072860389948, + 0.005895171780139208, + 0.007275739219039679, + 0.007866621017456055, + 0.007347282022237778, + 0.01532122865319252, + 0.003420752240344882, + 0.003551201894879341, + 0.0031775967217981815, + 0.0021382065024226904, + 0.004725506063550711, + 0.004947799723595381, + 0.004695645067840815, + 0.005796216893941164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.02980225533246994 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977491494268179 + ], + "opposite_logprob_tokens": [ + -4.634774684906006 + ], + "logprob_reward": -0.00977491494268179, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977491494268179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3010900318622589, + 0.10937707871198654, + 0.01973874866962433, + 0.028642456978559494, + 0.05139510706067085, + 0.016174858435988426, + 0.25167784094810486, + 0.08445502072572708, + 0.061597976833581924, + 0.07585086673498154 + ], + "credit_map": [ + 0.043550316244363785, + 0.02979455515742302, + 0.029660942032933235, + 0.02825017087161541, + 0.033292580395936966, + 0.03939899057149887, + 0.03811335936188698, + 0.059029143303632736, + 0.011740550398826599, + 0.012914961203932762, + 0.008979574777185917, + 0.009787295013666153, + 0.009546006098389626, + 0.015577604994177818, + 0.012207167223095894, + 0.028623932972550392, + 0.002107854001224041, + 0.0035578920505940914, + 0.0011124050943180919, + 0.0017916342476382852, + 0.0036161907482892275, + 0.002901589497923851, + 0.0023678173311054707, + 0.002283367794007063, + 0.003236851654946804, + 0.0030405924189835787, + 0.001720606698654592, + 0.001878057257272303, + 0.004054644610732794, + 0.0044717430137097836, + 0.0029471588786691427, + 0.0072928061708807945, + 0.004411660134792328, + 0.0038869346026331186, + 0.0049014040268957615, + 0.0037358258850872517, + 0.00897610280662775, + 0.0075457277707755566, + 0.00750406738370657, + 0.010433385148644447, + 0.0015083062462508678, + 0.0020765922963619232, + 0.001362278824672103, + 0.0016036592423915863, + 0.002330942079424858, + 0.0021926057524979115, + 0.0032827197574079037, + 0.0018177550518885255, + 0.019630542024970055, + 0.017060458660125732, + 0.015369356609880924, + 0.00861132238060236, + 0.015157842077314854, + 0.07187958061695099, + 0.018617264926433563, + 0.08535149693489075, + 0.010844645090401173, + 0.00809195451438427, + 0.009797276929020882, + 0.0068492102436721325, + 0.01470388937741518, + 0.01217001024633646, + 0.011283976025879383, + 0.010714063420891762, + 0.005748291499912739, + 0.006277758628129959, + 0.004244505427777767, + 0.005825369618833065, + 0.007203921675682068, + 0.008353829383850098, + 0.006968837231397629, + 0.01697547361254692, + 0.007956167683005333, + 0.007819307036697865, + 0.007235920988023281, + 0.005548384971916676, + 0.012913551181554794, + 0.010015404783189297, + 0.009726780466735363, + 0.014635355211794376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.00977491494268179 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a professor in his 50s, with thinning salt-and-pepper hair and wire-rimmed glasses, standing before a large, dark green blackboard in a traditional university lecture hall. He wears a classic brown tweed blazer with worn leather patches at the elbows, over a crisp light blue button-down shirt. His body is slightly turned away from the camera as he focuses intently on the blackboard, which is already partially covered with complex diagrams and mathematical symbols. With his right hand, he steadily and deliberately slides a small piece of white chalk across the surface, meticulously forming a new physics equation using Greek letters and symbols, each firm stroke releasing a faint burst of chalk dust. In the softly blurred foreground, the heads of the students are clearly visible, all hunched over their wooden tables, their hands moving in sync as they copy the professor's work into their notebooks. The scene is bathed in the even, practical glow of overhead lighting, creating an atmosphere of focused, tranquil, and attentive intelligence.", + "seed": 2399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_10/index_2254_seed_2399_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02575660585425794, + "vlm_logprob": -0.02575660585425794, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.02575660585425794, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04091762378811836 + ], + "opposite_logprob_tokens": [ + -3.2909176349639893 + ], + "logprob_reward": -0.04091762378811836, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04091762378811836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42219045758247375, + 0.07754398137331009, + 0.036389537155628204, + 0.1510201394557953, + 0.06309767812490463, + 0.0826754942536354, + 0.03154893219470978, + 0.05870978161692619, + 0.03993092104792595, + 0.03689311817288399 + ], + "credit_map": [ + 0.07325451821088791, + 0.03601736202836037, + 0.043765585869550705, + 0.05248744785785675, + 0.04917052760720253, + 0.0548936128616333, + 0.08171382546424866, + 0.030887577682733536, + 0.011110377497971058, + 0.010073335841298103, + 0.009120079688727856, + 0.0114606823772192, + 0.010065827518701553, + 0.008854239247739315, + 0.0068874964490532875, + 0.00997194554656744, + 0.0053547825664281845, + 0.005961982067674398, + 0.0048686605878174305, + 0.004918746650218964, + 0.003300711512565613, + 0.004044071771204472, + 0.004558968357741833, + 0.003381611779332161, + 0.028465203940868378, + 0.015776166692376137, + 0.025742316618561745, + 0.019167765974998474, + 0.013515258207917213, + 0.017980234697461128, + 0.012273788452148438, + 0.018099408596754074, + 0.008459334261715412, + 0.011980261653661728, + 0.0062415278516709805, + 0.010051718913018703, + 0.007559377700090408, + 0.007864203304052353, + 0.005199158098548651, + 0.005742099601775408, + 0.012872071005403996, + 0.013750405982136726, + 0.007078562397509813, + 0.010862819850444794, + 0.008808244951069355, + 0.010131660848855972, + 0.009255550801753998, + 0.009916173294186592, + 0.006536336615681648, + 0.003487901296466589, + 0.004228391218930483, + 0.003348169382661581, + 0.0028054825961589813, + 0.0042622084729373455, + 0.003960212226957083, + 0.00292022991925478, + 0.008774452842772007, + 0.005480057559907436, + 0.006262549199163914, + 0.008274839259684086, + 0.0041797286830842495, + 0.006929964292794466, + 0.013042195700109005, + 0.005765994545072317, + 0.005675490479916334, + 0.004495353437960148, + 0.004920399282127619, + 0.0053498768247663975, + 0.0037072186823934317, + 0.004074116703122854, + 0.0066238935105502605, + 0.005084571428596973, + 0.004032892175018787, + 0.004145736340433359, + 0.004850024823099375, + 0.004979768302291632, + 0.003247680841013789, + 0.003896377980709076, + 0.005366310011595488, + 0.006374327465891838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.04091762378811836 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020600290969014168 + ], + "opposite_logprob_tokens": [ + -3.8956003189086914 + ], + "logprob_reward": -0.020600290969014168, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020600290969014168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1725192815065384, + 0.14573730528354645, + 0.04455823823809624, + 0.1761174499988556, + 0.15316005051136017, + 0.09340740740299225, + 0.05539838969707489, + 0.1032007485628128, + 0.029617080464959145, + 0.026284029707312584 + ], + "credit_map": [ + 0.029076948761940002, + 0.01869858428835869, + 0.015955107286572456, + 0.01721031591296196, + 0.024812212213873863, + 0.024039866402745247, + 0.021843696013092995, + 0.020882567390799522, + 0.02370844967663288, + 0.020795654505491257, + 0.01507606916129589, + 0.026245325803756714, + 0.012766596861183643, + 0.017580723389983177, + 0.017327560111880302, + 0.012236930429935455, + 0.008754180744290352, + 0.006324646528810263, + 0.004482558462768793, + 0.005959779489785433, + 0.0047998856753110886, + 0.005644319113343954, + 0.003744733752682805, + 0.00484813516959548, + 0.017887234687805176, + 0.019057026132941246, + 0.030584920197725296, + 0.01976916193962097, + 0.021390795707702637, + 0.018636425957083702, + 0.015411404892802238, + 0.03338048234581947, + 0.021592887118458748, + 0.022479524835944176, + 0.00955956894904375, + 0.02581106312572956, + 0.01892337016761303, + 0.017383236438035965, + 0.018431281670928, + 0.01897912286221981, + 0.012359943240880966, + 0.01668532006442547, + 0.008152241818606853, + 0.01277468353509903, + 0.01005365140736103, + 0.01124345138669014, + 0.010010626167058945, + 0.012127491645514965, + 0.007249281741678715, + 0.005132355727255344, + 0.005236274562776089, + 0.007278850767761469, + 0.0059341685846447945, + 0.0064351544715464115, + 0.009724765084683895, + 0.00840754248201847, + 0.014720390550792217, + 0.011647730134427547, + 0.010299498215317726, + 0.014632481150329113, + 0.01419520191848278, + 0.012108881957828999, + 0.008702433668076992, + 0.016894131898880005, + 0.00471219839528203, + 0.003418278880417347, + 0.004588679876178503, + 0.003968533128499985, + 0.0024116470012813807, + 0.0032344581559300423, + 0.004239045083522797, + 0.0030442422721534967, + 0.004565421026200056, + 0.00335112982429564, + 0.002542403293773532, + 0.003260218072682619, + 0.002325694775208831, + 0.003136201063171029, + 0.0029370621778070927, + 0.0041658999398350716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.020600290969014168 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820563269779086 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.002820563269779086, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820563269779086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35153165459632874, + 0.11517062038183212, + 0.08563762158155441, + 0.1243446096777916, + 0.0967438593506813, + 0.040294099599123, + 0.06254573166370392, + 0.06128966435790062, + 0.028373906388878822, + 0.0340682715177536 + ], + "credit_map": [ + 0.058375466614961624, + 0.04893213510513306, + 0.035106103867292404, + 0.043224409222602844, + 0.03383949026465416, + 0.04385407269001007, + 0.05906170234084129, + 0.029138285666704178, + 0.02992069721221924, + 0.015246756374835968, + 0.00858784094452858, + 0.014207961037755013, + 0.009833154268562794, + 0.01568041741847992, + 0.013238970190286636, + 0.008454830385744572, + 0.02102225087583065, + 0.011907903477549553, + 0.006444119848310947, + 0.01167299598455429, + 0.005921101197600365, + 0.011628787964582443, + 0.009792199358344078, + 0.007248264737427235, + 0.0233763437718153, + 0.014067387208342552, + 0.01600492373108864, + 0.019962331280112267, + 0.00787663459777832, + 0.014867957681417465, + 0.014378312975168228, + 0.013810722157359123, + 0.017197037115693092, + 0.010943279601633549, + 0.01075452659279108, + 0.01336832158267498, + 0.009004335850477219, + 0.011271657422184944, + 0.013093442656099796, + 0.011111256666481495, + 0.008210370317101479, + 0.004654986783862114, + 0.004594468045979738, + 0.00596892274916172, + 0.00326913851313293, + 0.0048835319466888905, + 0.005928870756179094, + 0.002783812116831541, + 0.016066525131464005, + 0.006738448981195688, + 0.006639193743467331, + 0.008262655697762966, + 0.00507868779823184, + 0.009505022317171097, + 0.006458827294409275, + 0.003796370467171073, + 0.010793549008667469, + 0.0072766137309372425, + 0.006158739328384399, + 0.008106181398034096, + 0.004920997191220522, + 0.007392325438559055, + 0.010500963777303696, + 0.006140296813100576, + 0.005367559846490622, + 0.0034946652594953775, + 0.002997454721480608, + 0.005149153061211109, + 0.0021106209605932236, + 0.0031623172108083963, + 0.0032050651498138905, + 0.002887070644646883, + 0.005809708498418331, + 0.004261524882167578, + 0.003391200676560402, + 0.005042676813900471, + 0.003237227676436305, + 0.003663093550130725, + 0.004470088519155979, + 0.004192751366645098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.002820563269779086 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038106366991996765 + ], + "opposite_logprob_tokens": [ + -3.2881064414978027 + ], + "logprob_reward": -0.038106366991996765, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038106366991996765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3998710811138153, + 0.09877242892980576, + 0.053846269845962524, + 0.09024298191070557, + 0.03711218386888504, + 0.035252172499895096, + 0.08624698966741562, + 0.10059650242328644, + 0.06314797699451447, + 0.034911442548036575 + ], + "credit_map": [ + 0.055592697113752365, + 0.05748564377427101, + 0.039442818611860275, + 0.048743173480033875, + 0.044918276369571686, + 0.04573400318622589, + 0.0685965046286583, + 0.03935793787240982, + 0.01206018216907978, + 0.014400207437574863, + 0.012522215954959393, + 0.011939279735088348, + 0.012968608178198338, + 0.011047182604670525, + 0.008810902945697308, + 0.015023844316601753, + 0.009719029068946838, + 0.0092720752581954, + 0.0032642255537211895, + 0.008807402104139328, + 0.0050618830136954784, + 0.006818858440965414, + 0.0045080301351845264, + 0.006394767202436924, + 0.011664698831737041, + 0.012093049474060535, + 0.012779850512742996, + 0.01549132727086544, + 0.007734094746410847, + 0.009927747771143913, + 0.010801753960549831, + 0.009750460274517536, + 0.005993356928229332, + 0.00529180932790041, + 0.00435249088332057, + 0.0053351884707808495, + 0.004177355207502842, + 0.004279353190213442, + 0.00346040865406394, + 0.004222217481583357, + 0.004320694599300623, + 0.004683797247707844, + 0.003931484185159206, + 0.005459080450236797, + 0.004210342187434435, + 0.0035187765024602413, + 0.005992989521473646, + 0.0031350082717835903, + 0.02100491151213646, + 0.006707774009555578, + 0.010427550412714481, + 0.009165302850306034, + 0.008152526803314686, + 0.01183009147644043, + 0.011688406579196453, + 0.00727042555809021, + 0.012868239544332027, + 0.007693278603255749, + 0.014957132749259472, + 0.012398965656757355, + 0.0075095952488482, + 0.011029344983398914, + 0.025350848212838173, + 0.008789103478193283, + 0.008932424709200859, + 0.007339413743466139, + 0.008014649152755737, + 0.01018980797380209, + 0.007057045120745897, + 0.007019632495939732, + 0.008906222879886627, + 0.005688774865120649, + 0.004491693805903196, + 0.0036527973134070635, + 0.004338991362601519, + 0.004584884736686945, + 0.0037632877938449383, + 0.003883589990437031, + 0.003885784884914756, + 0.006310410797595978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.038106366991996765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338184252381325 + ], + "opposite_logprob_tokens": [ + -3.6513381004333496 + ], + "logprob_reward": -0.026338184252381325, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338184252381325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1908893883228302, + 0.05322146415710449, + 0.058201324194669724, + 0.5265782475471497, + 0.04624493047595024, + 0.034673213958740234, + 0.012047085911035538, + 0.027015723288059235, + 0.035034436732530594, + 0.016094181686639786 + ], + "credit_map": [ + 0.029956813901662827, + 0.025084225460886955, + 0.01764616370201111, + 0.021925844252109528, + 0.017376719042658806, + 0.02420295588672161, + 0.0395946130156517, + 0.015102054923772812, + 0.008189050480723381, + 0.008162333630025387, + 0.003923789132386446, + 0.005929552949965, + 0.006054698023945093, + 0.006281876005232334, + 0.009341330267488956, + 0.005338835064321756, + 0.008210625499486923, + 0.009783989749848843, + 0.005044755060225725, + 0.007134858053177595, + 0.005475255660712719, + 0.007233435288071632, + 0.008715355768799782, + 0.006603049114346504, + 0.07992439717054367, + 0.0489615872502327, + 0.09142379462718964, + 0.058976899832487106, + 0.055931396782398224, + 0.054685257375240326, + 0.04311622306704521, + 0.09355874359607697, + 0.00659940205514431, + 0.007013066206127405, + 0.006427795626223087, + 0.0055116028524935246, + 0.004991232417523861, + 0.005113035906106234, + 0.0069671147502958775, + 0.0036216771695762873, + 0.0041223024018108845, + 0.00366213358938694, + 0.005877029150724411, + 0.006001586560159922, + 0.0034025460481643677, + 0.003798767691478133, + 0.004933118354529142, + 0.002875732257962227, + 0.0020312946289777756, + 0.0013561868108808994, + 0.0013668009778484702, + 0.0015021625440567732, + 0.0012409872142598033, + 0.0013860829640179873, + 0.002006679540500045, + 0.0011568908812478185, + 0.003705553011968732, + 0.004213319625705481, + 0.002262072404846549, + 0.003919417038559914, + 0.003335462650284171, + 0.002953605493530631, + 0.002844407921656966, + 0.0037818853743374348, + 0.005872701294720173, + 0.005087466910481453, + 0.004005658905953169, + 0.005433813668787479, + 0.003923560027033091, + 0.0038117796648293734, + 0.0032680965960025787, + 0.00363136176019907, + 0.0023219448048621416, + 0.0021036930847913027, + 0.0013948027044534683, + 0.002519437810406089, + 0.0016003288328647614, + 0.0016999358776956797, + 0.0015642085345461965, + 0.0028898310847580433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.026338184252381325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a professor in his 50s, with thinning salt-and-pepper hair and wire-rimmed glasses, standing before a large, dark green blackboard in a traditional university lecture hall. He wears a classic brown tweed blazer with worn leather patches at the elbows, over a crisp light blue button-down shirt. His body is slightly turned away from the camera as he focuses intently on the blackboard, which is already partially covered with complex diagrams and mathematical symbols. With his right hand, he steadily and deliberately slides a small piece of white chalk across the surface, meticulously forming a new physics equation using Greek letters and symbols, each firm stroke releasing a faint burst of chalk dust. In the softly blurred foreground, the heads of the students are clearly visible, all hunched over their wooden tables, their hands moving in sync as they copy the professor's work into their notebooks. The scene is bathed in the even, practical glow of overhead lighting, creating an atmosphere of focused, tranquil, and attentive intelligence.", + "seed": 2400, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_10/index_2254_seed_2400_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09780493006110191, + "vlm_logprob": -0.09780493006110191, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.09780493006110191, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2850329577922821 + ], + "opposite_logprob_tokens": [ + -1.4100329875946045 + ], + "logprob_reward": -0.2850329577922821, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2850329577922821, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.334054172039032, + 0.11176542937755585, + 0.05955631285905838, + 0.04809393361210823, + 0.044682830572128296, + 0.1160680428147316, + 0.12812602519989014, + 0.04272471368312836, + 0.05953647568821907, + 0.05539208650588989 + ], + "credit_map": [ + 0.0440482534468174, + 0.042569566518068314, + 0.03152868524193764, + 0.0313849151134491, + 0.0571872740983963, + 0.058255087584257126, + 0.041170842945575714, + 0.027909545227885246, + 0.016079628840088844, + 0.02100064791738987, + 0.011429542675614357, + 0.01142896804958582, + 0.016106849536299706, + 0.014917653985321522, + 0.012981764040887356, + 0.007820378988981247, + 0.008763641119003296, + 0.010615795850753784, + 0.007578036282211542, + 0.0066835624165833, + 0.007312166038900614, + 0.0064671277068555355, + 0.0076980190351605415, + 0.004437961149960756, + 0.005856484640389681, + 0.009311523288488388, + 0.004402040969580412, + 0.005228482652455568, + 0.006411275360733271, + 0.007319045253098011, + 0.00520375045016408, + 0.004361328668892384, + 0.004900314379483461, + 0.007169140037149191, + 0.005628798622637987, + 0.006149948574602604, + 0.005638566333800554, + 0.006542575545608997, + 0.004278619308024645, + 0.004374872427433729, + 0.010232320986688137, + 0.013989127241075039, + 0.012019877322018147, + 0.013446781784296036, + 0.021123221144080162, + 0.020783806219697, + 0.014303161762654781, + 0.010169750079512596, + 0.010652486234903336, + 0.018574541434645653, + 0.013297860510647297, + 0.012718068435788155, + 0.02370835840702057, + 0.023152567446231842, + 0.014118017628788948, + 0.011904135346412659, + 0.004401232115924358, + 0.006401645019650459, + 0.005039521958678961, + 0.0037083709612488747, + 0.007572986651211977, + 0.005819874815642834, + 0.005784003995358944, + 0.0039970763027668, + 0.011694694869220257, + 0.011068279854953289, + 0.0054199909791350365, + 0.004488168749958277, + 0.00849310401827097, + 0.007656453642994165, + 0.006477714981883764, + 0.0042380718514323235, + 0.007790322881191969, + 0.007773969788104296, + 0.006421163212507963, + 0.006227631121873856, + 0.005735030863434076, + 0.008125118911266327, + 0.007291035261005163, + 0.006027812138199806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.2850329577922821 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07907778024673462 + ], + "opposite_logprob_tokens": [ + -2.57907772064209 + ], + "logprob_reward": -0.07907778024673462, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07907778024673462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3342588245868683, + 0.10779156535863876, + 0.08127282559871674, + 0.09689154475927353, + 0.044232260435819626, + 0.05946453660726547, + 0.061220213770866394, + 0.05064487084746361, + 0.061201587319374084, + 0.10302171111106873 + ], + "credit_map": [ + 0.04092017561197281, + 0.05975249782204628, + 0.03879636153578758, + 0.038177791982889175, + 0.043033894151449203, + 0.0458855926990509, + 0.03624057397246361, + 0.03145196661353111, + 0.013647891581058502, + 0.015539368614554405, + 0.01255084853619337, + 0.010960716754198074, + 0.0178042221814394, + 0.017957568168640137, + 0.012233210727572441, + 0.007097742520272732, + 0.009897789917886257, + 0.015363997779786587, + 0.007540897466242313, + 0.0067420233972370625, + 0.007303279358893633, + 0.017256101593375206, + 0.008046933449804783, + 0.009121808223426342, + 0.011335253715515137, + 0.016472188755869865, + 0.012716611847281456, + 0.012611146084964275, + 0.012616277672350407, + 0.012114174664020538, + 0.011182975023984909, + 0.00784291885793209, + 0.005766164045780897, + 0.008249015547335148, + 0.004189014434814453, + 0.004574159625917673, + 0.006640980951488018, + 0.00590478302910924, + 0.00579686788842082, + 0.003111274214461446, + 0.005498509854078293, + 0.008363709785044193, + 0.006016520317643881, + 0.0077482303604483604, + 0.00950951874256134, + 0.009109812788665295, + 0.00828077457845211, + 0.004937462043017149, + 0.007541375700384378, + 0.01287977583706379, + 0.005204720422625542, + 0.006089947186410427, + 0.010972985066473484, + 0.008204582147300243, + 0.0061815534718334675, + 0.004145272076129913, + 0.005032379645854235, + 0.008621280081570148, + 0.004733710549771786, + 0.005268539302051067, + 0.009219414554536343, + 0.005939324852079153, + 0.006916250102221966, + 0.004913975019007921, + 0.01030202116817236, + 0.008653233759105206, + 0.006215771194547415, + 0.006940324325114489, + 0.007721560541540384, + 0.007154515944421291, + 0.008675992488861084, + 0.005538167431950569, + 0.01686662994325161, + 0.015220219269394875, + 0.009888684377074242, + 0.009518217295408249, + 0.012149924412369728, + 0.01798977330327034, + 0.013810201548039913, + 0.007578068878501654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07907778024673462 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082679852843285 + ], + "opposite_logprob_tokens": [ + -4.141082763671875 + ], + "logprob_reward": -0.016082679852843285, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082679852843285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31998780369758606, + 0.2095547318458557, + 0.08840066939592361, + 0.09013449400663376, + 0.09282304346561432, + 0.0395229272544384, + 0.04269685223698616, + 0.03745894879102707, + 0.027919696643948555, + 0.051500797271728516 + ], + "credit_map": [ + 0.039507001638412476, + 0.05226251110434532, + 0.02682037092745304, + 0.02830106019973755, + 0.053114164620637894, + 0.06130033731460571, + 0.031506866216659546, + 0.027175478637218475, + 0.02397870272397995, + 0.036609847098588943, + 0.02531188353896141, + 0.021883247420191765, + 0.027152469381690025, + 0.03383374586701393, + 0.02159750461578369, + 0.019187336787581444, + 0.009494403377175331, + 0.012631834484636784, + 0.008521536365151405, + 0.009435821324586868, + 0.013277539052069187, + 0.015077735297381878, + 0.011261881329119205, + 0.008699918165802956, + 0.009677371941506863, + 0.014517620205879211, + 0.009629444219172001, + 0.010095002129673958, + 0.013693047687411308, + 0.012981826439499855, + 0.011482808738946915, + 0.008057370781898499, + 0.007887731306254864, + 0.01837952621281147, + 0.008803521282970905, + 0.008557286113500595, + 0.014074202626943588, + 0.01610713079571724, + 0.010216268710792065, + 0.008797372691333294, + 0.003958612680435181, + 0.005857198964804411, + 0.0038119573146104813, + 0.004232343286275864, + 0.0072473809123039246, + 0.006135296076536179, + 0.00475599430501461, + 0.00352414115332067, + 0.0033009247854351997, + 0.00598537502810359, + 0.004250063095241785, + 0.004547279328107834, + 0.008583838120102882, + 0.007152950391173363, + 0.004670456517487764, + 0.004205965902656317, + 0.0033866961020976305, + 0.005494918208569288, + 0.003323301440104842, + 0.003764866152778268, + 0.0071692862547934055, + 0.006044960580766201, + 0.00450816098600626, + 0.0037667611613869667, + 0.003560857381671667, + 0.004522575531154871, + 0.0029938272200524807, + 0.0033343564718961716, + 0.00393670704215765, + 0.004274384118616581, + 0.003104934236034751, + 0.002192054409533739, + 0.006998012773692608, + 0.008886436931788921, + 0.005186723545193672, + 0.004947098437696695, + 0.0063507528975605965, + 0.009416202083230019, + 0.005361455958336592, + 0.0043541137129068375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.016082679852843285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903383672237396 + ], + "opposite_logprob_tokens": [ + -2.579033851623535 + ], + "logprob_reward": -0.07903383672237396, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903383672237396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31287190318107605, + 0.13661490380764008, + 0.04475995525717735, + 0.11722924560308456, + 0.08067701011896133, + 0.11722629517316818, + 0.06861019879579544, + 0.04329786077141762, + 0.028521304950118065, + 0.05019128695130348 + ], + "credit_map": [ + 0.03523198887705803, + 0.04158748686313629, + 0.028020069003105164, + 0.0306309275329113, + 0.060961078852415085, + 0.05625545233488083, + 0.03393871709704399, + 0.026246197521686554, + 0.017286648973822594, + 0.023664813488721848, + 0.01462671160697937, + 0.01350195799022913, + 0.01737935096025467, + 0.02556474506855011, + 0.01285547949373722, + 0.011735193431377411, + 0.004973392002284527, + 0.009067564271390438, + 0.004476412199437618, + 0.0038964718114584684, + 0.007031367626041174, + 0.007158190477639437, + 0.004984487779438496, + 0.0031720695551484823, + 0.013209957629442215, + 0.017600223422050476, + 0.012187875807285309, + 0.015317581593990326, + 0.015584960579872131, + 0.018967965617775917, + 0.014786670915782452, + 0.00957400817424059, + 0.007773459888994694, + 0.015052401460707188, + 0.008173782378435135, + 0.00786628294736147, + 0.010934408754110336, + 0.014758508652448654, + 0.008291371166706085, + 0.00782679207623005, + 0.010185625404119492, + 0.015363495796918869, + 0.012690005823969841, + 0.012817510403692722, + 0.020106296986341476, + 0.020552482455968857, + 0.015584951266646385, + 0.009925922378897667, + 0.005704378243535757, + 0.01041764859110117, + 0.006108475383371115, + 0.007509833201766014, + 0.014310258440673351, + 0.011729677207767963, + 0.007153049111366272, + 0.005676876287907362, + 0.005540359299629927, + 0.006273873616009951, + 0.005020467098802328, + 0.005204171873629093, + 0.006052148528397083, + 0.005397138651460409, + 0.005998199339956045, + 0.0038115030620247126, + 0.004377542994916439, + 0.004700816702097654, + 0.0028658509254455566, + 0.0029835752211511135, + 0.004002493340522051, + 0.003521108068525791, + 0.0029596369713544846, + 0.0031102821230888367, + 0.006269445177167654, + 0.00720938527956605, + 0.006148248910903931, + 0.005580130033195019, + 0.006335200276225805, + 0.008672362193465233, + 0.005542552564293146, + 0.004433962982147932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07903383672237396 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professor in his 50s with thinning salt-and-pepper hair and wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the professor wearing a brown tweed blazer with worn leather patches at the elbows over a light blue button-down shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the professor standing before a large, dark green blackboard covered with complex diagrams and mathematical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the professor using his right hand to slide white chalk across the blackboard to form a physics equation with Greek letters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the heads of students visible in the softly blurred foreground, hunched over wooden tables and copying the professor's work?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797395691275597 + ], + "opposite_logprob_tokens": [ + -3.529797315597534 + ], + "logprob_reward": -0.029797395691275597, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797395691275597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14584727585315704, + 0.038330934941768646, + 0.04124041274189949, + 0.06330648809671402, + 0.02380618453025818, + 0.03430178761482239, + 0.022738849744200706, + 0.037171512842178345, + 0.05207919329404831, + 0.5411773324012756 + ], + "credit_map": [ + 0.011609084904193878, + 0.019963588565587997, + 0.01859842985868454, + 0.01656912826001644, + 0.021698161959648132, + 0.023935504257678986, + 0.02092585526406765, + 0.012547514401376247, + 0.004012949299067259, + 0.005387518089264631, + 0.0038237618282437325, + 0.0037017809227108955, + 0.005729710217565298, + 0.005873300600796938, + 0.005561937112361193, + 0.00423997500911355, + 0.006767536047846079, + 0.006003460846841335, + 0.004275845363736153, + 0.004775772336870432, + 0.004873664118349552, + 0.005367353092879057, + 0.005550914444029331, + 0.0036258643958717585, + 0.00968204066157341, + 0.009950927458703518, + 0.006605256348848343, + 0.00787933636456728, + 0.008765200152993202, + 0.007129549980163574, + 0.008187920786440372, + 0.005106260068714619, + 0.003833330934867263, + 0.003824962768703699, + 0.0021742258686572313, + 0.0027742378879338503, + 0.003776880446821451, + 0.0025593729224056005, + 0.003183158580213785, + 0.0016800153534859419, + 0.004050320480018854, + 0.004700944293290377, + 0.0034608556888997555, + 0.00341350631788373, + 0.0059034936130046844, + 0.005682205315679312, + 0.004356485325843096, + 0.002733974950388074, + 0.0021777362562716007, + 0.0037083954084664583, + 0.002402606653049588, + 0.00200238567776978, + 0.004319681320339441, + 0.003441351465880871, + 0.002928854199126363, + 0.001757840858772397, + 0.006008632481098175, + 0.006005159579217434, + 0.003490470815449953, + 0.003273200010880828, + 0.005491180811077356, + 0.0058577400632202625, + 0.004415306728333235, + 0.002629825146868825, + 0.007570213638246059, + 0.008585536852478981, + 0.006898741237819195, + 0.006811996456235647, + 0.004604359157383442, + 0.006850088480859995, + 0.006223998032510281, + 0.004534263629466295, + 0.08299639075994492, + 0.08455026149749756, + 0.06037888675928116, + 0.06079178303480148, + 0.05182554945349693, + 0.08876539766788483, + 0.06463341414928436, + 0.047235652804374695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.029797395691275597 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_11/index_tensor([4959, 4959])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_11/index_tensor([4959, 4959])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..52fa1193ee28b97ddac0647489c3bbe4778871f3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_11/index_tensor([4959, 4959])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in a surreal macro-style with a high-frame-rate camera, this static, low-angle, extreme close-up captures a moment of immense power in breathtaking slow motion. Within the industrial metal confines of a capital ship's hangar bay, the heavy magnetic launch clamps securing an Eta-2 Jedi starfighter mechanically snap open. This decisive action sends a fiery, brilliant burst of sparks across the frame, each particle a fleeting trace of fire. Simultaneously, the fighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma, filling the scene with intense, high-contrast light. This harsh lighting gleams across the starfighter's fuselage, its sharp focus revealing every intricate, weathered detail: a tapestry of burn marks, microscopic scratches in the paint, and subtle indentations in the metal plating. The powerful glow carves deep shadows into the surrounding launch bay floor, creating a dramatic and tense atmosphere of unleashed energy.", + "seed": 23111, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_11/index_4959_seed_23111_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2559706838801503, + "vlm_logprob": -0.2559706838801503, + "vlm_yes_no_margin": 2.4750000119209288, + "vlm_token_credit": -0.2559706838801503, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0823419094085693 + ], + "opposite_logprob_tokens": [ + -0.4573419690132141 + ], + "logprob_reward": -1.0823419094085693, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0823419094085693, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08363140374422073, + 0.06986227631568909, + 0.4152831733226776, + 0.04798354580998421, + 0.05878980830311775, + 0.05712150037288666, + 0.04425880312919617, + 0.13731281459331512, + 0.04351985827088356, + 0.04223678633570671 + ], + "credit_map": [ + 0.012851065956056118, + 0.011884478852152824, + 0.005592621862888336, + 0.012777605094015598, + 0.013266338035464287, + 0.010874461382627487, + 0.009071024134755135, + 0.00731382193043828, + 0.011158171109855175, + 0.009398497641086578, + 0.008283578790724277, + 0.007098868954926729, + 0.010152428410947323, + 0.009468561969697475, + 0.007256863638758659, + 0.007045311387628317, + 0.048893190920352936, + 0.04414582997560501, + 0.038495831191539764, + 0.06068703904747963, + 0.044654592871665955, + 0.041153308004140854, + 0.08150868117809296, + 0.055744726210832596, + 0.005126951262354851, + 0.004819628316909075, + 0.004227032419294119, + 0.005311624146997929, + 0.007361299823969603, + 0.00713002635166049, + 0.0077234311029314995, + 0.006283557508140802, + 0.006236087065190077, + 0.006452504545450211, + 0.004386504180729389, + 0.004111939575523138, + 0.005981643218547106, + 0.011691120453178883, + 0.010283736512064934, + 0.009646285325288773, + 0.006470447406172752, + 0.005527503788471222, + 0.006926683709025383, + 0.005419115535914898, + 0.005120222922414541, + 0.009821868501603603, + 0.008712050504982471, + 0.009123614989221096, + 0.0034497459419071674, + 0.004072492942214012, + 0.004014493431895971, + 0.003815307514742017, + 0.004426140803843737, + 0.010580077767372131, + 0.006822608411312103, + 0.007077943533658981, + 0.011646315455436707, + 0.012470372952520847, + 0.019333934411406517, + 0.01704009249806404, + 0.011828194372355938, + 0.022804470732808113, + 0.019831564277410507, + 0.02235788106918335, + 0.003113009501248598, + 0.0038974429480731487, + 0.00354402931407094, + 0.004890115931630135, + 0.004571243654936552, + 0.006865791045129299, + 0.008282552473247051, + 0.008355678990483284, + 0.004163376986980438, + 0.005478141829371452, + 0.0025167022831737995, + 0.003382754512131214, + 0.0038383614737540483, + 0.00684430543333292, + 0.00764503562822938, + 0.00836811400949955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -1.0823419094085693 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008995115756989 + ], + "opposite_logprob_tokens": [ + -2.6950900554656982 + ], + "logprob_reward": -0.07008995115756989, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008995115756989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16960008442401886, + 0.030874183401465416, + 0.12784692645072937, + 0.03728047013282776, + 0.0801193118095398, + 0.07988588511943817, + 0.07379499822854996, + 0.09514036029577255, + 0.021517537534236908, + 0.28394022583961487 + ], + "credit_map": [ + 0.02518349513411522, + 0.026216277852654457, + 0.01258831936866045, + 0.02334630861878395, + 0.022905802354216576, + 0.020540282130241394, + 0.01759546622633934, + 0.021224144846200943, + 0.007617054972797632, + 0.004775435663759708, + 0.0029449446592479944, + 0.002876281039789319, + 0.0035031025763601065, + 0.002678322372958064, + 0.0030160669703036547, + 0.0034629739820957184, + 0.009783647023141384, + 0.013726229779422283, + 0.010198220610618591, + 0.021352674812078476, + 0.01619645021855831, + 0.01326652243733406, + 0.026798153296113014, + 0.016525020822882652, + 0.0041891601867973804, + 0.0038643148727715015, + 0.004513887222856283, + 0.004170059226453304, + 0.0036916928365826607, + 0.006717969663441181, + 0.005693351849913597, + 0.004440035670995712, + 0.006460911128669977, + 0.009087505750358105, + 0.0068000550381839275, + 0.008335503749549389, + 0.005952082574367523, + 0.010722188279032707, + 0.01729360595345497, + 0.015467457473278046, + 0.007048619445413351, + 0.008271065540611744, + 0.011277684941887856, + 0.009356027469038963, + 0.007825353182852268, + 0.011464418843388557, + 0.01242880616337061, + 0.012213916517794132, + 0.00590387312695384, + 0.00778397498652339, + 0.005849975161254406, + 0.007080578710883856, + 0.0068971505388617516, + 0.018626170232892036, + 0.011078992858529091, + 0.010574286803603172, + 0.006360247265547514, + 0.0067386021837592125, + 0.008364610373973846, + 0.006900405045598745, + 0.011118550784885883, + 0.026654906570911407, + 0.01628795452415943, + 0.012715086340904236, + 0.0021528350189328194, + 0.0024791406467556953, + 0.002030501142144203, + 0.002040349179878831, + 0.002663983963429928, + 0.0028855048585683107, + 0.003485624212771654, + 0.003779600141569972, + 0.031088419258594513, + 0.048705510795116425, + 0.02513115666806698, + 0.029489140957593918, + 0.028003020212054253, + 0.029124949127435684, + 0.04746561869978905, + 0.04493242874741554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.07008995115756989 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008550316095352 + ], + "opposite_logprob_tokens": [ + -2.6950855255126953 + ], + "logprob_reward": -0.07008550316095352, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008550316095352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11930903792381287, + 0.05523700639605522, + 0.18587376177310944, + 0.05551093444228172, + 0.150636225938797, + 0.12559238076210022, + 0.08202847093343735, + 0.08417195826768875, + 0.04360737279057503, + 0.09803289920091629 + ], + "credit_map": [ + 0.018173351883888245, + 0.01873709261417389, + 0.0083216717466712, + 0.017949944362044334, + 0.017581094056367874, + 0.013393569737672806, + 0.012064442038536072, + 0.013087867759168148, + 0.010199290700256824, + 0.007342024240642786, + 0.005434549879282713, + 0.005814845208078623, + 0.008299889974296093, + 0.006861462723463774, + 0.005536472424864769, + 0.005748472176492214, + 0.024065418168902397, + 0.020139820873737335, + 0.022238442674279213, + 0.026067078113555908, + 0.022857876494526863, + 0.018511345610022545, + 0.034230977296829224, + 0.017762785777449608, + 0.007019094191491604, + 0.005288452375680208, + 0.004506511613726616, + 0.00755297252908349, + 0.005691115278750658, + 0.008360866457223892, + 0.009844363667070866, + 0.007247561123222113, + 0.013134412467479706, + 0.018263554200530052, + 0.013889864087104797, + 0.015427856706082821, + 0.010835585184395313, + 0.01854053884744644, + 0.032681602984666824, + 0.027862807735800743, + 0.011011466383934021, + 0.011069205589592457, + 0.02334100753068924, + 0.014566061086952686, + 0.009061954915523529, + 0.01638292707502842, + 0.018224548548460007, + 0.02193521149456501, + 0.006916490849107504, + 0.005846410058438778, + 0.008909626863896847, + 0.008239208720624447, + 0.006425534375011921, + 0.018617596477270126, + 0.013410653918981552, + 0.013662955723702908, + 0.007330084685236216, + 0.009419898502528667, + 0.008517704904079437, + 0.007529360242187977, + 0.008718626573681831, + 0.013543156906962395, + 0.014986229129135609, + 0.014126901514828205, + 0.004436003044247627, + 0.0038211627397686243, + 0.003335419809445739, + 0.00405128113925457, + 0.004182612989097834, + 0.007304711267352104, + 0.00943069439381361, + 0.0070454878732562065, + 0.006484700832515955, + 0.011350242421030998, + 0.005770009942352772, + 0.008938804268836975, + 0.007699769921600819, + 0.008790813386440277, + 0.025814129039645195, + 0.023184429854154587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.07008550316095352 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086548225954175 + ], + "opposite_logprob_tokens": [ + -4.758654594421387 + ], + "logprob_reward": -0.0086548225954175, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086548225954175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13190510869026184, + 0.041041299700737, + 0.16858160495758057, + 0.06237145885825157, + 0.06674502044916153, + 0.12586671113967896, + 0.11549423635005951, + 0.1090056523680687, + 0.0636984258890152, + 0.11529048532247543 + ], + "credit_map": [ + 0.022220414131879807, + 0.020433878526091576, + 0.008003674447536469, + 0.021959267556667328, + 0.02366836555302143, + 0.014489770866930485, + 0.009222344495356083, + 0.011907394975423813, + 0.005267980974167585, + 0.005882231518626213, + 0.003610388608649373, + 0.0055686053819954395, + 0.006557044107466936, + 0.005553479306399822, + 0.003975377418100834, + 0.0046261935494840145, + 0.012156392447650433, + 0.01610148698091507, + 0.013674427755177021, + 0.03169592097401619, + 0.01771889254450798, + 0.01796727068722248, + 0.03691817820072174, + 0.022349046543240547, + 0.007376089692115784, + 0.006189052015542984, + 0.004707677289843559, + 0.009904132224619389, + 0.006384120788425207, + 0.008336641825735569, + 0.011945732869207859, + 0.0075280144810676575, + 0.006876196246594191, + 0.010720230638980865, + 0.004932572599500418, + 0.0061334571801126, + 0.009244863875210285, + 0.007804151624441147, + 0.009793111123144627, + 0.011240442283451557, + 0.013155562803149223, + 0.013210444711148739, + 0.016756465658545494, + 0.012766681611537933, + 0.01235821470618248, + 0.020105058327317238, + 0.015492241829633713, + 0.02202204428613186, + 0.011071056127548218, + 0.01024754997342825, + 0.013934124261140823, + 0.013873495161533356, + 0.012103294022381306, + 0.020240120589733124, + 0.017637331038713455, + 0.016387280076742172, + 0.010160676203668118, + 0.01146033313125372, + 0.012897878885269165, + 0.012755392119288445, + 0.011059475131332874, + 0.022012151777744293, + 0.013959013856947422, + 0.014700740575790405, + 0.006479940377175808, + 0.005447132047265768, + 0.005667846649885178, + 0.005016081500798464, + 0.007542417850345373, + 0.012353571131825447, + 0.011416071094572544, + 0.009775367565453053, + 0.010644775815308094, + 0.012592427432537079, + 0.010701595805585384, + 0.01203695684671402, + 0.016575174406170845, + 0.013475128449499607, + 0.020936043933033943, + 0.018328385427594185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0086548225954175 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868123307824135 + ], + "opposite_logprob_tokens": [ + -3.0486812591552734 + ], + "logprob_reward": -0.04868123307824135, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868123307824135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12710709869861603, + 0.0669131651520729, + 0.15800680220127106, + 0.08507116883993149, + 0.2122178077697754, + 0.08174968510866165, + 0.11747672408819199, + 0.04405714571475983, + 0.051552869379520416, + 0.05584752932190895 + ], + "credit_map": [ + 0.023049240931868553, + 0.020356765016913414, + 0.006671992130577564, + 0.017831670120358467, + 0.022117383778095245, + 0.0172959603369236, + 0.00982328038662672, + 0.009960804134607315, + 0.015185842290520668, + 0.010037668980658054, + 0.007401124574244022, + 0.0074839177541434765, + 0.007954074069857597, + 0.008197072893381119, + 0.004488909617066383, + 0.006164551246911287, + 0.01691240817308426, + 0.014228827320039272, + 0.013941456563770771, + 0.024555640295147896, + 0.015915529802441597, + 0.018545128405094147, + 0.03573945537209511, + 0.018168358132243156, + 0.012247199192643166, + 0.008293472230434418, + 0.008607667870819569, + 0.010820754803717136, + 0.008610653690993786, + 0.009934935718774796, + 0.013573000207543373, + 0.01298348419368267, + 0.02164013311266899, + 0.03671838343143463, + 0.017864437773823738, + 0.01698899269104004, + 0.01302679069340229, + 0.018137214705348015, + 0.05057531222701073, + 0.03726654872298241, + 0.007051119580864906, + 0.006991133559495211, + 0.013097599148750305, + 0.008622693829238415, + 0.005725087132304907, + 0.009729397483170033, + 0.013788547366857529, + 0.016744105145335197, + 0.01095280610024929, + 0.0071822647005319595, + 0.011277185752987862, + 0.010449185967445374, + 0.008644786663353443, + 0.024848472326993942, + 0.017355797812342644, + 0.0267662201076746, + 0.0038425279781222343, + 0.004255084786564112, + 0.0036135129630565643, + 0.0049509406089782715, + 0.005263325292617083, + 0.007947438396513462, + 0.006375267170369625, + 0.00780904944986105, + 0.0036458526737987995, + 0.004199146758764982, + 0.006768120918422937, + 0.006279682740569115, + 0.002935240510851145, + 0.006966772023588419, + 0.01075307559221983, + 0.010004978626966476, + 0.004961661994457245, + 0.008312112651765347, + 0.0040112747810781, + 0.005169383250176907, + 0.005594499409198761, + 0.005371443461626768, + 0.011407745070755482, + 0.011019404977560043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.04868123307824135 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a surreal macro-style with a high-frame-rate camera, this static, low-angle, extreme close-up captures a moment of immense power in breathtaking slow motion. Within the industrial metal confines of a capital ship's hangar bay, the heavy magnetic launch clamps securing an Eta-2 Jedi starfighter mechanically snap open. This decisive action sends a fiery, brilliant burst of sparks across the frame, each particle a fleeting trace of fire. Simultaneously, the fighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma, filling the scene with intense, high-contrast light. This harsh lighting gleams across the starfighter's fuselage, its sharp focus revealing every intricate, weathered detail: a tapestry of burn marks, microscopic scratches in the paint, and subtle indentations in the metal plating. The powerful glow carves deep shadows into the surrounding launch bay floor, creating a dramatic and tense atmosphere of unleashed energy.", + "seed": 23112, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_11/index_4959_seed_23112_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.23444860465824605, + "vlm_logprob": -0.23444860465824605, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.23444860465824605, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.991994321346283 + ], + "opposite_logprob_tokens": [ + -0.49199432134628296 + ], + "logprob_reward": -0.991994321346283, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.991994321346283, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10152553021907806, + 0.04008778557181358, + 0.036571234464645386, + 0.2513929009437561, + 0.062223225831985474, + 0.0786837711930275, + 0.1966269612312317, + 0.03504475951194763, + 0.07975532114505768, + 0.11808852851390839 + ], + "credit_map": [ + 0.01155929546803236, + 0.029556743800640106, + 0.006079959217458963, + 0.009662577882409096, + 0.00973045825958252, + 0.009377467446029186, + 0.008815787732601166, + 0.016743240877985954, + 0.004514120053499937, + 0.005296709481626749, + 0.0030676894821226597, + 0.006440293975174427, + 0.005590763874351978, + 0.00622199522331357, + 0.004816936794668436, + 0.004139276221394539, + 0.0027751997113227844, + 0.005466042552143335, + 0.00316016492433846, + 0.009156511165201664, + 0.005289972759783268, + 0.003544165752828121, + 0.0030004146974533796, + 0.004178761970251799, + 0.016829006373882294, + 0.010208682157099247, + 0.08272215723991394, + 0.036195557564496994, + 0.07516809552907944, + 0.014376090839505196, + 0.0074732666835188866, + 0.008420056663453579, + 0.004020280670374632, + 0.008386882953345776, + 0.009077677503228188, + 0.018125468865036964, + 0.007132422644644976, + 0.005095056723803282, + 0.005696307402104139, + 0.00468912860378623, + 0.004629787057638168, + 0.008942767046391964, + 0.012933889403939247, + 0.011406566947698593, + 0.007778763305395842, + 0.010652408935129642, + 0.014105402864515781, + 0.008234185166656971, + 0.014648382551968098, + 0.026707293465733528, + 0.030174817889928818, + 0.05189599469304085, + 0.013882176950573921, + 0.018125159665942192, + 0.02366555854678154, + 0.017527583986520767, + 0.0023047062568366528, + 0.004131633788347244, + 0.003540477016940713, + 0.006780819967389107, + 0.004498030990362167, + 0.005051834974437952, + 0.005299228709191084, + 0.003438028274103999, + 0.007250696886330843, + 0.023391878232359886, + 0.007696541957557201, + 0.011865162290632725, + 0.0071195876225829124, + 0.008540964685380459, + 0.008607565425336361, + 0.005282921250909567, + 0.006706472486257553, + 0.009747207164764404, + 0.01198592409491539, + 0.02056972309947014, + 0.011998577043414116, + 0.009824147447943687, + 0.03427284210920334, + 0.01298364158719778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.991994321346283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029834190383553505 + ], + "opposite_logprob_tokens": [ + -3.529834270477295 + ], + "logprob_reward": -0.029834190383553505, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029834190383553505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13165807723999023, + 0.030381863936781883, + 0.10271322727203369, + 0.2897607088088989, + 0.029105201363563538, + 0.07553660869598389, + 0.21171987056732178, + 0.010780432261526585, + 0.010000942274928093, + 0.10834305733442307 + ], + "credit_map": [ + 0.014475701376795769, + 0.0369497574865818, + 0.009690029546618462, + 0.013408918865025043, + 0.01489245519042015, + 0.014622176997363567, + 0.013203039765357971, + 0.014416011050343513, + 0.002657084260135889, + 0.003560601966455579, + 0.0033380042295902967, + 0.0047462498769164085, + 0.0054802317172288895, + 0.003945264965295792, + 0.004036612343043089, + 0.0026178171392530203, + 0.007466303184628487, + 0.017117194831371307, + 0.007218461483716965, + 0.02444741502404213, + 0.016968142241239548, + 0.008225795812904835, + 0.008208120241761208, + 0.01306180004030466, + 0.017270252108573914, + 0.013613326475024223, + 0.09302929788827896, + 0.04504161328077316, + 0.07553331553936005, + 0.017308054491877556, + 0.013668338768184185, + 0.014296488836407661, + 0.0019068052060902119, + 0.0030504337046295404, + 0.0034584272652864456, + 0.00587130431085825, + 0.0039219302125275135, + 0.002683151513338089, + 0.005122528877109289, + 0.003090618411079049, + 0.004814589396119118, + 0.008877202868461609, + 0.012806505896151066, + 0.010976726189255714, + 0.006027724593877792, + 0.010014102794229984, + 0.013791069388389587, + 0.008228682912886143, + 0.01657775416970253, + 0.02950354292988777, + 0.03281848877668381, + 0.05760467052459717, + 0.013875375501811504, + 0.019217709079384804, + 0.024247407913208008, + 0.017874928191304207, + 0.0008304245420731604, + 0.0013132133753970265, + 0.0009639442432671785, + 0.00150417594704777, + 0.0016654353821650147, + 0.001700327848084271, + 0.001553247100673616, + 0.0012496642302721739, + 0.0009693673346191645, + 0.0018527075881138444, + 0.0008384775137528777, + 0.002128917956724763, + 0.0009383910219185054, + 0.000941194302868098, + 0.0013855795841664076, + 0.000946306623518467, + 0.007092037703841925, + 0.011239934712648392, + 0.014970125630497932, + 0.016516583040356636, + 0.010574588552117348, + 0.009855956770479679, + 0.02545996941626072, + 0.012633860111236572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.029834190383553505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636837214231491 + ], + "opposite_logprob_tokens": [ + -3.6513683795928955 + ], + "logprob_reward": -0.02636837214231491, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636837214231491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03778131678700447, + 0.0406254306435585, + 0.07770058512687683, + 0.054077520966529846, + 0.09784257411956787, + 0.05635150521993637, + 0.07963709533214569, + 0.030649304389953613, + 0.03899853676557541, + 0.4863361716270447 + ], + "credit_map": [ + 0.003581029362976551, + 0.005905099678784609, + 0.002458928618580103, + 0.005287740845233202, + 0.006921105552464724, + 0.0040473914705216885, + 0.005433341488242149, + 0.004146680701524019, + 0.004578061401844025, + 0.005444734822958708, + 0.004214462824165821, + 0.0063300831243395805, + 0.006783864926546812, + 0.004932167008519173, + 0.004979903809726238, + 0.0033621543552726507, + 0.005113441962748766, + 0.010542734526097775, + 0.00578621681779623, + 0.024259014055132866, + 0.007253855466842651, + 0.004731392953544855, + 0.006471704691648483, + 0.013542220927774906, + 0.003213135991245508, + 0.0059469398111104965, + 0.014202233403921127, + 0.011505872942507267, + 0.005713962949812412, + 0.004522216971963644, + 0.005785497836768627, + 0.0031876603607088327, + 0.008359928615391254, + 0.011919603683054447, + 0.011743891052901745, + 0.022481953725218773, + 0.018408769741654396, + 0.00809977762401104, + 0.008733582682907581, + 0.008095066994428635, + 0.0043940250761806965, + 0.006251585204154253, + 0.007735212333500385, + 0.007979679852724075, + 0.004235770553350449, + 0.009162550792098045, + 0.009600410237908363, + 0.006992268841713667, + 0.005711153615266085, + 0.01054896879941225, + 0.011937101371586323, + 0.020084746181964874, + 0.006557228974997997, + 0.007814576849341393, + 0.009799578227102757, + 0.007183730602264404, + 0.002562842797487974, + 0.004172886721789837, + 0.0027594463899731636, + 0.004997427109628916, + 0.003993721213191748, + 0.005421294830739498, + 0.0038151023909449577, + 0.0029265820048749447, + 0.003402559319511056, + 0.007063580211251974, + 0.004229012876749039, + 0.006929021328687668, + 0.0047372449189424515, + 0.004389458801597357, + 0.005934113170951605, + 0.0023135431110858917, + 0.02748899534344673, + 0.04524710029363632, + 0.07138923555612564, + 0.07626856118440628, + 0.044423043727874756, + 0.039870284497737885, + 0.12322275340557098, + 0.05842617526650429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.02636837214231491 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086530983448029 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086530983448029, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086530983448029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053829554468393326, + 0.042082011699676514, + 0.10255543887615204, + 0.3191860020160675, + 0.08948273956775665, + 0.08275388926267624, + 0.18362165987491608, + 0.03400995209813118, + 0.04287078604102135, + 0.04960796609520912 + ], + "credit_map": [ + 0.0063258870504796505, + 0.01265683677047491, + 0.0036071997601538897, + 0.007447164971381426, + 0.006922672502696514, + 0.006896121893078089, + 0.006595592480152845, + 0.003378077410161495, + 0.005069216247648001, + 0.007122304290533066, + 0.004469007719308138, + 0.005571942310780287, + 0.005268567707389593, + 0.005711334757506847, + 0.004590978845953941, + 0.004278662148863077, + 0.007974378764629364, + 0.016593635082244873, + 0.0076995971612632275, + 0.025283174589276314, + 0.015297167003154755, + 0.008277765475213528, + 0.008581026457250118, + 0.012848695740103722, + 0.017955251038074493, + 0.01757984794676304, + 0.10023023188114166, + 0.05147836357355118, + 0.07582207769155502, + 0.020657066255807877, + 0.018663497641682625, + 0.016799692064523697, + 0.008362981490790844, + 0.01224882248789072, + 0.010418672114610672, + 0.020464526489377022, + 0.012458392418920994, + 0.007215084508061409, + 0.010889556258916855, + 0.007424704730510712, + 0.006121429614722729, + 0.008444907143712044, + 0.014965904876589775, + 0.012347265146672726, + 0.006399210076779127, + 0.010623057372868061, + 0.016318626701831818, + 0.007533493917435408, + 0.013877090997993946, + 0.024880511686205864, + 0.029347438365221024, + 0.05235449597239494, + 0.011462886817753315, + 0.015490860678255558, + 0.02065526694059372, + 0.015553111210465431, + 0.003423444926738739, + 0.004271009936928749, + 0.003173206700012088, + 0.004747362807393074, + 0.0038617609534412622, + 0.0046864720061421394, + 0.0042460281401872635, + 0.005600668489933014, + 0.0028113126754760742, + 0.007543568499386311, + 0.004165110178291798, + 0.006797214969992638, + 0.004028184339404106, + 0.007369174621999264, + 0.005707868840545416, + 0.004448348190635443, + 0.0033323531970381737, + 0.0069951703771948814, + 0.0057684797793626785, + 0.00641690194606781, + 0.003999253734946251, + 0.0051275077275931835, + 0.011865322478115559, + 0.006102980114519596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.011086530983448029 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129596084356308 + ], + "opposite_logprob_tokens": [ + -2.237959623336792 + ], + "logprob_reward": -0.1129596084356308, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129596084356308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04319323971867561, + 0.026920799165964127, + 0.037451062351465225, + 0.3515579104423523, + 0.0421009287238121, + 0.057787731289863586, + 0.20458859205245972, + 0.02063298411667347, + 0.022753620520234108, + 0.19301311671733856 + ], + "credit_map": [ + 0.005298747681081295, + 0.011047063395380974, + 0.0030394275672733784, + 0.005138932261615992, + 0.0053154462948441505, + 0.005206348840147257, + 0.0054084076546132565, + 0.0027388676535338163, + 0.0030094441026449203, + 0.0038370664697140455, + 0.002681510057300329, + 0.003683144925162196, + 0.004425881430506706, + 0.0033498948905617, + 0.0033803663682192564, + 0.002553490921854973, + 0.002910585142672062, + 0.006033935118466616, + 0.0032090081367641687, + 0.009478720836341381, + 0.00514674698933959, + 0.0036101138684898615, + 0.002981762634590268, + 0.004080187529325485, + 0.020726727321743965, + 0.018120015040040016, + 0.11159282922744751, + 0.0542726032435894, + 0.09061878174543381, + 0.02173682302236557, + 0.01744801178574562, + 0.01704213209450245, + 0.0043551987037062645, + 0.005377346184104681, + 0.005485272500663996, + 0.0056590670719742775, + 0.009308484382927418, + 0.0036175046116113663, + 0.004009343218058348, + 0.004288712050765753, + 0.0033253526780754328, + 0.007052931934595108, + 0.009401715360581875, + 0.007703373208642006, + 0.005386063829064369, + 0.007890734821557999, + 0.010815275833010674, + 0.006212279200553894, + 0.015621554106473923, + 0.029097547754645348, + 0.03197907283902168, + 0.056356046348810196, + 0.013018002733588219, + 0.017813874408602715, + 0.023364903405308723, + 0.017337583005428314, + 0.0013535169418901205, + 0.0021248029079288244, + 0.002038415288552642, + 0.0031667957082390785, + 0.0017485845601186156, + 0.0035536489449441433, + 0.00309244473464787, + 0.0035547742154449224, + 0.0019472012063488364, + 0.006712624803185463, + 0.0030593019910156727, + 0.0022049599792808294, + 0.0024925479665398598, + 0.00270533817820251, + 0.0019237840315327048, + 0.0017078622477129102, + 0.011015619151294231, + 0.01811790093779564, + 0.030084729194641113, + 0.029582982882857323, + 0.017507534474134445, + 0.015775414183735847, + 0.047216590493917465, + 0.023712344467639923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.1129596084356308 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a surreal macro-style with a high-frame-rate camera, this static, low-angle, extreme close-up captures a moment of immense power in breathtaking slow motion. Within the industrial metal confines of a capital ship's hangar bay, the heavy magnetic launch clamps securing an Eta-2 Jedi starfighter mechanically snap open. This decisive action sends a fiery, brilliant burst of sparks across the frame, each particle a fleeting trace of fire. Simultaneously, the fighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma, filling the scene with intense, high-contrast light. This harsh lighting gleams across the starfighter's fuselage, its sharp focus revealing every intricate, weathered detail: a tapestry of burn marks, microscopic scratches in the paint, and subtle indentations in the metal plating. The powerful glow carves deep shadows into the surrounding launch bay floor, creating a dramatic and tense atmosphere of unleashed energy.", + "seed": 23113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_11/index_4959_seed_23113_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18784182090312243, + "vlm_logprob": -0.18784182090312243, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.18784182090312243, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7853119969367981 + ], + "opposite_logprob_tokens": [ + -0.6603119969367981 + ], + "logprob_reward": -0.7853119969367981, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7853119969367981, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08331041038036346, + 0.022967327386140823, + 0.02471565268933773, + 0.11949174106121063, + 0.16716767847537994, + 0.019871791824698448, + 0.08522162586450577, + 0.10073166340589523, + 0.29436707496643066, + 0.0821550115942955 + ], + "credit_map": [ + 0.009279570542275906, + 0.01366226002573967, + 0.015126507729291916, + 0.008462292142212391, + 0.011340242810547352, + 0.008038715459406376, + 0.011873762123286724, + 0.005527065135538578, + 0.001926571479998529, + 0.0018709965515881777, + 0.0024038306437432766, + 0.0032847137190401554, + 0.00409583980217576, + 0.0033976559061557055, + 0.003821159712970257, + 0.0021665608510375023, + 0.0025798415299504995, + 0.002112275455147028, + 0.003970961086452007, + 0.0029458438511937857, + 0.00431249663233757, + 0.003574859583750367, + 0.0028284420259296894, + 0.002390933455899358, + 0.01067338977009058, + 0.024152131751179695, + 0.017678910866379738, + 0.008792736567556858, + 0.01772475801408291, + 0.02113448455929756, + 0.0064239478670060635, + 0.012911384925246239, + 0.027507876977324486, + 0.0232455562800169, + 0.02340816892683506, + 0.02170603908598423, + 0.01908242516219616, + 0.017104703933000565, + 0.01888737641274929, + 0.01622554287314415, + 0.0024072190281003714, + 0.002379469573497772, + 0.0031728066969662905, + 0.0017677530413493514, + 0.003000832162797451, + 0.0031267174053937197, + 0.0019533017184585333, + 0.0020636923145502806, + 0.008218108676373959, + 0.02216271311044693, + 0.012566585093736649, + 0.008578598499298096, + 0.009077220223844051, + 0.008592473343014717, + 0.007596172392368317, + 0.00842975452542305, + 0.0077425641939044, + 0.011315817013382912, + 0.01251901127398014, + 0.013237975537776947, + 0.01344364695250988, + 0.010260719805955887, + 0.014891318045556545, + 0.017320625483989716, + 0.06630069017410278, + 0.02136603370308876, + 0.014582340605556965, + 0.024574177339673042, + 0.02743782289326191, + 0.098460853099823, + 0.017913121730089188, + 0.023732034489512444, + 0.009902099147439003, + 0.012813912704586983, + 0.006661387160420418, + 0.00802333652973175, + 0.014414101839065552, + 0.010878505185246468, + 0.008824569173157215, + 0.01063710916787386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.7853119969367981 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029833728447556496 + ], + "opposite_logprob_tokens": [ + -3.5298337936401367 + ], + "logprob_reward": -0.029833728447556496, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029833728447556496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19258588552474976, + 0.0644214078783989, + 0.11705822497606277, + 0.06965562701225281, + 0.1543765515089035, + 0.04101121798157692, + 0.08913704752922058, + 0.06600388884544373, + 0.09720589220523834, + 0.10854430496692657 + ], + "credit_map": [ + 0.014782161451876163, + 0.027194995433092117, + 0.03367079049348831, + 0.01816294714808464, + 0.02877822145819664, + 0.016016442328691483, + 0.029255671426653862, + 0.024724649265408516, + 0.011007064953446388, + 0.007212827913463116, + 0.007132487837225199, + 0.00686183525249362, + 0.009845582768321037, + 0.008658569306135178, + 0.006974095944315195, + 0.006728941574692726, + 0.012739474885165691, + 0.012939251028001308, + 0.01149181928485632, + 0.011882120743393898, + 0.02391759492456913, + 0.016980459913611412, + 0.013520653359591961, + 0.013586844317615032, + 0.004946492146700621, + 0.018327871337532997, + 0.004910110495984554, + 0.005529950838536024, + 0.008806481026113033, + 0.01639440283179283, + 0.003844386897981167, + 0.006895920727401972, + 0.02329140156507492, + 0.019981924444437027, + 0.022477028891444206, + 0.02132967673242092, + 0.01875213347375393, + 0.013259810395538807, + 0.02098030038177967, + 0.014304274693131447, + 0.003962646704167128, + 0.005469292402267456, + 0.008859249763190746, + 0.004069745074957609, + 0.007148051634430885, + 0.0047407387755811214, + 0.0034574430901557207, + 0.0033040472771972418, + 0.014744643121957779, + 0.008317510597407818, + 0.009824207983911037, + 0.008635844103991985, + 0.011587759479880333, + 0.020456314086914062, + 0.00820021703839302, + 0.007370543200522661, + 0.006928775459527969, + 0.013743234798312187, + 0.005532086826860905, + 0.009828616864979267, + 0.0049581569619476795, + 0.008187763392925262, + 0.010472617112100124, + 0.006352630443871021, + 0.013799340464174747, + 0.013453646562993526, + 0.006753487046808004, + 0.008128478191792965, + 0.012657055631279945, + 0.023481538519263268, + 0.00905571412295103, + 0.009876624681055546, + 0.010918408632278442, + 0.016331374645233154, + 0.008282586000859737, + 0.010677175596356392, + 0.021368741989135742, + 0.013660541735589504, + 0.013914814218878746, + 0.01339065469801426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.029833728447556496 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370462030172348 + ], + "opposite_logprob_tokens": [ + -3.6513705253601074 + ], + "logprob_reward": -0.026370462030172348, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370462030172348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28600892424583435, + 0.05798486992716789, + 0.08449947088956833, + 0.08089829981327057, + 0.13389535248279572, + 0.07801425457000732, + 0.06879553943872452, + 0.05258050933480263, + 0.0850483849644661, + 0.07227448374032974 + ], + "credit_map": [ + 0.02668765000998974, + 0.03561139851808548, + 0.04865391552448273, + 0.03150991350412369, + 0.03735797107219696, + 0.021026773378252983, + 0.056449469178915024, + 0.028711803257465363, + 0.007911580614745617, + 0.006974193267524242, + 0.004836779087781906, + 0.006269528064876795, + 0.009051312692463398, + 0.00745624303817749, + 0.007866915315389633, + 0.007618305739015341, + 0.008011754602193832, + 0.007382052950561047, + 0.008738002739846706, + 0.016091417521238327, + 0.013840840198099613, + 0.01122482493519783, + 0.011540914885699749, + 0.007669647689908743, + 0.006694943644106388, + 0.021352306008338928, + 0.0053019835613667965, + 0.007173079065978527, + 0.013133377768099308, + 0.015816427767276764, + 0.00416164705529809, + 0.007264525629580021, + 0.013295859098434448, + 0.019967159256339073, + 0.021372251212596893, + 0.01871977560222149, + 0.01789267174899578, + 0.012170094065368176, + 0.018126249313354492, + 0.012351279146969318, + 0.009328748099505901, + 0.012280251830816269, + 0.010568137280642986, + 0.009664071723818779, + 0.013153444975614548, + 0.008529731072485447, + 0.006530141923576593, + 0.007959707640111446, + 0.009179428219795227, + 0.0081432880833745, + 0.007123460527509451, + 0.009311539120972157, + 0.00872763805091381, + 0.01244888361543417, + 0.007798000704497099, + 0.0060632843524217606, + 0.006820415612310171, + 0.005534783937036991, + 0.004921297542750835, + 0.007234062068164349, + 0.004814688116312027, + 0.007898915559053421, + 0.006117960438132286, + 0.009238377213478088, + 0.018459506332874298, + 0.005753817968070507, + 0.005004437640309334, + 0.008367421105504036, + 0.007614349480718374, + 0.025010328739881516, + 0.007726043462753296, + 0.007112469058483839, + 0.00604807585477829, + 0.011080892756581306, + 0.0052072289399802685, + 0.008017195388674736, + 0.01530177891254425, + 0.00830526277422905, + 0.009635118767619133, + 0.008678916841745377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.026370462030172348 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645604364573956 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645604364573956, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645604364573956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12229949235916138, + 0.03807039186358452, + 0.03866540640592575, + 0.04361709952354431, + 0.2689119279384613, + 0.12616755068302155, + 0.05401468649506569, + 0.06853070855140686, + 0.20704792439937592, + 0.03267486020922661 + ], + "credit_map": [ + 0.011383503675460815, + 0.01742689311504364, + 0.019001226872205734, + 0.01354425959289074, + 0.016324836760759354, + 0.009409530088305473, + 0.022081132978200912, + 0.013128111138939857, + 0.004703654441982508, + 0.003911418374627829, + 0.003020171308889985, + 0.0072297160513699055, + 0.005805930122733116, + 0.0046661170199513435, + 0.004888833500444889, + 0.003844549646601081, + 0.0034056641161441803, + 0.003442204324528575, + 0.004585423972457647, + 0.006600768771022558, + 0.005783057305961847, + 0.004844181705266237, + 0.005449200980365276, + 0.0045549082569777966, + 0.004106912296265364, + 0.004763525445014238, + 0.009099051356315613, + 0.004789141472429037, + 0.0064996457658708096, + 0.006084291730076075, + 0.003378150053322315, + 0.0048963818699121475, + 0.03337448090314865, + 0.03326454386115074, + 0.05393313616514206, + 0.03874025493860245, + 0.03296637535095215, + 0.018327249214053154, + 0.03581535443663597, + 0.02249053306877613, + 0.012685065157711506, + 0.019245820119976997, + 0.016593769192695618, + 0.016389012336730957, + 0.02227565087378025, + 0.013004623353481293, + 0.00984957441687584, + 0.016124026849865913, + 0.005682822782546282, + 0.014100591652095318, + 0.005748190451413393, + 0.007914313115179539, + 0.005268124397844076, + 0.005462385714054108, + 0.004685172811150551, + 0.005153085105121136, + 0.004864866379648447, + 0.010470956563949585, + 0.008245707489550114, + 0.008742503821849823, + 0.00698815705254674, + 0.006802880205214024, + 0.007812726311385632, + 0.014602908864617348, + 0.047364261001348495, + 0.015031343325972557, + 0.010242045857012272, + 0.018491245806217194, + 0.019045166671276093, + 0.0673513263463974, + 0.01342182420194149, + 0.016100702807307243, + 0.004228869918733835, + 0.004513181746006012, + 0.00362377823330462, + 0.0030791512690484524, + 0.0037782746367156506, + 0.005270787980407476, + 0.0030623029451817274, + 0.005118511617183685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.008645604364573956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904731273651123 + ], + "opposite_logprob_tokens": [ + -2.464047431945801 + ], + "logprob_reward": -0.08904731273651123, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904731273651123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5132665038108826, + 0.04004683718085289, + 0.03454237058758736, + 0.06719159334897995, + 0.042651399970054626, + 0.1256851702928543, + 0.037818823009729385, + 0.04722794145345688, + 0.06113392859697342, + 0.030435409396886826 + ], + "credit_map": [ + 0.02793840691447258, + 0.037135928869247437, + 0.05548155680298805, + 0.08088813722133636, + 0.05390962213277817, + 0.03616106137633324, + 0.1647278368473053, + 0.05702399089932442, + 0.0035694758407771587, + 0.0036051629576832056, + 0.0037313431967049837, + 0.005702966824173927, + 0.007741780951619148, + 0.006202284712344408, + 0.005640662740916014, + 0.003853157628327608, + 0.003726694267243147, + 0.004222151357680559, + 0.004350368399173021, + 0.005535021889954805, + 0.0054987408220767975, + 0.004157769028097391, + 0.004053578712046146, + 0.002998047275468707, + 0.007325014099478722, + 0.010080275125801563, + 0.00933223869651556, + 0.007064326200634241, + 0.01002993993461132, + 0.010762445628643036, + 0.00619683088734746, + 0.006400523241609335, + 0.005338969640433788, + 0.007096890825778246, + 0.00516814598813653, + 0.005015842150896788, + 0.005974705796688795, + 0.003825390012934804, + 0.005087831057608128, + 0.00514362845569849, + 0.016284052282571793, + 0.02135850489139557, + 0.013276644051074982, + 0.011093302629888058, + 0.024811377748847008, + 0.019835567101836205, + 0.008097653277218342, + 0.010928080417215824, + 0.00513046607375145, + 0.004970737732946873, + 0.003203807631507516, + 0.005079352296888828, + 0.005021845921874046, + 0.004715705290436745, + 0.004447505809366703, + 0.005249402951449156, + 0.004405506420880556, + 0.007916123606264591, + 0.005025888793170452, + 0.005827066954225302, + 0.00409199483692646, + 0.00568049494177103, + 0.006892128381878138, + 0.007388743571937084, + 0.012343958020210266, + 0.006404506508260965, + 0.003430926939472556, + 0.0059137726202607155, + 0.005484772380441427, + 0.017491713166236877, + 0.004827850963920355, + 0.005236427299678326, + 0.004215719178318977, + 0.0040874783881008625, + 0.002970860805362463, + 0.0029945799615234137, + 0.004173646215349436, + 0.0030378259252756834, + 0.004800791852176189, + 0.004154508002102375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.08904731273651123 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a surreal macro-style with a high-frame-rate camera, this static, low-angle, extreme close-up captures a moment of immense power in breathtaking slow motion. Within the industrial metal confines of a capital ship's hangar bay, the heavy magnetic launch clamps securing an Eta-2 Jedi starfighter mechanically snap open. This decisive action sends a fiery, brilliant burst of sparks across the frame, each particle a fleeting trace of fire. Simultaneously, the fighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma, filling the scene with intense, high-contrast light. This harsh lighting gleams across the starfighter's fuselage, its sharp focus revealing every intricate, weathered detail: a tapestry of burn marks, microscopic scratches in the paint, and subtle indentations in the metal plating. The powerful glow carves deep shadows into the surrounding launch bay floor, creating a dramatic and tense atmosphere of unleashed energy.", + "seed": 23114, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_11/index_4959_seed_23114_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2087405424565077, + "vlm_logprob": -0.2087405424565077, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.2087405424565077, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25263115763664246 + ], + "opposite_logprob_tokens": [ + -1.6276311874389648 + ], + "logprob_reward": -0.25263115763664246, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.25263115763664246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13725602626800537, + 0.03891200199723244, + 0.07376393675804138, + 0.03375626355409622, + 0.052957091480493546, + 0.058469127863645554, + 0.07089363038539886, + 0.5009698271751404, + 0.019086720421910286, + 0.013935355469584465 + ], + "credit_map": [ + 0.013557888567447662, + 0.02174745313823223, + 0.01484597660601139, + 0.015982748940587044, + 0.017043080180883408, + 0.018094433471560478, + 0.014912247657775879, + 0.021072201430797577, + 0.0026082072872668505, + 0.00290116504766047, + 0.003504958003759384, + 0.01110233273357153, + 0.003083452582359314, + 0.004986537620425224, + 0.005162177141755819, + 0.0055631715804338455, + 0.0031394115649163723, + 0.0036319550126791, + 0.003056566696614027, + 0.0032578071113675833, + 0.002761473413556814, + 0.0268503800034523, + 0.0031006678473204374, + 0.027965672314167023, + 0.003543023718520999, + 0.002985089085996151, + 0.004125293344259262, + 0.006696464493870735, + 0.0030976817943155766, + 0.006421546917408705, + 0.003597377799451351, + 0.003289788728579879, + 0.0029299065936356783, + 0.014537184499204159, + 0.006365948356688023, + 0.01621120050549507, + 0.0032768086530268192, + 0.0037440285086631775, + 0.003068391000851989, + 0.0028236249927431345, + 0.006541054230183363, + 0.008600917644798756, + 0.007718739099800587, + 0.009524197317659855, + 0.008038233034312725, + 0.008798407390713692, + 0.004408891312777996, + 0.00483868969604373, + 0.008879048749804497, + 0.007332053501158953, + 0.012342904694378376, + 0.012991013005375862, + 0.009372584521770477, + 0.008518357761204243, + 0.007029194850474596, + 0.004428474232554436, + 0.048649862408638, + 0.05956492945551872, + 0.07255862653255463, + 0.09385013580322266, + 0.060951996594667435, + 0.06870301812887192, + 0.06350305676460266, + 0.033188190311193466, + 0.0026467847637832165, + 0.004225294571369886, + 0.002085241721943021, + 0.0022744848392903805, + 0.001999078318476677, + 0.002617089543491602, + 0.0017700825119391084, + 0.0014686650829389691, + 0.0015154898865148425, + 0.002188807586207986, + 0.0018810743931680918, + 0.0011917625088244677, + 0.0025383455213159323, + 0.0015291281742975116, + 0.0018161748303100467, + 0.0012745719868689775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.25263115763664246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233906507492065 + ], + "opposite_logprob_tokens": [ + -0.8983906507492065 + ], + "logprob_reward": -0.5233906507492065, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233906507492065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19919049739837646, + 0.05534923076629639, + 0.07180797308683395, + 0.11209575086832047, + 0.10896358639001846, + 0.06713955104351044, + 0.07592836022377014, + 0.2280903458595276, + 0.028831494972109795, + 0.05260324478149414 + ], + "credit_map": [ + 0.0384959951043129, + 0.03308262676000595, + 0.020602237433195114, + 0.02268201671540737, + 0.018556173890829086, + 0.021277185529470444, + 0.022456947714090347, + 0.022037295624613762, + 0.003582627745345235, + 0.00526106171309948, + 0.0046823411248624325, + 0.01570974849164486, + 0.005445409100502729, + 0.007051120977848768, + 0.006755772978067398, + 0.006861147005110979, + 0.002415965311229229, + 0.0032703750766813755, + 0.0025085473898798227, + 0.00343790790066123, + 0.0033863228745758533, + 0.02536427229642868, + 0.004806110169738531, + 0.026618463918566704, + 0.005906228441745043, + 0.006427264306694269, + 0.0045806411653757095, + 0.03455965220928192, + 0.00471552275121212, + 0.03826591372489929, + 0.0081972386687994, + 0.009443270973861217, + 0.006300681736320257, + 0.029149271547794342, + 0.01016495656222105, + 0.03849487006664276, + 0.00575670599937439, + 0.008781694807112217, + 0.005989209748804569, + 0.004326192196458578, + 0.006613109260797501, + 0.00914689153432846, + 0.00908686313778162, + 0.009938550181686878, + 0.010970070958137512, + 0.011359039694070816, + 0.004683509003371, + 0.005341515876352787, + 0.008846710436046124, + 0.008636568672955036, + 0.013046539388597012, + 0.01712595485150814, + 0.009843842126429081, + 0.006890229880809784, + 0.007064902223646641, + 0.0044736070558428764, + 0.022200550884008408, + 0.027529504150152206, + 0.03160066157579422, + 0.04258311539888382, + 0.02808736450970173, + 0.03164207562804222, + 0.030062898993492126, + 0.014384162612259388, + 0.0025938660837709904, + 0.004787944257259369, + 0.003068008925765753, + 0.00284917582757771, + 0.005167584866285324, + 0.005177698098123074, + 0.003256905125454068, + 0.00193030945956707, + 0.005906767677515745, + 0.00831310823559761, + 0.005059059243649244, + 0.00431634159758687, + 0.010390901938080788, + 0.007413714658468962, + 0.007048320956528187, + 0.0041550262831151485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.5233906507492065 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16040517389774323 + ], + "opposite_logprob_tokens": [ + -1.910405158996582 + ], + "logprob_reward": -0.16040517389774323, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16040517389774323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14460961520671844, + 0.10144955664873123, + 0.09732501208782196, + 0.06923066824674606, + 0.10595341771841049, + 0.04198462516069412, + 0.05201031267642975, + 0.2921582758426666, + 0.03195713460445404, + 0.06332138180732727 + ], + "credit_map": [ + 0.02191028743982315, + 0.022750582545995712, + 0.014996192418038845, + 0.01818753406405449, + 0.013670242391526699, + 0.01685784012079239, + 0.018897179514169693, + 0.01733975112438202, + 0.007330561522394419, + 0.009164460003376007, + 0.008517538197338581, + 0.032121691852808, + 0.007743257563561201, + 0.013556896708905697, + 0.012164374813437462, + 0.010850781574845314, + 0.004079862497746944, + 0.004533300641924143, + 0.004198333714157343, + 0.006790779996663332, + 0.005507136695086956, + 0.03319227695465088, + 0.005694119725376368, + 0.033329207450151443, + 0.005290402099490166, + 0.004877847619354725, + 0.006808554287999868, + 0.01790764555335045, + 0.005489924922585487, + 0.020014382898807526, + 0.004714422859251499, + 0.004127489868551493, + 0.007771981880068779, + 0.023791803047060966, + 0.012748945504426956, + 0.0322086475789547, + 0.007468377240002155, + 0.008939182385802269, + 0.007388287223875523, + 0.005636198446154594, + 0.006339508108794689, + 0.006794293411076069, + 0.005816024262458086, + 0.005424160975962877, + 0.005922830663621426, + 0.005665674805641174, + 0.0033003923017531633, + 0.0027217420283704996, + 0.007080511189997196, + 0.005373270716518164, + 0.009702782146632671, + 0.008021661080420017, + 0.006543384864926338, + 0.0063275741413235664, + 0.0052010067738592625, + 0.003760118503123522, + 0.02845430187880993, + 0.03459155559539795, + 0.04313553124666214, + 0.0545080229640007, + 0.03571310266852379, + 0.039300672709941864, + 0.037262532860040665, + 0.01919255405664444, + 0.0029847705736756325, + 0.003928975202143192, + 0.004616525489836931, + 0.0040670521557331085, + 0.0046181040816009045, + 0.006390209775418043, + 0.0034380273427814245, + 0.0019134687026962638, + 0.006488523446023464, + 0.009310301393270493, + 0.009487626142799854, + 0.004775737877935171, + 0.011727133765816689, + 0.006926832254976034, + 0.009132951498031616, + 0.005472276825457811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.16040517389774323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018206385895609856 + ], + "opposite_logprob_tokens": [ + -4.018206596374512 + ], + "logprob_reward": -0.018206385895609856, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018206385895609856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12191138416528702, + 0.054539091885089874, + 0.09299133718013763, + 0.09525110572576523, + 0.09982884675264359, + 0.08811783790588379, + 0.07160109281539917, + 0.3213443160057068, + 0.0291712936013937, + 0.025243710726499557 + ], + "credit_map": [ + 0.02213827706873417, + 0.02003731206059456, + 0.011397677473723888, + 0.013547251001000404, + 0.013674793764948845, + 0.014026626013219357, + 0.013814777135848999, + 0.013274655677378178, + 0.004907050170004368, + 0.00468213576823473, + 0.006599158979952335, + 0.005897624883800745, + 0.006566470488905907, + 0.009248951449990273, + 0.006610268726944923, + 0.010027428157627583, + 0.0026689572259783745, + 0.0036763702519237995, + 0.0026671316009014845, + 0.003256817813962698, + 0.003544979263097048, + 0.03715889900922775, + 0.0029398221522569656, + 0.03707835450768471, + 0.004660426639020443, + 0.005082433111965656, + 0.0036076258402317762, + 0.03324015066027641, + 0.004148073960095644, + 0.03812761232256889, + 0.003245468018576503, + 0.0031393093522638083, + 0.005315407644957304, + 0.025345126166939735, + 0.012513268738985062, + 0.037030354142189026, + 0.004749604035168886, + 0.007191549986600876, + 0.004825396928936243, + 0.0028581302613019943, + 0.009880726225674152, + 0.01377739105373621, + 0.010702477768063545, + 0.014538425952196121, + 0.01207407470792532, + 0.014345133677124977, + 0.006024695001542568, + 0.0067749121226370335, + 0.007733919657766819, + 0.008248110301792622, + 0.010555596090853214, + 0.013221711851656437, + 0.010278720408678055, + 0.008785026147961617, + 0.006619459018111229, + 0.006158541887998581, + 0.03200077265501022, + 0.04003974422812462, + 0.04706275835633278, + 0.06094472482800484, + 0.036959629505872726, + 0.042903125286102295, + 0.03975490853190422, + 0.021678633987903595, + 0.002601411659270525, + 0.006506789941340685, + 0.002868737792596221, + 0.0024908510968089104, + 0.005774055607616901, + 0.0034555611200630665, + 0.0025348595809191465, + 0.0029390251729637384, + 0.0027831317856907845, + 0.002709929598495364, + 0.0027544943150132895, + 0.0021906502079218626, + 0.004693048074841499, + 0.003800837090238929, + 0.003888108069077134, + 0.0024235087912529707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.018206385895609856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, low-angle, extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the industrial metal confines of a capital ship's hangar bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do heavy magnetic launch clamps mechanically snap open to release an Eta-2 Jedi starfighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the starfighter's twin engines erupt with a dazzling burst of brilliant blue-white plasma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a fiery burst of sparks and reveal weathered details like burn marks and scratches on the starfighter's fuselage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906934410333633 + ], + "opposite_logprob_tokens": [ + -2.464069366455078 + ], + "logprob_reward": -0.08906934410333633, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906934410333633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07218469679355621, + 0.02550208568572998, + 0.023900369182229042, + 0.043427225202322006, + 0.07461784034967422, + 0.059987623244524, + 0.12268644571304321, + 0.5592238903045654, + 0.00945046916604042, + 0.00901930034160614 + ], + "credit_map": [ + 0.007577723823487759, + 0.011911734938621521, + 0.00747004384174943, + 0.009243800304830074, + 0.0081839170306921, + 0.008874533697962761, + 0.00770892109721899, + 0.01121402159333229, + 0.0017088633030653, + 0.002663401188328862, + 0.0031450113747268915, + 0.0041391924023628235, + 0.0036968355998396873, + 0.0033578488510102034, + 0.0033614751882851124, + 0.003429458010941744, + 0.0017427419079467654, + 0.001938489731401205, + 0.0015689274296164513, + 0.0023930922616273165, + 0.00189490569755435, + 0.006041808519512415, + 0.0026140608824789524, + 0.005706342402845621, + 0.003520697122439742, + 0.003571550827473402, + 0.003971864003688097, + 0.010123152285814285, + 0.0034085751976817846, + 0.012379204854369164, + 0.003295598551630974, + 0.0031565839890390635, + 0.004065703600645065, + 0.021487215533852577, + 0.007474107202142477, + 0.02537766844034195, + 0.003611585358157754, + 0.005442374385893345, + 0.004058036487549543, + 0.0031011479441076517, + 0.006597651168704033, + 0.008979201316833496, + 0.007191540207713842, + 0.009947489015758038, + 0.008584899827837944, + 0.009723854251205921, + 0.0041655609384179115, + 0.004797426052391529, + 0.015765191987156868, + 0.013505574315786362, + 0.020910268649458885, + 0.02609388343989849, + 0.014281168580055237, + 0.013473227620124817, + 0.011295555159449577, + 0.007361583411693573, + 0.053835153579711914, + 0.06532159447669983, + 0.08206171542406082, + 0.10450565814971924, + 0.06827022135257721, + 0.07654904574155807, + 0.07151254266500473, + 0.0371679924428463, + 0.0009207589318975806, + 0.0016203686827793717, + 0.0012072011595591903, + 0.0011778771877288818, + 0.001032551284879446, + 0.0012835171073675156, + 0.0016062712529674172, + 0.0006019233842380345, + 0.0010439464822411537, + 0.001120955334044993, + 0.0013479223707690835, + 0.0008641023887321353, + 0.001205855398438871, + 0.0009520314633846283, + 0.0014649934601038694, + 0.0010194930946454406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.08906934410333633 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_12/index_tensor([482, 482])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_12/index_tensor([482, 482])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2392acae46597ac611ee46e791fd5bb989105844 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_12/index_tensor([482, 482])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a slightly tilted perspective, the camera moves in a slow, arcing motion, circling a scene set against an infinitely stretching, pure white backdrop. The image captures the signature high-contrast black-and-white portrait quality of Richard Avedon. A small, animated game robot, with a cubic body and spindly limbs, struggles to propel a highly polished black bowling ball, its size dramatically contrasting, down an abstract lane. Sharp studio lighting, directed from above, accentuates the robot's matte black exterior with distinct edges and casts harsh, almost overexposed white highlights on the ball's smooth surface, while deep shadows separate them from the pristine white ground. The robot leans forward, its legs pushing against the ground, its mechanical joints seemingly straining with immense force. The camera rotates, focusing in close-up on a small screen on the robot's face. The screen, composed of simple black and white pixels, captures a resolute frown, each pixel conveying unyielding effort. As the robot's legs struggled to move on the ground, the huge bowling ball rolled slowly inch by inch. The entire scene, in extreme black and white contrast, created a unique atmosphere that blended absurdity, concentration and epic struggle.", + "seed": 8942, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_12/index_482_seed_8942_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17891402021050454, + "vlm_logprob": -0.17891402021050454, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.17891402021050454, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13356387615203857 + ], + "opposite_logprob_tokens": [ + -2.133563995361328 + ], + "logprob_reward": -0.13356387615203857, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13356387615203857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3334139287471771, + 0.06283323466777802, + 0.1894146054983139, + 0.09344442933797836, + 0.009838934056460857, + 0.03440311178565025, + 0.040061239153146744, + 0.06537728756666183, + 0.11060894280672073, + 0.060604315251111984 + ], + "credit_map": [ + 0.04033941403031349, + 0.029727773740887642, + 0.028331760317087173, + 0.02957184985280037, + 0.048515141010284424, + 0.05375870689749718, + 0.053737543523311615, + 0.04943171516060829, + 0.010571988299489021, + 0.007710258476436138, + 0.006840644869953394, + 0.008249765262007713, + 0.007065269164741039, + 0.008424034342169762, + 0.007352849934250116, + 0.006618421990424395, + 0.018828270956873894, + 0.014110702089965343, + 0.013599533587694168, + 0.014731450006365776, + 0.01506056822836399, + 0.04319600388407707, + 0.02375328354537487, + 0.046134788542985916, + 0.014557089656591415, + 0.00929197110235691, + 0.010475415736436844, + 0.011430774815380573, + 0.008654942736029625, + 0.013584277592599392, + 0.010968135669827461, + 0.014481822960078716, + 0.0012728298315778375, + 0.001843673991970718, + 0.0009344032150693238, + 0.0010984695982187986, + 0.001211571041494608, + 0.0013026137603446841, + 0.0012136412551626563, + 0.0009617321193218231, + 0.0039018287789076567, + 0.004273108672350645, + 0.002810878213495016, + 0.003468169365078211, + 0.005747117567807436, + 0.003708417760208249, + 0.005337861832231283, + 0.005155730061233044, + 0.005598797928541899, + 0.0073768263682723045, + 0.00451129162684083, + 0.005186807829886675, + 0.006325907539576292, + 0.003961492795497179, + 0.0030600964091718197, + 0.004040021914988756, + 0.008866216987371445, + 0.00873223040252924, + 0.010771959088742733, + 0.009465287439525127, + 0.007984282448887825, + 0.006870357785373926, + 0.006097531411796808, + 0.006589418742805719, + 0.011051803827285767, + 0.015440591610968113, + 0.012245601043105125, + 0.022412685677409172, + 0.010839118622243404, + 0.0075382995419204235, + 0.015456813387572765, + 0.015624028630554676, + 0.012185419909656048, + 0.00563371367752552, + 0.009474178776144981, + 0.007245437242090702, + 0.004721720237284899, + 0.004398373421281576, + 0.005305848550051451, + 0.011639628559350967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.13356387615203857 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5760993957519531 + ], + "opposite_logprob_tokens": [ + -0.8260993957519531 + ], + "logprob_reward": -0.5760993957519531, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5760993957519531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17110228538513184, + 0.019921673461794853, + 0.06585647165775299, + 0.026400865986943245, + 0.03272649645805359, + 0.036524321883916855, + 0.1648629754781723, + 0.36636030673980713, + 0.09143923968076706, + 0.024805352091789246 + ], + "credit_map": [ + 0.019822511821985245, + 0.015554112382233143, + 0.015164474956691265, + 0.01425706222653389, + 0.024145804345607758, + 0.028751499950885773, + 0.024738596752285957, + 0.028668232262134552, + 0.00401681661605835, + 0.002392388181760907, + 0.0019145564874634147, + 0.0022396049462258816, + 0.0019276104867458344, + 0.0029571494087576866, + 0.0019125923281535506, + 0.002560954773798585, + 0.006645945366472006, + 0.004682017490267754, + 0.005242625251412392, + 0.005033114459365606, + 0.004395276308059692, + 0.01550584938377142, + 0.007982165552675724, + 0.01636948250234127, + 0.0043627419508993626, + 0.00338170537725091, + 0.0019199317321181297, + 0.003049558959901333, + 0.003675019135698676, + 0.0037651467137038708, + 0.0021816135849803686, + 0.004065149929374456, + 0.004379319027066231, + 0.0032751974649727345, + 0.0030607148073613644, + 0.0035173255018889904, + 0.003971701487898827, + 0.004617576953023672, + 0.005103667266666889, + 0.004800998140126467, + 0.005442622117698193, + 0.005009878892451525, + 0.003024849109351635, + 0.002352602779865265, + 0.007569057401269674, + 0.0037617161870002747, + 0.005421647801995277, + 0.003941949922591448, + 0.023332128301262856, + 0.020575683563947678, + 0.026161639019846916, + 0.019608277827501297, + 0.022772323340177536, + 0.018645839765667915, + 0.010307149961590767, + 0.02345992811024189, + 0.0561002716422081, + 0.05774502828717232, + 0.051918674260377884, + 0.08025748282670975, + 0.040327154099941254, + 0.022996002808213234, + 0.026849737390875816, + 0.030165953561663628, + 0.009888609871268272, + 0.012336691841483116, + 0.008236569352447987, + 0.02548270672559738, + 0.01062894519418478, + 0.005170709919184446, + 0.008056196384131908, + 0.011638819240033627, + 0.004152493551373482, + 0.0023700930178165436, + 0.004370545037090778, + 0.005109901539981365, + 0.0021668446715921164, + 0.001933611580170691, + 0.0018354018684476614, + 0.0028664611745625734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.5760993957519531 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294214427471161 + ], + "opposite_logprob_tokens": [ + -2.2379422187805176 + ], + "logprob_reward": -0.11294214427471161, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294214427471161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12629978358745575, + 0.021998586133122444, + 0.041774872690439224, + 0.018697239458560944, + 0.027486661449074745, + 0.046873122453689575, + 0.17995116114616394, + 0.3566129505634308, + 0.1143520176410675, + 0.0659535750746727 + ], + "credit_map": [ + 0.014329735189676285, + 0.011291692964732647, + 0.011733111925423145, + 0.010740573517978191, + 0.018083855509757996, + 0.021268853917717934, + 0.01775047369301319, + 0.02110149711370468, + 0.004502504598349333, + 0.0021857519168406725, + 0.002290442120283842, + 0.002918208483606577, + 0.0022343348246067762, + 0.003147717798128724, + 0.002233525039628148, + 0.002486100420355797, + 0.004310779273509979, + 0.0035669170320034027, + 0.003998695872724056, + 0.0037284852005541325, + 0.003432072466239333, + 0.00889571662992239, + 0.0045971921645104885, + 0.009245013818144798, + 0.0032997666858136654, + 0.003132226876914501, + 0.001579838339239359, + 0.0014518261887133121, + 0.002383616752922535, + 0.002562673296779394, + 0.001834633992984891, + 0.0024526582565158606, + 0.003581000491976738, + 0.002982201986014843, + 0.0025757828261703253, + 0.003147113136947155, + 0.0034826535265892744, + 0.003916936926543713, + 0.0038835094310343266, + 0.003917464055120945, + 0.006287742406129837, + 0.006071154493838549, + 0.002637533936649561, + 0.003934618551284075, + 0.010901113972067833, + 0.005140676163136959, + 0.00597057119011879, + 0.005929712671786547, + 0.02675590291619301, + 0.01984364539384842, + 0.02919837459921837, + 0.02034646086394787, + 0.024088233709335327, + 0.022236693650484085, + 0.011991853825747967, + 0.025490012019872665, + 0.05414414033293724, + 0.055035579949617386, + 0.04907401278614998, + 0.07172271609306335, + 0.04130401089787483, + 0.027427228167653084, + 0.027415165677666664, + 0.030490057542920113, + 0.013093536719679832, + 0.016489755362272263, + 0.012370346114039421, + 0.029005393385887146, + 0.013494588434696198, + 0.006902442313730717, + 0.011692317202687263, + 0.011303641833364964, + 0.00967870932072401, + 0.00643918476998806, + 0.00966794602572918, + 0.01413875911384821, + 0.007229141891002655, + 0.004658170510083437, + 0.006404918618500233, + 0.007736748084425926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.11294214427471161 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023298438638448715 + ], + "opposite_logprob_tokens": [ + -3.773298501968384 + ], + "logprob_reward": -0.023298438638448715, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023298438638448715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10027855634689331, + 0.059420276433229446, + 0.04095180705189705, + 0.07519087940454483, + 0.03648817539215088, + 0.09606818854808807, + 0.1272096037864685, + 0.14300601184368134, + 0.26981115341186523, + 0.05157539248466492 + ], + "credit_map": [ + 0.00968222226947546, + 0.004604162648320198, + 0.004183441400527954, + 0.006038821768015623, + 0.007518538739532232, + 0.025088761001825333, + 0.01152977254241705, + 0.03163282945752144, + 0.009975392371416092, + 0.006051157135516405, + 0.004029714502394199, + 0.006741025485098362, + 0.007498107850551605, + 0.00928550586104393, + 0.0071513899601995945, + 0.00868797767907381, + 0.004921905230730772, + 0.0039616855792701244, + 0.004563647788017988, + 0.0037819764111191034, + 0.0038594596553593874, + 0.00712970457971096, + 0.004918372258543968, + 0.007815057411789894, + 0.010960500687360764, + 0.010080758482217789, + 0.007908270694315434, + 0.009753439575433731, + 0.0071020242758095264, + 0.010307801887392998, + 0.009903093799948692, + 0.009174986742436886, + 0.003954493906348944, + 0.004923334810882807, + 0.00385161186568439, + 0.005001166835427284, + 0.004384508356451988, + 0.005079744383692741, + 0.004398939665406942, + 0.004894376732409, + 0.011667248792946339, + 0.01241199392825365, + 0.007782259490340948, + 0.008049764670431614, + 0.016054978594183922, + 0.011660551652312279, + 0.014325867407023907, + 0.01411551982164383, + 0.018628258258104324, + 0.022585604339838028, + 0.016441605985164642, + 0.012627286836504936, + 0.020604582503437996, + 0.012573027983307838, + 0.009549734182655811, + 0.014199486933648586, + 0.01885763742029667, + 0.02018757350742817, + 0.02323330193758011, + 0.024609673768281937, + 0.01584513485431671, + 0.01380979921668768, + 0.012986721470952034, + 0.013476146385073662, + 0.02897370606660843, + 0.03850407153367996, + 0.03148023039102554, + 0.04982086271047592, + 0.027106331661343575, + 0.019715409725904465, + 0.036900632083415985, + 0.03730989247560501, + 0.008834835141897202, + 0.006103360094130039, + 0.010596444830298424, + 0.0071517592296004295, + 0.005514579825103283, + 0.003788960864767432, + 0.004259852226823568, + 0.005325600039213896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.023298438638448715 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048666246235370636 + ], + "opposite_logprob_tokens": [ + -3.04866623878479 + ], + "logprob_reward": -0.048666246235370636, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048666246235370636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23270387947559357, + 0.022565050050616264, + 0.04821312800049782, + 0.031337231397628784, + 0.059541746973991394, + 0.03925945237278938, + 0.09026122093200684, + 0.2029159665107727, + 0.22710253298282623, + 0.04609985649585724 + ], + "credit_map": [ + 0.024064164608716965, + 0.018647218123078346, + 0.020992377772927284, + 0.01900637149810791, + 0.03232589364051819, + 0.04229546710848808, + 0.0298820361495018, + 0.04549034312367439, + 0.004249780438840389, + 0.002449420280754566, + 0.0016348239732906222, + 0.0026077169459313154, + 0.0022208981681615114, + 0.0032747609075158834, + 0.0028133566956967115, + 0.003314289962872863, + 0.006885657552629709, + 0.006435004062950611, + 0.005223600193858147, + 0.005072125233709812, + 0.004675497300922871, + 0.008095001801848412, + 0.003447617869824171, + 0.008378620259463787, + 0.004596429876983166, + 0.004797869827598333, + 0.002553025959059596, + 0.004224572330713272, + 0.0028926078230142593, + 0.0044676512479782104, + 0.0032105809077620506, + 0.004594492260366678, + 0.006465212907642126, + 0.006537267472594976, + 0.00778572540730238, + 0.011430855840444565, + 0.004614101257175207, + 0.007559450808912516, + 0.006570976693183184, + 0.008578152395784855, + 0.005842195823788643, + 0.003980016335844994, + 0.0028819767758250237, + 0.0029088121373206377, + 0.009924924932420254, + 0.003802678780630231, + 0.0047058528289198875, + 0.005212991498410702, + 0.016181642189621925, + 0.01136049721390009, + 0.013048293069005013, + 0.008077769540250301, + 0.009092149324715137, + 0.013492437079548836, + 0.005047994665801525, + 0.013960431329905987, + 0.02889167331159115, + 0.02831069566309452, + 0.024732578545808792, + 0.05388398468494415, + 0.022003697231411934, + 0.013606104999780655, + 0.014917662367224693, + 0.01656956598162651, + 0.028999466449022293, + 0.03495289012789726, + 0.02525867149233818, + 0.04103116691112518, + 0.025649067014455795, + 0.02436799183487892, + 0.021557800471782684, + 0.025285471230745316, + 0.010464855469763279, + 0.004180119838565588, + 0.007361605763435364, + 0.004397348500788212, + 0.004142341669648886, + 0.003790925955399871, + 0.005198596976697445, + 0.006564058363437653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.048666246235370636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a slightly tilted perspective, the camera moves in a slow, arcing motion, circling a scene set against an infinitely stretching, pure white backdrop. The image captures the signature high-contrast black-and-white portrait quality of Richard Avedon. A small, animated game robot, with a cubic body and spindly limbs, struggles to propel a highly polished black bowling ball, its size dramatically contrasting, down an abstract lane. Sharp studio lighting, directed from above, accentuates the robot's matte black exterior with distinct edges and casts harsh, almost overexposed white highlights on the ball's smooth surface, while deep shadows separate them from the pristine white ground. The robot leans forward, its legs pushing against the ground, its mechanical joints seemingly straining with immense force. The camera rotates, focusing in close-up on a small screen on the robot's face. The screen, composed of simple black and white pixels, captures a resolute frown, each pixel conveying unyielding effort. As the robot's legs struggled to move on the ground, the huge bowling ball rolled slowly inch by inch. The entire scene, in extreme black and white contrast, created a unique atmosphere that blended absurdity, concentration and epic struggle.", + "seed": 8943, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_12/index_482_seed_8943_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026695753494277597, + "vlm_logprob": -0.026695753494277597, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.026695753494277597, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0334593690931797 + ], + "opposite_logprob_tokens": [ + -3.533459424972534 + ], + "logprob_reward": -0.0334593690931797, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0334593690931797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21141554415225983, + 0.06928052008152008, + 0.11183103173971176, + 0.09297342598438263, + 0.040798962116241455, + 0.1467762291431427, + 0.1462087780237198, + 0.09222755581140518, + 0.07166480273008347, + 0.01682312600314617 + ], + "credit_map": [ + 0.03328199312090874, + 0.030717765912413597, + 0.024438094347715378, + 0.02717156894505024, + 0.021524034440517426, + 0.02406926453113556, + 0.023640966042876244, + 0.02657187543809414, + 0.009118442423641682, + 0.010078505612909794, + 0.009410269558429718, + 0.008527952246367931, + 0.011087595485150814, + 0.0045613632537424564, + 0.011439468711614609, + 0.005056921858340502, + 0.021996894851326942, + 0.017493749037384987, + 0.010933936573565006, + 0.016214128583669662, + 0.01656157337129116, + 0.007901525124907494, + 0.01078591775149107, + 0.00994331669062376, + 0.013937830924987793, + 0.016418421640992165, + 0.00887211225926876, + 0.025261227041482925, + 0.007270789705216885, + 0.004522137343883514, + 0.009239460341632366, + 0.007451451383531094, + 0.004861077293753624, + 0.005060212220996618, + 0.0052482555620372295, + 0.005444123409688473, + 0.003812385257333517, + 0.004037059377878904, + 0.006972916889935732, + 0.005362929776310921, + 0.017423901706933975, + 0.012282132171094418, + 0.015706866979599, + 0.01763501763343811, + 0.012252911925315857, + 0.0239638052880764, + 0.017977874726057053, + 0.029533714056015015, + 0.011645742692053318, + 0.011008254252374172, + 0.01999041996896267, + 0.019546935334801674, + 0.017176145687699318, + 0.021483464166522026, + 0.02164578251540661, + 0.02371203526854515, + 0.007698666304349899, + 0.008770771324634552, + 0.009754744358360767, + 0.010507801547646523, + 0.01176618691533804, + 0.012164014391601086, + 0.019440287724137306, + 0.012125089764595032, + 0.011172049678862095, + 0.010098709724843502, + 0.008603955619037151, + 0.007218693848699331, + 0.00872388668358326, + 0.008337977342307568, + 0.006487978622317314, + 0.011021559126675129, + 0.0022854451090097427, + 0.0017270236276090145, + 0.0014265634818002582, + 0.0021398160606622696, + 0.0015558156883344054, + 0.0022367737255990505, + 0.002091397764161229, + 0.003360291011631489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.0334593690931797 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022128161508589983 + ], + "opposite_logprob_tokens": [ + -6.127213001251221 + ], + "logprob_reward": -0.0022128161508589983, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022128161508589983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19028687477111816, + 0.06773104518651962, + 0.09218311309814453, + 0.02607676200568676, + 0.07407555729150772, + 0.059287458658218384, + 0.12005854398012161, + 0.06517723947763443, + 0.215974360704422, + 0.08914905041456223 + ], + "credit_map": [ + 0.031020168215036392, + 0.033119525760412216, + 0.024781346321105957, + 0.017847944051027298, + 0.014864775352180004, + 0.026257360354065895, + 0.02084323950111866, + 0.02155251055955887, + 0.010804946534335613, + 0.009285705164074898, + 0.00808032974600792, + 0.010622329078614712, + 0.009497019462287426, + 0.005170169286429882, + 0.009950296953320503, + 0.004320255946367979, + 0.014766305685043335, + 0.016878779977560043, + 0.008191485889256, + 0.011251071467995644, + 0.014614025130867958, + 0.007201381027698517, + 0.011461648158729076, + 0.007818415760993958, + 0.004275594372302294, + 0.005434251856058836, + 0.0025717029348015785, + 0.003737581428140402, + 0.0028083936776965857, + 0.0015996263828128576, + 0.0028916492592543364, + 0.0027579611632972956, + 0.010974263772368431, + 0.007036589551717043, + 0.008640363812446594, + 0.00686838710680604, + 0.0051902043633162975, + 0.008947346359491348, + 0.012560625560581684, + 0.01385777909308672, + 0.005988794844597578, + 0.007726722862571478, + 0.006568572949618101, + 0.0056212046183645725, + 0.005595799069851637, + 0.008605267852544785, + 0.010365631431341171, + 0.008815469220280647, + 0.008497695438563824, + 0.00908481702208519, + 0.011491159908473492, + 0.01233250554651022, + 0.01065516285598278, + 0.0211839210242033, + 0.017098037526011467, + 0.029715245589613914, + 0.007820799015462399, + 0.005120835732668638, + 0.007457053754478693, + 0.00659380154684186, + 0.007054958958178759, + 0.009397649206221104, + 0.011100775562226772, + 0.010631365701556206, + 0.023109063506126404, + 0.0167214535176754, + 0.01639425940811634, + 0.027213020250201225, + 0.017537808045744896, + 0.03406839445233345, + 0.03840525820851326, + 0.04252510145306587, + 0.020018955692648888, + 0.01045095082372427, + 0.009461226873099804, + 0.008458921685814857, + 0.008690902031958103, + 0.011241662316024303, + 0.008971120230853558, + 0.011855307035148144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0022128161508589983 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897159457206726 + ], + "opposite_logprob_tokens": [ + -2.5789716243743896 + ], + "logprob_reward": -0.07897159457206726, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897159457206726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19119934737682343, + 0.0788060799241066, + 0.052238959819078445, + 0.0344339981675148, + 0.268367201089859, + 0.0984991192817688, + 0.0527167022228241, + 0.04329105094075203, + 0.08809133619070053, + 0.09235614538192749 + ], + "credit_map": [ + 0.02422059141099453, + 0.04567126929759979, + 0.020517507568001747, + 0.02175237238407135, + 0.026013756170868874, + 0.020243363454937935, + 0.0156317837536335, + 0.01714870147407055, + 0.010336145758628845, + 0.011283351108431816, + 0.00856710970401764, + 0.010085539892315865, + 0.012488516978919506, + 0.006799597293138504, + 0.013415596447885036, + 0.005830223206430674, + 0.008586273528635502, + 0.010914601385593414, + 0.003953893668949604, + 0.006709118839353323, + 0.007563596125692129, + 0.003447159891948104, + 0.006636726204305887, + 0.004427589941769838, + 0.005686303600668907, + 0.005407489836215973, + 0.003285957733169198, + 0.005017525982111692, + 0.003851241897791624, + 0.0016954480670392513, + 0.0041839731857180595, + 0.0053060585632920265, + 0.02331773191690445, + 0.028523078188300133, + 0.06215498223900795, + 0.035796623677015305, + 0.02048836648464203, + 0.03036833181977272, + 0.031219566240906715, + 0.036498527973890305, + 0.013149941340088844, + 0.010340709239244461, + 0.013823975808918476, + 0.011403769254684448, + 0.009619870223104954, + 0.013955747708678246, + 0.010326283052563667, + 0.015878815203905106, + 0.006949935108423233, + 0.003918464295566082, + 0.006087734829634428, + 0.0048816390335559845, + 0.006013076286762953, + 0.008378407917916775, + 0.007722080685198307, + 0.008765369653701782, + 0.005967261269688606, + 0.003303305944427848, + 0.005284896120429039, + 0.005232124123722315, + 0.003618139075115323, + 0.0064148809760808945, + 0.005859152413904667, + 0.007611292880028486, + 0.012617154978215694, + 0.012125932611525059, + 0.01037727203220129, + 0.008766897954046726, + 0.009285719133913517, + 0.01004329789429903, + 0.011146657168865204, + 0.013728400692343712, + 0.021235965192317963, + 0.011560293845832348, + 0.009552433155477047, + 0.00892915204167366, + 0.009987629018723965, + 0.01068788394331932, + 0.006706220097839832, + 0.013696571812033653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.07897159457206726 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195887371897697 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195887371897697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195887371897697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17177443206310272, + 0.06977207958698273, + 0.0895681083202362, + 0.05471813306212425, + 0.09755240380764008, + 0.13818296790122986, + 0.2374136745929718, + 0.03719746321439743, + 0.06198926270008087, + 0.04183153808116913 + ], + "credit_map": [ + 0.029226668179035187, + 0.02330840937793255, + 0.020219767466187477, + 0.019757471978664398, + 0.022131068632006645, + 0.02041364274919033, + 0.017743021249771118, + 0.018974367529153824, + 0.009795003570616245, + 0.012696659192442894, + 0.008813517168164253, + 0.01205275859683752, + 0.008099102415144444, + 0.005784900858998299, + 0.007400386035442352, + 0.005129748024046421, + 0.016643065959215164, + 0.016435571014881134, + 0.010122661478817463, + 0.009481607005000114, + 0.011985449120402336, + 0.006904258392751217, + 0.009372308850288391, + 0.008623176254332066, + 0.007825703360140324, + 0.008446072228252888, + 0.005899448413401842, + 0.008585345931351185, + 0.007360418327152729, + 0.0035581195261329412, + 0.0072818538174033165, + 0.005761171691119671, + 0.010100288316607475, + 0.01059149019420147, + 0.01255502738058567, + 0.012215320952236652, + 0.006797397043555975, + 0.011954857967793941, + 0.01103231031447649, + 0.022305702790617943, + 0.017222512513399124, + 0.011827736161649227, + 0.01466738898307085, + 0.01572749763727188, + 0.010343480855226517, + 0.022223971784114838, + 0.01834147423505783, + 0.027828892692923546, + 0.020056167617440224, + 0.013830527663230896, + 0.032722026109695435, + 0.026621971279382706, + 0.024782247841358185, + 0.03774004429578781, + 0.034187231212854385, + 0.047473449259996414, + 0.002778255147859454, + 0.0034623469691723585, + 0.00394141161814332, + 0.004257844761013985, + 0.003766243113204837, + 0.005060313735157251, + 0.008496883325278759, + 0.005434161983430386, + 0.004867141600698233, + 0.0067354487255215645, + 0.004718909040093422, + 0.0060601606965065, + 0.007381013594567776, + 0.007704501505941153, + 0.012442068196833134, + 0.012080010958015919, + 0.00571396853774786, + 0.005600178614258766, + 0.007068505510687828, + 0.003976049832999706, + 0.003963223658502102, + 0.006190033629536629, + 0.004212947096675634, + 0.005106627941131592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.014195887371897697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639100283384323 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639100283384323, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639100283384323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0981476753950119, + 0.22006867825984955, + 0.16637487709522247, + 0.07085616886615753, + 0.13868632912635803, + 0.08633287250995636, + 0.04808266833424568, + 0.053269099444150925, + 0.06808241456747055, + 0.0500992015004158 + ], + "credit_map": [ + 0.011689345352351665, + 0.020992137491703033, + 0.010457601398229599, + 0.013166924938559532, + 0.012522551231086254, + 0.009786739014089108, + 0.008725428953766823, + 0.01080693956464529, + 0.03632320091128349, + 0.040128305554389954, + 0.019473891705274582, + 0.03503025323152542, + 0.04158662632107735, + 0.013572719879448414, + 0.02207806520164013, + 0.011875621043145657, + 0.02880287729203701, + 0.026663271710276604, + 0.018566176295280457, + 0.017903193831443787, + 0.028572969138622284, + 0.010275456123054028, + 0.01942337676882744, + 0.01616755872964859, + 0.010211721993982792, + 0.008255674503743649, + 0.006133191753178835, + 0.010910996235907078, + 0.013430776074528694, + 0.004199642222374678, + 0.011931434273719788, + 0.005782727617770433, + 0.012542525306344032, + 0.013666829094290733, + 0.032385047525167465, + 0.01864384301006794, + 0.011462423950433731, + 0.016208460554480553, + 0.01680806651711464, + 0.01696913316845894, + 0.012371521443128586, + 0.00768746342509985, + 0.005860951729118824, + 0.00802853237837553, + 0.007126308511942625, + 0.01290746871381998, + 0.016303876414895058, + 0.016046755015850067, + 0.0037508925888687372, + 0.0038865781389176846, + 0.005378432106226683, + 0.006174767855554819, + 0.005108844954520464, + 0.007409260608255863, + 0.008741735480725765, + 0.007632160559296608, + 0.004140011966228485, + 0.004018894862383604, + 0.00673332717269659, + 0.004915123339742422, + 0.006837228313088417, + 0.00748656177893281, + 0.0114402174949646, + 0.007697730790823698, + 0.012505987659096718, + 0.007456905674189329, + 0.007225204724818468, + 0.0057802628725767136, + 0.005935682915151119, + 0.009351629763841629, + 0.01052259560674429, + 0.009304147213697433, + 0.006738457828760147, + 0.004711242858320475, + 0.004459314048290253, + 0.004189764615148306, + 0.0053155566565692425, + 0.006947158370167017, + 0.008452709764242172, + 0.009284996427595615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.004639100283384323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a slightly tilted perspective, the camera moves in a slow, arcing motion, circling a scene set against an infinitely stretching, pure white backdrop. The image captures the signature high-contrast black-and-white portrait quality of Richard Avedon. A small, animated game robot, with a cubic body and spindly limbs, struggles to propel a highly polished black bowling ball, its size dramatically contrasting, down an abstract lane. Sharp studio lighting, directed from above, accentuates the robot's matte black exterior with distinct edges and casts harsh, almost overexposed white highlights on the ball's smooth surface, while deep shadows separate them from the pristine white ground. The robot leans forward, its legs pushing against the ground, its mechanical joints seemingly straining with immense force. The camera rotates, focusing in close-up on a small screen on the robot's face. The screen, composed of simple black and white pixels, captures a resolute frown, each pixel conveying unyielding effort. As the robot's legs struggled to move on the ground, the huge bowling ball rolled slowly inch by inch. The entire scene, in extreme black and white contrast, created a unique atmosphere that blended absurdity, concentration and epic struggle.", + "seed": 8944, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_12/index_482_seed_8944_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0454263910651207, + "vlm_logprob": -0.0454263910651207, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.0454263910651207, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03161389008164406 + ], + "opposite_logprob_tokens": [ + -3.656613826751709 + ], + "logprob_reward": -0.03161389008164406, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03161389008164406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2968149781227112, + 0.2829873859882355, + 0.049846578389406204, + 0.10488300025463104, + 0.07948987931013107, + 0.0219308752566576, + 0.06207916513085365, + 0.026365606114268303, + 0.037408988922834396, + 0.03819355368614197 + ], + "credit_map": [ + 0.02252350002527237, + 0.0274294875562191, + 0.035857122391462326, + 0.0281815305352211, + 0.04096706584095955, + 0.06741651147603989, + 0.02784232422709465, + 0.046597424894571304, + 0.01287585124373436, + 0.0067345681600272655, + 0.006582660600543022, + 0.006041375920176506, + 0.006577710621058941, + 0.12006854265928268, + 0.005337539128959179, + 0.11876913905143738, + 0.006416209973394871, + 0.0060711996629834175, + 0.005812874995172024, + 0.006161678582429886, + 0.005287962034344673, + 0.006740345153957605, + 0.006778482813388109, + 0.006577823311090469, + 0.01645067147910595, + 0.01604503206908703, + 0.01612907275557518, + 0.013477353379130363, + 0.007788960821926594, + 0.014611244201660156, + 0.007240036968141794, + 0.013140627183020115, + 0.006655083037912846, + 0.007183705922216177, + 0.0059050568379461765, + 0.006954510696232319, + 0.00534428795799613, + 0.0173355583101511, + 0.0060633448883891106, + 0.024048330262303352, + 0.003568253479897976, + 0.001940964488312602, + 0.0027798747178167105, + 0.0021807104349136353, + 0.002176375826820731, + 0.004358471836894751, + 0.002281677909195423, + 0.002644545864313841, + 0.021527793258428574, + 0.0030377765651792288, + 0.0039999173022806644, + 0.002441910095512867, + 0.0019663339480757713, + 0.003262109588831663, + 0.015309222973883152, + 0.010534103959798813, + 0.003453379264101386, + 0.00150547013618052, + 0.0028985789977014065, + 0.0020945884753018618, + 0.0017308293608948588, + 0.0024284753017127514, + 0.004903868772089481, + 0.007350415922701359, + 0.004157671704888344, + 0.0034669197630137205, + 0.005364449229091406, + 0.006025370210409164, + 0.0033200178295373917, + 0.002892072545364499, + 0.005542921833693981, + 0.006639564409852028, + 0.003051357576623559, + 0.0038873194716870785, + 0.0037662514951080084, + 0.004810640588402748, + 0.004297363106161356, + 0.003920155577361584, + 0.0080252755433321, + 0.006435188464820385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.03161389008164406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270439624786377 + ], + "opposite_logprob_tokens": [ + -2.1270439624786377 + ], + "logprob_reward": -0.1270439624786377, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270439624786377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06506887078285217, + 0.2897079586982727, + 0.04851275309920311, + 0.12024179846048355, + 0.24351678788661957, + 0.01554789487272501, + 0.0176025852560997, + 0.04454752802848816, + 0.05070189759135246, + 0.10455196350812912 + ], + "credit_map": [ + 0.009474013932049274, + 0.010831943713128567, + 0.0073699685744941235, + 0.009536216966807842, + 0.007230769377201796, + 0.007324715610593557, + 0.005754808429628611, + 0.007546429987996817, + 0.00999438762664795, + 0.005928517319262028, + 0.006399056408554316, + 0.005132276564836502, + 0.006009484641253948, + 0.12588734924793243, + 0.005965160671621561, + 0.12439171969890594, + 0.011818120256066322, + 0.009712936356663704, + 0.0048256488516926765, + 0.003654788713902235, + 0.00613386370241642, + 0.004341800697147846, + 0.0034350769128650427, + 0.004590511787682772, + 0.006123146042227745, + 0.005743206012994051, + 0.00457835104316473, + 0.025073716416954994, + 0.0034959041513502598, + 0.022118065506219864, + 0.004520615562796593, + 0.04858878254890442, + 0.024674678221344948, + 0.026853935793042183, + 0.01739429123699665, + 0.023714924231171608, + 0.02127566933631897, + 0.04985852912068367, + 0.01622595451772213, + 0.06351877748966217, + 0.002493135631084442, + 0.002092235255986452, + 0.0023492840118706226, + 0.001246814033947885, + 0.0015158865135163069, + 0.0024765306152403355, + 0.0017911088652908802, + 0.0015828979667276144, + 0.002397694392129779, + 0.0020215266849845648, + 0.001487205852754414, + 0.0013457250315696, + 0.0010051015997305512, + 0.0013973654713481665, + 0.004307310096919537, + 0.0036406537983566523, + 0.008019091561436653, + 0.005583552643656731, + 0.007018941920250654, + 0.0059743719175457954, + 0.00401151180267334, + 0.0033777167554944754, + 0.005141090601682663, + 0.005421245936304331, + 0.005282476544380188, + 0.005166077055037022, + 0.007120498921722174, + 0.009955980814993382, + 0.0033365990966558456, + 0.004260815214365721, + 0.006797404028475285, + 0.008782039396464825, + 0.006153085734695196, + 0.0060795689933001995, + 0.0077101620845496655, + 0.011324991472065449, + 0.014188706874847412, + 0.011264077387750149, + 0.027486782521009445, + 0.0203445702791214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.1270439624786377 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369717672467232 + ], + "opposite_logprob_tokens": [ + -3.4086971282958984 + ], + "logprob_reward": -0.03369717672467232, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369717672467232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11485899984836578, + 0.26581519842147827, + 0.05591609328985214, + 0.1672331541776657, + 0.12996557354927063, + 0.0394560731947422, + 0.019526218995451927, + 0.03647144138813019, + 0.10231634974479675, + 0.06844088435173035 + ], + "credit_map": [ + 0.012297295965254307, + 0.015337110497057438, + 0.013864202424883842, + 0.013500373810529709, + 0.01666480489075184, + 0.01549587957561016, + 0.011564522050321102, + 0.01613481156527996, + 0.012775466777384281, + 0.00902013573795557, + 0.012467768974602222, + 0.006979838479310274, + 0.013964598998427391, + 0.10105859488248825, + 0.0075132036581635475, + 0.10203558206558228, + 0.011915065348148346, + 0.009239626117050648, + 0.00569220632314682, + 0.004824316594749689, + 0.0073117478750646114, + 0.006384924985468388, + 0.0052124266512691975, + 0.005335785448551178, + 0.010868466459214687, + 0.011334705166518688, + 0.010096525773406029, + 0.03992566466331482, + 0.005662421230226755, + 0.019730357453227043, + 0.007223731372505426, + 0.06239129975438118, + 0.011835313402116299, + 0.014708403497934341, + 0.008904760703444481, + 0.010990331880748272, + 0.008421210572123528, + 0.026072096079587936, + 0.0102951405569911, + 0.03873831778764725, + 0.005950520280748606, + 0.0049651144072413445, + 0.004972423892468214, + 0.0033765181433409452, + 0.004011867567896843, + 0.00783100351691246, + 0.0042878128588199615, + 0.004060813225805759, + 0.0029033080209046602, + 0.0016952294390648603, + 0.0017078807577490807, + 0.001637189881876111, + 0.0015939973527565598, + 0.0028577803168445826, + 0.003974383696913719, + 0.0031564501114189625, + 0.005902914796024561, + 0.0028792929369956255, + 0.003833103459328413, + 0.003591321874409914, + 0.0032174692023545504, + 0.005299291107803583, + 0.005917103495448828, + 0.005830947309732437, + 0.010781806893646717, + 0.014988347887992859, + 0.014067137613892555, + 0.015467803925275803, + 0.0076226829551160336, + 0.009031466208398342, + 0.012058713473379612, + 0.018298393115401268, + 0.005779888015240431, + 0.006454048212617636, + 0.0072399009950459, + 0.008671150542795658, + 0.008272731676697731, + 0.005743115674704313, + 0.015010261908173561, + 0.011269786395132542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.03369717672467232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198355376720428 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198355376720428, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198355376720428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10808756947517395, + 0.08680703490972519, + 0.10870411247015, + 0.337451308965683, + 0.04643423110246658, + 0.05798559635877609, + 0.1256207376718521, + 0.03303531929850578, + 0.04335247352719307, + 0.05252166837453842 + ], + "credit_map": [ + 0.015953850001096725, + 0.009889500215649605, + 0.009605384431779385, + 0.01216807495802641, + 0.009744501672685146, + 0.01911323331296444, + 0.012078295461833477, + 0.019534718245267868, + 0.010244578123092651, + 0.00859053898602724, + 0.008672207593917847, + 0.006321905646473169, + 0.01231508981436491, + 0.016005223616957664, + 0.006569548975676298, + 0.01808793842792511, + 0.01595357246696949, + 0.014834936708211899, + 0.012330162338912487, + 0.01412858534604311, + 0.011475710198283195, + 0.012166019529104233, + 0.015285888686776161, + 0.012529238127171993, + 0.04352038353681564, + 0.04565691575407982, + 0.03842437267303467, + 0.04039383679628372, + 0.020114824175834656, + 0.07364807277917862, + 0.018968550488352776, + 0.05672434717416763, + 0.008590444922447205, + 0.0054910071194171906, + 0.005726052448153496, + 0.004672936163842678, + 0.0054760100319981575, + 0.006033900193870068, + 0.004548015538603067, + 0.0058958628214895725, + 0.007336556911468506, + 0.008343309164047241, + 0.007063913159072399, + 0.005236748140305281, + 0.007884540595114231, + 0.006253551691770554, + 0.010014028288424015, + 0.005852950271219015, + 0.023479651659727097, + 0.005784762091934681, + 0.008462340570986271, + 0.00488228490576148, + 0.006323010660707951, + 0.008826283738017082, + 0.0324094295501709, + 0.03545297309756279, + 0.0030986585188657045, + 0.0016736581455916166, + 0.0031401650048792362, + 0.0022771679796278477, + 0.0018293580505996943, + 0.0031312373466789722, + 0.007136870175600052, + 0.01074820663779974, + 0.005497888196259737, + 0.006269251462072134, + 0.005809868685901165, + 0.0061875744722783566, + 0.003110592719167471, + 0.0028628243599087, + 0.00559596810489893, + 0.008018509484827518, + 0.0037370326463133097, + 0.004604679066687822, + 0.00510910339653492, + 0.005051628220826387, + 0.005811849609017372, + 0.005059500224888325, + 0.012436582706868649, + 0.010711298324167728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.014198355376720428 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578570663928986 + ], + "opposite_logprob_tokens": [ + -3.895578622817993 + ], + "logprob_reward": -0.020578570663928986, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578570663928986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09535586833953857, + 0.12622085213661194, + 0.04066690802574158, + 0.04980618134140968, + 0.10937626659870148, + 0.09464748203754425, + 0.10531403124332428, + 0.053469058126211166, + 0.11406167596578598, + 0.21108168363571167 + ], + "credit_map": [ + 0.014120733365416527, + 0.010447600856423378, + 0.015065393410623074, + 0.010320443660020828, + 0.012193790636956692, + 0.011461194604635239, + 0.00792680960148573, + 0.01381990872323513, + 0.011675635352730751, + 0.007489192299544811, + 0.00828295573592186, + 0.005296420305967331, + 0.009471859782934189, + 0.039140500128269196, + 0.004753831773996353, + 0.04011046141386032, + 0.004763273522257805, + 0.0056539131328463554, + 0.005361157935112715, + 0.00426948769018054, + 0.00604550214484334, + 0.00551545200869441, + 0.00360340578481555, + 0.005454718135297298, + 0.005185052752494812, + 0.0036732458975166082, + 0.0028843169566243887, + 0.012562333606183529, + 0.002674028743058443, + 0.004674864467233419, + 0.0024059126153588295, + 0.015746427699923515, + 0.015742575749754906, + 0.014746326953172684, + 0.007851716130971909, + 0.011013754643499851, + 0.009705384261906147, + 0.016859352588653564, + 0.0074877869337797165, + 0.025969373062253, + 0.013897698372602463, + 0.015280217863619328, + 0.013991951011121273, + 0.014956262893974781, + 0.0065795788541436195, + 0.01168433204293251, + 0.007552333641797304, + 0.010705109685659409, + 0.020677048712968826, + 0.008332670666277409, + 0.0066550360061228275, + 0.006117550656199455, + 0.004750005900859833, + 0.005595563445240259, + 0.03989240527153015, + 0.013293767347931862, + 0.008322748355567455, + 0.0069450694136321545, + 0.007334993686527014, + 0.0071153463795781136, + 0.005662593524903059, + 0.004937468096613884, + 0.006594546604901552, + 0.006556294858455658, + 0.013897868804633617, + 0.014649692922830582, + 0.015020486898720264, + 0.02050590328872204, + 0.0072396304458379745, + 0.007995134219527245, + 0.013504673726856709, + 0.02124829962849617, + 0.01807563006877899, + 0.029046479612588882, + 0.022211499512195587, + 0.03246862068772316, + 0.022486092522740364, + 0.017216522246599197, + 0.036628156900405884, + 0.03294869139790535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.020578570663928986 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a slightly tilted perspective, the camera moves in a slow, arcing motion, circling a scene set against an infinitely stretching, pure white backdrop. The image captures the signature high-contrast black-and-white portrait quality of Richard Avedon. A small, animated game robot, with a cubic body and spindly limbs, struggles to propel a highly polished black bowling ball, its size dramatically contrasting, down an abstract lane. Sharp studio lighting, directed from above, accentuates the robot's matte black exterior with distinct edges and casts harsh, almost overexposed white highlights on the ball's smooth surface, while deep shadows separate them from the pristine white ground. The robot leans forward, its legs pushing against the ground, its mechanical joints seemingly straining with immense force. The camera rotates, focusing in close-up on a small screen on the robot's face. The screen, composed of simple black and white pixels, captures a resolute frown, each pixel conveying unyielding effort. As the robot's legs struggled to move on the ground, the huge bowling ball rolled slowly inch by inch. The entire scene, in extreme black and white contrast, created a unique atmosphere that blended absurdity, concentration and epic struggle.", + "seed": 8945, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_12/index_482_seed_8945_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03297988139092922, + "vlm_logprob": -0.03297988139092922, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.03297988139092922, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03101842664182186 + ], + "opposite_logprob_tokens": [ + -3.6560184955596924 + ], + "logprob_reward": -0.03101842664182186, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03101842664182186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10068900883197784, + 0.05150063335895538, + 0.09506997466087341, + 0.07885981351137161, + 0.051510997116565704, + 0.05621230974793434, + 0.09533080458641052, + 0.39850515127182007, + 0.04174871742725372, + 0.030572595074772835 + ], + "credit_map": [ + 0.015912888571619987, + 0.008460714481770992, + 0.017654113471508026, + 0.01027203444391489, + 0.01206857431679964, + 0.016117803752422333, + 0.007619030307978392, + 0.012583848088979721, + 0.007796324789524078, + 0.005422447342425585, + 0.00916375033557415, + 0.004489606712013483, + 0.005790218710899353, + 0.00771924015134573, + 0.0046092672273516655, + 0.006509777158498764, + 0.00895038340240717, + 0.00881736446171999, + 0.017805831506848335, + 0.015548400580883026, + 0.010787532664835453, + 0.013542630709707737, + 0.008087839931249619, + 0.011529985815286636, + 0.010514020919799805, + 0.006476937793195248, + 0.011879295110702515, + 0.006878742482513189, + 0.007338107563555241, + 0.0119253471493721, + 0.008787580765783787, + 0.015059774741530418, + 0.005545922089368105, + 0.005435301922261715, + 0.00944103579968214, + 0.008159606717526913, + 0.005415603052824736, + 0.007458158768713474, + 0.005420293193310499, + 0.004635075572878122, + 0.005581368692219257, + 0.007119218818843365, + 0.007473069708794355, + 0.010531867854297161, + 0.006301346700638533, + 0.007292189169675112, + 0.00650705024600029, + 0.0054061999544501305, + 0.015570418909192085, + 0.012415005825459957, + 0.007586588617414236, + 0.011210700497031212, + 0.013512255623936653, + 0.014865796081721783, + 0.008869219571352005, + 0.011300824582576752, + 0.12786991894245148, + 0.02366415597498417, + 0.028588639572262764, + 0.029598474502563477, + 0.02882613241672516, + 0.08934221416711807, + 0.017692141234874725, + 0.052923452109098434, + 0.006113920360803604, + 0.007569607347249985, + 0.004728577099740505, + 0.005079757887870073, + 0.004842695780098438, + 0.005234353244304657, + 0.0035431724973022938, + 0.0046366360038518906, + 0.004639548249542713, + 0.003874002955853939, + 0.005679767113178968, + 0.0019033156568184495, + 0.003608604660257697, + 0.004016313701868057, + 0.0031530652195215225, + 0.0036979764699935913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.03101842664182186 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307710379362106 + ], + "opposite_logprob_tokens": [ + -3.168076992034912 + ], + "logprob_reward": -0.04307710379362106, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307710379362106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10015706717967987, + 0.05119877681136131, + 0.07849855720996857, + 0.025282632559537888, + 0.1433197408914566, + 0.17625240981578827, + 0.15006141364574432, + 0.16040389239788055, + 0.06523653864860535, + 0.049588944762945175 + ], + "credit_map": [ + 0.011686771176755428, + 0.010492697358131409, + 0.01475572120398283, + 0.011294735595583916, + 0.009126204997301102, + 0.015085446648299694, + 0.008353075012564659, + 0.01936241425573826, + 0.0062225377187132835, + 0.005017756950110197, + 0.010565140284597874, + 0.005494292825460434, + 0.004994491580873728, + 0.007538237143307924, + 0.004777899011969566, + 0.006588421296328306, + 0.00923346821218729, + 0.007720121648162603, + 0.013948227278888226, + 0.008255531080067158, + 0.00800516176968813, + 0.011900573968887329, + 0.006835781503468752, + 0.012599687092006207, + 0.0022587415296584368, + 0.0019156844355165958, + 0.004578287247568369, + 0.0026722054462879896, + 0.0033585927449166775, + 0.003811132861301303, + 0.003492757910862565, + 0.003195230383425951, + 0.019748535007238388, + 0.012571249157190323, + 0.021277133375406265, + 0.021278923377394676, + 0.01688002049922943, + 0.021299712359905243, + 0.01539972797036171, + 0.01486444566398859, + 0.020895615220069885, + 0.024992680177092552, + 0.023793434724211693, + 0.027721386402845383, + 0.023110369220376015, + 0.023947205394506454, + 0.01681193709373474, + 0.014979783445596695, + 0.03173564746975899, + 0.015811163932085037, + 0.011892545968294144, + 0.01785344071686268, + 0.021906526759266853, + 0.025134960189461708, + 0.011879097670316696, + 0.013848032802343369, + 0.021007947623729706, + 0.01781964860856533, + 0.0160500630736351, + 0.02110622636973858, + 0.018187688663601875, + 0.020588714629411697, + 0.016619650647044182, + 0.029023952782154083, + 0.009348632767796516, + 0.011803669854998589, + 0.006497493479400873, + 0.007754527032375336, + 0.007760122884064913, + 0.007290520239621401, + 0.005822912789881229, + 0.008958658203482628, + 0.005347354803234339, + 0.007597749121487141, + 0.008908736519515514, + 0.003583880141377449, + 0.007139013614505529, + 0.006143173202872276, + 0.0049219559878110886, + 0.005947084631770849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.04307710379362106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635246515274048 + ], + "opposite_logprob_tokens": [ + -3.6513524055480957 + ], + "logprob_reward": -0.02635246515274048, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635246515274048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14661799371242523, + 0.13104994595050812, + 0.10067115724086761, + 0.05396629497408867, + 0.1250918209552765, + 0.12962551414966583, + 0.10140329599380493, + 0.09256903827190399, + 0.07645997405052185, + 0.04254494234919548 + ], + "credit_map": [ + 0.018821299076080322, + 0.011346837505698204, + 0.01870070770382881, + 0.013944792561233044, + 0.013840170577168465, + 0.021551234647631645, + 0.021129995584487915, + 0.02728295512497425, + 0.018987584859132767, + 0.015152018517255783, + 0.021401898935437202, + 0.01322737243026495, + 0.012060343287885189, + 0.01861797831952572, + 0.018274063244462013, + 0.013328680768609047, + 0.009971000254154205, + 0.010859338566660881, + 0.016340434551239014, + 0.011100181378424168, + 0.010371927171945572, + 0.014295287430286407, + 0.01308488566428423, + 0.014648106880486012, + 0.007169515825808048, + 0.005448952782899141, + 0.00784517265856266, + 0.007046394515782595, + 0.005451332312077284, + 0.007845401763916016, + 0.005498072598129511, + 0.007661449257284403, + 0.024108929559588432, + 0.009869717992842197, + 0.014314726926386356, + 0.012086881324648857, + 0.015496289357542992, + 0.017562393099069595, + 0.014136026613414288, + 0.017516862601041794, + 0.01815161295235157, + 0.015120050869882107, + 0.018920429050922394, + 0.020035255700349808, + 0.012652778066694736, + 0.01955702342092991, + 0.013241839595139027, + 0.011946531943976879, + 0.021312613040208817, + 0.01279446016997099, + 0.008895414881408215, + 0.013625423423945904, + 0.011107204481959343, + 0.016066303476691246, + 0.008578941226005554, + 0.009022935293614864, + 0.013071740046143532, + 0.010848408564925194, + 0.009874978102743626, + 0.01413889043033123, + 0.011473073624074459, + 0.012215331196784973, + 0.009235059842467308, + 0.011711560189723969, + 0.008312851190567017, + 0.008063895627856255, + 0.008160419762134552, + 0.006653045769780874, + 0.008207547478377819, + 0.010742557235062122, + 0.008782630786299706, + 0.017537031322717667, + 0.004044553264975548, + 0.005151417572051287, + 0.007759169675409794, + 0.002545654308050871, + 0.005257323384284973, + 0.005364859011024237, + 0.007368464022874832, + 0.005053500644862652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.02635246515274048 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355832815170288 + ], + "opposite_logprob_tokens": [ + -3.651355743408203 + ], + "logprob_reward": -0.026355832815170288, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355832815170288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09256362169981003, + 0.0884925127029419, + 0.09154597669839859, + 0.10560982674360275, + 0.06321956217288971, + 0.09791803359985352, + 0.1217040941119194, + 0.1468883454799652, + 0.14906997978687286, + 0.04298806190490723 + ], + "credit_map": [ + 0.012137616984546185, + 0.009280969388782978, + 0.012344197370111942, + 0.011632882989943027, + 0.009076165035367012, + 0.013409404084086418, + 0.012084550224244595, + 0.012597830034792423, + 0.00995406974107027, + 0.010368413291871548, + 0.016065623611211777, + 0.00910946261137724, + 0.00808279775083065, + 0.013265091925859451, + 0.011447503231465816, + 0.010199536569416523, + 0.011051039211452007, + 0.0094852140173316, + 0.01600033789873123, + 0.015222902409732342, + 0.008217084221541882, + 0.012649724259972572, + 0.007461941335350275, + 0.011457725428044796, + 0.011834116652607918, + 0.011419311165809631, + 0.02023540623486042, + 0.010966314002871513, + 0.009934204630553722, + 0.014979643747210503, + 0.011685704812407494, + 0.014555122703313828, + 0.006801125127822161, + 0.007830138318240643, + 0.00782393105328083, + 0.0069833784364163876, + 0.009638115763664246, + 0.009115040302276611, + 0.008905653841793537, + 0.006122182123363018, + 0.011691482737660408, + 0.011233326978981495, + 0.01263683196157217, + 0.01655999943614006, + 0.011735386215150356, + 0.014183794148266315, + 0.01056087389588356, + 0.009316331706941128, + 0.01724771410226822, + 0.018580783158540726, + 0.013550303876399994, + 0.015142603777348995, + 0.015769844874739647, + 0.01816762611269951, + 0.010414138436317444, + 0.012831076048314571, + 0.03263802453875542, + 0.018277348950505257, + 0.009932020679116249, + 0.014058372005820274, + 0.01738027296960354, + 0.02264218032360077, + 0.010505340062081814, + 0.021454785019159317, + 0.016535134986042976, + 0.02221638336777687, + 0.012614460662007332, + 0.012987539172172546, + 0.009990687482059002, + 0.024505505338311195, + 0.015089389868080616, + 0.03513087332248688, + 0.005755251739174128, + 0.006187126040458679, + 0.0053140693344175816, + 0.003118207212537527, + 0.005150855053216219, + 0.005574456416070461, + 0.006083160173147917, + 0.005804933607578278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.026355832815170288 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set against an infinitely stretching, pure white backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a small animated game robot with a cubic body and spindly limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the robot struggling to propel a highly polished black bowling ball down an abstract lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move in a slow, arcing motion circling the scene from a slightly tilted perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the robot lean forward with legs pushing against the ground while a close-up focuses on a small screen on its face showing a pixelated frown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809557855129242 + ], + "opposite_logprob_tokens": [ + -3.288095474243164 + ], + "logprob_reward": -0.03809557855129242, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809557855129242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060838066041469574, + 0.10941621661186218, + 0.03492050990462303, + 0.033493734896183014, + 0.08649609982967377, + 0.126265287399292, + 0.1111069843173027, + 0.12103155255317688, + 0.11210738867521286, + 0.2043241709470749 + ], + "credit_map": [ + 0.010304536670446396, + 0.005184609908610582, + 0.006771569140255451, + 0.005540931597352028, + 0.005421073641628027, + 0.009544281288981438, + 0.00843635480850935, + 0.009634709917008877, + 0.011973696760833263, + 0.011855396442115307, + 0.01818927563726902, + 0.01163354329764843, + 0.011237735860049725, + 0.015162481926381588, + 0.01631239987909794, + 0.013051684014499187, + 0.003899444593116641, + 0.003327904036268592, + 0.005703231319785118, + 0.004311864729970694, + 0.0041648633778095245, + 0.005465404130518436, + 0.003771962830796838, + 0.004275834187865257, + 0.004106713458895683, + 0.002631152980029583, + 0.0048914458602666855, + 0.005083517637103796, + 0.0040768529288470745, + 0.0046559968031942844, + 0.004091986920684576, + 0.003956067841500044, + 0.011102257296442986, + 0.008794604800641537, + 0.012429707683622837, + 0.008820547722280025, + 0.009099473245441914, + 0.012117047794163227, + 0.0110749751329422, + 0.013057486154139042, + 0.017529822885990143, + 0.013367627747356892, + 0.020268701016902924, + 0.019001543521881104, + 0.01310613565146923, + 0.020034033805131912, + 0.013415807858109474, + 0.009541615843772888, + 0.01528953481465578, + 0.017368566244840622, + 0.011756706051528454, + 0.016626909375190735, + 0.01456315815448761, + 0.01650090143084526, + 0.00925002433359623, + 0.009751178324222565, + 0.020376378670334816, + 0.014324923045933247, + 0.019533857703208923, + 0.012757639400660992, + 0.01868421770632267, + 0.0170420091599226, + 0.006956626195460558, + 0.011355899274349213, + 0.01416105031967163, + 0.020345410332083702, + 0.016271516680717468, + 0.011708838865160942, + 0.010014496743679047, + 0.015419527888298035, + 0.00978907197713852, + 0.014397482387721539, + 0.0232340507209301, + 0.03244205191731453, + 0.028067858889698982, + 0.018128518015146255, + 0.03180728852748871, + 0.026472490280866623, + 0.019846901297569275, + 0.024325016885995865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.03809557855129242 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_13/index_tensor([7606, 7606])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_13/index_tensor([7606, 7606])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fa83ee34a0257fb25b7fb51dd758439c9a95bcbf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_13/index_tensor([7606, 7606])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly pans over a U-shaped wooden chisel with a polished, sleek handle. Steadily steered by a pair of hands, marked by age but steady, the chisel's sharp metal blade slices precisely into the finely grained surface of light-colored oak. As the blade's steady and deliberate advance unfolds, a thin, curling shavings are elegantly detached from the wood, forming a perfect spiral in mid-air before cascading down, feather-light, to a workbench strewn with similar sawdust. The scene is bathed in a warm, vintage hue. Light comes from a vintage tungsten lamp suspended above the workbench, casting a concentrated beam with a distinct warm yellow hue. This light not only highlights the three-dimensionality of the sawdust and the delicate grain of the wood, but also casts a soft glow on the polished handle. In the background, a slightly blurred woodworking workshop appears, with faint glimpses of other handmade tools hanging on the wall, creating an atmosphere of focus, tranquility, and craftsmanship.", + "seed": 5570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_13/index_7606_seed_5570_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2727308213710785, + "vlm_logprob": -0.2727308213710785, + "vlm_yes_no_margin": 2.950000011920929, + "vlm_token_credit": -0.2727308213710785, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02068203315138817 + ], + "opposite_logprob_tokens": [ + -6.020681858062744 + ], + "logprob_reward": -0.02068203315138817, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.02068203315138817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1905672699213028, + 0.10104035586118698, + 0.11756820231676102, + 0.08812844008207321, + 0.09604135155677795, + 0.0839911475777626, + 0.07483866065740585, + 0.10214778780937195, + 0.05507856234908104, + 0.0905982181429863 + ], + "credit_map": [ + 0.015650315210223198, + 0.01103611197322607, + 0.010450107976794243, + 0.01453740056604147, + 0.0405506007373333, + 0.04533948749303818, + 0.028515733778476715, + 0.024487517774105072, + 0.013871748931705952, + 0.006518670357763767, + 0.011198824271559715, + 0.013164007104933262, + 0.011382829397916794, + 0.01672905869781971, + 0.01745631918311119, + 0.01071890164166689, + 0.007801568601280451, + 0.01172961387783289, + 0.009526485577225685, + 0.011243239976465702, + 0.03224465996026993, + 0.02270738035440445, + 0.016333764418959618, + 0.005981494206935167, + 0.015663214027881622, + 0.009534118697047234, + 0.0108920494094491, + 0.012060306966304779, + 0.015479924157261848, + 0.012609259225428104, + 0.005543012171983719, + 0.006346559152007103, + 0.01106381043791771, + 0.006751516833901405, + 0.009262694977223873, + 0.007152775302529335, + 0.024574361741542816, + 0.025091422721743584, + 0.005773601587861776, + 0.006371169351041317, + 0.016207220032811165, + 0.006727228872478008, + 0.009462939575314522, + 0.010051204822957516, + 0.012145974673330784, + 0.009579221718013287, + 0.010515019297599792, + 0.009302335791289806, + 0.012662255205214024, + 0.01051745843142271, + 0.008945446461439133, + 0.010951186530292034, + 0.010178696364164352, + 0.009274649433791637, + 0.006376317236572504, + 0.0059326509945094585, + 0.012198485434055328, + 0.008666351437568665, + 0.009565179236233234, + 0.012506149709224701, + 0.018940100446343422, + 0.01078091748058796, + 0.011275740340352058, + 0.018214859068393707, + 0.010191391222178936, + 0.006367207039147615, + 0.0073025724850595, + 0.0071151284500956535, + 0.00871234480291605, + 0.007253074087202549, + 0.0034620489459484816, + 0.004674792755395174, + 0.008905749768018723, + 0.010396357625722885, + 0.012252861633896828, + 0.010424606502056122, + 0.0188249871134758, + 0.015889551490545273, + 0.006369594018906355, + 0.007534513249993324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.02068203315138817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311535134911537 + ], + "opposite_logprob_tokens": [ + -3.1681153774261475 + ], + "logprob_reward": -0.04311535134911537, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311535134911537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3594866096973419, + 0.05684201419353485, + 0.2449098825454712, + 0.085384301841259, + 0.04154073819518089, + 0.07272583246231079, + 0.05165048688650131, + 0.03309755399823189, + 0.027656415477395058, + 0.02670620009303093 + ], + "credit_map": [ + 0.02386368438601494, + 0.019225824624300003, + 0.017132628709077835, + 0.028321800753474236, + 0.07463712990283966, + 0.08672964572906494, + 0.06079686060547829, + 0.048779044300317764, + 0.0065611302852630615, + 0.0048021115362644196, + 0.006432780530303717, + 0.004093472845852375, + 0.009124365635216236, + 0.009000380523502827, + 0.009194761514663696, + 0.007633011322468519, + 0.0529610812664032, + 0.02814369834959507, + 0.03641932085156441, + 0.01767958514392376, + 0.0279637910425663, + 0.045849431306123734, + 0.01815497875213623, + 0.01773800328373909, + 0.01657828502357006, + 0.006349410396069288, + 0.01076999306678772, + 0.007044200319796801, + 0.015024455264210701, + 0.015628604218363762, + 0.0060196234844625, + 0.007969734258949757, + 0.0029065159615129232, + 0.0022917611058801413, + 0.0032633512746542692, + 0.0032567481976002455, + 0.013510170392692089, + 0.012454823590815067, + 0.002012267243117094, + 0.0018450971692800522, + 0.010815940797328949, + 0.00574811827391386, + 0.0055019306018948555, + 0.0060724420472979546, + 0.01706789620220661, + 0.01376788504421711, + 0.00694649713113904, + 0.0068051256239414215, + 0.007846326567232609, + 0.005988781806081533, + 0.005270815920084715, + 0.00539914658293128, + 0.010196374729275703, + 0.007768468465656042, + 0.003889982821419835, + 0.005290588829666376, + 0.004597401712089777, + 0.00428084097802639, + 0.004700349643826485, + 0.00489790178835392, + 0.004097426775842905, + 0.003297646064311266, + 0.0038341693580150604, + 0.0033918186090886593, + 0.0019265715964138508, + 0.0032032353337854147, + 0.006640295963734388, + 0.0047346749342978, + 0.002831730991601944, + 0.002751437947154045, + 0.002689886139705777, + 0.0028785814065486193, + 0.004386940971016884, + 0.0034308522008359432, + 0.004059536848217249, + 0.0026221591979265213, + 0.004035752732306719, + 0.003382074646651745, + 0.0021181870251893997, + 0.002670695073902607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.04311535134911537 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038101889193058014 + ], + "opposite_logprob_tokens": [ + -3.2881019115448 + ], + "logprob_reward": -0.038101889193058014, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038101889193058014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3650062084197998, + 0.058626022189855576, + 0.23379236459732056, + 0.07045639306306839, + 0.08200700581073761, + 0.03031162917613983, + 0.016761040315032005, + 0.06816940009593964, + 0.03570552170276642, + 0.03916437551379204 + ], + "credit_map": [ + 0.030951565131545067, + 0.023386767134070396, + 0.020787915214896202, + 0.026309233158826828, + 0.06287563592195511, + 0.07263120263814926, + 0.08223450183868408, + 0.0458294078707695, + 0.005898192524909973, + 0.003457085695117712, + 0.005100498907268047, + 0.0038752614054828882, + 0.007695995736867189, + 0.014632444828748703, + 0.00706806406378746, + 0.010898481123149395, + 0.03541294485330582, + 0.02347700297832489, + 0.027171047404408455, + 0.017275797203183174, + 0.04459730535745621, + 0.04440552368760109, + 0.025955304503440857, + 0.015497446060180664, + 0.011405487544834614, + 0.005708602257072926, + 0.009789884090423584, + 0.00693736644461751, + 0.014360480010509491, + 0.01156260259449482, + 0.005304355174303055, + 0.005387617275118828, + 0.011759641580283642, + 0.004661696497350931, + 0.007513744290918112, + 0.0047182426787912846, + 0.022940101101994514, + 0.022191474214196205, + 0.0038863406516611576, + 0.00433577224612236, + 0.002341942396014929, + 0.0026409130077809095, + 0.003518318058922887, + 0.004228048492223024, + 0.007039714604616165, + 0.0058378977701067924, + 0.002578700426965952, + 0.0021260969806462526, + 0.0025195791386067867, + 0.001976432278752327, + 0.0023373581934720278, + 0.0018672073492780328, + 0.0026774273719638586, + 0.0019815866835415363, + 0.001354505424387753, + 0.0020469441078603268, + 0.006377373356372118, + 0.008404217660427094, + 0.012609549798071384, + 0.01583554595708847, + 0.008951353840529919, + 0.005405253265053034, + 0.006129696499556303, + 0.004456416238099337, + 0.0030886910390108824, + 0.00379024725407362, + 0.007265620399266481, + 0.004344176035374403, + 0.006902237888425589, + 0.004987272433936596, + 0.002279393607750535, + 0.003047883277758956, + 0.004146165680140257, + 0.003916661720722914, + 0.00636830972507596, + 0.007108849007636309, + 0.006019626744091511, + 0.003979350905865431, + 0.003750584088265896, + 0.003874830435961485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.038101889193058014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038114629685878754 + ], + "opposite_logprob_tokens": [ + -3.288114547729492 + ], + "logprob_reward": -0.038114629685878754, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038114629685878754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18059132993221283, + 0.10923799127340317, + 0.15866988897323608, + 0.09658361971378326, + 0.1955319494009018, + 0.052043113857507706, + 0.049570731818675995, + 0.08937786519527435, + 0.03078513592481613, + 0.037608351558446884 + ], + "credit_map": [ + 0.01571691781282425, + 0.00701104523614049, + 0.010225865058600903, + 0.01072751171886921, + 0.04968683421611786, + 0.05336746573448181, + 0.011276141740381718, + 0.02257956750690937, + 0.021184150129556656, + 0.01270681619644165, + 0.012334403581917286, + 0.008501917123794556, + 0.012563993223011494, + 0.015281634405255318, + 0.012296322733163834, + 0.014368769712746143, + 0.028307456523180008, + 0.015115687623620033, + 0.02359914407134056, + 0.013788913376629353, + 0.02388838864862919, + 0.025885790586471558, + 0.01576836034655571, + 0.012316161766648293, + 0.01880364678800106, + 0.007268564775586128, + 0.014314494095742702, + 0.011462584137916565, + 0.01769103668630123, + 0.013993539847433567, + 0.0052284961566329, + 0.007821265608072281, + 0.007375878281891346, + 0.006476145703345537, + 0.00847648549824953, + 0.004905703943222761, + 0.08120408654212952, + 0.07616209238767624, + 0.0038351258262991905, + 0.007096436806023121, + 0.006927249021828175, + 0.006379077676683664, + 0.0057748304679989815, + 0.0053592524491250515, + 0.009830843657255173, + 0.008396375924348831, + 0.004518403671681881, + 0.0048570833168923855, + 0.004310091957449913, + 0.005654015578329563, + 0.006637994199991226, + 0.009801249951124191, + 0.0062909964472055435, + 0.006351557560265064, + 0.00522235780954361, + 0.005302473437041044, + 0.008745911531150341, + 0.011061647906899452, + 0.01596040278673172, + 0.018357673659920692, + 0.011593755334615707, + 0.007471154909580946, + 0.007933191023766994, + 0.00825413502752781, + 0.0035989517346024513, + 0.0028689212631434202, + 0.004492148756980896, + 0.0031566352117806673, + 0.00636131688952446, + 0.004596765153110027, + 0.0021050027571618557, + 0.0036053939256817102, + 0.0064125400967895985, + 0.00494425930082798, + 0.005293733440339565, + 0.00449416832998395, + 0.004643288441002369, + 0.0032649594359099865, + 0.004031329415738583, + 0.004524072166532278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.038114629685878754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236402034759521 + ], + "opposite_logprob_tokens": [ + -0.3486402630805969 + ], + "logprob_reward": -1.2236402034759521, + "yes_no_margin_reward": -0.8749999403953552, + "token_credit_reward": -1.2236402034759521, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10856187343597412, + 0.08802797645330429, + 0.10285426676273346, + 0.07167381793260574, + 0.044876545667648315, + 0.11286211013793945, + 0.06104680895805359, + 0.22595740854740143, + 0.10126478224992752, + 0.08287446200847626 + ], + "credit_map": [ + 0.007582757622003555, + 0.005974673666059971, + 0.007138425018638372, + 0.00748911127448082, + 0.023340759798884392, + 0.026248827576637268, + 0.019283965229988098, + 0.011503340676426888, + 0.008285840973258018, + 0.007127285003662109, + 0.008741797879338264, + 0.0056321267038583755, + 0.012676248326897621, + 0.022505298256874084, + 0.009316391311585903, + 0.013742981478571892, + 0.018912354484200478, + 0.009797736071050167, + 0.01571127399802208, + 0.01061901357024908, + 0.013486653566360474, + 0.01633942686021328, + 0.011661885306239128, + 0.006325902882963419, + 0.01602276973426342, + 0.005179394967854023, + 0.009050310589373112, + 0.007658087648451328, + 0.01424841396510601, + 0.01028615701943636, + 0.004837047774344683, + 0.004391627851873636, + 0.008493018336594105, + 0.00419607013463974, + 0.004736186936497688, + 0.0058086044155061245, + 0.006348699796944857, + 0.006480591371655464, + 0.003972270525991917, + 0.004841102287173271, + 0.007909102365374565, + 0.013687670230865479, + 0.01479042787104845, + 0.01583487167954445, + 0.016892075538635254, + 0.016024157404899597, + 0.011533680371940136, + 0.016190124675631523, + 0.004432066809386015, + 0.006850540172308683, + 0.007234691642224789, + 0.009813397191464901, + 0.012581723742187023, + 0.009937640279531479, + 0.0049788812175393105, + 0.005217867437750101, + 0.019395975396037102, + 0.02318200096487999, + 0.02995106391608715, + 0.036361243575811386, + 0.04601443186402321, + 0.029690314084291458, + 0.016942670568823814, + 0.02441968396306038, + 0.01200052723288536, + 0.009597525000572205, + 0.014403761364519596, + 0.011323194019496441, + 0.0215604305267334, + 0.01594090647995472, + 0.006781396921724081, + 0.009657032787799835, + 0.009454192593693733, + 0.009419683367013931, + 0.011003512889146805, + 0.01270345225930214, + 0.016236813738942146, + 0.013525154441595078, + 0.00489708362147212, + 0.00563456118106842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -1.2236402034759521 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly pans over a U-shaped wooden chisel with a polished, sleek handle. Steadily steered by a pair of hands, marked by age but steady, the chisel's sharp metal blade slices precisely into the finely grained surface of light-colored oak. As the blade's steady and deliberate advance unfolds, a thin, curling shavings are elegantly detached from the wood, forming a perfect spiral in mid-air before cascading down, feather-light, to a workbench strewn with similar sawdust. The scene is bathed in a warm, vintage hue. Light comes from a vintage tungsten lamp suspended above the workbench, casting a concentrated beam with a distinct warm yellow hue. This light not only highlights the three-dimensionality of the sawdust and the delicate grain of the wood, but also casts a soft glow on the polished handle. In the background, a slightly blurred woodworking workshop appears, with faint glimpses of other handmade tools hanging on the wall, creating an atmosphere of focus, tranquility, and craftsmanship.", + "seed": 5571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_13/index_7606_seed_5571_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10417515970766544, + "vlm_logprob": -0.10417515970766544, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.10417515970766544, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03260253742337227 + ], + "opposite_logprob_tokens": [ + -5.907602310180664 + ], + "logprob_reward": -0.03260253742337227, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.03260253742337227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21335892379283905, + 0.0908297598361969, + 0.11034640669822693, + 0.08994774520397186, + 0.11108268797397614, + 0.08445204794406891, + 0.09134943038225174, + 0.10119009763002396, + 0.044911280274391174, + 0.06253165006637573 + ], + "credit_map": [ + 0.012897219508886337, + 0.010936387814581394, + 0.044970352202653885, + 0.014676940627396107, + 0.014578673057258129, + 0.018733831122517586, + 0.020217884331941605, + 0.07634763419628143, + 0.007349483203142881, + 0.012960169464349747, + 0.009530488401651382, + 0.007137018721550703, + 0.009755217470228672, + 0.019715314731001854, + 0.01245193462818861, + 0.011930132284760475, + 0.008527552708983421, + 0.011925268918275833, + 0.014367280527949333, + 0.009902698919177055, + 0.009352073073387146, + 0.028243916109204292, + 0.012183040380477905, + 0.015844574198126793, + 0.009327344596385956, + 0.0170760378241539, + 0.017833970487117767, + 0.02049580216407776, + 0.006906986236572266, + 0.00592731824144721, + 0.005508322734385729, + 0.006871961057186127, + 0.015088333748281002, + 0.011736281216144562, + 0.028783679008483887, + 0.0320155993103981, + 0.006706231739372015, + 0.0054518659599125385, + 0.006673510652035475, + 0.004627187270671129, + 0.009908360429108143, + 0.010811327025294304, + 0.012072627432644367, + 0.011066303588449955, + 0.008267014287412167, + 0.011978156864643097, + 0.012869483791291714, + 0.0074787745252251625, + 0.011926556006073952, + 0.014915362000465393, + 0.009167181327939034, + 0.012504838407039642, + 0.011719406582415104, + 0.012128496542572975, + 0.011087414808571339, + 0.00790016632527113, + 0.020925099030137062, + 0.010611002333462238, + 0.02675888128578663, + 0.011189047247171402, + 0.006312007084488869, + 0.0035676765255630016, + 0.016835210844874382, + 0.004991171881556511, + 0.004745647311210632, + 0.005008801352232695, + 0.007001631893217564, + 0.006917780265212059, + 0.005034150555729866, + 0.004557589069008827, + 0.005789399612694979, + 0.005856277886778116, + 0.006343733984977007, + 0.00907987542450428, + 0.006993014365434647, + 0.0077238790690898895, + 0.006938855163753033, + 0.00657395226880908, + 0.008376085199415684, + 0.010502257384359837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.03260253742337227 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054994914680719376 + ], + "opposite_logprob_tokens": [ + -2.929994821548462 + ], + "logprob_reward": -0.054994914680719376, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054994914680719376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05578078329563141, + 0.014741352759301662, + 0.017362279817461967, + 0.15306034684181213, + 0.6622082591056824, + 0.027196001261472702, + 0.021167222410440445, + 0.011466247960925102, + 0.013387175276875496, + 0.023630307987332344 + ], + "credit_map": [ + 0.005465840920805931, + 0.005060338415205479, + 0.009519841521978378, + 0.004958726931363344, + 0.006306879222393036, + 0.008169720880687237, + 0.006000235676765442, + 0.010299190878868103, + 0.0014531059423461556, + 0.0015296385390684009, + 0.0019913993310183287, + 0.001550436019897461, + 0.0014616400003433228, + 0.0020850624423474073, + 0.002416381612420082, + 0.0022536867763847113, + 0.0015326272696256638, + 0.003333770204335451, + 0.002137436531484127, + 0.0014352636644616723, + 0.00213998812250793, + 0.0018737739883363247, + 0.0018476077821105719, + 0.003061809344217181, + 0.01912621222436428, + 0.04194912314414978, + 0.02156064100563526, + 0.04123516380786896, + 0.012709404341876507, + 0.002721297787502408, + 0.009445871226489544, + 0.004312626551836729, + 0.07259995490312576, + 0.047831494361162186, + 0.21382899582386017, + 0.23694300651550293, + 0.013870961032807827, + 0.019233155995607376, + 0.03834507241845131, + 0.019555583596229553, + 0.0032522708643227816, + 0.002578487852588296, + 0.0036588057409971952, + 0.0031512512359768152, + 0.0029559049289673567, + 0.0029842713847756386, + 0.00568948732689023, + 0.002925517503172159, + 0.0027583218179643154, + 0.003091008635237813, + 0.0021564422640949488, + 0.0030484586022794247, + 0.0027862496208399534, + 0.0027464504819363356, + 0.0027924994938075542, + 0.0017877903301268816, + 0.0016077685868367553, + 0.0021846250165253878, + 0.0018931507365778089, + 0.0011463259579613805, + 0.001181730767711997, + 0.0013720423448830843, + 0.0011281834449619055, + 0.0009524197084829211, + 0.001867808517999947, + 0.0018213114235550165, + 0.0014845477417111397, + 0.0017048562876880169, + 0.002158344956114888, + 0.001741896034218371, + 0.0014166224282234907, + 0.0011917861411347985, + 0.0026080512907356024, + 0.0025121846701949835, + 0.0022342922165989876, + 0.0026626335456967354, + 0.004211935680359602, + 0.002192888641729951, + 0.003920132759958506, + 0.00328818685375154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.054994914680719376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540886178612709 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540886178612709, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540886178612709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09990890324115753, + 0.09495120495557785, + 0.05179411172866821, + 0.30139783024787903, + 0.2588164806365967, + 0.03242313861846924, + 0.035384126007556915, + 0.05667867138981819, + 0.014984481036663055, + 0.05366106703877449 + ], + "credit_map": [ + 0.010572453960776329, + 0.01007787138223648, + 0.025610661134123802, + 0.008329766802489758, + 0.009823077358305454, + 0.008250911720097065, + 0.0072056809440255165, + 0.020038483664393425, + 0.00710604153573513, + 0.008813902735710144, + 0.010931962169706821, + 0.008750803768634796, + 0.009679239243268967, + 0.017185784876346588, + 0.01877521723508835, + 0.013708260841667652, + 0.005415627732872963, + 0.00883426796644926, + 0.005198282655328512, + 0.0045942566357553005, + 0.005693037062883377, + 0.006969247478991747, + 0.005025714635848999, + 0.010063677094876766, + 0.03666975349187851, + 0.07983610779047012, + 0.043970752507448196, + 0.08277170360088348, + 0.02448439784348011, + 0.005953567568212748, + 0.018396593630313873, + 0.009314942173659801, + 0.03178766742348671, + 0.019436638802289963, + 0.07978793978691101, + 0.09126559644937515, + 0.005384812597185373, + 0.009057139977812767, + 0.014921814203262329, + 0.007174868136644363, + 0.003962298389524221, + 0.003096535336226225, + 0.003720394568517804, + 0.003344515338540077, + 0.003912738990038633, + 0.0035947784781455994, + 0.006847842130810022, + 0.003944034222513437, + 0.004922502674162388, + 0.002729568397626281, + 0.008985159918665886, + 0.0043247174471616745, + 0.0021437755785882473, + 0.003033200278878212, + 0.006212475709617138, + 0.0030327218119055033, + 0.010509836487472057, + 0.005860814359039068, + 0.015122747048735619, + 0.007237959653139114, + 0.0037093183491379023, + 0.0026109281461685896, + 0.008941135369241238, + 0.002685929648578167, + 0.0017968370812013745, + 0.0019174758344888687, + 0.001667015254497528, + 0.00336804729886353, + 0.0013596671633422375, + 0.0013267849572002888, + 0.00217332667671144, + 0.0013753267703577876, + 0.0055387564934790134, + 0.007148039061576128, + 0.005509621929377317, + 0.0070566702634096146, + 0.004621739499270916, + 0.0057534221559762955, + 0.007654249668121338, + 0.01037856750190258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.012540886178612709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370340168476105 + ], + "opposite_logprob_tokens": [ + -3.408703327178955 + ], + "logprob_reward": -0.03370340168476105, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370340168476105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05221806466579437, + 0.0601358637213707, + 0.08744056522846222, + 0.025811711326241493, + 0.7018424868583679, + 0.011235198006033897, + 0.023228826001286507, + 0.016490140929818153, + 0.00412304000928998, + 0.017474129796028137 + ], + "credit_map": [ + 0.005887866951525211, + 0.004832737147808075, + 0.01503788586705923, + 0.004128963220864534, + 0.004940294194966555, + 0.004444896709173918, + 0.0037451081443578005, + 0.009200315922498703, + 0.004720934666693211, + 0.008306601084768772, + 0.005553977098315954, + 0.005968963727355003, + 0.005819293670356274, + 0.008232367224991322, + 0.010133347474038601, + 0.011400379240512848, + 0.009821576997637749, + 0.019053606316447258, + 0.008867173455655575, + 0.007500907871872187, + 0.01305476389825344, + 0.016351066529750824, + 0.005704736802726984, + 0.007086741738021374, + 0.002261137356981635, + 0.003962404560297728, + 0.005868997424840927, + 0.005941002629697323, + 0.0019682087004184723, + 0.001691598678007722, + 0.001657669898122549, + 0.0024606906808912754, + 0.0677272379398346, + 0.044520869851112366, + 0.23546673357486725, + 0.2571599781513214, + 0.01622369885444641, + 0.01946377195417881, + 0.04096505790948868, + 0.020315151661634445, + 0.0011783834779635072, + 0.001274257549084723, + 0.0013861856423318386, + 0.0017084976425394416, + 0.0009337667725048959, + 0.0013065115781500936, + 0.001625060336664319, + 0.0018225355306640267, + 0.003578518284484744, + 0.003272231435403228, + 0.002917318372055888, + 0.002988688414916396, + 0.003110098885372281, + 0.0023512488696724176, + 0.0031881132163107395, + 0.0018226081738248467, + 0.002930260030552745, + 0.0022109795827418566, + 0.0040200925432145596, + 0.0022425809875130653, + 0.0011613323586061597, + 0.0009831404313445091, + 0.0020788274705410004, + 0.0008629276417195797, + 0.0004737992421723902, + 0.0006742087425664067, + 0.0005575435934588313, + 0.0005690817488357425, + 0.0003470178344286978, + 0.0004751347005367279, + 0.0005938917747698724, + 0.0004323624016251415, + 0.0016297224210575223, + 0.0018410920165479183, + 0.001978693762794137, + 0.002551811281591654, + 0.0018235845491290092, + 0.0022048410028219223, + 0.0027115456759929657, + 0.0027328412979841232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.03370340168476105 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870340585708618 + ], + "opposite_logprob_tokens": [ + -1.1370340585708618 + ], + "logprob_reward": -0.3870340585708618, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870340585708618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0819118544459343, + 0.09532935172319412, + 0.11340125650167465, + 0.3275338113307953, + 0.2662658989429474, + 0.02395092323422432, + 0.026051191613078117, + 0.033576104789972305, + 0.010892173275351524, + 0.021087436005473137 + ], + "credit_map": [ + 0.007566952612251043, + 0.00847592018544674, + 0.018978141248226166, + 0.004632878582924604, + 0.008909894153475761, + 0.0076485611498355865, + 0.009012837894260883, + 0.016686664894223213, + 0.008319802582263947, + 0.010917925275862217, + 0.01022508554160595, + 0.009454282000660896, + 0.010505844838917255, + 0.01959838531911373, + 0.013464546762406826, + 0.012843475677073002, + 0.012546513229608536, + 0.015493364073336124, + 0.010334459133446217, + 0.008969428017735481, + 0.013621348887681961, + 0.027979319915175438, + 0.012027976103127003, + 0.012428844347596169, + 0.0404023714363575, + 0.08715376257896423, + 0.04699225351214409, + 0.08925395458936691, + 0.02669670060276985, + 0.007064369507133961, + 0.020387284457683563, + 0.009583109989762306, + 0.03028024174273014, + 0.01907147653400898, + 0.0855947881937027, + 0.0956893190741539, + 0.005203560926020145, + 0.007828651927411556, + 0.01532434206455946, + 0.007273528259247541, + 0.003256121650338173, + 0.0037310621701180935, + 0.003730107331648469, + 0.002768471371382475, + 0.0025093224830925465, + 0.0023626061156392097, + 0.00280576990917325, + 0.002787461271509528, + 0.0032674171961843967, + 0.003340214490890503, + 0.003429755801334977, + 0.0037826120387762785, + 0.0029345310758799314, + 0.0031067044474184513, + 0.003750754753127694, + 0.002439204603433609, + 0.0077999671921133995, + 0.0033574746921658516, + 0.010770435445010662, + 0.002805974567309022, + 0.0014245407655835152, + 0.0013794327387586236, + 0.004331602714955807, + 0.0017066734144464135, + 0.0011883764527738094, + 0.0019314411329105496, + 0.001053303712978959, + 0.0014688237570226192, + 0.001582170370966196, + 0.0011036645155400038, + 0.0013414081186056137, + 0.001222986145876348, + 0.002088886918500066, + 0.002453417982906103, + 0.0023527878802269697, + 0.0017375241732224822, + 0.001861188211478293, + 0.0035845539532601833, + 0.003581283614039421, + 0.0034277935046702623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.3870340585708618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly pans over a U-shaped wooden chisel with a polished, sleek handle. Steadily steered by a pair of hands, marked by age but steady, the chisel's sharp metal blade slices precisely into the finely grained surface of light-colored oak. As the blade's steady and deliberate advance unfolds, a thin, curling shavings are elegantly detached from the wood, forming a perfect spiral in mid-air before cascading down, feather-light, to a workbench strewn with similar sawdust. The scene is bathed in a warm, vintage hue. Light comes from a vintage tungsten lamp suspended above the workbench, casting a concentrated beam with a distinct warm yellow hue. This light not only highlights the three-dimensionality of the sawdust and the delicate grain of the wood, but also casts a soft glow on the polished handle. In the background, a slightly blurred woodworking workshop appears, with faint glimpses of other handmade tools hanging on the wall, creating an atmosphere of focus, tranquility, and craftsmanship.", + "seed": 5572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_13/index_7606_seed_5572_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20933796130120755, + "vlm_logprob": -0.20933796130120755, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.20933796130120755, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026479944586753845 + ], + "opposite_logprob_tokens": [ + -5.776479721069336 + ], + "logprob_reward": -0.026479944586753845, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.026479944586753845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3027041554450989, + 0.07884617149829865, + 0.17220771312713623, + 0.02549969218671322, + 0.042606838047504425, + 0.140287384390831, + 0.09687820076942444, + 0.03740646690130234, + 0.03766665235161781, + 0.0658966675400734 + ], + "credit_map": [ + 0.01871255226433277, + 0.025804273784160614, + 0.01527534518390894, + 0.027153747156262398, + 0.020037822425365448, + 0.09427446126937866, + 0.017611276358366013, + 0.08383467048406601, + 0.007831472903490067, + 0.009288385510444641, + 0.010566461831331253, + 0.008490574546158314, + 0.009185213595628738, + 0.01200959924608469, + 0.010769248008728027, + 0.010705222375690937, + 0.005369166377931833, + 0.006541743408888578, + 0.006210066843777895, + 0.005528038367629051, + 0.00831648986786604, + 0.0726039931178093, + 0.008575709536671638, + 0.05906250700354576, + 0.002544835675507784, + 0.002722534816712141, + 0.0034786954056471586, + 0.0025220676325261593, + 0.002528525423258543, + 0.003957694862037897, + 0.004120246507227421, + 0.003625091165304184, + 0.0039660087786614895, + 0.008681402541697025, + 0.0034777666442096233, + 0.0050743697211146355, + 0.0030475985258817673, + 0.007039512507617474, + 0.004738034680485725, + 0.006582146510481834, + 0.009401698596775532, + 0.04368845745921135, + 0.011146370321512222, + 0.014033361338078976, + 0.012001174502074718, + 0.018310321494936943, + 0.012587686069309711, + 0.019118310883641243, + 0.012967856600880623, + 0.011315586045384407, + 0.010925373062491417, + 0.014938841573894024, + 0.005555399227887392, + 0.01234496757388115, + 0.01882280595600605, + 0.010007370263338089, + 0.005000169388949871, + 0.005124887451529503, + 0.006034440360963345, + 0.004646742716431618, + 0.0033899694681167603, + 0.004814034793525934, + 0.004474798683077097, + 0.0039214277639985085, + 0.006067623384296894, + 0.0036399608943611383, + 0.004910219460725784, + 0.0056811366230249405, + 0.002823658986017108, + 0.0038995901122689247, + 0.005305672064423561, + 0.0053387912921607494, + 0.010654211975634098, + 0.00634771166369319, + 0.006370737683027983, + 0.012535346671938896, + 0.00550235016271472, + 0.006255731452256441, + 0.009830658324062824, + 0.008399919606745243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.026479944586753845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870723560452461 + ], + "opposite_logprob_tokens": [ + -3.0487072467803955 + ], + "logprob_reward": -0.04870723560452461, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870723560452461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15479251742362976, + 0.09266118705272675, + 0.20263436436653137, + 0.030770104378461838, + 0.1209871768951416, + 0.0709005668759346, + 0.13355925679206848, + 0.09761746227741241, + 0.031494658440351486, + 0.06458272784948349 + ], + "credit_map": [ + 0.017312591895461082, + 0.012225309386849403, + 0.02436079829931259, + 0.021515144035220146, + 0.01858442835509777, + 0.019570399075746536, + 0.021130001172423363, + 0.02009386196732521, + 0.007774734869599342, + 0.006206209305673838, + 0.013660930097103119, + 0.01244833692908287, + 0.00881069153547287, + 0.018283354118466377, + 0.013713777996599674, + 0.01176315825432539, + 0.007404741365462542, + 0.00809723325073719, + 0.008618601597845554, + 0.010233602486550808, + 0.01006362121552229, + 0.08125633746385574, + 0.014441488310694695, + 0.06251874566078186, + 0.0024269174318760633, + 0.0023111170157790184, + 0.004941013641655445, + 0.003933367319405079, + 0.0025036686565726995, + 0.004925519227981567, + 0.005675121210515499, + 0.004053380340337753, + 0.0116573516279459, + 0.030854491516947746, + 0.012134496122598648, + 0.010327012278139591, + 0.012177979573607445, + 0.017649183049798012, + 0.010440800338983536, + 0.015745859593153, + 0.0055230665020644665, + 0.008247710764408112, + 0.006165339611470699, + 0.008610749617218971, + 0.005072630010545254, + 0.007907071150839329, + 0.019199101254343987, + 0.010174902155995369, + 0.017952660098671913, + 0.016353929415345192, + 0.015470054000616074, + 0.020849689841270447, + 0.008854235522449017, + 0.0141171645373106, + 0.02514784224331379, + 0.01481368113309145, + 0.011267878115177155, + 0.01046282984316349, + 0.015807323157787323, + 0.017776599153876305, + 0.006399484351277351, + 0.013540094718337059, + 0.009959462098777294, + 0.012403789907693863, + 0.005940503440797329, + 0.0023935833014547825, + 0.005858709570020437, + 0.0052201715297997, + 0.0018157007871195674, + 0.0023597388062626123, + 0.003790058195590973, + 0.004116191063076258, + 0.01178405899554491, + 0.006779599469155073, + 0.007026601582765579, + 0.013358167372643948, + 0.004909668117761612, + 0.005601025652140379, + 0.008269239217042923, + 0.006854362320154905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.04870723560452461 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495508760213852 + ], + "opposite_logprob_tokens": [ + -2.929955005645752 + ], + "logprob_reward": -0.05495508760213852, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495508760213852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19243071973323822, + 0.1569546014070511, + 0.26166462898254395, + 0.06407851725816727, + 0.0536123663187027, + 0.09965607523918152, + 0.06628185510635376, + 0.034466393291950226, + 0.04236504063010216, + 0.02848980389535427 + ], + "credit_map": [ + 0.012303000316023827, + 0.016081582754850388, + 0.018263088539242744, + 0.017380261793732643, + 0.013641123659908772, + 0.045428693294525146, + 0.02291669324040413, + 0.0464162677526474, + 0.008533806540071964, + 0.011717919260263443, + 0.014575883746147156, + 0.017576195299625397, + 0.010030853562057018, + 0.040413495153188705, + 0.021425530314445496, + 0.0326809287071228, + 0.011544195003807545, + 0.009376198053359985, + 0.007252848707139492, + 0.008977035991847515, + 0.012210052460432053, + 0.11126409471035004, + 0.012089605443179607, + 0.08895058929920197, + 0.0059648738242685795, + 0.004956848453730345, + 0.007392039056867361, + 0.0075246659107506275, + 0.006269155535846949, + 0.011946349404752254, + 0.01208312064409256, + 0.007941470481455326, + 0.004220497794449329, + 0.0057243169285357, + 0.008900536224246025, + 0.004974284674972296, + 0.004225707147270441, + 0.01049056090414524, + 0.007016345392912626, + 0.008060112595558167, + 0.008155398070812225, + 0.024440210312604904, + 0.01010642759501934, + 0.01041778177022934, + 0.009400907903909683, + 0.01059679128229618, + 0.013116154819726944, + 0.013422396965324879, + 0.0051261805929243565, + 0.009100627154111862, + 0.007151448633521795, + 0.009263763204216957, + 0.004423825070261955, + 0.010163193568587303, + 0.011758005246520042, + 0.009294814430177212, + 0.0030901285354048014, + 0.004031851422041655, + 0.004904292058199644, + 0.005245123989880085, + 0.0036848029121756554, + 0.004932626150548458, + 0.004586522001773119, + 0.003991046454757452, + 0.009070200845599174, + 0.004171271808445454, + 0.005895854905247688, + 0.005785481072962284, + 0.0029005093965679407, + 0.005587002262473106, + 0.0038072518073022366, + 0.005147471558302641, + 0.0033970691729336977, + 0.0033482469152659178, + 0.0037448168732225895, + 0.006301713176071644, + 0.0018132711993530393, + 0.002054758369922638, + 0.0048959567211568356, + 0.0029339692555367947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.05495508760213852 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198074772953987 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018198074772953987, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018198074772953987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2506117820739746, + 0.133911594748497, + 0.18747536838054657, + 0.019995955750346184, + 0.07357846945524216, + 0.16306893527507782, + 0.022155562415719032, + 0.06766494363546371, + 0.03890350088477135, + 0.042633917182683945 + ], + "credit_map": [ + 0.014795489609241486, + 0.011673917062580585, + 0.01409930270165205, + 0.017960350960493088, + 0.021190356463193893, + 0.08412496745586395, + 0.01291299145668745, + 0.07385438680648804, + 0.008730927482247353, + 0.011058772914111614, + 0.014545769430696964, + 0.01295086182653904, + 0.011557450518012047, + 0.03819281980395317, + 0.00783313438296318, + 0.02904186211526394, + 0.006479599047452211, + 0.007081529125571251, + 0.00522495387122035, + 0.006328742019832134, + 0.008794249035418034, + 0.08222775906324387, + 0.0044281138107180595, + 0.06691041588783264, + 0.0017920441459864378, + 0.0019632347393780947, + 0.002800846705213189, + 0.0029337480664253235, + 0.0016827236395329237, + 0.0025323217269033194, + 0.003593619680032134, + 0.0026974172797054052, + 0.006830418016761541, + 0.0189918614923954, + 0.0054459040984511375, + 0.00726531445980072, + 0.00804015714675188, + 0.010884258896112442, + 0.005607684142887592, + 0.010512875393033028, + 0.011498622596263885, + 0.04729432612657547, + 0.011701022274792194, + 0.015646150335669518, + 0.016422340646386147, + 0.020850136876106262, + 0.017816128209233284, + 0.02184019796550274, + 0.0025870520621538162, + 0.0024875253438949585, + 0.0029726254288107157, + 0.0030289904680103064, + 0.0013813310069963336, + 0.0028531684074550867, + 0.004441419150680304, + 0.002403451129794121, + 0.009952711872756481, + 0.007256773300468922, + 0.011101821437478065, + 0.01141100749373436, + 0.0043613300658762455, + 0.008477630093693733, + 0.007396116852760315, + 0.0077075534500181675, + 0.006860761903226376, + 0.004138554912060499, + 0.0049340371042490005, + 0.005533934570848942, + 0.0025727127213031054, + 0.004247345961630344, + 0.005357522517442703, + 0.005258628167212009, + 0.005925897508859634, + 0.006486327853053808, + 0.005537659395486116, + 0.009198217652738094, + 0.0028227977454662323, + 0.003958513028919697, + 0.004789206199347973, + 0.003915296867489815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.018198074772953987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983494639396667 + ], + "opposite_logprob_tokens": [ + -0.5233494639396667 + ], + "logprob_reward": -0.8983494639396667, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983494639396667, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2062569409608841, + 0.11028604209423065, + 0.10139262676239014, + 0.03532499447464943, + 0.07266846299171448, + 0.25149786472320557, + 0.03561517223715782, + 0.060503385961055756, + 0.07759733498096466, + 0.04885720834136009 + ], + "credit_map": [ + 0.01414563786238432, + 0.016566315665841103, + 0.023768188431859016, + 0.010743265971541405, + 0.013462549075484276, + 0.06427807360887527, + 0.008753528818488121, + 0.05453937128186226, + 0.00969520304352045, + 0.012764225713908672, + 0.013876860029995441, + 0.014648432843387127, + 0.011410965584218502, + 0.021865354850888252, + 0.009909112937748432, + 0.016115887090563774, + 0.004924448672682047, + 0.0036230292171239853, + 0.0032498168293386698, + 0.004749410320073366, + 0.005052903667092323, + 0.041859425604343414, + 0.004224835429340601, + 0.033708758652210236, + 0.002926323562860489, + 0.002718579489737749, + 0.004546366631984711, + 0.004652722738683224, + 0.002922418527305126, + 0.006097205448895693, + 0.006632259581238031, + 0.004829120822250843, + 0.005889309104532003, + 0.015384393744170666, + 0.005922860000282526, + 0.008975034579634666, + 0.007607190404087305, + 0.011411808431148529, + 0.006609464529901743, + 0.010868402197957039, + 0.021255385130643845, + 0.04171440750360489, + 0.026377571746706963, + 0.030844498425722122, + 0.024569638073444366, + 0.03175554424524307, + 0.04055475443601608, + 0.034426070749759674, + 0.002783730858936906, + 0.004254012368619442, + 0.003600920317694545, + 0.006001697853207588, + 0.002761001465842128, + 0.005403958261013031, + 0.006521861534565687, + 0.004287991672754288, + 0.005846114829182625, + 0.008269098587334156, + 0.0074266414158046246, + 0.009515427984297276, + 0.004290703218430281, + 0.005286407191306353, + 0.013421861454844475, + 0.006447129417210817, + 0.013377190567553043, + 0.009030248038470745, + 0.010882365517318249, + 0.011447922326624393, + 0.004941507242619991, + 0.00922421645373106, + 0.00910729169845581, + 0.009586594998836517, + 0.005768774542957544, + 0.007531717419624329, + 0.00614975905045867, + 0.011523609049618244, + 0.0030391484033316374, + 0.00479547493159771, + 0.00505432253703475, + 0.004994404036551714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.8983494639396667 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, the camera slowly pans over a U-shaped wooden chisel with a polished, sleek handle. Steadily steered by a pair of hands, marked by age but steady, the chisel's sharp metal blade slices precisely into the finely grained surface of light-colored oak. As the blade's steady and deliberate advance unfolds, a thin, curling shavings are elegantly detached from the wood, forming a perfect spiral in mid-air before cascading down, feather-light, to a workbench strewn with similar sawdust. The scene is bathed in a warm, vintage hue. Light comes from a vintage tungsten lamp suspended above the workbench, casting a concentrated beam with a distinct warm yellow hue. This light not only highlights the three-dimensionality of the sawdust and the delicate grain of the wood, but also casts a soft glow on the polished handle. In the background, a slightly blurred woodworking workshop appears, with faint glimpses of other handmade tools hanging on the wall, creating an atmosphere of focus, tranquility, and craftsmanship.", + "seed": 5573, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_13/index_7606_seed_5573_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14868858382105826, + "vlm_logprob": -0.14868858382105826, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.14868858382105826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03395786136388779 + ], + "opposite_logprob_tokens": [ + -4.9089579582214355 + ], + "logprob_reward": -0.03395786136388779, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.03395786136388779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12777955830097198, + 0.1264859139919281, + 0.06600875407457352, + 0.1365373432636261, + 0.07957922667264938, + 0.09908313304185867, + 0.10320678353309631, + 0.07332193106412888, + 0.10170648992061615, + 0.08629090338945389 + ], + "credit_map": [ + 0.02392859384417534, + 0.008438427932560444, + 0.017771001905202866, + 0.02100416272878647, + 0.010316978208720684, + 0.018037952482700348, + 0.011475983075797558, + 0.016806451603770256, + 0.01845097914338112, + 0.010760674253106117, + 0.01819463074207306, + 0.01170438528060913, + 0.007164257112890482, + 0.023992935195565224, + 0.013941113837063313, + 0.02227691002190113, + 0.009197494946420193, + 0.004611065611243248, + 0.008036010898649693, + 0.007226570975035429, + 0.00835632998496294, + 0.006783933844417334, + 0.01065672654658556, + 0.011140614748001099, + 0.0325453095138073, + 0.012363585643470287, + 0.009523240849375725, + 0.005175463855266571, + 0.009423657320439816, + 0.014151881448924541, + 0.008770542219281197, + 0.044583648443222046, + 0.013059207238256931, + 0.009677115827798843, + 0.011089597828686237, + 0.004242271650582552, + 0.014532451517879963, + 0.007006863597780466, + 0.014025097712874413, + 0.005946613382548094, + 0.012441038154065609, + 0.014612765051424503, + 0.015520193614065647, + 0.011778715997934341, + 0.00740272318944335, + 0.010913604870438576, + 0.010443951934576035, + 0.015970127657055855, + 0.01703125424683094, + 0.032980818301439285, + 0.00920812226831913, + 0.007116269785910845, + 0.006028421688824892, + 0.01025352906435728, + 0.00924901943653822, + 0.011339343152940273, + 0.0104170897975564, + 0.013346642255783081, + 0.006719803437590599, + 0.007942205294966698, + 0.005791381932795048, + 0.014130032621324062, + 0.006211556028574705, + 0.008763214573264122, + 0.012912344187498093, + 0.015088952146470547, + 0.011806190013885498, + 0.013057528994977474, + 0.010147206485271454, + 0.0119776064530015, + 0.007256398908793926, + 0.019460253417491913, + 0.013123568147420883, + 0.020440567284822464, + 0.010183035396039486, + 0.013171805068850517, + 0.0050412798300385475, + 0.01320420578122139, + 0.0051171984523534775, + 0.006009229458868504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.03395786136388779 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07905751466751099 + ], + "opposite_logprob_tokens": [ + -2.579057455062866 + ], + "logprob_reward": -0.07905751466751099, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905751466751099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07797159999608994, + 0.03091336041688919, + 0.05847838521003723, + 0.07186336070299149, + 0.3117867112159729, + 0.029388466849923134, + 0.0560283288359642, + 0.20213693380355835, + 0.021869394928216934, + 0.13956347107887268 + ], + "credit_map": [ + 0.013184884563088417, + 0.0067422883585095406, + 0.009485922753810883, + 0.010741199366748333, + 0.005417880602180958, + 0.007068163249641657, + 0.012790787033736706, + 0.01254047080874443, + 0.005322245415300131, + 0.0032776815351098776, + 0.0038230076897889376, + 0.0033668403048068285, + 0.0026774813886731863, + 0.006713666953146458, + 0.002692891051992774, + 0.003039546078070998, + 0.010520543903112411, + 0.0029018742498010397, + 0.0059602162800729275, + 0.005423988215625286, + 0.0071241408586502075, + 0.004948620218783617, + 0.005814910400658846, + 0.015784092247486115, + 0.016341378912329674, + 0.0036158463917672634, + 0.004652057308703661, + 0.0030981386080384254, + 0.006124361418187618, + 0.007573353126645088, + 0.004828698933124542, + 0.025629522278904915, + 0.05511261522769928, + 0.03140885382890701, + 0.01676531322300434, + 0.008313985541462898, + 0.11531557142734528, + 0.056791264563798904, + 0.013941654935479164, + 0.014137450605630875, + 0.00397133594378829, + 0.0030098760034888983, + 0.003143146401271224, + 0.0026075048372149467, + 0.002305507892742753, + 0.006196840200573206, + 0.00295437965542078, + 0.005199875216931105, + 0.006697258912026882, + 0.01221806276589632, + 0.006258130539208651, + 0.00739900441840291, + 0.0030314424075186253, + 0.005644645541906357, + 0.005966043099761009, + 0.008813740685582161, + 0.04577488452196121, + 0.029745014384388924, + 0.010167363099753857, + 0.013104833662509918, + 0.015265418216586113, + 0.06903979927301407, + 0.007852854207158089, + 0.011186767369508743, + 0.002560541732236743, + 0.0032577570527791977, + 0.0029358980245888233, + 0.004422700498253107, + 0.0014321071794256568, + 0.0028969584964215755, + 0.0018100306624546647, + 0.002553400583565235, + 0.03747662529349327, + 0.05740368738770485, + 0.005159228108823299, + 0.005043333396315575, + 0.006633426994085312, + 0.02100166119635105, + 0.0034842928871512413, + 0.00336122396402061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.07905751466751099 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370559215545654 + ], + "opposite_logprob_tokens": [ + -3.408705711364746 + ], + "logprob_reward": -0.03370559215545654, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370559215545654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2000298649072647, + 0.04015841707587242, + 0.05787275731563568, + 0.05554879084229469, + 0.2070663869380951, + 0.05176667496562004, + 0.19896356761455536, + 0.05724266171455383, + 0.09959474951028824, + 0.03175617381930351 + ], + "credit_map": [ + 0.01816573739051819, + 0.01295690517872572, + 0.015894195064902306, + 0.03318709135055542, + 0.016530925408005714, + 0.032340992242097855, + 0.03313435986638069, + 0.03781965747475624, + 0.00514173274859786, + 0.003951664548367262, + 0.005753362085670233, + 0.006259183399379253, + 0.002937629586085677, + 0.005584755912423134, + 0.005868915468454361, + 0.004661175422370434, + 0.012961646541953087, + 0.002777541521936655, + 0.00432283291593194, + 0.0033746229019016027, + 0.006083471700549126, + 0.004274775739759207, + 0.008207726292312145, + 0.015870139002799988, + 0.006298783700913191, + 0.0037629546131938696, + 0.0049210223369300365, + 0.005740959197282791, + 0.009511128067970276, + 0.006625862792134285, + 0.0064337546937167645, + 0.012254326604306698, + 0.0416119284927845, + 0.0192128773778677, + 0.008593020960688591, + 0.005028499290347099, + 0.07871178537607193, + 0.03526410460472107, + 0.0096796415746212, + 0.008964527398347855, + 0.0069373189471662045, + 0.007779646664857864, + 0.006249884609133005, + 0.007678118534386158, + 0.0036444447468966246, + 0.007635793648660183, + 0.006420080084353685, + 0.005421388894319534, + 0.03071865253150463, + 0.049579739570617676, + 0.025462348014116287, + 0.025847818702459335, + 0.013809258118271828, + 0.016205323860049248, + 0.017189176753163338, + 0.020151246339082718, + 0.010016903281211853, + 0.006826465949416161, + 0.004978008568286896, + 0.005145909730345011, + 0.004457367118448019, + 0.013463260605931282, + 0.006126577500253916, + 0.00622817175462842, + 0.013597236014902592, + 0.009830794297158718, + 0.00865224190056324, + 0.011300729587674141, + 0.007473930716514587, + 0.02064884826540947, + 0.0073828850872814655, + 0.020708082243800163, + 0.0045950342901051044, + 0.00415476318448782, + 0.0036432931665331125, + 0.0037129458505660295, + 0.002206259174272418, + 0.006259719841182232, + 0.002777577843517065, + 0.004406583495438099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.03370559215545654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020593050867319107 + ], + "opposite_logprob_tokens": [ + -3.8955931663513184 + ], + "logprob_reward": -0.020593050867319107, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020593050867319107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042444709688425064, + 0.03630002588033676, + 0.072885662317276, + 0.28624093532562256, + 0.2067178636789322, + 0.09129881858825684, + 0.013732237741351128, + 0.12482620030641556, + 0.018243912607431412, + 0.10730967670679092 + ], + "credit_map": [ + 0.007947484962642193, + 0.003684750758111477, + 0.005092066712677479, + 0.0031787094194442034, + 0.00515578780323267, + 0.008408188819885254, + 0.0038104099221527576, + 0.005167313851416111, + 0.0038974357303231955, + 0.002462704200297594, + 0.005079076159745455, + 0.009109975770115852, + 0.00234229420311749, + 0.0034828137140721083, + 0.004938055761158466, + 0.004987672436982393, + 0.010841385461390018, + 0.005397280212491751, + 0.009345123544335365, + 0.0061124698258936405, + 0.005646064877510071, + 0.010870635509490967, + 0.008060331456363201, + 0.016612373292446136, + 0.07040545344352722, + 0.010772851295769215, + 0.013364629819989204, + 0.007829345762729645, + 0.02355610579252243, + 0.02939562126994133, + 0.012735848315060139, + 0.11818108707666397, + 0.032274626195430756, + 0.015088253654539585, + 0.009117579087615013, + 0.005093642510473728, + 0.07607565075159073, + 0.04086090251803398, + 0.016317378729581833, + 0.011889825575053692, + 0.013106961734592915, + 0.006901558022946119, + 0.006743319798260927, + 0.005041991360485554, + 0.007708759978413582, + 0.018377410247921944, + 0.01086898148059845, + 0.02254982851445675, + 0.0015705482801422477, + 0.0025618914514780045, + 0.0014381034998223186, + 0.001842276193201542, + 0.0012954487465322018, + 0.001311768894083798, + 0.0018417083192616701, + 0.0018704929389059544, + 0.03282845392823219, + 0.011039930395781994, + 0.004760055802762508, + 0.003939400892704725, + 0.008394546806812286, + 0.055063776671886444, + 0.0032231181394308805, + 0.00557691091671586, + 0.002386565553024411, + 0.002432408044114709, + 0.0024467625189572573, + 0.0039641642943024635, + 0.0012741262326017022, + 0.0017426151316612959, + 0.0017212929669767618, + 0.0022759782150387764, + 0.03024863637983799, + 0.04601958394050598, + 0.0029005641117691994, + 0.0034884491469711065, + 0.004781495314091444, + 0.015255008824169636, + 0.0021350665483623743, + 0.0024808698799461126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.020593050867319107 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a U-shaped wooden chisel with a polished, sleek handle being steered by a pair of aged hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chisel's metal blade slice into light-colored oak, creating thin, curling shavings that form a spiral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a vintage tungsten lamp casting a warm yellow beam and a warm, vintage hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan over the scene while the background shows a blurred woodworking workshop with tools on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761289000511169 + ], + "opposite_logprob_tokens": [ + -0.8261289000511169 + ], + "logprob_reward": -0.5761289000511169, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761289000511169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10972560197114944, + 0.05463118851184845, + 0.026851139962673187, + 0.2434496134519577, + 0.0842495858669281, + 0.023304292932152748, + 0.027621600776910782, + 0.2195669412612915, + 0.10689456760883331, + 0.10370546579360962 + ], + "credit_map": [ + 0.015806131064891815, + 0.012846932746469975, + 0.0077186524868011475, + 0.010766143910586834, + 0.0170594435185194, + 0.020697467029094696, + 0.014017528854310513, + 0.010813304223120213, + 0.00702808378264308, + 0.004885091912001371, + 0.00839153490960598, + 0.006050093099474907, + 0.0059546297416090965, + 0.010410476475954056, + 0.006155550479888916, + 0.005755734629929066, + 0.005207634065300226, + 0.0018144063651561737, + 0.002948251785710454, + 0.0024233332369476557, + 0.0021983799524605274, + 0.0023868398275226355, + 0.0035187164321541786, + 0.006353581324219704, + 0.05809076130390167, + 0.009351016953587532, + 0.010781937278807163, + 0.008019594475626945, + 0.022201061248779297, + 0.022163167595863342, + 0.011922113597393036, + 0.10091998428106308, + 0.01579011045396328, + 0.009946204721927643, + 0.004172471351921558, + 0.002099568722769618, + 0.030721081420779228, + 0.014008786529302597, + 0.003893044078722596, + 0.0036183258052915335, + 0.003328830935060978, + 0.0038714101538062096, + 0.004271888639777899, + 0.0022356323897838593, + 0.001541503588669002, + 0.003136447397992015, + 0.0025247950106859207, + 0.002393786795437336, + 0.003526644315570593, + 0.007453157100826502, + 0.0028689869213849306, + 0.0040078600868582726, + 0.0019938750192523003, + 0.003172993427142501, + 0.002138308482244611, + 0.0024597784504294395, + 0.06021536886692047, + 0.020734624937176704, + 0.007294524926692247, + 0.0056950394064188, + 0.014573239721357822, + 0.09664515405893326, + 0.0051435502246022224, + 0.009265436790883541, + 0.014239583164453506, + 0.011868449859321117, + 0.01464440394192934, + 0.018730752170085907, + 0.008385122753679752, + 0.01909778267145157, + 0.008548437617719173, + 0.011380046606063843, + 0.029924217611551285, + 0.04268642142415047, + 0.0032859828788787127, + 0.003641705960035324, + 0.004656709264963865, + 0.013918810524046421, + 0.0024467185139656067, + 0.003144902642816305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.5761289000511169 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_14/index_tensor([895, 895])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_14/index_tensor([895, 895])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..258ad9a834e76cb8d9dbaea8430e720a4cfd5266 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_14/index_tensor([895, 895])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up, the camera remains still, focused on the deeply wrinkled face of an elderly monk. A ray of warm afternoon sunlight streams in from the side, sculpting his features with precise precision, casting bright highlights on his nose, cheekbones, and forehead, while the other side of his face is immersed in soft shadow, revealing every wrinkle of his skin, like the ridges of time. Even the finest pores and freckles of time are visible on his bronzed complexion. His shiny head reflects the soft light, and he wears a faded ochre-colored, coarse cotton and linen robe. His expression is serene and solemn. His eyelids droop slowly, almost imperceptibly, casting a faint shadow in the light before closing completely, shutting out the outside world, as if he has entered a deep meditative state. As his eyes close, the muscles on his face relax slightly, and his entire face conveys a sense of ultimate peace and detached wisdom, creating an atmosphere of tranquility, dignity, and Zen.", + "seed": 1146, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_14/index_895_seed_1146_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18872277364134787, + "vlm_logprob": -0.18872277364134787, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.18872277364134787, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005032965913414955 + ], + "opposite_logprob_tokens": [ + -6.130033016204834 + ], + "logprob_reward": -0.005032965913414955, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.005032965913414955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19514207541942596, + 0.049486082047224045, + 0.044584646821022034, + 0.09762436151504517, + 0.062479741871356964, + 0.07713943719863892, + 0.04578576609492302, + 0.1364002823829651, + 0.18280182778835297, + 0.10855579376220703 + ], + "credit_map": [ + 0.022548140957951546, + 0.02735796570777893, + 0.01398695819079876, + 0.03054889105260372, + 0.023935340344905853, + 0.030110346153378487, + 0.018786607310175896, + 0.027867812663316727, + 0.007651825435459614, + 0.005870639346539974, + 0.004507394041866064, + 0.007763972040265799, + 0.006432313472032547, + 0.006015397608280182, + 0.006162688136100769, + 0.005081850104033947, + 0.0056357914581894875, + 0.006602749228477478, + 0.004091065842658281, + 0.005277628544718027, + 0.005004505626857281, + 0.006563154980540276, + 0.005444355309009552, + 0.005965396761894226, + 0.016129741445183754, + 0.012572362087666988, + 0.008579676039516926, + 0.011940778233110905, + 0.008912239223718643, + 0.01579747535288334, + 0.0077672419138252735, + 0.015924841165542603, + 0.00791096594184637, + 0.01031380146741867, + 0.005358789581805468, + 0.011243744753301144, + 0.008299902081489563, + 0.006370245944708586, + 0.0051560490392148495, + 0.007826242595911026, + 0.0095442496240139, + 0.012342485599219799, + 0.007677457295358181, + 0.012871896848082542, + 0.0075927129946649075, + 0.009492287412285805, + 0.0059891799464821815, + 0.011629165150225163, + 0.008367616683244705, + 0.005493965931236744, + 0.005740431137382984, + 0.007712849881500006, + 0.0030641767662018538, + 0.006008832715451717, + 0.0034541466739028692, + 0.005943745374679565, + 0.016151675954461098, + 0.011068849824368954, + 0.005925378296524286, + 0.009282305836677551, + 0.009282357059419155, + 0.039810944348573685, + 0.006078390870243311, + 0.03880038857460022, + 0.021194541826844215, + 0.016047772020101547, + 0.010099702514708042, + 0.019854949787259102, + 0.011548787355422974, + 0.046895429491996765, + 0.011042838916182518, + 0.04611779749393463, + 0.006826306693255901, + 0.008770394138991833, + 0.005095453932881355, + 0.00796806626021862, + 0.006248359568417072, + 0.03485969454050064, + 0.004935288801789284, + 0.0338522270321846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.005032965913414955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934298872947693 + ], + "opposite_logprob_tokens": [ + -0.6934298872947693 + ], + "logprob_reward": -0.6934298872947693, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934298872947693, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08412390202283859, + 0.049067966639995575, + 0.022788366302847862, + 0.051834940910339355, + 0.14272619783878326, + 0.0814494714140892, + 0.04694280028343201, + 0.03511040657758713, + 0.07878872007131577, + 0.4071672856807709 + ], + "credit_map": [ + 0.011874804273247719, + 0.009195619262754917, + 0.005838301498442888, + 0.010515783913433552, + 0.007883081212639809, + 0.016297807916998863, + 0.006483843550086021, + 0.01603465899825096, + 0.005094020161777735, + 0.005368495360016823, + 0.0033633485436439514, + 0.00554678775370121, + 0.005232139956206083, + 0.009375667199492455, + 0.005286253523081541, + 0.00980125367641449, + 0.00450145686045289, + 0.002596876583993435, + 0.0017957512754946947, + 0.0029066624119877815, + 0.002358446130529046, + 0.003310202853754163, + 0.001964109018445015, + 0.0033548614010214806, + 0.009238651022315025, + 0.007075052242726088, + 0.00583691569045186, + 0.006931503769010305, + 0.005255755968391895, + 0.006630283314734697, + 0.004167411010712385, + 0.006699366495013237, + 0.013799158856272697, + 0.01580485701560974, + 0.013161005452275276, + 0.03221749886870384, + 0.01221806462854147, + 0.020363405346870422, + 0.014453532174229622, + 0.02070866897702217, + 0.01117025688290596, + 0.011476052924990654, + 0.0064612459391355515, + 0.016693949699401855, + 0.008702965453267097, + 0.009890181012451649, + 0.008553290739655495, + 0.008501531556248665, + 0.006475416477769613, + 0.004760967567563057, + 0.0045904554426670074, + 0.007092474959790707, + 0.005463053472340107, + 0.006361327599734068, + 0.005878381431102753, + 0.006320723332464695, + 0.004306606482714415, + 0.003579071257263422, + 0.002824329072609544, + 0.00410447409376502, + 0.0026523962151259184, + 0.007240565028041601, + 0.0030225729569792747, + 0.007380393333733082, + 0.008847098797559738, + 0.012486296705901623, + 0.0063363672234117985, + 0.007843374274671078, + 0.008040867745876312, + 0.01527214515954256, + 0.005364205688238144, + 0.01459836307913065, + 0.0427512526512146, + 0.028515346348285675, + 0.016242755576968193, + 0.02806132473051548, + 0.01609162986278534, + 0.13147567212581635, + 0.015609445050358772, + 0.12841986119747162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.6934298872947693 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049111589789390564 + ], + "opposite_logprob_tokens": [ + -3.0491116046905518 + ], + "logprob_reward": -0.049111589789390564, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.049111589789390564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1254756599664688, + 0.09765323251485825, + 0.09716581553220749, + 0.07099036127328873, + 0.07743891328573227, + 0.10574964433908463, + 0.05384716019034386, + 0.06998257339000702, + 0.08077877759933472, + 0.22091783583164215 + ], + "credit_map": [ + 0.014748015441000462, + 0.01897185482084751, + 0.01007059309631586, + 0.01915634050965309, + 0.016251342371106148, + 0.01688140071928501, + 0.013073374517261982, + 0.01632273942232132, + 0.009680005721747875, + 0.01011674851179123, + 0.009393803775310516, + 0.012185824103653431, + 0.007083437871187925, + 0.018497202545404434, + 0.012161457911133766, + 0.01853475719690323, + 0.01147956307977438, + 0.012323323637247086, + 0.008486227132380009, + 0.011919701471924782, + 0.010039222426712513, + 0.015688061714172363, + 0.01071922667324543, + 0.01651049219071865, + 0.00921997893601656, + 0.01103585958480835, + 0.007234930992126465, + 0.00854340847581625, + 0.008427082560956478, + 0.009797438979148865, + 0.006641901098191738, + 0.010089759714901447, + 0.008547341451048851, + 0.010713446885347366, + 0.006701208185404539, + 0.01539541408419609, + 0.008565046824514866, + 0.009280506521463394, + 0.007560001686215401, + 0.010675949044525623, + 0.012996247038245201, + 0.01826004683971405, + 0.009559666737914085, + 0.012843110598623753, + 0.00987517274916172, + 0.01619831472635269, + 0.008077516220510006, + 0.017939576879143715, + 0.00898957159370184, + 0.00598871847614646, + 0.00678671570494771, + 0.008594418875873089, + 0.0037126634269952774, + 0.0076354737393558025, + 0.004522870294749737, + 0.007616724818944931, + 0.009055602364242077, + 0.008308147080242634, + 0.00863777194172144, + 0.011711533181369305, + 0.0072958217933773994, + 0.009538679383695126, + 0.005089913960546255, + 0.010345110669732094, + 0.01011957973241806, + 0.010409743525087833, + 0.005199715029448271, + 0.010949166491627693, + 0.007876983843743801, + 0.014335159212350845, + 0.007645690347999334, + 0.014242740347981453, + 0.021725036203861237, + 0.014174028299748898, + 0.010604268871247768, + 0.014759366400539875, + 0.008013765327632427, + 0.07244067639112473, + 0.008222068659961224, + 0.07097864151000977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.049111589789390564 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160838533192873 + ], + "opposite_logprob_tokens": [ + -4.141083717346191 + ], + "logprob_reward": -0.0160838533192873, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160838533192873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10954374819993973, + 0.06839603185653687, + 0.057309217751026154, + 0.23329827189445496, + 0.18190431594848633, + 0.05303046852350235, + 0.03092256560921669, + 0.043574459850788116, + 0.07828139513731003, + 0.14373955130577087 + ], + "credit_map": [ + 0.014438164420425892, + 0.017062734812498093, + 0.011423466727137566, + 0.015571258030831814, + 0.012902668677270412, + 0.013588775880634785, + 0.010674824938178062, + 0.013881854712963104, + 0.012078131549060345, + 0.009458920918405056, + 0.005547648295760155, + 0.008553451858460903, + 0.008798778057098389, + 0.009174656122922897, + 0.00773383304476738, + 0.007050613407045603, + 0.007722389418631792, + 0.004994938150048256, + 0.004893386270850897, + 0.006195339374244213, + 0.004515782464295626, + 0.012304096482694149, + 0.005240066908299923, + 0.011443222872912884, + 0.02445496805012226, + 0.023009635508060455, + 0.024233901873230934, + 0.021948641166090965, + 0.019893625751137733, + 0.04502907022833824, + 0.022465214133262634, + 0.05226321145892143, + 0.021664736792445183, + 0.03180666267871857, + 0.016288014128804207, + 0.02436176873743534, + 0.021612301468849182, + 0.021823201328516006, + 0.015917832031846046, + 0.028429819270968437, + 0.0068271346390247345, + 0.00695376330986619, + 0.005436057224869728, + 0.00971645675599575, + 0.004551062826067209, + 0.007295400369912386, + 0.0036650151014328003, + 0.008585579693317413, + 0.0036282099317759275, + 0.0035360290203243494, + 0.004113186150789261, + 0.004658053629100323, + 0.0020926976576447487, + 0.005322709213942289, + 0.0024832256603986025, + 0.005088454112410545, + 0.004072810523211956, + 0.003485763445496559, + 0.0027148318476974964, + 0.0036430538166314363, + 0.0031397484708577394, + 0.012080805376172066, + 0.002617832040414214, + 0.011819613166153431, + 0.00948515348136425, + 0.008448666892945766, + 0.005945870187133551, + 0.00806912686675787, + 0.006248157005757093, + 0.018093537539243698, + 0.0043143825605511665, + 0.017676502466201782, + 0.015396383590996265, + 0.008909646421670914, + 0.010322163812816143, + 0.009708737954497337, + 0.006007860414683819, + 0.04363005608320236, + 0.006219995208084583, + 0.0435447171330452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.0160838533192873 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17995557188987732 + ], + "opposite_logprob_tokens": [ + -1.8049556016921997 + ], + "logprob_reward": -0.17995557188987732, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17995557188987732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10740048438310623, + 0.1017264872789383, + 0.10823690891265869, + 0.0710534006357193, + 0.08471710234880447, + 0.07544123381376266, + 0.0421132929623127, + 0.03986215218901634, + 0.08325818926095963, + 0.2861907482147217 + ], + "credit_map": [ + 0.011946120299398899, + 0.013548039831221104, + 0.008129354566335678, + 0.017641589045524597, + 0.010784079320728779, + 0.018626686185598373, + 0.00933714397251606, + 0.017387472093105316, + 0.011741762049496174, + 0.01235545240342617, + 0.009426632896065712, + 0.013052531518042088, + 0.009033529087901115, + 0.01658492162823677, + 0.012246855534613132, + 0.017284797504544258, + 0.011642377823591232, + 0.012248110957443714, + 0.009612135589122772, + 0.012922411784529686, + 0.009432506747543812, + 0.01968088187277317, + 0.013830047100782394, + 0.01886843703687191, + 0.005894800648093224, + 0.0074735465459525585, + 0.004789897706359625, + 0.010020270012319088, + 0.0052236514165997505, + 0.016563961282372475, + 0.004977210890501738, + 0.01611006073653698, + 0.009211087599396706, + 0.011279795318841934, + 0.005956572014838457, + 0.01605057902634144, + 0.008449641056358814, + 0.013291451148688793, + 0.007814867421984673, + 0.01266311015933752, + 0.00919155590236187, + 0.011425836011767387, + 0.007684405893087387, + 0.015020365826785564, + 0.006866175681352615, + 0.00880737416446209, + 0.005334066692739725, + 0.011111455969512463, + 0.00616631330922246, + 0.004811371210962534, + 0.004685255698859692, + 0.006425849627703428, + 0.0026589510962367058, + 0.007712758611887693, + 0.0021648418623954058, + 0.007487952709197998, + 0.00612297561019659, + 0.006350207608193159, + 0.0032624653540551662, + 0.004584294743835926, + 0.0032835551537573338, + 0.006056091748178005, + 0.003644560696557164, + 0.006558004301041365, + 0.011423703283071518, + 0.009376577101647854, + 0.006255004554986954, + 0.009351806715130806, + 0.006798646878451109, + 0.01754075288772583, + 0.005119698587805033, + 0.017392002046108246, + 0.028773751109838486, + 0.01969689130783081, + 0.012846570461988449, + 0.019778788089752197, + 0.011518767103552818, + 0.09209419786930084, + 0.011127560399472713, + 0.09035421907901764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.17995557188987732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera remains still, focused on the deeply wrinkled face of an elderly monk. A ray of warm afternoon sunlight streams in from the side, sculpting his features with precise precision, casting bright highlights on his nose, cheekbones, and forehead, while the other side of his face is immersed in soft shadow, revealing every wrinkle of his skin, like the ridges of time. Even the finest pores and freckles of time are visible on his bronzed complexion. His shiny head reflects the soft light, and he wears a faded ochre-colored, coarse cotton and linen robe. His expression is serene and solemn. His eyelids droop slowly, almost imperceptibly, casting a faint shadow in the light before closing completely, shutting out the outside world, as if he has entered a deep meditative state. As his eyes close, the muscles on his face relax slightly, and his entire face conveys a sense of ultimate peace and detached wisdom, creating an atmosphere of tranquility, dignity, and Zen.", + "seed": 1147, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_14/index_895_seed_1147_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.20738607477396725, + "vlm_logprob": -0.20738607477396725, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.20738607477396725, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00721787940710783 + ], + "opposite_logprob_tokens": [ + -6.257217884063721 + ], + "logprob_reward": -0.00721787940710783, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00721787940710783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28549838066101074, + 0.09301400184631348, + 0.06386696547269821, + 0.047634199261665344, + 0.07338040322065353, + 0.05601435527205467, + 0.1637096107006073, + 0.09809207916259766, + 0.06518416106700897, + 0.0536058247089386 + ], + "credit_map": [ + 0.020690923556685448, + 0.05412593483924866, + 0.0385061614215374, + 0.02703329361975193, + 0.029859913513064384, + 0.04374301806092262, + 0.027074027806520462, + 0.04446512088179588, + 0.012922587804496288, + 0.013991409912705421, + 0.012551769614219666, + 0.00778996292501688, + 0.009406178258359432, + 0.013371261768043041, + 0.0071557858027517796, + 0.015825051814317703, + 0.00866208877414465, + 0.008478832431137562, + 0.00846079457551241, + 0.005387466866523027, + 0.005789680872112513, + 0.009197724051773548, + 0.00473362673074007, + 0.013156753033399582, + 0.005467350594699383, + 0.007972423918545246, + 0.005681625567376614, + 0.004828677512705326, + 0.007371555548161268, + 0.006159395445138216, + 0.003798467107117176, + 0.006354709155857563, + 0.007770909462124109, + 0.01124486792832613, + 0.009829485788941383, + 0.006384427659213543, + 0.010437466204166412, + 0.010880948975682259, + 0.005204989109188318, + 0.011627309024333954, + 0.006612707860767841, + 0.008811953477561474, + 0.008167965337634087, + 0.00424669124186039, + 0.008062266744673252, + 0.008030309341847897, + 0.0038383568171411753, + 0.008244113996624947, + 0.01635008305311203, + 0.022398384287953377, + 0.019805315881967545, + 0.014763640239834785, + 0.032470475882291794, + 0.02245492674410343, + 0.012545336969196796, + 0.02292146347463131, + 0.010739949531853199, + 0.011975377798080444, + 0.013638673350214958, + 0.007902465760707855, + 0.01159098744392395, + 0.01391136646270752, + 0.008297339081764221, + 0.02003592811524868, + 0.004725536331534386, + 0.00948482658714056, + 0.007409586571156979, + 0.005847694352269173, + 0.011199827305972576, + 0.009555468335747719, + 0.00729322386905551, + 0.009667997248470783, + 0.006579288747161627, + 0.00776510126888752, + 0.0074043432250618935, + 0.005254420917481184, + 0.007260531187057495, + 0.007611460983753204, + 0.0041962843388319016, + 0.007534398697316647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.00721787940710783 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7579070329666138 + ], + "opposite_logprob_tokens": [ + -0.6329070329666138 + ], + "logprob_reward": -0.7579070329666138, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7579070329666138, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2836282551288605, + 0.12409631162881851, + 0.06716512143611908, + 0.09233901649713516, + 0.06278794258832932, + 0.0835275948047638, + 0.12170067429542542, + 0.052565086632966995, + 0.06856068968772888, + 0.043629344552755356 + ], + "credit_map": [ + 0.025443006306886673, + 0.049285851418972015, + 0.038118697702884674, + 0.02762320637702942, + 0.03886288404464722, + 0.035427045077085495, + 0.02930968813598156, + 0.03955785185098648, + 0.014723852276802063, + 0.019067781046032906, + 0.015051549300551414, + 0.00996679998934269, + 0.02040247805416584, + 0.016864201053977013, + 0.010497434996068478, + 0.017522217705845833, + 0.004528895020484924, + 0.010331726633012295, + 0.008439089171588421, + 0.007159829139709473, + 0.008835213258862495, + 0.008806031197309494, + 0.007861267775297165, + 0.011203068308532238, + 0.008105508051812649, + 0.014542442746460438, + 0.00940889772027731, + 0.009168640710413456, + 0.016215462237596512, + 0.01189876813441515, + 0.009801986627280712, + 0.013197312131524086, + 0.005424686707556248, + 0.007602319121360779, + 0.01180438045412302, + 0.004769426770508289, + 0.00812199991196394, + 0.009035507217049599, + 0.004771751817315817, + 0.011257865466177464, + 0.008232006803154945, + 0.010259438306093216, + 0.010661473497748375, + 0.00499344989657402, + 0.009960257448256016, + 0.012585166841745377, + 0.005965863820165396, + 0.02086993306875229, + 0.006374693941324949, + 0.01385563611984253, + 0.015956779941916466, + 0.007696399465203285, + 0.01742769218981266, + 0.02103680744767189, + 0.011099173687398434, + 0.028253493830561638, + 0.006151535548269749, + 0.006728285923600197, + 0.007308255881071091, + 0.003605285659432411, + 0.005274937488138676, + 0.00770599814131856, + 0.0038151987828314304, + 0.011975591070950031, + 0.005755867809057236, + 0.009084777906537056, + 0.009921529330313206, + 0.005311555229127407, + 0.010066092945635319, + 0.010162497870624065, + 0.005500133614987135, + 0.012758233584463596, + 0.003912028856575489, + 0.005835131276398897, + 0.006714517250657082, + 0.0036172433756291866, + 0.004839987959712744, + 0.005986395757645369, + 0.0038622848223894835, + 0.008861757814884186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.7579070329666138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030044058337807655 + ], + "opposite_logprob_tokens": [ + -3.5300440788269043 + ], + "logprob_reward": -0.030044058337807655, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.030044058337807655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4578152596950531, + 0.18358658254146576, + 0.07378905266523361, + 0.021064961329102516, + 0.0358569398522377, + 0.06657274812459946, + 0.055705536156892776, + 0.06374549865722656, + 0.021968377754092216, + 0.019895076751708984 + ], + "credit_map": [ + 0.026295583695173264, + 0.04526791721582413, + 0.0628076121211052, + 0.036124926060438156, + 0.05192660912871361, + 0.06685783714056015, + 0.05527494475245476, + 0.11325981467962265, + 0.013186552561819553, + 0.017032824456691742, + 0.02525652013719082, + 0.01555542554706335, + 0.019988641142845154, + 0.026099324226379395, + 0.022079430520534515, + 0.04438785836100578, + 0.00498896325007081, + 0.006728888023644686, + 0.010903168469667435, + 0.005979257170110941, + 0.008539971895515919, + 0.01025751605629921, + 0.008041898719966412, + 0.018349383026361465, + 0.002141037257388234, + 0.0024557504802942276, + 0.003071961458772421, + 0.001985683338716626, + 0.0026094927452504635, + 0.003441344015300274, + 0.0020912017207592726, + 0.0032684882171452045, + 0.0024560526944696903, + 0.0036173691041767597, + 0.004663463681936264, + 0.003168407129123807, + 0.004235685802996159, + 0.004922183696180582, + 0.004253664519637823, + 0.008540108799934387, + 0.003490146715193987, + 0.006889504846185446, + 0.008794859983026981, + 0.005322225391864777, + 0.007178931497037411, + 0.009859425015747547, + 0.009101205505430698, + 0.015936441719532013, + 0.003232076996937394, + 0.005617859307676554, + 0.007545079104602337, + 0.004605131223797798, + 0.006414428353309631, + 0.008387126959860325, + 0.006953847128897905, + 0.012949979864060879, + 0.004724075086414814, + 0.006805088371038437, + 0.008172739297151566, + 0.0043861884623765945, + 0.007645511534065008, + 0.00899673905223608, + 0.007684399373829365, + 0.015330752357840538, + 0.003538519609719515, + 0.002911507850512862, + 0.003325516125187278, + 0.0013947290135547519, + 0.0025908472016453743, + 0.0035707708448171616, + 0.0017394691240042448, + 0.002897014608606696, + 0.0029700640588998795, + 0.0023725517094135284, + 0.0032432666048407555, + 0.00196729670278728, + 0.0020886678248643875, + 0.0028654199559241533, + 0.0017804792150855064, + 0.002607327653095126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.030044058337807655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016090186312794685 + ], + "opposite_logprob_tokens": [ + -4.141090393066406 + ], + "logprob_reward": -0.016090186312794685, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016090186312794685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18469829857349396, + 0.11037196218967438, + 0.10046229511499405, + 0.0815691351890564, + 0.20627853274345398, + 0.07829643785953522, + 0.041348110884428024, + 0.045753128826618195, + 0.04842440411448479, + 0.10279762744903564 + ], + "credit_map": [ + 0.019455239176750183, + 0.022876130416989326, + 0.02522670291364193, + 0.0163718368858099, + 0.021003352478146553, + 0.027146970853209496, + 0.017333228141069412, + 0.03528484329581261, + 0.009285260923206806, + 0.011426903307437897, + 0.015616185031831264, + 0.009618374519050121, + 0.011217436753213406, + 0.016167832538485527, + 0.013883078470826149, + 0.023156890645623207, + 0.007251329254359007, + 0.010523800738155842, + 0.014780690893530846, + 0.00847967341542244, + 0.010642742738127708, + 0.013693577609956264, + 0.012021135538816452, + 0.023069338873028755, + 0.011727096512913704, + 0.010507135652005672, + 0.010956760495901108, + 0.0063114226795732975, + 0.009147142991423607, + 0.011638741008937359, + 0.006376382429152727, + 0.014904449693858624, + 0.02815929427742958, + 0.03421834856271744, + 0.025034354999661446, + 0.015202410519123077, + 0.023809805512428284, + 0.032764337956905365, + 0.016439124941825867, + 0.03065086342394352, + 0.010546551086008549, + 0.011729370802640915, + 0.009110995568335056, + 0.005902869626879692, + 0.008597242645919323, + 0.012192018330097198, + 0.00666659465059638, + 0.013550790958106518, + 0.005323524121195078, + 0.005776959005743265, + 0.006051446311175823, + 0.003234857926145196, + 0.004797000903636217, + 0.006083992309868336, + 0.0025635394267737865, + 0.007516789250075817, + 0.004760917276144028, + 0.008060947991907597, + 0.004909323528409004, + 0.0033361604437232018, + 0.007726890966296196, + 0.0066906968131661415, + 0.004291403107345104, + 0.005976793356239796, + 0.005098262801766396, + 0.007482380140572786, + 0.005771220661699772, + 0.004813681356608868, + 0.007718067150563002, + 0.006765452679246664, + 0.004810785409063101, + 0.005964558571577072, + 0.013888077810406685, + 0.016191016882658005, + 0.012580132111907005, + 0.007929515093564987, + 0.01186851691454649, + 0.015646424144506454, + 0.008312162943184376, + 0.01638178527355194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.016090186312794685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256712168455124 + ], + "opposite_logprob_tokens": [ + -1.6006711721420288 + ], + "logprob_reward": -0.2256712168455124, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256712168455124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28465667366981506, + 0.1558823138475418, + 0.09962106496095657, + 0.07123613357543945, + 0.12432172894477844, + 0.03492932394146919, + 0.10599938035011292, + 0.023620612919330597, + 0.05898378789424896, + 0.040749043226242065 + ], + "credit_map": [ + 0.019876914098858833, + 0.04319550096988678, + 0.040718067437410355, + 0.0225661750882864, + 0.038301289081573486, + 0.04219453036785126, + 0.022383134812116623, + 0.05542103573679924, + 0.011956549249589443, + 0.021738450974225998, + 0.018935156986117363, + 0.01570185087621212, + 0.02063487283885479, + 0.021573735401034355, + 0.025843700394034386, + 0.019497981294989586, + 0.007024317514151335, + 0.013997207395732403, + 0.01309006754308939, + 0.010531972162425518, + 0.010716554708778858, + 0.013769339770078659, + 0.01538880169391632, + 0.015102790668606758, + 0.006810801103711128, + 0.011006503365933895, + 0.00946857687085867, + 0.005811142735183239, + 0.011026888154447079, + 0.009002325125038624, + 0.0067887804470956326, + 0.011321105062961578, + 0.009788723662495613, + 0.012324371375143528, + 0.021096495911478996, + 0.010638203471899033, + 0.013664508238434792, + 0.01764393411576748, + 0.013848040252923965, + 0.025317439809441566, + 0.0028523108921945095, + 0.0056394413113594055, + 0.004750591237097979, + 0.002277375664561987, + 0.004332320764660835, + 0.0049062129110097885, + 0.002425865735858679, + 0.0077452026307582855, + 0.008779915980994701, + 0.013909619301557541, + 0.016694577410817146, + 0.007740723434835672, + 0.014530553482472897, + 0.01763669028878212, + 0.007748198229819536, + 0.018959088250994682, + 0.001997184008359909, + 0.003428125288337469, + 0.00287753832526505, + 0.0014999056002125144, + 0.002832654397934675, + 0.0033296889159828424, + 0.00209446856752038, + 0.0055610439740121365, + 0.005308501422405243, + 0.007886647246778011, + 0.008998451754450798, + 0.004407041240483522, + 0.00765110831707716, + 0.008582117967307568, + 0.0057417177595198154, + 0.01040819101035595, + 0.004683551378548145, + 0.00722274836152792, + 0.0040048109367489815, + 0.0029111814219504595, + 0.007534991484135389, + 0.0056216842494904995, + 0.0037912384141236544, + 0.0049788313917815685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.2256712168455124 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera remains still, focused on the deeply wrinkled face of an elderly monk. A ray of warm afternoon sunlight streams in from the side, sculpting his features with precise precision, casting bright highlights on his nose, cheekbones, and forehead, while the other side of his face is immersed in soft shadow, revealing every wrinkle of his skin, like the ridges of time. Even the finest pores and freckles of time are visible on his bronzed complexion. His shiny head reflects the soft light, and he wears a faded ochre-colored, coarse cotton and linen robe. His expression is serene and solemn. His eyelids droop slowly, almost imperceptibly, casting a faint shadow in the light before closing completely, shutting out the outside world, as if he has entered a deep meditative state. As his eyes close, the muscles on his face relax slightly, and his entire face conveys a sense of ultimate peace and detached wisdom, creating an atmosphere of tranquility, dignity, and Zen.", + "seed": 1148, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_14/index_895_seed_1148_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1756633988581598, + "vlm_logprob": -0.1756633988581598, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.1756633988581598, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0038713286630809307 + ], + "opposite_logprob_tokens": [ + -6.878871440887451 + ], + "logprob_reward": -0.0038713286630809307, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0038713286630809307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14375312626361847, + 0.0704493373632431, + 0.04832759127020836, + 0.08934582024812698, + 0.0859440341591835, + 0.15229664742946625, + 0.2388056069612503, + 0.038620561361312866, + 0.09800475090742111, + 0.03445252403616905 + ], + "credit_map": [ + 0.014970256015658379, + 0.017515597864985466, + 0.00957697443664074, + 0.010550757870078087, + 0.024525878950953484, + 0.017983272671699524, + 0.0295942984521389, + 0.01903608813881874, + 0.007783535402268171, + 0.006949061527848244, + 0.0059449030086398125, + 0.005739736370742321, + 0.014558953233063221, + 0.008875544182956219, + 0.011224268935620785, + 0.009373335167765617, + 0.005429988261312246, + 0.006209334824234247, + 0.0033269121777266264, + 0.005576516967266798, + 0.007512614596635103, + 0.0065920487977564335, + 0.007425015326589346, + 0.0062551642768085, + 0.010671311989426613, + 0.008891735225915909, + 0.00493937311694026, + 0.009202243760228157, + 0.015010139904916286, + 0.013049234636127949, + 0.016901619732379913, + 0.010680166073143482, + 0.008527372032403946, + 0.0077263242565095425, + 0.00501515856012702, + 0.006874071434140205, + 0.018858294934034348, + 0.010958450846374035, + 0.014926397241652012, + 0.013057961128652096, + 0.012995812110602856, + 0.015962112694978714, + 0.01180764939635992, + 0.014096192084252834, + 0.029613811522722244, + 0.019319863989949226, + 0.02089526504278183, + 0.02760593593120575, + 0.030375169590115547, + 0.026313399896025658, + 0.016842572018504143, + 0.017006346955895424, + 0.043734360486269, + 0.03582358360290527, + 0.03431779891252518, + 0.034392379224300385, + 0.003479608567431569, + 0.004322141408920288, + 0.001648553879931569, + 0.003968413453549147, + 0.007238056510686874, + 0.005575918592512608, + 0.007969173602759838, + 0.0044186911545693874, + 0.010867095552384853, + 0.012481574900448322, + 0.010111815296113491, + 0.009710684418678284, + 0.015021482482552528, + 0.011744687333703041, + 0.014612076804041862, + 0.013455337844789028, + 0.0025879116728901863, + 0.00396345928311348, + 0.0018992348341271281, + 0.0036083816085010767, + 0.006439992226660252, + 0.0036711657885462046, + 0.008095570839941502, + 0.004186810459941626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0038713286630809307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762465596199036 + ], + "opposite_logprob_tokens": [ + -0.8262465596199036 + ], + "logprob_reward": -0.5762465596199036, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762465596199036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23636557161808014, + 0.07873643189668655, + 0.09962760657072067, + 0.11640233546495438, + 0.05775754153728485, + 0.042773324996232986, + 0.08452444523572922, + 0.12479595094919205, + 0.11004646122455597, + 0.04897031933069229 + ], + "credit_map": [ + 0.02703046426177025, + 0.02160056307911873, + 0.01818348653614521, + 0.02302832342684269, + 0.036305300891399384, + 0.030681412667036057, + 0.05318010598421097, + 0.02635592222213745, + 0.004436424467712641, + 0.00998152606189251, + 0.005940182134509087, + 0.006018355488777161, + 0.015364686027169228, + 0.00934769120067358, + 0.018856598064303398, + 0.008790971711277962, + 0.010767832398414612, + 0.011079042218625546, + 0.00827972125262022, + 0.009861172176897526, + 0.018922707065939903, + 0.011799954809248447, + 0.013851594179868698, + 0.015065586194396019, + 0.010365169495344162, + 0.010287094861268997, + 0.012681032530963421, + 0.013581539504230022, + 0.014383521862328053, + 0.01799647882580757, + 0.021008864045143127, + 0.016098638996481895, + 0.003748369636014104, + 0.006692754570394754, + 0.0037451451644301414, + 0.005402110517024994, + 0.014147372916340828, + 0.006310673430562019, + 0.011634750291705132, + 0.006076365243643522, + 0.00392866600304842, + 0.004434411879628897, + 0.0027110367082059383, + 0.0036513982340693474, + 0.008035199716687202, + 0.0059753479436039925, + 0.007465369068086147, + 0.00657189404591918, + 0.009596006013453007, + 0.009360644966363907, + 0.007604627870023251, + 0.00907977856695652, + 0.013184752315282822, + 0.012471022084355354, + 0.012342573143541813, + 0.010885041207075119, + 0.015965396538376808, + 0.014792976900935173, + 0.006921603810042143, + 0.011461847461760044, + 0.024302994832396507, + 0.018087275326251984, + 0.019535327330231667, + 0.013728532008826733, + 0.01268682535737753, + 0.01691775768995285, + 0.006703774910420179, + 0.01098704431205988, + 0.016760770231485367, + 0.01173622626811266, + 0.019585028290748596, + 0.014669036492705345, + 0.0037092699203640223, + 0.005684994626790285, + 0.003497071797028184, + 0.0048878388479352, + 0.0072796531021595, + 0.006172035355120897, + 0.010800343006849289, + 0.006939114071428776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.5762465596199036 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016180040314793587 + ], + "opposite_logprob_tokens": [ + -4.141180038452148 + ], + "logprob_reward": -0.016180040314793587, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016180040314793587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22256453335285187, + 0.0861591026186943, + 0.11534766852855682, + 0.17497673630714417, + 0.058759160339832306, + 0.03241492807865143, + 0.04565195366740227, + 0.05964092165231705, + 0.08674436807632446, + 0.11774057894945145 + ], + "credit_map": [ + 0.028094951063394547, + 0.02322656288743019, + 0.015944359824061394, + 0.02151605114340782, + 0.03572564944624901, + 0.03509934991598129, + 0.03210243582725525, + 0.030855195596814156, + 0.010280133225023746, + 0.010091697797179222, + 0.0072184354066848755, + 0.008308464661240578, + 0.015199307352304459, + 0.01242206059396267, + 0.012056716717779636, + 0.010582291521131992, + 0.008673603646457195, + 0.012156493030488491, + 0.009201082400977612, + 0.009996160864830017, + 0.02328256145119667, + 0.016002332791686058, + 0.018395686522126198, + 0.017639752477407455, + 0.017663365229964256, + 0.017295805737376213, + 0.017213016748428345, + 0.014327168464660645, + 0.03325505554676056, + 0.021870635449886322, + 0.02611587755382061, + 0.02723582275211811, + 0.0034093777649104595, + 0.005742284003645182, + 0.004939856007695198, + 0.005252772942185402, + 0.00992611050605774, + 0.00925726443529129, + 0.012592736631631851, + 0.007638758514076471, + 0.004824576433748007, + 0.004343034699559212, + 0.002126916777342558, + 0.0030294270254671574, + 0.005602676421403885, + 0.004412114154547453, + 0.004515326581895351, + 0.0035608564503490925, + 0.00426883902400732, + 0.004211551509797573, + 0.004340668208897114, + 0.0037703453563153744, + 0.007965960539877415, + 0.0066037471406161785, + 0.006212596315890551, + 0.008278250694274902, + 0.007731112185865641, + 0.005887070670723915, + 0.0059792716056108475, + 0.004790955688804388, + 0.011159969493746758, + 0.00831578765064478, + 0.008969129994511604, + 0.006807632278650999, + 0.006981049198657274, + 0.012690244242548943, + 0.0041587576270103455, + 0.007649144623428583, + 0.013990753330290318, + 0.010199264623224735, + 0.017233332619071007, + 0.013841832987964153, + 0.00965665653347969, + 0.014074664562940598, + 0.007041678763926029, + 0.010803412646055222, + 0.02228071354329586, + 0.012476542964577675, + 0.025387762114405632, + 0.016019154340028763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.016180040314793587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982412464916706 + ], + "opposite_logprob_tokens": [ + -3.5298240184783936 + ], + "logprob_reward": -0.02982412464916706, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982412464916706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21474583446979523, + 0.16084414720535278, + 0.18745996057987213, + 0.08930763602256775, + 0.08620352298021317, + 0.02742161974310875, + 0.05731208994984627, + 0.06579887121915817, + 0.03700641542673111, + 0.07389984279870987 + ], + "credit_map": [ + 0.01844516023993492, + 0.027451902627944946, + 0.017534634098410606, + 0.02029378153383732, + 0.04015456140041351, + 0.026455853134393692, + 0.03645947948098183, + 0.027950484305620193, + 0.012213102541863918, + 0.016488492488861084, + 0.011460806243121624, + 0.014322432689368725, + 0.03132890537381172, + 0.02205745503306389, + 0.035827990621328354, + 0.01714499481022358, + 0.01925710216164589, + 0.020394064486026764, + 0.012241481803357601, + 0.01689828932285309, + 0.033520109951496124, + 0.022680770605802536, + 0.04084828868508339, + 0.021619874984025955, + 0.0069654639810323715, + 0.008505693636834621, + 0.007702429313212633, + 0.007011789362877607, + 0.01828114129602909, + 0.012385095469653606, + 0.015987861901521683, + 0.01246817409992218, + 0.007766712922602892, + 0.008201384916901588, + 0.008551690727472305, + 0.00856170617043972, + 0.01266541425138712, + 0.01346609927713871, + 0.01400156132876873, + 0.01298896037042141, + 0.0023188928607851267, + 0.003121939254924655, + 0.002207663143053651, + 0.0023165580350905657, + 0.005455735605210066, + 0.0034359698183834553, + 0.005004324018955231, + 0.0035605398006737232, + 0.0051439059898257256, + 0.005440968554466963, + 0.005227929912507534, + 0.004962512291967869, + 0.010970006696879864, + 0.007454144302755594, + 0.008157922886312008, + 0.009954704903066158, + 0.0069319405592978, + 0.005843642633408308, + 0.006960754748433828, + 0.006441482808440924, + 0.011157657951116562, + 0.008479188196361065, + 0.01165052130818367, + 0.008333686739206314, + 0.004889365751296282, + 0.0045484029687941074, + 0.002289567841216922, + 0.0032896811608225107, + 0.006844634190201759, + 0.004423562902957201, + 0.005169300362467766, + 0.0055519030429422855, + 0.007206551730632782, + 0.010700570419430733, + 0.0036701816134154797, + 0.006763432174921036, + 0.014446211978793144, + 0.008610434830188751, + 0.012969420291483402, + 0.009533049538731575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.02982412464916706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25219494104385376 + ], + "opposite_logprob_tokens": [ + -1.502194881439209 + ], + "logprob_reward": -0.25219494104385376, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25219494104385376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.165318101644516, + 0.20402291417121887, + 0.07513569295406342, + 0.06478501111268997, + 0.15512613952159882, + 0.05493621900677681, + 0.095795638859272, + 0.08445131033658981, + 0.07333496958017349, + 0.027093948796391487 + ], + "credit_map": [ + 0.01264004223048687, + 0.01915394514799118, + 0.0077908397652208805, + 0.018477806821465492, + 0.028909724205732346, + 0.020970091223716736, + 0.031439829617738724, + 0.02593582682311535, + 0.012884845025837421, + 0.02556166797876358, + 0.011924179270863533, + 0.01917889155447483, + 0.04601294919848442, + 0.028762560337781906, + 0.03872257098555565, + 0.020975247025489807, + 0.004775333683937788, + 0.0073500266298651695, + 0.004777084570378065, + 0.007308031897991896, + 0.015402470715343952, + 0.01033860631287098, + 0.017410550266504288, + 0.007773591671139002, + 0.0071082874201238155, + 0.007134079933166504, + 0.006137089803814888, + 0.00541918259114027, + 0.010776427574455738, + 0.008547071367502213, + 0.0120693976059556, + 0.007593480404466391, + 0.0156788919121027, + 0.014460471458733082, + 0.0110630476847291, + 0.010986864566802979, + 0.033782731741666794, + 0.018334465101361275, + 0.028102444484829903, + 0.022717224434018135, + 0.0049821422435343266, + 0.007160269655287266, + 0.004050360526889563, + 0.0059671360068023205, + 0.010467498563230038, + 0.007414068095386028, + 0.007807751651853323, + 0.007086996920406818, + 0.011613759212195873, + 0.010199020616710186, + 0.007679942529648542, + 0.007258536759763956, + 0.01605401374399662, + 0.012322609312832355, + 0.014544853009283543, + 0.016122903674840927, + 0.010872771963477135, + 0.010037695057690144, + 0.0041040475480258465, + 0.007844599895179272, + 0.01636177860200405, + 0.01154354028403759, + 0.014944230206310749, + 0.008742650039494038, + 0.008840198628604412, + 0.010865099728107452, + 0.004526533652096987, + 0.007744475733488798, + 0.010629607364535332, + 0.008277089335024357, + 0.011724225245416164, + 0.010727740824222565, + 0.003300470532849431, + 0.0027944515459239483, + 0.0013913015136495233, + 0.0015297519275918603, + 0.005944379605352879, + 0.0029710454400628805, + 0.005268766079097986, + 0.0038937791250646114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.25219494104385376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera remains still, focused on the deeply wrinkled face of an elderly monk. A ray of warm afternoon sunlight streams in from the side, sculpting his features with precise precision, casting bright highlights on his nose, cheekbones, and forehead, while the other side of his face is immersed in soft shadow, revealing every wrinkle of his skin, like the ridges of time. Even the finest pores and freckles of time are visible on his bronzed complexion. His shiny head reflects the soft light, and he wears a faded ochre-colored, coarse cotton and linen robe. His expression is serene and solemn. His eyelids droop slowly, almost imperceptibly, casting a faint shadow in the light before closing completely, shutting out the outside world, as if he has entered a deep meditative state. As his eyes close, the muscles on his face relax slightly, and his entire face conveys a sense of ultimate peace and detached wisdom, creating an atmosphere of tranquility, dignity, and Zen.", + "seed": 1149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_14/index_895_seed_1149_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2302621129900217, + "vlm_logprob": -0.2302621129900217, + "vlm_yes_no_margin": 3.4000000298023223, + "vlm_token_credit": -0.2302621129900217, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00869311299175024 + ], + "opposite_logprob_tokens": [ + -6.883693218231201 + ], + "logprob_reward": -0.00869311299175024, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.00869311299175024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1601308137178421, + 0.058602724224328995, + 0.05401236563920975, + 0.23351718485355377, + 0.21914270520210266, + 0.058508675545454025, + 0.03876504302024841, + 0.06252965331077576, + 0.0622578002512455, + 0.05253300070762634 + ], + "credit_map": [ + 0.015971515327692032, + 0.015643425285816193, + 0.01962524652481079, + 0.01069472823292017, + 0.030928602442145348, + 0.009615331888198853, + 0.029262181371450424, + 0.028389785438776016, + 0.006473931018263102, + 0.004021255765110254, + 0.006197022274136543, + 0.005064632743597031, + 0.010201194323599339, + 0.0033261470962315798, + 0.011441325768828392, + 0.011877213604748249, + 0.005044830963015556, + 0.003301308024674654, + 0.004440016578882933, + 0.003653943771496415, + 0.010423206724226475, + 0.0032983217388391495, + 0.009416460059583187, + 0.014434274286031723, + 0.01801316812634468, + 0.032133396714925766, + 0.015577173791825771, + 0.0263358186930418, + 0.050829529762268066, + 0.012131900526583195, + 0.03766581788659096, + 0.04083036258816719, + 0.01847466453909874, + 0.01131685171276331, + 0.008854768238961697, + 0.007331698201596737, + 0.04369121044874191, + 0.006166350096464157, + 0.087001733481884, + 0.036305446177721024, + 0.004513172898441553, + 0.007273807190358639, + 0.006411330308765173, + 0.004114415030926466, + 0.008644849061965942, + 0.006636388599872589, + 0.012427851557731628, + 0.008486862294375896, + 0.0034986885730177164, + 0.0057546477764844894, + 0.004176520276814699, + 0.00555465929210186, + 0.004105802159756422, + 0.0034458779264241457, + 0.004645890556275845, + 0.007582959719002247, + 0.007120572961866856, + 0.00883795227855444, + 0.006400233134627342, + 0.006267460063099861, + 0.0067653413861989975, + 0.005732795223593712, + 0.0081699313595891, + 0.013235368765890598, + 0.003723612753674388, + 0.008354046382009983, + 0.004498336464166641, + 0.0041123200207948685, + 0.006559835746884346, + 0.003949983976781368, + 0.011070244945585728, + 0.019989417865872383, + 0.0026167877949774265, + 0.007540194317698479, + 0.0033081239089369774, + 0.004597160499542952, + 0.0071568735875189304, + 0.0041060796938836575, + 0.00793229229748249, + 0.015275489538908005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.00869311299175024 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743374586105347 + ], + "opposite_logprob_tokens": [ + -0.47433748841285706 + ], + "logprob_reward": -0.9743374586105347, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9743374586105347, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.030006621032953262, + 0.0348716676235199, + 0.0745263621211052, + 0.07299591600894928, + 0.6349369287490845, + 0.045928530395030975, + 0.05209844931960106, + 0.038963768631219864, + 0.006728705018758774, + 0.0089430445805192 + ], + "credit_map": [ + 0.0036226266529411077, + 0.0036822007969021797, + 0.0031580261420458555, + 0.004154451657086611, + 0.004687408916652203, + 0.002000739797949791, + 0.005016818642616272, + 0.0036843493580818176, + 0.0038267646450549364, + 0.0025738996919244528, + 0.002476741326972842, + 0.0019466161029413342, + 0.006833313498646021, + 0.0023365248925983906, + 0.005196957848966122, + 0.009680848568677902, + 0.006837329361587763, + 0.0041318172588944435, + 0.005875736009329557, + 0.003944311290979385, + 0.014707189984619617, + 0.004772864747792482, + 0.012717636302113533, + 0.0215394776314497, + 0.005435968283563852, + 0.009589440189301968, + 0.005447886884212494, + 0.00748869962990284, + 0.018068090081214905, + 0.004315612372010946, + 0.01226877886801958, + 0.010381441563367844, + 0.06273916363716125, + 0.06648939847946167, + 0.07069160044193268, + 0.05523071810603142, + 0.07913936674594879, + 0.04524432495236397, + 0.11375255882740021, + 0.14164981245994568, + 0.00438445620238781, + 0.006919041741639376, + 0.005585878621786833, + 0.004008588846772909, + 0.006162027362734079, + 0.002130253007635474, + 0.010778965428471565, + 0.005959320347756147, + 0.0046396818943321705, + 0.007395007181912661, + 0.0056883604265749454, + 0.00632506376132369, + 0.006628325209021568, + 0.0056085544638335705, + 0.006958868820220232, + 0.008854586631059647, + 0.0031700623221695423, + 0.0034891278482973576, + 0.0035435587633401155, + 0.002960560377687216, + 0.003756290301680565, + 0.003382663708180189, + 0.005823870189487934, + 0.012837637215852737, + 0.0005201022140681744, + 0.0010052020661532879, + 0.000816615647636354, + 0.000644478655885905, + 0.0008544561569578946, + 0.0007729626377113163, + 0.0009613900911062956, + 0.001153497607447207, + 0.0006684233667328954, + 0.0011949650943279266, + 0.0008544440497644246, + 0.0007353744003921747, + 0.0010495636379346251, + 0.001204559812322259, + 0.0011090774787589908, + 0.0021266366820782423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.9743374586105347 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014313879422843456 + ], + "opposite_logprob_tokens": [ + -4.264313697814941 + ], + "logprob_reward": -0.014313879422843456, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014313879422843456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20976664125919342, + 0.28252336382865906, + 0.10560642927885056, + 0.10362272709608078, + 0.04954327270388603, + 0.05613311007618904, + 0.043905045837163925, + 0.041032176464796066, + 0.033241014927625656, + 0.07462619245052338 + ], + "credit_map": [ + 0.021255798637866974, + 0.016143321990966797, + 0.026519667357206345, + 0.014096101745963097, + 0.03964819386601448, + 0.011334837414324284, + 0.03687397390604019, + 0.04389474168419838, + 0.028905058279633522, + 0.02316310629248619, + 0.03441363200545311, + 0.02382773905992508, + 0.04883928596973419, + 0.01623503305017948, + 0.05863136798143387, + 0.0485081672668457, + 0.008272659033536911, + 0.008470091037452221, + 0.00694371247664094, + 0.0068055917508900166, + 0.021888768300414085, + 0.006484935525804758, + 0.017024733126163483, + 0.02971593476831913, + 0.007142378017306328, + 0.01820632629096508, + 0.006338527426123619, + 0.016896093264222145, + 0.01898587867617607, + 0.004613755736500025, + 0.014432553201913834, + 0.017007211223244667, + 0.004143227823078632, + 0.004309146199375391, + 0.004351234063506126, + 0.004160710144788027, + 0.008297160267829895, + 0.002916032914072275, + 0.014534987509250641, + 0.006830774713307619, + 0.00430955970659852, + 0.007828926667571068, + 0.006612005643546581, + 0.004196908324956894, + 0.006671739276498556, + 0.00392200518399477, + 0.011655556969344616, + 0.01093640923500061, + 0.005549746099859476, + 0.005071297287940979, + 0.0055093648843467236, + 0.006700778845697641, + 0.004512253683060408, + 0.0034550961572676897, + 0.0057580797001719475, + 0.007348431274294853, + 0.003507478628307581, + 0.004574080929160118, + 0.004237081855535507, + 0.0027426541782915592, + 0.004572909791022539, + 0.0032436714973300695, + 0.0061182742938399315, + 0.01203602459281683, + 0.0020326944068074226, + 0.004374472424387932, + 0.0035854224115610123, + 0.001976655563339591, + 0.0035414418671280146, + 0.002703222446143627, + 0.005061846226453781, + 0.009965258650481701, + 0.006238965783268213, + 0.009129387326538563, + 0.008961323648691177, + 0.005854865536093712, + 0.007514141034334898, + 0.005829418543726206, + 0.01013292744755745, + 0.02096516825258732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.014313879422843456 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091247200965881 + ], + "opposite_logprob_tokens": [ + -4.511091232299805 + ], + "logprob_reward": -0.011091247200965881, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091247200965881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07168536633253098, + 0.0646604672074318, + 0.09549834579229355, + 0.053073637187480927, + 0.20222324132919312, + 0.03884505480527878, + 0.06037416681647301, + 0.12470140308141708, + 0.05999639630317688, + 0.2289418876171112 + ], + "credit_map": [ + 0.007686229422688484, + 0.010430882684886456, + 0.007891492918133736, + 0.0083455890417099, + 0.011231821030378342, + 0.004629548639059067, + 0.012127385474741459, + 0.009342418983578682, + 0.005512504372745752, + 0.008900423534214497, + 0.0050037014298141, + 0.006783223710954189, + 0.010962835513055325, + 0.003536742413416505, + 0.009066539816558361, + 0.014894491992890835, + 0.007869835011661053, + 0.011131723411381245, + 0.006927571725100279, + 0.010311334393918514, + 0.016769839450716972, + 0.005896462593227625, + 0.01448341365903616, + 0.022108163684606552, + 0.004153131041675806, + 0.007919450290501118, + 0.004009926225990057, + 0.008065576665103436, + 0.010528869926929474, + 0.0027729785069823265, + 0.007900213822722435, + 0.007723487913608551, + 0.019579827785491943, + 0.018009666353464127, + 0.02128174714744091, + 0.016938088461756706, + 0.027600735425949097, + 0.01269938051700592, + 0.04571250081062317, + 0.04040129482746124, + 0.0036894320510327816, + 0.0033792343456298113, + 0.0029169812332838774, + 0.0026091851759701967, + 0.0047086249105632305, + 0.0035640897694975138, + 0.010020861402153969, + 0.007956643588840961, + 0.0061080921441316605, + 0.00675225118175149, + 0.005572741851210594, + 0.004511859267950058, + 0.007665799465030432, + 0.007785354740917683, + 0.008625992573797703, + 0.013352074660360813, + 0.011689476668834686, + 0.015337471850216389, + 0.012860099785029888, + 0.009791768155992031, + 0.012336332350969315, + 0.01258988119661808, + 0.01792512834072113, + 0.03217124938964844, + 0.005030459724366665, + 0.005703584756702185, + 0.004548251163214445, + 0.005459008738398552, + 0.005732950288802385, + 0.00785494688898325, + 0.008470745757222176, + 0.017196452245116234, + 0.018883781507611275, + 0.020658820867538452, + 0.018994498997926712, + 0.017910970374941826, + 0.023433255031704903, + 0.023667432367801666, + 0.03341865912079811, + 0.07197447121143341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.011091247200965881 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused on the face of an elderly monk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the monk wearing a faded ochre-colored coarse cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a ray of warm afternoon sunlight stream from the side, creating bright highlights on one side of the face and soft shadows on the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the monk's eyelids droop slowly and close completely while the facial muscles relax?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14287486672401428 + ], + "opposite_logprob_tokens": [ + -2.0178749561309814 + ], + "logprob_reward": -0.14287486672401428, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14287486672401428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12829084694385529, + 0.04358532652258873, + 0.08003278821706772, + 0.09500762820243835, + 0.19277533888816833, + 0.10105574131011963, + 0.2242339551448822, + 0.033509183675050735, + 0.04331482574343681, + 0.058194298297166824 + ], + "credit_map": [ + 0.010640537366271019, + 0.019221000373363495, + 0.010759287513792515, + 0.013426225632429123, + 0.02722514234483242, + 0.007224864326417446, + 0.01943683624267578, + 0.02035696990787983, + 0.004335443489253521, + 0.006317254155874252, + 0.0038312359247356653, + 0.004655240103602409, + 0.008652422577142715, + 0.002315566875040531, + 0.006696220487356186, + 0.006781947799026966, + 0.0074486020021140575, + 0.009965022094547749, + 0.006896835286170244, + 0.006324275862425566, + 0.01691359654068947, + 0.004003110807389021, + 0.013449590653181076, + 0.015031766146421432, + 0.007640636060386896, + 0.007977336645126343, + 0.005726141855120659, + 0.007239063270390034, + 0.028264475986361504, + 0.004665581975132227, + 0.01933392696082592, + 0.014160473830997944, + 0.015752345323562622, + 0.01419230829924345, + 0.014075723476707935, + 0.01041271723806858, + 0.03482162952423096, + 0.007717796601355076, + 0.06379799544811249, + 0.03200483322143555, + 0.01019787322729826, + 0.013949703425168991, + 0.010715967044234276, + 0.011664368212223053, + 0.011653123423457146, + 0.0057405754923820496, + 0.023189205676317215, + 0.01394493319094181, + 0.019414816051721573, + 0.03958567604422569, + 0.022555265575647354, + 0.034503024071455, + 0.02600233256816864, + 0.01581619866192341, + 0.03442850708961487, + 0.0319281667470932, + 0.002601426560431719, + 0.0039932336658239365, + 0.0039411853067576885, + 0.004836501087993383, + 0.004341741558164358, + 0.0031017246656119823, + 0.004240353591740131, + 0.006453019101172686, + 0.0029143185820430517, + 0.006723260041326284, + 0.0030597809236496687, + 0.004514402709901333, + 0.004299630410969257, + 0.004820127505809069, + 0.005881716497242451, + 0.01110159046947956, + 0.003866124665364623, + 0.0054959626868367195, + 0.005421200767159462, + 0.00604658666998148, + 0.006247612182050943, + 0.004241386894136667, + 0.008671794086694717, + 0.018203632906079292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.14287486672401428 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_15/index_tensor([852, 852])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_15/index_tensor([852, 852])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cc874be26a1ebca3dd98f839026e4f7c8d51dab4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_15/index_tensor([852, 852])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, majestic wide shot begins, the camera slowly arcing and pulling back to reveal a complex deep-space probe tumbling uncontrollably in the silent void. The probe's design is a terrifyingly beautiful spectacle, reminiscent of H. R. Giger, composed of a ribbed, skeletal biomechanical structure that fuses polished obsidian-like chitin with dark, visceral sinew and metal alloys. It spins slowly and erratically, end-to-end, in a state of catastrophic failure. Its massive solar panels shatter into countless crystalline shards that drift away like frozen tears, catching the faint, cold glow of a distant sun and revealing sparkling, frayed internal wiring. As the camera continues to pull back to an extreme wide shot, Jupiter's vast, stormy backdrop is revealed, its turbulent clouds swirling in shades of deep ochre, cream, and intense red, the Great Red Spot like a vast, unblinking eye. The scene is illuminated by a harsh directional light from a distant sun, casting harsh highlights and deep, cavernous shadows on the probe's alien form, while the planet itself casts a faint, ominous, reddish-orange glow across its belly, creating a sublime and terrifying atmosphere of cosmic isolation and magnificent decay.", + "seed": 16012, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_15/index_852_seed_16012_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06222373824566603, + "vlm_logprob": -0.06222373824566603, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.06222373824566603, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04042476788163185 + ], + "opposite_logprob_tokens": [ + -3.6654248237609863 + ], + "logprob_reward": -0.04042476788163185, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04042476788163185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23371897637844086, + 0.06867071986198425, + 0.05875779315829277, + 0.0633850246667862, + 0.1227613091468811, + 0.08962473273277283, + 0.0192280113697052, + 0.03841199725866318, + 0.1422617882490158, + 0.16317960619926453 + ], + "credit_map": [ + 0.019694559276103973, + 0.016676584258675575, + 0.020011499524116516, + 0.015303189866244793, + 0.03003464825451374, + 0.04250930994749069, + 0.03886428475379944, + 0.05062490701675415, + 0.006421232596039772, + 0.007917866110801697, + 0.011260696686804295, + 0.007747614290565252, + 0.004954272881150246, + 0.008805534802377224, + 0.009562142193317413, + 0.012001359835267067, + 0.004429667256772518, + 0.008362935855984688, + 0.009104056283831596, + 0.005440504755824804, + 0.005375852808356285, + 0.008137875236570835, + 0.010435870848596096, + 0.007471031043678522, + 0.005398503039032221, + 0.0067052836529910564, + 0.006122709717601538, + 0.007783798035234213, + 0.0058869426138699055, + 0.011815892532467842, + 0.010417432524263859, + 0.009254460223019123, + 0.009196016006171703, + 0.013912062160670757, + 0.016845205798745155, + 0.013344489969313145, + 0.0224782582372427, + 0.02006908506155014, + 0.013446548953652382, + 0.013469640165567398, + 0.013593747280538082, + 0.004734793212264776, + 0.021447062492370605, + 0.005939373280853033, + 0.006990515626966953, + 0.006821704097092152, + 0.007762224413454533, + 0.02233530767261982, + 0.00226327171549201, + 0.0018600051989778876, + 0.0019552637822926044, + 0.002628665417432785, + 0.002345436019822955, + 0.002462939592078328, + 0.0025198860093951225, + 0.0031925439834594727, + 0.0057662962935864925, + 0.004327764734625816, + 0.004397327546030283, + 0.006279750261455774, + 0.004433642607182264, + 0.004873073659837246, + 0.004783425480127335, + 0.0035507185384631157, + 0.025868134573101997, + 0.022736871615052223, + 0.01953275315463543, + 0.015857171267271042, + 0.016196919605135918, + 0.01692378520965576, + 0.013064682483673096, + 0.012081467546522617, + 0.020062841475009918, + 0.007771121338009834, + 0.05176561325788498, + 0.02477039024233818, + 0.00909293070435524, + 0.019986707717180252, + 0.007307914085686207, + 0.02242208831012249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.04042476788163185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332662232220173 + ], + "opposite_logprob_tokens": [ + -3.7733266353607178 + ], + "logprob_reward": -0.02332662232220173, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332662232220173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12729744613170624, + 0.11505567282438278, + 0.06163676455616951, + 0.0686187595129013, + 0.11888154596090317, + 0.13195398449897766, + 0.025825046002864838, + 0.0916607677936554, + 0.08824655413627625, + 0.17082339525222778 + ], + "credit_map": [ + 0.017307421192526817, + 0.00943025667220354, + 0.011079860851168633, + 0.007285913918167353, + 0.01610540598630905, + 0.019726017490029335, + 0.0188404843211174, + 0.027522079646587372, + 0.01180926151573658, + 0.01186454202979803, + 0.0177757628262043, + 0.011472100391983986, + 0.008626931346952915, + 0.016497327014803886, + 0.016614170745015144, + 0.02039557509124279, + 0.00561250327154994, + 0.00979161448776722, + 0.005919469520449638, + 0.005916670896112919, + 0.006442981772124767, + 0.006408571265637875, + 0.012829535640776157, + 0.008715416304767132, + 0.008815441280603409, + 0.004797918722033501, + 0.009433194063603878, + 0.005735814105719328, + 0.007075837347656488, + 0.010067163966596127, + 0.012360899709165096, + 0.010332489386200905, + 0.011406207457184792, + 0.011518510058522224, + 0.015471186488866806, + 0.011669830419123173, + 0.018610509112477303, + 0.019705751910805702, + 0.014266922138631344, + 0.016232630237936974, + 0.02054942399263382, + 0.00500181270763278, + 0.03474222868680954, + 0.008526720106601715, + 0.008659682236611843, + 0.00802003126591444, + 0.011017655953764915, + 0.03543642535805702, + 0.004193329717963934, + 0.002338739112019539, + 0.0030385323334485292, + 0.003993337042629719, + 0.002645627362653613, + 0.002935249824076891, + 0.002980760997161269, + 0.0036994710098952055, + 0.007355903275310993, + 0.00983330700546503, + 0.007633984088897705, + 0.013780183158814907, + 0.009707925841212273, + 0.017802705988287926, + 0.013272121548652649, + 0.01227463036775589, + 0.010765167884528637, + 0.010915438644587994, + 0.010703985579311848, + 0.01199387852102518, + 0.010106420144438744, + 0.013317464850842953, + 0.011941513046622276, + 0.008502686396241188, + 0.021239526569843292, + 0.009459578432142735, + 0.054669324308633804, + 0.0270759928971529, + 0.010090126655995846, + 0.018306085839867592, + 0.008510864339768887, + 0.021471887826919556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.02332662232220173 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496704578399658 + ], + "opposite_logprob_tokens": [ + -2.929966926574707 + ], + "logprob_reward": -0.05496704578399658, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496704578399658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05942478030920029, + 0.06366331875324249, + 0.035910122096538544, + 0.023458097130060196, + 0.02237117290496826, + 0.08887521177530289, + 0.013837405480444431, + 0.060600120574235916, + 0.10014402121305466, + 0.5317156910896301 + ], + "credit_map": [ + 0.006097387056797743, + 0.005050189793109894, + 0.006273383740335703, + 0.004591252654790878, + 0.010426308959722519, + 0.009672565385699272, + 0.010095960460603237, + 0.0072177378460764885, + 0.00586446700617671, + 0.006901936139911413, + 0.008845613338053226, + 0.006480650510638952, + 0.005130726844072342, + 0.010715760290622711, + 0.009173144586384296, + 0.010551019571721554, + 0.0025903587229549885, + 0.006024757865816355, + 0.0068777077831327915, + 0.0036099969875067472, + 0.00377942924387753, + 0.004604178015142679, + 0.0035016017500311136, + 0.004922093823552132, + 0.002980089746415615, + 0.0032859493512660265, + 0.0024189476389437914, + 0.0023952953051775694, + 0.0020573686342686415, + 0.0026951611507683992, + 0.004029470030218363, + 0.0035958178341388702, + 0.00247464282438159, + 0.001948200399056077, + 0.004025994334369898, + 0.0022954470477998257, + 0.003334277542307973, + 0.0029235067777335644, + 0.0026786483358591795, + 0.00269045727327466, + 0.014628073200583458, + 0.0027844137512147427, + 0.024254271760582924, + 0.005554609000682831, + 0.006146165542304516, + 0.005727837793529034, + 0.006486704107373953, + 0.02329314313828945, + 0.0017044610576704144, + 0.0014099014224484563, + 0.0015705367550253868, + 0.0019625802524387836, + 0.0014956917148083448, + 0.001839321805164218, + 0.0016000018222257495, + 0.0022549114655703306, + 0.007217221427708864, + 0.006392761133611202, + 0.007961193099617958, + 0.005739275831729174, + 0.005245997570455074, + 0.01190126221626997, + 0.00781505461782217, + 0.008327357470989227, + 0.011133470572531223, + 0.014721237123012543, + 0.0073390319012105465, + 0.011143220588564873, + 0.01351742260158062, + 0.019009098410606384, + 0.012040619738399982, + 0.011239924468100071, + 0.07204902172088623, + 0.027756715193390846, + 0.17735646665096283, + 0.08579330146312714, + 0.027067339047789574, + 0.05059153586626053, + 0.02483951859176159, + 0.06626179814338684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.05496704578399658 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543358840048313 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543358840048313, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543358840048313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12439047545194626, + 0.06607189029455185, + 0.04326717555522919, + 0.04731303080916405, + 0.053625959903001785, + 0.07849515229463577, + 0.0538826547563076, + 0.17458610236644745, + 0.2550491392612457, + 0.10331837087869644 + ], + "credit_map": [ + 0.014370854943990707, + 0.008044383488595486, + 0.008732764981687069, + 0.007717792876064777, + 0.020062074065208435, + 0.020475544035434723, + 0.02297140657901764, + 0.022015657275915146, + 0.008670005947351456, + 0.0069661675952374935, + 0.008683629333972931, + 0.006184958852827549, + 0.005393835250288248, + 0.008645559661090374, + 0.008663109503686428, + 0.012864628806710243, + 0.003677073633298278, + 0.006253697909414768, + 0.005625301040709019, + 0.0038080138619989157, + 0.006461411248892546, + 0.0060120378620922565, + 0.006485613062977791, + 0.004944028332829475, + 0.00598848657682538, + 0.004937102552503347, + 0.004749619867652655, + 0.00575309107080102, + 0.004555493127554655, + 0.006367849186062813, + 0.008358330465853214, + 0.0066030556336045265, + 0.005467723589390516, + 0.005191175267100334, + 0.008617730811238289, + 0.005883521866053343, + 0.005440804176032543, + 0.007121396716684103, + 0.007472696248441935, + 0.008430910296738148, + 0.011113084852695465, + 0.0058885845355689526, + 0.01494190189987421, + 0.006986516993492842, + 0.006805031094700098, + 0.00627994304522872, + 0.008634136989712715, + 0.01784595474600792, + 0.00988815538585186, + 0.005666754208505154, + 0.0059521133080124855, + 0.005686149001121521, + 0.0046052485704422, + 0.006088393274694681, + 0.006426021456718445, + 0.009569820947945118, + 0.02467876859009266, + 0.02693084254860878, + 0.02214748226106167, + 0.02797679603099823, + 0.01808365248143673, + 0.015136124566197395, + 0.02688848413527012, + 0.012743953615427017, + 0.0376313291490078, + 0.035596225410699844, + 0.03363161161541939, + 0.030080972239375114, + 0.024477927014231682, + 0.03881947696208954, + 0.026385996490716934, + 0.02842560224235058, + 0.010152335278689861, + 0.0075858598574995995, + 0.023175029084086418, + 0.011110851541161537, + 0.0107913538813591, + 0.018206702545285225, + 0.00727826589718461, + 0.01501797791570425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.012543358840048313 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985689640045166 + ], + "opposite_logprob_tokens": [ + -1.8048568964004517 + ], + "logprob_reward": -0.17985689640045166, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985689640045166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21305246651172638, + 0.06644454598426819, + 0.09804211556911469, + 0.17000429332256317, + 0.05523065850138664, + 0.0608937032520771, + 0.06980089098215103, + 0.07533206045627594, + 0.11193442344665527, + 0.07926484197378159 + ], + "credit_map": [ + 0.020133132115006447, + 0.019560014829039574, + 0.01971275359392166, + 0.015598069876432419, + 0.03484645485877991, + 0.03383176028728485, + 0.03006264753639698, + 0.039307642728090286, + 0.008585871197283268, + 0.006878701038658619, + 0.008099989965558052, + 0.006089036352932453, + 0.005634058732539415, + 0.010212158784270287, + 0.009662609547376633, + 0.011282125487923622, + 0.008540615439414978, + 0.010206914506852627, + 0.012567284516990185, + 0.008299296721816063, + 0.009559454396367073, + 0.016632825136184692, + 0.019393743947148323, + 0.012841983698308468, + 0.02192179299890995, + 0.018016647547483444, + 0.02311512641608715, + 0.016790039837360382, + 0.013251560740172863, + 0.022311413660645485, + 0.025880491361021996, + 0.028717240318655968, + 0.005470709875226021, + 0.005903748795390129, + 0.0075457836501300335, + 0.00695852329954505, + 0.009646536782383919, + 0.005760496947914362, + 0.007020710967481136, + 0.00692415377125144, + 0.009551478549838066, + 0.009548714384436607, + 0.005145674105733633, + 0.006281606387346983, + 0.008105230517685413, + 0.00564909540116787, + 0.009520534425973892, + 0.0070913792587816715, + 0.011862250976264477, + 0.006440824829041958, + 0.007569960784167051, + 0.008713655173778534, + 0.00710254767909646, + 0.008104690350592136, + 0.008617537096142769, + 0.011389431543648243, + 0.008295401930809021, + 0.009768291376531124, + 0.00862693041563034, + 0.011550725437700748, + 0.0074334763921797276, + 0.008256928063929081, + 0.012236357666552067, + 0.009163960814476013, + 0.018993720412254333, + 0.019846821203827858, + 0.025152618065476418, + 0.010985661298036575, + 0.008938908576965332, + 0.011625433340668678, + 0.008012359961867332, + 0.008378908969461918, + 0.013436862267553806, + 0.006709425710141659, + 0.017630772665143013, + 0.012388692237436771, + 0.005572706926614046, + 0.009570077992975712, + 0.005291561130434275, + 0.0086647504940629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.17985689640045166 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, majestic wide shot begins, the camera slowly arcing and pulling back to reveal a complex deep-space probe tumbling uncontrollably in the silent void. The probe's design is a terrifyingly beautiful spectacle, reminiscent of H. R. Giger, composed of a ribbed, skeletal biomechanical structure that fuses polished obsidian-like chitin with dark, visceral sinew and metal alloys. It spins slowly and erratically, end-to-end, in a state of catastrophic failure. Its massive solar panels shatter into countless crystalline shards that drift away like frozen tears, catching the faint, cold glow of a distant sun and revealing sparkling, frayed internal wiring. As the camera continues to pull back to an extreme wide shot, Jupiter's vast, stormy backdrop is revealed, its turbulent clouds swirling in shades of deep ochre, cream, and intense red, the Great Red Spot like a vast, unblinking eye. The scene is illuminated by a harsh directional light from a distant sun, casting harsh highlights and deep, cavernous shadows on the probe's alien form, while the planet itself casts a faint, ominous, reddish-orange glow across its belly, creating a sublime and terrifying atmosphere of cosmic isolation and magnificent decay.", + "seed": 16013, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_15/index_852_seed_16013_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09515766650438309, + "vlm_logprob": -0.09515766650438309, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.09515766650438309, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08234453946352005 + ], + "opposite_logprob_tokens": [ + -2.7073445320129395 + ], + "logprob_reward": -0.08234453946352005, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08234453946352005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15466244518756866, + 0.12029025703668594, + 0.04883163794875145, + 0.11893098801374435, + 0.09441890567541122, + 0.11468201130628586, + 0.0773993730545044, + 0.16027431190013885, + 0.053077612072229385, + 0.05743243917822838 + ], + "credit_map": [ + 0.012517271563410759, + 0.02700786478817463, + 0.021173056215047836, + 0.022412339225411415, + 0.02209738828241825, + 0.013110578991472721, + 0.021463850513100624, + 0.014880094677209854, + 0.008315358310937881, + 0.012628010474145412, + 0.012313834391534328, + 0.02044067345559597, + 0.01915575936436653, + 0.009550211951136589, + 0.02354314923286438, + 0.014343265444040298, + 0.005289535038173199, + 0.0046801334246993065, + 0.009043239057064056, + 0.006431615445762873, + 0.003843493526801467, + 0.004581310786306858, + 0.0070561314933001995, + 0.007906180806457996, + 0.010683990083634853, + 0.011872467584908009, + 0.0219709575176239, + 0.019302936270833015, + 0.013909258879721165, + 0.009597335942089558, + 0.015449704602360725, + 0.016144340857863426, + 0.007379607297480106, + 0.0057612089440226555, + 0.028995120897889137, + 0.010306203737854958, + 0.004884615540504456, + 0.013217016123235226, + 0.0065080258063972, + 0.017367107793688774, + 0.014648395590484142, + 0.008810642175376415, + 0.02348000928759575, + 0.015574749559164047, + 0.00820406898856163, + 0.013933070935308933, + 0.01276413444429636, + 0.01726694405078888, + 0.011354716494679451, + 0.005393192172050476, + 0.011286726221442223, + 0.009793057106435299, + 0.008643629029393196, + 0.009695668704807758, + 0.010399444960057735, + 0.010832938365638256, + 0.018455028533935547, + 0.020282043144106865, + 0.019963741302490234, + 0.015400340780615807, + 0.020641259849071503, + 0.025357594713568687, + 0.017948754131793976, + 0.022225553169846535, + 0.005471032112836838, + 0.006200375966727734, + 0.006803075782954693, + 0.006703217979520559, + 0.00783700030297041, + 0.009374989196658134, + 0.005037527997046709, + 0.005650391802191734, + 0.007345374207943678, + 0.005762768909335136, + 0.007823863066732883, + 0.006689634174108505, + 0.009108703583478928, + 0.010084094479680061, + 0.00478381710126996, + 0.005834182258695364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.08234453946352005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04873687028884888 + ], + "opposite_logprob_tokens": [ + -3.048736810684204 + ], + "logprob_reward": -0.04873687028884888, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04873687028884888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08117520809173584, + 0.305427223443985, + 0.032569997012615204, + 0.17912621796131134, + 0.07467740029096603, + 0.10830240696668625, + 0.05594532564282417, + 0.0638437271118164, + 0.06787002086639404, + 0.031062480062246323 + ], + "credit_map": [ + 0.007637299131602049, + 0.01326169352978468, + 0.009680643677711487, + 0.01622646115720272, + 0.011171557940542698, + 0.005808867048472166, + 0.009738331660628319, + 0.007650358602404594, + 0.020439568907022476, + 0.04703211039304733, + 0.029128195717930794, + 0.04499213397502899, + 0.04709985479712486, + 0.023354068398475647, + 0.05619899928569794, + 0.03718230128288269, + 0.0027064441237598658, + 0.003680052701383829, + 0.004824836738407612, + 0.005558661185204983, + 0.004427954088896513, + 0.002383058425039053, + 0.004470848012715578, + 0.004518142901360989, + 0.01746045984327793, + 0.016573181375861168, + 0.03772904351353645, + 0.025232622399926186, + 0.02226523868739605, + 0.014911720529198647, + 0.020914558321237564, + 0.02403939887881279, + 0.007439499720931053, + 0.0065130917355418205, + 0.019925439730286598, + 0.0070737008936703205, + 0.005518920719623566, + 0.009912543930113316, + 0.004966205917298794, + 0.01332799345254898, + 0.013426660560071468, + 0.008627394214272499, + 0.02326681651175022, + 0.014681966044008732, + 0.006024776492267847, + 0.01085074432194233, + 0.013711178675293922, + 0.017712870612740517, + 0.007302315905690193, + 0.004084865562617779, + 0.00826745480298996, + 0.007559795863926411, + 0.007468926254659891, + 0.007460078690201044, + 0.006403170060366392, + 0.007398718968033791, + 0.008180062286555767, + 0.006536077708005905, + 0.0063630579970777035, + 0.006420953664928675, + 0.008419881574809551, + 0.012679236941039562, + 0.005505292676389217, + 0.009739171713590622, + 0.007036763709038496, + 0.010191383771598339, + 0.008296462707221508, + 0.010476420633494854, + 0.009950948879122734, + 0.010595450177788734, + 0.005702391266822815, + 0.005620202980935574, + 0.0025220976676791906, + 0.0032275074627250433, + 0.0053711882792413235, + 0.004495024681091309, + 0.0036204131320118904, + 0.0042060017585754395, + 0.0036507202312350273, + 0.003969525918364525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.04873687028884888 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160570740699768 + ], + "opposite_logprob_tokens": [ + -1.7016056776046753 + ], + "logprob_reward": -0.20160570740699768, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160570740699768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0947193056344986, + 0.06186765432357788, + 0.06009736284613609, + 0.2879360616207123, + 0.25784504413604736, + 0.05160227790474892, + 0.07550941407680511, + 0.050158169120550156, + 0.03781929239630699, + 0.022445499897003174 + ], + "credit_map": [ + 0.007974660955369473, + 0.017428778111934662, + 0.012628593482077122, + 0.012788147665560246, + 0.012585875578224659, + 0.00912691093981266, + 0.01369304209947586, + 0.008493280969560146, + 0.0050272210501134396, + 0.007805845234543085, + 0.006201766896992922, + 0.010517306625843048, + 0.007603888399899006, + 0.006202037911862135, + 0.010516115464270115, + 0.007993465289473534, + 0.004550011362880468, + 0.008503803052008152, + 0.009405040182173252, + 0.008547154255211353, + 0.006109883543103933, + 0.005773126147687435, + 0.008721561171114445, + 0.008486776612699032, + 0.02447865717113018, + 0.03151065483689308, + 0.052169281989336014, + 0.045722417533397675, + 0.0377233624458313, + 0.023718351498246193, + 0.03882599622011185, + 0.0337873250246048, + 0.02082604356110096, + 0.011575383134186268, + 0.09040581434965134, + 0.024464813992381096, + 0.011448657140135765, + 0.024582942947745323, + 0.019207457080483437, + 0.05533391609787941, + 0.005678154993802309, + 0.003844078164547682, + 0.0076208943501114845, + 0.009938828647136688, + 0.004061666317284107, + 0.005436754785478115, + 0.00649212347343564, + 0.008529768325388432, + 0.01120709627866745, + 0.004676593001931906, + 0.01144515909254551, + 0.01219864096492529, + 0.008080852217972279, + 0.010396425612270832, + 0.008049610070884228, + 0.00945502519607544, + 0.006232707295566797, + 0.006870657671242952, + 0.005094058346003294, + 0.004582136403769255, + 0.00740679819136858, + 0.009506890550255775, + 0.004438033793121576, + 0.006026883609592915, + 0.0029308749362826347, + 0.006294399965554476, + 0.0034191894810646772, + 0.004553184378892183, + 0.006214121822267771, + 0.008172228932380676, + 0.0033556220587342978, + 0.0028796661645174026, + 0.0020007893908768892, + 0.004098060540854931, + 0.001988332951441407, + 0.0037898619193583727, + 0.0033140117302536964, + 0.0034925532527267933, + 0.0019525287207216024, + 0.0018093585968017578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.20160570740699768 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077518463134766 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.016077518463134766, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077518463134766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07167138159275055, + 0.10867936909198761, + 0.05101469159126282, + 0.09048758447170258, + 0.1020609512925148, + 0.2782391905784607, + 0.04425732418894768, + 0.06984734535217285, + 0.12490401417016983, + 0.05883815139532089 + ], + "credit_map": [ + 0.005959700793027878, + 0.013022726401686668, + 0.010507799685001373, + 0.007296348921954632, + 0.009595988318324089, + 0.007653513923287392, + 0.008593177422881126, + 0.00904212985187769, + 0.009443770162761211, + 0.023269232362508774, + 0.013021516613662243, + 0.013398447073996067, + 0.015148194506764412, + 0.009342486038804054, + 0.011260979808866978, + 0.01379474438726902, + 0.00684475339949131, + 0.004374013748019934, + 0.00920707918703556, + 0.006070631090551615, + 0.006696898955851793, + 0.0042528617195785046, + 0.005538607481867075, + 0.008029844611883163, + 0.010625324212014675, + 0.008474110625684261, + 0.01501582469791174, + 0.01167284045368433, + 0.012488661333918571, + 0.009257695637643337, + 0.009726294316351414, + 0.013226832263171673, + 0.010024186223745346, + 0.004947707522660494, + 0.02598544955253601, + 0.010188408195972443, + 0.007270737085491419, + 0.01945166103541851, + 0.008309059776365757, + 0.015883740037679672, + 0.030857710167765617, + 0.0218550655990839, + 0.05917392298579216, + 0.03104989603161812, + 0.015963353216648102, + 0.02960999310016632, + 0.03793039545416832, + 0.05179886147379875, + 0.0060741775669157505, + 0.0039486633613705635, + 0.00689590023830533, + 0.006055617239326239, + 0.004339574370533228, + 0.005625225603580475, + 0.005076074041426182, + 0.006242090370506048, + 0.008131674490869045, + 0.008542977273464203, + 0.009543079882860184, + 0.008613468147814274, + 0.009486224502325058, + 0.011580624617636204, + 0.006008474621921778, + 0.007940823212265968, + 0.013958141207695007, + 0.019329117611050606, + 0.014125506393611431, + 0.019320204854011536, + 0.018505077809095383, + 0.018968142569065094, + 0.010443213395774364, + 0.010254607535898685, + 0.004782930016517639, + 0.007860219106078148, + 0.006897921208292246, + 0.009837445802986622, + 0.008455600589513779, + 0.007868909277021885, + 0.007973060943186283, + 0.0051620653830468655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.016077518463134766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702369689941406 + ], + "opposite_logprob_tokens": [ + -2.127023696899414 + ], + "logprob_reward": -0.12702369689941406, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702369689941406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17860355973243713, + 0.14416253566741943, + 0.05861291661858559, + 0.08570412546396255, + 0.2522951364517212, + 0.1486371010541916, + 0.024098264053463936, + 0.041317641735076904, + 0.041690338402986526, + 0.02487836964428425 + ], + "credit_map": [ + 0.012233481742441654, + 0.03057953715324402, + 0.02442166768014431, + 0.02717166766524315, + 0.02597360499203205, + 0.015323019586503506, + 0.025037268176674843, + 0.017863307148218155, + 0.011965280398726463, + 0.028264660388231277, + 0.01727021485567093, + 0.02086903713643551, + 0.019053801894187927, + 0.009516223333775997, + 0.01987485960125923, + 0.017348453402519226, + 0.006459586322307587, + 0.007422204129397869, + 0.0074126264080405235, + 0.008309593424201012, + 0.009451327845454216, + 0.005681802053004503, + 0.005938397720456123, + 0.007937376387417316, + 0.008146923966705799, + 0.009803149849176407, + 0.01122092455625534, + 0.010908594354987144, + 0.015127786435186863, + 0.008636659942567348, + 0.011129613034427166, + 0.010730474255979061, + 0.02992737852036953, + 0.01628817990422249, + 0.07657532393932343, + 0.022570276632905006, + 0.01703912578523159, + 0.016706325113773346, + 0.020054250955581665, + 0.05313427746295929, + 0.01223937887698412, + 0.015094330534338951, + 0.020711535587906837, + 0.02718292735517025, + 0.0114098796620965, + 0.013680223375558853, + 0.024680333212018013, + 0.023638496175408363, + 0.0031003609765321016, + 0.0020398832857608795, + 0.003348676720634103, + 0.002905831439420581, + 0.002305316971614957, + 0.0033061958383768797, + 0.002967189298942685, + 0.004124808590859175, + 0.0036482554860413074, + 0.005040347576141357, + 0.0038539422675967216, + 0.0039357030764222145, + 0.005649195052683353, + 0.008645673282444477, + 0.003913835622370243, + 0.006630691234022379, + 0.0045698764733970165, + 0.005233918782323599, + 0.005677126348018646, + 0.005692698992788792, + 0.005101941525936127, + 0.006459142547100782, + 0.004230632912367582, + 0.004725002683699131, + 0.00246468186378479, + 0.0021497856359928846, + 0.0041514053009450436, + 0.003101502312347293, + 0.0032960304524749517, + 0.00371662899851799, + 0.002869785763323307, + 0.0031285493168979883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.12702369689941406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, majestic wide shot begins, the camera slowly arcing and pulling back to reveal a complex deep-space probe tumbling uncontrollably in the silent void. The probe's design is a terrifyingly beautiful spectacle, reminiscent of H. R. Giger, composed of a ribbed, skeletal biomechanical structure that fuses polished obsidian-like chitin with dark, visceral sinew and metal alloys. It spins slowly and erratically, end-to-end, in a state of catastrophic failure. Its massive solar panels shatter into countless crystalline shards that drift away like frozen tears, catching the faint, cold glow of a distant sun and revealing sparkling, frayed internal wiring. As the camera continues to pull back to an extreme wide shot, Jupiter's vast, stormy backdrop is revealed, its turbulent clouds swirling in shades of deep ochre, cream, and intense red, the Great Red Spot like a vast, unblinking eye. The scene is illuminated by a harsh directional light from a distant sun, casting harsh highlights and deep, cavernous shadows on the probe's alien form, while the planet itself casts a faint, ominous, reddish-orange glow across its belly, creating a sublime and terrifying atmosphere of cosmic isolation and magnificent decay.", + "seed": 16014, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_15/index_852_seed_16014_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09945527110248804, + "vlm_logprob": -0.09945527110248804, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.09945527110248804, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03732689470052719 + ], + "opposite_logprob_tokens": [ + -3.6623268127441406 + ], + "logprob_reward": -0.03732689470052719, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03732689470052719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13897575438022614, + 0.10224466025829315, + 0.07942651212215424, + 0.16010774672031403, + 0.07133512198925018, + 0.0834946483373642, + 0.11256206035614014, + 0.10031596571207047, + 0.09886582940816879, + 0.05267171561717987 + ], + "credit_map": [ + 0.017318151891231537, + 0.02476038783788681, + 0.01842993125319481, + 0.01178060658276081, + 0.015390899032354355, + 0.012937611900269985, + 0.01903199777007103, + 0.01932617649435997, + 0.020193450152873993, + 0.01507263071835041, + 0.015384873375296593, + 0.010649257339537144, + 0.009974882937967777, + 0.011632933281362057, + 0.009521564468741417, + 0.009815058670938015, + 0.012423355132341385, + 0.012928501702845097, + 0.009391113184392452, + 0.010999036952853203, + 0.008372313342988491, + 0.009289349429309368, + 0.007652975153177977, + 0.008369872346520424, + 0.005855686962604523, + 0.005378005560487509, + 0.0056731621734797955, + 0.00814334861934185, + 0.005378848873078823, + 0.060155145823955536, + 0.008093581534922123, + 0.061429958790540695, + 0.010459544137120247, + 0.006065082736313343, + 0.009258504956960678, + 0.008214617148041725, + 0.006526408717036247, + 0.00908861868083477, + 0.013807220384478569, + 0.007915128953754902, + 0.009332851506769657, + 0.012362688779830933, + 0.009541867300868034, + 0.008492768742144108, + 0.01043733861297369, + 0.011241471394896507, + 0.011157105676829815, + 0.010928559117019176, + 0.009167192503809929, + 0.016718681901693344, + 0.0072360653430223465, + 0.014141103252768517, + 0.018284063786268234, + 0.015569447539746761, + 0.01692649908363819, + 0.014518999494612217, + 0.009003286249935627, + 0.011660346761345863, + 0.007431503385305405, + 0.011942801997065544, + 0.016895951703190804, + 0.011621399782598019, + 0.017697297036647797, + 0.014063382521271706, + 0.00886629056185484, + 0.017160922288894653, + 0.010535147972404957, + 0.01272672601044178, + 0.012392132543027401, + 0.010787713341414928, + 0.011633267626166344, + 0.014763633720576763, + 0.004612304735928774, + 0.005220768041908741, + 0.004421593155711889, + 0.007292748894542456, + 0.005797456484287977, + 0.00626808637753129, + 0.007050239481031895, + 0.012008522637188435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.03732689470052719 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233280211687088 + ], + "opposite_logprob_tokens": [ + -3.7733280658721924 + ], + "logprob_reward": -0.0233280211687088, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233280211687088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1493089646100998, + 0.17362593114376068, + 0.04347187280654907, + 0.050519973039627075, + 0.09383203834295273, + 0.061574216932058334, + 0.10354568809270859, + 0.06478844583034515, + 0.18960894644260406, + 0.069723941385746 + ], + "credit_map": [ + 0.019341446459293365, + 0.018684735521674156, + 0.02044025808572769, + 0.018972070887684822, + 0.0218674186617136, + 0.012776615098118782, + 0.017870362848043442, + 0.019356047734618187, + 0.02480206824839115, + 0.02046968787908554, + 0.024371951818466187, + 0.022436106577515602, + 0.016001440584659576, + 0.02972332015633583, + 0.01798088476061821, + 0.017840461805462837, + 0.0057271323166787624, + 0.006286626681685448, + 0.0074734813533723354, + 0.004676018375903368, + 0.0036094284150749445, + 0.005251350346952677, + 0.004277813248336315, + 0.0061700185760855675, + 0.004563651978969574, + 0.006036724895238876, + 0.005483882501721382, + 0.008695367723703384, + 0.004306950606405735, + 0.006936919875442982, + 0.006466819904744625, + 0.008029656484723091, + 0.01047698687762022, + 0.009550844319164753, + 0.011380314826965332, + 0.010187797248363495, + 0.009804577566683292, + 0.014378647319972515, + 0.017183596268296242, + 0.010869267396628857, + 0.010318208485841751, + 0.006507926620543003, + 0.006236090324819088, + 0.006777805741876364, + 0.007140939123928547, + 0.008463638834655285, + 0.008566293865442276, + 0.007563314400613308, + 0.00672947708517313, + 0.014425503090023994, + 0.006527996156364679, + 0.011884285137057304, + 0.015391076914966106, + 0.017341820523142815, + 0.015349742956459522, + 0.015895795077085495, + 0.00514125544577837, + 0.00856083258986473, + 0.006823948118835688, + 0.009035401977598667, + 0.006779641844332218, + 0.008132997900247574, + 0.007945208810269833, + 0.012369154952466488, + 0.015447861514985561, + 0.016634270548820496, + 0.019427161663770676, + 0.017033111304044724, + 0.017396196722984314, + 0.023622805252671242, + 0.04291429743170738, + 0.03713323548436165, + 0.006185164209455252, + 0.007504916284233332, + 0.007614664267748594, + 0.013026287779211998, + 0.007831985130906105, + 0.00731182936578989, + 0.008315340615808964, + 0.01193375326693058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0233280211687088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992252111434937 + ], + "opposite_logprob_tokens": [ + -1.8049225807189941 + ], + "logprob_reward": -0.17992252111434937, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992252111434937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.091275155544281, + 0.08236619830131531, + 0.12229689955711365, + 0.11626599729061127, + 0.06570827960968018, + 0.09365390986204147, + 0.11363191157579422, + 0.09545572847127914, + 0.17368975281715393, + 0.04565617814660072 + ], + "credit_map": [ + 0.011681937612593174, + 0.016104303300380707, + 0.013792955316603184, + 0.009517249651253223, + 0.01292143203318119, + 0.0076197674497962, + 0.011280260980129242, + 0.008357253856956959, + 0.020137017592787743, + 0.010935530997812748, + 0.014260011725127697, + 0.00767406914383173, + 0.005162602290511131, + 0.007567314431071281, + 0.007966183125972748, + 0.008663459680974483, + 0.015383987687528133, + 0.01987285353243351, + 0.0156877264380455, + 0.02102268859744072, + 0.010007730685174465, + 0.013314064592123032, + 0.011849252507090569, + 0.015158599242568016, + 0.009305182844400406, + 0.012905307114124298, + 0.011490793898701668, + 0.014187419787049294, + 0.0074187982827425, + 0.025058621540665627, + 0.010258564725518227, + 0.02564130909740925, + 0.009937082417309284, + 0.006648029200732708, + 0.008620018139481544, + 0.006528086494654417, + 0.007189047988504171, + 0.009271111339330673, + 0.011566641740500927, + 0.005948255304247141, + 0.007518994156271219, + 0.010945169255137444, + 0.008057472296059132, + 0.011828286573290825, + 0.013780694454908371, + 0.015515141189098358, + 0.014173150062561035, + 0.011834999546408653, + 0.014454381540417671, + 0.014073862694203854, + 0.009102008305490017, + 0.014404438436031342, + 0.0179266519844532, + 0.012617186643183231, + 0.018914680927991867, + 0.01213869545608759, + 0.008335135877132416, + 0.011223279871046543, + 0.006219451315701008, + 0.012291066348552704, + 0.01649978756904602, + 0.010720808058977127, + 0.016620784997940063, + 0.013545410707592964, + 0.015291835181415081, + 0.024933714419603348, + 0.016424737870693207, + 0.023589450865983963, + 0.022316012531518936, + 0.01883280649781227, + 0.027234245091676712, + 0.025066949427127838, + 0.0036857451777905226, + 0.005222346633672714, + 0.0040542567148804665, + 0.006347873248159885, + 0.005102940835058689, + 0.005284508690237999, + 0.008714648894965649, + 0.007243860047310591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.17992252111434937 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638862796127796 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638862796127796, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638862796127796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1375739425420761, + 0.09157942235469818, + 0.06663402915000916, + 0.05945147946476936, + 0.08389551937580109, + 0.1110868752002716, + 0.1295381784439087, + 0.167380690574646, + 0.10214359313249588, + 0.050716277211904526 + ], + "credit_map": [ + 0.015378051437437534, + 0.01854109950363636, + 0.01816674880683422, + 0.014718777500092983, + 0.019089415669441223, + 0.011843651533126831, + 0.016818491742014885, + 0.02301771380007267, + 0.0160147026181221, + 0.008269201032817364, + 0.013490117155015469, + 0.010307218879461288, + 0.006564858835190535, + 0.012561067007482052, + 0.00878134649246931, + 0.015590899623930454, + 0.012879163958132267, + 0.0091458884999156, + 0.013195579871535301, + 0.007155582308769226, + 0.00615377863869071, + 0.005824575200676918, + 0.006067412905395031, + 0.006212047766894102, + 0.0032548145391047, + 0.0035050592850893736, + 0.0031747305765748024, + 0.003833235939964652, + 0.002808899851515889, + 0.019373467192053795, + 0.00385855115018785, + 0.019642719998955727, + 0.013239613734185696, + 0.007200443185865879, + 0.009379558265209198, + 0.00943222176283598, + 0.009400439448654652, + 0.01263200119137764, + 0.01348048634827137, + 0.009130759164690971, + 0.013101442717015743, + 0.011459443718194962, + 0.008914932608604431, + 0.00812467746436596, + 0.010579166933894157, + 0.0239864494651556, + 0.013270138762891293, + 0.02165062166750431, + 0.015474154613912106, + 0.01847945712506771, + 0.01041777990758419, + 0.018502812832593918, + 0.015399916097521782, + 0.01514799240976572, + 0.01956120878458023, + 0.016554854810237885, + 0.014997225254774094, + 0.02036934159696102, + 0.01109367236495018, + 0.02098216861486435, + 0.02623419091105461, + 0.01982019655406475, + 0.02972782775759697, + 0.02415607124567032, + 0.009445768781006336, + 0.011884004808962345, + 0.010272533632814884, + 0.010446025989949703, + 0.00993089098483324, + 0.012499154545366764, + 0.02008417434990406, + 0.017581036314368248, + 0.004276537336409092, + 0.006700344383716583, + 0.004825342912226915, + 0.009593186900019646, + 0.006245273165404797, + 0.004751857835799456, + 0.008080706000328064, + 0.0062430305406451225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.004638862796127796 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25206005573272705 + ], + "opposite_logprob_tokens": [ + -1.502060055732727 + ], + "logprob_reward": -0.25206005573272705, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25206005573272705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06784181296825409, + 0.02245977707207203, + 0.02895650453865528, + 0.08639844506978989, + 0.06388343125581741, + 0.17894138395786285, + 0.20595575869083405, + 0.11382585763931274, + 0.20467405021190643, + 0.02706293575465679 + ], + "credit_map": [ + 0.009317790158092976, + 0.012186919339001179, + 0.009740099310874939, + 0.007514213677495718, + 0.008688855916261673, + 0.005709771066904068, + 0.008761437609791756, + 0.005922724958509207, + 0.002488047583028674, + 0.004463974852114916, + 0.003023449331521988, + 0.0020125305745750666, + 0.0027176502626389265, + 0.0027675579767674208, + 0.0026625709142535925, + 0.002323996974155307, + 0.003813608782365918, + 0.005041319411247969, + 0.0034203974064439535, + 0.004985363222658634, + 0.0037113195285201073, + 0.0022978137712925673, + 0.0030792634934186935, + 0.0026074189227074385, + 0.0034600072540342808, + 0.003377809189260006, + 0.0034745612647384405, + 0.0031665244605392218, + 0.0023714073467999697, + 0.032675500959157944, + 0.004234762862324715, + 0.033637870103120804, + 0.009121718816459179, + 0.004936903715133667, + 0.008066573180258274, + 0.007109632715582848, + 0.005646197125315666, + 0.008463219739496708, + 0.013900761492550373, + 0.006638420280069113, + 0.017043940722942352, + 0.01770055666565895, + 0.014938725158572197, + 0.013741230592131615, + 0.019579894840717316, + 0.039225365966558456, + 0.02288978174328804, + 0.03382188826799393, + 0.02618030644953251, + 0.022446541115641594, + 0.017431216314435005, + 0.028264502063393593, + 0.024364806711673737, + 0.025725241750478745, + 0.02901286445558071, + 0.0325302891433239, + 0.008758461102843285, + 0.015042508952319622, + 0.006406690459698439, + 0.013533007353544235, + 0.01898210123181343, + 0.013276358135044575, + 0.02412567473948002, + 0.013701059855520725, + 0.016793593764305115, + 0.02691180445253849, + 0.017408575862646103, + 0.024731148034334183, + 0.02370450273156166, + 0.024175506085157394, + 0.03831816837191582, + 0.03263074532151222, + 0.002573894802480936, + 0.003074382897466421, + 0.003032968146726489, + 0.0046180193312466145, + 0.0031644159462302923, + 0.0031723943538963795, + 0.00406286446377635, + 0.0033639948815107346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.25206005573272705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, majestic wide shot begins, the camera slowly arcing and pulling back to reveal a complex deep-space probe tumbling uncontrollably in the silent void. The probe's design is a terrifyingly beautiful spectacle, reminiscent of H. R. Giger, composed of a ribbed, skeletal biomechanical structure that fuses polished obsidian-like chitin with dark, visceral sinew and metal alloys. It spins slowly and erratically, end-to-end, in a state of catastrophic failure. Its massive solar panels shatter into countless crystalline shards that drift away like frozen tears, catching the faint, cold glow of a distant sun and revealing sparkling, frayed internal wiring. As the camera continues to pull back to an extreme wide shot, Jupiter's vast, stormy backdrop is revealed, its turbulent clouds swirling in shades of deep ochre, cream, and intense red, the Great Red Spot like a vast, unblinking eye. The scene is illuminated by a harsh directional light from a distant sun, casting harsh highlights and deep, cavernous shadows on the probe's alien form, while the planet itself casts a faint, ominous, reddish-orange glow across its belly, creating a sublime and terrifying atmosphere of cosmic isolation and magnificent decay.", + "seed": 16015, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_15/index_852_seed_16015_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11369290817528963, + "vlm_logprob": -0.11369290817528963, + "vlm_yes_no_margin": 2.524999976158142, + "vlm_token_credit": -0.11369290817528963, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20959392189979553 + ], + "opposite_logprob_tokens": [ + -1.7095938920974731 + ], + "logprob_reward": -0.20959392189979553, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20959392189979553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05837498605251312, + 0.09178755432367325, + 0.05917222425341606, + 0.0981476902961731, + 0.0454879067838192, + 0.43035000562667847, + 0.061904434114694595, + 0.07341416925191879, + 0.044663600623607635, + 0.03669743984937668 + ], + "credit_map": [ + 0.008990099653601646, + 0.005879699252545834, + 0.005015747155994177, + 0.007967576384544373, + 0.007302054204046726, + 0.008712555281817913, + 0.008707051165401936, + 0.005800202954560518, + 0.007780550513416529, + 0.011783028952777386, + 0.00906496774405241, + 0.006908973678946495, + 0.007016293238848448, + 0.02972358651459217, + 0.008142775855958462, + 0.011367373168468475, + 0.008299136534333229, + 0.008469679392874241, + 0.010664308443665504, + 0.005009407177567482, + 0.005467094015330076, + 0.00902711320668459, + 0.006178756710141897, + 0.00605672225356102, + 0.013427838683128357, + 0.022918490692973137, + 0.010526389814913273, + 0.007381717674434185, + 0.009268403053283691, + 0.00830891914665699, + 0.008168692700564861, + 0.018147233873605728, + 0.005912160035222769, + 0.00428300304338336, + 0.00606145802885294, + 0.004087983164936304, + 0.00424152659252286, + 0.006145365070551634, + 0.005376005079597235, + 0.009380404837429523, + 0.16263379156589508, + 0.015570196323096752, + 0.18884536623954773, + 0.017580032348632812, + 0.006834982894361019, + 0.0066859046928584576, + 0.017352450639009476, + 0.014847255311906338, + 0.006619438994675875, + 0.01352835725992918, + 0.006518560461699963, + 0.007199198007583618, + 0.007052680943161249, + 0.007323972415179014, + 0.007487691007554531, + 0.006174526643007994, + 0.008849676698446274, + 0.013489979319274426, + 0.00555798364803195, + 0.012005380354821682, + 0.0155325373634696, + 0.00573696568608284, + 0.007828720845282078, + 0.004412920214235783, + 0.004337043035775423, + 0.008945857174694538, + 0.0019376240670681, + 0.003549755085259676, + 0.009415263310074806, + 0.006631119176745415, + 0.0041353474371135235, + 0.005711586680263281, + 0.003340762108564377, + 0.004879369400441647, + 0.004121337551623583, + 0.0021337333600968122, + 0.00861787237226963, + 0.0050193024799227715, + 0.0033615664578974247, + 0.0052234940230846405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.20959392189979553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0431344211101532 + ], + "opposite_logprob_tokens": [ + -3.1681344509124756 + ], + "logprob_reward": -0.0431344211101532, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0431344211101532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08759710192680359, + 0.10551155358552933, + 0.0888836607336998, + 0.043916188180446625, + 0.06084847450256348, + 0.334910124540329, + 0.05944650620222092, + 0.11964242905378342, + 0.048922743648290634, + 0.05032122880220413 + ], + "credit_map": [ + 0.01499656680971384, + 0.010122778825461864, + 0.007171762175858021, + 0.01213148981332779, + 0.007996517233550549, + 0.015521089546382427, + 0.010824207216501236, + 0.008832688443362713, + 0.00924087967723608, + 0.014337901026010513, + 0.011354089714586735, + 0.0074856700375676155, + 0.007028087507933378, + 0.03389794006943703, + 0.009015101008117199, + 0.013151876628398895, + 0.010318121872842312, + 0.012482182122766972, + 0.015817129984498024, + 0.00824209488928318, + 0.0099466722458601, + 0.012348445132374763, + 0.00993356667459011, + 0.009795455262064934, + 0.005957955028861761, + 0.007254548370838165, + 0.00480410223826766, + 0.004005030728876591, + 0.0035885919351130724, + 0.006838877219706774, + 0.004927163012325764, + 0.006539918947964907, + 0.006994863506406546, + 0.008673205971717834, + 0.0072316452860832214, + 0.005939628463238478, + 0.005799561273306608, + 0.008283854462206364, + 0.008532209321856499, + 0.009393508546054363, + 0.12048279494047165, + 0.012319923378527164, + 0.14273713529109955, + 0.014820930548012257, + 0.0070626260712742805, + 0.007517720106989145, + 0.014529653824865818, + 0.015439347364008427, + 0.0071035949513316154, + 0.012644387781620026, + 0.006033892743289471, + 0.006989387795329094, + 0.007344337645918131, + 0.006240217015147209, + 0.007294181268662214, + 0.005796508397907019, + 0.014232548885047436, + 0.021767811849713326, + 0.008824135176837444, + 0.01978902332484722, + 0.02476421929895878, + 0.009468578733503819, + 0.013332941569387913, + 0.007463171612471342, + 0.004616972059011459, + 0.00999148003757, + 0.0022089609410613775, + 0.0032447923440486193, + 0.012585368007421494, + 0.005889815744012594, + 0.004352209623903036, + 0.0060331388376653194, + 0.003958098124712706, + 0.00652135768905282, + 0.005117499269545078, + 0.002920502331107855, + 0.011885198764503002, + 0.008282807655632496, + 0.004134983289986849, + 0.007500783074647188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.0431344211101532 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428440660238266 + ], + "opposite_logprob_tokens": [ + -2.0178439617156982 + ], + "logprob_reward": -0.1428440660238266, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428440660238266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026779677718877792, + 0.038566816598176956, + 0.04638632386922836, + 0.03915027901530266, + 0.05399659648537636, + 0.7021891474723816, + 0.04689323902130127, + 0.028431681916117668, + 0.009568200446665287, + 0.008038056083023548 + ], + "credit_map": [ + 0.0034240109380334616, + 0.003280997509136796, + 0.0028407846111804247, + 0.003950726240873337, + 0.0027280705980956554, + 0.004772098269313574, + 0.003260866506025195, + 0.0025221197865903378, + 0.0031998518388718367, + 0.005875949282199144, + 0.004255713429301977, + 0.0029418605845421553, + 0.0026598151307553053, + 0.011567752808332443, + 0.0034972759895026684, + 0.004568594042211771, + 0.005376389250159264, + 0.006958316080272198, + 0.0074035851284861565, + 0.0037545792292803526, + 0.004464523866772652, + 0.008178399875760078, + 0.005145783070474863, + 0.005104736424982548, + 0.005486471578478813, + 0.00876621250063181, + 0.004311306402087212, + 0.0027380818501114845, + 0.004205911885946989, + 0.004181012976914644, + 0.0038347369991242886, + 0.00562654435634613, + 0.00590581214055419, + 0.005852664355188608, + 0.006788999307900667, + 0.004968281369656324, + 0.0054067885503172874, + 0.007712592836469412, + 0.008267635479569435, + 0.0090938163921237, + 0.2701523005962372, + 0.025910135358572006, + 0.3082665503025055, + 0.028811469674110413, + 0.011749891564249992, + 0.010772139765322208, + 0.02718096785247326, + 0.019345607608556747, + 0.006681544706225395, + 0.007917109876871109, + 0.004986404906958342, + 0.006613716948777437, + 0.006543560419231653, + 0.003629587823525071, + 0.006264088675379753, + 0.00425722124055028, + 0.0036080307327210903, + 0.00533657381311059, + 0.0024301125667989254, + 0.0045925830490887165, + 0.00559265585616231, + 0.0021312155295163393, + 0.0028879810124635696, + 0.0018525253981351852, + 0.0007287997868843377, + 0.0012746318243443966, + 0.00044490411528386176, + 0.0009170770645141602, + 0.002616948913782835, + 0.0014984100125730038, + 0.0010986624984070659, + 0.0009887655032798648, + 0.0007185505237430334, + 0.001030542072840035, + 0.0006252185557968915, + 0.000564315530937165, + 0.0018494161777198315, + 0.0014337741304188967, + 0.0006559173343703151, + 0.0011603212915360928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.1428440660238266 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544536031782627 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544536031782627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544536031782627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05254669487476349, + 0.040394555777311325, + 0.029545877128839493, + 0.04279177635908127, + 0.05921897292137146, + 0.5585823059082031, + 0.07011492550373077, + 0.06567401438951492, + 0.04103546962141991, + 0.040095433592796326 + ], + "credit_map": [ + 0.007934888824820518, + 0.0053574563935399055, + 0.004354780539870262, + 0.005853300914168358, + 0.0050094276666641235, + 0.011179002933204174, + 0.006777230184525251, + 0.006080602761358023, + 0.00364492554217577, + 0.0050943573005497456, + 0.004143340978771448, + 0.002936927368864417, + 0.004239521920681, + 0.010448884218931198, + 0.005183772183954716, + 0.004702827427536249, + 0.0029401895590126514, + 0.0032838790211826563, + 0.005558670032769442, + 0.0021747048012912273, + 0.002441425109282136, + 0.005083537194877863, + 0.0038805119693279266, + 0.004182959441095591, + 0.005596063565462828, + 0.005607780534774065, + 0.005416859406977892, + 0.002467523794621229, + 0.004273363389074802, + 0.007130686193704605, + 0.006818442139774561, + 0.005481057800352573, + 0.009413884952664375, + 0.004278123844414949, + 0.007863965816795826, + 0.006565147545188665, + 0.00390064250677824, + 0.006510969717055559, + 0.004817970562726259, + 0.01586826890707016, + 0.20153886079788208, + 0.02807234413921833, + 0.2154284119606018, + 0.026273535564541817, + 0.015277734957635403, + 0.02303023263812065, + 0.027722397819161415, + 0.021238816902041435, + 0.009788090363144875, + 0.012721464037895203, + 0.01102882158011198, + 0.007169527467340231, + 0.006019063759595156, + 0.00739018339663744, + 0.008858906105160713, + 0.007138870190829039, + 0.008000173605978489, + 0.011968892998993397, + 0.0055593024007976055, + 0.011370192281901836, + 0.012449388392269611, + 0.005192344076931477, + 0.006918294355273247, + 0.004215425346046686, + 0.005344522651284933, + 0.01209226343780756, + 0.001764960354194045, + 0.0032804571092128754, + 0.005009694956243038, + 0.004273575264960527, + 0.002816637745127082, + 0.006453359965234995, + 0.0029543668497353792, + 0.004390624817460775, + 0.002569554140791297, + 0.00251969532109797, + 0.006778457667678595, + 0.010040025226771832, + 0.0034509520046412945, + 0.007391757797449827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.012544536031782627 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a deep-space probe with a ribbed, skeletal biomechanical structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the probe's material composed of polished obsidian-like chitin, dark sinew, and metal alloys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the probe's massive solar panels shattering into crystalline shards that drift away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background Jupiter with turbulent clouds in shades of deep ochre, cream, and intense red, including the Great Red Spot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a wide shot and slowly arc and pull back to an extreme wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603475958108902 + ], + "opposite_logprob_tokens": [ + -1.910347580909729 + ], + "logprob_reward": -0.1603475958108902, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603475958108902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08877101540565491, + 0.0684751570224762, + 0.04187578335404396, + 0.15789741277694702, + 0.2026464194059372, + 0.2261088788509369, + 0.07076581567525864, + 0.0737682580947876, + 0.036877021193504333, + 0.03281427547335625 + ], + "credit_map": [ + 0.011571311391890049, + 0.010353044606745243, + 0.00603767903521657, + 0.011825497262179852, + 0.00903036817908287, + 0.019761158153414726, + 0.011986654251813889, + 0.0082052918151021, + 0.00520560285076499, + 0.010483909398317337, + 0.006900404114276171, + 0.005485307425260544, + 0.008659936487674713, + 0.01558650191873312, + 0.008402740582823753, + 0.007750748656690121, + 0.005543087143450975, + 0.005750924348831177, + 0.005782948341220617, + 0.003314947010949254, + 0.003544821171090007, + 0.007558450568467379, + 0.004639456048607826, + 0.005741145461797714, + 0.020886290818452835, + 0.0369873009622097, + 0.01678377017378807, + 0.012551293708384037, + 0.01386866346001625, + 0.012665091082453728, + 0.015687495470046997, + 0.028467489406466484, + 0.044541135430336, + 0.015785157680511475, + 0.022131385281682014, + 0.020686136558651924, + 0.0111232940107584, + 0.019507326185703278, + 0.016276227310299873, + 0.05259574204683304, + 0.059415847063064575, + 0.01945747621357441, + 0.0764661431312561, + 0.015295340679585934, + 0.010941851884126663, + 0.010005014017224312, + 0.016487494111061096, + 0.018039707094430923, + 0.009168921038508415, + 0.011891878210008144, + 0.007597519084811211, + 0.009677533991634846, + 0.011071781627833843, + 0.0057366290129721165, + 0.009410648606717587, + 0.0062108999118208885, + 0.007997151464223862, + 0.013390690088272095, + 0.007287943735718727, + 0.010079387575387955, + 0.010737568140029907, + 0.008380470797419548, + 0.009221554733812809, + 0.0066734859719872475, + 0.0032721618190407753, + 0.007116108201444149, + 0.0015524771297350526, + 0.0038168742321431637, + 0.009027502499520779, + 0.0040647005662322044, + 0.003770943498238921, + 0.00425625080242753, + 0.0023789117112755775, + 0.004868150223046541, + 0.002227101009339094, + 0.002269730670377612, + 0.0071436273865401745, + 0.006051215808838606, + 0.002567624906077981, + 0.005307910032570362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.1603475958108902 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_2/index_tensor([2294, 2294])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_2/index_tensor([2294, 2294])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8c769fd21a5ead3af9ec2cb6822f5d8448a73c0a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_2/index_tensor([2294, 2294])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly advancing tracking shot, tilted slightly upward, draws the viewer into a bottomless, dark universe or mysterious deep sea, surrounded by a shimmering, ethereal aura. At the center of the frame, a mountain-sized sea anemone buds, its closed form resembling the heart of a sleeping star. Its surface possesses a translucent jade texture, and a soft glow pulsates faintly within. As the camera steadily approaches, the giant anemone suddenly and suddenly blossoms, imbued with the grand, fantastical aesthetic of The Legend of Shu: countless crystal-clear tentacles gracefully and swiftly unfurl, each emanating a rainbow spectrum, smoothly transitioning from deep indigo and mysterious violet to brilliant turquoise and warm gold, its surface shimmering with stardust-like specks of light. These radiant tentacles danced in all directions as smoothly as silk, stirring up the surrounding immortal energy and forming a dreamy vortex of light and shadow. The sacred light emanating from its core illuminated the void, creating a solemn, vast and fantastical atmosphere full of miracles of life.", + "seed": 3872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_2/index_2294_seed_3872_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03653611897025257, + "vlm_logprob": -0.03653611897025257, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.03653611897025257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10693453252315521 + ], + "opposite_logprob_tokens": [ + -2.3569345474243164 + ], + "logprob_reward": -0.10693453252315521, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10693453252315521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2890385687351227, + 0.10411214083433151, + 0.05303199589252472, + 0.332882285118103, + 0.04583657532930374, + 0.027883408591151237, + 0.037894200533628464, + 0.019227419048547745, + 0.029776085168123245, + 0.06031734123826027 + ], + "credit_map": [ + 0.02894899807870388, + 0.019598089158535004, + 0.0895688384771347, + 0.014174862764775753, + 0.028384314849972725, + 0.014962658286094666, + 0.02011270262300968, + 0.07328811287879944, + 0.008188963867723942, + 0.011246315203607082, + 0.015715081244707108, + 0.009647678583860397, + 0.018794303759932518, + 0.012476514093577862, + 0.012619138695299625, + 0.015424147248268127, + 0.006141417194157839, + 0.007996638305485249, + 0.008503805845975876, + 0.006053023971617222, + 0.006126325111836195, + 0.007587294559925795, + 0.005719351582229137, + 0.00490414397791028, + 0.025172226130962372, + 0.014491920359432697, + 0.1161503717303276, + 0.012033888138830662, + 0.012658073566854, + 0.11921083182096481, + 0.011383445002138615, + 0.021781526505947113, + 0.0034687931183725595, + 0.005614671390503645, + 0.0061332592740654945, + 0.006316513754427433, + 0.012294420041143894, + 0.004212646279484034, + 0.004513646475970745, + 0.0032826231326907873, + 0.0025374828837811947, + 0.005086266435682774, + 0.0020844906102865934, + 0.004161162301898003, + 0.0036285242531448603, + 0.004147633910179138, + 0.0029119914397597313, + 0.003325855126604438, + 0.00400546146556735, + 0.005953560583293438, + 0.002108878456056118, + 0.00457249628379941, + 0.005429218988865614, + 0.008234675973653793, + 0.0036192687693983316, + 0.003970638383179903, + 0.002655668882653117, + 0.0025795253459364176, + 0.00177224341314286, + 0.0021826850716024637, + 0.0036098510026931763, + 0.0024053279776126146, + 0.002077993471175432, + 0.001944124000146985, + 0.004143080674111843, + 0.004206942860037088, + 0.0031294890213757753, + 0.003146060975268483, + 0.004797372035682201, + 0.002565124537795782, + 0.004147404804825783, + 0.0036406121216714382, + 0.005264196544885635, + 0.007946004159748554, + 0.004224422387778759, + 0.0050695487298071384, + 0.015221689827740192, + 0.007789223454892635, + 0.009330413304269314, + 0.005471843760460615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.10693453252315521 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550774961709976 + ], + "opposite_logprob_tokens": [ + -4.3875508308410645 + ], + "logprob_reward": -0.012550774961709976, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550774961709976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11604534089565277, + 0.0649120956659317, + 0.07617367804050446, + 0.5269677639007568, + 0.07107161730527878, + 0.0313744880259037, + 0.04653412476181984, + 0.01237882487475872, + 0.028788018971681595, + 0.025754012167453766 + ], + "credit_map": [ + 0.011119947768747807, + 0.0065142386592924595, + 0.036158591508865356, + 0.00637707719579339, + 0.008587568998336792, + 0.005590411368757486, + 0.01137448288500309, + 0.030323022976517677, + 0.005651301704347134, + 0.004523997660726309, + 0.013285279273986816, + 0.005955567117780447, + 0.006195030175149441, + 0.005191324278712273, + 0.008946525864303112, + 0.01516307145357132, + 0.010574554093182087, + 0.0049402001313865185, + 0.024466702714562416, + 0.004133737646043301, + 0.006433914881199598, + 0.006088819354772568, + 0.004226784221827984, + 0.015308966860175133, + 0.03746861591935158, + 0.025701358914375305, + 0.17219989001750946, + 0.02123778499662876, + 0.018204281106591225, + 0.19723837077617645, + 0.01875481754541397, + 0.036162640899419785, + 0.005411633290350437, + 0.008385931141674519, + 0.006900890730321407, + 0.01867636851966381, + 0.011488187126815319, + 0.0063992817886173725, + 0.009263948537409306, + 0.004545372445136309, + 0.0030132082756608725, + 0.006005643401294947, + 0.0018579510506242514, + 0.004419286735355854, + 0.005936033092439175, + 0.0028047380037605762, + 0.003845458384603262, + 0.00349217071197927, + 0.00619989400729537, + 0.00819664727896452, + 0.003359529422596097, + 0.004020281136035919, + 0.009641503915190697, + 0.0049286289140582085, + 0.0036652921698987484, + 0.006522348150610924, + 0.0015808038879185915, + 0.001574025140143931, + 0.0010108130518347025, + 0.001232800423167646, + 0.0025018765591084957, + 0.0020304073113948107, + 0.0011570273200049996, + 0.0012910709483548999, + 0.004362151958048344, + 0.0041957697831094265, + 0.003559251083061099, + 0.00256710359826684, + 0.004900080151855946, + 0.003114345483481884, + 0.002938749734312296, + 0.0031505662482231855, + 0.002913437783718109, + 0.00412162346765399, + 0.0016218872042372823, + 0.0026036493945866823, + 0.005130702629685402, + 0.005213922820985317, + 0.002509387908503413, + 0.001639399561099708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.012550774961709976 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001520192832686007 + ], + "opposite_logprob_tokens": [ + -6.501520156860352 + ], + "logprob_reward": -0.001520192832686007, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001520192832686007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07068738341331482, + 0.1834939867258072, + 0.10795065015554428, + 0.2504037618637085, + 0.032641153782606125, + 0.0802423283457756, + 0.0790715366601944, + 0.06463989615440369, + 0.033556919544935226, + 0.09731236845254898 + ], + "credit_map": [ + 0.007061900570988655, + 0.006674343254417181, + 0.01196646224707365, + 0.005806060507893562, + 0.010901765897870064, + 0.007068513426929712, + 0.01122791413217783, + 0.009980420581996441, + 0.012175026349723339, + 0.01492578536272049, + 0.04131273925304413, + 0.01224622130393982, + 0.027392489835619926, + 0.014873378910124302, + 0.014982214197516441, + 0.04558612033724785, + 0.014711637049913406, + 0.009383692406117916, + 0.03369760140776634, + 0.005698163527995348, + 0.009184347465634346, + 0.006320452783256769, + 0.007771215867251158, + 0.021183544769883156, + 0.018991651013493538, + 0.010169082321226597, + 0.0855022519826889, + 0.008906304836273193, + 0.010929698124527931, + 0.0906560942530632, + 0.009087472222745419, + 0.01616123504936695, + 0.0037290637847036123, + 0.00427329633384943, + 0.0034261078108102083, + 0.003979096654802561, + 0.007004630286246538, + 0.004244277719408274, + 0.003383347298949957, + 0.0026013320311903954, + 0.009151680395007133, + 0.010512158274650574, + 0.0060719335451722145, + 0.015967609360814095, + 0.014066165313124657, + 0.00894917268306017, + 0.008597999811172485, + 0.006925616879016161, + 0.0072874329052865505, + 0.019358500838279724, + 0.004113106522709131, + 0.008481833152472973, + 0.01689959689974785, + 0.008672431111335754, + 0.00801645964384079, + 0.006242176052182913, + 0.0054324036464095116, + 0.013531403616070747, + 0.003980559296905994, + 0.005111322738230228, + 0.01504401583224535, + 0.007275585550814867, + 0.00846902932971716, + 0.005795573350042105, + 0.004739709198474884, + 0.0052851526997983456, + 0.002690050518140197, + 0.003502187319099903, + 0.003872602479532361, + 0.0042092641815543175, + 0.003432184923440218, + 0.00582577008754015, + 0.010713286697864532, + 0.01610257662832737, + 0.006515911314636469, + 0.009173020720481873, + 0.02186433970928192, + 0.01621789112687111, + 0.01112997904419899, + 0.005595361348241568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.001520192832686007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739978212863207 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006739978212863207, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739978212863207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24278953671455383, + 0.06002623215317726, + 0.06620653718709946, + 0.21554730832576752, + 0.07350236922502518, + 0.10328013449907303, + 0.06421978026628494, + 0.03239988535642624, + 0.09994852542877197, + 0.04207969829440117 + ], + "credit_map": [ + 0.022883575409650803, + 0.0074395062401890755, + 0.08904924243688583, + 0.011077545583248138, + 0.010360893793404102, + 0.011921906843781471, + 0.01338996272534132, + 0.07666690647602081, + 0.004708771128207445, + 0.0033988088835030794, + 0.015100804157555103, + 0.004515551030635834, + 0.005950955208390951, + 0.003558632219210267, + 0.0047868299297988415, + 0.01800587773323059, + 0.006931563373655081, + 0.004532115068286657, + 0.017086371779441833, + 0.0037230653688311577, + 0.00862329825758934, + 0.006508287508040667, + 0.008257836103439331, + 0.010544005781412125, + 0.015761850401759148, + 0.026164578273892403, + 0.020369114354252815, + 0.022467685863375664, + 0.011848135851323605, + 0.0704827755689621, + 0.017029758542776108, + 0.03142339363694191, + 0.00568421371281147, + 0.007954269647598267, + 0.009208652190864086, + 0.017062976956367493, + 0.012163105420768261, + 0.008651889860630035, + 0.007795840501785278, + 0.004981422331184149, + 0.009907210245728493, + 0.02145932987332344, + 0.0054979813285171986, + 0.01729290746152401, + 0.022260578349232674, + 0.007248863112181425, + 0.011458320543169975, + 0.008154948242008686, + 0.0061565604992210865, + 0.01948634535074234, + 0.0029499002266675234, + 0.003575820242986083, + 0.0181948933750391, + 0.00476960651576519, + 0.0033806231804192066, + 0.005706028547137976, + 0.003003777004778385, + 0.0051312558352947235, + 0.002976643620058894, + 0.004448520485311747, + 0.008143981918692589, + 0.0029463290702551603, + 0.00373574229888618, + 0.0020136372186243534, + 0.010498383082449436, + 0.01895294524729252, + 0.008389794267714024, + 0.008885675109922886, + 0.012549752369523048, + 0.008971646428108215, + 0.014829505234956741, + 0.01687082275748253, + 0.005752600263804197, + 0.004950553644448519, + 0.004971481394022703, + 0.0035102511756122112, + 0.007201423868536949, + 0.004046075511723757, + 0.006195486057549715, + 0.005451827775686979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.006739978212863207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054935116320848465 + ], + "opposite_logprob_tokens": [ + -2.9299352169036865 + ], + "logprob_reward": -0.054935116320848465, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054935116320848465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2286558300256729, + 0.08982939273118973, + 0.07327604293823242, + 0.27930203080177307, + 0.07541633397340775, + 0.06681335717439651, + 0.08769341558218002, + 0.022030292078852654, + 0.06882736086845398, + 0.00815589353442192 + ], + "credit_map": [ + 0.020785707980394363, + 0.010415630415081978, + 0.0781177282333374, + 0.01094589289277792, + 0.012687812559306622, + 0.011927727609872818, + 0.017335157841444016, + 0.06644018739461899, + 0.006766930688172579, + 0.004788278602063656, + 0.024538282305002213, + 0.005992037709802389, + 0.005800197366625071, + 0.006498442497104406, + 0.007207415997982025, + 0.028237801045179367, + 0.010106871835887432, + 0.003895103232935071, + 0.02440527267754078, + 0.004191573243588209, + 0.004778486676514149, + 0.005993629805743694, + 0.005277553107589483, + 0.01462754886597395, + 0.019245943054556847, + 0.022562971338629723, + 0.06846366077661514, + 0.01620691455900669, + 0.01194706279784441, + 0.10145943611860275, + 0.015586148016154766, + 0.02382989600300789, + 0.005706797819584608, + 0.008596946485340595, + 0.012531968764960766, + 0.011395806446671486, + 0.014840693213045597, + 0.009009338915348053, + 0.007485899608582258, + 0.005848878063261509, + 0.005545012187212706, + 0.011955369263887405, + 0.004075688309967518, + 0.015631310641765594, + 0.009927823208272457, + 0.005304424092173576, + 0.008402065373957157, + 0.0059716603718698025, + 0.006169595755636692, + 0.028344834223389626, + 0.0033054447267204523, + 0.005219300277531147, + 0.025416603311896324, + 0.007301487028598785, + 0.004837706685066223, + 0.0070984382182359695, + 0.002040362684056163, + 0.003629703540354967, + 0.0016236166702583432, + 0.0022410417441278696, + 0.00616929866373539, + 0.0023791808634996414, + 0.002471909625455737, + 0.001475178636610508, + 0.009969227947294712, + 0.01112352404743433, + 0.008914691396057606, + 0.0072038015350699425, + 0.010680150240659714, + 0.007173215504735708, + 0.00716958474367857, + 0.006593166384845972, + 0.0010621603578329086, + 0.0013925002422183752, + 0.000705794314853847, + 0.0008351205615326762, + 0.0014672246761620045, + 0.0008619988220743835, + 0.0009593017748557031, + 0.0008717917953617871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.054935116320848465 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing tracking shot, tilted slightly upward, draws the viewer into a bottomless, dark universe or mysterious deep sea, surrounded by a shimmering, ethereal aura. At the center of the frame, a mountain-sized sea anemone buds, its closed form resembling the heart of a sleeping star. Its surface possesses a translucent jade texture, and a soft glow pulsates faintly within. As the camera steadily approaches, the giant anemone suddenly and suddenly blossoms, imbued with the grand, fantastical aesthetic of The Legend of Shu: countless crystal-clear tentacles gracefully and swiftly unfurl, each emanating a rainbow spectrum, smoothly transitioning from deep indigo and mysterious violet to brilliant turquoise and warm gold, its surface shimmering with stardust-like specks of light. These radiant tentacles danced in all directions as smoothly as silk, stirring up the surrounding immortal energy and forming a dreamy vortex of light and shadow. The sacred light emanating from its core illuminated the void, creating a solemn, vast and fantastical atmosphere full of miracles of life.", + "seed": 3873, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_2/index_2294_seed_3873_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0288400087505579, + "vlm_logprob": -0.0288400087505579, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.0288400087505579, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03505170717835426 + ], + "opposite_logprob_tokens": [ + -3.5350518226623535 + ], + "logprob_reward": -0.03505170717835426, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03505170717835426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23954062163829803, + 0.09941955655813217, + 0.05687868967652321, + 0.11993668228387833, + 0.02860400825738907, + 0.19228972494602203, + 0.08225180208683014, + 0.07093241065740585, + 0.07502271234989166, + 0.03512372821569443 + ], + "credit_map": [ + 0.020050005987286568, + 0.037794142961502075, + 0.026263173669576645, + 0.0246739462018013, + 0.03655022382736206, + 0.03109484352171421, + 0.02643098495900631, + 0.036683302372694016, + 0.008864050731062889, + 0.014838624745607376, + 0.01854236237704754, + 0.008421453647315502, + 0.008736046962440014, + 0.014590623788535595, + 0.007634449750185013, + 0.01779194362461567, + 0.004610019735991955, + 0.006377921439707279, + 0.009865469299256802, + 0.004299885127693415, + 0.003726034890860319, + 0.00938965193927288, + 0.003959171008318663, + 0.014650538563728333, + 0.014040257781744003, + 0.017943767830729485, + 0.017930636182427406, + 0.009837795980274677, + 0.010046305134892464, + 0.01920785754919052, + 0.007895412854850292, + 0.023034634068608284, + 0.0030570155940949917, + 0.0036104810424149036, + 0.0030292728915810585, + 0.0027590913232415915, + 0.002685176907107234, + 0.005407219752669334, + 0.003457109211012721, + 0.0045986417680978775, + 0.014961449429392815, + 0.009776146151125431, + 0.015571522526443005, + 0.010907355695962906, + 0.009121481329202652, + 0.07346393167972565, + 0.010784068144857883, + 0.047703783959150314, + 0.006015400867909193, + 0.015023529529571533, + 0.004911503754556179, + 0.006333222612738609, + 0.02703738957643509, + 0.01152193546295166, + 0.005791420117020607, + 0.00561739644035697, + 0.005766970105469227, + 0.012695387937128544, + 0.006618062499910593, + 0.011536426842212677, + 0.01220961194485426, + 0.007561072241514921, + 0.0071098641492426395, + 0.007435016334056854, + 0.006272526457905769, + 0.009237581863999367, + 0.010690038092434406, + 0.008793446235358715, + 0.011807316914200783, + 0.008274644613265991, + 0.009722895920276642, + 0.01022426038980484, + 0.0024058478884398937, + 0.005538363009691238, + 0.00255110883153975, + 0.004586934577673674, + 0.00904476922005415, + 0.002856347244232893, + 0.004395214840769768, + 0.003745140740647912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.03505170717835426 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619787283241749 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619787283241749, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619787283241749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10877049714326859, + 0.05710313469171524, + 0.12606099247932434, + 0.10389669984579086, + 0.05710456147789955, + 0.14798769354820251, + 0.12096311151981354, + 0.10585580766201019, + 0.12977638840675354, + 0.04248105362057686 + ], + "credit_map": [ + 0.016211939975619316, + 0.011908327229321003, + 0.012205258011817932, + 0.010366188362240791, + 0.013699248433113098, + 0.01484514120966196, + 0.009245435707271099, + 0.020288962870836258, + 0.0033538215793669224, + 0.008762460201978683, + 0.008318841457366943, + 0.004168177023530006, + 0.006727586500346661, + 0.008761712349951267, + 0.0034370499197393656, + 0.013573483563959599, + 0.010051186196506023, + 0.023798486217856407, + 0.01120449136942625, + 0.017394786700606346, + 0.015608888119459152, + 0.015153863467276096, + 0.00971477385610342, + 0.023134516552090645, + 0.012862710282206535, + 0.0155604537576437, + 0.015039862133562565, + 0.009917468763887882, + 0.010211262851953506, + 0.014621851034462452, + 0.008102616295218468, + 0.01758047379553318, + 0.0051770624704658985, + 0.00746484799310565, + 0.008869515731930733, + 0.006811304017901421, + 0.006707018706947565, + 0.006474282592535019, + 0.0061766901053488255, + 0.009423838928341866, + 0.012694623321294785, + 0.010731526650488377, + 0.02890876866877079, + 0.011298037134110928, + 0.006883721332997084, + 0.032744355499744415, + 0.01259501464664936, + 0.03213164582848549, + 0.010856605134904385, + 0.01342390850186348, + 0.013945743441581726, + 0.01828388124704361, + 0.01836138404905796, + 0.010404398664832115, + 0.017329612746834755, + 0.01835758052766323, + 0.0077299317345023155, + 0.023020990192890167, + 0.010243133641779423, + 0.018058624118566513, + 0.01804528757929802, + 0.009240242652595043, + 0.00980969425290823, + 0.009707903489470482, + 0.010073844343423843, + 0.021398259326815605, + 0.012289161793887615, + 0.01983438991010189, + 0.02998299151659012, + 0.01275861356407404, + 0.013284672051668167, + 0.01015445776283741, + 0.003250966314226389, + 0.0050961971282958984, + 0.005787068512290716, + 0.005130672361701727, + 0.007308154366910458, + 0.005188451148569584, + 0.005010210443288088, + 0.005709331948310137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.003619787283241749 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739504635334015 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739504635334015, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739504635334015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1410398781299591, + 0.07984551042318344, + 0.06125233694911003, + 0.14840950071811676, + 0.034506190568208694, + 0.20106977224349976, + 0.07700171321630478, + 0.09655524045228958, + 0.06448252499103546, + 0.09583730250597 + ], + "credit_map": [ + 0.015245793387293816, + 0.024917837232351303, + 0.01409753505140543, + 0.012274830602109432, + 0.019793406128883362, + 0.01667807251214981, + 0.01860121637582779, + 0.019431181252002716, + 0.0057661812752485275, + 0.01099823135882616, + 0.012384979985654354, + 0.007504122797399759, + 0.007377962116152048, + 0.011919536627829075, + 0.0065774694085121155, + 0.0173170268535614, + 0.005040041171014309, + 0.00888562947511673, + 0.009828407317399979, + 0.00483781797811389, + 0.005226236302405596, + 0.0094824293628335, + 0.003876974806189537, + 0.01407480426132679, + 0.017743684351444244, + 0.012401577085256577, + 0.0169034693390131, + 0.014475037343800068, + 0.01468177605420351, + 0.03551269695162773, + 0.013087584637105465, + 0.02360367774963379, + 0.0029890136793255806, + 0.003983893897384405, + 0.0072173550724983215, + 0.003528001718223095, + 0.0035428639966994524, + 0.005640463903546333, + 0.0029644707683473825, + 0.004640128929167986, + 0.015674611553549767, + 0.012724190019071102, + 0.016084380447864532, + 0.010785438120365143, + 0.010996314696967602, + 0.0775311216711998, + 0.007427003234624863, + 0.04984671249985695, + 0.004560513887554407, + 0.015479116700589657, + 0.004875162150710821, + 0.006937254220247269, + 0.026307465508580208, + 0.009031970985233784, + 0.004850012715905905, + 0.0049602193757891655, + 0.008493735454976559, + 0.014765189960598946, + 0.00941503420472145, + 0.011290385387837887, + 0.017764192074537277, + 0.012313428334891796, + 0.009876471012830734, + 0.01263680960983038, + 0.005616217851638794, + 0.009570456109941006, + 0.008666685782372952, + 0.006786472164094448, + 0.013233600184321404, + 0.006958933547139168, + 0.006944555789232254, + 0.006705599837005138, + 0.006566873285919428, + 0.010540787130594254, + 0.010021806694567204, + 0.012406360357999802, + 0.019831638783216476, + 0.010847273282706738, + 0.014010696671903133, + 0.011611866764724255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.006739504635334015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783060289919376 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783060289919376, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783060289919376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1593305617570877, + 0.07522325217723846, + 0.06919533014297485, + 0.3274063169956207, + 0.03306359052658081, + 0.05701884627342224, + 0.09959225356578827, + 0.06759683042764664, + 0.08323229104280472, + 0.028340689837932587 + ], + "credit_map": [ + 0.030786840245127678, + 0.01738480105996132, + 0.02163228951394558, + 0.011883295141160488, + 0.0175704937428236, + 0.022435346618294716, + 0.01218392699956894, + 0.025453578680753708, + 0.004949034191668034, + 0.007744042202830315, + 0.012626126408576965, + 0.009367258287966251, + 0.006294426042586565, + 0.010534076020121574, + 0.009007042273879051, + 0.014701247215270996, + 0.007632504682987928, + 0.006673439405858517, + 0.010042792186141014, + 0.006279746536165476, + 0.006408736575394869, + 0.009537252597510815, + 0.005591998342424631, + 0.017028862610459328, + 0.03780937194824219, + 0.04977115988731384, + 0.05011161044239998, + 0.0281232837587595, + 0.029880737885832787, + 0.04685986042022705, + 0.021844148635864258, + 0.06300614774227142, + 0.003253389149904251, + 0.003921319264918566, + 0.0034693845082074404, + 0.0037152650766074657, + 0.0028762395959347486, + 0.008298140950500965, + 0.003015333553776145, + 0.0045145172625780106, + 0.004906123038381338, + 0.005845837295055389, + 0.009102017618715763, + 0.006411456968635321, + 0.004771427717059851, + 0.009714730083942413, + 0.006833986844867468, + 0.009433263912796974, + 0.007044192403554916, + 0.0170381311327219, + 0.009147313423454762, + 0.00917783286422491, + 0.028891155496239662, + 0.0062922327779233456, + 0.008978812023997307, + 0.013022584840655327, + 0.005825148429721594, + 0.013026965782046318, + 0.007384844124317169, + 0.010556651279330254, + 0.010998198762536049, + 0.00601916154846549, + 0.006589590106159449, + 0.0071962676011025906, + 0.007330188527703285, + 0.010064165107905865, + 0.009262488223612309, + 0.011928552761673927, + 0.013967417180538177, + 0.012609900906682014, + 0.008289720863103867, + 0.009779859334230423, + 0.002326719230040908, + 0.003845510771498084, + 0.002506506396457553, + 0.003284525591880083, + 0.005758421495556831, + 0.0033912525977939367, + 0.0036194627173244953, + 0.003608289174735546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.009783060289919376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890059843659401 + ], + "opposite_logprob_tokens": [ + -2.464005947113037 + ], + "logprob_reward": -0.0890059843659401, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890059843659401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05547740310430527, + 0.031628742814064026, + 0.02135990560054779, + 0.18922212719917297, + 0.057148277759552, + 0.2624073028564453, + 0.2822469174861908, + 0.02753947675228119, + 0.049846842885017395, + 0.02312302589416504 + ], + "credit_map": [ + 0.006427670828998089, + 0.005822600796818733, + 0.007658455520868301, + 0.0048711118288338184, + 0.00866423174738884, + 0.007022865116596222, + 0.004261454101651907, + 0.01074900571256876, + 0.003226796630769968, + 0.0035661691799759865, + 0.005731847137212753, + 0.002407749881967902, + 0.0027674599550664425, + 0.004923237022012472, + 0.0025513898581266403, + 0.0064540887251496315, + 0.0016843994380906224, + 0.002688422566279769, + 0.0033864397555589676, + 0.0028971561696380377, + 0.0017577973194420338, + 0.003054959699511528, + 0.0021686141844838858, + 0.0037221135571599007, + 0.021582890301942825, + 0.02581753022968769, + 0.028952304273843765, + 0.015646861866116524, + 0.01598259061574936, + 0.03218280151486397, + 0.012795459479093552, + 0.0362616702914238, + 0.005553107708692551, + 0.00780839566141367, + 0.0065420218743383884, + 0.0074525014497339725, + 0.005899046082049608, + 0.00751890242099762, + 0.006758594885468483, + 0.0096157006919384, + 0.020045749843120575, + 0.01507810689508915, + 0.03803851827979088, + 0.01527358964085579, + 0.012120692990720272, + 0.08193950355052948, + 0.015138243325054646, + 0.06477287411689758, + 0.01439784187823534, + 0.04295258969068527, + 0.010679444298148155, + 0.016352970153093338, + 0.08767082542181015, + 0.06606447696685791, + 0.01243062037974596, + 0.03169811889529228, + 0.001629852456972003, + 0.004262958653271198, + 0.0031990515999495983, + 0.0051955594681203365, + 0.004480777308344841, + 0.002635751385241747, + 0.0037623655516654253, + 0.0023731582332402468, + 0.00393409188836813, + 0.008148317225277424, + 0.005793818738311529, + 0.006978246383368969, + 0.008362647145986557, + 0.005370571278035641, + 0.0064275674521923065, + 0.004831573460251093, + 0.0018960790475830436, + 0.0029449686408042908, + 0.0026329217944294214, + 0.003096893895417452, + 0.004630898125469685, + 0.0018061127047985792, + 0.0029157602693885565, + 0.003199388738721609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.0890059843659401 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing tracking shot, tilted slightly upward, draws the viewer into a bottomless, dark universe or mysterious deep sea, surrounded by a shimmering, ethereal aura. At the center of the frame, a mountain-sized sea anemone buds, its closed form resembling the heart of a sleeping star. Its surface possesses a translucent jade texture, and a soft glow pulsates faintly within. As the camera steadily approaches, the giant anemone suddenly and suddenly blossoms, imbued with the grand, fantastical aesthetic of The Legend of Shu: countless crystal-clear tentacles gracefully and swiftly unfurl, each emanating a rainbow spectrum, smoothly transitioning from deep indigo and mysterious violet to brilliant turquoise and warm gold, its surface shimmering with stardust-like specks of light. These radiant tentacles danced in all directions as smoothly as silk, stirring up the surrounding immortal energy and forming a dreamy vortex of light and shadow. The sacred light emanating from its core illuminated the void, creating a solemn, vast and fantastical atmosphere full of miracles of life.", + "seed": 3874, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_2/index_2294_seed_3874_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018572296341881157, + "vlm_logprob": -0.018572296341881157, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.018572296341881157, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01476162951439619 + ], + "opposite_logprob_tokens": [ + -4.389761447906494 + ], + "logprob_reward": -0.01476162951439619, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01476162951439619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13334180414676666, + 0.04509762302041054, + 0.06382717937231064, + 0.09334059804677963, + 0.3806701600551605, + 0.04612583667039871, + 0.04470443353056908, + 0.037166811525821686, + 0.10319839417934418, + 0.05252719298005104 + ], + "credit_map": [ + 0.01370101235806942, + 0.016834747046232224, + 0.012875551357865334, + 0.016031283885240555, + 0.019580861553549767, + 0.013275139033794403, + 0.02373450994491577, + 0.01730869524180889, + 0.003418481443077326, + 0.005914179142564535, + 0.0029559768736362457, + 0.004140953533351421, + 0.004512915387749672, + 0.010226416401565075, + 0.007668095175176859, + 0.006260606460273266, + 0.004819066263735294, + 0.008134208619594574, + 0.007710450794547796, + 0.008875423111021519, + 0.011464656330645084, + 0.008595786057412624, + 0.00727608660236001, + 0.006951503921300173, + 0.009486525319516659, + 0.014899971894919872, + 0.013659135438501835, + 0.014833367429673672, + 0.01313852146267891, + 0.008268805220723152, + 0.011940008960664272, + 0.007114264648407698, + 0.028374511748552322, + 0.01869598589837551, + 0.1549292355775833, + 0.11660008132457733, + 0.01791948452591896, + 0.012344221584498882, + 0.018237223848700523, + 0.01356938760727644, + 0.004082667175680399, + 0.005503946449607611, + 0.005902951583266258, + 0.005503493826836348, + 0.008745302446186543, + 0.005744424182921648, + 0.00684075802564621, + 0.0038022908847779036, + 0.0038405624218285084, + 0.004665242042392492, + 0.005371968727558851, + 0.006657741032540798, + 0.006360106635838747, + 0.0045328205451369286, + 0.007309150882065296, + 0.005966844502836466, + 0.0024524268228560686, + 0.0029279584996402264, + 0.006324982736259699, + 0.005199162755161524, + 0.00449320487678051, + 0.003932000603526831, + 0.00727320509031415, + 0.004563869908452034, + 0.008041728287935257, + 0.010784018784761429, + 0.013669492676854134, + 0.013038796372711658, + 0.013519790023565292, + 0.009079452604055405, + 0.01943589560687542, + 0.01562921702861786, + 0.004331235773861408, + 0.006897719111293554, + 0.009217551909387112, + 0.007190897595137358, + 0.00693786283954978, + 0.0043853106908500195, + 0.005752908065915108, + 0.007813706994056702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.01476162951439619 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097163211554289 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004097163211554289, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097163211554289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05570834130048752, + 0.06683169305324554, + 0.0939989984035492, + 0.02971068024635315, + 0.5535046458244324, + 0.10075443238019943, + 0.030723825097084045, + 0.02168496884405613, + 0.028184790164232254, + 0.018897652626037598 + ], + "credit_map": [ + 0.004869399126619101, + 0.005781759042292833, + 0.005547350738197565, + 0.00755402073264122, + 0.00940782856196165, + 0.008041712455451488, + 0.005929356906563044, + 0.008576908148825169, + 0.0062840403988957405, + 0.009278562851250172, + 0.006216054782271385, + 0.008436361327767372, + 0.01126482430845499, + 0.0057121035642921925, + 0.012102952226996422, + 0.007536787539720535, + 0.007104227785021067, + 0.010693087242543697, + 0.006889526732265949, + 0.007677404675632715, + 0.03296877071261406, + 0.007791381329298019, + 0.014847583137452602, + 0.006027010269463062, + 0.0023296652361750603, + 0.004091889597475529, + 0.00262659415602684, + 0.006411545444279909, + 0.0036529614590108395, + 0.0033591969404369593, + 0.004121195990592241, + 0.0031176316551864147, + 0.042630717158317566, + 0.02535696141421795, + 0.22563236951828003, + 0.18148702383041382, + 0.01958230510354042, + 0.016749273985624313, + 0.022976575419306755, + 0.019089389592409134, + 0.01154040452092886, + 0.01081036776304245, + 0.011443015187978745, + 0.011923267506062984, + 0.01677476055920124, + 0.009952682070434093, + 0.01781509444117546, + 0.010494837537407875, + 0.0028489730320870876, + 0.0036904714070260525, + 0.00428602984175086, + 0.004946134518831968, + 0.0031520333141088486, + 0.0028458177112042904, + 0.004897772800177336, + 0.004056592471897602, + 0.0016874022549018264, + 0.0014848127029836178, + 0.0024288902059197426, + 0.002848969306796789, + 0.002843053312972188, + 0.0026558630634099245, + 0.004160239361226559, + 0.003575735492631793, + 0.002997184870764613, + 0.003124288283288479, + 0.004270665813237429, + 0.0036673748400062323, + 0.003423972986638546, + 0.0028803374152630568, + 0.004898385144770145, + 0.002922578016296029, + 0.0020605672616511583, + 0.002409154549241066, + 0.002768897684291005, + 0.002370350994169712, + 0.0025134929455816746, + 0.0015391201013699174, + 0.0025663617998361588, + 0.0026697053108364344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.004097163211554289 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022005646023899317 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.0022005646023899317, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022005646023899317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1039988100528717, + 0.02313457429409027, + 0.02824128046631813, + 0.0567905455827713, + 0.5581845641136169, + 0.02684861607849598, + 0.059238795191049576, + 0.03583384305238724, + 0.05082577466964722, + 0.05690324306488037 + ], + "credit_map": [ + 0.008570815436542034, + 0.014572042971849442, + 0.014463989995419979, + 0.01423153467476368, + 0.013196916319429874, + 0.010079389438033104, + 0.015500415116548538, + 0.013383697718381882, + 0.0026143547147512436, + 0.0040572634898126125, + 0.0025076158344745636, + 0.0029819912742823362, + 0.0025460205506533384, + 0.002466590143740177, + 0.003008544212207198, + 0.0029521912802010775, + 0.002285994589328766, + 0.004084596876055002, + 0.0036754130851477385, + 0.004212635103613138, + 0.0031246852595359087, + 0.004428913351148367, + 0.0033284742385149, + 0.0031005649361759424, + 0.006577590946108103, + 0.00790386088192463, + 0.008249253034591675, + 0.008937539532780647, + 0.007437824737280607, + 0.0057592084631323814, + 0.007431515026837587, + 0.0044937473721802235, + 0.04255591705441475, + 0.031139936298131943, + 0.21660609543323517, + 0.18460547924041748, + 0.02084968239068985, + 0.018114890903234482, + 0.023751402273774147, + 0.020561132580041885, + 0.003746806876733899, + 0.003262808546423912, + 0.004025849048048258, + 0.0030038391705602407, + 0.002748115686699748, + 0.0031791001092642546, + 0.0034868516959249973, + 0.003395242616534233, + 0.004229011479765177, + 0.005262917838990688, + 0.006997790653258562, + 0.0072196912951767445, + 0.005403121002018452, + 0.011564351618289948, + 0.011173827573657036, + 0.007388079538941383, + 0.0034338010009378195, + 0.003036494366824627, + 0.0050742290914058685, + 0.003887087805196643, + 0.0046259742230176926, + 0.006580782122910023, + 0.005178398918360472, + 0.0040170722641050816, + 0.0034445191267877817, + 0.005302132572978735, + 0.005365455523133278, + 0.005782062187790871, + 0.011947279796004295, + 0.004031190183013678, + 0.010042110458016396, + 0.004911019466817379, + 0.005288834683597088, + 0.008341261185705662, + 0.006472056265920401, + 0.0060613080859184265, + 0.009362083859741688, + 0.006500985007733107, + 0.007695604115724564, + 0.007181102875620127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.0022005646023899317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781289845705032 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.009781289845705032, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781289845705032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057512324303388596, + 0.02606748230755329, + 0.05807610601186752, + 0.20023974776268005, + 0.4057880938053131, + 0.04473092779517174, + 0.054144639521837234, + 0.024316994473338127, + 0.08037733286619186, + 0.04874630644917488 + ], + "credit_map": [ + 0.008056516759097576, + 0.008036114275455475, + 0.005726046394556761, + 0.0074043250642716885, + 0.006546703167259693, + 0.007657600101083517, + 0.007210236508399248, + 0.0068747797049582005, + 0.002399875083938241, + 0.004584718029946089, + 0.0023115091025829315, + 0.003077146830037236, + 0.0036387802101671696, + 0.004078931175172329, + 0.0032764344941824675, + 0.002700089244171977, + 0.0037129472475498915, + 0.008207160048186779, + 0.008127447217702866, + 0.013250813819468021, + 0.006255702115595341, + 0.005869904533028603, + 0.007091083563864231, + 0.005561047699302435, + 0.020619846880435944, + 0.026644837111234665, + 0.024235814809799194, + 0.021144287660717964, + 0.03781250864267349, + 0.026758598163723946, + 0.027133045718073845, + 0.0158908162266016, + 0.030908675864338875, + 0.020399857312440872, + 0.16770756244659424, + 0.12403038144111633, + 0.01589384116232395, + 0.013016624376177788, + 0.01930396631360054, + 0.014527186751365662, + 0.005387142766267061, + 0.0048646084032952785, + 0.005680963397026062, + 0.004977671429514885, + 0.006171168759465218, + 0.004967830143868923, + 0.006599206477403641, + 0.0060823336243629456, + 0.00540184834972024, + 0.004945836961269379, + 0.006612251978367567, + 0.006413220427930355, + 0.006877965293824673, + 0.008319118991494179, + 0.007435904815793037, + 0.00813849363476038, + 0.0020585483871400356, + 0.0023306922521442175, + 0.0036317924968898296, + 0.002876227954402566, + 0.0036231493577361107, + 0.003987371921539307, + 0.003048722166568041, + 0.0027604899369180202, + 0.006225398741662502, + 0.009132783859968185, + 0.008222945965826511, + 0.008092994801700115, + 0.013851472176611423, + 0.008973990567028522, + 0.011649861931800842, + 0.014227880164980888, + 0.005683703813701868, + 0.0047691501677036285, + 0.007277241442352533, + 0.005735231097787619, + 0.006707453168928623, + 0.005876602604985237, + 0.005484810099005699, + 0.007212119176983833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.009781289845705032 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062020834535360336 + ], + "opposite_logprob_tokens": [ + -2.812020778656006 + ], + "logprob_reward": -0.062020834535360336, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062020834535360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054524779319763184, + 0.038606900721788406, + 0.03294629231095314, + 0.03050363063812256, + 0.6730031967163086, + 0.03568090870976448, + 0.05097309127449989, + 0.029790140688419342, + 0.019970862194895744, + 0.03400013595819473 + ], + "credit_map": [ + 0.007035117596387863, + 0.007043205667287111, + 0.007992314174771309, + 0.005895585753023624, + 0.005040029529482126, + 0.006962253246456385, + 0.005348432343453169, + 0.009207844734191895, + 0.002927261870354414, + 0.006022985093295574, + 0.0038969458546489477, + 0.00491772685199976, + 0.00366347748786211, + 0.005800311919301748, + 0.006488994229584932, + 0.004889199044555426, + 0.0026865736581385136, + 0.0036532655358314514, + 0.0025911431293934584, + 0.0034570565912872553, + 0.009875522926449776, + 0.0030917776748538017, + 0.004875852260738611, + 0.0027151021640747786, + 0.0019584482070058584, + 0.0041826032102108, + 0.00320426351390779, + 0.006702035199850798, + 0.0039764889515936375, + 0.002815563464537263, + 0.004451781511306763, + 0.003212445415556431, + 0.051701679825782776, + 0.032256629317998886, + 0.27741384506225586, + 0.21419045329093933, + 0.023372935131192207, + 0.021153083071112633, + 0.029126416891813278, + 0.023788167163729668, + 0.004057334735989571, + 0.004168798215687275, + 0.006637990940362215, + 0.004370659124106169, + 0.004493724089115858, + 0.0037180606741458178, + 0.004737962037324905, + 0.003496378194540739, + 0.004261287860572338, + 0.004942307714372873, + 0.005012822337448597, + 0.006957865320146084, + 0.006734566763043404, + 0.012053068727254868, + 0.006414490286260843, + 0.00459667993709445, + 0.0021390446927398443, + 0.0028892355039715767, + 0.0038935584016144276, + 0.004230793099850416, + 0.005036344286054373, + 0.0033529438078403473, + 0.004687167704105377, + 0.0035610508639365435, + 0.0016830845270305872, + 0.001967682270333171, + 0.002557742642238736, + 0.0023037493228912354, + 0.003289716085419059, + 0.002151559107005596, + 0.002775955945253372, + 0.0032413718290627003, + 0.002533373422920704, + 0.004512258339673281, + 0.004736617207527161, + 0.004089105408638716, + 0.005118330009281635, + 0.0037632030434906483, + 0.0047439513728022575, + 0.00450329901650548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.062020834535360336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing tracking shot, tilted slightly upward, draws the viewer into a bottomless, dark universe or mysterious deep sea, surrounded by a shimmering, ethereal aura. At the center of the frame, a mountain-sized sea anemone buds, its closed form resembling the heart of a sleeping star. Its surface possesses a translucent jade texture, and a soft glow pulsates faintly within. As the camera steadily approaches, the giant anemone suddenly and suddenly blossoms, imbued with the grand, fantastical aesthetic of The Legend of Shu: countless crystal-clear tentacles gracefully and swiftly unfurl, each emanating a rainbow spectrum, smoothly transitioning from deep indigo and mysterious violet to brilliant turquoise and warm gold, its surface shimmering with stardust-like specks of light. These radiant tentacles danced in all directions as smoothly as silk, stirring up the surrounding immortal energy and forming a dreamy vortex of light and shadow. The sacred light emanating from its core illuminated the void, creating a solemn, vast and fantastical atmosphere full of miracles of life.", + "seed": 3875, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_2/index_2294_seed_3875_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023939373344182967, + "vlm_logprob": -0.023939373344182967, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.023939373344182967, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031606730073690414 + ], + "opposite_logprob_tokens": [ + -3.656606674194336 + ], + "logprob_reward": -0.031606730073690414, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.031606730073690414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05891148000955582, + 0.043162573128938675, + 0.20713040232658386, + 0.44772422313690186, + 0.01867910660803318, + 0.0906575471162796, + 0.02868484891951084, + 0.05848390981554985, + 0.02923686057329178, + 0.017329039052128792 + ], + "credit_map": [ + 0.004679091740399599, + 0.009732081554830074, + 0.00607925932854414, + 0.008314826525747776, + 0.00891852006316185, + 0.006633529905229807, + 0.010460927151143551, + 0.004093236289918423, + 0.004225593525916338, + 0.003625073004513979, + 0.00549793615937233, + 0.006094309501349926, + 0.006114994175732136, + 0.005936105269938707, + 0.006094264332205057, + 0.005574291571974754, + 0.031561627984046936, + 0.007578589953482151, + 0.007225667126476765, + 0.007456679828464985, + 0.00996982678771019, + 0.0824437364935875, + 0.008259843103587627, + 0.05263439938426018, + 0.03512174263596535, + 0.05908386409282684, + 0.042475391179323196, + 0.06638731807470322, + 0.05789905786514282, + 0.058015670627355576, + 0.08532853424549103, + 0.043412625789642334, + 0.0021888238843530416, + 0.0023566214367747307, + 0.002801730763167143, + 0.003147263778373599, + 0.0017976178787648678, + 0.002583913505077362, + 0.0018965484341606498, + 0.0019065836677327752, + 0.008459887467324734, + 0.006232115440070629, + 0.03195161744952202, + 0.006513987202197313, + 0.006525545846670866, + 0.016774067655205727, + 0.008000650443136692, + 0.006199663504958153, + 0.002632505027577281, + 0.002976037561893463, + 0.0039428058080375195, + 0.0039013498462736607, + 0.003040524898096919, + 0.004298735409975052, + 0.003598022973164916, + 0.004294865299016237, + 0.004731948021799326, + 0.01736663095653057, + 0.00671607768163085, + 0.0053314585238695145, + 0.010709481313824654, + 0.004316485952585936, + 0.00470554456114769, + 0.004606275819242001, + 0.002016924088820815, + 0.0031531467102468014, + 0.0032551540061831474, + 0.003272727131843567, + 0.004906992428004742, + 0.0033069956116378307, + 0.00519942631945014, + 0.004125491715967655, + 0.0012941481545567513, + 0.0027668671682476997, + 0.0024145469069480896, + 0.0020241697784513235, + 0.0025502003263682127, + 0.0019455390283837914, + 0.0018430354539304972, + 0.0024905314203351736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.031606730073690414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292848840355873 + ], + "opposite_logprob_tokens": [ + -3.7732927799224854 + ], + "logprob_reward": -0.023292848840355873, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292848840355873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07169870287179947, + 0.0661313384771347, + 0.11447571218013763, + 0.10592986643314362, + 0.0639052763581276, + 0.3144230246543884, + 0.08099071681499481, + 0.08769059181213379, + 0.07477902621030807, + 0.019975736737251282 + ], + "credit_map": [ + 0.0063097113743424416, + 0.011102595366537571, + 0.010040136985480785, + 0.009976583532989025, + 0.00737224705517292, + 0.008147119544446468, + 0.010374055244028568, + 0.008376256562769413, + 0.009002099744975567, + 0.006638344377279282, + 0.00758777791634202, + 0.009535716846585274, + 0.004761780146509409, + 0.012627735733985901, + 0.007427876815199852, + 0.008550005033612251, + 0.02397448569536209, + 0.004928840324282646, + 0.0048691267147660255, + 0.006091849412769079, + 0.005416386760771275, + 0.044110409915447235, + 0.004680450074374676, + 0.020404163748025894, + 0.0117136025801301, + 0.01186252199113369, + 0.013114572502672672, + 0.01404530368745327, + 0.013000125996768475, + 0.010069654323160648, + 0.020337197929620743, + 0.011786885559558868, + 0.004925350192934275, + 0.010751155205070972, + 0.00780487759038806, + 0.01097782887518406, + 0.006579707842320204, + 0.006811731494963169, + 0.008775287307798862, + 0.007279342971742153, + 0.034244854003190994, + 0.023690495640039444, + 0.11884103715419769, + 0.021552762016654015, + 0.01829630322754383, + 0.03316064551472664, + 0.029927678406238556, + 0.03470923751592636, + 0.007434772793203592, + 0.01270750630646944, + 0.020828183740377426, + 0.007008279673755169, + 0.008594506420195103, + 0.004664733074605465, + 0.013859457336366177, + 0.005893271882086992, + 0.007373327389359474, + 0.025999439880251884, + 0.0093698026612401, + 0.008005642332136631, + 0.016004569828510284, + 0.006727177649736404, + 0.006606155075132847, + 0.007604483049362898, + 0.00514827249571681, + 0.006155753042548895, + 0.009066860191524029, + 0.008726154454052448, + 0.011546730995178223, + 0.008144319988787174, + 0.013799622654914856, + 0.012191319838166237, + 0.0017413337482139468, + 0.0030458918772637844, + 0.0038382024504244328, + 0.0033258574549108744, + 0.002164896810427308, + 0.001784390420652926, + 0.0022437800653278828, + 0.0018313847249373794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.023292848840355873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036215689033269882 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036215689033269882, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036215689033269882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06346315145492554, + 0.02963910438120365, + 0.2540615499019623, + 0.05732700601220131, + 0.07886809855699539, + 0.30263593792915344, + 0.03973795101046562, + 0.09531030803918839, + 0.04874474182724953, + 0.03021210804581642 + ], + "credit_map": [ + 0.007816672325134277, + 0.005599861033260822, + 0.011944412253797054, + 0.008896438404917717, + 0.005545058753341436, + 0.008157762698829174, + 0.0063593569211661816, + 0.009143590927124023, + 0.003633965738117695, + 0.002481158822774887, + 0.0034991949796676636, + 0.004701301921159029, + 0.003528354223817587, + 0.004289984703063965, + 0.002966825384646654, + 0.004538319539278746, + 0.06716015189886093, + 0.010921045206487179, + 0.01113476138561964, + 0.009852925315499306, + 0.013087590225040913, + 0.10226623713970184, + 0.00777039397507906, + 0.03186846151947975, + 0.00465337885543704, + 0.007030999753624201, + 0.006204233504831791, + 0.00914714578539133, + 0.006610481068491936, + 0.007889221422374249, + 0.011859970167279243, + 0.003931577317416668, + 0.007761568762362003, + 0.00762210926041007, + 0.012352246791124344, + 0.012390568852424622, + 0.005396381486207247, + 0.009170018136501312, + 0.008854402229189873, + 0.01532080490142107, + 0.03171823173761368, + 0.027722375467419624, + 0.106906458735466, + 0.02056201919913292, + 0.01987469755113125, + 0.05901488661766052, + 0.01731713116168976, + 0.01952013000845909, + 0.004963997285813093, + 0.006163912825286388, + 0.004516843240708113, + 0.005286751314997673, + 0.005920655559748411, + 0.004436854738742113, + 0.004487284924834967, + 0.003961650654673576, + 0.008101136423647404, + 0.025859372690320015, + 0.01056588813662529, + 0.00876852497458458, + 0.0160929337143898, + 0.006956165656447411, + 0.008426251821219921, + 0.010540039278566837, + 0.0037253291811794043, + 0.007000901270657778, + 0.007256146054714918, + 0.006676069926470518, + 0.008693112060427666, + 0.004460384603589773, + 0.0052223908714950085, + 0.005710408557206392, + 0.0022253969218581915, + 0.0052190483547747135, + 0.003402263391762972, + 0.003395978594198823, + 0.004697924479842186, + 0.0034130041021853685, + 0.003283239435404539, + 0.004575252067297697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.0036215689033269882 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541946023702621 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541946023702621, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541946023702621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07394109666347504, + 0.02437029778957367, + 0.1718681901693344, + 0.1863917112350464, + 0.04312668740749359, + 0.18530040979385376, + 0.1590348184108734, + 0.08174129575490952, + 0.04084985703229904, + 0.03337565064430237 + ], + "credit_map": [ + 0.01015123538672924, + 0.008611248806118965, + 0.007877036929130554, + 0.00844558421522379, + 0.009845329448580742, + 0.009615635499358177, + 0.01102428324520588, + 0.008370744064450264, + 0.0032535286154597998, + 0.0024213334545493126, + 0.0028662821277976036, + 0.003407243639230728, + 0.002222242532297969, + 0.0034511983394622803, + 0.0033032321371138096, + 0.003445238107815385, + 0.026532389223575592, + 0.006466086953878403, + 0.005810635630041361, + 0.0071561397053301334, + 0.00872432254254818, + 0.06773040443658829, + 0.007503083907067776, + 0.04194512963294983, + 0.014789968729019165, + 0.024102196097373962, + 0.01820923574268818, + 0.02691478095948696, + 0.026532946154475212, + 0.024567609652876854, + 0.031141148880124092, + 0.020133811980485916, + 0.003109685843810439, + 0.00628307182341814, + 0.00633219163864851, + 0.008150727488100529, + 0.0048378994688391685, + 0.004423644859343767, + 0.004506728146225214, + 0.0054827420972287655, + 0.017455028370022774, + 0.011104202829301357, + 0.06298486888408661, + 0.011804469861090183, + 0.012945706956088543, + 0.039991866797208786, + 0.01501863170415163, + 0.013995630666613579, + 0.012815545313060284, + 0.029658442363142967, + 0.04016193747520447, + 0.01160420197993517, + 0.02053556591272354, + 0.009230907075107098, + 0.02611260488629341, + 0.008915618062019348, + 0.0057004583068192005, + 0.02372434176504612, + 0.009800505824387074, + 0.008561985567212105, + 0.014737945050001144, + 0.00575742544606328, + 0.008034156635403633, + 0.005424485541880131, + 0.0031388322822749615, + 0.006455857772380114, + 0.0048097786493599415, + 0.0055449227802455425, + 0.007402511313557625, + 0.0038337884470820427, + 0.004749325104057789, + 0.004914841614663601, + 0.0019707728642970324, + 0.0033555026166141033, + 0.003495091339573264, + 0.003367749275639653, + 0.0038128457963466644, + 0.003316420130431652, + 0.010748523287475109, + 0.0033087441697716713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.012541946023702621 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a mountain-sized sea anemone with a translucent jade texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the anemone start in a closed form resembling a sleeping star and then suddenly blossom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless crystal-clear tentacles unfurl from the anemone, displaying a rainbow spectrum from indigo/violet to turquoise/gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a bottomless dark universe or deep sea with a shimmering ethereal aura?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slowly advancing tracking shot that is tilted slightly upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048633772879838943 + ], + "opposite_logprob_tokens": [ + -3.0486338138580322 + ], + "logprob_reward": -0.048633772879838943, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048633772879838943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04081811010837555, + 0.03488278388977051, + 0.08638874441385269, + 0.10740502923727036, + 0.07615288347005844, + 0.13412025570869446, + 0.24209998548030853, + 0.1895047277212143, + 0.06958525627851486, + 0.01904219575226307 + ], + "credit_map": [ + 0.005065957084298134, + 0.004973331931978464, + 0.00500145461410284, + 0.0055334619246423244, + 0.0036451818887144327, + 0.004465154372155666, + 0.005391757469624281, + 0.006741810590028763, + 0.0029131779447197914, + 0.003696971805766225, + 0.0049756611697375774, + 0.004431481473147869, + 0.0039133112877607346, + 0.00414323853328824, + 0.005740043241530657, + 0.005068897735327482, + 0.012173627503216267, + 0.003273971378803253, + 0.0033764245454221964, + 0.004687611013650894, + 0.004495968110859394, + 0.03176390379667282, + 0.003907541278749704, + 0.02270970307290554, + 0.009659940376877785, + 0.014513807371258736, + 0.0098714055493474, + 0.01447032019495964, + 0.012932260520756245, + 0.016149956732988358, + 0.0189890768378973, + 0.01081826537847519, + 0.0053715636022388935, + 0.010971026495099068, + 0.009043695405125618, + 0.017241697758436203, + 0.007430123630911112, + 0.007037540432065725, + 0.009336931630969048, + 0.009720304980874062, + 0.010904118418693542, + 0.02179192192852497, + 0.02933717705309391, + 0.009189669974148273, + 0.013579939492046833, + 0.028029851615428925, + 0.00855228304862976, + 0.012735296040773392, + 0.015705103054642677, + 0.04086759313941002, + 0.06458014994859695, + 0.015630880370736122, + 0.023641962558031082, + 0.011540870182216167, + 0.05702982842922211, + 0.013103591278195381, + 0.014706047251820564, + 0.056940458714962006, + 0.021610131487250328, + 0.017295366153120995, + 0.03586422652006149, + 0.012482792139053345, + 0.016171598806977272, + 0.01443412248045206, + 0.004758467897772789, + 0.009857886470854282, + 0.009419660083949566, + 0.008120663464069366, + 0.00853828527033329, + 0.007345190737396479, + 0.01109502837061882, + 0.010450071655213833, + 0.0012421871069818735, + 0.003688492113724351, + 0.0018614995060488582, + 0.0026600693818181753, + 0.002287429291754961, + 0.0014576491666957736, + 0.0038699365686625242, + 0.0019749333150684834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.048633772879838943 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_5/index_tensor([243, 243])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_5/index_tensor([243, 243])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..55e73a886916bf813bf06dec91f852126d6cf935 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_5/index_tensor([243, 243])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, long shot, rising slowly from a low angle, slowly reveals a boundless, thick, swirling sea of clouds. Above this sea of frozen white, a massive wooden warship, carved with dragons, silently floats. Constructed of dark, weathered wood, the ship is covered in intricate, majestic relief carvings of oriental dragons. Its head gazes fiercely forward, its towering prow like the jaws of a behemoth. The ship remains almost motionless, only a few ancient flags fluttering gently from the towering masts in the breeze. As the sun sets, a massive sunset hangs high in the sky, its blazing golden backlight casting a dazzling, sacred silhouette across the ship, shrouding the camera-facing side in deep shadow. Yet, this golden light subtly brushes against the edges of the intricate dragon carvings, making them appear and disappear amidst the darkness, creating a striking sense of three-dimensionality. The afterglow of the setting sun dyed the top of the sea of clouds a warm orange-red, and the whole picture presented a solemn, tranquil and epic fantasy atmosphere, as if time had frozen at this moment.", + "seed": 16961, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_5/index_243_seed_16961_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06204475528793409, + "vlm_logprob": -0.06204475528793409, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.06204475528793409, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039614103734493256 + ], + "opposite_logprob_tokens": [ + -3.414614200592041 + ], + "logprob_reward": -0.039614103734493256, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.039614103734493256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08622175455093384, + 0.46357467770576477, + 0.08365622907876968, + 0.11249694973230362, + 0.0511380098760128, + 0.038522493094205856, + 0.06385648250579834, + 0.06193127483129501, + 0.021611960604786873, + 0.016990121454000473 + ], + "credit_map": [ + 0.012618327513337135, + 0.010592863894999027, + 0.011580551974475384, + 0.011752433143556118, + 0.008353661745786667, + 0.011006947606801987, + 0.007555229589343071, + 0.012761738151311874, + 0.03372783213853836, + 0.02831234596669674, + 0.04046424478292465, + 0.03292723000049591, + 0.035055238753557205, + 0.08288419991731644, + 0.0809711143374443, + 0.1292324662208557, + 0.007821771316230297, + 0.004810314625501633, + 0.007367215119302273, + 0.014433195814490318, + 0.0065684630535542965, + 0.014701979234814644, + 0.008688537403941154, + 0.01926475763320923, + 0.010211092419922352, + 0.00644611194729805, + 0.017900165170431137, + 0.026433732360601425, + 0.006922104395925999, + 0.01398131251335144, + 0.015787916257977486, + 0.014814517460763454, + 0.005512131378054619, + 0.0036395350471138954, + 0.00453981151804328, + 0.008300317451357841, + 0.004498555790632963, + 0.006554530933499336, + 0.0097498195245862, + 0.00834331102669239, + 0.004615738522261381, + 0.0036886546295136213, + 0.007255780976265669, + 0.004316875245422125, + 0.0036072004586458206, + 0.004496241919696331, + 0.003119455184787512, + 0.007422550115734339, + 0.007623929996043444, + 0.004521974828094244, + 0.011942225508391857, + 0.006567283067852259, + 0.004658205900341272, + 0.008690329268574715, + 0.004933505319058895, + 0.01491902582347393, + 0.007859732024371624, + 0.003675521118566394, + 0.011245792731642723, + 0.005476599093526602, + 0.0042150444351136684, + 0.008824880234897137, + 0.006845301948487759, + 0.013788406737148762, + 0.0025967422407120466, + 0.0013265515444800258, + 0.002737032948061824, + 0.001995345111936331, + 0.001794658601284027, + 0.0034714164212346077, + 0.0020304438658058643, + 0.005659769754856825, + 0.0037379739806056023, + 0.0011033902410417795, + 0.0036515486426651478, + 0.0016982404049485922, + 0.0010634817881509662, + 0.0021702474914491177, + 0.0013640092220157385, + 0.00220122910104692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.039614103734493256 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011954312212765217 + ], + "opposite_logprob_tokens": [ + -6.751195430755615 + ], + "logprob_reward": -0.0011954312212765217, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011954312212765217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12548059225082397, + 0.449017196893692, + 0.05517740920186043, + 0.11290434002876282, + 0.061793938279151917, + 0.025677725672721863, + 0.07908431440591812, + 0.02781577780842781, + 0.028401533141732216, + 0.03464716672897339 + ], + "credit_map": [ + 0.019463514909148216, + 0.017975343391299248, + 0.010246030054986477, + 0.012247522361576557, + 0.015345091931521893, + 0.016150135546922684, + 0.013591933064162731, + 0.020461037755012512, + 0.0346786193549633, + 0.038864582777023315, + 0.045272596180438995, + 0.03302375599741936, + 0.02988222986459732, + 0.07388310134410858, + 0.07758741825819016, + 0.11582490801811218, + 0.005577475763857365, + 0.004888135474175215, + 0.005102220457047224, + 0.0069356984458863735, + 0.00657032523304224, + 0.009360458701848984, + 0.006147984880954027, + 0.010595114901661873, + 0.01236653607338667, + 0.010284407064318657, + 0.01731666550040245, + 0.022948522120714188, + 0.007979207672178745, + 0.013138394802808762, + 0.014249780215322971, + 0.014620834961533546, + 0.010820365510880947, + 0.010365663096308708, + 0.006878083571791649, + 0.0057800039649009705, + 0.00515369139611721, + 0.0070535107515752316, + 0.007279231678694487, + 0.008463393896818161, + 0.0026915099006146193, + 0.003181657986715436, + 0.0038621029816567898, + 0.003198012476786971, + 0.0024243544321507215, + 0.0031582978554069996, + 0.002296481980010867, + 0.004865311551839113, + 0.010327602736651897, + 0.00504094734787941, + 0.013797242194414139, + 0.01011995505541563, + 0.005758452694863081, + 0.009821521118283272, + 0.007085081189870834, + 0.01713351532816887, + 0.0022262209095060825, + 0.0017225733026862144, + 0.005961032584309578, + 0.0031824326142668724, + 0.002337059937417507, + 0.0036974726244807243, + 0.003052182961255312, + 0.005636804737150669, + 0.004198071081191301, + 0.002191650914028287, + 0.0048112934455275536, + 0.0028768323827534914, + 0.002590687945485115, + 0.003947277553379536, + 0.002200078684836626, + 0.00558564392849803, + 0.009359072893857956, + 0.0025210289750248194, + 0.0053323376923799515, + 0.0039537642151117325, + 0.002582218963652849, + 0.003797668032348156, + 0.0036733464803546667, + 0.0034277327358722687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0011954312212765217 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22553794085979462 + ], + "opposite_logprob_tokens": [ + -1.600537896156311 + ], + "logprob_reward": -0.22553794085979462, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22553794085979462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1581524759531021, + 0.07283774763345718, + 0.05038096755743027, + 0.07615840435028076, + 0.14410173892974854, + 0.06902413070201874, + 0.13586746156215668, + 0.21427415311336517, + 0.06070685014128685, + 0.018496088683605194 + ], + "credit_map": [ + 0.02890203334391117, + 0.01936226524412632, + 0.025408677756786346, + 0.018511053174734116, + 0.013310256414115429, + 0.01933293230831623, + 0.015032364055514336, + 0.018292903900146484, + 0.012583687901496887, + 0.007570667192339897, + 0.00876174308359623, + 0.0079279700294137, + 0.006809824146330357, + 0.009952941909432411, + 0.010086174122989178, + 0.009144734591245651, + 0.005273980554193258, + 0.006153074558824301, + 0.00452625984326005, + 0.008513555862009525, + 0.006512382533401251, + 0.00667374674230814, + 0.004920077510178089, + 0.007807889953255653, + 0.010367183946073055, + 0.011720122769474983, + 0.008977760560810566, + 0.00928925909101963, + 0.007319960743188858, + 0.010423128493130207, + 0.007898938842117786, + 0.010162055492401123, + 0.010745505802333355, + 0.012864264659583569, + 0.014042656868696213, + 0.02058982662856579, + 0.010528963059186935, + 0.020611081272363663, + 0.020023778080940247, + 0.03469565883278847, + 0.010818245820701122, + 0.006482345517724752, + 0.013242659159004688, + 0.012799786403775215, + 0.004024796653538942, + 0.007091381587088108, + 0.0053699626587331295, + 0.009194948710501194, + 0.022617002949118614, + 0.007306130137294531, + 0.025309713557362556, + 0.013559307903051376, + 0.008930669166147709, + 0.018178019672632217, + 0.009353465400636196, + 0.03061315231025219, + 0.041660375893116, + 0.02209046296775341, + 0.0357273668050766, + 0.021410979330539703, + 0.018296044319868088, + 0.027994215488433838, + 0.015919890254735947, + 0.031174801290035248, + 0.008480721153318882, + 0.0035575837828218937, + 0.011887182481586933, + 0.006218889728188515, + 0.005387135315686464, + 0.008320163935422897, + 0.0062352679669857025, + 0.010619905777275562, + 0.0035731967072933912, + 0.0009617914911359549, + 0.0033866323065012693, + 0.0015877169789746404, + 0.0012988732196390629, + 0.0025696642696857452, + 0.001952294958755374, + 0.003165918169543147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.22553794085979462 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008152975351549685 + ], + "opposite_logprob_tokens": [ + -7.125815391540527 + ], + "logprob_reward": -0.0008152975351549685, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008152975351549685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.155908465385437, + 0.20740102231502533, + 0.0460163839161396, + 0.10419636219739914, + 0.23800605535507202, + 0.10071945190429688, + 0.023469112813472748, + 0.03802372142672539, + 0.040674839168787, + 0.045584626495838165 + ], + "credit_map": [ + 0.01984339952468872, + 0.01646704226732254, + 0.018226219341158867, + 0.02169315330684185, + 0.023779263719916344, + 0.020644987002015114, + 0.013825441710650921, + 0.021428946405649185, + 0.017150115221738815, + 0.019502686336636543, + 0.02207932621240616, + 0.022829337045550346, + 0.019554343074560165, + 0.032291848212480545, + 0.043154649436473846, + 0.030838701874017715, + 0.005582490935921669, + 0.0038689516950398684, + 0.005967745091766119, + 0.006750865373760462, + 0.00468803895637393, + 0.006330604664981365, + 0.004455674439668655, + 0.008372005075216293, + 0.010933233425021172, + 0.006896281149238348, + 0.012163056991994381, + 0.011015360243618488, + 0.008035028353333473, + 0.013971352949738503, + 0.014514533802866936, + 0.026667503640055656, + 0.055384039878845215, + 0.03032585233449936, + 0.028191203251481056, + 0.02499871514737606, + 0.0164257250726223, + 0.03009594976902008, + 0.02660025842487812, + 0.025984277948737144, + 0.014276755973696709, + 0.009683516807854176, + 0.022370588034391403, + 0.017466505989432335, + 0.005955332890152931, + 0.01209555845707655, + 0.009120753966271877, + 0.00975042860955, + 0.004019643645733595, + 0.0020671007223427296, + 0.0029365229420363903, + 0.0031406243797391653, + 0.0021749839652329683, + 0.003011151682585478, + 0.0021483704913407564, + 0.003970712423324585, + 0.00550075713545084, + 0.003163220128044486, + 0.006435367278754711, + 0.004288724157959223, + 0.0026031755842268467, + 0.00521570909768343, + 0.004931915085762739, + 0.0058848499320447445, + 0.005735562648624182, + 0.0028177136555314064, + 0.0063253045082092285, + 0.004150262102484703, + 0.003354873973876238, + 0.005484688561409712, + 0.004595279227942228, + 0.008211149834096432, + 0.006731739733368158, + 0.003245000960305333, + 0.006002289243042469, + 0.0042925504967570305, + 0.0055002616718411446, + 0.005693464074283838, + 0.005690316203981638, + 0.008429000154137611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0008152975351549685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306100308895111 + ], + "opposite_logprob_tokens": [ + -3.1680610179901123 + ], + "logprob_reward": -0.04306100308895111, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306100308895111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19222183525562286, + 0.19591328501701355, + 0.05871376767754555, + 0.11573740839958191, + 0.11337528377771378, + 0.10983395576477051, + 0.04107731953263283, + 0.037171684205532074, + 0.025344962254166603, + 0.11061049997806549 + ], + "credit_map": [ + 0.03457891196012497, + 0.02319624088704586, + 0.01612435281276703, + 0.017574740573763847, + 0.018899133428931236, + 0.023423688486218452, + 0.02355692721903324, + 0.03486785292625427, + 0.023836713284254074, + 0.020047912374138832, + 0.02526230923831463, + 0.022822678089141846, + 0.016134576871991158, + 0.0234188511967659, + 0.036077551543712616, + 0.02831270731985569, + 0.007399482186883688, + 0.006951196119189262, + 0.008466765284538269, + 0.009714283049106598, + 0.004777397960424423, + 0.006871185731142759, + 0.005463919602334499, + 0.009069536812603474, + 0.019418790936470032, + 0.013486981391906738, + 0.01676984690129757, + 0.018939197063446045, + 0.011953715234994888, + 0.014137336052954197, + 0.009157558903098106, + 0.011873980052769184, + 0.022021420300006866, + 0.015921898186206818, + 0.012939234264194965, + 0.013237664476037025, + 0.008777393028140068, + 0.01397295854985714, + 0.014986148104071617, + 0.011518564075231552, + 0.01439740788191557, + 0.009932959452271461, + 0.023671943694353104, + 0.018761377781629562, + 0.007157376501709223, + 0.012631645426154137, + 0.006985613610595465, + 0.01629563793540001, + 0.0071665337309241295, + 0.0035608718171715736, + 0.007082303985953331, + 0.005560524761676788, + 0.003466630820184946, + 0.004857137333601713, + 0.0028057817835360765, + 0.006577537395060062, + 0.0043581570498645306, + 0.0025478980969637632, + 0.006647903006523848, + 0.003174420678988099, + 0.0029955736827105284, + 0.0055490462109446526, + 0.003194473683834076, + 0.008704210631549358, + 0.003352233674377203, + 0.0018161307089030743, + 0.005191797856241465, + 0.002486777026206255, + 0.002445806050673127, + 0.0032263891771435738, + 0.00236538820900023, + 0.00446044048294425, + 0.013901874423027039, + 0.009181618690490723, + 0.020980101078748703, + 0.0135794747620821, + 0.015513415448367596, + 0.01201536599546671, + 0.012056732550263405, + 0.01338190957903862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.04306100308895111 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, long shot, rising slowly from a low angle, slowly reveals a boundless, thick, swirling sea of clouds. Above this sea of frozen white, a massive wooden warship, carved with dragons, silently floats. Constructed of dark, weathered wood, the ship is covered in intricate, majestic relief carvings of oriental dragons. Its head gazes fiercely forward, its towering prow like the jaws of a behemoth. The ship remains almost motionless, only a few ancient flags fluttering gently from the towering masts in the breeze. As the sun sets, a massive sunset hangs high in the sky, its blazing golden backlight casting a dazzling, sacred silhouette across the ship, shrouding the camera-facing side in deep shadow. Yet, this golden light subtly brushes against the edges of the intricate dragon carvings, making them appear and disappear amidst the darkness, creating a striking sense of three-dimensionality. The afterglow of the setting sun dyed the top of the sea of clouds a warm orange-red, and the whole picture presented a solemn, tranquil and epic fantasy atmosphere, as if time had frozen at this moment.", + "seed": 16962, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_5/index_243_seed_16962_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04412257213552948, + "vlm_logprob": -0.04412257213552948, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.04412257213552948, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13010133802890778 + ], + "opposite_logprob_tokens": [ + -2.130101442337036 + ], + "logprob_reward": -0.13010133802890778, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13010133802890778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30300530791282654, + 0.09371011704206467, + 0.1864825189113617, + 0.0183867160230875, + 0.06514773517847061, + 0.014532409608364105, + 0.05436865612864494, + 0.015502968803048134, + 0.07669713348150253, + 0.17216642200946808 + ], + "credit_map": [ + 0.018956594169139862, + 0.03536108508706093, + 0.04253101721405983, + 0.048357151448726654, + 0.02021193690598011, + 0.022049663588404655, + 0.050347767770290375, + 0.06519008427858353, + 0.015840627253055573, + 0.014111775904893875, + 0.015452264808118343, + 0.015789223834872246, + 0.010923154652118683, + 0.006691752001643181, + 0.007260981947183609, + 0.007640344090759754, + 0.006868994794785976, + 0.004052382428199053, + 0.07303476333618164, + 0.053393006324768066, + 0.005146844312548637, + 0.00782314594835043, + 0.0077204289846122265, + 0.02844293601810932, + 0.0019522204529494047, + 0.0014896064531058073, + 0.002422721590846777, + 0.0024852280039340258, + 0.0024532745592296124, + 0.003764360211789608, + 0.0012507778592407703, + 0.0025685266591608524, + 0.011469895020127296, + 0.002535418840125203, + 0.004516530316323042, + 0.010102343745529652, + 0.0032216226682066917, + 0.02407197095453739, + 0.003028612583875656, + 0.006201339419931173, + 0.0015723477117717266, + 0.0013202149420976639, + 0.0011863791150972247, + 0.00133096007630229, + 0.0021571118850260973, + 0.001946829492226243, + 0.0016438907478004694, + 0.003374676452949643, + 0.004849277902394533, + 0.0053041670471429825, + 0.004637283273041248, + 0.0044713192619383335, + 0.007185216993093491, + 0.006699450314044952, + 0.00749706756323576, + 0.013724872842431068, + 0.0020248088985681534, + 0.0013723507290706038, + 0.001440738094970584, + 0.0013983231037855148, + 0.002507739933207631, + 0.0020305044017732143, + 0.002466233680024743, + 0.002262270078063011, + 0.006362665444612503, + 0.0062528569251298904, + 0.009114472195506096, + 0.006294307764619589, + 0.01625564508140087, + 0.012395375408232212, + 0.010478277690708637, + 0.009543530642986298, + 0.01530325785279274, + 0.016671326011419296, + 0.01373939961194992, + 0.01770237274467945, + 0.026121970266103745, + 0.023361606523394585, + 0.02661389857530594, + 0.032652597874403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.13010133802890778 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059982344508171 + ], + "opposite_logprob_tokens": [ + -3.895599842071533 + ], + "logprob_reward": -0.02059982344508171, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059982344508171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1523447036743164, + 0.2929140329360962, + 0.14220376312732697, + 0.035341527312994, + 0.15729695558547974, + 0.024072028696537018, + 0.030482282862067223, + 0.021880077198147774, + 0.06548746675252914, + 0.07797712087631226 + ], + "credit_map": [ + 0.014392640441656113, + 0.022438591346144676, + 0.02614462375640869, + 0.022391248494386673, + 0.011264681816101074, + 0.009697058238089085, + 0.02098006196320057, + 0.025035811588168144, + 0.05672205984592438, + 0.03845663368701935, + 0.048552434891462326, + 0.050618693232536316, + 0.031518399715423584, + 0.020002152770757675, + 0.022770851850509644, + 0.024272827431559563, + 0.008698484860360622, + 0.012043335475027561, + 0.035906415432691574, + 0.025998102501034737, + 0.015740415081381798, + 0.01514089573174715, + 0.010116413235664368, + 0.018559705466032028, + 0.003349801991134882, + 0.004088398534804583, + 0.004517476540058851, + 0.006076031364500523, + 0.0035777960438281298, + 0.003231774317100644, + 0.0031176216434687376, + 0.007382629439234734, + 0.027233239263296127, + 0.008893565274775028, + 0.01599087379872799, + 0.03064918890595436, + 0.007916076108813286, + 0.04807809367775917, + 0.00844398234039545, + 0.010091960430145264, + 0.0031717829406261444, + 0.0026802148204296827, + 0.003189121838659048, + 0.0027719687204807997, + 0.0017769525293260813, + 0.0030711654108017683, + 0.0031063430942595005, + 0.004304481204599142, + 0.0026979369577020407, + 0.002201998606324196, + 0.0030348694417625666, + 0.0033695711754262447, + 0.0044656777754426, + 0.003927632700651884, + 0.004669282119721174, + 0.00611531687900424, + 0.0029131874907761812, + 0.0018219795310869813, + 0.003122213063761592, + 0.002523099072277546, + 0.0032506168354302645, + 0.002610970288515091, + 0.002408805536106229, + 0.0032292089890688658, + 0.007203948684036732, + 0.0072547076269984245, + 0.007214766927063465, + 0.005390437785536051, + 0.016033563762903214, + 0.010890001431107521, + 0.00610068254172802, + 0.005399365909397602, + 0.006987385917454958, + 0.006881776265799999, + 0.006287600379437208, + 0.007299068383872509, + 0.011536240577697754, + 0.010700143873691559, + 0.012119033373892307, + 0.016165878623723984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.02059982344508171 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062040217220783234 + ], + "opposite_logprob_tokens": [ + -2.812040328979492 + ], + "logprob_reward": -0.062040217220783234, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062040217220783234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21027900278568268, + 0.09608768671751022, + 0.10887646675109863, + 0.12797869741916656, + 0.12277865409851074, + 0.044245485216379166, + 0.08700250089168549, + 0.08106033504009247, + 0.055749449878931046, + 0.06594173610210419 + ], + "credit_map": [ + 0.0199184138327837, + 0.026086585596203804, + 0.051828838884830475, + 0.03767888620495796, + 0.012877562083303928, + 0.01391957513988018, + 0.015768203884363174, + 0.032200947403907776, + 0.015697656199336052, + 0.014081393368542194, + 0.014379121363162994, + 0.015317168086767197, + 0.011326403357088566, + 0.005637837573885918, + 0.008708165027201176, + 0.010939941741526127, + 0.005801335442811251, + 0.008265157230198383, + 0.03368331491947174, + 0.02523702196776867, + 0.008260462433099747, + 0.006324793212115765, + 0.006743160076439381, + 0.014561214484274387, + 0.011638383381068707, + 0.011146371252834797, + 0.017220664769411087, + 0.024700868874788284, + 0.010646343231201172, + 0.016419189050793648, + 0.01381700299680233, + 0.022389862686395645, + 0.00868301372975111, + 0.008518077433109283, + 0.019306117668747902, + 0.038982708007097244, + 0.006868841592222452, + 0.012717255391180515, + 0.008361882530152798, + 0.01934075728058815, + 0.005765553563833237, + 0.004610175732523203, + 0.004490836523473263, + 0.0066300430335104465, + 0.0038373565766960382, + 0.006744216661900282, + 0.00584339490160346, + 0.006323903799057007, + 0.009321038611233234, + 0.01112062856554985, + 0.006983153056353331, + 0.007115707732737064, + 0.019058790057897568, + 0.011369804851710796, + 0.009785132482647896, + 0.012248233892023563, + 0.008591221645474434, + 0.007911274209618568, + 0.008489612489938736, + 0.008002202026546001, + 0.0106483343988657, + 0.010935946367681026, + 0.01321865152567625, + 0.013263082131743431, + 0.004845793824642897, + 0.006549065932631493, + 0.010004883632063866, + 0.006094019394367933, + 0.010409659706056118, + 0.008619838394224644, + 0.004465555772185326, + 0.0047606294974684715, + 0.005639984272420406, + 0.006164733320474625, + 0.003995359409600496, + 0.00600478844717145, + 0.011392755433917046, + 0.009321920573711395, + 0.01015259139239788, + 0.013269605115056038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.062040217220783234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00023862851958256215 + ], + "opposite_logprob_tokens": [ + -8.375238418579102 + ], + "logprob_reward": -0.00023862851958256215, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.00023862851958256215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13145487010478973, + 0.13176670670509338, + 0.4378296732902527, + 0.04642260819673538, + 0.028183849528431892, + 0.014571278356015682, + 0.04106713831424713, + 0.021347256377339363, + 0.06667239218950272, + 0.08068427443504333 + ], + "credit_map": [ + 0.010268677957355976, + 0.021878870204091072, + 0.017346546053886414, + 0.017726732417941093, + 0.010696887038648129, + 0.00917032640427351, + 0.019472606480121613, + 0.024894213303923607, + 0.018297765403985977, + 0.01970437914133072, + 0.022411994636058807, + 0.022192902863025665, + 0.019055945798754692, + 0.009442401118576527, + 0.009871350601315498, + 0.010789966210722923, + 0.016788622364401817, + 0.008002306334674358, + 0.17674961686134338, + 0.13079732656478882, + 0.006131432019174099, + 0.01471409760415554, + 0.015993807464838028, + 0.06865247339010239, + 0.005016996059566736, + 0.003984643612056971, + 0.006400018464773893, + 0.007854428142309189, + 0.004029588308185339, + 0.0063252197578549385, + 0.005949802231043577, + 0.006861909758299589, + 0.004452857188880444, + 0.0014005956472828984, + 0.001950185396708548, + 0.0033154876437038183, + 0.0022414608392864466, + 0.009382092393934727, + 0.002337606158107519, + 0.0031035644933581352, + 0.002444378798827529, + 0.0016953842714428902, + 0.0014536600792780519, + 0.0013649130705744028, + 0.0017690517706796527, + 0.002541483147069812, + 0.0014974632067605853, + 0.0018049436621367931, + 0.0037364771123975515, + 0.0040592774748802185, + 0.003379115602001548, + 0.0034160015638917685, + 0.005102506838738918, + 0.0049008834175765514, + 0.005548614542931318, + 0.010924264788627625, + 0.0024040467105805874, + 0.003090084996074438, + 0.0018561766482889652, + 0.001897377078421414, + 0.003780621336773038, + 0.0028547304682433605, + 0.002779667731374502, + 0.002684549428522587, + 0.005794684402644634, + 0.006540143862366676, + 0.007737026084214449, + 0.005394423380494118, + 0.016035370528697968, + 0.011399277485907078, + 0.006937227677553892, + 0.006834235042333603, + 0.0065913572907447815, + 0.007814150303602219, + 0.006746864411979914, + 0.008690929971635342, + 0.012010516598820686, + 0.010849491693079472, + 0.013332483358681202, + 0.014648479409515858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.00023862851958256215 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632853463292122 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632853463292122, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632853463292122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.172663614153862, + 0.15423884987831116, + 0.3443220257759094, + 0.025056390091776848, + 0.041255105286836624, + 0.012954145669937134, + 0.07571544498205185, + 0.08352033793926239, + 0.04410311579704285, + 0.04617098718881607 + ], + "credit_map": [ + 0.017450407147407532, + 0.01661107689142227, + 0.050439320504665375, + 0.0374782420694828, + 0.010097421705722809, + 0.012102169916033745, + 0.010427571833133698, + 0.018057385459542274, + 0.02718154340982437, + 0.018375249579548836, + 0.027164947241544724, + 0.027416540309786797, + 0.017707332968711853, + 0.011192694306373596, + 0.012371324002742767, + 0.012829206883907318, + 0.013739338144659996, + 0.009633238427340984, + 0.13454456627368927, + 0.0987015962600708, + 0.008518929593265057, + 0.013032565824687481, + 0.013557313941419125, + 0.05259447172284126, + 0.002763299271464348, + 0.002690297784283757, + 0.0029270348604768515, + 0.002947635017335415, + 0.003668224671855569, + 0.003656637854874134, + 0.0025707348249852657, + 0.0038325227797031403, + 0.0052040559239685535, + 0.0024014031514525414, + 0.004845741670578718, + 0.009111895225942135, + 0.002413405803963542, + 0.009388376027345657, + 0.0035892026498913765, + 0.00430102227255702, + 0.0017098820535466075, + 0.001217929762788117, + 0.0019645600114017725, + 0.0017426992999389768, + 0.0012055424740538, + 0.0016685067676007748, + 0.0016854442656040192, + 0.0017595805693417788, + 0.007393990643322468, + 0.008337817154824734, + 0.006262362003326416, + 0.006013743579387665, + 0.015586497262120247, + 0.010915299877524376, + 0.008638697676360607, + 0.012567026540637016, + 0.008466235361993313, + 0.008015601895749569, + 0.00895246583968401, + 0.010850992985069752, + 0.010190185159444809, + 0.009810203686356544, + 0.011778797954320908, + 0.015455849468708038, + 0.005367303732782602, + 0.004615272860974073, + 0.0036908250767737627, + 0.0037159512285143137, + 0.008954948745667934, + 0.007108308840543032, + 0.005215440411120653, + 0.005435059778392315, + 0.004931162111461163, + 0.005645898170769215, + 0.004111364018172026, + 0.0036696542520076036, + 0.01074485294520855, + 0.0074920617043972015, + 0.004535870160907507, + 0.005040121730417013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.007632853463292122 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, long shot, rising slowly from a low angle, slowly reveals a boundless, thick, swirling sea of clouds. Above this sea of frozen white, a massive wooden warship, carved with dragons, silently floats. Constructed of dark, weathered wood, the ship is covered in intricate, majestic relief carvings of oriental dragons. Its head gazes fiercely forward, its towering prow like the jaws of a behemoth. The ship remains almost motionless, only a few ancient flags fluttering gently from the towering masts in the breeze. As the sun sets, a massive sunset hangs high in the sky, its blazing golden backlight casting a dazzling, sacred silhouette across the ship, shrouding the camera-facing side in deep shadow. Yet, this golden light subtly brushes against the edges of the intricate dragon carvings, making them appear and disappear amidst the darkness, creating a striking sense of three-dimensionality. The afterglow of the setting sun dyed the top of the sea of clouds a warm orange-red, and the whole picture presented a solemn, tranquil and epic fantasy atmosphere, as if time had frozen at this moment.", + "seed": 16963, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_5/index_243_seed_16963_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.5406937710475177, + "vlm_logprob": -0.5406937710475177, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.5406937710475177, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.3514456748962402 + ], + "opposite_logprob_tokens": [ + -0.10144557803869247 + ], + "logprob_reward": -2.3514456748962402, + "yes_no_margin_reward": -2.25, + "token_credit_reward": -2.3514456748962402, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.192228302359581, + 0.14192348718643188, + 0.1277734935283661, + 0.030240921303629875, + 0.043233342468738556, + 0.029218627139925957, + 0.02104761265218258, + 0.08525358140468597, + 0.29870113730430603, + 0.030379552394151688 + ], + "credit_map": [ + 0.016871927306056023, + 0.02364516444504261, + 0.025519732385873795, + 0.02482539787888527, + 0.02804403379559517, + 0.019165119156241417, + 0.020452603697776794, + 0.03370431810617447, + 0.0125749995931983, + 0.012567603960633278, + 0.02415066957473755, + 0.010346335358917713, + 0.018323957920074463, + 0.039340853691101074, + 0.011800363659858704, + 0.012818712741136551, + 0.0145137133076787, + 0.009335841052234173, + 0.016455229371786118, + 0.015296408906579018, + 0.01703501120209694, + 0.015433931723237038, + 0.017406972125172615, + 0.02229638583958149, + 0.00291659333743155, + 0.0022252099588513374, + 0.004423959646373987, + 0.0034497228916734457, + 0.004225695971399546, + 0.004202976822853088, + 0.004240123089402914, + 0.004556639119982719, + 0.006051041651517153, + 0.0025294525548815727, + 0.006159890443086624, + 0.004079877864569426, + 0.005946060176938772, + 0.004467714577913284, + 0.005829954985529184, + 0.008169349282979965, + 0.0033274327870458364, + 0.0022923308424651623, + 0.004711343441158533, + 0.0032432193402200937, + 0.003336007008329034, + 0.0031386837363243103, + 0.005289962515234947, + 0.0038796488661319017, + 0.0025364391040056944, + 0.0023136278614401817, + 0.0027070094365626574, + 0.00245727994479239, + 0.002979771699756384, + 0.0027162309270352125, + 0.0026205587200820446, + 0.00271669402718544, + 0.013177845627069473, + 0.007513240911066532, + 0.011285564862191677, + 0.00981988850980997, + 0.010223403573036194, + 0.016763905063271523, + 0.008645816706120968, + 0.007823905907571316, + 0.024016661569476128, + 0.015187209472060204, + 0.011400312185287476, + 0.009182216599583626, + 0.012900980189442635, + 0.12069777399301529, + 0.024922607466578484, + 0.08039338141679764, + 0.005426386836916208, + 0.0031584305688738823, + 0.004430751316249371, + 0.002393292495980859, + 0.003862874349579215, + 0.003919167909771204, + 0.003945725504308939, + 0.0032429227139800787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -2.3514456748962402 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06213051825761795 + ], + "opposite_logprob_tokens": [ + -2.8121304512023926 + ], + "logprob_reward": -0.06213051825761795, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06213051825761795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17976844310760498, + 0.09847363829612732, + 0.18630780279636383, + 0.07182497531175613, + 0.05135883390903473, + 0.11593518406152725, + 0.035820238292217255, + 0.10691507160663605, + 0.12887725234031677, + 0.024718506261706352 + ], + "credit_map": [ + 0.018744314089417458, + 0.02524605765938759, + 0.02368638664484024, + 0.020557411015033722, + 0.024185512214899063, + 0.018392877653241158, + 0.01688752882182598, + 0.03206836059689522, + 0.006395947188138962, + 0.010967174544930458, + 0.012298722751438618, + 0.009409157559275627, + 0.020276427268981934, + 0.015832237899303436, + 0.01179692056030035, + 0.011497054249048233, + 0.01928696408867836, + 0.014687643386423588, + 0.021616782993078232, + 0.024024106562137604, + 0.02997499890625477, + 0.022693807259202003, + 0.024372387677431107, + 0.029651107266545296, + 0.007043006364256144, + 0.004473701119422913, + 0.0073358528316020966, + 0.008892916142940521, + 0.013795582577586174, + 0.009916332550346851, + 0.009454135783016682, + 0.010913445614278316, + 0.006846096832305193, + 0.0029038614593446255, + 0.006691183429211378, + 0.005607428029179573, + 0.005602745804935694, + 0.0057001723907887936, + 0.0076232049614191055, + 0.01038414053618908, + 0.012937277555465698, + 0.012095971032977104, + 0.009737495332956314, + 0.012370116077363491, + 0.019556954503059387, + 0.01530018076300621, + 0.016975363716483116, + 0.0169618371874094, + 0.006057100836187601, + 0.005391946993768215, + 0.003824855200946331, + 0.003934652078896761, + 0.003624277887865901, + 0.004768433980643749, + 0.004137958399951458, + 0.004081017337739468, + 0.017817186191678047, + 0.014872717671096325, + 0.010705349035561085, + 0.013724999502301216, + 0.011339344084262848, + 0.015849493443965912, + 0.012197419069707394, + 0.010408569127321243, + 0.009215389378368855, + 0.006791045423597097, + 0.004783733282238245, + 0.0036247374955564737, + 0.004597595427185297, + 0.052306607365608215, + 0.010134735144674778, + 0.03742341697216034, + 0.004226028453558683, + 0.0023367921821773052, + 0.0033542471937835217, + 0.002634650096297264, + 0.0026501698885113, + 0.0036843279376626015, + 0.0029282153118401766, + 0.0029040740337222815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.06213051825761795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560614347457886 + ], + "opposite_logprob_tokens": [ + -1.6006062030792236 + ], + "logprob_reward": -0.22560614347457886, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560614347457886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11945568025112152, + 0.09805406630039215, + 0.026583949103951454, + 0.02493179775774479, + 0.0472303070127964, + 0.03478512167930603, + 0.030909106135368347, + 0.04738180339336395, + 0.5322225093841553, + 0.03844563290476799 + ], + "credit_map": [ + 0.009450865909457207, + 0.02264067716896534, + 0.013386663980782032, + 0.014509364031255245, + 0.02140272781252861, + 0.012723175808787346, + 0.01379922591149807, + 0.011542990803718567, + 0.008423480205237865, + 0.011172976344823837, + 0.016425954177975655, + 0.00824584998190403, + 0.0184817872941494, + 0.020977534353733063, + 0.006655627861618996, + 0.007670865394175053, + 0.002759251045063138, + 0.0024050280917435884, + 0.0030383188277482986, + 0.0035900594666600227, + 0.005317479372024536, + 0.0035174607764929533, + 0.003389779245480895, + 0.0025665746070444584, + 0.002846037270501256, + 0.00237007113173604, + 0.0030336559284478426, + 0.00286216102540493, + 0.004303793888539076, + 0.0027126085478812456, + 0.004237705375999212, + 0.002565765753388405, + 0.005656912922859192, + 0.003090999089181423, + 0.00791850220412016, + 0.005480927415192127, + 0.007689035497605801, + 0.0048030284233391285, + 0.0055990396067500114, + 0.00699186185374856, + 0.004068233538419008, + 0.002431791741400957, + 0.004959567915648222, + 0.0034299474209547043, + 0.004907742142677307, + 0.005309143103659153, + 0.005562909878790379, + 0.004115786403417587, + 0.00498887337744236, + 0.0025942756328731775, + 0.003093746490776539, + 0.004232614301145077, + 0.0036484033335000277, + 0.003862007288262248, + 0.0035913533065468073, + 0.004897834733128548, + 0.00954071618616581, + 0.006057119928300381, + 0.006857309024780989, + 0.005982121918350458, + 0.004900629166513681, + 0.005355587229132652, + 0.003808933310210705, + 0.004879389889538288, + 0.04043750837445259, + 0.026779787614941597, + 0.017962535843253136, + 0.012718217447400093, + 0.020437544211745262, + 0.21941407024860382, + 0.04386717826128006, + 0.1506056934595108, + 0.004944126587361097, + 0.00402451679110527, + 0.0052701132372021675, + 0.003330662613734603, + 0.005694075953215361, + 0.003677546279504895, + 0.004294125828891993, + 0.0072104670107364655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.22560614347457886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022067499812692404 + ], + "opposite_logprob_tokens": [ + -6.127206802368164 + ], + "logprob_reward": -0.0022067499812692404, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022067499812692404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17072395980358124, + 0.10567368566989899, + 0.13765345513820648, + 0.039186205714941025, + 0.07076887041330338, + 0.06055762991309166, + 0.059759315103292465, + 0.07329054176807404, + 0.19747774302959442, + 0.08490864932537079 + ], + "credit_map": [ + 0.015949593856930733, + 0.016178874298930168, + 0.024304896593093872, + 0.01716456189751625, + 0.019890809431672096, + 0.020083973184227943, + 0.022846639156341553, + 0.03430462256073952, + 0.012387330643832684, + 0.008877629414200783, + 0.015597322955727577, + 0.00806447584182024, + 0.012386491522192955, + 0.027222013100981712, + 0.00953422486782074, + 0.011604195460677147, + 0.01955260895192623, + 0.011316893622279167, + 0.01912843808531761, + 0.012728565372526646, + 0.01833381876349449, + 0.0176409762352705, + 0.015441883355379105, + 0.023510269820690155, + 0.004064494743943214, + 0.003536693751811981, + 0.003905734047293663, + 0.0047197784297168255, + 0.007446625269949436, + 0.004466362297534943, + 0.006017375271767378, + 0.005029142834246159, + 0.010029063560068607, + 0.004704209044575691, + 0.008910913951694965, + 0.005890446715056896, + 0.007562762591987848, + 0.006905821617692709, + 0.010283448733389378, + 0.01648220792412758, + 0.007946728728711605, + 0.006510394625365734, + 0.009087077341973782, + 0.006460416596382856, + 0.008010217919945717, + 0.0068725296296179295, + 0.008020571433007717, + 0.007649694103747606, + 0.0063558537513017654, + 0.004395211581140757, + 0.008528086356818676, + 0.011821587570011616, + 0.0054640150628983974, + 0.007893179543316364, + 0.008399576880037785, + 0.006901800632476807, + 0.010780923999845982, + 0.0047676339745521545, + 0.010343580506742, + 0.005320894997566938, + 0.010701067745685577, + 0.017198652029037476, + 0.007840371690690517, + 0.006337422411888838, + 0.014028884470462799, + 0.010403241962194443, + 0.00709671201184392, + 0.0057500991970300674, + 0.008388708345592022, + 0.07787340134382248, + 0.018562281504273415, + 0.05537441000342369, + 0.013176634907722473, + 0.008728516288101673, + 0.012897171080112457, + 0.007095262408256531, + 0.009384860284626484, + 0.009630154818296432, + 0.010230034589767456, + 0.013766017742455006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0022067499812692404 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062079768627882004 + ], + "opposite_logprob_tokens": [ + -2.812079668045044 + ], + "logprob_reward": -0.062079768627882004, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062079768627882004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05469544976949692, + 0.04599478840827942, + 0.10131281614303589, + 0.06945592910051346, + 0.09000149369239807, + 0.04155094921588898, + 0.011555833742022514, + 0.1224353089928627, + 0.4423902928829193, + 0.020607102662324905 + ], + "credit_map": [ + 0.003461292712017894, + 0.006430978421121836, + 0.0068812016397714615, + 0.0050743259489536285, + 0.009141438640654087, + 0.007698598317801952, + 0.006972170900553465, + 0.0090354448184371, + 0.00574930664151907, + 0.005929393693804741, + 0.0066023822873830795, + 0.005269576329737902, + 0.007577584125101566, + 0.005684266332536936, + 0.004083387553691864, + 0.005098895635455847, + 0.014498111791908741, + 0.008423231542110443, + 0.014484918676316738, + 0.01052454300224781, + 0.014285211451351643, + 0.01145731471478939, + 0.01475568488240242, + 0.012883808463811874, + 0.009087283164262772, + 0.0070989103987813, + 0.007351024076342583, + 0.007295385468751192, + 0.01579856500029564, + 0.007114709820598364, + 0.008905473165214062, + 0.006804578006267548, + 0.008039707317948341, + 0.007316829636693001, + 0.01306391041725874, + 0.007754407823085785, + 0.01766541227698326, + 0.0094435540959239, + 0.012713639065623283, + 0.014004037715494633, + 0.004827423952519894, + 0.0032670169603079557, + 0.004147857893258333, + 0.004601724445819855, + 0.009279314428567886, + 0.005243049468845129, + 0.0053460258059203625, + 0.0048385378904640675, + 0.0016727738548070192, + 0.001339637441560626, + 0.0015118573792278767, + 0.002266703639179468, + 0.0013286542380228639, + 0.0011802070075646043, + 0.0014114293735474348, + 0.0008445727289654315, + 0.016565022990107536, + 0.017793867737054825, + 0.012562721967697144, + 0.013192023150622845, + 0.01849967986345291, + 0.01385035365819931, + 0.01709163188934326, + 0.012880013324320316, + 0.03287544474005699, + 0.0221969336271286, + 0.014068427495658398, + 0.010175634175539017, + 0.016041113063693047, + 0.18400661647319794, + 0.03613889217376709, + 0.12688729166984558, + 0.0021603198256343603, + 0.0021875868551433086, + 0.002549573080614209, + 0.002707309555262327, + 0.003162745153531432, + 0.001898408168926835, + 0.002241149777546525, + 0.0037000116426497698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.062079768627882004 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, long shot, rising slowly from a low angle, slowly reveals a boundless, thick, swirling sea of clouds. Above this sea of frozen white, a massive wooden warship, carved with dragons, silently floats. Constructed of dark, weathered wood, the ship is covered in intricate, majestic relief carvings of oriental dragons. Its head gazes fiercely forward, its towering prow like the jaws of a behemoth. The ship remains almost motionless, only a few ancient flags fluttering gently from the towering masts in the breeze. As the sun sets, a massive sunset hangs high in the sky, its blazing golden backlight casting a dazzling, sacred silhouette across the ship, shrouding the camera-facing side in deep shadow. Yet, this golden light subtly brushes against the edges of the intricate dragon carvings, making them appear and disappear amidst the darkness, creating a striking sense of three-dimensionality. The afterglow of the setting sun dyed the top of the sea of clouds a warm orange-red, and the whole picture presented a solemn, tranquil and epic fantasy atmosphere, as if time had frozen at this moment.", + "seed": 16964, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_5/index_243_seed_16964_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16061974680051208, + "vlm_logprob": -0.16061974680051208, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.16061974680051208, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6350600719451904 + ], + "opposite_logprob_tokens": [ + -0.7600600719451904 + ], + "logprob_reward": -0.6350600719451904, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6350600719451904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19429907202720642, + 0.033219944685697556, + 0.21036268770694733, + 0.1604970097541809, + 0.07737871259450912, + 0.12190787494182587, + 0.06328148394823074, + 0.016858907416462898, + 0.014729849062860012, + 0.10746456682682037 + ], + "credit_map": [ + 0.016484733670949936, + 0.017948197200894356, + 0.030587302520871162, + 0.04157880321145058, + 0.015221793204545975, + 0.013194833882153034, + 0.032993484288454056, + 0.026289910078048706, + 0.0026784702204167843, + 0.0025502871721982956, + 0.0044267126359045506, + 0.006637632381170988, + 0.00414206413552165, + 0.0029276437126100063, + 0.005179759114980698, + 0.004677377175539732, + 0.00787961296737194, + 0.006084740627557039, + 0.017548637464642525, + 0.027692079544067383, + 0.011988085694611073, + 0.009706166572868824, + 0.074381984770298, + 0.055081386119127274, + 0.015596313402056694, + 0.008727222681045532, + 0.016545113176107407, + 0.017728326842188835, + 0.009835274890065193, + 0.008824069984257221, + 0.05023910105228424, + 0.033001575618982315, + 0.0035540766548365355, + 0.0031493327114731073, + 0.021591927856206894, + 0.012638214975595474, + 0.0059022605419158936, + 0.004312694072723389, + 0.01090426929295063, + 0.015325933694839478, + 0.006770933512598276, + 0.004912266507744789, + 0.028990233317017555, + 0.022804345935583115, + 0.009010185487568378, + 0.009480424225330353, + 0.028847862035036087, + 0.011091619729995728, + 0.006591729819774628, + 0.004140499047935009, + 0.010947289876639843, + 0.006810834165662527, + 0.007802681531757116, + 0.010594798251986504, + 0.0102787921205163, + 0.00611486192792654, + 0.0012789975153282285, + 0.002053443342447281, + 0.0019395331619307399, + 0.0022666531149297953, + 0.0019147160928696394, + 0.002650198759511113, + 0.0031957291066646576, + 0.0015596358571201563, + 0.0024511704687029123, + 0.001931249862536788, + 0.0018494665855541825, + 0.0015202630311250687, + 0.0018853198271244764, + 0.002165032783523202, + 0.0015013315714895725, + 0.0014260149328038096, + 0.009621679782867432, + 0.00802967045456171, + 0.006117547396570444, + 0.0068411328829824924, + 0.0196173507720232, + 0.04187217727303505, + 0.007128794677555561, + 0.008236217312514782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.6350600719451904 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023307057097554207 + ], + "opposite_logprob_tokens": [ + -3.7733070850372314 + ], + "logprob_reward": -0.023307057097554207, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023307057097554207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19129212200641632, + 0.03671516850590706, + 0.15152579545974731, + 0.04555805027484894, + 0.08506397157907486, + 0.15923036634922028, + 0.074238121509552, + 0.012244476936757565, + 0.010177357122302055, + 0.2339545488357544 + ], + "credit_map": [ + 0.017443805932998657, + 0.017314350232481956, + 0.02810821682214737, + 0.04217888042330742, + 0.016036909073591232, + 0.013833167962729931, + 0.029879195615649223, + 0.026497595012187958, + 0.0031582636293023825, + 0.002846551826223731, + 0.004852866288274527, + 0.00760400015860796, + 0.004601658787578344, + 0.0033502618316560984, + 0.004333008546382189, + 0.005968555808067322, + 0.006453465670347214, + 0.004289440810680389, + 0.012305766344070435, + 0.01816486194729805, + 0.007898109033703804, + 0.005213409196585417, + 0.054745983332395554, + 0.042454756796360016, + 0.002883874112740159, + 0.0023223222233355045, + 0.004790456499904394, + 0.009957321919500828, + 0.004106507636606693, + 0.003629333106800914, + 0.012884176336228848, + 0.004984057042747736, + 0.003249245695769787, + 0.0021791800390928984, + 0.03246478736400604, + 0.015960311517119408, + 0.003967862576246262, + 0.003601295407861471, + 0.008145039901137352, + 0.015496250241994858, + 0.00716197257861495, + 0.008092460222542286, + 0.02890363149344921, + 0.022302521392703056, + 0.01533357985317707, + 0.027639735490083694, + 0.030660485848784447, + 0.019135968759655952, + 0.010990996845066547, + 0.00547985453158617, + 0.014314486645162106, + 0.009679931215941906, + 0.006798630114644766, + 0.0068472339771687984, + 0.01222640834748745, + 0.007900582626461983, + 0.0009022272424772382, + 0.0012460649013519287, + 0.001404965529218316, + 0.0015814463840797544, + 0.0017405148828402162, + 0.0018127873772755265, + 0.002161602256819606, + 0.0013948682462796569, + 0.0018337261863052845, + 0.0012173072900623083, + 0.001232316717505455, + 0.001301195239648223, + 0.0014419660437852144, + 0.0012950237141922116, + 0.0009640700300224125, + 0.0008917520171962678, + 0.021315399557352066, + 0.01618979498744011, + 0.012097237631678581, + 0.013102380558848381, + 0.04360123723745346, + 0.09684250503778458, + 0.01480904035270214, + 0.015996966511011124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.023307057097554207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549561008810997 + ], + "opposite_logprob_tokens": [ + -2.9299561977386475 + ], + "logprob_reward": -0.0549561008810997, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549561008810997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09688591212034225, + 0.017861397936940193, + 0.26052558422088623, + 0.03568824753165245, + 0.2883557677268982, + 0.16103845834732056, + 0.08201456815004349, + 0.017930051311850548, + 0.023484978824853897, + 0.01621505618095398 + ], + "credit_map": [ + 0.008383979089558125, + 0.011409331113100052, + 0.017111850902438164, + 0.013737685047090054, + 0.008563325740396976, + 0.007444461341947317, + 0.015072660520672798, + 0.015162612311542034, + 0.0017988934414461255, + 0.0016418902669101954, + 0.002106960630044341, + 0.0025622746907174587, + 0.002008294453844428, + 0.0023586859460920095, + 0.002217348897829652, + 0.0031670487951487303, + 0.01120651513338089, + 0.007794845849275589, + 0.02250330150127411, + 0.03364120423793793, + 0.013567172922194004, + 0.008863801136612892, + 0.09189704805612564, + 0.07105165719985962, + 0.004957239609211683, + 0.0036292041186243296, + 0.004387696273624897, + 0.003242787905037403, + 0.0024169757962226868, + 0.003671127837151289, + 0.007057836279273033, + 0.006325379945337772, + 0.00959701742976904, + 0.0044149053283035755, + 0.12917067110538483, + 0.06678946316242218, + 0.004784967750310898, + 0.008001542650163174, + 0.031066302210092545, + 0.0345308743417263, + 0.009141866117715836, + 0.006800191011279821, + 0.035808879882097244, + 0.033086664974689484, + 0.011586688458919525, + 0.011387779377400875, + 0.03825290873646736, + 0.014973469078540802, + 0.004304422996938229, + 0.004846677184104919, + 0.013227149844169617, + 0.009560225531458855, + 0.011599704623222351, + 0.015528672374784946, + 0.015188920311629772, + 0.007758784107863903, + 0.0030675679445266724, + 0.003338641719892621, + 0.001851121662184596, + 0.0018314395565539598, + 0.0016533007146790624, + 0.0022868909873068333, + 0.002201259834691882, + 0.0016998269129544497, + 0.0053963614627718925, + 0.0022219785023480654, + 0.00302028376609087, + 0.0024304897524416447, + 0.0026610163040459156, + 0.002716970629990101, + 0.002833589445799589, + 0.0022042870987206697, + 0.0022932933643460274, + 0.0010340516455471516, + 0.001234668307006359, + 0.0012813194189220667, + 0.0022911913692951202, + 0.005402953829616308, + 0.001578479423187673, + 0.001099096960388124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.0549561008810997 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007201223634183407 + ], + "opposite_logprob_tokens": [ + -7.250720024108887 + ], + "logprob_reward": -0.0007201223634183407, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007201223634183407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21900811791419983, + 0.029172811657190323, + 0.13270626962184906, + 0.1969880759716034, + 0.08804892003536224, + 0.11375301331281662, + 0.038337551057338715, + 0.02221331186592579, + 0.01948530785739422, + 0.14028652012348175 + ], + "credit_map": [ + 0.01731211692094803, + 0.020772961899638176, + 0.03648575022816658, + 0.044116031378507614, + 0.017599055543541908, + 0.017610516399145126, + 0.03639320656657219, + 0.028718475252389908, + 0.0023615951649844646, + 0.002067891415208578, + 0.0038165131118148565, + 0.0051677171140909195, + 0.003372253617271781, + 0.0028815974947065115, + 0.005297275260090828, + 0.004207966849207878, + 0.005294076167047024, + 0.0042053088545799255, + 0.010614166967570782, + 0.01594642922282219, + 0.007457716856151819, + 0.006505596451461315, + 0.04600369930267334, + 0.03667926788330078, + 0.022871969267725945, + 0.014855336397886276, + 0.02209697850048542, + 0.01622089557349682, + 0.012636813335120678, + 0.011126652359962463, + 0.05546688660979271, + 0.041712552309036255, + 0.005213575903326273, + 0.004105087369680405, + 0.022093256935477257, + 0.014850751496851444, + 0.00864013284444809, + 0.00463024852797389, + 0.01540185697376728, + 0.01311401091516018, + 0.0055860597640275955, + 0.004282818641513586, + 0.029023274779319763, + 0.021161530166864395, + 0.007485967595130205, + 0.009114202111959457, + 0.02731403522193432, + 0.009785132482647896, + 0.0037607273552566767, + 0.0023621758446097374, + 0.006156092043966055, + 0.0037987036630511284, + 0.004976032767444849, + 0.006609370931982994, + 0.006892094388604164, + 0.00378235150128603, + 0.002268474781885743, + 0.0021743481047451496, + 0.0027586042415350676, + 0.0023512316402047873, + 0.002372784772887826, + 0.0034132341388612986, + 0.003861404489725828, + 0.0030132292304188013, + 0.0027590456884354353, + 0.0018996335566043854, + 0.00198759650811553, + 0.0017747003585100174, + 0.0024587076622992754, + 0.002934361109510064, + 0.0027775196358561516, + 0.0028937438037246466, + 0.013916767202317715, + 0.008604886941611767, + 0.007105506956577301, + 0.007752981968224049, + 0.026793986558914185, + 0.05952583998441696, + 0.008337927982211113, + 0.008248626254498959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0007201223634183407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a boundless sea of thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive wooden warship with dragon carvings floating above the clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot a low-angle long shot that rises slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a massive sunset with golden backlight creating a silhouette on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ancient flags fluttering gently from the ship's masts?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890553817152977 + ], + "opposite_logprob_tokens": [ + -2.464055299758911 + ], + "logprob_reward": -0.0890553817152977, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890553817152977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09406870603561401, + 0.023483458906412125, + 0.10602040588855743, + 0.22767622768878937, + 0.13560697436332703, + 0.10151629894971848, + 0.09058231860399246, + 0.02701965905725956, + 0.027134014293551445, + 0.166891947388649 + ], + "credit_map": [ + 0.009263570420444012, + 0.01000779028981924, + 0.010997738689184189, + 0.01640370860695839, + 0.011577571742236614, + 0.009867190383374691, + 0.012993760406970978, + 0.01295736338943243, + 0.0018323496915400028, + 0.001965033356100321, + 0.0023805727250874043, + 0.003113627200946212, + 0.004208100959658623, + 0.003521206323057413, + 0.003056890796869993, + 0.0034056769218295813, + 0.0053935302421450615, + 0.0036001161206513643, + 0.009515330195426941, + 0.009495682083070278, + 0.00518216285854578, + 0.004242866765707731, + 0.035861220210790634, + 0.03272949531674385, + 0.03208480402827263, + 0.0261161457747221, + 0.028695201501250267, + 0.039508916437625885, + 0.012171010486781597, + 0.0120993135496974, + 0.031763896346092224, + 0.04523691534996033, + 0.004712384659796953, + 0.0023897651117295027, + 0.05968872457742691, + 0.030613698065280914, + 0.0025317457038909197, + 0.003818538272753358, + 0.01582128368318081, + 0.016030829399824142, + 0.005411997437477112, + 0.005145727191120386, + 0.01569097861647606, + 0.012284045107662678, + 0.013937830924987793, + 0.01837700419127941, + 0.01717265509068966, + 0.013496051542460918, + 0.006581704132258892, + 0.005196648184210062, + 0.01853998750448227, + 0.012503151781857014, + 0.013210948556661606, + 0.013293263502418995, + 0.012322777882218361, + 0.008933821693062782, + 0.0030034708324819803, + 0.00405157869681716, + 0.0032305491622537374, + 0.0026354389265179634, + 0.0025204920675605536, + 0.004226407036185265, + 0.0037852085661143064, + 0.003566512605175376, + 0.004516857210546732, + 0.0023696892894804478, + 0.0038510889280587435, + 0.002746256999671459, + 0.0030169193632900715, + 0.003760069841518998, + 0.003484623273834586, + 0.0033885068260133266, + 0.015925506129860878, + 0.011192812584340572, + 0.00833178497850895, + 0.008717796765267849, + 0.03167880326509476, + 0.07073818892240524, + 0.009763618931174278, + 0.01054343581199646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.0890553817152977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_6/index_tensor([5013, 5013])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_6/index_tensor([5013, 5013])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a479ea36b9ea555262816aaf6ccce84cb4445bdb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_6/index_tensor([5013, 5013])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "single, fixed, medium-long shot, shot from a level angle, captures the breathtaking moment of two massive, adult Siberian tigers locked in a desperate struggle in a vast, snow-covered coniferous forest clearing. Against a cold, gray sky, towering, snow-laden pine trees rise in the background, and a thick blanket of powdery snow blankets the ground. The two tigers' muscular bodies clash fiercely, their thick, shaggy orange and black fur striking in the cold light, their white manes on their necks and bellies blending seamlessly into the snow. One pushes off its hind legs, lowering its body, muscles tensed, poised to strike, its maw open wide, baring sharp canines and emitting a deep growl. The other rears up, one massive front paw raised high, claws fully extended, poised to deliver a fatal blow. Each of their powerful movements stirs up the deep snow beneath their feet. Their massive claws and powerful tails whip up vast expanses of powdered snow, creating a whirlwind of dancing white that instantly blurs the contours of their bodies, leaving only dynamic silhouettes. The entire scene is imbued with a sense of primal and brutal power. The stark white and gray tones highlight the vibrant colors of the tigers, creating a breathtaking atmosphere of intense confrontation filled with the beauty of wildness.", + "seed": 32869, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_6/index_5013_seed_32869_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.023746134992688895, + "vlm_logprob": -0.023746134992688895, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.023746134992688895, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02737080119550228 + ], + "opposite_logprob_tokens": [ + -3.9023706912994385 + ], + "logprob_reward": -0.02737080119550228, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02737080119550228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14446881413459778, + 0.21113435924053192, + 0.07266023010015488, + 0.09736773371696472, + 0.14525721967220306, + 0.1305149793624878, + 0.03325383737683296, + 0.07531890273094177, + 0.037022385746240616, + 0.05300155654549599 + ], + "credit_map": [ + 0.01891481876373291, + 0.011402899399399757, + 0.022045284509658813, + 0.018630631268024445, + 0.016075652092695236, + 0.010005060583353043, + 0.030938509851694107, + 0.016455957666039467, + 0.01842130906879902, + 0.017702320590615273, + 0.030439559370279312, + 0.031356338411569595, + 0.016401346772909164, + 0.019391683861613274, + 0.05192989856004715, + 0.02549191378057003, + 0.007148568984121084, + 0.012722921557724476, + 0.009767744690179825, + 0.011254672892391682, + 0.010579409077763557, + 0.007493470795452595, + 0.009403830394148827, + 0.004289609845727682, + 0.011884808540344238, + 0.01076517067849636, + 0.014616594649851322, + 0.017638860270380974, + 0.012058514170348644, + 0.006627601571381092, + 0.013991507701575756, + 0.009784682653844357, + 0.019140642136335373, + 0.019169140607118607, + 0.01904158480465412, + 0.022334741428494453, + 0.012597194872796535, + 0.016436448320746422, + 0.022551748901605606, + 0.013985718600451946, + 0.020168939605355263, + 0.009965813718736172, + 0.017365194857120514, + 0.017396604642271996, + 0.014848760329186916, + 0.011298991739749908, + 0.012323862873017788, + 0.027146805077791214, + 0.005093802697956562, + 0.0029275785200297832, + 0.0033293713349848986, + 0.005355206783860922, + 0.003291146829724312, + 0.003672039369121194, + 0.00530975591391325, + 0.004274934530258179, + 0.008918412029743195, + 0.004666287451982498, + 0.009437288157641888, + 0.01712646335363388, + 0.009155269712209702, + 0.007621866650879383, + 0.010421971790492535, + 0.007971341721713543, + 0.004387439228594303, + 0.0028719143010675907, + 0.00532294437289238, + 0.00614976417273283, + 0.0032130868639796972, + 0.003077551256865263, + 0.006355940829962492, + 0.005643743555992842, + 0.005578210111707449, + 0.004622629377990961, + 0.004486422520130873, + 0.00862143188714981, + 0.005586311221122742, + 0.003595868358388543, + 0.01359121035784483, + 0.006919478997588158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.02737080119550228 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005973581690341234 + ], + "opposite_logprob_tokens": [ + -5.130973815917969 + ], + "logprob_reward": -0.005973581690341234, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005973581690341234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19652406871318817, + 0.13957567512989044, + 0.07527068257331848, + 0.16559892892837524, + 0.12474530190229416, + 0.06615383923053741, + 0.08434128016233444, + 0.07201411575078964, + 0.03713800385594368, + 0.03863803297281265 + ], + "credit_map": [ + 0.028237808495759964, + 0.01923995465040207, + 0.027411581948399544, + 0.023536933586001396, + 0.024956971406936646, + 0.020904526114463806, + 0.029377706348896027, + 0.02285861037671566, + 0.014260580763220787, + 0.014014304615557194, + 0.021894220262765884, + 0.021042663604021072, + 0.010926367715001106, + 0.014782972633838654, + 0.02105070650577545, + 0.021603869274258614, + 0.01398004125803709, + 0.005496088415384293, + 0.010530857369303703, + 0.007774324156343937, + 0.011103849858045578, + 0.009040290489792824, + 0.009813813492655754, + 0.007531433831900358, + 0.024815622717142105, + 0.017580285668373108, + 0.020310085266828537, + 0.021121198311448097, + 0.029396694153547287, + 0.013192023150622845, + 0.021726984530687332, + 0.017456041648983955, + 0.016976643353700638, + 0.01825524866580963, + 0.017108788713812828, + 0.02305515483021736, + 0.01082040835171938, + 0.013560115359723568, + 0.013974906876683235, + 0.010994057171046734, + 0.009595139883458614, + 0.007387401536107063, + 0.009260901249945164, + 0.010622095316648483, + 0.008657735772430897, + 0.0057973721995949745, + 0.006446602754294872, + 0.008386597968637943, + 0.012483702041208744, + 0.007883541285991669, + 0.006583638023585081, + 0.010744430124759674, + 0.007658736314624548, + 0.01537275779992342, + 0.011344329454004765, + 0.012270152568817139, + 0.014587581157684326, + 0.005293697584420443, + 0.00437692878767848, + 0.020717762410640717, + 0.005693390499800444, + 0.007322497665882111, + 0.007169153541326523, + 0.006853117141872644, + 0.005463939160108566, + 0.002335061552003026, + 0.00451293308287859, + 0.004192864987999201, + 0.002766721649095416, + 0.005129366181790829, + 0.0038273585960268974, + 0.00890976283699274, + 0.006706275045871735, + 0.0035220624413341284, + 0.0037110501434653997, + 0.004707282409071922, + 0.004150350112468004, + 0.004344078246504068, + 0.007709216326475143, + 0.003787725931033492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.005973581690341234 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306728392839432 + ], + "opposite_logprob_tokens": [ + -3.168067216873169 + ], + "logprob_reward": -0.04306728392839432, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306728392839432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07317923754453659, + 0.10371995717287064, + 0.12075741589069366, + 0.32940673828125, + 0.08631252497434616, + 0.014092712663114071, + 0.024556202813982964, + 0.0923079326748848, + 0.05527302995324135, + 0.10039425641298294 + ], + "credit_map": [ + 0.008232164196670055, + 0.006117852404713631, + 0.010933052748441696, + 0.0075584352016448975, + 0.008350895717740059, + 0.00467843608930707, + 0.02130001038312912, + 0.006008396856486797, + 0.010813736356794834, + 0.008409461006522179, + 0.01666731759905815, + 0.014470607042312622, + 0.0098893903195858, + 0.010209517553448677, + 0.017392542213201523, + 0.015867391601204872, + 0.018098216503858566, + 0.01885559782385826, + 0.016620812937617302, + 0.01118231751024723, + 0.013742217794060707, + 0.018827352672815323, + 0.01506959181278944, + 0.00836130976676941, + 0.05446944385766983, + 0.036030691117048264, + 0.042166952043771744, + 0.03668484091758728, + 0.06337746232748032, + 0.025030257180333138, + 0.03650015592575073, + 0.035146910697221756, + 0.012101578526198864, + 0.01240304671227932, + 0.012731092050671577, + 0.015451976098120213, + 0.007486393209546804, + 0.010055185295641422, + 0.009430818259716034, + 0.006652432959526777, + 0.002321807434782386, + 0.0013552799355238676, + 0.0018478883430361748, + 0.0019030219409614801, + 0.0012834768276661634, + 0.0017817193875089288, + 0.0014647983480244875, + 0.0021347207948565483, + 0.00407833606004715, + 0.0018473758827894926, + 0.002226422540843487, + 0.0042207371443510056, + 0.0030255895107984543, + 0.0024203055072575808, + 0.0028667619917541742, + 0.0038706737104803324, + 0.010088145732879639, + 0.004908464848995209, + 0.012069480493664742, + 0.025791630148887634, + 0.009581414051353931, + 0.008524485863745213, + 0.010974431410431862, + 0.010369883850216866, + 0.006337513215839863, + 0.00398642010986805, + 0.0058926683850586414, + 0.005259850528091192, + 0.00640742015093565, + 0.007255500182509422, + 0.007759042549878359, + 0.012374616228044033, + 0.01095594372600317, + 0.009149202145636082, + 0.008390858769416809, + 0.015903344377875328, + 0.013527565635740757, + 0.008443259634077549, + 0.023570606485009193, + 0.010453475639224052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.04306728392839432 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537001632153988 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012537001632153988, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537001632153988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11287262290716171, + 0.0857490673661232, + 0.11816972494125366, + 0.057631030678749084, + 0.3738032281398773, + 0.032753169536590576, + 0.020322993397712708, + 0.05922143906354904, + 0.026856904849410057, + 0.11261981725692749 + ], + "credit_map": [ + 0.014768767170608044, + 0.00784469023346901, + 0.01563686691224575, + 0.013113124296069145, + 0.013570765033364296, + 0.009281646460294724, + 0.0209841076284647, + 0.017672667279839516, + 0.010705960914492607, + 0.007574634626507759, + 0.012563048861920834, + 0.0111484844237566, + 0.006974386516958475, + 0.007987946271896362, + 0.016764912754297256, + 0.012029699981212616, + 0.020160479471087456, + 0.014940176159143448, + 0.016025589779019356, + 0.00960618071258068, + 0.013950355350971222, + 0.02071390487253666, + 0.012975262477993965, + 0.009797781705856323, + 0.005391423124819994, + 0.005772181786596775, + 0.008292180486023426, + 0.010648949071764946, + 0.008134116418659687, + 0.005073958542197943, + 0.008807157166302204, + 0.005511065479367971, + 0.04955621436238289, + 0.04406050592660904, + 0.052347321063280106, + 0.06913133710622787, + 0.042234331369400024, + 0.03590346500277519, + 0.04980947822332382, + 0.030760565772652626, + 0.00597069738432765, + 0.003953567240387201, + 0.0051604826003313065, + 0.004292897880077362, + 0.0035309172235429287, + 0.0037972130812704563, + 0.0028869302477687597, + 0.0031604645773768425, + 0.0026237929705530405, + 0.0020608650520443916, + 0.0018523441394791007, + 0.002601040294393897, + 0.0019019844476133585, + 0.0037321096751838923, + 0.00347356079146266, + 0.0020772970747202635, + 0.008819409646093845, + 0.004148722626268864, + 0.008175491355359554, + 0.010397983714938164, + 0.0071679106913506985, + 0.00576894311234355, + 0.007276797667145729, + 0.0074661788530647755, + 0.0029321766924113035, + 0.0019341125153005123, + 0.0034912563860416412, + 0.0036789337173104286, + 0.0026601727586239576, + 0.002453672233968973, + 0.004138210788369179, + 0.005568373948335648, + 0.013010744005441666, + 0.009050514549016953, + 0.007942426018416882, + 0.018819184973835945, + 0.009386667050421238, + 0.011021139100193977, + 0.030148038640618324, + 0.01324111595749855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.012537001632153988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978200651705265 + ], + "opposite_logprob_tokens": [ + -3.5297820568084717 + ], + "logprob_reward": -0.02978200651705265, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978200651705265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08702774345874786, + 0.2501024901866913, + 0.1188354417681694, + 0.05710308626294136, + 0.2639700472354889, + 0.04529165104031563, + 0.03898747265338898, + 0.04180712625384331, + 0.047588471323251724, + 0.049286480993032455 + ], + "credit_map": [ + 0.013475896790623665, + 0.00941148865967989, + 0.013197008520364761, + 0.009446762502193451, + 0.010259008035063744, + 0.012286331504583359, + 0.010390153154730797, + 0.008561094291508198, + 0.02715354599058628, + 0.021293966099619865, + 0.038046944886446, + 0.03402986750006676, + 0.016996892169117928, + 0.02706805430352688, + 0.04129664599895477, + 0.044216565787792206, + 0.02227761223912239, + 0.01118071936070919, + 0.016976308077573776, + 0.010193919762969017, + 0.01382511481642723, + 0.019364504143595695, + 0.013435433618724346, + 0.011581835336983204, + 0.007462329231202602, + 0.00571154011413455, + 0.008149831555783749, + 0.009219592437148094, + 0.008591808378696442, + 0.004198454786092043, + 0.009550410322844982, + 0.004219120368361473, + 0.03810020908713341, + 0.028635820373892784, + 0.03825785964727402, + 0.05072692781686783, + 0.027468277141451836, + 0.022774076089262962, + 0.036680664867162704, + 0.021326223388314247, + 0.009570117108523846, + 0.003628073027357459, + 0.005409559700638056, + 0.00784300547093153, + 0.003933236002922058, + 0.002705365652218461, + 0.0035034818574786186, + 0.008698814548552036, + 0.006031847558915615, + 0.003703563241288066, + 0.003602387383580208, + 0.006986913271248341, + 0.004063691012561321, + 0.005269405897706747, + 0.0038506619166582823, + 0.005479004234075546, + 0.005046626087278128, + 0.002317117527127266, + 0.00633591553196311, + 0.013261393643915653, + 0.002934208605438471, + 0.0027435130905359983, + 0.004534080158919096, + 0.004634274635463953, + 0.00654574716463685, + 0.003632795996963978, + 0.008260867558419704, + 0.00911956001073122, + 0.0034589278511703014, + 0.004275164566934109, + 0.0044869850389659405, + 0.007808424532413483, + 0.005865954328328371, + 0.0037280824035406113, + 0.00409467238932848, + 0.007229940965771675, + 0.0036620062310248613, + 0.0038747338112443686, + 0.012841245159506798, + 0.007989847101271152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.02978200651705265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, fixed, medium-long shot, shot from a level angle, captures the breathtaking moment of two massive, adult Siberian tigers locked in a desperate struggle in a vast, snow-covered coniferous forest clearing. Against a cold, gray sky, towering, snow-laden pine trees rise in the background, and a thick blanket of powdery snow blankets the ground. The two tigers' muscular bodies clash fiercely, their thick, shaggy orange and black fur striking in the cold light, their white manes on their necks and bellies blending seamlessly into the snow. One pushes off its hind legs, lowering its body, muscles tensed, poised to strike, its maw open wide, baring sharp canines and emitting a deep growl. The other rears up, one massive front paw raised high, claws fully extended, poised to deliver a fatal blow. Each of their powerful movements stirs up the deep snow beneath their feet. Their massive claws and powerful tails whip up vast expanses of powdered snow, creating a whirlwind of dancing white that instantly blurs the contours of their bodies, leaving only dynamic silhouettes. The entire scene is imbued with a sense of primal and brutal power. The stark white and gray tones highlight the vibrant colors of the tigers, creating a breathtaking atmosphere of intense confrontation filled with the beauty of wildness.", + "seed": 32870, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_6/index_5013_seed_32870_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.046968520153313875, + "vlm_logprob": -0.046968520153313875, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.046968520153313875, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09428209066390991 + ], + "opposite_logprob_tokens": [ + -2.4692821502685547 + ], + "logprob_reward": -0.09428209066390991, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09428209066390991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10846984386444092, + 0.13227593898773193, + 0.08914557099342346, + 0.1242455393075943, + 0.2139192372560501, + 0.17142674326896667, + 0.07294253259897232, + 0.037125248461961746, + 0.04209762066602707, + 0.008351719006896019 + ], + "credit_map": [ + 0.012379098683595657, + 0.010108393616974354, + 0.011548381298780441, + 0.019319789484143257, + 0.010579262860119343, + 0.007508524227887392, + 0.01801329106092453, + 0.019013099372386932, + 0.011199157685041428, + 0.01269963476806879, + 0.016269700601696968, + 0.020401939749717712, + 0.021256359294056892, + 0.008963722735643387, + 0.029346434399485588, + 0.012138999998569489, + 0.01028949674218893, + 0.014791559427976608, + 0.011145928874611855, + 0.016135763376951218, + 0.009567772038280964, + 0.006636255886405706, + 0.012198222801089287, + 0.008380567654967308, + 0.01069840881973505, + 0.014418643899261951, + 0.01711840182542801, + 0.019469130784273148, + 0.013477430678904057, + 0.01053428091108799, + 0.024203909561038017, + 0.014325334690511227, + 0.023459386080503464, + 0.013202229514718056, + 0.039795536547899246, + 0.04119556397199631, + 0.019741076976060867, + 0.017092391848564148, + 0.03332684561610222, + 0.02610618993639946, + 0.023159202188253403, + 0.01849447377026081, + 0.024189136922359467, + 0.018635673448443413, + 0.019035158678889275, + 0.014769069850444794, + 0.02784152142703533, + 0.025302499532699585, + 0.005997771397233009, + 0.0022476401645690203, + 0.019478902220726013, + 0.00646384758874774, + 0.0035430637653917074, + 0.0023499939125031233, + 0.028871525079011917, + 0.00398979103192687, + 0.003540415782481432, + 0.002481867792084813, + 0.005155045073479414, + 0.006144477520138025, + 0.0033694859594106674, + 0.0023085614666342735, + 0.008198822848498821, + 0.005926573183387518, + 0.003912453074008226, + 0.002886497648432851, + 0.006478011142462492, + 0.005495225545018911, + 0.005766492802649736, + 0.002521292306482792, + 0.00946677103638649, + 0.005570875946432352, + 0.0008918982348404825, + 0.0005105963791720569, + 0.0008161955629475415, + 0.0013751076767221093, + 0.0016529938438907266, + 0.0009619806078262627, + 0.0010548137361183763, + 0.0010881328489631414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.09428209066390991 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005278934258967638 + ], + "opposite_logprob_tokens": [ + -5.255279064178467 + ], + "logprob_reward": -0.005278934258967638, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005278934258967638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1474921703338623, + 0.0567658357322216, + 0.2172950953245163, + 0.10329514741897583, + 0.05798327922821045, + 0.06528477370738983, + 0.1696939468383789, + 0.038845423609018326, + 0.08982937037944794, + 0.053515031933784485 + ], + "credit_map": [ + 0.013818892650306225, + 0.010070035234093666, + 0.01592189073562622, + 0.02843482419848442, + 0.018277054652571678, + 0.013389029540121555, + 0.027396759018301964, + 0.020183665677905083, + 0.007591744419187307, + 0.005075015593320131, + 0.006441144738346338, + 0.00910694058984518, + 0.00647530984133482, + 0.00559820793569088, + 0.011482041329145432, + 0.004995424300432205, + 0.022056005895137787, + 0.030520303174853325, + 0.03044803999364376, + 0.0444926954805851, + 0.021336032077670097, + 0.015784380957484245, + 0.03303905203938484, + 0.01961856707930565, + 0.010196683928370476, + 0.016432754695415497, + 0.011744525283575058, + 0.010380322113633156, + 0.014389045536518097, + 0.011880134232342243, + 0.014380427077412605, + 0.013891251757740974, + 0.005989256780594587, + 0.004116276279091835, + 0.00992912519723177, + 0.008709598332643509, + 0.007152434438467026, + 0.0057364581152796745, + 0.009700482711195946, + 0.0066496399231255054, + 0.006257563829421997, + 0.006820353679358959, + 0.011964619159698486, + 0.010421558283269405, + 0.005269291345030069, + 0.005108239594846964, + 0.01230583991855383, + 0.007137302774935961, + 0.0144613953307271, + 0.005536743439733982, + 0.045206766575574875, + 0.015422348864376545, + 0.011117645539343357, + 0.0065327961929142475, + 0.06314916163682938, + 0.008267063647508621, + 0.004378431476652622, + 0.005047767423093319, + 0.004782691597938538, + 0.005435549654066563, + 0.00455275597050786, + 0.0038894887547940016, + 0.006762955337762833, + 0.003995777573436499, + 0.011015888303518295, + 0.008009728044271469, + 0.010550616309046745, + 0.007981613278388977, + 0.010031897574663162, + 0.010880514979362488, + 0.01312616653740406, + 0.018232939764857292, + 0.006343475542962551, + 0.004755096044391394, + 0.00611460953950882, + 0.00844745896756649, + 0.006968976464122534, + 0.008152643218636513, + 0.0053270901553332806, + 0.007405674085021019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.005278934258967638 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899737358093262 + ], + "opposite_logprob_tokens": [ + -2.5789973735809326 + ], + "logprob_reward": -0.07899737358093262, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899737358093262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18135520815849304, + 0.07734373956918716, + 0.06323093175888062, + 0.049239397048950195, + 0.06160154193639755, + 0.13844658434391022, + 0.3058352470397949, + 0.05575127154588699, + 0.04220199212431908, + 0.02499408833682537 + ], + "credit_map": [ + 0.02097276784479618, + 0.015712795779109, + 0.013638373464345932, + 0.030526287853717804, + 0.014595866203308105, + 0.012372640892863274, + 0.03957560285925865, + 0.03396085649728775, + 0.008520365692675114, + 0.006190343759953976, + 0.007442591246217489, + 0.011035568080842495, + 0.01730172149837017, + 0.006531982216984034, + 0.012080914340913296, + 0.00824025459587574, + 0.009050664491951466, + 0.013159053400158882, + 0.00610100943595171, + 0.007125671021640301, + 0.007169083226472139, + 0.0060144406743347645, + 0.007813870906829834, + 0.0067971390672028065, + 0.004559656139463186, + 0.00505627179518342, + 0.006272229831665754, + 0.009002828039228916, + 0.0061920806765556335, + 0.004312123637646437, + 0.008819270879030228, + 0.005024935118854046, + 0.008102646097540855, + 0.004008682444691658, + 0.010984210297465324, + 0.012139789760112762, + 0.005586754065006971, + 0.004678421653807163, + 0.008923676796257496, + 0.007177356630563736, + 0.02014446072280407, + 0.011301578022539616, + 0.019154494628310204, + 0.01602132059633732, + 0.015360680408775806, + 0.010778112336993217, + 0.029251428321003914, + 0.01643451489508152, + 0.024561256170272827, + 0.008712144568562508, + 0.07976452261209488, + 0.027698451653122902, + 0.018323376774787903, + 0.011902650818228722, + 0.1183595284819603, + 0.016513315960764885, + 0.004997415468096733, + 0.004395280033349991, + 0.007311251014471054, + 0.009358969517052174, + 0.0059853168204426765, + 0.0037215095944702625, + 0.011653288267552853, + 0.008328240364789963, + 0.0036740852519869804, + 0.0030178844463080168, + 0.005704302340745926, + 0.005888062063604593, + 0.0064096455462276936, + 0.002834254177287221, + 0.009067622013390064, + 0.005606137681752443, + 0.003241592552512884, + 0.002405818086117506, + 0.004370345268398523, + 0.0032443657983094454, + 0.0023728448431938887, + 0.0019364285981282592, + 0.004442465957254171, + 0.0029802266508340836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.07899737358093262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810567781329155 + ], + "opposite_logprob_tokens": [ + -3.2881057262420654 + ], + "logprob_reward": -0.03810567781329155, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810567781329155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07154318690299988, + 0.08841414749622345, + 0.07553049921989441, + 0.09244080632925034, + 0.026761779561638832, + 0.24985726177692413, + 0.26335686445236206, + 0.039846763014793396, + 0.021388189867138863, + 0.07086049765348434 + ], + "credit_map": [ + 0.008630828000605106, + 0.008794101886451244, + 0.007204442285001278, + 0.01485404185950756, + 0.006559956353157759, + 0.006962894927710295, + 0.009696785360574722, + 0.00884013157337904, + 0.009754212573170662, + 0.010509197600185871, + 0.007962122559547424, + 0.008536833338439465, + 0.025575434789061546, + 0.008900217711925507, + 0.009023898281157017, + 0.008152234368026257, + 0.010324348695576191, + 0.010897777043282986, + 0.01111501082777977, + 0.012272450141608715, + 0.007713424041867256, + 0.006466064136475325, + 0.009522244334220886, + 0.007219184655696154, + 0.008892943151295185, + 0.011271703988313675, + 0.011085127480328083, + 0.014636716805398464, + 0.010324825532734394, + 0.008271411992609501, + 0.018196359276771545, + 0.009761715307831764, + 0.003573428373783827, + 0.001991493394598365, + 0.0037272365298122168, + 0.0035740449093282223, + 0.004176025278866291, + 0.002891859272494912, + 0.004246990662068129, + 0.0025807011406868696, + 0.025487491860985756, + 0.02906791865825653, + 0.03656689077615738, + 0.03248303756117821, + 0.025086211040616035, + 0.019295183941721916, + 0.048117175698280334, + 0.03375336155295372, + 0.021642854437232018, + 0.005997936706990004, + 0.07255402207374573, + 0.02430068701505661, + 0.013596980832517147, + 0.009698850102722645, + 0.10221974551677704, + 0.013345788232982159, + 0.0049803731963038445, + 0.004227468278259039, + 0.003999160602688789, + 0.004689868539571762, + 0.007509454619139433, + 0.003117177402600646, + 0.005215243436396122, + 0.006108016707003117, + 0.001884425524622202, + 0.0024604233913123608, + 0.0029150729533284903, + 0.002028106478974223, + 0.003307278733700514, + 0.001901974668726325, + 0.0041581070981919765, + 0.0027328026480972767, + 0.007583015598356724, + 0.0049498057924211025, + 0.006829464808106422, + 0.011663969606161118, + 0.010717337019741535, + 0.006148921325802803, + 0.008710579946637154, + 0.014257400296628475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.03810567781329155 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817852444946766 + ], + "opposite_logprob_tokens": [ + -4.018178462982178 + ], + "logprob_reward": -0.01817852444946766, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817852444946766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1337374448776245, + 0.04126732423901558, + 0.0514909103512764, + 0.043367162346839905, + 0.056293319910764694, + 0.3431849479675293, + 0.11697462201118469, + 0.027702614665031433, + 0.11690636724233627, + 0.06907526403665543 + ], + "credit_map": [ + 0.014324343763291836, + 0.011043338105082512, + 0.01200453843921423, + 0.018873853608965874, + 0.011405512690544128, + 0.01187165267765522, + 0.035418830811977386, + 0.01879539154469967, + 0.006112446542829275, + 0.00364297884516418, + 0.004644045606255531, + 0.005447126924991608, + 0.008053827099502087, + 0.0054025002755224705, + 0.004483513999730349, + 0.0034808896016329527, + 0.007480595726519823, + 0.004073480609804392, + 0.007526590954512358, + 0.008591452613472939, + 0.005471864715218544, + 0.005270638503134251, + 0.007798680569976568, + 0.0052776094526052475, + 0.004160073585808277, + 0.0049200900830328465, + 0.005618078634142876, + 0.007487852592021227, + 0.0045492141507565975, + 0.0041802613995969296, + 0.007586516905575991, + 0.004865078255534172, + 0.010251238010823727, + 0.006814983673393726, + 0.006089712493121624, + 0.009747451171278954, + 0.005565715488046408, + 0.004186511039733887, + 0.0072220442816615105, + 0.006415673065930605, + 0.03988195210695267, + 0.039596036076545715, + 0.04676029458642006, + 0.04518360272049904, + 0.03631236031651497, + 0.027732068672776222, + 0.06280151009559631, + 0.044917140156030655, + 0.00916469655930996, + 0.003407298354431987, + 0.03246339410543442, + 0.010550301522016525, + 0.006062423810362816, + 0.004408514592796564, + 0.04475390911102295, + 0.006164093967527151, + 0.0031573993619531393, + 0.002388585824519396, + 0.0032776049338281155, + 0.004097004886716604, + 0.003324928693473339, + 0.002022388856858015, + 0.005757567938417196, + 0.0036771364975720644, + 0.011693648993968964, + 0.013930011540651321, + 0.014602736569941044, + 0.012369344010949135, + 0.014913070946931839, + 0.00963010173290968, + 0.022615309804677963, + 0.017152149230241776, + 0.007930280640721321, + 0.005070962011814117, + 0.009013411588966846, + 0.009600933641195297, + 0.009165025316178799, + 0.007421270478516817, + 0.009343964979052544, + 0.011529417708516121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.01817852444946766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, fixed, medium-long shot, shot from a level angle, captures the breathtaking moment of two massive, adult Siberian tigers locked in a desperate struggle in a vast, snow-covered coniferous forest clearing. Against a cold, gray sky, towering, snow-laden pine trees rise in the background, and a thick blanket of powdery snow blankets the ground. The two tigers' muscular bodies clash fiercely, their thick, shaggy orange and black fur striking in the cold light, their white manes on their necks and bellies blending seamlessly into the snow. One pushes off its hind legs, lowering its body, muscles tensed, poised to strike, its maw open wide, baring sharp canines and emitting a deep growl. The other rears up, one massive front paw raised high, claws fully extended, poised to deliver a fatal blow. Each of their powerful movements stirs up the deep snow beneath their feet. Their massive claws and powerful tails whip up vast expanses of powdered snow, creating a whirlwind of dancing white that instantly blurs the contours of their bodies, leaving only dynamic silhouettes. The entire scene is imbued with a sense of primal and brutal power. The stark white and gray tones highlight the vibrant colors of the tigers, creating a breathtaking atmosphere of intense confrontation filled with the beauty of wildness.", + "seed": 32871, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_6/index_5013_seed_32871_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028620212711393832, + "vlm_logprob": -0.028620212711393832, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.028620212711393832, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041765306144952774 + ], + "opposite_logprob_tokens": [ + -3.2917652130126953 + ], + "logprob_reward": -0.041765306144952774, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.041765306144952774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13226217031478882, + 0.038651738315820694, + 0.031466465443372726, + 0.1173039972782135, + 0.1825781762599945, + 0.08288685232400894, + 0.2376987636089325, + 0.0300136785954237, + 0.033906638622283936, + 0.11323151737451553 + ], + "credit_map": [ + 0.01491304486989975, + 0.013153324835002422, + 0.022641312330961227, + 0.03531385958194733, + 0.006831330247223377, + 0.007829742506146431, + 0.01952570676803589, + 0.012053845450282097, + 0.004136852454394102, + 0.0034089076798409224, + 0.005862227641046047, + 0.008802080526947975, + 0.0030629229731857777, + 0.002658562269061804, + 0.005548631772398949, + 0.005171553231775761, + 0.003513054456561804, + 0.002610259922221303, + 0.004949466325342655, + 0.005842939950525761, + 0.0039134579710662365, + 0.00157552445307374, + 0.00560124171897769, + 0.003460518317297101, + 0.013537716120481491, + 0.015248923562467098, + 0.016086554154753685, + 0.019117826595902443, + 0.016712574288249016, + 0.008043256588280201, + 0.013777540996670723, + 0.014779597520828247, + 0.014241018332540989, + 0.0146345105022192, + 0.01743197813630104, + 0.026739079505205154, + 0.00942263100296259, + 0.017087290063500404, + 0.05800505727529526, + 0.02501661516726017, + 0.00894093420356512, + 0.006995575036853552, + 0.007153801154345274, + 0.017735954374074936, + 0.005212624557316303, + 0.00910896435379982, + 0.009842058643698692, + 0.017896946519613266, + 0.023047730326652527, + 0.01921045407652855, + 0.028365256264805794, + 0.030512169003486633, + 0.02197309024631977, + 0.02357451803982258, + 0.038774482905864716, + 0.05224107578396797, + 0.006012353580445051, + 0.004242964554578066, + 0.0037417495623230934, + 0.0028131913859397173, + 0.0025602872483432293, + 0.004185644909739494, + 0.003440721658989787, + 0.003016767557710409, + 0.0037034284323453903, + 0.004098760429769754, + 0.004407625645399094, + 0.0038822456263005733, + 0.004595013801008463, + 0.004190001171082258, + 0.0050136856734752655, + 0.004015875514596701, + 0.014566614292562008, + 0.01167891826480627, + 0.010193386115133762, + 0.016227547079324722, + 0.013726181350648403, + 0.009754872880876064, + 0.026686931028962135, + 0.010397062636911869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.041765306144952774 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653167635202408 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.008653167635202408, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653167635202408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2342241108417511, + 0.1302628070116043, + 0.1202395111322403, + 0.029840240254998207, + 0.17088522017002106, + 0.0626426562666893, + 0.11287104338407516, + 0.02666497975587845, + 0.037670522928237915, + 0.07469889521598816 + ], + "credit_map": [ + 0.02079622820019722, + 0.01562244351953268, + 0.04225606098771095, + 0.06317423284053802, + 0.01723446510732174, + 0.012752614915370941, + 0.03640000894665718, + 0.025988051667809486, + 0.02387045882642269, + 0.014417879283428192, + 0.021036075428128242, + 0.023981476202607155, + 0.015373199246823788, + 0.00798953790217638, + 0.012601529248058796, + 0.01099264807999134, + 0.017588963732123375, + 0.01670767180621624, + 0.0175496693700552, + 0.01690041832625866, + 0.009961165487766266, + 0.01031270157545805, + 0.017050139605998993, + 0.014168775640428066, + 0.003793115261942148, + 0.0038068899884819984, + 0.004076478537172079, + 0.00498020276427269, + 0.0033447726164013147, + 0.0024360253009945154, + 0.003920009825378656, + 0.0034827478229999542, + 0.022539883852005005, + 0.018243474885821342, + 0.01064425241202116, + 0.018968096002936363, + 0.009861249476671219, + 0.01651996374130249, + 0.05779724568128586, + 0.0163110364228487, + 0.007513837888836861, + 0.006753640249371529, + 0.005280723329633474, + 0.013747177086770535, + 0.0046151201240718365, + 0.007192776072770357, + 0.006872464902698994, + 0.010666911490261555, + 0.013691351749002934, + 0.015179365873336792, + 0.009499083273112774, + 0.013351642526686192, + 0.011858385056257248, + 0.015022724866867065, + 0.015816479921340942, + 0.018452003598213196, + 0.0036649424582719803, + 0.003233615541830659, + 0.0035915325861424208, + 0.002095249015837908, + 0.0025593142490833998, + 0.00483864126726985, + 0.004560123197734356, + 0.0021215626038610935, + 0.00356223713606596, + 0.004642333369702101, + 0.005255620460957289, + 0.005293725524097681, + 0.005188404116779566, + 0.004332688171416521, + 0.0057983240112662315, + 0.0035971887409687042, + 0.010105296969413757, + 0.006742340046912432, + 0.009571264497935772, + 0.010663378983736038, + 0.00876052025705576, + 0.007633690256625414, + 0.01344830822199583, + 0.007774097844958305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.008653167635202408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048652395606040955 + ], + "opposite_logprob_tokens": [ + -3.048652410507202 + ], + "logprob_reward": -0.048652395606040955, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048652395606040955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3078332841396332, + 0.05632900074124336, + 0.04053245857357979, + 0.07479073852300644, + 0.10055708885192871, + 0.04638560861349106, + 0.18347471952438354, + 0.0870257169008255, + 0.066146120429039, + 0.03692527115345001 + ], + "credit_map": [ + 0.023009981960058212, + 0.03867191821336746, + 0.0506008081138134, + 0.0777188092470169, + 0.015866773203015327, + 0.01056068018078804, + 0.05173367261886597, + 0.0396706685423851, + 0.00755594577640295, + 0.006960241124033928, + 0.00834670290350914, + 0.009830406866967678, + 0.00443563936278224, + 0.0047413865104317665, + 0.007165325805544853, + 0.00729335518553853, + 0.005080780014395714, + 0.004140560515224934, + 0.006196152418851852, + 0.005544650834053755, + 0.003694811835885048, + 0.0026742133777588606, + 0.006634180434048176, + 0.006567107513546944, + 0.00935788732022047, + 0.009284606203436852, + 0.01033925823867321, + 0.010480138473212719, + 0.01036685798317194, + 0.00433982303366065, + 0.011409292928874493, + 0.009212871082127094, + 0.00921708345413208, + 0.00851607695221901, + 0.012331072241067886, + 0.023235885426402092, + 0.004221978131681681, + 0.010018336586654186, + 0.01142258569598198, + 0.021594077348709106, + 0.007797685451805592, + 0.004540367983281612, + 0.004854967352002859, + 0.008271981962025166, + 0.00465188967064023, + 0.004744433797895908, + 0.005738153588026762, + 0.0057861278764903545, + 0.024015089496970177, + 0.016060197725892067, + 0.02454969473183155, + 0.023149892687797546, + 0.020990237593650818, + 0.012841888703405857, + 0.035826098173856735, + 0.026041612029075623, + 0.008010201156139374, + 0.006799164228141308, + 0.019185645505785942, + 0.007391124032437801, + 0.0050431955605745316, + 0.006433774717152119, + 0.025174705311655998, + 0.008987904526293278, + 0.006958429701626301, + 0.008354579098522663, + 0.010319838300347328, + 0.0073221828788518906, + 0.010287878103554249, + 0.005483000539243221, + 0.008903923444449902, + 0.008516283705830574, + 0.004450857173651457, + 0.003918409813195467, + 0.004422041587531567, + 0.004095937591046095, + 0.006385827902704477, + 0.0042321146465837955, + 0.004849029704928398, + 0.004571053199470043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.048652395606040955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949762649834156 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005949762649834156, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949762649834156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48527151346206665, + 0.03781052678823471, + 0.02740352228283882, + 0.12218807637691498, + 0.04237251356244087, + 0.045597318559885025, + 0.08151362091302872, + 0.09212055802345276, + 0.019685303792357445, + 0.046037059277296066 + ], + "credit_map": [ + 0.04588426277041435, + 0.03640636056661606, + 0.0818096250295639, + 0.12452489882707596, + 0.04565237835049629, + 0.046363070607185364, + 0.06386131793260574, + 0.04076959192752838, + 0.006716195028275251, + 0.005836492404341698, + 0.0058169858530163765, + 0.005339998751878738, + 0.003811344737187028, + 0.0027734979521483183, + 0.0038345863576978445, + 0.003681428264826536, + 0.004970523063093424, + 0.002827684162184596, + 0.004894528072327375, + 0.004213125444948673, + 0.0020924126729369164, + 0.0017302645137533545, + 0.003636398818343878, + 0.003038583090528846, + 0.01660086400806904, + 0.017555493861436844, + 0.018230080604553223, + 0.019925326108932495, + 0.014621228910982609, + 0.010139786638319492, + 0.013254566118121147, + 0.011860727332532406, + 0.003647211007773876, + 0.005141228437423706, + 0.004881001077592373, + 0.006549576297402382, + 0.0026119675021618605, + 0.005843843799084425, + 0.008500526659190655, + 0.005197158548980951, + 0.004644040018320084, + 0.003838551929220557, + 0.003927142824977636, + 0.007204845082014799, + 0.003624809207394719, + 0.004621429834514856, + 0.007236171513795853, + 0.010500327683985233, + 0.009639418683946133, + 0.008285061456263065, + 0.010310085490345955, + 0.009954272769391537, + 0.010302703827619553, + 0.0063630640506744385, + 0.015960652381181717, + 0.010698363184928894, + 0.007983856834471226, + 0.0064437855035066605, + 0.023837286978960037, + 0.006644761189818382, + 0.005286445375531912, + 0.006089585833251476, + 0.027705572545528412, + 0.008129255846142769, + 0.002017152961343527, + 0.0029794257134199142, + 0.002923820400610566, + 0.0024295123293995857, + 0.002140630967915058, + 0.0022267368622124195, + 0.0025482187047600746, + 0.0024198065511882305, + 0.007079325616359711, + 0.004683713428676128, + 0.0043002828024327755, + 0.008210897445678711, + 0.0033557757269591093, + 0.004064497537910938, + 0.010971303097903728, + 0.0033712598960846663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.005949762649834156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808043152093887 + ], + "opposite_logprob_tokens": [ + -3.2880804538726807 + ], + "logprob_reward": -0.03808043152093887, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808043152093887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09839949011802673, + 0.028371326625347137, + 0.028847895562648773, + 0.24548552930355072, + 0.10226340591907501, + 0.06090027838945389, + 0.04743708670139313, + 0.10839042067527771, + 0.05248028039932251, + 0.22742430865764618 + ], + "credit_map": [ + 0.011305050924420357, + 0.010744651779532433, + 0.016133248805999756, + 0.01806384138762951, + 0.008206261321902275, + 0.009055031463503838, + 0.014840357005596161, + 0.010051042772829533, + 0.004114755429327488, + 0.0032878071069717407, + 0.004437551833689213, + 0.00425153411924839, + 0.002607909496873617, + 0.0029298977460712194, + 0.003760349238291383, + 0.0029815216548740864, + 0.0037136452738195658, + 0.002649665344506502, + 0.004838470835238695, + 0.005312275607138872, + 0.0022339157294481993, + 0.001734501332975924, + 0.0046128807589411736, + 0.0037525398656725883, + 0.030748752877116203, + 0.03221246227622032, + 0.035328932106494904, + 0.042154569178819656, + 0.03302531689405441, + 0.017399176955223083, + 0.029123511165380478, + 0.02549281343817711, + 0.012698634527623653, + 0.011543074622750282, + 0.011785976588726044, + 0.01451010350137949, + 0.005455693230032921, + 0.015779031440615654, + 0.01626264862716198, + 0.014228243380784988, + 0.006848717108368874, + 0.00589380506426096, + 0.004466296639293432, + 0.011118659749627113, + 0.00416700029745698, + 0.005122039932757616, + 0.0077996812760829926, + 0.015484075993299484, + 0.004536289256066084, + 0.005547238048166037, + 0.007185815833508968, + 0.006056076381355524, + 0.005937389563769102, + 0.00441043172031641, + 0.0068212817423045635, + 0.006942563224583864, + 0.011379682458937168, + 0.010629374533891678, + 0.02373747155070305, + 0.007579443976283073, + 0.0072814710438251495, + 0.011442212387919426, + 0.021925413981080055, + 0.014415349811315536, + 0.0054761613719165325, + 0.006228350102901459, + 0.007290191948413849, + 0.00749660050496459, + 0.006613105535507202, + 0.005166143178939819, + 0.007611589040607214, + 0.006598139181733131, + 0.030013449490070343, + 0.024338675662875175, + 0.0227070152759552, + 0.02911508083343506, + 0.02710072509944439, + 0.019930211827158928, + 0.05053529515862465, + 0.023683855310082436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.03808043152093887 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, fixed, medium-long shot, shot from a level angle, captures the breathtaking moment of two massive, adult Siberian tigers locked in a desperate struggle in a vast, snow-covered coniferous forest clearing. Against a cold, gray sky, towering, snow-laden pine trees rise in the background, and a thick blanket of powdery snow blankets the ground. The two tigers' muscular bodies clash fiercely, their thick, shaggy orange and black fur striking in the cold light, their white manes on their necks and bellies blending seamlessly into the snow. One pushes off its hind legs, lowering its body, muscles tensed, poised to strike, its maw open wide, baring sharp canines and emitting a deep growl. The other rears up, one massive front paw raised high, claws fully extended, poised to deliver a fatal blow. Each of their powerful movements stirs up the deep snow beneath their feet. Their massive claws and powerful tails whip up vast expanses of powdered snow, creating a whirlwind of dancing white that instantly blurs the contours of their bodies, leaving only dynamic silhouettes. The entire scene is imbued with a sense of primal and brutal power. The stark white and gray tones highlight the vibrant colors of the tigers, creating a breathtaking atmosphere of intense confrontation filled with the beauty of wildness.", + "seed": 32872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_6/index_5013_seed_32872_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05324266096577048, + "vlm_logprob": -0.05324266096577048, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.05324266096577048, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10549498349428177 + ], + "opposite_logprob_tokens": [ + -2.355494976043701 + ], + "logprob_reward": -0.10549498349428177, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10549498349428177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07100467383861542, + 0.060557372868061066, + 0.029698951169848442, + 0.04717736691236496, + 0.30337536334991455, + 0.08314701914787292, + 0.09596245735883713, + 0.09238632023334503, + 0.17743533849716187, + 0.03925511613488197 + ], + "credit_map": [ + 0.012093073688447475, + 0.006604834459722042, + 0.009061607532203197, + 0.01321208942681551, + 0.006215492729097605, + 0.0067830332554876804, + 0.008982925675809383, + 0.00805161613970995, + 0.005784170236438513, + 0.005345728248357773, + 0.008120102807879448, + 0.011556351557374, + 0.007252553012222052, + 0.003848866792395711, + 0.013558377511799335, + 0.00509122246876359, + 0.004103729035705328, + 0.002451936714351177, + 0.003340719034895301, + 0.0051479400135576725, + 0.004007833544164896, + 0.002333875512704253, + 0.004842421505600214, + 0.0034704972058534622, + 0.010556104592978954, + 0.0044999136589467525, + 0.006924327928572893, + 0.00675454456359148, + 0.004303273279219866, + 0.0056435479782521725, + 0.004968416411429644, + 0.0035272431559860706, + 0.06820657104253769, + 0.021588550880551338, + 0.10680196434259415, + 0.03473440557718277, + 0.01421909499913454, + 0.014169605448842049, + 0.026217937469482422, + 0.01743723638355732, + 0.009671051055192947, + 0.008207941427826881, + 0.010391846299171448, + 0.015314002521336079, + 0.009442845359444618, + 0.007892372086644173, + 0.010801843367516994, + 0.011425118893384933, + 0.023145930841565132, + 0.006849796511232853, + 0.021664535626769066, + 0.01147822942584753, + 0.007345533464103937, + 0.0068544442765414715, + 0.013040105812251568, + 0.005583880469202995, + 0.012447082437574863, + 0.00656107859686017, + 0.010632055811583996, + 0.011380248703062534, + 0.02319290116429329, + 0.00672617182135582, + 0.012124861590564251, + 0.009321920573711395, + 0.023613354191184044, + 0.015314525924623013, + 0.015688493847846985, + 0.03231566771864891, + 0.02054372988641262, + 0.01885550282895565, + 0.030212843790650368, + 0.02089122124016285, + 0.007680179085582495, + 0.002309560077264905, + 0.010120914317667484, + 0.005313734523952007, + 0.0026564947329461575, + 0.0027878503315150738, + 0.004518142435699701, + 0.003868240164592862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.10549498349428177 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067563182674348354 + ], + "opposite_logprob_tokens": [ + -5.00675630569458 + ], + "logprob_reward": -0.0067563182674348354, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067563182674348354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04578603804111481, + 0.025655927136540413, + 0.034373749047517776, + 0.03016127087175846, + 0.6658179759979248, + 0.035997044295072556, + 0.047690290957689285, + 0.03876485303044319, + 0.06318163126707077, + 0.012571156024932861 + ], + "credit_map": [ + 0.006213585380464792, + 0.005440272856503725, + 0.005796661600470543, + 0.006656127981841564, + 0.004776134621351957, + 0.004805267788469791, + 0.005839768797159195, + 0.006258225068449974, + 0.0030501619912683964, + 0.003138133790344, + 0.0031608333811163902, + 0.00579387741163373, + 0.0025710451882332563, + 0.001647197175770998, + 0.0038705707993358374, + 0.002424111356958747, + 0.004433401860296726, + 0.006470727734267712, + 0.0036168841179460287, + 0.004396808333694935, + 0.004695129580795765, + 0.0023983882274478674, + 0.004365059081465006, + 0.00399735476821661, + 0.006472705397754908, + 0.002697213087230921, + 0.0035716022830456495, + 0.003246128326281905, + 0.004861120134592056, + 0.003852883819490671, + 0.0033872928470373154, + 0.0020723268389701843, + 0.07366843521595001, + 0.02722236141562462, + 0.27099570631980896, + 0.22740381956100464, + 0.017820551991462708, + 0.015327610075473785, + 0.022397322580218315, + 0.010982228443026543, + 0.0050685834139585495, + 0.002591436030343175, + 0.006852527614682913, + 0.00606065709143877, + 0.004092249553650618, + 0.002534921281039715, + 0.004749487154185772, + 0.004047185182571411, + 0.008462456986308098, + 0.0028763674199581146, + 0.016939887776970863, + 0.006816092878580093, + 0.002448431681841612, + 0.0020367675460875034, + 0.005891249515116215, + 0.0022190415766090155, + 0.0064027560874819756, + 0.004462649114429951, + 0.004308800678700209, + 0.0049200644716620445, + 0.005452118813991547, + 0.004608821589499712, + 0.00414946349337697, + 0.004460183437913656, + 0.008227351121604443, + 0.006456063594669104, + 0.00535535579547286, + 0.010924724861979485, + 0.0060514588840305805, + 0.007557572331279516, + 0.010754621587693691, + 0.00785448495298624, + 0.0024856615345925093, + 0.0012406104942783713, + 0.0024036027025431395, + 0.0015670235734432936, + 0.0010531328152865171, + 0.0012761252000927925, + 0.0014811706496402621, + 0.0010638314997777343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.0067563182674348354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898415625095367 + ], + "opposite_logprob_tokens": [ + -2.578984260559082 + ], + "logprob_reward": -0.07898415625095367, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898415625095367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05756982043385506, + 0.04637584090232849, + 0.03257172182202339, + 0.06761462241411209, + 0.4121788740158081, + 0.17007751762866974, + 0.06963589787483215, + 0.033488016575574875, + 0.029254885390400887, + 0.08123281598091125 + ], + "credit_map": [ + 0.00657610734924674, + 0.006346939597278833, + 0.006640222854912281, + 0.009949985891580582, + 0.009249608963727951, + 0.0044169239699840546, + 0.007891129702329636, + 0.006498902570456266, + 0.005060779862105846, + 0.0045592873357236385, + 0.005880328826606274, + 0.010866223834455013, + 0.004364984575659037, + 0.0026492089964449406, + 0.007600498851388693, + 0.005394527688622475, + 0.004270443692803383, + 0.0031286263838410378, + 0.004105670843273401, + 0.004238402470946312, + 0.004796810448169708, + 0.0025357792619615793, + 0.0045345197431743145, + 0.004961471538990736, + 0.01564835198223591, + 0.00633274158462882, + 0.009797883220016956, + 0.010004473850131035, + 0.006532940082252026, + 0.0091255446895957, + 0.0056695411913096905, + 0.00450314162299037, + 0.029543263837695122, + 0.016815122216939926, + 0.15187279880046844, + 0.16299089789390564, + 0.012907137162983418, + 0.011300621554255486, + 0.01833978109061718, + 0.008409279398620129, + 0.024919316172599792, + 0.012718123383820057, + 0.03801186755299568, + 0.028882650658488274, + 0.014857574366033077, + 0.013597923330962658, + 0.01911456510424614, + 0.017975490540266037, + 0.012282912619411945, + 0.004996610805392265, + 0.01632845774292946, + 0.013175638392567635, + 0.0054821656085550785, + 0.004504346754401922, + 0.009076764807105064, + 0.003789003938436508, + 0.006531088147312403, + 0.0046895211562514305, + 0.00391388451680541, + 0.0034738609101623297, + 0.0031518260948359966, + 0.0048499987460672855, + 0.003137633902952075, + 0.0037402028683573008, + 0.0038253897801041603, + 0.003672380466014147, + 0.0029799193143844604, + 0.004246451426297426, + 0.00292569724842906, + 0.0036221682094037533, + 0.004418755415827036, + 0.0035641221329569817, + 0.015859628096222878, + 0.005421600304543972, + 0.023896349593997, + 0.009312647394835949, + 0.005915504414588213, + 0.005638051778078079, + 0.007975154556334019, + 0.007213874254375696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.07898415625095367 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347821578383446 + ], + "opposite_logprob_tokens": [ + -3.6513478755950928 + ], + "logprob_reward": -0.026347821578383446, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347821578383446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07746084779500961, + 0.03624723479151726, + 0.036977000534534454, + 0.020471518859267235, + 0.528128445148468, + 0.11409085243940353, + 0.09545720368623734, + 0.03439764305949211, + 0.03449307382106781, + 0.02227623388171196 + ], + "credit_map": [ + 0.012507468461990356, + 0.006595788523554802, + 0.009663903154432774, + 0.01308070495724678, + 0.010264175944030285, + 0.006872947793453932, + 0.010123668238520622, + 0.008352185599505901, + 0.0032930339220911264, + 0.0038611649069935083, + 0.00433809170499444, + 0.008249481208622456, + 0.0040751174092292786, + 0.001769702648743987, + 0.007094854488968849, + 0.003565784078091383, + 0.008017532527446747, + 0.006715159863233566, + 0.004362083040177822, + 0.004386166576296091, + 0.00372665049508214, + 0.002945889951661229, + 0.0030272877775132656, + 0.0037962293718010187, + 0.004084118641912937, + 0.0018747528083622456, + 0.0025233514606952667, + 0.004236238077282906, + 0.0017125904560089111, + 0.0018586524529382586, + 0.0025147281121462584, + 0.001667084638029337, + 0.040463097393512726, + 0.020768942311406136, + 0.20319855213165283, + 0.20857840776443481, + 0.015551324002444744, + 0.012369353324174881, + 0.018802473321557045, + 0.008396266028285027, + 0.022490244358778, + 0.006142219062894583, + 0.03149562329053879, + 0.019769901409745216, + 0.007060076575726271, + 0.007907065562903881, + 0.009711485356092453, + 0.009514221921563148, + 0.014117513783276081, + 0.009489962831139565, + 0.015247912146151066, + 0.01921422965824604, + 0.007622475735843182, + 0.006825292948633432, + 0.016158603131771088, + 0.006781203672289848, + 0.004343469627201557, + 0.002356936689466238, + 0.0038894549943506718, + 0.004455358721315861, + 0.007594428025186062, + 0.0030680729541927576, + 0.003926556557416916, + 0.004763361532241106, + 0.00596251804381609, + 0.004574315156787634, + 0.0022685453295707703, + 0.004640887025743723, + 0.004203315358608961, + 0.005751012824475765, + 0.004297101870179176, + 0.002795376116409898, + 0.003909194376319647, + 0.0011370948050171137, + 0.006239153444766998, + 0.0030988832004368305, + 0.0014133783988654613, + 0.0013610312016680837, + 0.0027194609865546227, + 0.0023980368860065937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.026347821578383446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two adult Siberian tigers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a vast, snow-covered coniferous forest clearing with towering pine trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the tigers engaged in a fierce struggle with one pushing off its hind legs and the other rearing up with a raised paw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tigers' movements stir up deep snow, creating a whirlwind that blurs their body contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, fixed, medium-long shot taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863002523779869 + ], + "opposite_logprob_tokens": [ + -3.0486299991607666 + ], + "logprob_reward": -0.04863002523779869, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863002523779869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06493272632360458, + 0.04238731414079666, + 0.026783667504787445, + 0.061811503022909164, + 0.4144256114959717, + 0.08609198778867722, + 0.12393505871295929, + 0.0936756581068039, + 0.056579943746328354, + 0.029376571998000145 + ], + "credit_map": [ + 0.007811896502971649, + 0.006568522192537785, + 0.0071957786567509174, + 0.010154726915061474, + 0.008516687899827957, + 0.0038552272599190474, + 0.014825134538114071, + 0.006004754453897476, + 0.004795978777110577, + 0.0035412833094596863, + 0.0053589302115142345, + 0.009884222410619259, + 0.004380649887025356, + 0.0030524577014148235, + 0.007201572880148888, + 0.004172217100858688, + 0.006219233386218548, + 0.002990683075040579, + 0.003308525774627924, + 0.0024912087246775627, + 0.003618073184043169, + 0.0024575558491051197, + 0.003210473572835326, + 0.0024879146367311478, + 0.011883769184350967, + 0.005993615370243788, + 0.009066098369657993, + 0.008506816811859608, + 0.006892967503517866, + 0.008465098217129707, + 0.005702475551515818, + 0.00530066154897213, + 0.042196642607450485, + 0.016653841361403465, + 0.16880901157855988, + 0.1465589851140976, + 0.011095240712165833, + 0.00957755371928215, + 0.013369150459766388, + 0.006165207363665104, + 0.012626443058252335, + 0.006722701713442802, + 0.016193605959415436, + 0.013045971281826496, + 0.011627476662397385, + 0.007306121755391359, + 0.010810865089297295, + 0.007758799009025097, + 0.02460550144314766, + 0.005377627443522215, + 0.045582614839076996, + 0.015491100959479809, + 0.00844500120729208, + 0.004268658347427845, + 0.015577631071209908, + 0.004586918279528618, + 0.018777115270495415, + 0.015484169125556946, + 0.010939219035208225, + 0.009506132453680038, + 0.008472751826047897, + 0.014219878241419792, + 0.007733568083494902, + 0.00854281336069107, + 0.007481633685529232, + 0.006618362385779619, + 0.005422294605523348, + 0.007495387457311153, + 0.0047064367681741714, + 0.01011659111827612, + 0.008662880398333073, + 0.006076352670788765, + 0.006056350190192461, + 0.0019211218459531665, + 0.007218648679554462, + 0.003717681858688593, + 0.0021532264072448015, + 0.0026070899330079556, + 0.002998810028657317, + 0.0027036452665925026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.04863002523779869 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_tensor([220, 220])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_tensor([220, 220])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4d0fdd0743a0e3c6375907fa459acf4de68d0081 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_tensor([220, 220])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Rendered in a bright and playful 3D cartoon style, a static medium close-up shot from a slightly upward angle focuses on a plump, shaggy raccoon, comically wearing a classic black cloth bandit's eye mask. The raccoon lies on a smooth, light-wood kitchen countertop, before an open, large, pale yellow ceramic cookie jar painted with a simple floral pattern. To carry out its \"theft,\" the raccoon leans forward, slightly bending its hind legs in a cautious, stealthy stance. Steadily and slowly, it inserts one of its furry, pink-padded front paws into the wide jar's opening, attempting to reach the tempting chocolate chip cookies stacked within. Beneath the comical eye mask, its round, black eyes gleam with cunning anticipation, its nose twitching slightly, its entire body stiff with intense concentration. The soft morning light pours in through the window outside the picture, providing the entire scene with warm and even lighting. The surfaces of all objects appear smooth and highly saturated, with almost no sharp shadows, perfectly creating an atmosphere full of childlike fun, light-hearted humor, and dramatic tension.", + "seed": 12920, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_220_seed_12920_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010833828675094992, + "vlm_logprob": -0.010833828675094992, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.010833828675094992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012271380983293056 + ], + "opposite_logprob_tokens": [ + -4.637271404266357 + ], + "logprob_reward": -0.012271380983293056, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012271380983293056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23119758069515228, + 0.09538903832435608, + 0.3891003429889679, + 0.029667701572179794, + 0.08229425549507141, + 0.07584445923566818, + 0.02967590093612671, + 0.01953868567943573, + 0.02021646685898304, + 0.027075601741671562 + ], + "credit_map": [ + 0.033679887652397156, + 0.022005220875144005, + 0.025222714990377426, + 0.021105872467160225, + 0.04967641085386276, + 0.026705441996455193, + 0.02136923000216484, + 0.03143279626965523, + 0.01181460078805685, + 0.009728464297950268, + 0.014837435446679592, + 0.013128810562193394, + 0.013032523915171623, + 0.016148395836353302, + 0.00640977593138814, + 0.0102890245616436, + 0.0548504963517189, + 0.03679727762937546, + 0.04405801743268967, + 0.07134392857551575, + 0.034592848271131516, + 0.04449889808893204, + 0.04897717386484146, + 0.05398167669773102, + 0.003013704437762499, + 0.003093740437179804, + 0.0037054489366710186, + 0.005310434382408857, + 0.003658873261883855, + 0.0039512356743216515, + 0.0027692685835063457, + 0.004164994228631258, + 0.009281066246330738, + 0.007605565246194601, + 0.009131602942943573, + 0.009468014352023602, + 0.01251041330397129, + 0.010810839012265205, + 0.007207402493804693, + 0.016279349103569984, + 0.006874600425362587, + 0.0053072706796228886, + 0.008594517596065998, + 0.008437173441052437, + 0.01659279689192772, + 0.010669509880244732, + 0.008702748455107212, + 0.010665835812687874, + 0.0036987534258514643, + 0.0021869365591555834, + 0.0032588124740868807, + 0.005335064604878426, + 0.004651987925171852, + 0.0035187380854040384, + 0.0034212202299386263, + 0.0036043873988091946, + 0.0027712334413081408, + 0.0015587506350129843, + 0.00227736821398139, + 0.0023798062466084957, + 0.0030803403351455927, + 0.0027327178977429867, + 0.001423918642103672, + 0.003314550733193755, + 0.002202240051701665, + 0.001391950878314674, + 0.002107026055455208, + 0.004076766781508923, + 0.0020953132770955563, + 0.0024843604769557714, + 0.0030779759399592876, + 0.00278083304874599, + 0.0025075620505958796, + 0.0018399000400677323, + 0.0028499499894678593, + 0.004140567034482956, + 0.0046406639739871025, + 0.00410084193572402, + 0.002720761112868786, + 0.004275353625416756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.012271380983293056 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201516181230545 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201516181230545, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201516181230545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13207118213176727, + 0.11707442253828049, + 0.35588133335113525, + 0.11121590435504913, + 0.06351135671138763, + 0.030231498181819916, + 0.07653502374887466, + 0.04129616171121597, + 0.029544653370976448, + 0.04263846576213837 + ], + "credit_map": [ + 0.018873021006584167, + 0.015532462857663631, + 0.017953069880604744, + 0.023273607715964317, + 0.014036692678928375, + 0.011980049312114716, + 0.015547697432339191, + 0.014874592423439026, + 0.021575376391410828, + 0.013197981752455235, + 0.015135444700717926, + 0.021220268681645393, + 0.011942065320909023, + 0.015913134440779686, + 0.00925743393599987, + 0.008832723833620548, + 0.045251525938510895, + 0.03004002571105957, + 0.04124748334288597, + 0.08125550299882889, + 0.02682245895266533, + 0.03069497086107731, + 0.04799311235547066, + 0.052576273679733276, + 0.010516083799302578, + 0.010710805654525757, + 0.014311770908534527, + 0.013871186412870884, + 0.015102018602192402, + 0.020581215620040894, + 0.009876828640699387, + 0.016246002167463303, + 0.00889088585972786, + 0.0055330777540802956, + 0.008026916533708572, + 0.015939613804221153, + 0.0051852683536708355, + 0.005040624178946018, + 0.006687848828732967, + 0.00820712000131607, + 0.004521636292338371, + 0.0022494951263070107, + 0.003744877642020583, + 0.006034679710865021, + 0.003574099624529481, + 0.003059237962588668, + 0.002551722340285778, + 0.0044957492500543594, + 0.009661947377026081, + 0.0051276846788823605, + 0.007948985323309898, + 0.015442909672856331, + 0.009843663312494755, + 0.009018823504447937, + 0.007958284579217434, + 0.011532730422914028, + 0.004680066369473934, + 0.0030497414991259575, + 0.004668862093240023, + 0.010092228651046753, + 0.00598890520632267, + 0.003486417233943939, + 0.004839352797716856, + 0.0044905878603458405, + 0.0042264689691364765, + 0.0025265326257795095, + 0.00306316209025681, + 0.005689295008778572, + 0.003985944669693708, + 0.003069146303460002, + 0.002579650143161416, + 0.00440445588901639, + 0.005163892172276974, + 0.0034110022243112326, + 0.0048279715701937675, + 0.01119750365614891, + 0.0032139744143933058, + 0.0036214198917150497, + 0.00528488727286458, + 0.005917817819863558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.002201516181230545 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026345383375883102 + ], + "opposite_logprob_tokens": [ + -3.6513454914093018 + ], + "logprob_reward": -0.026345383375883102, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026345383375883102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21509306132793427, + 0.1460230052471161, + 0.2289482206106186, + 0.0587695948779583, + 0.08596401661634445, + 0.04583057761192322, + 0.06921939551830292, + 0.04088282212615013, + 0.06521650403738022, + 0.04405279457569122 + ], + "credit_map": [ + 0.030814459547400475, + 0.0180156659334898, + 0.023221872746944427, + 0.01979449950158596, + 0.04523632675409317, + 0.023652618750929832, + 0.02476777881383896, + 0.02958984114229679, + 0.022899797186255455, + 0.013037577271461487, + 0.01596941240131855, + 0.01597318798303604, + 0.022535113617777824, + 0.02178098075091839, + 0.013384237885475159, + 0.020442696288228035, + 0.03997482359409332, + 0.018414229154586792, + 0.02716539241373539, + 0.032743290066719055, + 0.018668517470359802, + 0.018488911911845207, + 0.029971497133374214, + 0.043521538376808167, + 0.005871601868420839, + 0.004923503380268812, + 0.00663136038929224, + 0.006196783389896154, + 0.008279871195554733, + 0.011061059311032295, + 0.005900747608393431, + 0.009904668666422367, + 0.007921166718006134, + 0.008154738694429398, + 0.009148193523287773, + 0.013310198672115803, + 0.016060039401054382, + 0.010192393325269222, + 0.010923688299953938, + 0.0102536017075181, + 0.004823985509574413, + 0.0033649869728833437, + 0.004893751349300146, + 0.00786867830902338, + 0.008475448936223984, + 0.005346972495317459, + 0.00544878002256155, + 0.005607972852885723, + 0.006862409878522158, + 0.005034566856920719, + 0.007304571568965912, + 0.008755052462220192, + 0.013280792161822319, + 0.009648781269788742, + 0.007102309260517359, + 0.011230909265577793, + 0.003975070081651211, + 0.00317848427221179, + 0.004739291965961456, + 0.009122353978455067, + 0.0053036450408399105, + 0.004570214543491602, + 0.003952914383262396, + 0.006040849722921848, + 0.007694370578974485, + 0.005335825029760599, + 0.006417582277208567, + 0.007820057682693005, + 0.011818904429674149, + 0.008752783760428429, + 0.0061179897747933865, + 0.011258987709879875, + 0.004761020187288523, + 0.003897117916494608, + 0.004231820348650217, + 0.005720159970223904, + 0.008492950350046158, + 0.00569868041202426, + 0.003518644254654646, + 0.007732399273663759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.026345383375883102 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008132726070471108 + ], + "opposite_logprob_tokens": [ + -7.1258134841918945 + ], + "logprob_reward": -0.0008132726070471108, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008132726070471108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13381162285804749, + 0.10287456214427948, + 0.3721115291118622, + 0.0772993341088295, + 0.07324176281690598, + 0.06106621026992798, + 0.05294939503073692, + 0.04669126868247986, + 0.05424029007554054, + 0.025714069604873657 + ], + "credit_map": [ + 0.015910150483250618, + 0.011168218217790127, + 0.015764113515615463, + 0.011463016271591187, + 0.020718654617667198, + 0.02218804508447647, + 0.015546138398349285, + 0.02105327881872654, + 0.009877250529825687, + 0.009134063497185707, + 0.011246075853705406, + 0.012973298318684101, + 0.0197522584348917, + 0.018895281478762627, + 0.01045946590602398, + 0.010536855086684227, + 0.04870225116610527, + 0.033419959247112274, + 0.04430879279971123, + 0.0866483598947525, + 0.027584677562117577, + 0.029882818460464478, + 0.0510907806456089, + 0.050473857671022415, + 0.0061195227317512035, + 0.00809455942362547, + 0.010569573380053043, + 0.01052370760589838, + 0.008631829172372818, + 0.01721653901040554, + 0.00656702509149909, + 0.009576572105288506, + 0.009644904173910618, + 0.00645224517211318, + 0.009147866629064083, + 0.012171815149486065, + 0.009977268986403942, + 0.008849321864545345, + 0.005839820485562086, + 0.011158517561852932, + 0.00715402839705348, + 0.003297284245491028, + 0.007249964866787195, + 0.009161669760942459, + 0.008048458024859428, + 0.009030941873788834, + 0.007306915707886219, + 0.009816944599151611, + 0.006508909165859222, + 0.0044155665673315525, + 0.00580941466614604, + 0.006595910526812077, + 0.009575306437909603, + 0.0066249980591237545, + 0.00600615656003356, + 0.007413122337311506, + 0.005428873933851719, + 0.0035168009344488382, + 0.005365683697164059, + 0.009837896563112736, + 0.007128341589123011, + 0.0040162065997719765, + 0.005607166327536106, + 0.005790294148027897, + 0.005849790759384632, + 0.0037559892516583204, + 0.0054735345765948296, + 0.010042502544820309, + 0.007943498902022839, + 0.006486057303845882, + 0.006460517179220915, + 0.008228396065533161, + 0.0026165808085352182, + 0.001896768226288259, + 0.002727222628891468, + 0.005992128048092127, + 0.002512461505830288, + 0.0034040166065096855, + 0.0027598701417446136, + 0.00380502175539732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0008132726070471108 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537590228021145 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537590228021145, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537590228021145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2697846293449402, + 0.07617450505495071, + 0.3482455909252167, + 0.02985154651105404, + 0.025965310633182526, + 0.03621261939406395, + 0.061199598014354706, + 0.053731244057416916, + 0.06704245507717133, + 0.03179255127906799 + ], + "credit_map": [ + 0.04979526624083519, + 0.04003355652093887, + 0.032432276755571365, + 0.03193751722574234, + 0.020187098532915115, + 0.028925567865371704, + 0.02922343648970127, + 0.03724990040063858, + 0.007759008090943098, + 0.007569967769086361, + 0.010375824756920338, + 0.011970855295658112, + 0.01178552582859993, + 0.011306008324027061, + 0.0077541847713291645, + 0.007653130684047937, + 0.043793708086013794, + 0.028985492885112762, + 0.041598111391067505, + 0.08033227175474167, + 0.02543075941503048, + 0.024851420894265175, + 0.04808855056762695, + 0.05516524985432625, + 0.006212981883436441, + 0.00321611063554883, + 0.004145202226936817, + 0.003939853981137276, + 0.003076971508562565, + 0.0036793651524931192, + 0.0027001728303730488, + 0.0028808866627514362, + 0.0041511342860758305, + 0.0022229687310755253, + 0.002918716985732317, + 0.003863748162984848, + 0.0036160775925964117, + 0.0031130623538047075, + 0.0024730628356337547, + 0.0036065394524484873, + 0.003416743827983737, + 0.003126193303614855, + 0.003854047507047653, + 0.0061168731190264225, + 0.006349112838506699, + 0.004777607508003712, + 0.0034731943160295486, + 0.005098844412714243, + 0.008274954743683338, + 0.004712332971394062, + 0.006548059172928333, + 0.011014305055141449, + 0.007200493477284908, + 0.006369204726070166, + 0.0071853878907859325, + 0.009894858114421368, + 0.006483481265604496, + 0.003999530337750912, + 0.006478080525994301, + 0.01418759860098362, + 0.004194321576505899, + 0.004211242776364088, + 0.00750565342605114, + 0.006671334616839886, + 0.009936796501278877, + 0.0056879837065935135, + 0.006699319928884506, + 0.011012280359864235, + 0.00993861723691225, + 0.008271127007901669, + 0.0054760402999818325, + 0.01002028863877058, + 0.004428401589393616, + 0.0025910737458616495, + 0.0033205184154212475, + 0.007215474266558886, + 0.004112502094358206, + 0.002844669623300433, + 0.003069843165576458, + 0.004210067447274923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.012537590228021145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a bright and playful 3D cartoon style, a static medium close-up shot from a slightly upward angle focuses on a plump, shaggy raccoon, comically wearing a classic black cloth bandit's eye mask. The raccoon lies on a smooth, light-wood kitchen countertop, before an open, large, pale yellow ceramic cookie jar painted with a simple floral pattern. To carry out its \"theft,\" the raccoon leans forward, slightly bending its hind legs in a cautious, stealthy stance. Steadily and slowly, it inserts one of its furry, pink-padded front paws into the wide jar's opening, attempting to reach the tempting chocolate chip cookies stacked within. Beneath the comical eye mask, its round, black eyes gleam with cunning anticipation, its nose twitching slightly, its entire body stiff with intense concentration. The soft morning light pours in through the window outside the picture, providing the entire scene with warm and even lighting. The surfaces of all objects appear smooth and highly saturated, with almost no sharp shadows, perfectly creating an atmosphere full of childlike fun, light-hearted humor, and dramatic tension.", + "seed": 12921, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_220_seed_12921_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.030965645797550678, + "vlm_logprob": -0.030965645797550678, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.030965645797550678, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10244063287973404 + ], + "opposite_logprob_tokens": [ + -2.352440595626831 + ], + "logprob_reward": -0.10244063287973404, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10244063287973404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2841460406780243, + 0.07081733644008636, + 0.039676740765571594, + 0.11591684818267822, + 0.05104508996009827, + 0.04828711226582527, + 0.06992685794830322, + 0.06854414939880371, + 0.03113865852355957, + 0.2205011546611786 + ], + "credit_map": [ + 0.019608819857239723, + 0.04054766893386841, + 0.042248714715242386, + 0.03382064774632454, + 0.05077748745679855, + 0.026699604466557503, + 0.040548425167798996, + 0.029894674196839333, + 0.005398263689130545, + 0.007108741439878941, + 0.009841632097959518, + 0.013911127112805843, + 0.011780348606407642, + 0.010589180514216423, + 0.005727772135287523, + 0.006460272241383791, + 0.006952337455004454, + 0.0029266083147376776, + 0.004898055922240019, + 0.0029793076682835817, + 0.003602894488722086, + 0.011379665695130825, + 0.002636605640873313, + 0.004301265347748995, + 0.01702207140624523, + 0.010519685223698616, + 0.01234368421137333, + 0.011611472815275192, + 0.011327614076435566, + 0.028336917981505394, + 0.013658291660249233, + 0.011097104288637638, + 0.004318884573876858, + 0.007505204528570175, + 0.006816989742219448, + 0.006483783479779959, + 0.008316337130963802, + 0.005899210926145315, + 0.005221852101385593, + 0.006482826080173254, + 0.0038061910308897495, + 0.004558855202049017, + 0.005741936154663563, + 0.003893005196005106, + 0.009941832162439823, + 0.0048867883160710335, + 0.006111505441367626, + 0.009347000159323215, + 0.005905194208025932, + 0.009442542679607868, + 0.007225606590509415, + 0.005952933803200722, + 0.014761344529688358, + 0.00635141134262085, + 0.011655064299702644, + 0.00863275770097971, + 0.005942388437688351, + 0.006748566869646311, + 0.00856032595038414, + 0.0075384145602583885, + 0.007841477170586586, + 0.006130889058113098, + 0.017607470974326134, + 0.008174610324203968, + 0.002704059239476919, + 0.002701248275116086, + 0.004748785402625799, + 0.0035303307231515646, + 0.003309016115963459, + 0.0024458440020680428, + 0.008223450742661953, + 0.003475923789665103, + 0.021165339276194572, + 0.02475164458155632, + 0.030486606061458588, + 0.020242106169462204, + 0.037040725350379944, + 0.018120093271136284, + 0.035120803862810135, + 0.03357381746172905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.10244063287973404 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070596873760223 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070596873760223, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070596873760223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12651975452899933, + 0.07757632434368134, + 0.20343753695487976, + 0.09013909846544266, + 0.12599408626556396, + 0.07723163813352585, + 0.09158274531364441, + 0.12494807690382004, + 0.022476959973573685, + 0.060093823820352554 + ], + "credit_map": [ + 0.009099435992538929, + 0.015679433941841125, + 0.017794739454984665, + 0.012460066005587578, + 0.020938238129019737, + 0.01487136259675026, + 0.019114481285214424, + 0.01656198501586914, + 0.0087790722027421, + 0.006030360236763954, + 0.010955195873975754, + 0.007386424578726292, + 0.009761263616383076, + 0.01372054684907198, + 0.008812042884528637, + 0.012131411582231522, + 0.048444364219903946, + 0.008145890198647976, + 0.011547724716365337, + 0.007876245304942131, + 0.008452228270471096, + 0.08612017333507538, + 0.006032099016010761, + 0.026818793267011642, + 0.02159029059112072, + 0.0037211626768112183, + 0.0059081679210066795, + 0.004741664510220289, + 0.0039755674079060555, + 0.03758791834115982, + 0.003712913254275918, + 0.00890140701085329, + 0.011951223015785217, + 0.02002628892660141, + 0.017193332314491272, + 0.01552477478981018, + 0.020208729431033134, + 0.015804316848516464, + 0.01312188059091568, + 0.01216353103518486, + 0.006105219945311546, + 0.007548521272838116, + 0.011575847864151001, + 0.007517985068261623, + 0.010357216000556946, + 0.007817677222192287, + 0.010310642421245575, + 0.01599852368235588, + 0.009084463119506836, + 0.009064173325896263, + 0.017632532864809036, + 0.01071253139525652, + 0.013229521922767162, + 0.007970153354108334, + 0.015183043666183949, + 0.008706311695277691, + 0.007454992737621069, + 0.007514868397265673, + 0.015600611455738544, + 0.010037414729595184, + 0.009575900621712208, + 0.013341536745429039, + 0.03762239217758179, + 0.02380034141242504, + 0.002054407261312008, + 0.002501202980056405, + 0.0035567698068916798, + 0.002322535729035735, + 0.002313465578481555, + 0.002095594769343734, + 0.0037425789050757885, + 0.0038904023822396994, + 0.005448981653898954, + 0.0060703628696501255, + 0.006469938438385725, + 0.005203280597925186, + 0.008934295736253262, + 0.005181015934795141, + 0.012107624672353268, + 0.010678316466510296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.016070596873760223 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058861404657364 + ], + "opposite_logprob_tokens": [ + -3.8955886363983154 + ], + "logprob_reward": -0.02058861404657364, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058861404657364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07122321426868439, + 0.07699310779571533, + 0.2308831512928009, + 0.0510760173201561, + 0.035775959491729736, + 0.062100350856781006, + 0.15153205394744873, + 0.1372191160917282, + 0.11258066445589066, + 0.07061641663312912 + ], + "credit_map": [ + 0.0059703653678298, + 0.010247823782265186, + 0.00901330728083849, + 0.007598233874887228, + 0.009148924611508846, + 0.007646921556442976, + 0.008677653037011623, + 0.012919985689222813, + 0.004802834242582321, + 0.011957664042711258, + 0.008358676917850971, + 0.011655286885797977, + 0.013060429133474827, + 0.0088422866538167, + 0.008058924227952957, + 0.010257009416818619, + 0.055790286511182785, + 0.008738977834582329, + 0.01569167710840702, + 0.009597517549991608, + 0.010057282634079456, + 0.09666649252176285, + 0.006566775497049093, + 0.0277741439640522, + 0.005978170782327652, + 0.006787738297134638, + 0.007122171577066183, + 0.006295417435467243, + 0.008205804042518139, + 0.005481095984578133, + 0.0065400805324316025, + 0.004665539134293795, + 0.003129717893898487, + 0.0051611848175525665, + 0.0051251547411084175, + 0.005513363983482122, + 0.004771709442138672, + 0.0038528554141521454, + 0.003807866247370839, + 0.004414106719195843, + 0.004407881759107113, + 0.007214983459562063, + 0.006075181066989899, + 0.005641214549541473, + 0.010154563933610916, + 0.005515577271580696, + 0.012646242044866085, + 0.010444707237184048, + 0.012388695031404495, + 0.016271259635686874, + 0.01734166219830513, + 0.01684451289474964, + 0.018040377646684647, + 0.013707058504223824, + 0.038551390171051025, + 0.018387096002697945, + 0.01219965424388647, + 0.017329556867480278, + 0.02130468375980854, + 0.019261211156845093, + 0.023395031690597534, + 0.012463536113500595, + 0.01626676321029663, + 0.01499867346137762, + 0.010302026756107807, + 0.012604853138327599, + 0.020868351683020592, + 0.015507079660892487, + 0.014419341459870338, + 0.010742285288870335, + 0.015005926601588726, + 0.013130789622664452, + 0.006111613474786282, + 0.008123856037855148, + 0.008777378126978874, + 0.007447682321071625, + 0.012066937051713467, + 0.005798812489956617, + 0.012169980444014072, + 0.010120157152414322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.02058861404657364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191736526787281 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003191736526787281, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191736526787281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18176361918449402, + 0.08473864197731018, + 0.18523657321929932, + 0.06481868028640747, + 0.056056417524814606, + 0.1235937774181366, + 0.06905344873666763, + 0.11905650794506073, + 0.03266110643744469, + 0.08302123844623566 + ], + "credit_map": [ + 0.014519020915031433, + 0.035321254283189774, + 0.022174948826432228, + 0.01983800157904625, + 0.03889121860265732, + 0.017691679298877716, + 0.018558358773589134, + 0.01476913783699274, + 0.005613821558654308, + 0.013601141050457954, + 0.009203873574733734, + 0.013924325816333294, + 0.014238298870623112, + 0.009461715817451477, + 0.009222431108355522, + 0.00947303231805563, + 0.041921984404325485, + 0.006243501789867878, + 0.011173252016305923, + 0.00785939022898674, + 0.008504500612616539, + 0.07725460082292557, + 0.007863330654799938, + 0.024416012689471245, + 0.012136546894907951, + 0.0037110503762960434, + 0.005988661199808121, + 0.006495888810604811, + 0.003091351827606559, + 0.023424923419952393, + 0.0036312539596110582, + 0.006339003331959248, + 0.005759925115853548, + 0.00791823212057352, + 0.007326997350901365, + 0.00631271256133914, + 0.004386098124086857, + 0.012957866303622723, + 0.004997937940061092, + 0.0063966489396989346, + 0.006143950391560793, + 0.012364287860691547, + 0.0122087886556983, + 0.011429443955421448, + 0.017824575304985046, + 0.011769934557378292, + 0.027011966332793236, + 0.02484082616865635, + 0.0066297962330281734, + 0.009310588240623474, + 0.010642783716320992, + 0.00829988531768322, + 0.01097294595092535, + 0.00611632876098156, + 0.009302590973675251, + 0.00777853187173605, + 0.010661854408681393, + 0.014723064377903938, + 0.016101893037557602, + 0.012584124691784382, + 0.014848587103188038, + 0.01250812690705061, + 0.0187250804156065, + 0.01890377700328827, + 0.0028049442917108536, + 0.003472143318504095, + 0.005560973193496466, + 0.003394327824935317, + 0.003922307398170233, + 0.003260266501456499, + 0.005650099832564592, + 0.004596045706421137, + 0.007588694337755442, + 0.008831452578306198, + 0.010703031904995441, + 0.007436821237206459, + 0.012962618842720985, + 0.006316022481769323, + 0.01714620739221573, + 0.012036388739943504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.003191736526787281 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536648660898209 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536648660898209, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536648660898209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06669388711452484, + 0.09733448177576065, + 0.09847525507211685, + 0.1255914717912674, + 0.07701879739761353, + 0.1368759423494339, + 0.2304849624633789, + 0.07778449356555939, + 0.061866067349910736, + 0.027874676510691643 + ], + "credit_map": [ + 0.00552022410556674, + 0.01020103320479393, + 0.010340191423892975, + 0.006612514611333609, + 0.00977378711104393, + 0.008435314521193504, + 0.008917780593037605, + 0.006893044337630272, + 0.009798871353268623, + 0.013793679885566235, + 0.013915547169744968, + 0.0103353476151824, + 0.017917528748512268, + 0.00806525256484747, + 0.012529062107205391, + 0.01097919326275587, + 0.014236382208764553, + 0.012571858242154121, + 0.008623297326266766, + 0.01041476707905531, + 0.014209311455488205, + 0.022531529888510704, + 0.006723378319293261, + 0.00916473101824522, + 0.021475588902831078, + 0.013352951034903526, + 0.011934231966733932, + 0.011012514121830463, + 0.014224709011614323, + 0.037775855511426926, + 0.006282752379775047, + 0.009532874450087547, + 0.007099453825503588, + 0.007549499627202749, + 0.01187397725880146, + 0.010394941084086895, + 0.009982498362660408, + 0.011086968705058098, + 0.010844513773918152, + 0.008186941966414452, + 0.009844658896327019, + 0.01682843081653118, + 0.010252083651721478, + 0.01807963289320469, + 0.03740429878234863, + 0.010888616554439068, + 0.015148022212088108, + 0.018430190160870552, + 0.02623264119029045, + 0.029993100091814995, + 0.03472292423248291, + 0.029804013669490814, + 0.032444097101688385, + 0.022617990151047707, + 0.026205692440271378, + 0.028464507311582565, + 0.007175873499363661, + 0.007128148805350065, + 0.009264596737921238, + 0.008600990287959576, + 0.007965929806232452, + 0.008455189876258373, + 0.017231738194823265, + 0.011962025426328182, + 0.0037038882728666067, + 0.004110286477953196, + 0.007835828699171543, + 0.004610815085470676, + 0.005660719238221645, + 0.006036174483597279, + 0.02188616804778576, + 0.008022187277674675, + 0.003074601525440812, + 0.002787922974675894, + 0.005717465654015541, + 0.0034706161823123693, + 0.0026974377688020468, + 0.0025672113988548517, + 0.00392047269269824, + 0.003638948779553175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.012536648660898209 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a bright and playful 3D cartoon style, a static medium close-up shot from a slightly upward angle focuses on a plump, shaggy raccoon, comically wearing a classic black cloth bandit's eye mask. The raccoon lies on a smooth, light-wood kitchen countertop, before an open, large, pale yellow ceramic cookie jar painted with a simple floral pattern. To carry out its \"theft,\" the raccoon leans forward, slightly bending its hind legs in a cautious, stealthy stance. Steadily and slowly, it inserts one of its furry, pink-padded front paws into the wide jar's opening, attempting to reach the tempting chocolate chip cookies stacked within. Beneath the comical eye mask, its round, black eyes gleam with cunning anticipation, its nose twitching slightly, its entire body stiff with intense concentration. The soft morning light pours in through the window outside the picture, providing the entire scene with warm and even lighting. The surfaces of all objects appear smooth and highly saturated, with almost no sharp shadows, perfectly creating an atmosphere full of childlike fun, light-hearted humor, and dramatic tension.", + "seed": 12922, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_220_seed_12922_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02043082471936941, + "vlm_logprob": -0.02043082471936941, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02043082471936941, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01739601232111454 + ], + "opposite_logprob_tokens": [ + -4.267395973205566 + ], + "logprob_reward": -0.01739601232111454, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01739601232111454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08188336342573166, + 0.03250908851623535, + 0.02164541743695736, + 0.03698034584522247, + 0.49594855308532715, + 0.027647634968161583, + 0.07045737653970718, + 0.06220816448330879, + 0.15111136436462402, + 0.01960870251059532 + ], + "credit_map": [ + 0.00838080607354641, + 0.018021436408162117, + 0.005523056257516146, + 0.0074744331650435925, + 0.0119093656539917, + 0.014077889733016491, + 0.007797138765454292, + 0.008699238300323486, + 0.0031153166200965643, + 0.00485257338732481, + 0.0029378696344792843, + 0.004305321257561445, + 0.005533723626285791, + 0.005550767295062542, + 0.003466220572590828, + 0.0027472958900034428, + 0.0021341892424970865, + 0.0024945593904703856, + 0.001628568395972252, + 0.0014651349047198892, + 0.002841877518221736, + 0.002915427554398775, + 0.004683358129113913, + 0.0034823031164705753, + 0.004194166511297226, + 0.004468868020921946, + 0.003922574687749147, + 0.004381358157843351, + 0.0037006952334195375, + 0.0061858342960476875, + 0.004629725590348244, + 0.005497124046087265, + 0.023874515667557716, + 0.011855262331664562, + 0.0066850678995251656, + 0.007457773666828871, + 0.016440030187368393, + 0.22053410112857819, + 0.029444802552461624, + 0.1796569973230362, + 0.0032259237486869097, + 0.004015789832919836, + 0.0019513653824105859, + 0.0023920510429888964, + 0.0035934506449848413, + 0.004723475314676762, + 0.0033908109180629253, + 0.004354770295321941, + 0.012301837094128132, + 0.009070776402950287, + 0.010394318960607052, + 0.0063399565406143665, + 0.008428932167589664, + 0.008626685477793217, + 0.007864011451601982, + 0.007430861704051495, + 0.007458142936229706, + 0.0076974788680672646, + 0.0059928204864263535, + 0.006656826939433813, + 0.006905269809067249, + 0.008471796289086342, + 0.008809962309896946, + 0.01021586824208498, + 0.01220452506095171, + 0.01985359936952591, + 0.011687343940138817, + 0.02233036234974861, + 0.0354984812438488, + 0.021574465557932854, + 0.015068457461893559, + 0.012894111685454845, + 0.0024568631779402494, + 0.0017889387672767043, + 0.002062377752736211, + 0.0015145192155614495, + 0.004141025710850954, + 0.0031600266229361296, + 0.0017160156276077032, + 0.0027689344715327024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.01739601232111454 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040996563620865345 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.0040996563620865345, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040996563620865345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11988487094640732, + 0.06245678663253784, + 0.05752190202474594, + 0.06250113248825073, + 0.249225452542305, + 0.09910259395837784, + 0.053431686013936996, + 0.08496503531932831, + 0.048672452569007874, + 0.16223806142807007 + ], + "credit_map": [ + 0.015466347336769104, + 0.010191801004111767, + 0.006812501233071089, + 0.008352052420377731, + 0.018101368099451065, + 0.022906312718987465, + 0.019409948959946632, + 0.018644535914063454, + 0.004195514600723982, + 0.01076896209269762, + 0.0049071055836975574, + 0.007395164109766483, + 0.009537813253700733, + 0.006489180028438568, + 0.009070503525435925, + 0.010092544369399548, + 0.004225024022161961, + 0.004686365835368633, + 0.004426836501806974, + 0.003790389047935605, + 0.0052388813346624374, + 0.009556399658322334, + 0.010630317032337189, + 0.014967693947255611, + 0.008150916546583176, + 0.005479114595800638, + 0.007132875733077526, + 0.0038480344228446484, + 0.007535228971391916, + 0.007368093356490135, + 0.014098587445914745, + 0.008888274431228638, + 0.02160484530031681, + 0.011617152020335197, + 0.009073532186448574, + 0.008096045814454556, + 0.01766384020447731, + 0.0834120512008667, + 0.02565792016685009, + 0.07210007309913635, + 0.008310715667903423, + 0.0111911091953516, + 0.0065903440117836, + 0.00883575901389122, + 0.012245997786521912, + 0.020696841180324554, + 0.012343390844762325, + 0.01888844184577465, + 0.007410078309476376, + 0.006535322871059179, + 0.008033132180571556, + 0.003067459212616086, + 0.004762235097587109, + 0.009066295810043812, + 0.00615012226626277, + 0.008407040499150753, + 0.0067131975665688515, + 0.005409992765635252, + 0.005032591987401247, + 0.005044525023549795, + 0.012179099023342133, + 0.014878949150443077, + 0.016254747286438942, + 0.019451934844255447, + 0.003639302682131529, + 0.005343767814338207, + 0.0033807323779910803, + 0.005188400857150555, + 0.009053651243448257, + 0.007529395166784525, + 0.004783006850630045, + 0.009754194878041744, + 0.012273591011762619, + 0.009514999575912952, + 0.010628603398799896, + 0.008974597789347172, + 0.024032825604081154, + 0.03150037303566933, + 0.021520178765058517, + 0.04379291087388992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.0040996563620865345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0298079252243042 + ], + "opposite_logprob_tokens": [ + -3.5298080444335938 + ], + "logprob_reward": -0.0298079252243042, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298079252243042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10904170572757721, + 0.09619641304016113, + 0.07066751271486282, + 0.06878361105918884, + 0.3061642646789551, + 0.10112006217241287, + 0.04093604162335396, + 0.0960303470492363, + 0.05987134203314781, + 0.05118868872523308 + ], + "credit_map": [ + 0.010156642645597458, + 0.012647506780922413, + 0.00729939341545105, + 0.008396284654736519, + 0.01827019825577736, + 0.014762128703296185, + 0.020569901913404465, + 0.01693965122103691, + 0.006934082601219416, + 0.014245431870222092, + 0.008166514337062836, + 0.0108070382848382, + 0.01276872307062149, + 0.010318776592612267, + 0.01717464253306389, + 0.01578119583427906, + 0.007160032168030739, + 0.004987654276192188, + 0.008017457090318203, + 0.006427925545722246, + 0.004738609306514263, + 0.0065377503633499146, + 0.017493242397904396, + 0.01530484389513731, + 0.0061402893625199795, + 0.006610812619328499, + 0.004356880206614733, + 0.004452671390026808, + 0.007864238694310188, + 0.01507281418889761, + 0.010223476216197014, + 0.014062430709600449, + 0.01745602674782276, + 0.009619307704269886, + 0.00716373510658741, + 0.007143205497413874, + 0.010516515001654625, + 0.13050678372383118, + 0.022168606519699097, + 0.101590096950531, + 0.009393623098731041, + 0.013656221330165863, + 0.00834487285465002, + 0.007671348284929991, + 0.021365921944379807, + 0.012015780434012413, + 0.013326862826943398, + 0.015345425345003605, + 0.005451148375868797, + 0.005049359053373337, + 0.005770575255155563, + 0.0035965293645858765, + 0.005303090903908014, + 0.005833217408508062, + 0.0047768414951860905, + 0.005155281163752079, + 0.011655035428702831, + 0.013669947162270546, + 0.006806381046772003, + 0.00862444844096899, + 0.021080777049064636, + 0.01606466807425022, + 0.007948268204927444, + 0.010180817916989326, + 0.006003681104630232, + 0.007507652975618839, + 0.005614951718598604, + 0.0068681323900818825, + 0.007409590762108564, + 0.008716989308595657, + 0.00659103924408555, + 0.011159298941493034, + 0.005320270545780659, + 0.006521984469145536, + 0.005454526282846928, + 0.005102708004415035, + 0.00794793851673603, + 0.007372820749878883, + 0.005785288754850626, + 0.007683149538934231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0298079252243042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021996130235493183 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021996130235493183, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021996130235493183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05238524451851845, + 0.05664545297622681, + 0.06510861217975616, + 0.026761146262288094, + 0.6202385425567627, + 0.09377753734588623, + 0.04146496206521988, + 0.01599743962287903, + 0.014527871273458004, + 0.013093153014779091 + ], + "credit_map": [ + 0.007101630326360464, + 0.0069193653762340546, + 0.005160638131201267, + 0.005198417231440544, + 0.00798907782882452, + 0.006871945224702358, + 0.007260403595864773, + 0.00588376447558403, + 0.00558821763843298, + 0.00830432865768671, + 0.0068962606601417065, + 0.006489255931228399, + 0.008768253959715366, + 0.009203756228089333, + 0.006280460860580206, + 0.0051149227656424046, + 0.005690676160156727, + 0.007291166111826897, + 0.007029354106634855, + 0.007714993320405483, + 0.008486997336149216, + 0.009814620949327946, + 0.010085794143378735, + 0.008995010517537594, + 0.0023778360337018967, + 0.002528670011088252, + 0.002515228232368827, + 0.003173679346218705, + 0.0028313633520156145, + 0.004136589355766773, + 0.004173098597675562, + 0.005024681333452463, + 0.03207005187869072, + 0.01592816412448883, + 0.009890194982290268, + 0.009608971886336803, + 0.024151096120476723, + 0.2687416672706604, + 0.036801498383283615, + 0.22304686903953552, + 0.007777109742164612, + 0.010913161560893059, + 0.006135573610663414, + 0.007375268265604973, + 0.009776754304766655, + 0.023703917860984802, + 0.01012105867266655, + 0.017974691465497017, + 0.008517022244632244, + 0.0050394474528729916, + 0.006975040305405855, + 0.0033007669262588024, + 0.004294134210795164, + 0.005888096056878567, + 0.003344941884279251, + 0.0041055153124034405, + 0.0016531411092728376, + 0.0020096180960536003, + 0.001434819190762937, + 0.0013322249287739396, + 0.0018020188435912132, + 0.002208030316978693, + 0.0025660654064267874, + 0.0029915212653577328, + 0.00185483752284199, + 0.0017809993587434292, + 0.0011996888788416982, + 0.001344143645837903, + 0.0026249869260936975, + 0.0020780209451913834, + 0.001661383779719472, + 0.0019838104490190744, + 0.0012415720848366618, + 0.0009313062764704227, + 0.001218105899170041, + 0.0009725886629894376, + 0.0018668592674657702, + 0.0023608740884810686, + 0.0013949660351499915, + 0.0031068818643689156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.0021996130235493183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048650916665792465 + ], + "opposite_logprob_tokens": [ + -3.0486509799957275 + ], + "logprob_reward": -0.048650916665792465, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048650916665792465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03437026962637901, + 0.042006734758615494, + 0.04466195032000542, + 0.026118017733097076, + 0.7267999053001404, + 0.09381023794412613, + 0.01491751242429018, + 0.007927202619612217, + 0.004730698186904192, + 0.0046574706211686134 + ], + "credit_map": [ + 0.007746262475848198, + 0.0037319681141525507, + 0.0029815449379384518, + 0.0020912799518555403, + 0.0038304615300148726, + 0.005124923773109913, + 0.0048373183235526085, + 0.004026513081043959, + 0.003984570037573576, + 0.004613985773175955, + 0.0031715696677565575, + 0.003216247307136655, + 0.005788907408714294, + 0.005501833278685808, + 0.007603006903082132, + 0.008126620203256607, + 0.0035256047267466784, + 0.00435987813398242, + 0.003728811861947179, + 0.004041383042931557, + 0.006177108269184828, + 0.0068707033060491085, + 0.00961383618414402, + 0.00634462246671319, + 0.004029593896120787, + 0.002775590168312192, + 0.0023330410476773977, + 0.0023030289448797703, + 0.0026485994458198547, + 0.0032013205345720053, + 0.00558040477335453, + 0.003246439853683114, + 0.03669939190149307, + 0.018842555582523346, + 0.010937195271253586, + 0.010873242281377316, + 0.027297712862491608, + 0.3174452483654022, + 0.04316359758377075, + 0.261540949344635, + 0.007806222885847092, + 0.010947021655738354, + 0.006166150327771902, + 0.007598796859383583, + 0.009502040222287178, + 0.023556191474199295, + 0.010203571058809757, + 0.018030239269137383, + 0.0027263606898486614, + 0.002102210186421871, + 0.001865487894974649, + 0.0013692566426470876, + 0.0013793789548799396, + 0.002149995882064104, + 0.0015258415369316936, + 0.0017989807529374957, + 0.0009205469978041947, + 0.001289397245272994, + 0.0005844149854965508, + 0.0007684367592446506, + 0.001849402324296534, + 0.0012055507395416498, + 0.0006001171423122287, + 0.0007093364838510752, + 0.0004278086998965591, + 0.0005709746037609875, + 0.0004595111240632832, + 0.0006880373111926019, + 0.0007373487460426986, + 0.0006638040649704635, + 0.0004609893658198416, + 0.0007222241256386042, + 0.00039250653935596347, + 0.0004561245150398463, + 0.00038004142697900534, + 0.0003820404235739261, + 0.0011901287361979485, + 0.0007681585266254842, + 0.0004533860192168504, + 0.0006350845214910805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.048650916665792465 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a bright and playful 3D cartoon style, a static medium close-up shot from a slightly upward angle focuses on a plump, shaggy raccoon, comically wearing a classic black cloth bandit's eye mask. The raccoon lies on a smooth, light-wood kitchen countertop, before an open, large, pale yellow ceramic cookie jar painted with a simple floral pattern. To carry out its \"theft,\" the raccoon leans forward, slightly bending its hind legs in a cautious, stealthy stance. Steadily and slowly, it inserts one of its furry, pink-padded front paws into the wide jar's opening, attempting to reach the tempting chocolate chip cookies stacked within. Beneath the comical eye mask, its round, black eyes gleam with cunning anticipation, its nose twitching slightly, its entire body stiff with intense concentration. The soft morning light pours in through the window outside the picture, providing the entire scene with warm and even lighting. The surfaces of all objects appear smooth and highly saturated, with almost no sharp shadows, perfectly creating an atmosphere full of childlike fun, light-hearted humor, and dramatic tension.", + "seed": 12923, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_7/index_220_seed_12923_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.017115185409784316, + "vlm_logprob": -0.017115185409784316, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.017115185409784316, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014692206867039204 + ], + "opposite_logprob_tokens": [ + -4.514692306518555 + ], + "logprob_reward": -0.014692206867039204, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014692206867039204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18397270143032074, + 0.08900371193885803, + 0.03750317916274071, + 0.0717361718416214, + 0.38886791467666626, + 0.0483325719833374, + 0.0436665378510952, + 0.06090690195560455, + 0.0363696850836277, + 0.03964070975780487 + ], + "credit_map": [ + 0.021266575902700424, + 0.023502470925450325, + 0.022063560783863068, + 0.02187427692115307, + 0.017936643213033676, + 0.017811905592679977, + 0.034669581800699234, + 0.02484765835106373, + 0.009046007879078388, + 0.013798020780086517, + 0.013643301092088223, + 0.012951922602951527, + 0.008197649382054806, + 0.009599137119948864, + 0.012992871925234795, + 0.008774794638156891, + 0.00771497143432498, + 0.004375914111733437, + 0.004933447111397982, + 0.0033549147192388773, + 0.003668571123853326, + 0.0036348707508295774, + 0.005481415428221226, + 0.004339068196713924, + 0.008019613102078438, + 0.00895869079977274, + 0.015654219314455986, + 0.012286344543099403, + 0.007456309162080288, + 0.0047965943813323975, + 0.007921562530100346, + 0.006642827298492193, + 0.06480755656957626, + 0.06675228476524353, + 0.04260297864675522, + 0.03990659862756729, + 0.0465938001871109, + 0.031118087470531464, + 0.0593966469168663, + 0.03768991306424141, + 0.008903997018933296, + 0.007540533784776926, + 0.0035949377343058586, + 0.004245164804160595, + 0.006024692673236132, + 0.003978289198130369, + 0.008562173694372177, + 0.005482778884470463, + 0.00843022856861353, + 0.0061276135966181755, + 0.004232244100421667, + 0.00490538077428937, + 0.005840648896992207, + 0.003839861834421754, + 0.005928055848926306, + 0.004362496081739664, + 0.007127047516405582, + 0.010793536901473999, + 0.0055128601379692554, + 0.007976323366165161, + 0.008067247457802296, + 0.005191939417272806, + 0.009382343851029873, + 0.006855600979179144, + 0.004458904266357422, + 0.006560403387993574, + 0.0037562670186161995, + 0.003917895257472992, + 0.004784236662089825, + 0.002955249510705471, + 0.005166640039533377, + 0.004770082887262106, + 0.0044891079887747765, + 0.0049680559895932674, + 0.005351145751774311, + 0.006654100492596626, + 0.0051160226576030254, + 0.0027041444554924965, + 0.006569515913724899, + 0.0037886116188019514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.014692206867039204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642531931400299 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642531931400299, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642531931400299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15552937984466553, + 0.04173087328672409, + 0.05117290839552879, + 0.05889727547764778, + 0.15814976394176483, + 0.23092114925384521, + 0.09383557736873627, + 0.08122613281011581, + 0.027343301102519035, + 0.10119359940290451 + ], + "credit_map": [ + 0.023779448121786118, + 0.014780465513467789, + 0.013243714347481728, + 0.012288250960409641, + 0.02357964962720871, + 0.022905366495251656, + 0.027357595041394234, + 0.017594901844859123, + 0.005722006782889366, + 0.004775788635015488, + 0.00446337228640914, + 0.004573578480631113, + 0.0045881750993430614, + 0.004886261187493801, + 0.006126767490059137, + 0.006594923324882984, + 0.007880797609686852, + 0.007709227968007326, + 0.010487671010196209, + 0.004697143100202084, + 0.0037160785868763924, + 0.0048237452283501625, + 0.0054422905668616295, + 0.006415958516299725, + 0.0056607709266245365, + 0.010022422298789024, + 0.010680010542273521, + 0.009375148452818394, + 0.005880103912204504, + 0.003803064813837409, + 0.007121537812054157, + 0.006354221608489752, + 0.030743256211280823, + 0.027926919981837273, + 0.016243668273091316, + 0.015557462349534035, + 0.013781668618321419, + 0.013123748824000359, + 0.025112850591540337, + 0.015660196542739868, + 0.05545014515519142, + 0.021969329565763474, + 0.020207585766911507, + 0.023369502276182175, + 0.019547520205378532, + 0.02199726551771164, + 0.04130304977297783, + 0.027076754719018936, + 0.015246031805872917, + 0.01171909086406231, + 0.00938641931861639, + 0.014286934398114681, + 0.010878171771764755, + 0.007612515706568956, + 0.014262175187468529, + 0.010444249957799911, + 0.015505225397646427, + 0.0114093367010355, + 0.009125718846917152, + 0.009069498628377914, + 0.006973219569772482, + 0.007512247655540705, + 0.012867356650531292, + 0.008763530291616917, + 0.0038935220800340176, + 0.004243608098477125, + 0.002990950597450137, + 0.0036425809375941753, + 0.00312244170345366, + 0.00228114053606987, + 0.0034764751326292753, + 0.0036925848107784986, + 0.009745697490870953, + 0.007916267029941082, + 0.011616150848567486, + 0.012393693439662457, + 0.00675977161154151, + 0.007364239543676376, + 0.02770256996154785, + 0.01769520901143551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.008642531931400299 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337010957300663 + ], + "opposite_logprob_tokens": [ + -3.408701181411743 + ], + "logprob_reward": -0.0337010957300663, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337010957300663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09325335919857025, + 0.06989736109972, + 0.044154610484838486, + 0.1521701216697693, + 0.227369487285614, + 0.07899811863899231, + 0.10102519392967224, + 0.10859764367341995, + 0.06840631365776062, + 0.05612775683403015 + ], + "credit_map": [ + 0.013149159029126167, + 0.014011825434863567, + 0.013014189898967743, + 0.009897404350340366, + 0.009109417907893658, + 0.008732369169592857, + 0.014442411251366138, + 0.010896584019064903, + 0.0066621205769479275, + 0.011542066931724548, + 0.010695445351302624, + 0.00841373112052679, + 0.006403149105608463, + 0.008618075400590897, + 0.010028218850493431, + 0.0075345514342188835, + 0.00607508746907115, + 0.005650077015161514, + 0.006390773691236973, + 0.00467694690451026, + 0.004847519565373659, + 0.004299030173569918, + 0.008117623627185822, + 0.004097553435713053, + 0.016607141122221947, + 0.021381044760346413, + 0.023512836545705795, + 0.02893752232193947, + 0.018295200541615486, + 0.012563945725560188, + 0.016471998766064644, + 0.01440043281763792, + 0.040476009249687195, + 0.041933573782444, + 0.021964702755212784, + 0.02453133836388588, + 0.02730133756995201, + 0.018538305535912514, + 0.030670227482914925, + 0.021953990682959557, + 0.015442419797182083, + 0.009706386364996433, + 0.009232152253389359, + 0.00974125787615776, + 0.006490486674010754, + 0.007093019783496857, + 0.012900869362056255, + 0.008391525596380234, + 0.018733518198132515, + 0.014925641939043999, + 0.00968595314770937, + 0.01196268294006586, + 0.01068313792347908, + 0.00973537378013134, + 0.015330003574490547, + 0.009968886151909828, + 0.019886502996087074, + 0.01686798594892025, + 0.011655965819954872, + 0.011122455820441246, + 0.011334369890391827, + 0.010211898945271969, + 0.016862446442246437, + 0.010656014084815979, + 0.009481146931648254, + 0.011045138351619244, + 0.0074961804784834385, + 0.009376166388392448, + 0.008655100129544735, + 0.00499517610296607, + 0.010743631981313229, + 0.006613774225115776, + 0.007738971617072821, + 0.008125881664454937, + 0.006049171555787325, + 0.00560183497145772, + 0.007437191437929869, + 0.0041123502887785435, + 0.010083399713039398, + 0.006978955585509539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.0337010957300663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198542468249798 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198542468249798, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198542468249798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10803056508302689, + 0.050231870263814926, + 0.049873679876327515, + 0.038055576384067535, + 0.11234167218208313, + 0.3466995656490326, + 0.12135893106460571, + 0.1011543869972229, + 0.03522811457514763, + 0.03702567145228386 + ], + "credit_map": [ + 0.0148845249786973, + 0.012483276426792145, + 0.01172987837344408, + 0.00907851941883564, + 0.014135426841676235, + 0.010354672558605671, + 0.01945394277572632, + 0.015910319983959198, + 0.006227650213986635, + 0.0067664263769984245, + 0.005987249780446291, + 0.004296387545764446, + 0.005899737123399973, + 0.0057180277071893215, + 0.010235336609184742, + 0.0051010483875870705, + 0.008845429867506027, + 0.006934721488505602, + 0.008259869180619717, + 0.006043247878551483, + 0.004827680066227913, + 0.004780546296387911, + 0.0063452403992414474, + 0.0038369381800293922, + 0.0034250577446073294, + 0.005732009653002024, + 0.0032397909089922905, + 0.004854551982134581, + 0.007186275906860828, + 0.0034456963185220957, + 0.005407444667071104, + 0.004764743614941835, + 0.016938887536525726, + 0.0188257098197937, + 0.0114132696762681, + 0.01311615388840437, + 0.015269797295331955, + 0.010054628364741802, + 0.017370767891407013, + 0.009352453984320164, + 0.0782819539308548, + 0.043841615319252014, + 0.027553075924515724, + 0.03597493842244148, + 0.0339067243039608, + 0.032259292900562286, + 0.0580756850540638, + 0.03680625185370445, + 0.019953548908233643, + 0.017516644671559334, + 0.012912796810269356, + 0.010351544246077538, + 0.015414929017424583, + 0.010331935249269009, + 0.019578522071242332, + 0.015298987738788128, + 0.019234104081988335, + 0.01935918815433979, + 0.01105488184839487, + 0.012051491066813469, + 0.012830456718802452, + 0.008889622054994106, + 0.00978829525411129, + 0.007946333847939968, + 0.004233585670590401, + 0.0047589195892214775, + 0.005922364071011543, + 0.0034749202895909548, + 0.00437607616186142, + 0.0028590690344572067, + 0.005964050069451332, + 0.0036391227040439844, + 0.004791787825524807, + 0.004038240760564804, + 0.0060216570273041725, + 0.007209471892565489, + 0.003252730006352067, + 0.0026345010846853256, + 0.005679191090166569, + 0.0033980889711529016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.002198542468249798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, shaggy raccoon wearing a black cloth bandit's eye mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the raccoon lying on a smooth, light-wood kitchen countertop next to a pale yellow ceramic cookie jar with a floral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the raccoon lean forward with bent hind legs and insert a pink-padded front paw into the cookie jar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a bright, playful 3D cartoon style with smooth, highly saturated surfaces and almost no sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium close-up from a slightly upward angle with soft morning light coming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341550052165985 + ], + "opposite_logprob_tokens": [ + -3.651341438293457 + ], + "logprob_reward": -0.026341550052165985, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341550052165985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.076316699385643, + 0.028130941092967987, + 0.020785901695489883, + 0.04676467925310135, + 0.08851955085992813, + 0.35972222685813904, + 0.23724089562892914, + 0.10811782628297806, + 0.013304800726473331, + 0.021096432581543922 + ], + "credit_map": [ + 0.010402902960777283, + 0.009988727048039436, + 0.010769137181341648, + 0.006724670063704252, + 0.007752399425953627, + 0.006901879329234362, + 0.012656115926802158, + 0.0111208725720644, + 0.003269467270001769, + 0.004643294028937817, + 0.003874319139868021, + 0.003521790262311697, + 0.003412635065615177, + 0.003426500828936696, + 0.003629051148891449, + 0.0023538863752037287, + 0.0032591274939477444, + 0.003558500437065959, + 0.002574439626187086, + 0.0026927513536065817, + 0.0019542695954442024, + 0.0018951001111418009, + 0.002984840888530016, + 0.0018668724223971367, + 0.004767262376844883, + 0.005819011479616165, + 0.0067133535631000996, + 0.005149772856384516, + 0.006225434597581625, + 0.004383235238492489, + 0.007767250295728445, + 0.00593936163932085, + 0.011534650810062885, + 0.01456284150481224, + 0.009363698773086071, + 0.010798502713441849, + 0.013593392446637154, + 0.007201221771538258, + 0.013107214123010635, + 0.008358036167919636, + 0.07960662245750427, + 0.04860948771238327, + 0.028324969112873077, + 0.03752848133444786, + 0.03670627996325493, + 0.032854702323675156, + 0.057711824774742126, + 0.03837989270687103, + 0.04440750926733017, + 0.02706119418144226, + 0.024180566892027855, + 0.021577106788754463, + 0.02632492035627365, + 0.022998454049229622, + 0.04203306511044502, + 0.02865811064839363, + 0.02719411440193653, + 0.014889970421791077, + 0.01273486576974392, + 0.009441196918487549, + 0.009812161326408386, + 0.010842031799256802, + 0.013426452875137329, + 0.009777048602700233, + 0.0015876911347731948, + 0.0017699365271255374, + 0.0016176198842003942, + 0.001311831409111619, + 0.0016951893921941519, + 0.0010592769831418991, + 0.002648371970281005, + 0.0016148857539519668, + 0.0041688960045576096, + 0.0021673657465726137, + 0.0028663468547165394, + 0.003080929396674037, + 0.0027071977965533733, + 0.0016863779164850712, + 0.0024455266539007425, + 0.0019737945403903723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.026341550052165985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_8/index_tensor([402, 402])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_8/index_tensor([402, 402])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..64a399dc453c29b0ecc1533e6cf99593927bf66f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_8/index_tensor([402, 402])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, arcing tracking shot, shot from a low angle, rotates with a solemn and ponderous rhythm around an ancient stone monument standing alone amidst a raging thunderstorm. The entire scene is rendered in a stark, high-contrast black-and-white film texture, its rough grain dancing across the frame, imbuing the image with a raw, unsettling quality. Thick storm clouds swirl beneath the inky night sky, while torrential rain lashes the stone monument's weathered surface, creating glistening streams across it. Carved from dark rock, the monument is riddled with spiderweb-like cracks and mottled with moss, its edges battered by the erosion of countless years. At the moment the camera was moving smoothly, a pale, forked lightning bolt tore through the sky without warning, dragging the entire world from extreme darkness into dazzling daylight in less than a second, instantly illuminating the blurred inscriptions on the stone tablets that had been worn down by rain and time to the point of being almost unrecognizable; in this brief moment of light, ancient runes flashed by, and then the light quickly faded, and the picture was once again swallowed up by the deep darkness and the sound of continuous heavy rain, leaving only the fleeting sense of mystery and the heavy oppression of history.", + "seed": 29364, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_8/index_402_seed_29364_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01759658716619015, + "vlm_logprob": -0.01759658716619015, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.01759658716619015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005400591529905796 + ], + "opposite_logprob_tokens": [ + -5.755400657653809 + ], + "logprob_reward": -0.005400591529905796, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.005400591529905796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12154440581798553, + 0.06848687678575516, + 0.06155350059270859, + 0.09956740587949753, + 0.04777342826128006, + 0.3833124339580536, + 0.03402640298008919, + 0.05158104747533798, + 0.05873952805995941, + 0.07341493666172028 + ], + "credit_map": [ + 0.011850947514176369, + 0.010917946696281433, + 0.03071722760796547, + 0.015357403084635735, + 0.013857658952474594, + 0.015188224613666534, + 0.0122377322986722, + 0.011417262256145477, + 0.0038361286278814077, + 0.006242258939892054, + 0.008015736006200314, + 0.007815413177013397, + 0.01346101239323616, + 0.015704259276390076, + 0.0075938086956739426, + 0.005818255711346865, + 0.00247763330116868, + 0.0027826365549117327, + 0.008831217885017395, + 0.006857169326394796, + 0.006558121647685766, + 0.015830019488930702, + 0.011115151457488537, + 0.007101551163941622, + 0.00650834571570158, + 0.006184609606862068, + 0.007185459602624178, + 0.01250309869647026, + 0.02051171474158764, + 0.013952652923762798, + 0.011395729146897793, + 0.021325794979929924, + 0.006238764151930809, + 0.006792279426008463, + 0.007336453069001436, + 0.007126018404960632, + 0.0047064595855772495, + 0.0029157563112676144, + 0.006885188166052103, + 0.005772508680820465, + 0.03301010653376579, + 0.0370553582906723, + 0.01028959359973669, + 0.025288503617048264, + 0.05383432283997536, + 0.029230773448944092, + 0.11995375901460648, + 0.07465001940727234, + 0.0029743933118879795, + 0.005183700937777758, + 0.0051560550928115845, + 0.0050247046165168285, + 0.006386734079569578, + 0.003858923679217696, + 0.003210986265912652, + 0.00223090429790318, + 0.0038359875325113535, + 0.008504647761583328, + 0.00868502352386713, + 0.008448136039078236, + 0.005558354314416647, + 0.006302359513938427, + 0.005488774739205837, + 0.004757764749228954, + 0.004313855431973934, + 0.006383261177688837, + 0.013897070661187172, + 0.006479702424257994, + 0.003963706083595753, + 0.010727659799158573, + 0.00504689384251833, + 0.007927375845611095, + 0.006006236188113689, + 0.008195440284907818, + 0.01113600842654705, + 0.00876597873866558, + 0.0072231232188642025, + 0.012069002725183964, + 0.011996835470199585, + 0.008022312074899673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.005400591529905796 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205877669155598 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.014205877669155598, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205877669155598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05369274318218231, + 0.1727914810180664, + 0.09983032941818237, + 0.03804469481110573, + 0.1397840529680252, + 0.08198487758636475, + 0.10353881120681763, + 0.03752734512090683, + 0.05335540696978569, + 0.2194502055644989 + ], + "credit_map": [ + 0.0044730668887495995, + 0.005856817122548819, + 0.009389905259013176, + 0.00946727953851223, + 0.007001628167927265, + 0.005632607266306877, + 0.00761635135859251, + 0.0042550875805318356, + 0.007183377631008625, + 0.013594403862953186, + 0.018167713657021523, + 0.0184397641569376, + 0.0435510091483593, + 0.03096812032163143, + 0.02772475779056549, + 0.01316234190016985, + 0.004835057072341442, + 0.004489599727094173, + 0.012456733733415604, + 0.010736693628132343, + 0.011735296808183193, + 0.02746563032269478, + 0.0174596905708313, + 0.010651636868715286, + 0.004482547752559185, + 0.0035771969705820084, + 0.004905425477772951, + 0.005133714992552996, + 0.003222731640562415, + 0.00589766725897789, + 0.005564451217651367, + 0.005260958336293697, + 0.02336028218269348, + 0.020585980266332626, + 0.012048343196511269, + 0.017486058175563812, + 0.017677363008260727, + 0.00657022837549448, + 0.019003186374902725, + 0.02305261231958866, + 0.008942347951233387, + 0.01023898459970951, + 0.004583136644214392, + 0.011474722996354103, + 0.010168925859034061, + 0.007527797482907772, + 0.011181439273059368, + 0.017867516726255417, + 0.01183957327157259, + 0.008462104946374893, + 0.01631624810397625, + 0.007664871402084827, + 0.029842540621757507, + 0.012348845601081848, + 0.01029213797301054, + 0.006772489286959171, + 0.0025963103398680687, + 0.005073221866041422, + 0.006688669789582491, + 0.005665434058755636, + 0.004014964681118727, + 0.0050882562063634396, + 0.0037388228811323643, + 0.0046616666950285435, + 0.00426093116402626, + 0.006904074922204018, + 0.01138411182910204, + 0.004460359923541546, + 0.004452415741980076, + 0.013178776949644089, + 0.0034736318048089743, + 0.005241105332970619, + 0.016781503334641457, + 0.0218596663326025, + 0.03466993570327759, + 0.021738184615969658, + 0.015820713713765144, + 0.04246509447693825, + 0.04178408533334732, + 0.024331023916602135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.014205877669155598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420164667069912 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.01420164667069912, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420164667069912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20121259987354279, + 0.07168689370155334, + 0.06500032544136047, + 0.1617298573255539, + 0.04879958555102348, + 0.09574083238840103, + 0.022491520270705223, + 0.102312371134758, + 0.04245040938258171, + 0.1885756105184555 + ], + "credit_map": [ + 0.014927108772099018, + 0.018467852845788002, + 0.04752017557621002, + 0.02538549154996872, + 0.029600443318486214, + 0.023592311888933182, + 0.01756640523672104, + 0.02415280230343342, + 0.004344712942838669, + 0.005214320961385965, + 0.009031604044139385, + 0.010278834961354733, + 0.012717142701148987, + 0.016759026795625687, + 0.007112682797014713, + 0.006228571757674217, + 0.0027986466884613037, + 0.003400467336177826, + 0.011263140477240086, + 0.008789047598838806, + 0.009673386812210083, + 0.010223710909485817, + 0.012410963885486126, + 0.0064409635961055756, + 0.013003258034586906, + 0.010859621688723564, + 0.012324887327849865, + 0.01685081236064434, + 0.04278681054711342, + 0.024922551587224007, + 0.025458378717303276, + 0.01552354171872139, + 0.006685601081699133, + 0.007278783246874809, + 0.006005784496665001, + 0.006670712027698755, + 0.005787416826933622, + 0.0025150487199425697, + 0.005582873243838549, + 0.008273366838693619, + 0.007805489469319582, + 0.015401839278638363, + 0.00565107399597764, + 0.016849685460329056, + 0.011308738961815834, + 0.017133483663201332, + 0.010491455905139446, + 0.011099065653979778, + 0.0022490147966891527, + 0.003319042967632413, + 0.002582903252914548, + 0.0027338743675500154, + 0.00494439946487546, + 0.002612914890050888, + 0.002438799012452364, + 0.0016105722170323133, + 0.009948620572686195, + 0.01006323378533125, + 0.024537697434425354, + 0.008731271140277386, + 0.012950933538377285, + 0.019860124215483665, + 0.008075658231973648, + 0.008144823834300041, + 0.0026776217855513096, + 0.003965921234339476, + 0.010569180361926556, + 0.004436214454472065, + 0.0029802871868014336, + 0.007790135219693184, + 0.003768214024603367, + 0.00626283697783947, + 0.013313530012965202, + 0.020708810538053513, + 0.028948361054062843, + 0.02201630361378193, + 0.015767276287078857, + 0.030069595202803612, + 0.03590622544288635, + 0.021845506504178047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.01420164667069912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072357073426247 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072357073426247, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072357073426247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13657748699188232, + 0.12491745501756668, + 0.052072398364543915, + 0.059952396899461746, + 0.042459506541490555, + 0.2831999361515045, + 0.05662636086344719, + 0.08095511049032211, + 0.0584290437400341, + 0.10481029748916626 + ], + "credit_map": [ + 0.012968488037586212, + 0.012830938212573528, + 0.04071652889251709, + 0.02321045845746994, + 0.013060640543699265, + 0.01411732379347086, + 0.011060560122132301, + 0.008612548932433128, + 0.004951043054461479, + 0.008958933874964714, + 0.020806970074772835, + 0.03022409789264202, + 0.017972426488995552, + 0.017762895673513412, + 0.015798769891262054, + 0.00844231341034174, + 0.002376348478719592, + 0.0032394323498010635, + 0.007775426376610994, + 0.0059574320912361145, + 0.011318917386233807, + 0.008334523998200893, + 0.007851607166230679, + 0.005218713078647852, + 0.005949020851403475, + 0.004026153590530157, + 0.007377132307738066, + 0.00899525173008442, + 0.004486401099711657, + 0.006104563362896442, + 0.013362371362745762, + 0.009651508182287216, + 0.006506717298179865, + 0.006383978761732578, + 0.00668315077200532, + 0.007241513580083847, + 0.004697010386735201, + 0.002129064407199621, + 0.004448551218956709, + 0.00436951732262969, + 0.024998661130666733, + 0.03058134764432907, + 0.012812511995434761, + 0.031298037618398666, + 0.029048528522253036, + 0.02584891766309738, + 0.07350926101207733, + 0.055102672427892685, + 0.0062700239941477776, + 0.00661616213619709, + 0.009090258739888668, + 0.006455948110669851, + 0.00940367579460144, + 0.007175171282142401, + 0.006081509869545698, + 0.005533612333238125, + 0.005776062607765198, + 0.013590961694717407, + 0.015226881951093674, + 0.011068773455917835, + 0.009051364846527576, + 0.01180553063750267, + 0.008151696994900703, + 0.006283838767558336, + 0.003977774176746607, + 0.005669070407748222, + 0.011305204592645168, + 0.0082467719912529, + 0.004063000436872244, + 0.01137740258127451, + 0.0057485150173306465, + 0.008041305467486382, + 0.008093181066215038, + 0.011673708446323872, + 0.01783733069896698, + 0.010770132765173912, + 0.008428297005593777, + 0.019428784027695656, + 0.017726365476846695, + 0.010852488689124584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.016072357073426247 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810246288776398 + ], + "opposite_logprob_tokens": [ + -3.288102388381958 + ], + "logprob_reward": -0.03810246288776398, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810246288776398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09718579798936844, + 0.08326170593500137, + 0.04678083956241608, + 0.03216603770852089, + 0.040589675307273865, + 0.27233317494392395, + 0.12345004826784134, + 0.04010170325636864, + 0.165556862950325, + 0.09857425093650818 + ], + "credit_map": [ + 0.00992805976420641, + 0.008543666452169418, + 0.014553843066096306, + 0.011986632831394672, + 0.012204146012663841, + 0.012270687147974968, + 0.013292917050421238, + 0.014405847527086735, + 0.0039515565149486065, + 0.005765043199062347, + 0.011671601794660091, + 0.012322580441832542, + 0.013840981759130955, + 0.0179226603358984, + 0.010760942474007607, + 0.00702633848413825, + 0.0017336372984573245, + 0.0023951432667672634, + 0.007914799265563488, + 0.005303668323904276, + 0.004566287621855736, + 0.009112461470067501, + 0.010232532396912575, + 0.005522312596440315, + 0.0034053949639201164, + 0.002850443124771118, + 0.004743961151689291, + 0.004323466215282679, + 0.0021018246188759804, + 0.0024356809444725513, + 0.006769758649170399, + 0.005535507574677467, + 0.005477136932313442, + 0.006246916949748993, + 0.006709537468850613, + 0.00811490137130022, + 0.00360552198253572, + 0.002158206421881914, + 0.005000299774110317, + 0.0032771555706858635, + 0.024511346593499184, + 0.023860961198806763, + 0.008299124427139759, + 0.016905661672353745, + 0.03651069477200508, + 0.02096380479633808, + 0.08360821008682251, + 0.05767334997653961, + 0.015033056028187275, + 0.012538332492113113, + 0.017018979415297508, + 0.012492276728153229, + 0.032574087381362915, + 0.013126752339303493, + 0.012646234594285488, + 0.008020333014428616, + 0.003020518459379673, + 0.005550211761146784, + 0.009526081383228302, + 0.005782528314739466, + 0.004069296643137932, + 0.0054235076531767845, + 0.0029157977551221848, + 0.003813759656623006, + 0.01247281115502119, + 0.015446359291672707, + 0.03950301930308342, + 0.022399622946977615, + 0.011698896996676922, + 0.027512909844517708, + 0.013930863700807095, + 0.022592375054955482, + 0.006872795056551695, + 0.011050743982195854, + 0.015553997829556465, + 0.014626662246882915, + 0.006073288153856993, + 0.020905539393424988, + 0.012283073738217354, + 0.01120815984904766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.03810246288776398 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, arcing tracking shot, shot from a low angle, rotates with a solemn and ponderous rhythm around an ancient stone monument standing alone amidst a raging thunderstorm. The entire scene is rendered in a stark, high-contrast black-and-white film texture, its rough grain dancing across the frame, imbuing the image with a raw, unsettling quality. Thick storm clouds swirl beneath the inky night sky, while torrential rain lashes the stone monument's weathered surface, creating glistening streams across it. Carved from dark rock, the monument is riddled with spiderweb-like cracks and mottled with moss, its edges battered by the erosion of countless years. At the moment the camera was moving smoothly, a pale, forked lightning bolt tore through the sky without warning, dragging the entire world from extreme darkness into dazzling daylight in less than a second, instantly illuminating the blurred inscriptions on the stone tablets that had been worn down by rain and time to the point of being almost unrecognizable; in this brief moment of light, ancient runes flashed by, and then the light quickly faded, and the picture was once again swallowed up by the deep darkness and the sound of continuous heavy rain, leaving only the fleeting sense of mystery and the heavy oppression of history.", + "seed": 29365, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_8/index_402_seed_29365_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020784533023834227, + "vlm_logprob": -0.020784533023834227, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.020784533023834227, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04720474034547806 + ], + "opposite_logprob_tokens": [ + -3.1722047328948975 + ], + "logprob_reward": -0.04720474034547806, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04720474034547806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1017264649271965, + 0.025955185294151306, + 0.030491674318909645, + 0.035913094878196716, + 0.10793677717447281, + 0.022767825052142143, + 0.2794685363769531, + 0.12357890605926514, + 0.08251386135816574, + 0.18964757025241852 + ], + "credit_map": [ + 0.008889026008546352, + 0.01409902237355709, + 0.010115738958120346, + 0.015509242191910744, + 0.010964764282107353, + 0.02274012565612793, + 0.007779954932630062, + 0.011628597043454647, + 0.0021064772736281157, + 0.005527946632355452, + 0.0017266490031033754, + 0.004628082271665335, + 0.0029245007317513227, + 0.0024678607005625963, + 0.0027428618632256985, + 0.0038308093789964914, + 0.002099275356158614, + 0.005365247838199139, + 0.0023482153192162514, + 0.004222130868583918, + 0.005046191159635782, + 0.004838492721319199, + 0.003510355716571212, + 0.0030617685988545418, + 0.003867235966026783, + 0.0038604396395385265, + 0.0023183252196758986, + 0.010935762897133827, + 0.0033972496166825294, + 0.0053200190886855125, + 0.0029257857240736485, + 0.0032882774248719215, + 0.009497617371380329, + 0.007709277328103781, + 0.0038931146264076233, + 0.019900647923350334, + 0.006404420360922813, + 0.012700885534286499, + 0.007156745996326208, + 0.040674079209566116, + 0.0022119744680821896, + 0.0045859781093895435, + 0.001665950519964099, + 0.002834947779774666, + 0.003190362825989723, + 0.0024243704974651337, + 0.0029675308614969254, + 0.0028867118526250124, + 0.01189623586833477, + 0.010730368085205555, + 0.007621897384524345, + 0.005730354692786932, + 0.013401215896010399, + 0.10882987827062607, + 0.00790536031126976, + 0.11335325986146927, + 0.010296152904629707, + 0.021211905404925346, + 0.009104962460696697, + 0.01095051970332861, + 0.009554984979331493, + 0.02348732203245163, + 0.013578384183347225, + 0.02539469301700592, + 0.01572403311729431, + 0.005640811286866665, + 0.003957195673137903, + 0.006570514757186174, + 0.007105107419192791, + 0.01731487549841404, + 0.008037757128477097, + 0.01816357672214508, + 0.01649942621588707, + 0.018344270065426826, + 0.023577505722641945, + 0.028240256011486053, + 0.029101727530360222, + 0.02442149631679058, + 0.03223308548331261, + 0.017229823395609856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.04720474034547806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746016908437014 + ], + "opposite_logprob_tokens": [ + -5.0067458152771 + ], + "logprob_reward": -0.006746016908437014, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746016908437014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07877440750598907, + 0.04212595149874687, + 0.0822630524635315, + 0.05485067516565323, + 0.08339980989694595, + 0.037573229521512985, + 0.2939821481704712, + 0.15111954510211945, + 0.08889806270599365, + 0.0870131328701973 + ], + "credit_map": [ + 0.007186254020780325, + 0.014819508418440819, + 0.008401213213801384, + 0.012581407092511654, + 0.01076613087207079, + 0.014051270671188831, + 0.004469019360840321, + 0.006499608978629112, + 0.0033445528242737055, + 0.008055939339101315, + 0.002642751205712557, + 0.007083402015268803, + 0.006908134091645479, + 0.006925733759999275, + 0.0031724697910249233, + 0.003992971032857895, + 0.0046466500498354435, + 0.006957644131034613, + 0.007210609037429094, + 0.012636001221835613, + 0.012731295078992844, + 0.017980074509978294, + 0.007473303005099297, + 0.012627474963665009, + 0.004658449906855822, + 0.0040881335735321045, + 0.0027273220475763083, + 0.013000347651541233, + 0.005954404827207327, + 0.007055320776998997, + 0.0039824265986680984, + 0.013384271413087845, + 0.0069807060062885284, + 0.005255769472569227, + 0.0036220988258719444, + 0.015701614320278168, + 0.00453633489087224, + 0.009843613021075726, + 0.0049017807468771935, + 0.0325578935444355, + 0.002500572009012103, + 0.005578125827014446, + 0.0032698248978704214, + 0.004255735315382481, + 0.007637982256710529, + 0.004020899534225464, + 0.006236284971237183, + 0.00407380098477006, + 0.01364438608288765, + 0.005898274481296539, + 0.007245034445077181, + 0.006416175048798323, + 0.01321688387542963, + 0.11605561524629593, + 0.00943491980433464, + 0.12207084894180298, + 0.01618579402565956, + 0.01461691316217184, + 0.00938230101019144, + 0.007921869866549969, + 0.012667921371757984, + 0.03937290608882904, + 0.011358017101883888, + 0.039613813161849976, + 0.015468585304915905, + 0.006477641873061657, + 0.004115853924304247, + 0.0048666284419596195, + 0.008327305316925049, + 0.019762655720114708, + 0.008031470701098442, + 0.02184792049229145, + 0.008025562390685081, + 0.008818057365715504, + 0.009281794540584087, + 0.0129024563357234, + 0.014142746105790138, + 0.011415105313062668, + 0.014632739126682281, + 0.007794670294970274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.006746016908437014 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369959816336632 + ], + "opposite_logprob_tokens": [ + -3.4086995124816895 + ], + "logprob_reward": -0.03369959816336632, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369959816336632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0631863996386528, + 0.02431662194430828, + 0.013736696913838387, + 0.015942422673106194, + 0.010950261726975441, + 0.03203254193067551, + 0.6419885754585266, + 0.10451150685548782, + 0.03691688925027847, + 0.056418076157569885 + ], + "credit_map": [ + 0.005087411031126976, + 0.009678314439952374, + 0.006045099813491106, + 0.009900104254484177, + 0.007146923337131739, + 0.014262677170336246, + 0.003984357230365276, + 0.007081517018377781, + 0.0015245452523231506, + 0.003477957332506776, + 0.0016668490134179592, + 0.003407879499718547, + 0.003963470458984375, + 0.003533954033628106, + 0.002833244390785694, + 0.003908723127096891, + 0.0010007943492382765, + 0.0025060942862182856, + 0.0008677593432366848, + 0.0020901344250887632, + 0.002331457333639264, + 0.0021659964695572853, + 0.0012222931254655123, + 0.0015521679306402802, + 0.0013035776792094111, + 0.001209240173920989, + 0.0008614192483946681, + 0.0027459438424557447, + 0.002940211910754442, + 0.0029881305526942015, + 0.0013665871229022741, + 0.0025273114442825317, + 0.0005479453830048442, + 0.0021188841201364994, + 0.0007468742551282048, + 0.0016520596109330654, + 0.002152665052562952, + 0.0014200430596247315, + 0.0012009921483695507, + 0.0011107976315543056, + 0.0030072196386754513, + 0.005772449541836977, + 0.0035302084870636463, + 0.0032324958592653275, + 0.004340062383562326, + 0.004326967988163233, + 0.0038463708478957415, + 0.003976764623075724, + 0.02834227681159973, + 0.012712365947663784, + 0.016524115577340126, + 0.011975840665400028, + 0.029779905453324318, + 0.2566083073616028, + 0.018679143860936165, + 0.2673666179180145, + 0.009923317469656467, + 0.010423199273645878, + 0.007638446521013975, + 0.0054989103227853775, + 0.015839222818613052, + 0.023887163028120995, + 0.009057729505002499, + 0.022243525832891464, + 0.005618198309093714, + 0.003437864128500223, + 0.0019136922201141715, + 0.0035751499235630035, + 0.003768158843740821, + 0.0068548088893294334, + 0.003491887589916587, + 0.008257129229605198, + 0.006094267591834068, + 0.005193729884922504, + 0.004154743626713753, + 0.00677858293056488, + 0.010119694285094738, + 0.00749205844476819, + 0.008213804103434086, + 0.00837118923664093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.03369959816336632 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630250882357359 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630250882357359, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630250882357359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0200494471937418, + 0.010320906527340412, + 0.008929102681577206, + 0.09485774487257004, + 0.06693834811449051, + 0.01705629751086235, + 0.5365200042724609, + 0.13265185058116913, + 0.038017842918634415, + 0.07465852797031403 + ], + "credit_map": [ + 0.0018720432417467237, + 0.004387123975902796, + 0.001830719062127173, + 0.0034045062493532896, + 0.0021972982212901115, + 0.00265299528837204, + 0.0014625736512243748, + 0.0022421882022172213, + 0.0006035802653059363, + 0.0027381223626434803, + 0.0007214404758997262, + 0.001721475855447352, + 0.0014708996750414371, + 0.0012340977555140853, + 0.0007129922159947455, + 0.0011182976886630058, + 0.0006248881109058857, + 0.0020352783612906933, + 0.0007400789181701839, + 0.0013037592871114612, + 0.001149787800386548, + 0.0012339870445430279, + 0.0008667422225698829, + 0.0009745804709382355, + 0.00852117221802473, + 0.005878536030650139, + 0.004192649386823177, + 0.03121596947312355, + 0.007117665372788906, + 0.012877905741333961, + 0.006475924514234066, + 0.01857791841030121, + 0.005947633180767298, + 0.003947367891669273, + 0.002536727348342538, + 0.01271478645503521, + 0.0031112185679376125, + 0.008123059757053852, + 0.003339116694405675, + 0.02721843682229519, + 0.0010105345863848925, + 0.003286807332187891, + 0.001303719007410109, + 0.0018850063206627965, + 0.002617586636915803, + 0.0021086265332996845, + 0.0023241103626787663, + 0.0025199074298143387, + 0.023629069328308105, + 0.012340210378170013, + 0.013927535153925419, + 0.009797582402825356, + 0.026120947673916817, + 0.21287555992603302, + 0.01651054620742798, + 0.2213185727596283, + 0.013542640954256058, + 0.03336980938911438, + 0.00888818595558405, + 0.010579019784927368, + 0.007911767810583115, + 0.024866078048944473, + 0.009676434099674225, + 0.02381790801882744, + 0.004753197077661753, + 0.0051467446610331535, + 0.002182172378525138, + 0.005946692079305649, + 0.004009513184428215, + 0.006219377275556326, + 0.0051728482358157635, + 0.004587299656122923, + 0.008283644914627075, + 0.007483129855245352, + 0.006762627977877855, + 0.013386757113039494, + 0.007892568595707417, + 0.010700997896492481, + 0.010113473981618881, + 0.010035324841737747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.007630250882357359 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642058819532394 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642058819532394, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642058819532394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12260899692773819, + 0.05535915493965149, + 0.032523028552532196, + 0.05676436796784401, + 0.20217300951480865, + 0.04207376390695572, + 0.14787082374095917, + 0.1909402459859848, + 0.12176542729139328, + 0.027921229600906372 + ], + "credit_map": [ + 0.008587982505559921, + 0.017272168770432472, + 0.012017087079584599, + 0.02013278752565384, + 0.014614924788475037, + 0.028132611885666847, + 0.00926487147808075, + 0.01258654985576868, + 0.003728671232238412, + 0.011749655939638615, + 0.005365272983908653, + 0.008278601802885532, + 0.008308150805532932, + 0.006739647127687931, + 0.0037207340355962515, + 0.007468419149518013, + 0.0017316964222118258, + 0.004323827568441629, + 0.002470278413966298, + 0.00489651458337903, + 0.005095250904560089, + 0.007119435351341963, + 0.002854949561879039, + 0.004031071439385414, + 0.0041568991728127, + 0.005614362191408873, + 0.0027005267329514027, + 0.011526279151439667, + 0.006961627397686243, + 0.009239624254405499, + 0.004037164151668549, + 0.01252787746489048, + 0.01911388710141182, + 0.010061399079859257, + 0.007237367331981659, + 0.039172522723674774, + 0.007681047078222036, + 0.02527410350739956, + 0.008910582400858402, + 0.08472207933664322, + 0.0029700829181820154, + 0.005542232654988766, + 0.004244477488100529, + 0.004898174665868282, + 0.007927239872515202, + 0.005306157749146223, + 0.005819469690322876, + 0.005365924444049597, + 0.008987880311906338, + 0.011705314740538597, + 0.00816018134355545, + 0.006022136192768812, + 0.012213415466248989, + 0.04365789517760277, + 0.010630493052303791, + 0.04649350419640541, + 0.02005677856504917, + 0.036850497126579285, + 0.010688167996704578, + 0.015165434218943119, + 0.01472784485667944, + 0.03480688855051994, + 0.020400838926434517, + 0.038243770599365234, + 0.018760541453957558, + 0.009347125887870789, + 0.00787744577974081, + 0.010585904121398926, + 0.015948012471199036, + 0.022859584540128708, + 0.009812615811824799, + 0.026574192568659782, + 0.002874004188925028, + 0.0022792858071625233, + 0.0033529908396303654, + 0.0022408231161534786, + 0.003462526947259903, + 0.004288308788090944, + 0.004112252965569496, + 0.00531103415414691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.008642058819532394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, arcing tracking shot, shot from a low angle, rotates with a solemn and ponderous rhythm around an ancient stone monument standing alone amidst a raging thunderstorm. The entire scene is rendered in a stark, high-contrast black-and-white film texture, its rough grain dancing across the frame, imbuing the image with a raw, unsettling quality. Thick storm clouds swirl beneath the inky night sky, while torrential rain lashes the stone monument's weathered surface, creating glistening streams across it. Carved from dark rock, the monument is riddled with spiderweb-like cracks and mottled with moss, its edges battered by the erosion of countless years. At the moment the camera was moving smoothly, a pale, forked lightning bolt tore through the sky without warning, dragging the entire world from extreme darkness into dazzling daylight in less than a second, instantly illuminating the blurred inscriptions on the stone tablets that had been worn down by rain and time to the point of being almost unrecognizable; in this brief moment of light, ancient runes flashed by, and then the light quickly faded, and the picture was once again swallowed up by the deep darkness and the sound of continuous heavy rain, leaving only the fleeting sense of mystery and the heavy oppression of history.", + "seed": 29366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_8/index_402_seed_29366_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017901441548019648, + "vlm_logprob": -0.017901441548019648, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.017901441548019648, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004542983137071133 + ], + "opposite_logprob_tokens": [ + -5.879542827606201 + ], + "logprob_reward": -0.004542983137071133, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004542983137071133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2866389751434326, + 0.10009697824716568, + 0.03717488795518875, + 0.06812973320484161, + 0.0609988272190094, + 0.07666850835084915, + 0.1916137933731079, + 0.11580269783735275, + 0.04382048919796944, + 0.019055066630244255 + ], + "credit_map": [ + 0.02195156179368496, + 0.012833633460104465, + 0.02725602127611637, + 0.013516243547201157, + 0.017054451629519463, + 0.1121068149805069, + 0.010113540105521679, + 0.07180669903755188, + 0.006763475481420755, + 0.014938480220735073, + 0.013401052914559841, + 0.015413838438689709, + 0.01655869372189045, + 0.012319629080593586, + 0.014909938909113407, + 0.005791863892227411, + 0.003029098268598318, + 0.002484958153218031, + 0.00582536868751049, + 0.0030709700658917427, + 0.003915747627615929, + 0.007714347448199987, + 0.005853555630892515, + 0.005280840676277876, + 0.0063637252897024155, + 0.004715850111097097, + 0.00918217096477747, + 0.006725550629198551, + 0.006533466745167971, + 0.011521331034600735, + 0.011026911437511444, + 0.012060727924108505, + 0.0051619200967252254, + 0.0023429079446941614, + 0.010332743637263775, + 0.0028714430518448353, + 0.0028720705304294825, + 0.01969807967543602, + 0.003784272586926818, + 0.013935389928519726, + 0.008769278414547443, + 0.005235838703811169, + 0.0132027268409729, + 0.005516780540347099, + 0.0063497573137283325, + 0.013702807016670704, + 0.012926764786243439, + 0.010964551940560341, + 0.007588018197566271, + 0.0030442404095083475, + 0.0077158138155937195, + 0.00436696782708168, + 0.008421928621828556, + 0.0762147605419159, + 0.011560946702957153, + 0.07270113378763199, + 0.006185569334775209, + 0.009105049073696136, + 0.003981729503720999, + 0.006086267530918121, + 0.007293837610632181, + 0.04019784554839134, + 0.006405345164239407, + 0.03654704988002777, + 0.005039404146373272, + 0.005968173500150442, + 0.005170082673430443, + 0.004823480732738972, + 0.004758499562740326, + 0.006503242067992687, + 0.0033063108567148447, + 0.008251294493675232, + 0.0031720944680273533, + 0.0022458031307905912, + 0.003322063246741891, + 0.003099427791312337, + 0.001756468671374023, + 0.0027707775589078665, + 0.0015379000687971711, + 0.0011505320435389876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.004542983137071133 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189996480941772 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018189996480941772, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189996480941772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13826066255569458, + 0.11573746055364609, + 0.03162280097603798, + 0.030604088678956032, + 0.14834104478359222, + 0.045612018555402756, + 0.09227580577135086, + 0.029077667742967606, + 0.33969399333000183, + 0.028774432837963104 + ], + "credit_map": [ + 0.011460484005510807, + 0.008195914328098297, + 0.015533472411334515, + 0.011006536893546581, + 0.008616380393505096, + 0.04691045358777046, + 0.0072236736305058, + 0.029313748702406883, + 0.010101024992763996, + 0.015595735050737858, + 0.015441665425896645, + 0.012855746783316135, + 0.021420706063508987, + 0.01970408484339714, + 0.013578807935118675, + 0.007039687596261501, + 0.0027556458953768015, + 0.0014801077777519822, + 0.0048845866695046425, + 0.0027060124557465315, + 0.0020328499376773834, + 0.007696370128542185, + 0.0036687194369733334, + 0.006398508325219154, + 0.004493759945034981, + 0.0011960359988734126, + 0.0044147246517241, + 0.0018533802358433604, + 0.001330629806034267, + 0.007802847307175398, + 0.0025339596904814243, + 0.006978751625865698, + 0.005578869022428989, + 0.003517693607136607, + 0.009937948547303677, + 0.004670053254812956, + 0.0028305905871093273, + 0.061782073229551315, + 0.0030494234524667263, + 0.0569743886590004, + 0.0059276157990098, + 0.0028568864800035954, + 0.009486931376159191, + 0.0042130351066589355, + 0.004370264708995819, + 0.006523780524730682, + 0.004958298522979021, + 0.007275201380252838, + 0.013612760230898857, + 0.004802262410521507, + 0.011785304173827171, + 0.00784266833215952, + 0.012935547158122063, + 0.018492596223950386, + 0.009149787947535515, + 0.013654880225658417, + 0.004140518605709076, + 0.005532986484467983, + 0.0023611572105437517, + 0.0029501195531338453, + 0.0035237353295087814, + 0.004715531133115292, + 0.0028746547177433968, + 0.0029789626132696867, + 0.04525579884648323, + 0.041535623371601105, + 0.06039055809378624, + 0.034140683710575104, + 0.026761429384350777, + 0.04814048483967781, + 0.03184428811073303, + 0.05162513256072998, + 0.003487049136310816, + 0.0021113979164510965, + 0.007368429098278284, + 0.0034764136653393507, + 0.002421468961983919, + 0.005567957181483507, + 0.001619053422473371, + 0.002722664037719369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.018189996480941772 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369544818997383 + ], + "opposite_logprob_tokens": [ + -3.4086954593658447 + ], + "logprob_reward": -0.03369544818997383, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369544818997383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31433066725730896, + 0.06401755660772324, + 0.07447818666696548, + 0.040367692708969116, + 0.07519443333148956, + 0.08524120599031448, + 0.06054355949163437, + 0.15857361257076263, + 0.06037479266524315, + 0.06687839329242706 + ], + "credit_map": [ + 0.023077143356204033, + 0.018740272149443626, + 0.03141620382666588, + 0.022385580465197563, + 0.017597468569874763, + 0.11468452960252762, + 0.01169961504638195, + 0.07472983002662659, + 0.005604276899248362, + 0.007812537252902985, + 0.009635101072490215, + 0.009174623526632786, + 0.012230398133397102, + 0.006256855558604002, + 0.00960040558129549, + 0.003703352529555559, + 0.005482127424329519, + 0.002891774754971266, + 0.022578608244657516, + 0.00605587475001812, + 0.003532610833644867, + 0.012620031833648682, + 0.00729760667309165, + 0.014019553549587727, + 0.00595686724409461, + 0.0015538420993834734, + 0.006910623051226139, + 0.0027247897814959288, + 0.0016547878040000796, + 0.008526184596121311, + 0.0038221857976168394, + 0.009218405932188034, + 0.0038756737485527992, + 0.0018844383303076029, + 0.007534136530011892, + 0.0021674735471606255, + 0.0019187903963029385, + 0.028472667559981346, + 0.0026837345212697983, + 0.02665751241147518, + 0.005921524949371815, + 0.0035439354833215475, + 0.008578531444072723, + 0.00470426632091403, + 0.005219554994255304, + 0.02334403432905674, + 0.007313653361052275, + 0.026615696027874947, + 0.005275411531329155, + 0.0023717458825558424, + 0.0055503929033875465, + 0.00395390298217535, + 0.006364903878420591, + 0.016737233847379684, + 0.005151661578565836, + 0.015138305723667145, + 0.0097886947914958, + 0.008250932209193707, + 0.005557789001613855, + 0.004834337159991264, + 0.009382855147123337, + 0.062057316303253174, + 0.006087725516408682, + 0.05261394381523132, + 0.007664130534976721, + 0.006824849639087915, + 0.010900718159973621, + 0.008028977550566196, + 0.0058105443604290485, + 0.007433463353663683, + 0.005167095921933651, + 0.008545011281967163, + 0.007096493151038885, + 0.004833736456930637, + 0.015479051508009434, + 0.008633527904748917, + 0.006047085393220186, + 0.013063263148069382, + 0.004550348035991192, + 0.007174880243837833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.03369544818997383 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735952105373144 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735952105373144, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735952105373144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18121419847011566, + 0.03327525407075882, + 0.06455692648887634, + 0.11022541671991348, + 0.15239062905311584, + 0.10933269560337067, + 0.07936309278011322, + 0.08112435042858124, + 0.1333678513765335, + 0.0551496185362339 + ], + "credit_map": [ + 0.013194185681641102, + 0.008726519532501698, + 0.019848033785820007, + 0.01074904389679432, + 0.008731376379728317, + 0.06752903014421463, + 0.008896490558981895, + 0.04353950545191765, + 0.003171508898958564, + 0.0034613353200256824, + 0.00535448407754302, + 0.003900049487128854, + 0.005201367661356926, + 0.005055426154285669, + 0.004603181965649128, + 0.002527899108827114, + 0.005438289605081081, + 0.004304022528231144, + 0.010162269696593285, + 0.007776058744639158, + 0.006903266068547964, + 0.012883208692073822, + 0.007164396345615387, + 0.009925409220159054, + 0.014629027806222439, + 0.006847226992249489, + 0.021836284548044205, + 0.009255791082978249, + 0.009223618544638157, + 0.021064631640911102, + 0.01684536412358284, + 0.010523461736738682, + 0.008362484164536, + 0.003274139016866684, + 0.01166126225143671, + 0.005952396895736456, + 0.003019362688064575, + 0.059000469744205475, + 0.006013630889356136, + 0.05510688200592995, + 0.006169924978166819, + 0.0065233903005719185, + 0.008349000476300716, + 0.005776106379926205, + 0.0075532738119363785, + 0.02974620833992958, + 0.011284071020781994, + 0.033930715173482895, + 0.004283042624592781, + 0.002868100767955184, + 0.004413536749780178, + 0.0036734244786202908, + 0.00534324673935771, + 0.02549455501139164, + 0.005576162599027157, + 0.027711020782589912, + 0.004751995671540499, + 0.007292372640222311, + 0.004491487052291632, + 0.0072148023173213005, + 0.007072013802826405, + 0.023145481944084167, + 0.007470890413969755, + 0.019685301929712296, + 0.017406033352017403, + 0.01735703833401203, + 0.022497227415442467, + 0.012728759087622166, + 0.011354726739227772, + 0.019460180774331093, + 0.01322640385478735, + 0.019337469711899757, + 0.00701409624889493, + 0.005507326684892178, + 0.010401439853012562, + 0.006506212521344423, + 0.005469511728733778, + 0.009952071122825146, + 0.003978805150836706, + 0.00632015336304903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.006735952105373144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342827826738358 + ], + "opposite_logprob_tokens": [ + -3.6513428688049316 + ], + "logprob_reward": -0.026342827826738358, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342827826738358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13958072662353516, + 0.05603528395295143, + 0.044363778084516525, + 0.04485444352030754, + 0.1575823426246643, + 0.2297019213438034, + 0.22591373324394226, + 0.03087637573480606, + 0.03291028365492821, + 0.03818107768893242 + ], + "credit_map": [ + 0.009976801462471485, + 0.007107520010322332, + 0.01432886067777872, + 0.006672235671430826, + 0.00888307485729456, + 0.05315841734409332, + 0.005242367275059223, + 0.03421145677566528, + 0.006225616671144962, + 0.006920811720192432, + 0.008892541751265526, + 0.009320363402366638, + 0.007715270388871431, + 0.0070150624960660934, + 0.006949781905859709, + 0.0029958421364426613, + 0.004232149571180344, + 0.002097621327266097, + 0.009671999141573906, + 0.005004298407584429, + 0.003282576333731413, + 0.007795002777129412, + 0.006289496552199125, + 0.005990638863295317, + 0.006486528553068638, + 0.0021332886535674334, + 0.008983468636870384, + 0.003138098167255521, + 0.002206051256507635, + 0.00788276456296444, + 0.0059991078451275826, + 0.008025139570236206, + 0.00913300085812807, + 0.0032041745726019144, + 0.016063451766967773, + 0.004250162281095982, + 0.0027692080475389957, + 0.06160294637084007, + 0.004863522481173277, + 0.05569588020443916, + 0.023598534986376762, + 0.01282606739550829, + 0.04035953804850578, + 0.016135530546307564, + 0.016918709501624107, + 0.0453912727534771, + 0.022806357592344284, + 0.05166592448949814, + 0.02059011161327362, + 0.008821245282888412, + 0.020513013005256653, + 0.013498102314770222, + 0.01751740090548992, + 0.06217733770608902, + 0.01612936705350876, + 0.06666718423366547, + 0.002337651327252388, + 0.0030124709010124207, + 0.0022142331581562757, + 0.0017311557894572616, + 0.0026491505559533834, + 0.00913457665592432, + 0.0024820538237690926, + 0.007315084338188171, + 0.004231933504343033, + 0.0038084902334958315, + 0.004340307787060738, + 0.004453766625374556, + 0.0028741667047142982, + 0.004523586481809616, + 0.0024944492615759373, + 0.006183587480336428, + 0.004065986722707748, + 0.0019599327351897955, + 0.008543787524104118, + 0.004606526345014572, + 0.003419793676584959, + 0.008568254299461842, + 0.002489950507879257, + 0.004526851233094931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.026342827826738358 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, arcing tracking shot, shot from a low angle, rotates with a solemn and ponderous rhythm around an ancient stone monument standing alone amidst a raging thunderstorm. The entire scene is rendered in a stark, high-contrast black-and-white film texture, its rough grain dancing across the frame, imbuing the image with a raw, unsettling quality. Thick storm clouds swirl beneath the inky night sky, while torrential rain lashes the stone monument's weathered surface, creating glistening streams across it. Carved from dark rock, the monument is riddled with spiderweb-like cracks and mottled with moss, its edges battered by the erosion of countless years. At the moment the camera was moving smoothly, a pale, forked lightning bolt tore through the sky without warning, dragging the entire world from extreme darkness into dazzling daylight in less than a second, instantly illuminating the blurred inscriptions on the stone tablets that had been worn down by rain and time to the point of being almost unrecognizable; in this brief moment of light, ancient runes flashed by, and then the light quickly faded, and the picture was once again swallowed up by the deep darkness and the sound of continuous heavy rain, leaving only the fleeting sense of mystery and the heavy oppression of history.", + "seed": 29367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_8/index_402_seed_29367_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014425215031951665, + "vlm_logprob": -0.014425215031951665, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.014425215031951665, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004957288969308138 + ], + "opposite_logprob_tokens": [ + -6.62995719909668 + ], + "logprob_reward": -0.004957288969308138, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.004957288969308138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06724380701780319, + 0.02984810806810856, + 0.03776497393846512, + 0.04870979115366936, + 0.1875140368938446, + 0.17746186256408691, + 0.05564204230904579, + 0.11048182100057602, + 0.17125025391578674, + 0.11408329010009766 + ], + "credit_map": [ + 0.007116205058991909, + 0.01082952693104744, + 0.013245166279375553, + 0.0066809942945837975, + 0.006993603426963091, + 0.004064067732542753, + 0.01065716240555048, + 0.00765709113329649, + 0.004088056739419699, + 0.004347062204033136, + 0.003861432196572423, + 0.004343455657362938, + 0.003821206744760275, + 0.0016965734539553523, + 0.004917101934552193, + 0.002773221815004945, + 0.005839648190885782, + 0.00378432753495872, + 0.004482480697333813, + 0.007270549889653921, + 0.0019973544403910637, + 0.005082791671156883, + 0.004508760757744312, + 0.004799062851816416, + 0.006116403266787529, + 0.003473087679594755, + 0.0046165441162884235, + 0.007315296679735184, + 0.004450247157365084, + 0.007942255586385727, + 0.006608880124986172, + 0.00818707887083292, + 0.0177084319293499, + 0.009020406752824783, + 0.0388559065759182, + 0.017756907269358635, + 0.01838827133178711, + 0.035030148923397064, + 0.027436360716819763, + 0.023317620158195496, + 0.015558826737105846, + 0.016884319484233856, + 0.035502251237630844, + 0.019866572692990303, + 0.018979070708155632, + 0.023718485608696938, + 0.026131901890039444, + 0.020820438861846924, + 0.004313244484364986, + 0.00445195147767663, + 0.006038557272404432, + 0.004327718168497086, + 0.009046104736626148, + 0.009954526089131832, + 0.009506205096840858, + 0.008003736846148968, + 0.008730648085474968, + 0.0036204340867698193, + 0.029455209150910378, + 0.005196578800678253, + 0.013592269271612167, + 0.024951566010713577, + 0.01353534683585167, + 0.011399777606129646, + 0.012024953030049801, + 0.00653468444943428, + 0.0164530947804451, + 0.01130148395895958, + 0.025534790009260178, + 0.022188274189829826, + 0.05444975569844246, + 0.02276322804391384, + 0.013623975217342377, + 0.005641288589686155, + 0.01312862616032362, + 0.0060951365157961845, + 0.03203645348548889, + 0.014641755260527134, + 0.016986940056085587, + 0.011929125525057316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.004957288969308138 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370270878076553 + ], + "opposite_logprob_tokens": [ + -3.4087026119232178 + ], + "logprob_reward": -0.03370270878076553, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370270878076553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07225187867879868, + 0.04386015236377716, + 0.12543503940105438, + 0.15389031171798706, + 0.35168492794036865, + 0.12447281181812286, + 0.03605568781495094, + 0.03764212504029274, + 0.028763914480805397, + 0.025943094864487648 + ], + "credit_map": [ + 0.006860048975795507, + 0.010183569975197315, + 0.016072804108262062, + 0.007935340516269207, + 0.007149775046855211, + 0.005207720212638378, + 0.011479753069579601, + 0.0073628611862659454, + 0.005485419649630785, + 0.007223127409815788, + 0.005465247202664614, + 0.006227313540875912, + 0.006518526468425989, + 0.0027251930441707373, + 0.005143699701875448, + 0.005071624182164669, + 0.0186807569116354, + 0.01037697121500969, + 0.017154498025774956, + 0.019255440682172775, + 0.00949054304510355, + 0.02367117628455162, + 0.014972374774515629, + 0.011833279393613338, + 0.020914118736982346, + 0.008391031064093113, + 0.015228742733597755, + 0.018517710268497467, + 0.009374458342790604, + 0.02230210043489933, + 0.020806701853871346, + 0.03835545852780342, + 0.03540903702378273, + 0.016731640323996544, + 0.07813607156276703, + 0.027731789276003838, + 0.03167250007390976, + 0.06465509533882141, + 0.05215081200003624, + 0.0451979897916317, + 0.009702366776764393, + 0.01027057133615017, + 0.02112736366689205, + 0.013040227815508842, + 0.01292400062084198, + 0.0204269178211689, + 0.01695547066628933, + 0.02002590522170067, + 0.003296722425147891, + 0.0027945293113589287, + 0.004953543655574322, + 0.003140268614515662, + 0.005004114005714655, + 0.0067375958897173405, + 0.004663646221160889, + 0.005465273279696703, + 0.003309524618089199, + 0.001471388153731823, + 0.010734012350440025, + 0.0019894633442163467, + 0.003733422141522169, + 0.008507121354341507, + 0.004510069731622934, + 0.0033871240448206663, + 0.0020553271751850843, + 0.0017376696923747659, + 0.003112866310402751, + 0.0024112900719046593, + 0.003645665245130658, + 0.0037763267755508423, + 0.00856114737689495, + 0.0034636242780834436, + 0.003024290082976222, + 0.0017564980080351233, + 0.0035354008432477713, + 0.0015620049089193344, + 0.006129023618996143, + 0.003392973216250539, + 0.004336804151535034, + 0.002206101082265377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.03370270878076553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258300807327032 + ], + "opposite_logprob_tokens": [ + -5.255258083343506 + ], + "logprob_reward": -0.005258300807327032, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258300807327032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.128680020570755, + 0.06355736404657364, + 0.1642594337463379, + 0.04737357050180435, + 0.06245966628193855, + 0.10316368192434311, + 0.05827375128865242, + 0.04771598428487778, + 0.20666073262691498, + 0.1178557351231575 + ], + "credit_map": [ + 0.010193347930908203, + 0.027292899787425995, + 0.020304007455706596, + 0.014043078757822514, + 0.014865323901176453, + 0.008421129547059536, + 0.017719736322760582, + 0.015840496867895126, + 0.005670088808983564, + 0.01540808193385601, + 0.010491060093045235, + 0.00808258168399334, + 0.005783688742667437, + 0.0045891087502241135, + 0.008727779611945152, + 0.0048049758188426495, + 0.026461219415068626, + 0.013517637737095356, + 0.022527050226926804, + 0.024443918839097023, + 0.010722266510128975, + 0.03147418051958084, + 0.019474726170301437, + 0.015638435259461403, + 0.004411645233631134, + 0.002945284591987729, + 0.005324586294591427, + 0.00951144378632307, + 0.0038570405449718237, + 0.007042510900646448, + 0.007492974400520325, + 0.006788086146116257, + 0.0027822076808661222, + 0.0034678159281611443, + 0.0037084585055708885, + 0.006926799193024635, + 0.00892599206417799, + 0.014985156245529652, + 0.01379774883389473, + 0.007865493185818195, + 0.005025594960898161, + 0.007893066853284836, + 0.00891790073364973, + 0.012471862137317657, + 0.0138295479118824, + 0.020049292594194412, + 0.015290163457393646, + 0.019686268642544746, + 0.004319915547966957, + 0.006966733373701572, + 0.0058713145554065704, + 0.007271853741258383, + 0.009585532359778881, + 0.007521829567849636, + 0.008822919800877571, + 0.00791365746408701, + 0.004942175466567278, + 0.002572887111455202, + 0.009563514031469822, + 0.0037868788931518793, + 0.006224074866622686, + 0.006094672717154026, + 0.009962811134755611, + 0.004568978678435087, + 0.015682358294725418, + 0.009826021268963814, + 0.022563382983207703, + 0.016004454344511032, + 0.029092131182551384, + 0.029866473749279976, + 0.04846266284584999, + 0.03516325354576111, + 0.014546578750014305, + 0.0063078440725803375, + 0.014405507594347, + 0.007553884759545326, + 0.03415749594569206, + 0.015123918652534485, + 0.014768279157578945, + 0.010992231778800488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.005258300807327032 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577052608132362 + ], + "opposite_logprob_tokens": [ + -3.8955769538879395 + ], + "logprob_reward": -0.020577052608132362, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577052608132362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07944338023662567, + 0.037701498717069626, + 0.05657641962170601, + 0.08039271086454391, + 0.22539418935775757, + 0.14739510416984558, + 0.054454293102025986, + 0.08451661467552185, + 0.08253034204244614, + 0.15159545838832855 + ], + "credit_map": [ + 0.0071486420929431915, + 0.009643209166824818, + 0.018142055720090866, + 0.009196002036333084, + 0.008463614620268345, + 0.005157953593879938, + 0.011701071634888649, + 0.009990832768380642, + 0.004396719858050346, + 0.005505860783159733, + 0.005028941668570042, + 0.005226516630500555, + 0.005892685614526272, + 0.0018478387501090765, + 0.005098253954201937, + 0.00470467796549201, + 0.008293157443404198, + 0.005025404971092939, + 0.0077910213731229305, + 0.008619251661002636, + 0.0043222662061452866, + 0.009464548900723457, + 0.0074479831382632256, + 0.005612790118902922, + 0.010602448135614395, + 0.005161828361451626, + 0.006621579639613628, + 0.012818634510040283, + 0.008113809861242771, + 0.013890475034713745, + 0.010998046025633812, + 0.012185891158878803, + 0.019129449501633644, + 0.00933280773460865, + 0.04508401080965996, + 0.022886523976922035, + 0.024268832057714462, + 0.0436578243970871, + 0.03771737590432167, + 0.0233173705637455, + 0.014075900427997112, + 0.013929443433880806, + 0.030741417780518532, + 0.01467928383499384, + 0.018724815919995308, + 0.019685540348291397, + 0.01893816888332367, + 0.01662052795290947, + 0.0052100555039942265, + 0.005405620206147432, + 0.008676208555698395, + 0.0046921661123633385, + 0.008113052695989609, + 0.00827916618436575, + 0.008141439408063889, + 0.005936584901064634, + 0.01078319363296032, + 0.004723692778497934, + 0.01624874584376812, + 0.005819244310259819, + 0.0118695218116045, + 0.013789120130240917, + 0.00893975980579853, + 0.012343336828052998, + 0.005672216881066561, + 0.0056561073288321495, + 0.009500792250037193, + 0.006533455103635788, + 0.014982469379901886, + 0.012009155005216599, + 0.016946088522672653, + 0.011230056174099445, + 0.01631977967917919, + 0.008690724149346352, + 0.01959773153066635, + 0.014745848253369331, + 0.025320986285805702, + 0.021443163976073265, + 0.034074801951646805, + 0.011402420699596405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.020577052608132362 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient stone monument with weathered surfaces, cracks, and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a raging thunderstorm with thick storm clouds, torrential rain, and an inky night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video rendered in a stark, high-contrast black-and-white film texture with visible rough grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a slow, arcing tracking shot from a low angle that rotates around the monument?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a pale, forked lightning bolt flash through the sky, briefly illuminating the monument and inscriptions before fading back into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630723994225264 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630723994225264, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630723994225264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08367466926574707, + 0.03383327275514603, + 0.03029126115143299, + 0.09104517847299576, + 0.10853058844804764, + 0.17251330614089966, + 0.0792136862874031, + 0.10255924612283707, + 0.15911933779716492, + 0.1392194628715515 + ], + "credit_map": [ + 0.009553764015436172, + 0.015503247268497944, + 0.011948646046221256, + 0.009671857580542564, + 0.010197644121944904, + 0.007231202907860279, + 0.01055203564465046, + 0.009016264230012894, + 0.002937674755230546, + 0.005136349704116583, + 0.004657484125345945, + 0.006916607730090618, + 0.004092436749488115, + 0.00312390411272645, + 0.0043007913045585155, + 0.002668023342266679, + 0.0032865204848349094, + 0.002372246002778411, + 0.0037856092676520348, + 0.006483931560069323, + 0.002489614998921752, + 0.004518069792538881, + 0.003765677334740758, + 0.0035895926412194967, + 0.009413711726665497, + 0.005025174468755722, + 0.008158163167536259, + 0.014386473223567009, + 0.005324510857462883, + 0.01581914909183979, + 0.012733391486108303, + 0.020184602588415146, + 0.012741431593894958, + 0.009357449598610401, + 0.025366639718413353, + 0.01191666815429926, + 0.011941957287490368, + 0.011613751761615276, + 0.010736696422100067, + 0.014855993911623955, + 0.011055029928684235, + 0.013915118761360645, + 0.019095933064818382, + 0.019343189895153046, + 0.019449427723884583, + 0.0334840789437294, + 0.025289317592978477, + 0.030881201848387718, + 0.006424659863114357, + 0.0065451133996248245, + 0.010160133242607117, + 0.008115649223327637, + 0.014624636620283127, + 0.012034047394990921, + 0.011010522954165936, + 0.01029891986399889, + 0.008463645353913307, + 0.0035448772832751274, + 0.02657109871506691, + 0.005408948287367821, + 0.010803962126374245, + 0.022549044340848923, + 0.015018385834991932, + 0.010199285112321377, + 0.011068174615502357, + 0.007505609188228846, + 0.01686331257224083, + 0.01052013784646988, + 0.023557865992188454, + 0.021797098219394684, + 0.04353538528084755, + 0.024271761998534203, + 0.017299199476838112, + 0.008051415905356407, + 0.020519176498055458, + 0.009470592252910137, + 0.037318576127290726, + 0.018206577748060226, + 0.0153829799965024, + 0.012970944866538048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.007630723994225264 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_0/index_tensor([4889, 4889])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_0/index_tensor([4889, 4889])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dd844f6b8e5b6e267d02d83f31e8adacb1a6124d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_0/index_tensor([4889, 4889])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the style of photorealistic CGI cinematic realism, a steady medium shot follows a massive landing ship dock, its powerful hull painted a non-reflective navy gray and streaked with rust, as it plunges into the tsunami. The ship struggles in the fierce, turbulent sea, pitching and rolling violently amidst the chaotic, deep blue waves, crashing onto its port side and sending a massive spray of white water cascading down onto the deck. On the horizon, a vast, dark, roiling wall of water rises at a terrifying speed, a liquid mountain whose sheer surface is a vortex of shadows and movement. Above, an ominous, scarred sky seethes with heavy storm clouds of ashen gray and deep purple, pierced only by a cold, steely glow. This bleak lighting captures the foaming, wind-ripped crests of the approaching waves, casting a sinister glow as the naval vessel pushes forward, its structure groaning under immense pressure, creating an atmosphere of tension and despair, of impending, overwhelming disaster.", + "seed": 4284, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_0/index_4889_seed_4284_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04657363146543503, + "vlm_logprob": -0.04657363146543503, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.04657363146543503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10424253344535828 + ], + "opposite_logprob_tokens": [ + -2.3542425632476807 + ], + "logprob_reward": -0.10424253344535828, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10424253344535828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14510835707187653, + 0.06162649393081665, + 0.03480031341314316, + 0.06234847009181976, + 0.08546879142522812, + 0.09555249661207199, + 0.31682267785072327, + 0.033217959105968475, + 0.11465201526880264, + 0.05040237307548523 + ], + "credit_map": [ + 0.015728948637843132, + 0.018040252849459648, + 0.019913287833333015, + 0.01893608272075653, + 0.01761910691857338, + 0.015914063900709152, + 0.01725914515554905, + 0.02169748954474926, + 0.010835286229848862, + 0.00960920937359333, + 0.006630594376474619, + 0.008108437992632389, + 0.008952354080975056, + 0.004913498647511005, + 0.007577209733426571, + 0.004999913275241852, + 0.0035395740997046232, + 0.005644906312227249, + 0.003258543089032173, + 0.0033835049252957106, + 0.004407222848385572, + 0.0056283315643668175, + 0.004507918376475573, + 0.004430316388607025, + 0.00807634461671114, + 0.010112171992659569, + 0.007096279878169298, + 0.008996930904686451, + 0.006990443915128708, + 0.006757375318557024, + 0.00727162417024374, + 0.00704730162397027, + 0.01448110956698656, + 0.011982005089521408, + 0.005682369228452444, + 0.009089666418731213, + 0.009286674670875072, + 0.007536693010479212, + 0.010955671779811382, + 0.016454612836241722, + 0.011016094125807285, + 0.014956922270357609, + 0.007680636364966631, + 0.012606151401996613, + 0.017588971182703972, + 0.008813352324068546, + 0.011005948297679424, + 0.011884435079991817, + 0.022564511746168137, + 0.1134968101978302, + 0.014226474799215794, + 0.01810303144156933, + 0.04490910843014717, + 0.021885640919208527, + 0.02300094999372959, + 0.05863618105649948, + 0.0033115148544311523, + 0.005757817067205906, + 0.0034234272316098213, + 0.0028912632260471582, + 0.004789308644831181, + 0.004119141958653927, + 0.004484756384044886, + 0.004440730903297663, + 0.010104252025485039, + 0.01594952680170536, + 0.007662682794034481, + 0.012925376184284687, + 0.026565901935100555, + 0.020251590758562088, + 0.010729565285146236, + 0.010463141836225986, + 0.005450190976262093, + 0.006874506361782551, + 0.0029951483011245728, + 0.004967887420207262, + 0.013561940751969814, + 0.007585674524307251, + 0.00538030406460166, + 0.0035867264959961176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.10424253344535828 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496637150645256 + ], + "opposite_logprob_tokens": [ + -2.929966449737549 + ], + "logprob_reward": -0.05496637150645256, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496637150645256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1440802663564682, + 0.12896190583705902, + 0.01837511546909809, + 0.05831205099821091, + 0.02682432159781456, + 0.15631833672523499, + 0.295400470495224, + 0.06529974192380905, + 0.05539276823401451, + 0.051035016775131226 + ], + "credit_map": [ + 0.015813421458005905, + 0.022832011803984642, + 0.01077983621507883, + 0.01777610369026661, + 0.01399722509086132, + 0.018984103575348854, + 0.018420878797769547, + 0.025476709008216858, + 0.014639977365732193, + 0.01387731172144413, + 0.013832909986376762, + 0.016421599313616753, + 0.019937120378017426, + 0.014584941789507866, + 0.022635452449321747, + 0.013032608665525913, + 0.0017985620070248842, + 0.0033209535758942366, + 0.0014529064064845443, + 0.001606615143828094, + 0.0027097673155367374, + 0.002241387264803052, + 0.002471379702910781, + 0.00277354521676898, + 0.005169275216758251, + 0.009901919402182102, + 0.0044404976069927216, + 0.005835604853928089, + 0.016786660999059677, + 0.005538456607609987, + 0.005097358953207731, + 0.005542278289794922, + 0.003960825502872467, + 0.004341446794569492, + 0.002329017501324415, + 0.0035342995543032885, + 0.004641438834369183, + 0.0030042894650250673, + 0.0029733795672655106, + 0.002039626007899642, + 0.013459478504955769, + 0.02091483771800995, + 0.01161899883300066, + 0.019740048795938492, + 0.024307643994688988, + 0.017568601295351982, + 0.031279485672712326, + 0.01742924191057682, + 0.021793855354189873, + 0.10416223853826523, + 0.013651326298713684, + 0.020682156085968018, + 0.04923323914408684, + 0.0191299207508564, + 0.025031650438904762, + 0.04171610623598099, + 0.009264755062758923, + 0.007387964520603418, + 0.008751899935305119, + 0.006042138207703829, + 0.007774170022457838, + 0.00787125714123249, + 0.009788415394723415, + 0.008419142104685307, + 0.01073293387889862, + 0.0078171631321311, + 0.004768904764205217, + 0.005781358107924461, + 0.010366398841142654, + 0.006529645062983036, + 0.003991199657320976, + 0.005405170377343893, + 0.006649183575063944, + 0.007483148016035557, + 0.004546365235000849, + 0.005906226113438606, + 0.009075313806533813, + 0.006691176909953356, + 0.0052315592765808105, + 0.005452049896121025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.05496637150645256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981266938149929 + ], + "opposite_logprob_tokens": [ + -3.5298125743865967 + ], + "logprob_reward": -0.02981266938149929, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981266938149929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13703453540802002, + 0.0653824731707573, + 0.02941087633371353, + 0.038600340485572815, + 0.052662912756204605, + 0.08534979075193405, + 0.37640923261642456, + 0.1495657116174698, + 0.03921763226389885, + 0.02636650763452053 + ], + "credit_map": [ + 0.012174723669886589, + 0.017354816198349, + 0.015233620069921017, + 0.013741997070610523, + 0.01587623916566372, + 0.019426271319389343, + 0.016069183126091957, + 0.027157671749591827, + 0.008513069711625576, + 0.007131713442504406, + 0.005212478339672089, + 0.009221071377396584, + 0.008999117650091648, + 0.0064399996772408485, + 0.012672160752117634, + 0.007192861754447222, + 0.0031406066846102476, + 0.0036876299418509007, + 0.0029402333311736584, + 0.00343603384681046, + 0.0037138215266168118, + 0.005237825680524111, + 0.0033280018251389265, + 0.003926721401512623, + 0.00383313512429595, + 0.005802758038043976, + 0.004616287536919117, + 0.004818453453481197, + 0.0070807854644954205, + 0.0035424011293798685, + 0.004234022460877895, + 0.004672495182603598, + 0.006655792705714703, + 0.008093508891761303, + 0.004495765548199415, + 0.007494943682104349, + 0.007595832925289869, + 0.005603510420769453, + 0.006502047181129456, + 0.0062215132638812065, + 0.008851398713886738, + 0.009532167576253414, + 0.00553928641602397, + 0.007744595408439636, + 0.01075004693120718, + 0.008555829524993896, + 0.018602436408400536, + 0.015774020925164223, + 0.030196906998753548, + 0.13067834079265594, + 0.014130118303000927, + 0.02418636903166771, + 0.04385436698794365, + 0.0270798709243536, + 0.03099842742085457, + 0.07528480887413025, + 0.0173594132065773, + 0.027241775766015053, + 0.016298538073897362, + 0.01008942537009716, + 0.01935555599629879, + 0.02437754161655903, + 0.014777814038097858, + 0.020065637305378914, + 0.005300039425492287, + 0.005280793644487858, + 0.0029855873435735703, + 0.0040888735093176365, + 0.00952846184372902, + 0.0055739800445735455, + 0.0034887047950178385, + 0.0029711874667555094, + 0.003378375666216016, + 0.0037115535233169794, + 0.0017114118672907352, + 0.002867489354684949, + 0.006507370620965958, + 0.0034097174648195505, + 0.0026621536817401648, + 0.0021184312645345926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.02981266938149929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327875606715679 + ], + "opposite_logprob_tokens": [ + -3.7732787132263184 + ], + "logprob_reward": -0.02327875606715679, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327875606715679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30996155738830566, + 0.09596662223339081, + 0.03534412384033203, + 0.08043274283409119, + 0.11033736914396286, + 0.09483188390731812, + 0.1244087964296341, + 0.021342206746339798, + 0.1008111983537674, + 0.026563581079244614 + ], + "credit_map": [ + 0.027170564979314804, + 0.040131956338882446, + 0.026678144931793213, + 0.03391500189900398, + 0.0488821379840374, + 0.04469946026802063, + 0.03866579383611679, + 0.04981844499707222, + 0.012623775750398636, + 0.009889937937259674, + 0.00983668677508831, + 0.013944508507847786, + 0.017309192568063736, + 0.009098521433770657, + 0.013956637121737003, + 0.009307356551289558, + 0.003292817622423172, + 0.005979899782687426, + 0.0026908672880381346, + 0.0037347106263041496, + 0.004144996404647827, + 0.0058695971965789795, + 0.005436020437628031, + 0.004195208661258221, + 0.009174286387860775, + 0.014230797998607159, + 0.006182951387017965, + 0.009306156076490879, + 0.009126773104071617, + 0.009056618437170982, + 0.01429519522935152, + 0.009059958159923553, + 0.013204232789576054, + 0.019499674439430237, + 0.007851671427488327, + 0.013373248279094696, + 0.01586160436272621, + 0.010559819638729095, + 0.013260490261018276, + 0.016726622357964516, + 0.008575798943638802, + 0.009660121984779835, + 0.00930351298302412, + 0.009930537082254887, + 0.012778494507074356, + 0.011967972852289677, + 0.015778440982103348, + 0.016836998984217644, + 0.00870975200086832, + 0.013488423079252243, + 0.007731861900538206, + 0.008151279762387276, + 0.04323608800768852, + 0.013874265365302563, + 0.009343537501990795, + 0.01987357996404171, + 0.0019358632853254676, + 0.002943602157756686, + 0.0019799505826085806, + 0.001995163969695568, + 0.0030815855134278536, + 0.0027000135742127895, + 0.0034715915098786354, + 0.003234432078897953, + 0.00968003086745739, + 0.015146507881581783, + 0.0072326334193348885, + 0.010759231634438038, + 0.020542429760098457, + 0.01599854603409767, + 0.00927061215043068, + 0.012181198224425316, + 0.0030934163369238377, + 0.003524234052747488, + 0.0020012406166642904, + 0.0031976874452084303, + 0.006783498451113701, + 0.0028698178939521313, + 0.0029156673699617386, + 0.002178015885874629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.02327875606715679 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02056782692670822 + ], + "opposite_logprob_tokens": [ + -3.8955678939819336 + ], + "logprob_reward": -0.02056782692670822, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02056782692670822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09920583665370941, + 0.15554526448249817, + 0.06523767858743668, + 0.2547979950904846, + 0.057708874344825745, + 0.15550561249256134, + 0.1174103319644928, + 0.03078177198767662, + 0.03522445261478424, + 0.028582165017724037 + ], + "credit_map": [ + 0.0075396704487502575, + 0.012090072967112064, + 0.011678780429065228, + 0.013607078231871128, + 0.009282056242227554, + 0.014150760136544704, + 0.013792935758829117, + 0.01706448756158352, + 0.026508569717407227, + 0.03225993365049362, + 0.010306790471076965, + 0.014750362373888493, + 0.02702333591878414, + 0.018029386177659035, + 0.013819223269820213, + 0.012847693637013435, + 0.006591869052499533, + 0.009381677024066448, + 0.006374870892614126, + 0.006176808848977089, + 0.008560958318412304, + 0.011225773952901363, + 0.009279808960855007, + 0.007645911071449518, + 0.026293961331248283, + 0.05084981396794319, + 0.015875162556767464, + 0.020136211067438126, + 0.07475002110004425, + 0.021903060376644135, + 0.024181412532925606, + 0.020808372646570206, + 0.008876536972820759, + 0.008752239868044853, + 0.0027438122779130936, + 0.005935016553848982, + 0.007367318496108055, + 0.005283972714096308, + 0.006979641504585743, + 0.011770342476665974, + 0.018859419971704483, + 0.024910183623433113, + 0.011062155477702618, + 0.017470089718699455, + 0.030386878177523613, + 0.01348300650715828, + 0.019269367679953575, + 0.02006452903151512, + 0.008027354255318642, + 0.016540173441171646, + 0.007163926027715206, + 0.007416976615786552, + 0.04298559948801994, + 0.013957645744085312, + 0.007017566356807947, + 0.014301101677119732, + 0.002886319300159812, + 0.003919763956218958, + 0.002505664248019457, + 0.0027540058363229036, + 0.0073217120952904224, + 0.004068087320774794, + 0.0035924988333135843, + 0.003733724122866988, + 0.003371273400261998, + 0.004626895766705275, + 0.0037390077486634254, + 0.005979085806757212, + 0.006755260284990072, + 0.0039028325118124485, + 0.003380289301276207, + 0.0034698094241321087, + 0.0027905451133847237, + 0.0043495893478393555, + 0.0013904297957196832, + 0.0026311958208680153, + 0.008272329345345497, + 0.00425222422927618, + 0.0029100680258125067, + 0.001985785784199834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.02056782692670822 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of photorealistic CGI cinematic realism, a steady medium shot follows a massive landing ship dock, its powerful hull painted a non-reflective navy gray and streaked with rust, as it plunges into the tsunami. The ship struggles in the fierce, turbulent sea, pitching and rolling violently amidst the chaotic, deep blue waves, crashing onto its port side and sending a massive spray of white water cascading down onto the deck. On the horizon, a vast, dark, roiling wall of water rises at a terrifying speed, a liquid mountain whose sheer surface is a vortex of shadows and movement. Above, an ominous, scarred sky seethes with heavy storm clouds of ashen gray and deep purple, pierced only by a cold, steely glow. This bleak lighting captures the foaming, wind-ripped crests of the approaching waves, casting a sinister glow as the naval vessel pushes forward, its structure groaning under immense pressure, creating an atmosphere of tension and despair, of impending, overwhelming disaster.", + "seed": 4285, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_0/index_4889_seed_4285_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1353078931570053, + "vlm_logprob": -0.1353078931570053, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.1353078931570053, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5777320265769958 + ], + "opposite_logprob_tokens": [ + -0.8277320265769958 + ], + "logprob_reward": -0.5777320265769958, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5777320265769958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1118997260928154, + 0.10771884769201279, + 0.05352143198251724, + 0.5541170835494995, + 0.022196175530552864, + 0.030064716935157776, + 0.020929893478751183, + 0.05862940847873688, + 0.021516626700758934, + 0.01940610259771347 + ], + "credit_map": [ + 0.016562502831220627, + 0.016641806811094284, + 0.014468087814748287, + 0.015287656337022781, + 0.013334623537957668, + 0.00843651033937931, + 0.011690405197441578, + 0.015478130429983139, + 0.009249813854694366, + 0.013362090103328228, + 0.01633254997432232, + 0.012443403713405132, + 0.015712251886725426, + 0.010425560176372528, + 0.016515616327524185, + 0.013677559792995453, + 0.006828657817095518, + 0.004302124492824078, + 0.0091047128662467, + 0.014185148291289806, + 0.0036193332634866238, + 0.005152103956788778, + 0.0034879830200225115, + 0.006841369438916445, + 0.1551307737827301, + 0.03439867123961449, + 0.0294658076018095, + 0.04274113103747368, + 0.047048624604940414, + 0.1756129264831543, + 0.017609450966119766, + 0.052109718322753906, + 0.0028960146009922028, + 0.0025094004813581705, + 0.001982351066544652, + 0.0019145276164636016, + 0.006788919679820538, + 0.0018960874294862151, + 0.0020514370407909155, + 0.0021574383135885, + 0.002733650617301464, + 0.0032223586458712816, + 0.002784950425848365, + 0.004655672237277031, + 0.005171767435967922, + 0.0016624957788735628, + 0.007201903499662876, + 0.0026319159660488367, + 0.0023298279847949743, + 0.0023894933983683586, + 0.003191945841535926, + 0.003630089107900858, + 0.004257793538272381, + 0.0016269603511318564, + 0.0019157824572175741, + 0.0015880003338679671, + 0.0070525347255170345, + 0.006424630992114544, + 0.008909560739994049, + 0.009930978529155254, + 0.00715799443423748, + 0.005779390223324299, + 0.006667370442301035, + 0.006706944666802883, + 0.0032733893021941185, + 0.0030151554383337498, + 0.0022253142669796944, + 0.0028985408134758472, + 0.003661909606307745, + 0.002469416242092848, + 0.0023479675874114037, + 0.0016249327454715967, + 0.002161468379199505, + 0.002857644110918045, + 0.0024693142622709274, + 0.0025756454560905695, + 0.0029363767243921757, + 0.001321336254477501, + 0.003347816178575158, + 0.0017364997183904052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.5777320265769958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300884291529655 + ], + "opposite_logprob_tokens": [ + -3.773300886154175 + ], + "logprob_reward": -0.023300884291529655, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300884291529655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13045121729373932, + 0.20538069307804108, + 0.16483406722545624, + 0.250580757856369, + 0.02061823382973671, + 0.01556921936571598, + 0.04234439879655838, + 0.12434656172990799, + 0.013537833467125893, + 0.0323369987308979 + ], + "credit_map": [ + 0.031561899930238724, + 0.013752562925219536, + 0.019699016585946083, + 0.015110267326235771, + 0.01144452765583992, + 0.011989498510956764, + 0.0135323042050004, + 0.013361147604882717, + 0.01554480753839016, + 0.010503475554287434, + 0.0681803748011589, + 0.05822893977165222, + 0.009432448074221611, + 0.022557616233825684, + 0.009809192270040512, + 0.011123842559754848, + 0.020717516541481018, + 0.016612332314252853, + 0.029078327119350433, + 0.04516351595520973, + 0.013029484078288078, + 0.014225970022380352, + 0.011940857395529747, + 0.014066062867641449, + 0.035510919988155365, + 0.02011205069720745, + 0.027223721146583557, + 0.02915678545832634, + 0.022537538781762123, + 0.05922234430909157, + 0.014390461146831512, + 0.0424269363284111, + 0.002392911585047841, + 0.0024940334260463715, + 0.0018193612340837717, + 0.0029623652808368206, + 0.006046508904546499, + 0.0015377121744677424, + 0.0010709648486226797, + 0.002294376492500305, + 0.0020738651510328054, + 0.0017680588643997908, + 0.0015176244778558612, + 0.0033900372218340635, + 0.0024198899045586586, + 0.0014742467319592834, + 0.001641764771193266, + 0.0012837323592975736, + 0.005350225139409304, + 0.003893445013090968, + 0.005448810290545225, + 0.004020979627966881, + 0.015148937702178955, + 0.002285369671881199, + 0.003922536503523588, + 0.002274093683809042, + 0.016206983476877213, + 0.016125546768307686, + 0.014445440843701363, + 0.02234671451151371, + 0.014861810952425003, + 0.010190408676862717, + 0.01309395395219326, + 0.01707569696009159, + 0.0019808479119092226, + 0.0021741085220128298, + 0.0015636620810255408, + 0.0016057834727689624, + 0.002008771989494562, + 0.001369541510939598, + 0.0015897434204816818, + 0.0012453739764168859, + 0.004349233582615852, + 0.005600269418209791, + 0.005257558077573776, + 0.004476645961403847, + 0.003942776937037706, + 0.002108733868226409, + 0.003374119522050023, + 0.0032276627607643604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.023300884291529655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809913620352745 + ], + "opposite_logprob_tokens": [ + -3.2880990505218506 + ], + "logprob_reward": -0.03809913620352745, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809913620352745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17237338423728943, + 0.2662191092967987, + 0.1028401106595993, + 0.2107308954000473, + 0.05001005157828331, + 0.04001428559422493, + 0.029198000207543373, + 0.05384981632232666, + 0.025922490283846855, + 0.04884181544184685 + ], + "credit_map": [ + 0.04687930643558502, + 0.024018311873078346, + 0.024846559390425682, + 0.017596155405044556, + 0.016034262254834175, + 0.011786389164626598, + 0.019931262359023094, + 0.011281134560704231, + 0.019012246280908585, + 0.014033298008143902, + 0.0881769135594368, + 0.07145164161920547, + 0.015758289024233818, + 0.03221551328897476, + 0.013208873569965363, + 0.012362323701381683, + 0.012157348915934563, + 0.009979482740163803, + 0.016579948365688324, + 0.028672242537140846, + 0.006522076204419136, + 0.010967150330543518, + 0.007823978550732136, + 0.010137888602912426, + 0.05234096944332123, + 0.014796212315559387, + 0.014839965850114822, + 0.017225027084350586, + 0.018010355532169342, + 0.06275734305381775, + 0.008227014914155006, + 0.022534022107720375, + 0.004585572052747011, + 0.005739220883697271, + 0.0076700374484062195, + 0.008671619929373264, + 0.009714105166494846, + 0.0031252503395080566, + 0.004360137041658163, + 0.006144106388092041, + 0.005771723110228777, + 0.004936761222779751, + 0.0032810457050800323, + 0.0036991958040744066, + 0.010251261293888092, + 0.0026096717920154333, + 0.0062332418747246265, + 0.0032313831616193056, + 0.0032023543026298285, + 0.002379827667027712, + 0.0053213732317090034, + 0.0023627891205251217, + 0.010930174961686134, + 0.001621165662072599, + 0.001988309435546398, + 0.0013920057099312544, + 0.006987882778048515, + 0.0055774725042283535, + 0.006554929539561272, + 0.008247237652540207, + 0.0064840903505682945, + 0.004258702974766493, + 0.008557901717722416, + 0.007181602995842695, + 0.004095432348549366, + 0.003426152979955077, + 0.002727722516283393, + 0.003729886142536998, + 0.0034935108851641417, + 0.00264738779515028, + 0.0033983413595706224, + 0.0024040555581450462, + 0.005645482800900936, + 0.007450111676007509, + 0.0059254365041852, + 0.008393245749175549, + 0.005984061397612095, + 0.003225996857509017, + 0.006800761446356773, + 0.005416723899543285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.03809913620352745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978675067424774 + ], + "opposite_logprob_tokens": [ + -3.5297868251800537 + ], + "logprob_reward": -0.02978675067424774, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978675067424774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2180541604757309, + 0.06696302443742752, + 0.1504267156124115, + 0.3583574891090393, + 0.06742538511753082, + 0.04741871729493141, + 0.010424474254250526, + 0.028136923909187317, + 0.025386150926351547, + 0.027407007291913033 + ], + "credit_map": [ + 0.05793851613998413, + 0.026490362361073494, + 0.031222468242049217, + 0.019000371918082237, + 0.021398942917585373, + 0.0240354984998703, + 0.022108739241957664, + 0.015859253704547882, + 0.006017332896590233, + 0.0055060600861907005, + 0.019179299473762512, + 0.015031836926937103, + 0.004473254084587097, + 0.007956000044941902, + 0.004816572181880474, + 0.0039826650172472, + 0.018279002979397774, + 0.014490983448922634, + 0.02406957931816578, + 0.04722720384597778, + 0.009516417980194092, + 0.015778910368680954, + 0.009385161101818085, + 0.011679447256028652, + 0.09264631569385529, + 0.022576352581381798, + 0.021192288026213646, + 0.02920140139758587, + 0.031139248982071877, + 0.11183235794305801, + 0.012378290295600891, + 0.03739123418927193, + 0.006952631287276745, + 0.008728238753974438, + 0.004182217642664909, + 0.00801441166549921, + 0.021925456821918488, + 0.006181949749588966, + 0.0037941867485642433, + 0.0076462868601083755, + 0.006384051870554686, + 0.005619578994810581, + 0.003195613157004118, + 0.006356728263199329, + 0.012149226851761341, + 0.0037391381338238716, + 0.006650443188846111, + 0.003323937766253948, + 0.0011854800395667553, + 0.0014985962770879269, + 0.0013250012416392565, + 0.0013017815072089434, + 0.0015361282275989652, + 0.0010551076848059893, + 0.0016217837110161781, + 0.0009005949250422418, + 0.003787682857364416, + 0.002873170655220747, + 0.0035167180467396975, + 0.0037855575792491436, + 0.003449162235483527, + 0.003126468975096941, + 0.0035297025460749865, + 0.004068458452820778, + 0.0047479975037276745, + 0.006082769017666578, + 0.0018431632779538631, + 0.0024531022645533085, + 0.004475089721381664, + 0.0022932833526283503, + 0.001871940097771585, + 0.001618801848962903, + 0.003310843836516142, + 0.0033156771678477526, + 0.003975117579102516, + 0.00369308446533978, + 0.0034644785337150097, + 0.0023309311363846064, + 0.004342311527580023, + 0.0029745621141046286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.02978675067424774 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007620668038725853 + ], + "opposite_logprob_tokens": [ + -4.882620811462402 + ], + "logprob_reward": -0.007620668038725853, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007620668038725853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16115927696228027, + 0.1348661482334137, + 0.05368698388338089, + 0.29968395829200745, + 0.05016390606760979, + 0.04635876417160034, + 0.03702224791049957, + 0.08450824022293091, + 0.05555230751633644, + 0.07699817419052124 + ], + "credit_map": [ + 0.0260649211704731, + 0.01633644662797451, + 0.02003365568816662, + 0.015178213827311993, + 0.01981978863477707, + 0.02572363242506981, + 0.019977742806077003, + 0.018024887889623642, + 0.012186175212264061, + 0.009600892663002014, + 0.03802061080932617, + 0.0318463072180748, + 0.010056092403829098, + 0.0092166131362319, + 0.012606067582964897, + 0.011333384551107883, + 0.0065688565373420715, + 0.006499323528259993, + 0.008001510985195637, + 0.011134025640785694, + 0.004850585479289293, + 0.0067483833990991116, + 0.0037901331670582294, + 0.00609416700899601, + 0.07705354690551758, + 0.018893012776970863, + 0.019006259739398956, + 0.024432823061943054, + 0.0259972233325243, + 0.09327787905931473, + 0.010366261005401611, + 0.030656959861516953, + 0.005027539096772671, + 0.00613735057413578, + 0.0044349548406898975, + 0.006793181411921978, + 0.01304447278380394, + 0.006013651378452778, + 0.0033434557262808084, + 0.005369302351027727, + 0.007330633234232664, + 0.005806977394968271, + 0.00269633112475276, + 0.004957376979291439, + 0.011713490821421146, + 0.004078435245901346, + 0.006733387243002653, + 0.003042127937078476, + 0.004885146394371986, + 0.003974958322942257, + 0.003560770070180297, + 0.004453151486814022, + 0.011434238404035568, + 0.0032102344557642937, + 0.002937489189207554, + 0.0025662556290626526, + 0.011432518251240253, + 0.008254768326878548, + 0.010510742664337158, + 0.01715625450015068, + 0.01067887432873249, + 0.008769244886934757, + 0.010933289304375648, + 0.00677254144102335, + 0.010313291102647781, + 0.012576056644320488, + 0.004065883345901966, + 0.005318822804838419, + 0.010872162878513336, + 0.0052133542485535145, + 0.0039640869945287704, + 0.003228649729862809, + 0.009415186941623688, + 0.009586498141288757, + 0.010748651809990406, + 0.01011864934116602, + 0.00851492676883936, + 0.00819496437907219, + 0.011819443665444851, + 0.00859985314309597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.007620668038725853 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of photorealistic CGI cinematic realism, a steady medium shot follows a massive landing ship dock, its powerful hull painted a non-reflective navy gray and streaked with rust, as it plunges into the tsunami. The ship struggles in the fierce, turbulent sea, pitching and rolling violently amidst the chaotic, deep blue waves, crashing onto its port side and sending a massive spray of white water cascading down onto the deck. On the horizon, a vast, dark, roiling wall of water rises at a terrifying speed, a liquid mountain whose sheer surface is a vortex of shadows and movement. Above, an ominous, scarred sky seethes with heavy storm clouds of ashen gray and deep purple, pierced only by a cold, steely glow. This bleak lighting captures the foaming, wind-ripped crests of the approaching waves, casting a sinister glow as the naval vessel pushes forward, its structure groaning under immense pressure, creating an atmosphere of tension and despair, of impending, overwhelming disaster.", + "seed": 4286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_0/index_4889_seed_4286_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09906914662569762, + "vlm_logprob": -0.09906914662569762, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.09906914662569762, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4311511516571045 + ], + "opposite_logprob_tokens": [ + -1.0561511516571045 + ], + "logprob_reward": -0.4311511516571045, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4311511516571045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1683439463376999, + 0.07693105936050415, + 0.0967489629983902, + 0.26336392760276794, + 0.023551318794488907, + 0.06298430263996124, + 0.16009566187858582, + 0.0685843899846077, + 0.045368779450654984, + 0.03402763232588768 + ], + "credit_map": [ + 0.015063153579831123, + 0.026864370331168175, + 0.017767349258065224, + 0.022529063746333122, + 0.019365280866622925, + 0.018818873912096024, + 0.024178331717848778, + 0.02375752106308937, + 0.00733167864382267, + 0.00900521781295538, + 0.00861932709813118, + 0.013792355544865131, + 0.011803065426647663, + 0.008228808641433716, + 0.010764651000499725, + 0.007385958917438984, + 0.00898692011833191, + 0.007311561610549688, + 0.026788823306560516, + 0.006169116124510765, + 0.015112249180674553, + 0.018827056512236595, + 0.006268872879445553, + 0.007284369319677353, + 0.02176804468035698, + 0.028129596263170242, + 0.012954877689480782, + 0.009238144382834435, + 0.09439890086650848, + 0.05166191607713699, + 0.0253111831843853, + 0.019901277497410774, + 0.0023330592084676027, + 0.0027869853656738997, + 0.0025353494565933943, + 0.0022653909400105476, + 0.005557690281420946, + 0.003667777869850397, + 0.002123187994584441, + 0.0022818762809038162, + 0.005883328150957823, + 0.007549819070845842, + 0.004090431611984968, + 0.006889011710882187, + 0.016941174864768982, + 0.007414680439978838, + 0.009301834739744663, + 0.004914029035717249, + 0.014300430193543434, + 0.01686987839639187, + 0.01691295951604843, + 0.030115440487861633, + 0.018760688602924347, + 0.017337452620267868, + 0.031027095392346382, + 0.014771714806556702, + 0.0052735693752765656, + 0.009779139421880245, + 0.01088367123156786, + 0.008864848874509335, + 0.009606271982192993, + 0.007926232181489468, + 0.008519977331161499, + 0.007730682846158743, + 0.002531174337491393, + 0.004670622292906046, + 0.002007102593779564, + 0.0010997789213433862, + 0.0167442224919796, + 0.013227316550910473, + 0.0023910205345600843, + 0.0026975441724061966, + 0.002751985564827919, + 0.004520960617810488, + 0.003440862288698554, + 0.004391604568809271, + 0.0084726857021451, + 0.004236630629748106, + 0.003155242884531617, + 0.0030576593708246946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.4311511516571045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648086339235306 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008648086339235306, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648086339235306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10990530997514725, + 0.05600251629948616, + 0.09948702156543732, + 0.21545195579528809, + 0.0491257943212986, + 0.12748585641384125, + 0.11018361896276474, + 0.03562253713607788, + 0.11746396869421005, + 0.07927136868238449 + ], + "credit_map": [ + 0.013536698184907436, + 0.021595127880573273, + 0.01171860285103321, + 0.01249067671597004, + 0.014246068894863129, + 0.014871516264975071, + 0.01227402500808239, + 0.009172597900032997, + 0.006941597443073988, + 0.011478536762297153, + 0.005300776567310095, + 0.006307101342827082, + 0.00806252658367157, + 0.00598697317764163, + 0.006928290240466595, + 0.004996722564101219, + 0.007527387700974941, + 0.010072488337755203, + 0.010165759362280369, + 0.010405651293694973, + 0.020487627014517784, + 0.013144588097929955, + 0.010591675527393818, + 0.01709185726940632, + 0.013307587243616581, + 0.017878178507089615, + 0.005537915509194136, + 0.005145274102687836, + 0.09373865276575089, + 0.06693985313177109, + 0.005059349816292524, + 0.007845156826078892, + 0.0056297434493899345, + 0.006886654067784548, + 0.0058812727220356464, + 0.007705449592322111, + 0.00966318417340517, + 0.005571622401475906, + 0.004521518014371395, + 0.0032663531601428986, + 0.013559885323047638, + 0.016150210052728653, + 0.009061193093657494, + 0.017699137330055237, + 0.027802973985671997, + 0.015086796134710312, + 0.01614856719970703, + 0.011977104470133781, + 0.012417130172252655, + 0.01378659438341856, + 0.013574613258242607, + 0.01744428090751171, + 0.00958977546542883, + 0.011231647804379463, + 0.023426661267876625, + 0.008712922222912312, + 0.002250846242532134, + 0.004715554416179657, + 0.003864822443574667, + 0.002696473151445389, + 0.010025909170508385, + 0.0073097264394164085, + 0.002141200006008148, + 0.0026180073618888855, + 0.007574909832328558, + 0.011326340027153492, + 0.002689318498596549, + 0.0025569009594619274, + 0.04502193257212639, + 0.03562949225306511, + 0.006481369026005268, + 0.006183713208884001, + 0.00712957140058279, + 0.013585447333753109, + 0.006311257369816303, + 0.009044699370861053, + 0.01884601265192032, + 0.008282137103378773, + 0.008373432792723179, + 0.007698822300881147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.008648086339235306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634642831981182 + ], + "opposite_logprob_tokens": [ + -3.651346445083618 + ], + "logprob_reward": -0.02634642831981182, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634642831981182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10361828655004501, + 0.06213273108005524, + 0.05491179972887039, + 0.39900562167167664, + 0.037981484085321426, + 0.07541745156049728, + 0.10665826499462128, + 0.023722415789961815, + 0.05692724511027336, + 0.07962466776371002 + ], + "credit_map": [ + 0.0116031588986516, + 0.017283102497458458, + 0.011827053502202034, + 0.012298825196921825, + 0.013334928080439568, + 0.01301548257470131, + 0.010225548408925533, + 0.014030184596776962, + 0.007322241552174091, + 0.013103905133903027, + 0.006378344260156155, + 0.007573981303721666, + 0.010682113468647003, + 0.007510923780500889, + 0.005427561234682798, + 0.0041336556896567345, + 0.004548751283437014, + 0.005875424947589636, + 0.008543474599719048, + 0.00525665795430541, + 0.008840958587825298, + 0.011941657401621342, + 0.003980387933552265, + 0.005924489814788103, + 0.025595085695385933, + 0.035130105912685394, + 0.009344576857984066, + 0.006012402940541506, + 0.1803497076034546, + 0.12032332271337509, + 0.00894220545887947, + 0.013308241032063961, + 0.004619300365447998, + 0.005011158529669046, + 0.003175990888848901, + 0.0032356567680835724, + 0.009105286560952663, + 0.006631950847804546, + 0.0039044723380357027, + 0.0022976656910032034, + 0.007575678173452616, + 0.011753574945032597, + 0.00659182108938694, + 0.008637046441435814, + 0.017421092838048935, + 0.008044415153563023, + 0.010062230750918388, + 0.005331592168658972, + 0.010978383012115955, + 0.011480146087706089, + 0.010712153278291225, + 0.01922939531505108, + 0.01463779341429472, + 0.011728890240192413, + 0.01723482273519039, + 0.010656682774424553, + 0.0018548668595030904, + 0.003802661783993244, + 0.0032142263371497393, + 0.0028518992476165295, + 0.003481198102235794, + 0.0032894040923565626, + 0.0029086312279105186, + 0.002319527557119727, + 0.004426495637744665, + 0.007054606452584267, + 0.003439293010160327, + 0.001648829667828977, + 0.022085584700107574, + 0.011895084753632545, + 0.0034985302481800318, + 0.002878819592297077, + 0.007356912828981876, + 0.01517847552895546, + 0.008366188034415245, + 0.009547200985252857, + 0.015129663050174713, + 0.0070237996987998486, + 0.009385312907397747, + 0.007637114729732275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.02634642831981182 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020568644627928734 + ], + "opposite_logprob_tokens": [ + -3.895568609237671 + ], + "logprob_reward": -0.020568644627928734, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020568644627928734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13849776983261108, + 0.18032507598400116, + 0.11989626288414001, + 0.26900044083595276, + 0.020409420132637024, + 0.04157644510269165, + 0.03061782196164131, + 0.08588622510433197, + 0.08852613717317581, + 0.02526441402733326 + ], + "credit_map": [ + 0.01707860268652439, + 0.0218423493206501, + 0.015436865389347076, + 0.015388686209917068, + 0.015165668912231922, + 0.01780412718653679, + 0.016129810363054276, + 0.01965167000889778, + 0.01508790161460638, + 0.023732982575893402, + 0.019116666167974472, + 0.02858283556997776, + 0.03699735924601555, + 0.02346181683242321, + 0.01651480421423912, + 0.016830703243613243, + 0.009869539178907871, + 0.007441704627126455, + 0.03118719719350338, + 0.008042804896831512, + 0.015652235597372055, + 0.031197788193821907, + 0.007616643328219652, + 0.008888349868357182, + 0.016638660803437233, + 0.022554874420166016, + 0.006644837092608213, + 0.004403560888022184, + 0.11868863552808762, + 0.08234099298715591, + 0.007746899966150522, + 0.00998197216540575, + 0.001975106541067362, + 0.002313359174877405, + 0.002547807525843382, + 0.0020783699583262205, + 0.005423394963145256, + 0.0025695054791867733, + 0.001842134865000844, + 0.001659739064052701, + 0.004506720695644617, + 0.004438783507794142, + 0.003234128002077341, + 0.005781140178442001, + 0.009362127631902695, + 0.004802820272743702, + 0.00535175483673811, + 0.004098969511687756, + 0.003179969498887658, + 0.004076170269399881, + 0.0031217229552567005, + 0.004952852614223957, + 0.0037524013314396143, + 0.00401532044634223, + 0.0043641021475195885, + 0.003155285259708762, + 0.005616146605461836, + 0.01523787621408701, + 0.00883612222969532, + 0.0082874009385705, + 0.01961231231689453, + 0.010500261560082436, + 0.011372899636626244, + 0.006423197686672211, + 0.0062910825945436954, + 0.008488640189170837, + 0.0023808383848518133, + 0.002048793016001582, + 0.032353684306144714, + 0.025962479412555695, + 0.005012252368032932, + 0.005988366901874542, + 0.002231174148619175, + 0.0031154814641922712, + 0.002883853390812874, + 0.002447438193485141, + 0.004191803280264139, + 0.003998671192675829, + 0.004510215017944574, + 0.001885776873677969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.020568644627928734 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631422184407711 + ], + "opposite_logprob_tokens": [ + -4.758631229400635 + ], + "logprob_reward": -0.008631422184407711, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631422184407711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21804367005825043, + 0.08565135300159454, + 0.18042461574077606, + 0.13040925562381744, + 0.020893126726150513, + 0.017089569941163063, + 0.047026149928569794, + 0.1944553405046463, + 0.05314278602600098, + 0.05286417528986931 + ], + "credit_map": [ + 0.034588657319545746, + 0.0435471273958683, + 0.021823003888130188, + 0.022227896377444267, + 0.02490953356027603, + 0.025146054103970528, + 0.021116316318511963, + 0.024685049429535866, + 0.007944345474243164, + 0.010660595260560513, + 0.00906104315072298, + 0.012966955080628395, + 0.016877111047506332, + 0.013253921642899513, + 0.008328001014888287, + 0.006559364032000303, + 0.016714923083782196, + 0.009063877165317535, + 0.06333108991384506, + 0.010995399206876755, + 0.015500680543482304, + 0.04869111627340317, + 0.006344739813357592, + 0.009782776236534119, + 0.008379960432648659, + 0.01215960830450058, + 0.007547301240265369, + 0.00418211380019784, + 0.054031845182180405, + 0.03427843004465103, + 0.004881818778812885, + 0.004948161542415619, + 0.0022400273010134697, + 0.002870691940188408, + 0.002797536551952362, + 0.0023987572640180588, + 0.004599932115525007, + 0.002275951439514756, + 0.0018960420275107026, + 0.0018141861073672771, + 0.002042554086074233, + 0.0017362545477226377, + 0.001953774830326438, + 0.0020031952299177647, + 0.0038855026941746473, + 0.001820992212742567, + 0.001933773630298674, + 0.0017135217785835266, + 0.006217021029442549, + 0.006212892010807991, + 0.004300528205931187, + 0.00801287405192852, + 0.005571454763412476, + 0.005495281424373388, + 0.007449695840477943, + 0.003766395850107074, + 0.012653091922402382, + 0.02636340633034706, + 0.01761181466281414, + 0.014803875237703323, + 0.05920248478651047, + 0.034886039793491364, + 0.01586145907640457, + 0.013073145411908627, + 0.004104192368686199, + 0.005918694194406271, + 0.002026698784902692, + 0.0017629326321184635, + 0.017571428790688515, + 0.01360936276614666, + 0.0032739974558353424, + 0.004875470418483019, + 0.005491445306688547, + 0.007493718061596155, + 0.005412892438471317, + 0.00513049028813839, + 0.01148972101509571, + 0.0068285176530480385, + 0.007390265818685293, + 0.003627119120210409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.008631422184407711 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of photorealistic CGI cinematic realism, a steady medium shot follows a massive landing ship dock, its powerful hull painted a non-reflective navy gray and streaked with rust, as it plunges into the tsunami. The ship struggles in the fierce, turbulent sea, pitching and rolling violently amidst the chaotic, deep blue waves, crashing onto its port side and sending a massive spray of white water cascading down onto the deck. On the horizon, a vast, dark, roiling wall of water rises at a terrifying speed, a liquid mountain whose sheer surface is a vortex of shadows and movement. Above, an ominous, scarred sky seethes with heavy storm clouds of ashen gray and deep purple, pierced only by a cold, steely glow. This bleak lighting captures the foaming, wind-ripped crests of the approaching waves, casting a sinister glow as the naval vessel pushes forward, its structure groaning under immense pressure, creating an atmosphere of tension and despair, of impending, overwhelming disaster.", + "seed": 4287, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_0/index_4889_seed_4287_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08360697319731117, + "vlm_logprob": -0.08360697319731117, + "vlm_yes_no_margin": 3.550000011920929, + "vlm_token_credit": -0.08360697319731117, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35323137044906616 + ], + "opposite_logprob_tokens": [ + -1.228231430053711 + ], + "logprob_reward": -0.35323137044906616, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.35323137044906616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07247721403837204, + 0.0294888224452734, + 0.2484411895275116, + 0.14690609276294708, + 0.09711019694805145, + 0.035561222583055496, + 0.14517521858215332, + 0.02986830286681652, + 0.12737971544265747, + 0.06759199500083923 + ], + "credit_map": [ + 0.008532630279660225, + 0.008434047922492027, + 0.014657700434327126, + 0.008138860575854778, + 0.00788817461580038, + 0.007907810620963573, + 0.008943093940615654, + 0.007974891923367977, + 0.003173119155690074, + 0.004142286721616983, + 0.004858309868723154, + 0.005326003767549992, + 0.003837013151496649, + 0.0028734998777508736, + 0.003110900055617094, + 0.0021676907781511545, + 0.02580387331545353, + 0.03085867129266262, + 0.02960234135389328, + 0.057368408888578415, + 0.03182464465498924, + 0.0206123236566782, + 0.030688835307955742, + 0.021682100370526314, + 0.017007900401949883, + 0.019130581989884377, + 0.032988958060741425, + 0.021255934610962868, + 0.01990673318505287, + 0.009777289815247059, + 0.013290025293827057, + 0.01354867685586214, + 0.012328330427408218, + 0.013423379510641098, + 0.015686700120568275, + 0.012356816790997982, + 0.013762121088802814, + 0.0127842016518116, + 0.008241873234510422, + 0.008526775054633617, + 0.005460883025079966, + 0.005236375145614147, + 0.0034183927346020937, + 0.004211325664073229, + 0.004674642812460661, + 0.0037370151840150356, + 0.004517761990427971, + 0.004304827190935612, + 0.027615472674369812, + 0.024051548913121223, + 0.009385421872138977, + 0.019106319174170494, + 0.012505201622843742, + 0.025690468028187752, + 0.013562425971031189, + 0.013258368708193302, + 0.005623256787657738, + 0.009793279692530632, + 0.0018897328991442919, + 0.002667726017534733, + 0.002639054786413908, + 0.003019674215465784, + 0.0017580684507265687, + 0.0024775101337581873, + 0.018386784940958023, + 0.018516257405281067, + 0.01284126564860344, + 0.015104304999113083, + 0.020306959748268127, + 0.016036907210946083, + 0.014308816753327847, + 0.011878417804837227, + 0.009970351122319698, + 0.014777852222323418, + 0.006997773423790932, + 0.007299263030290604, + 0.008773937821388245, + 0.007125530391931534, + 0.005592449102550745, + 0.0070548406802117825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.35323137044906616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209050685167313 + ], + "opposite_logprob_tokens": [ + -4.264209270477295 + ], + "logprob_reward": -0.014209050685167313, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209050685167313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11708765476942062, + 0.06322883069515228, + 0.05362379178404808, + 0.07586998492479324, + 0.16819453239440918, + 0.04730985686182976, + 0.22104467451572418, + 0.027088245376944542, + 0.10816355049610138, + 0.11838886886835098 + ], + "credit_map": [ + 0.015456686727702618, + 0.021626019850373268, + 0.016631200909614563, + 0.014216210693120956, + 0.016415255144238472, + 0.010504970327019691, + 0.011649339459836483, + 0.010587972588837147, + 0.005879516713321209, + 0.0073556783609092236, + 0.008357318118214607, + 0.011884680017828941, + 0.008152480237185955, + 0.004292124882340431, + 0.011942433193325996, + 0.005364605691283941, + 0.005407338496297598, + 0.005391852464526892, + 0.008327685296535492, + 0.012457158416509628, + 0.005537332035601139, + 0.004630522336810827, + 0.005155995953828096, + 0.006715909577906132, + 0.009137448854744434, + 0.009033143520355225, + 0.009674298577010632, + 0.014858507551252842, + 0.011322389356791973, + 0.006441969890147448, + 0.008679717779159546, + 0.006722515448927879, + 0.025642069056630135, + 0.02377372235059738, + 0.02028256468474865, + 0.014258398674428463, + 0.022476710379123688, + 0.021072788164019585, + 0.018894044682383537, + 0.02179422788321972, + 0.006458744406700134, + 0.0054787928238511086, + 0.006327689625322819, + 0.007302767131477594, + 0.004775510635226965, + 0.004753377754241228, + 0.006810386665165424, + 0.005402594804763794, + 0.040042802691459656, + 0.03291452303528786, + 0.01307052280753851, + 0.02475048415362835, + 0.023570748046040535, + 0.05367707088589668, + 0.014414256438612938, + 0.018604272976517677, + 0.004234995227307081, + 0.005866102408617735, + 0.0017747217789292336, + 0.003348149359226227, + 0.0027889783959835768, + 0.0028092197608202696, + 0.003031235421076417, + 0.003234845120459795, + 0.01843188889324665, + 0.018065307289361954, + 0.012473540380597115, + 0.01147092692553997, + 0.012968371622264385, + 0.010667924769222736, + 0.009607363492250443, + 0.01447824016213417, + 0.016363942995667458, + 0.02271297387778759, + 0.01534150168299675, + 0.017930351197719574, + 0.009151852689683437, + 0.011942694894969463, + 0.010777616873383522, + 0.01416794117540121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.014209050685167313 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329215034842491 + ], + "opposite_logprob_tokens": [ + -3.773292064666748 + ], + "logprob_reward": -0.02329215034842491, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329215034842491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09819430857896805, + 0.055601567029953, + 0.09310466051101685, + 0.14028292894363403, + 0.15674784779548645, + 0.060836613178253174, + 0.08930733799934387, + 0.1063428744673729, + 0.10226032137870789, + 0.0973215401172638 + ], + "credit_map": [ + 0.010156305506825447, + 0.012634704820811749, + 0.019124602898955345, + 0.018745144829154015, + 0.011613331735134125, + 0.009976351633667946, + 0.008367275819182396, + 0.007576595526188612, + 0.0058677783235907555, + 0.007571091875433922, + 0.009322498925030231, + 0.008947852067649364, + 0.00707936892285943, + 0.004552698228508234, + 0.0062219249084591866, + 0.006038358900696039, + 0.008713720366358757, + 0.01110802125185728, + 0.01593812182545662, + 0.01925073005259037, + 0.011559322476387024, + 0.008075710386037827, + 0.008796224370598793, + 0.009662805125117302, + 0.014836104586720467, + 0.0171167254447937, + 0.02047048509120941, + 0.02630394510924816, + 0.01989392191171646, + 0.010195656679570675, + 0.016705937683582306, + 0.014760158024728298, + 0.023535914719104767, + 0.022161543369293213, + 0.016724979504942894, + 0.013941789045929909, + 0.02228310890495777, + 0.023570356890559196, + 0.016103804111480713, + 0.018426362425088882, + 0.013001534156501293, + 0.01627459190785885, + 0.005298469681292772, + 0.005116569343954325, + 0.0058020479045808315, + 0.005195776931941509, + 0.006658163387328386, + 0.0034894610289484262, + 0.015241912566125393, + 0.012616141699254513, + 0.005859149154275656, + 0.010347845032811165, + 0.010165245272219181, + 0.0210255179554224, + 0.006577088963240385, + 0.007474442943930626, + 0.018543703481554985, + 0.01769299805164337, + 0.01199342031031847, + 0.012318672612309456, + 0.007807621732354164, + 0.014868143945932388, + 0.010907766409218311, + 0.012210550718009472, + 0.016624722629785538, + 0.01878858357667923, + 0.011346935294568539, + 0.013713357038795948, + 0.011196897365152836, + 0.008273700252175331, + 0.008758313953876495, + 0.013557816855609417, + 0.015985295176506042, + 0.017511077225208282, + 0.008917365223169327, + 0.009913429617881775, + 0.012418710626661777, + 0.01477835513651371, + 0.007977326400578022, + 0.00981997698545456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.02329215034842491 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057238109409809 + ], + "opposite_logprob_tokens": [ + -3.8955724239349365 + ], + "logprob_reward": -0.02057238109409809, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057238109409809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1778043955564499, + 0.05873057618737221, + 0.10536924004554749, + 0.09604848921298981, + 0.14284609258174896, + 0.06359042972326279, + 0.14772789180278778, + 0.05011092871427536, + 0.11267828196287155, + 0.04509364068508148 + ], + "credit_map": [ + 0.018045838922262192, + 0.02835964225232601, + 0.024755993857979774, + 0.016762783750891685, + 0.02883949689567089, + 0.018981043249368668, + 0.019975891336798668, + 0.022083701565861702, + 0.007438375148922205, + 0.01275546196848154, + 0.006022090092301369, + 0.005135688465088606, + 0.007739849854260683, + 0.0056729912757873535, + 0.009074444882571697, + 0.004891674034297466, + 0.01115968357771635, + 0.013043832965195179, + 0.012072484940290451, + 0.024100646376609802, + 0.013506588526070118, + 0.012961769476532936, + 0.011261933483183384, + 0.007262296974658966, + 0.010204908438026905, + 0.01205953024327755, + 0.018457850441336632, + 0.016801925376057625, + 0.012469611130654812, + 0.00682447524741292, + 0.009881989099085331, + 0.009348197840154171, + 0.021124079823493958, + 0.02165309339761734, + 0.013632837682962418, + 0.014982614666223526, + 0.016289016231894493, + 0.027894549071788788, + 0.012183054350316525, + 0.015086857602000237, + 0.011247873306274414, + 0.012256195768713951, + 0.00622600270435214, + 0.006380097474902868, + 0.007091468200087547, + 0.010721542872488499, + 0.003947226796299219, + 0.0057200235314667225, + 0.029586462303996086, + 0.023433629423379898, + 0.01099394541233778, + 0.022472228854894638, + 0.013429886661469936, + 0.024463824927806854, + 0.011591953225433826, + 0.011755971238017082, + 0.008276104927062988, + 0.013304389081895351, + 0.00397123396396637, + 0.006400193087756634, + 0.004413427319377661, + 0.004198061302304268, + 0.0037437458522617817, + 0.005803775042295456, + 0.016226815059781075, + 0.022511499002575874, + 0.009871689602732658, + 0.014055229723453522, + 0.013046362437307835, + 0.012305689975619316, + 0.011713909916579723, + 0.012947085313498974, + 0.006333936005830765, + 0.007185178808867931, + 0.004922004416584969, + 0.008415469899773598, + 0.004513893276453018, + 0.004544101655483246, + 0.0044185929000377655, + 0.004760465119034052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.02057238109409809 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive landing ship dock with a non-reflective navy gray hull streaked with rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship shown plunging into a tsunami and struggling by pitching and rolling violently in deep blue waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ship crash onto its port side, sending a massive spray of white water cascading down onto the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast, dark, roiling wall of water on the horizon rising at speed, appearing as a vortex of shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky ominous with heavy storm clouds of ashen gray and deep purple, pierced by a cold, steely glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006729913409799337 + ], + "opposite_logprob_tokens": [ + -5.006730079650879 + ], + "logprob_reward": -0.006729913409799337, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006729913409799337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18021425604820251, + 0.08668030053377151, + 0.1276080161333084, + 0.10332267731428146, + 0.07338723540306091, + 0.17506413161754608, + 0.0409073531627655, + 0.04914327710866928, + 0.07536575943231583, + 0.0883069783449173 + ], + "credit_map": [ + 0.020159488543868065, + 0.019513817504048347, + 0.028412071987986565, + 0.02275296486914158, + 0.025025147944688797, + 0.019866101443767548, + 0.02377052791416645, + 0.020714161917567253, + 0.009048250503838062, + 0.016672849655151367, + 0.01007835939526558, + 0.01143536064773798, + 0.013545053079724312, + 0.006294430233538151, + 0.013390007428824902, + 0.006215999368578196, + 0.014148370362818241, + 0.015902532264590263, + 0.013688536360859871, + 0.020786326378583908, + 0.017289483919739723, + 0.01908990740776062, + 0.013859796337783337, + 0.012843074277043343, + 0.010609125718474388, + 0.013715986162424088, + 0.013605941087007523, + 0.014087608084082603, + 0.015093548223376274, + 0.016658678650856018, + 0.00989561527967453, + 0.009656180627644062, + 0.009603261947631836, + 0.01126924343407154, + 0.00697358837351203, + 0.008510760962963104, + 0.008240047842264175, + 0.014636571519076824, + 0.006748199928551912, + 0.007405569311231375, + 0.03456367924809456, + 0.04006141796708107, + 0.012218011543154716, + 0.012744289822876453, + 0.01759319007396698, + 0.033962275832891464, + 0.010907327756285667, + 0.013013946823775768, + 0.008713265880942345, + 0.008573468774557114, + 0.0033580539748072624, + 0.004909744951874018, + 0.003299192525446415, + 0.003885864047333598, + 0.004586293827742338, + 0.0035814722068607807, + 0.008889530785381794, + 0.010249019600450993, + 0.0039033384528011084, + 0.005193898919969797, + 0.004503624513745308, + 0.0045859492383897305, + 0.0053262352012097836, + 0.006491686217486858, + 0.011054654605686665, + 0.017019249498844147, + 0.005724005401134491, + 0.010545124299824238, + 0.007009638007730246, + 0.006519503891468048, + 0.006747052539139986, + 0.010746538639068604, + 0.012815846130251884, + 0.01636948063969612, + 0.008839065209031105, + 0.013504084199666977, + 0.007973081432282925, + 0.008590939454734325, + 0.007064867299050093, + 0.013149627484381199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.006729913409799337 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_1/index_tensor([8578, 8578])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_1/index_tensor([8578, 8578])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..13721d932cef5afc0d98cfd56e39e749b9bea9f2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_1/index_tensor([8578, 8578])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from an extremely low, upward angle, the camera slowly and steadily moves forward, locking the viewer's gaze squarely on the face of a majestic, adult white tiger. The beast's snow-white fur is dotted with barely perceptible pale gray streaks, and in the dappled light filtering through the gloomy forest, each individual hair appears distinctly textured. Facing the camera, its muscular maw opens wide to its fullest, revealing a moist, pink interior and a barbed tongue. Two massive, ivory-textured canines gleam with a deadly coldness in the light, surrounded by smaller teeth of equal razor-sharpness. Its icy blue eyes narrow into murderous slits of intense rage, and its wiry white beard stretches forward with intense tension. With a silent yet powerful roar, emanating from its chest, the muscles of its neck and shoulders bulge, the force of which seems to shake the very air, and a stream of glistening saliva scatters from between its teeth. The background is a completely out-of-focus, deep and humid green jungle, creating a primitive, awe-inspiring sense of wild power and extremely tense oppression.", + "seed": 7526, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_1/index_8578_seed_7526_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05994423711672425, + "vlm_logprob": -0.05994423711672425, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.05994423711672425, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008152540773153305 + ], + "opposite_logprob_tokens": [ + -6.133152484893799 + ], + "logprob_reward": -0.008152540773153305, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.008152540773153305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19153863191604614, + 0.10841825604438782, + 0.0957164540886879, + 0.03318708389997482, + 0.09158182144165039, + 0.1388964205980301, + 0.05552023649215698, + 0.021375484764575958, + 0.1379772424697876, + 0.12578831613063812 + ], + "credit_map": [ + 0.030832812190055847, + 0.025542620569467545, + 0.029343316331505775, + 0.024570772424340248, + 0.02648293413221836, + 0.018469808623194695, + 0.015203578397631645, + 0.02109280414879322, + 0.010744700208306313, + 0.009504435583949089, + 0.015204239636659622, + 0.013680288568139076, + 0.021267453208565712, + 0.01071010623127222, + 0.018675070255994797, + 0.008631966076791286, + 0.021989282220602036, + 0.010237038135528564, + 0.015336007811129093, + 0.009548380970954895, + 0.008079152554273605, + 0.007210570387542248, + 0.012201907113194466, + 0.01111411303281784, + 0.004182003904134035, + 0.003458562307059765, + 0.00434931879863143, + 0.004328742623329163, + 0.006151555571705103, + 0.002536618849262595, + 0.005288147367537022, + 0.0028921328485012054, + 0.02411092445254326, + 0.004977802745997906, + 0.032272059470415115, + 0.007387511432170868, + 0.005223166663199663, + 0.005151066463440657, + 0.0076223695650696754, + 0.004836919717490673, + 0.046231262385845184, + 0.008611645549535751, + 0.019067805260419846, + 0.007455618120729923, + 0.009909753687679768, + 0.011566607281565666, + 0.01125688012689352, + 0.0247968640178442, + 0.006183440797030926, + 0.004470227286219597, + 0.0074964906089007854, + 0.008374200202524662, + 0.006782612297683954, + 0.005326287820935249, + 0.00889632012695074, + 0.007990656420588493, + 0.003483199281617999, + 0.0015452540246769786, + 0.0034404003527015448, + 0.0022417805157601833, + 0.002030297415331006, + 0.0011664088815450668, + 0.0047947801649570465, + 0.0026733647100627422, + 0.021638892590999603, + 0.011326393112540245, + 0.01962761953473091, + 0.013933063484728336, + 0.02038707584142685, + 0.011438127607107162, + 0.024863282218575478, + 0.01476279180496931, + 0.01534269005060196, + 0.011856666766107082, + 0.017592133954167366, + 0.012056373991072178, + 0.015774480998516083, + 0.0193430595099926, + 0.01785934343934059, + 0.015963556244969368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.008152540773153305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557751834392548 + ], + "opposite_logprob_tokens": [ + -1.600577473640442 + ], + "logprob_reward": -0.22557751834392548, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557751834392548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2196890264749527, + 0.03621438518166542, + 0.04667387902736664, + 0.08792467415332794, + 0.16495266556739807, + 0.053233303129673004, + 0.04236557334661484, + 0.014652944169938564, + 0.2599385380744934, + 0.07435508072376251 + ], + "credit_map": [ + 0.02819168195128441, + 0.021198585629463196, + 0.031784944236278534, + 0.029661618173122406, + 0.0317075215280056, + 0.026761164888739586, + 0.021128162741661072, + 0.029255356639623642, + 0.003496491815894842, + 0.003979288041591644, + 0.0049064322374761105, + 0.004558588843792677, + 0.005722690373659134, + 0.005083215422928333, + 0.00522179901599884, + 0.0032458784990012646, + 0.009113287553191185, + 0.004440962802618742, + 0.006734246388077736, + 0.003473584307357669, + 0.006361854262650013, + 0.004394543822854757, + 0.008036782033741474, + 0.004118616227060556, + 0.01211115624755621, + 0.007092313840985298, + 0.012166788801550865, + 0.011797926388680935, + 0.013513724319636822, + 0.007589367683976889, + 0.01650463603436947, + 0.00714875990524888, + 0.0615159347653389, + 0.005740078166127205, + 0.06748594343662262, + 0.007060330361127853, + 0.0038762062322348356, + 0.007636981550604105, + 0.005813857540488243, + 0.00582332257181406, + 0.019288813695311546, + 0.003023147815838456, + 0.006774642504751682, + 0.002548432443290949, + 0.0017904946580529213, + 0.004113487899303436, + 0.0028695366345345974, + 0.01282474584877491, + 0.005324762314558029, + 0.00371017144061625, + 0.005747335031628609, + 0.00460016867145896, + 0.007945570163428783, + 0.004914971999824047, + 0.005983342416584492, + 0.004139249678701162, + 0.0021943198516964912, + 0.0011156962718814611, + 0.002101980848237872, + 0.0016358362045139074, + 0.0012093697441741824, + 0.0007741957670077682, + 0.003961021546274424, + 0.0016605242853984237, + 0.04252662509679794, + 0.015504752285778522, + 0.042053643614053726, + 0.02094966173171997, + 0.024106493219733238, + 0.015819817781448364, + 0.07700515538454056, + 0.02197238616645336, + 0.015358134172856808, + 0.007392781786620617, + 0.01018238253891468, + 0.009776407852768898, + 0.010191673412919044, + 0.005840142257511616, + 0.008228341117501259, + 0.007385216653347015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.22557751834392548 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634515054523945 + ], + "opposite_logprob_tokens": [ + -3.6513452529907227 + ], + "logprob_reward": -0.02634515054523945, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634515054523945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13184703886508942, + 0.042199209332466125, + 0.0798245370388031, + 0.07499044388532639, + 0.11970119923353195, + 0.025404062122106552, + 0.050547875463962555, + 0.08148005604743958, + 0.34318235516548157, + 0.05082327499985695 + ], + "credit_map": [ + 0.015936551615595818, + 0.010836632922291756, + 0.023378411307930946, + 0.015753725543618202, + 0.01600944995880127, + 0.013294536620378494, + 0.016140183433890343, + 0.020497547462582588, + 0.00788755901157856, + 0.0037473635748028755, + 0.006228610873222351, + 0.0049080029129981995, + 0.005586699582636356, + 0.003495449898764491, + 0.005801894702017307, + 0.004543627612292767, + 0.010994178242981434, + 0.011383363977074623, + 0.011173414066433907, + 0.00939446222037077, + 0.008835944347083569, + 0.007494853809475899, + 0.011593322269618511, + 0.00895499624311924, + 0.010559937916696072, + 0.0063955942168831825, + 0.010336562059819698, + 0.011687948368489742, + 0.010994347743690014, + 0.004836662206798792, + 0.013247588649392128, + 0.006931801792234182, + 0.04413403570652008, + 0.0046261027455329895, + 0.04796118289232254, + 0.00491657480597496, + 0.0033080868888646364, + 0.00545587157830596, + 0.004920019302517176, + 0.00437932088971138, + 0.00457542110234499, + 0.0028868410736322403, + 0.0027357812505215406, + 0.00250165443867445, + 0.0030594614800065756, + 0.004055588040500879, + 0.0032266525086015463, + 0.0023626629263162613, + 0.008325324393808842, + 0.003564420621842146, + 0.007915698923170567, + 0.006885104347020388, + 0.006697200704365969, + 0.004758000373840332, + 0.007333293091505766, + 0.005068831145763397, + 0.011955664493143559, + 0.00489546824246645, + 0.012201632373034954, + 0.009858398698270321, + 0.0082173403352499, + 0.004798846319317818, + 0.01883699558675289, + 0.010715710930526257, + 0.04363621026277542, + 0.016198841854929924, + 0.05250990018248558, + 0.02594836987555027, + 0.022923005744814873, + 0.0206166822463274, + 0.13267065584659576, + 0.028678663074970245, + 0.008453418500721455, + 0.004885703790932894, + 0.00744742713868618, + 0.006284391973167658, + 0.005646382924169302, + 0.004945811815559864, + 0.006881493143737316, + 0.006278643850237131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.02634515054523945 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370201960206032 + ], + "opposite_logprob_tokens": [ + -3.4087021350860596 + ], + "logprob_reward": -0.03370201960206032, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370201960206032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18915316462516785, + 0.04158708453178406, + 0.10731995850801468, + 0.09817706048488617, + 0.0819612368941307, + 0.044696658849716187, + 0.047680892050266266, + 0.015426167286932468, + 0.3122606575489044, + 0.06173720583319664 + ], + "credit_map": [ + 0.023195410147309303, + 0.02303403988480568, + 0.02826898917555809, + 0.03146733343601227, + 0.02251218445599079, + 0.023129193112254143, + 0.015615783631801605, + 0.02193022333085537, + 0.005367164965718985, + 0.0029443392995744944, + 0.006400310434401035, + 0.005659640766680241, + 0.007854784838855267, + 0.0032938497606664896, + 0.006806212943047285, + 0.003260782454162836, + 0.017199760302901268, + 0.01188886072486639, + 0.01520234439522028, + 0.009621604345738888, + 0.015123320743441582, + 0.01240867841988802, + 0.01280703954398632, + 0.013068347238004208, + 0.013635489158332348, + 0.008773380890488625, + 0.011745911091566086, + 0.01418137364089489, + 0.012630444951355457, + 0.006306944414973259, + 0.021168896928429604, + 0.009734616614878178, + 0.025564489886164665, + 0.004395872820168734, + 0.02929474413394928, + 0.004945327527821064, + 0.003608226077631116, + 0.0043115075677633286, + 0.00552269397303462, + 0.004318375140428543, + 0.013548399321734905, + 0.0027507843915373087, + 0.005569008179008961, + 0.0028732295613735914, + 0.0022162713576108217, + 0.004757136572152376, + 0.003583746962249279, + 0.0093980822712183, + 0.0067177568562328815, + 0.003776209196075797, + 0.00717255799099803, + 0.0059480927884578705, + 0.006874163635075092, + 0.004026827402412891, + 0.009840872138738632, + 0.003324410878121853, + 0.0025624355766922235, + 0.0013230619952082634, + 0.002632890362292528, + 0.0017697059083729982, + 0.002020476385951042, + 0.0009497145074419677, + 0.003009028034284711, + 0.001158853992819786, + 0.04126989468932152, + 0.015678899362683296, + 0.04860497638583183, + 0.024524817243218422, + 0.03007444739341736, + 0.019796673208475113, + 0.1044975221157074, + 0.027813415974378586, + 0.007548633962869644, + 0.005479838233441114, + 0.0079946992918849, + 0.005030789412558079, + 0.010324464179575443, + 0.009648703038692474, + 0.009043273515999317, + 0.006666801404207945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.03370201960206032 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943956319242716 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943956319242716, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943956319242716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23673878610134125, + 0.06446293741464615, + 0.053884852677583694, + 0.060071855783462524, + 0.11069570481777191, + 0.17533865571022034, + 0.0812910795211792, + 0.032050758600234985, + 0.13841521739959717, + 0.04705013707280159 + ], + "credit_map": [ + 0.03504974767565727, + 0.01788443513214588, + 0.04027146100997925, + 0.028217151761054993, + 0.03498968482017517, + 0.027075130492448807, + 0.024279966950416565, + 0.028971241787075996, + 0.007436410989612341, + 0.006313743535429239, + 0.008972018025815487, + 0.007301785051822662, + 0.00912886206060648, + 0.007929345592856407, + 0.00968949869275093, + 0.007691284641623497, + 0.010450669564306736, + 0.006149119231849909, + 0.008071623742580414, + 0.004198229871690273, + 0.007437092252075672, + 0.004788391292095184, + 0.006615474354475737, + 0.0061742570251226425, + 0.009658698923885822, + 0.004846824333071709, + 0.010135120712220669, + 0.008194077759981155, + 0.008740742690861225, + 0.003359031630679965, + 0.010190225206315517, + 0.004947141278535128, + 0.03365734964609146, + 0.007718461100012064, + 0.03652331978082657, + 0.005818321835249662, + 0.005358744412660599, + 0.007583863567560911, + 0.00664208410307765, + 0.007393582724034786, + 0.06711374223232269, + 0.00917817559093237, + 0.02541017159819603, + 0.007741113658994436, + 0.005418593529611826, + 0.014603563584387302, + 0.008681795559823513, + 0.03719150647521019, + 0.013675997033715248, + 0.006662669591605663, + 0.0115820886567235, + 0.00950897578150034, + 0.01358036883175373, + 0.008141281083226204, + 0.010206818580627441, + 0.007932888343930244, + 0.007664043456315994, + 0.0024906927719712257, + 0.005568638909608126, + 0.003362111048772931, + 0.003580842399969697, + 0.0033473188523203135, + 0.0033259147312492132, + 0.0027112006209790707, + 0.029374221339821815, + 0.010393020696938038, + 0.02375430054962635, + 0.011098367162048817, + 0.013918732292950153, + 0.00757019966840744, + 0.030814548954367638, + 0.011491844430565834, + 0.006525830831378698, + 0.005267315078526735, + 0.00630229152739048, + 0.007925794459879398, + 0.006063371431082487, + 0.004817166831344366, + 0.005514852236956358, + 0.004633518401533365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.005943956319242716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low, upward angle, the camera slowly and steadily moves forward, locking the viewer's gaze squarely on the face of a majestic, adult white tiger. The beast's snow-white fur is dotted with barely perceptible pale gray streaks, and in the dappled light filtering through the gloomy forest, each individual hair appears distinctly textured. Facing the camera, its muscular maw opens wide to its fullest, revealing a moist, pink interior and a barbed tongue. Two massive, ivory-textured canines gleam with a deadly coldness in the light, surrounded by smaller teeth of equal razor-sharpness. Its icy blue eyes narrow into murderous slits of intense rage, and its wiry white beard stretches forward with intense tension. With a silent yet powerful roar, emanating from its chest, the muscles of its neck and shoulders bulge, the force of which seems to shake the very air, and a stream of glistening saliva scatters from between its teeth. The background is a completely out-of-focus, deep and humid green jungle, creating a primitive, awe-inspiring sense of wild power and extremely tense oppression.", + "seed": 7527, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_1/index_8578_seed_7527_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07155788466334342, + "vlm_logprob": -0.07155788466334342, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.07155788466334342, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067752632312476635 + ], + "opposite_logprob_tokens": [ + -6.506775379180908 + ], + "logprob_reward": -0.0067752632312476635, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0067752632312476635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11666151136159897, + 0.08239106088876724, + 0.05879596993327141, + 0.03975594416260719, + 0.10033935308456421, + 0.16449958086013794, + 0.14351500570774078, + 0.14912444353103638, + 0.07599563896656036, + 0.06892146915197372 + ], + "credit_map": [ + 0.008653290569782257, + 0.02574325166642666, + 0.010558105073869228, + 0.018511410802602768, + 0.015906939283013344, + 0.011597742326557636, + 0.011694272048771381, + 0.013996506109833717, + 0.007346119731664658, + 0.007690458558499813, + 0.007051086984574795, + 0.011969679966568947, + 0.01658373884856701, + 0.009516355581581593, + 0.009324311278760433, + 0.012909319251775742, + 0.00478001032024622, + 0.005752003751695156, + 0.004548931028693914, + 0.007677101995795965, + 0.015873005613684654, + 0.007193074561655521, + 0.0063234842382371426, + 0.006648360751569271, + 0.003064970253035426, + 0.00434589758515358, + 0.004425721243023872, + 0.005791878327727318, + 0.005172219127416611, + 0.006143899634480476, + 0.00543324276804924, + 0.005378116853535175, + 0.008124511688947678, + 0.012863175012171268, + 0.011087221093475819, + 0.01672281138598919, + 0.013441260904073715, + 0.013850774616003036, + 0.01283994596451521, + 0.011409657076001167, + 0.012645155191421509, + 0.02121696248650551, + 0.01770426146686077, + 0.03425942361354828, + 0.015326296910643578, + 0.021919850260019302, + 0.02008218504488468, + 0.02134547010064125, + 0.011389798484742641, + 0.02201792597770691, + 0.013694667257368565, + 0.021417275071144104, + 0.018858185037970543, + 0.022584743797779083, + 0.019562676548957825, + 0.013989757746458054, + 0.011814218945801258, + 0.01883872039616108, + 0.021972090005874634, + 0.026274975389242172, + 0.022609584033489227, + 0.014678329229354858, + 0.017228348180651665, + 0.01570819318294525, + 0.005183482076972723, + 0.011078229174017906, + 0.008913944475352764, + 0.014967628754675388, + 0.009074627421796322, + 0.007600992452353239, + 0.010967973619699478, + 0.008208765648305416, + 0.006670241244137287, + 0.007590880151838064, + 0.007874751463532448, + 0.010017013177275658, + 0.01226134691387415, + 0.008070842362940311, + 0.00928752962499857, + 0.0071488674730062485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.0067752632312476635 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813222408294678 + ], + "opposite_logprob_tokens": [ + -1.4063222408294678 + ], + "logprob_reward": -0.2813222408294678, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813222408294678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23288697004318237, + 0.04216115549206734, + 0.125991091132164, + 0.04647253826260567, + 0.08491653949022293, + 0.11456909775733948, + 0.04436380788683891, + 0.08680310100317001, + 0.11692477017641068, + 0.10491088777780533 + ], + "credit_map": [ + 0.015490616671741009, + 0.032117653638124466, + 0.022169675678014755, + 0.04382850602269173, + 0.03290242701768875, + 0.025940945371985435, + 0.029628688469529152, + 0.03080846555531025, + 0.003137606894597411, + 0.006128562614321709, + 0.004518379457294941, + 0.008111786097288132, + 0.004853708203881979, + 0.0047936514019966125, + 0.0050238510593771935, + 0.005593614187091589, + 0.00846568588167429, + 0.020568454638123512, + 0.01156036090105772, + 0.015661224722862244, + 0.032836418598890305, + 0.014827616512775421, + 0.011825584806501865, + 0.010245757177472115, + 0.003752404358237982, + 0.005188975017517805, + 0.005240281578153372, + 0.0061912317760288715, + 0.009044361300766468, + 0.005352505017071962, + 0.0058333552442491055, + 0.005869425367563963, + 0.005987192038446665, + 0.012488694861531258, + 0.008974299766123295, + 0.013548162765800953, + 0.008460000157356262, + 0.013917938806116581, + 0.010567845776677132, + 0.010972417891025543, + 0.00942420307546854, + 0.0147666921839118, + 0.01313920971006155, + 0.020848829299211502, + 0.016274327412247658, + 0.013003792613744736, + 0.01253314595669508, + 0.01457891147583723, + 0.0045557706616818905, + 0.006448644679039717, + 0.005529709160327911, + 0.006780774798244238, + 0.0049410369247198105, + 0.006679351907223463, + 0.004151165019720793, + 0.005277362652122974, + 0.007104627788066864, + 0.008708156645298004, + 0.010068279691040516, + 0.016633586958050728, + 0.007809402421116829, + 0.012146152555942535, + 0.011192240752279758, + 0.01314066257327795, + 0.009247655048966408, + 0.015040291473269463, + 0.01675519160926342, + 0.023446809500455856, + 0.009720251895487309, + 0.01511075533926487, + 0.012965629808604717, + 0.014638198539614677, + 0.008029730059206486, + 0.008337765000760555, + 0.011403338983654976, + 0.017712995409965515, + 0.01307893916964531, + 0.009232460521161556, + 0.0222789254039526, + 0.014836740680038929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.2813222408294678 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634398825466633 + ], + "opposite_logprob_tokens": [ + -3.651344060897827 + ], + "logprob_reward": -0.02634398825466633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634398825466633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1976756602525711, + 0.1717779040336609, + 0.11402249336242676, + 0.04584681615233421, + 0.08021268993616104, + 0.03993869200348854, + 0.08426345139741898, + 0.04432090371847153, + 0.07893802970647812, + 0.14300335943698883 + ], + "credit_map": [ + 0.01526658795773983, + 0.030859971418976784, + 0.019599514082074165, + 0.02129204012453556, + 0.04665684700012207, + 0.02067946270108223, + 0.020844003185629845, + 0.022477226331830025, + 0.012582060880959034, + 0.018176676705479622, + 0.020870717242360115, + 0.03349916264414787, + 0.023530539125204086, + 0.018465695902705193, + 0.024662740528583527, + 0.019990315660834312, + 0.0075643109157681465, + 0.013423332013189793, + 0.01832490786910057, + 0.0140959108248353, + 0.022358490154147148, + 0.014336913824081421, + 0.012656139209866524, + 0.01126248948276043, + 0.0034787654876708984, + 0.005160832777619362, + 0.004365871194750071, + 0.00809406116604805, + 0.008809792809188366, + 0.0059450967237353325, + 0.004357333295047283, + 0.00563505943864584, + 0.005934093147516251, + 0.010314750485122204, + 0.010016221553087234, + 0.010422123596072197, + 0.010700777173042297, + 0.0123223215341568, + 0.01033961120992899, + 0.010162788443267345, + 0.003083517076447606, + 0.0045509785413742065, + 0.00536416657269001, + 0.008246690966188908, + 0.0037456382997334003, + 0.004264068324118853, + 0.005539180710911751, + 0.005144452676177025, + 0.005728948395699263, + 0.010117296129465103, + 0.013679304160177708, + 0.01463302318006754, + 0.009046724066138268, + 0.009872671216726303, + 0.01195905264467001, + 0.009226428344845772, + 0.003184514818713069, + 0.006614969111979008, + 0.003011989174410701, + 0.009382449090480804, + 0.007245994638651609, + 0.005231172312051058, + 0.004934846889227629, + 0.004714967682957649, + 0.006104812491685152, + 0.006904078647494316, + 0.011143038049340248, + 0.01592778041958809, + 0.008713369257748127, + 0.010135279037058353, + 0.009992660023272038, + 0.010017006658017635, + 0.00877122301608324, + 0.01778293214738369, + 0.015526138246059418, + 0.02171189896762371, + 0.03059878572821617, + 0.01712668128311634, + 0.017273077741265297, + 0.01421262975782156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02634398825466633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038099709898233414 + ], + "opposite_logprob_tokens": [ + -3.288099765777588 + ], + "logprob_reward": -0.038099709898233414, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038099709898233414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15234814584255219, + 0.061294086277484894, + 0.133792445063591, + 0.08854636549949646, + 0.10582864284515381, + 0.029031341895461082, + 0.08095524460077286, + 0.10291207581758499, + 0.10045351088047028, + 0.1448381394147873 + ], + "credit_map": [ + 0.010621781460940838, + 0.021300753578543663, + 0.013516206294298172, + 0.023016657680273056, + 0.03391873463988304, + 0.01494310051202774, + 0.018350806087255478, + 0.016680104658007622, + 0.004709895234555006, + 0.0057655577547848225, + 0.00800544023513794, + 0.012695283629000187, + 0.007025410886853933, + 0.0071799540892243385, + 0.008038898929953575, + 0.00787364412099123, + 0.009373713284730911, + 0.02353084273636341, + 0.011226535774767399, + 0.014666223898530006, + 0.033018749207258224, + 0.015338083729147911, + 0.013620973564684391, + 0.01301732286810875, + 0.0056788185611367226, + 0.01256642211228609, + 0.012781992554664612, + 0.01652849279344082, + 0.01046138908714056, + 0.012497331015765667, + 0.008445666171610355, + 0.009586256928741932, + 0.008703846484422684, + 0.014545207843184471, + 0.013485905714333057, + 0.018303479999303818, + 0.013202249072492123, + 0.012913163751363754, + 0.012900407426059246, + 0.011774382553994656, + 0.0024596538860350847, + 0.003350700717419386, + 0.0025548767298460007, + 0.005739891901612282, + 0.0026975723449140787, + 0.00422225845977664, + 0.0042477198876440525, + 0.0037586677353829145, + 0.006468860898166895, + 0.009215357713401318, + 0.008631390519440174, + 0.015344790183007717, + 0.011546025983989239, + 0.00996552873402834, + 0.009752222336828709, + 0.010031070560216904, + 0.006943821907043457, + 0.010904470458626747, + 0.008384652435779572, + 0.025738606229424477, + 0.008691051043570042, + 0.014786992222070694, + 0.013776530511677265, + 0.013685949146747589, + 0.006630399264395237, + 0.015810860320925713, + 0.0113085201010108, + 0.020355064421892166, + 0.012147012166678905, + 0.007665819488465786, + 0.01440636720508337, + 0.012129463255405426, + 0.012421770952641964, + 0.012143709696829319, + 0.015553629957139492, + 0.025946160778403282, + 0.015859555453062057, + 0.019059086218476295, + 0.027066336944699287, + 0.016787894070148468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.038099709898233414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248221103101969 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005248221103101969, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248221103101969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19274485111236572, + 0.08988270163536072, + 0.045439060777425766, + 0.08807495981454849, + 0.08461922407150269, + 0.04295456409454346, + 0.1288669854402542, + 0.2262130230665207, + 0.04089106619358063, + 0.06031355634331703 + ], + "credit_map": [ + 0.01777409203350544, + 0.029355520382523537, + 0.01658426970243454, + 0.019687123596668243, + 0.035824134945869446, + 0.02226593717932701, + 0.0178327988833189, + 0.03342097997665405, + 0.0068297795951366425, + 0.009805996902287006, + 0.009093351662158966, + 0.020154599100351334, + 0.007450798526406288, + 0.010545377619564533, + 0.012828588485717773, + 0.013174207881093025, + 0.003498067380860448, + 0.004833569750189781, + 0.004689402412623167, + 0.007724843453615904, + 0.005808897316455841, + 0.0061004869639873505, + 0.006765015888959169, + 0.0060187773779034615, + 0.0063141086138784885, + 0.008565695025026798, + 0.01404678076505661, + 0.016425184905529022, + 0.012675468809902668, + 0.012532101944088936, + 0.008283248171210289, + 0.009232372045516968, + 0.007502967957407236, + 0.008002940565347672, + 0.006628270260989666, + 0.008000078611075878, + 0.008371188305318356, + 0.024871213361620903, + 0.00838468223810196, + 0.012857885099947453, + 0.003910926170647144, + 0.005579257849603891, + 0.005176660604774952, + 0.006878382991999388, + 0.0054762233048677444, + 0.005454393103718758, + 0.004991966765373945, + 0.00548675237223506, + 0.009847634471952915, + 0.015094730071723461, + 0.016209105029702187, + 0.020241713151335716, + 0.02066456340253353, + 0.01746169477701187, + 0.014597042463719845, + 0.014750500209629536, + 0.019639553502202034, + 0.02746124565601349, + 0.03408181294798851, + 0.041600193828344345, + 0.034303195774555206, + 0.018532514572143555, + 0.02243172749876976, + 0.028162779286503792, + 0.0034231022000312805, + 0.0039283097721636295, + 0.005375937093049288, + 0.007804151624441147, + 0.005247716326266527, + 0.004894875455647707, + 0.005791248753666878, + 0.004425722174346447, + 0.004607388749718666, + 0.00606371508911252, + 0.007742653600871563, + 0.008952164091169834, + 0.008014002814888954, + 0.005836627911776304, + 0.010567215271294117, + 0.008529787883162498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.005248221103101969 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low, upward angle, the camera slowly and steadily moves forward, locking the viewer's gaze squarely on the face of a majestic, adult white tiger. The beast's snow-white fur is dotted with barely perceptible pale gray streaks, and in the dappled light filtering through the gloomy forest, each individual hair appears distinctly textured. Facing the camera, its muscular maw opens wide to its fullest, revealing a moist, pink interior and a barbed tongue. Two massive, ivory-textured canines gleam with a deadly coldness in the light, surrounded by smaller teeth of equal razor-sharpness. Its icy blue eyes narrow into murderous slits of intense rage, and its wiry white beard stretches forward with intense tension. With a silent yet powerful roar, emanating from its chest, the muscles of its neck and shoulders bulge, the force of which seems to shake the very air, and a stream of glistening saliva scatters from between its teeth. The background is a completely out-of-focus, deep and humid green jungle, creating a primitive, awe-inspiring sense of wild power and extremely tense oppression.", + "seed": 7528, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_1/index_8578_seed_7528_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05322200618684292, + "vlm_logprob": -0.05322200618684292, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.05322200618684292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007468162104487419 + ], + "opposite_logprob_tokens": [ + -5.882468223571777 + ], + "logprob_reward": -0.007468162104487419, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.007468162104487419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06712887436151505, + 0.12132821977138519, + 0.054038155823946, + 0.16765695810317993, + 0.22321143746376038, + 0.04251084849238396, + 0.05585680156946182, + 0.02959330566227436, + 0.09681747108697891, + 0.14185790717601776 + ], + "credit_map": [ + 0.01139142643660307, + 0.006853936705738306, + 0.012475024908781052, + 0.007242106832563877, + 0.008753121830523014, + 0.005653093568980694, + 0.0068078055046498775, + 0.007952360436320305, + 0.02589254267513752, + 0.011528612114489079, + 0.026566272601485252, + 0.011781004257500172, + 0.013170155696570873, + 0.009503967128694057, + 0.010809340514242649, + 0.012076327577233315, + 0.010298753157258034, + 0.007902372628450394, + 0.005590979941189289, + 0.006599363870918751, + 0.007259794045239687, + 0.003318159142509103, + 0.00915629044175148, + 0.00391243863850832, + 0.01727110706269741, + 0.028706975281238556, + 0.01622644253075123, + 0.01831093430519104, + 0.029412513598799706, + 0.020045656710863113, + 0.01779390685260296, + 0.01988942176103592, + 0.033276643604040146, + 0.03153038024902344, + 0.030182482674717903, + 0.03496013209223747, + 0.030440036207437515, + 0.016716834157705307, + 0.025843773037195206, + 0.02026115544140339, + 0.0055623529478907585, + 0.004289030563086271, + 0.006241771392524242, + 0.005347891710698605, + 0.0047103650867938995, + 0.0037828870117664337, + 0.006735088303685188, + 0.005841460078954697, + 0.006234981119632721, + 0.008969924412667751, + 0.004448218736797571, + 0.008874560706317425, + 0.007546298671513796, + 0.004559580236673355, + 0.009120259433984756, + 0.006102978717535734, + 0.004600569140166044, + 0.003998153377324343, + 0.0036967333871871233, + 0.00429931003600359, + 0.003554154420271516, + 0.0023690122179687023, + 0.0040575237944722176, + 0.0030178476590663195, + 0.0075172726064920425, + 0.007683233357965946, + 0.03496682271361351, + 0.007171473000198603, + 0.005736460909247398, + 0.025621531531214714, + 0.0036562918685376644, + 0.004464380443096161, + 0.01996682956814766, + 0.007964672520756721, + 0.03435350954532623, + 0.01075135450810194, + 0.04491294175386429, + 0.0066363210789859295, + 0.012776863761246204, + 0.004495413042604923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.007468162104487419 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156730711460114 + ], + "opposite_logprob_tokens": [ + -1.70156729221344 + ], + "logprob_reward": -0.20156730711460114, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156730711460114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0795641541481018, + 0.06514257192611694, + 0.04480694234371185, + 0.17460376024246216, + 0.06791158020496368, + 0.09540148079395294, + 0.09459742903709412, + 0.04115113243460655, + 0.13454926013946533, + 0.2022717446088791 + ], + "credit_map": [ + 0.012741472572088242, + 0.0076010203920304775, + 0.019606998190283775, + 0.006659006234258413, + 0.011673316359519958, + 0.006500763352960348, + 0.007783190347254276, + 0.006998377852141857, + 0.013640115968883038, + 0.007804835215210915, + 0.014221659861505032, + 0.007589593064039946, + 0.0048868050798773766, + 0.0060067614540457726, + 0.006284632720053196, + 0.004708165302872658, + 0.008000371046364307, + 0.005837973207235336, + 0.0061868224292993546, + 0.005594741087406874, + 0.004680195823311806, + 0.003337889676913619, + 0.006381392013281584, + 0.004787548910826445, + 0.019022205844521523, + 0.029959913343191147, + 0.015603342093527317, + 0.020734364166855812, + 0.02771102637052536, + 0.021803585812449455, + 0.019171888008713722, + 0.020597416907548904, + 0.008468982763588428, + 0.011521308682858944, + 0.007762187626212835, + 0.009880236349999905, + 0.009694133885204792, + 0.005204176530241966, + 0.008313518948853016, + 0.007067025639116764, + 0.011942517012357712, + 0.015074252150952816, + 0.010577574372291565, + 0.012717347592115402, + 0.014897161163389683, + 0.006640810985118151, + 0.010916581377387047, + 0.012635229155421257, + 0.007093111053109169, + 0.015057478100061417, + 0.008866599760949612, + 0.02373887225985527, + 0.007813537493348122, + 0.006329341791570187, + 0.018733840435743332, + 0.006964633706957102, + 0.0055112652480602264, + 0.005009182263165712, + 0.005170869175344706, + 0.006413501221686602, + 0.004661223851144314, + 0.003401107620447874, + 0.007439522538334131, + 0.003544455859810114, + 0.011387736536562443, + 0.008020089007914066, + 0.04543010890483856, + 0.009430882520973682, + 0.007763989735394716, + 0.042028289288282394, + 0.004451042041182518, + 0.006037105806171894, + 0.03370634466409683, + 0.007803167216479778, + 0.03262880817055702, + 0.007235872559249401, + 0.07611355185508728, + 0.00822142232209444, + 0.02048955298960209, + 0.016073007136583328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.20156730711460114 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819128356873989 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.01819128356873989, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819128356873989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0471317358314991, + 0.019197843968868256, + 0.07494296878576279, + 0.11702622473239899, + 0.025314124301075935, + 0.030734531581401825, + 0.10494749993085861, + 0.021661438047885895, + 0.48463505506515503, + 0.07440858334302902 + ], + "credit_map": [ + 0.0072034322656691074, + 0.004972211550921202, + 0.00786073599010706, + 0.00610749889165163, + 0.006326457019895315, + 0.0038650829810649157, + 0.005825011059641838, + 0.004971309565007687, + 0.004667611327022314, + 0.0020281097386032343, + 0.00300388946197927, + 0.0022243435960263014, + 0.0018703152891248465, + 0.0020776528399437666, + 0.0018358755623921752, + 0.0014900458045303822, + 0.010749582201242447, + 0.01080307923257351, + 0.008777416311204433, + 0.008345087058842182, + 0.0068116323091089725, + 0.003959546331316233, + 0.019837459549307823, + 0.005659162066876888, + 0.012488373555243015, + 0.020072802901268005, + 0.010151391848921776, + 0.017210446298122406, + 0.016522090882062912, + 0.008538499474525452, + 0.02224312163889408, + 0.009799499064683914, + 0.0031459510792046785, + 0.004091049078851938, + 0.0020928310696035624, + 0.003599828574806452, + 0.0034839557483792305, + 0.0020006869453936815, + 0.002757911803200841, + 0.004141912795603275, + 0.0030228151008486748, + 0.0041073886677622795, + 0.0035139992833137512, + 0.00414392352104187, + 0.0042665088549256325, + 0.0021937116980552673, + 0.005954023916274309, + 0.0035321610048413277, + 0.011019222438335419, + 0.01508190669119358, + 0.010293981991708279, + 0.024942811578512192, + 0.009385786019265652, + 0.0055299014784395695, + 0.01833195798099041, + 0.010361934080719948, + 0.0026640030555427074, + 0.0038295623380690813, + 0.0030439626425504684, + 0.00225706584751606, + 0.0031051631085574627, + 0.001886657439172268, + 0.0027857760433107615, + 0.002089246641844511, + 0.03324661776423454, + 0.028908047825098038, + 0.18782919645309448, + 0.03478534519672394, + 0.024076206609606743, + 0.14083173871040344, + 0.013320468366146088, + 0.021637430414557457, + 0.010117671452462673, + 0.0035992921330034733, + 0.02127356268465519, + 0.004342752043157816, + 0.023316264152526855, + 0.003057608613744378, + 0.005465764086693525, + 0.003235666546970606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.01819128356873989 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348749175667763 + ], + "opposite_logprob_tokens": [ + -3.651348829269409 + ], + "logprob_reward": -0.026348749175667763, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348749175667763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06337350606918335, + 0.05588267371058464, + 0.06930788606405258, + 0.13633139431476593, + 0.10314512252807617, + 0.10803212970495224, + 0.1143067479133606, + 0.04804641753435135, + 0.15054936707019806, + 0.151024729013443 + ], + "credit_map": [ + 0.009550623595714569, + 0.0066194068640470505, + 0.010509383864700794, + 0.006687805987894535, + 0.00964572373777628, + 0.0050666965544223785, + 0.007216057274490595, + 0.008077808655798435, + 0.0072129275649785995, + 0.006928826682269573, + 0.006170193199068308, + 0.008450624532997608, + 0.006238560192286968, + 0.004689666908234358, + 0.012723798863589764, + 0.0034680769313126802, + 0.008929785341024399, + 0.010579322464764118, + 0.007190977223217487, + 0.007753243204206228, + 0.008335922844707966, + 0.005421044770628214, + 0.01307469978928566, + 0.008022882975637913, + 0.012716496363282204, + 0.03289667144417763, + 0.010603617876768112, + 0.022504912689328194, + 0.021515775471925735, + 0.009401299059391022, + 0.017944350838661194, + 0.008748278953135014, + 0.013517331331968307, + 0.018020857125520706, + 0.011790032498538494, + 0.015413816086947918, + 0.014499709941446781, + 0.008084961213171482, + 0.011617296375334263, + 0.010201117023825645, + 0.010002590715885162, + 0.016148744150996208, + 0.011350873857736588, + 0.015491587109863758, + 0.013781271874904633, + 0.007855940610170364, + 0.019532332196831703, + 0.013868783600628376, + 0.012532445602118969, + 0.017487088218331337, + 0.008944566361606121, + 0.022968260571360588, + 0.013395310379564762, + 0.010342815890908241, + 0.01914391480386257, + 0.009492341428995132, + 0.007197114638984203, + 0.007478621788322926, + 0.0046684048138558865, + 0.007376287132501602, + 0.004963408224284649, + 0.004419920966029167, + 0.006832090672105551, + 0.00511056836694479, + 0.014192258007824421, + 0.017687391489744186, + 0.03904739394783974, + 0.0126466890797019, + 0.011556011624634266, + 0.03610401973128319, + 0.009830604307353497, + 0.009485011920332909, + 0.01669047586619854, + 0.0079332385212183, + 0.04420352727174759, + 0.01159864105284214, + 0.03809503838419914, + 0.006976062431931496, + 0.013410087674856186, + 0.012117667123675346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.026348749175667763 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534528970718384 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534528970718384, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534528970718384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06504223495721817, + 0.1475069522857666, + 0.11581753939390182, + 0.027144573628902435, + 0.04863414540886879, + 0.035260073840618134, + 0.022393520921468735, + 0.034596048295497894, + 0.0800074115395546, + 0.4235974848270416 + ], + "credit_map": [ + 0.00893132109194994, + 0.007264315616339445, + 0.011726870201528072, + 0.007197782397270203, + 0.009867435321211815, + 0.005665266886353493, + 0.006323597859591246, + 0.008065649308264256, + 0.018715813755989075, + 0.029706766828894615, + 0.018742991611361504, + 0.02264370769262314, + 0.015974264591932297, + 0.014910142868757248, + 0.016016516834497452, + 0.010796760208904743, + 0.01999690942466259, + 0.012875487096607685, + 0.013903025537729263, + 0.017196321859955788, + 0.009088252671062946, + 0.008789694868028164, + 0.02103959582746029, + 0.012928257696330547, + 0.0026533303316682577, + 0.004422048572450876, + 0.0020357188768684864, + 0.004172259941697121, + 0.004144931677728891, + 0.0030621185433119535, + 0.00330494181253016, + 0.0033492252696305513, + 0.006570659577846527, + 0.006489336956292391, + 0.006146267522126436, + 0.005174535326659679, + 0.006005700211971998, + 0.004168076906353235, + 0.005330306012183428, + 0.008749266155064106, + 0.005677617620676756, + 0.004385882057249546, + 0.004311884753406048, + 0.0038008466362953186, + 0.006665646564215422, + 0.0032766673248261213, + 0.004329254385083914, + 0.002812277991324663, + 0.001652904786169529, + 0.0029137434903532267, + 0.0024285661056637764, + 0.003958271816372871, + 0.0023525438737124205, + 0.0018659149063751101, + 0.003093292936682701, + 0.004128285218030214, + 0.0045066894963383675, + 0.005427179392427206, + 0.0029559799004346132, + 0.005281468387693167, + 0.0035318683367222548, + 0.003558180760592222, + 0.00642017787322402, + 0.0029145057778805494, + 0.007886332459747791, + 0.00606977753341198, + 0.022054867818951607, + 0.0061638131737709045, + 0.005408298224210739, + 0.025610366836190224, + 0.003100927686318755, + 0.0037130306009203196, + 0.07085617631673813, + 0.014064308255910873, + 0.07998376339673996, + 0.016728416085243225, + 0.159694105386734, + 0.016352513805031776, + 0.03881775215268135, + 0.027100451290607452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.012534528970718384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low, upward angle, the camera slowly and steadily moves forward, locking the viewer's gaze squarely on the face of a majestic, adult white tiger. The beast's snow-white fur is dotted with barely perceptible pale gray streaks, and in the dappled light filtering through the gloomy forest, each individual hair appears distinctly textured. Facing the camera, its muscular maw opens wide to its fullest, revealing a moist, pink interior and a barbed tongue. Two massive, ivory-textured canines gleam with a deadly coldness in the light, surrounded by smaller teeth of equal razor-sharpness. Its icy blue eyes narrow into murderous slits of intense rage, and its wiry white beard stretches forward with intense tension. With a silent yet powerful roar, emanating from its chest, the muscles of its neck and shoulders bulge, the force of which seems to shake the very air, and a stream of glistening saliva scatters from between its teeth. The background is a completely out-of-focus, deep and humid green jungle, creating a primitive, awe-inspiring sense of wild power and extremely tense oppression.", + "seed": 7529, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_1/index_8578_seed_7529_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12701771911233664, + "vlm_logprob": -0.12701771911233664, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.12701771911233664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006767448503524065 + ], + "opposite_logprob_tokens": [ + -7.131767272949219 + ], + "logprob_reward": -0.006767448503524065, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.006767448503524065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01870737038552761, + 0.1135735809803009, + 0.3743876516819, + 0.013667816296219826, + 0.017309749498963356, + 0.04887323081493378, + 0.07297360152006149, + 0.015927396714687347, + 0.08221345394849777, + 0.24236610531806946 + ], + "credit_map": [ + 0.003801619401201606, + 0.0010725281899794936, + 0.004602788016200066, + 0.0021452223882079124, + 0.0016891293926164508, + 0.0014641564339399338, + 0.0019114093156531453, + 0.0020205180626362562, + 0.030353030189871788, + 0.009684552438557148, + 0.01183358021080494, + 0.0073099154978990555, + 0.029591849073767662, + 0.011280244216322899, + 0.008663070388138294, + 0.00485734548419714, + 0.0707225650548935, + 0.020453589037060738, + 0.036763809621334076, + 0.08645334094762802, + 0.013179787434637547, + 0.02481810189783573, + 0.014749884605407715, + 0.10724660754203796, + 0.00187182507943362, + 0.002437615068629384, + 0.0014548467006534338, + 0.0013257804093882442, + 0.0021094384137541056, + 0.0012754867784678936, + 0.001304860576055944, + 0.0018879640847444534, + 0.002623628592118621, + 0.0017154753440991044, + 0.0022944225929677486, + 0.002569653792306781, + 0.0017174085369333625, + 0.0020816917531192303, + 0.002175242407247424, + 0.0021322285756468773, + 0.004088832996785641, + 0.003249478992074728, + 0.012902775779366493, + 0.0033227542880922556, + 0.0028909998945891857, + 0.007076173089444637, + 0.0023412820883095264, + 0.013000943697988987, + 0.01180747989565134, + 0.006033525802195072, + 0.009372780099511147, + 0.009811379946768284, + 0.007849755696952343, + 0.009045448154211044, + 0.011569839902222157, + 0.007483393419533968, + 0.0011383877135813236, + 0.001414161641150713, + 0.001331126200966537, + 0.001570759224705398, + 0.0016790680820122361, + 0.004331797827035189, + 0.0015689352294430137, + 0.0028931605629622936, + 0.00575526338070631, + 0.006520879454910755, + 0.005985208787024021, + 0.004421165678650141, + 0.006351591553539038, + 0.03111962229013443, + 0.005523835774511099, + 0.016535887494683266, + 0.08925706148147583, + 0.010073881596326828, + 0.014925064519047737, + 0.011540967039763927, + 0.006745191756635904, + 0.01592530496418476, + 0.00871914904564619, + 0.08517950028181076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.006767448503524065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761404633522034 + ], + "opposite_logprob_tokens": [ + -0.8261404633522034 + ], + "logprob_reward": -0.5761404633522034, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761404633522034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009888675063848495, + 0.130113884806633, + 0.5272791385650635, + 0.027194730937480927, + 0.013007472269237041, + 0.07296261191368103, + 0.036588799208402634, + 0.030939610674977303, + 0.1042051911354065, + 0.047819834202528 + ], + "credit_map": [ + 0.0014205483021214604, + 0.0005671079852618277, + 0.0020507541485130787, + 0.0011989664053544402, + 0.0012420094572007656, + 0.0006739181699231267, + 0.00160612678155303, + 0.0011292436392977834, + 0.03590898960828781, + 0.011402999050915241, + 0.013352668844163418, + 0.0069700260646641254, + 0.033102136105298996, + 0.012770475819706917, + 0.01052788458764553, + 0.006078716367483139, + 0.0942084863781929, + 0.03171558678150177, + 0.051911987364292145, + 0.12153084576129913, + 0.028799433261156082, + 0.034302715212106705, + 0.021597903221845627, + 0.14321216940879822, + 0.002784386510029435, + 0.00373444682918489, + 0.0023964201100170612, + 0.00426209531724453, + 0.004168559797108173, + 0.0030975823756307364, + 0.0030150029342621565, + 0.0037362370640039444, + 0.001040155068039894, + 0.0016264563892036676, + 0.001825125073082745, + 0.0026277860160917044, + 0.0017432071035727859, + 0.0014242875622585416, + 0.0013502975925803185, + 0.0013701568823307753, + 0.01111690979450941, + 0.006012789439409971, + 0.007549276575446129, + 0.006898378487676382, + 0.006809355691075325, + 0.019847877323627472, + 0.0053787389770150185, + 0.009349281899631023, + 0.007041211239993572, + 0.0033335948828607798, + 0.0037584900856018066, + 0.0044203102588653564, + 0.003567357547581196, + 0.006832887884229422, + 0.0033213484566658735, + 0.004313599783927202, + 0.002812717342749238, + 0.0029697685968130827, + 0.004151100758463144, + 0.004019379615783691, + 0.0036115157417953014, + 0.002643575193360448, + 0.004588040988892317, + 0.006143511738628149, + 0.007352057378739119, + 0.008068090304732323, + 0.029434481635689735, + 0.006337576080113649, + 0.009147929027676582, + 0.022485623136162758, + 0.005613324698060751, + 0.01576610654592514, + 0.017477454617619514, + 0.0021896965336054564, + 0.0029693436808884144, + 0.0024488363415002823, + 0.0015269016148522496, + 0.0030065136961638927, + 0.001886459649540484, + 0.016314629465341568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.5761404633522034 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347937062382698 + ], + "opposite_logprob_tokens": [ + -3.6513478755950928 + ], + "logprob_reward": -0.026347937062382698, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347937062382698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017964530736207962, + 0.07243963330984116, + 0.335116446018219, + 0.033569350838661194, + 0.018169471994042397, + 0.07831832021474838, + 0.04398373141884804, + 0.015405744314193726, + 0.25361552834510803, + 0.1314171999692917 + ], + "credit_map": [ + 0.0030701342038810253, + 0.0014092240016907454, + 0.002705940743908286, + 0.0030296624172478914, + 0.001680118846707046, + 0.0014039566740393639, + 0.002936078468337655, + 0.0017294154968112707, + 0.01920342817902565, + 0.0062807281501591206, + 0.00717479782178998, + 0.005884639918804169, + 0.017824294045567513, + 0.007382902782410383, + 0.0057185739278793335, + 0.0029702719766646624, + 0.06077269837260246, + 0.02025233767926693, + 0.033428389579057693, + 0.07766696065664291, + 0.016732990741729736, + 0.021921774372458458, + 0.013683811761438847, + 0.09065748006105423, + 0.0029041622765362263, + 0.003067671786993742, + 0.0037458601873368025, + 0.0033058938570320606, + 0.003843142418190837, + 0.007047576364129782, + 0.003172454657033086, + 0.006482589058578014, + 0.001548471045680344, + 0.0018845460144802928, + 0.0027301586233079433, + 0.0029911284800618887, + 0.002454254776239395, + 0.0022696061059832573, + 0.0013733567902818322, + 0.002917950041592121, + 0.0056059998460114, + 0.003921807743608952, + 0.025398677214980125, + 0.005058564245700836, + 0.0029867098201066256, + 0.007599352393299341, + 0.0032359049655497074, + 0.0245113056153059, + 0.0036409618332982063, + 0.003259720280766487, + 0.0038151496555656195, + 0.010120256803929806, + 0.0032417287584394217, + 0.004729186184704304, + 0.011979028582572937, + 0.003197699086740613, + 0.0014820218784734607, + 0.001115247723646462, + 0.0015086354687809944, + 0.0021150007378309965, + 0.0014173564268276095, + 0.003196189645677805, + 0.0022640444803982973, + 0.002307248068973422, + 0.017198974266648293, + 0.017658092081546783, + 0.011319911107420921, + 0.011742881499230862, + 0.013675056397914886, + 0.09914296120405197, + 0.012821095064282417, + 0.07005655765533447, + 0.048865657299757004, + 0.005771363619714975, + 0.00830611027777195, + 0.006239994894713163, + 0.003807113040238619, + 0.008451702073216438, + 0.004589451942592859, + 0.04538580775260925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.026347937062382698 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584525540471077 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.020584525540471077, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584525540471077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01494480948895216, + 0.066847063601017, + 0.42133834958076477, + 0.025985388085246086, + 0.018292026594281197, + 0.07248225808143616, + 0.04239882156252861, + 0.02553139254450798, + 0.15667296946048737, + 0.1555069535970688 + ], + "credit_map": [ + 0.002486994955688715, + 0.0011347366962581873, + 0.0026365702506154776, + 0.0019479282200336456, + 0.0011131087085232139, + 0.0014501751866191626, + 0.0023982205893844366, + 0.001777074416168034, + 0.018452443182468414, + 0.005942132789641619, + 0.006669329479336739, + 0.003947535529732704, + 0.017187828198075294, + 0.0065331063233315945, + 0.005198931321501732, + 0.0029157614335417747, + 0.07721532881259918, + 0.02400028519332409, + 0.04251693934202194, + 0.09834130108356476, + 0.018454041332006454, + 0.02738562412559986, + 0.016786936670541763, + 0.11663787811994553, + 0.002180659445002675, + 0.0032149485778063536, + 0.0021404605358839035, + 0.0031514675356447697, + 0.0041073886677622795, + 0.00439597200602293, + 0.0026739949826151133, + 0.004120497964322567, + 0.001234638039022684, + 0.0016182875260710716, + 0.0025582751259207726, + 0.003306036815047264, + 0.0015981316100805998, + 0.0030050647910684347, + 0.0019270622869953513, + 0.0030445298179984093, + 0.009750453755259514, + 0.006639017723500729, + 0.007202759850770235, + 0.007109863683581352, + 0.007990510202944279, + 0.01820257492363453, + 0.005627383477985859, + 0.009959698654711246, + 0.005160989705473185, + 0.0030725565738976, + 0.0029169078916311264, + 0.00879601202905178, + 0.0048637790605425835, + 0.004369877744466066, + 0.010001346468925476, + 0.003217351855710149, + 0.0018705630209296942, + 0.0024047186598181725, + 0.0019382977625355124, + 0.0027266903780400753, + 0.0028417573776096106, + 0.006684813182801008, + 0.0026700724847614765, + 0.004394479561597109, + 0.010901341214776039, + 0.010557928122580051, + 0.006280847359448671, + 0.007053009234368801, + 0.006967402063310146, + 0.06542512029409409, + 0.008211610838770866, + 0.041275717318058014, + 0.057177044451236725, + 0.00660982821136713, + 0.009328911080956459, + 0.007524040527641773, + 0.004335291218012571, + 0.010207466781139374, + 0.005535057745873928, + 0.05478932335972786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.020584525540471077 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an adult white tiger facing the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low, upward angle moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger open its maw wide to reveal a pink interior, barbed tongue, and gleaming ivory canines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the tiger's eyes narrowed into slits while its neck and shoulder muscles bulge as if roaring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, deep green jungle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248221103101969 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005248221103101969, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248221103101969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026085857301950455, + 0.08866006880998611, + 0.12217225134372711, + 0.060659147799015045, + 0.014534300193190575, + 0.06716857105493546, + 0.07532144337892532, + 0.06486187130212784, + 0.1082414910197258, + 0.37229499220848083 + ], + "credit_map": [ + 0.0047692181542515755, + 0.001620918745175004, + 0.006682738661766052, + 0.0030820793472230434, + 0.0017982167191803455, + 0.00216309423558414, + 0.00256926822476089, + 0.003400323214009404, + 0.023817161098122597, + 0.007763441186398268, + 0.009424486197531223, + 0.005622992757707834, + 0.021625978872179985, + 0.008336495608091354, + 0.008532959036529064, + 0.0035365549847483635, + 0.01920902729034424, + 0.0075433021411299706, + 0.011417598463594913, + 0.03084646910429001, + 0.00876211654394865, + 0.007001430727541447, + 0.005425978451967239, + 0.031966328620910645, + 0.004823404364287853, + 0.004144613165408373, + 0.007189927157014608, + 0.004662740044295788, + 0.005741564556956291, + 0.018406398594379425, + 0.003585511352866888, + 0.012104989029467106, + 0.0011738932225853205, + 0.0014418864157050848, + 0.0023976045195013285, + 0.0029264348559081554, + 0.0014129264745861292, + 0.0015602803323417902, + 0.0014506744919344783, + 0.002170599764212966, + 0.010633900761604309, + 0.007266850210726261, + 0.009206853806972504, + 0.004770202562212944, + 0.005505725275725126, + 0.014308028854429722, + 0.00506474357098341, + 0.010412267409265041, + 0.010855521075427532, + 0.0048351832665503025, + 0.006148441229015589, + 0.013472020626068115, + 0.005597839131951332, + 0.010511755011975765, + 0.017821939662098885, + 0.006078746635466814, + 0.0029413094744086266, + 0.003111413214355707, + 0.006693494971841574, + 0.003597419476136565, + 0.003331858664751053, + 0.019517479464411736, + 0.0036596618592739105, + 0.022009234875440598, + 0.007777119521051645, + 0.007504441309720278, + 0.007366012781858444, + 0.005189929157495499, + 0.0051628281362354755, + 0.04510285705327988, + 0.005234625190496445, + 0.024903679266572, + 0.13004787266254425, + 0.015450231730937958, + 0.0214939434081316, + 0.018318522721529007, + 0.0114498445764184, + 0.029281221330165863, + 0.013970639556646347, + 0.13228271901607513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.005248221103101969 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_10/index_tensor([7152, 7152])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_10/index_tensor([7152, 7152])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1f52c9a1687a91f6761aa241539a618bddde67aa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_10/index_tensor([7152, 7152])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a surreal, high-definition digital aesthetic, the film begins with a slow, graceful, doll-like movement, sliding horizontally across the magnificent, ornate facade of a sprawling limestone mansion. Soft, golden morning light casts long shadows across the intricate carvings and towering pillars, rendering every detail with crisp 4K clarity. As the camera continues its steady, uninterrupted path, it perfectly aligns with a spotless, large bay window, revealing a brightly lit, luxurious bathroom. Through the pristine glass, a woman with dark black hair pulled back into a sleek, elegant bun can be seen leaning against a large, ornate mirror. She wears a lustrous, blush-tinged silk robe, the fabric striking in its allure. She meticulously shapes her eyebrows with precise, deliberate movements, her expression intense and unwavering. In her right hand, she holds a pair of pointed metal tweezers, while with her left, she gently pinches the skin above her eyes. The soft morning light from the window and the warm, even glow from the bulb in the mirror-framed dressing table combine to illuminate the scene, creating a focused halo around her face, while the luxurious surroundings of white marble countertops, gleaming gilded fixtures, and crystal accents fade into the soft-focus background, establishing an atmosphere of quiet, personal ritual, and focused perfection.", + "seed": 29972, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_10/index_7152_seed_29972_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06495748572051525, + "vlm_logprob": -0.06495748572051525, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.06495748572051525, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0745200514793396 + ], + "opposite_logprob_tokens": [ + -2.8245201110839844 + ], + "logprob_reward": -0.0745200514793396, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0745200514793396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34438401460647583, + 0.24508537352085114, + 0.09871436655521393, + 0.044638413935899734, + 0.098673015832901, + 0.03839820623397827, + 0.05100468918681145, + 0.02987990342080593, + 0.041860468685626984, + 0.007361581549048424 + ], + "credit_map": [ + 0.01808389090001583, + 0.10405507683753967, + 0.01459505409002304, + 0.02418859489262104, + 0.059366416186094284, + 0.02636725641787052, + 0.024361154064536095, + 0.07336656749248505, + 0.06160903722047806, + 0.014315922744572163, + 0.07925290614366531, + 0.01357963401824236, + 0.01548161543905735, + 0.01948676072061062, + 0.02319173887372017, + 0.01816774532198906, + 0.019174497574567795, + 0.016352983191609383, + 0.011740704998373985, + 0.019181987270712852, + 0.009671200066804886, + 0.005263819359242916, + 0.010354644618928432, + 0.006974523421376944, + 0.004102024249732494, + 0.0028724349103868008, + 0.01179027184844017, + 0.003998386673629284, + 0.005501291714608669, + 0.00371003826148808, + 0.009454963728785515, + 0.003209003247320652, + 0.006072262767702341, + 0.00537329725921154, + 0.006457616109400988, + 0.006761680822819471, + 0.005574468988925219, + 0.029468543827533722, + 0.006543139927089214, + 0.032422006130218506, + 0.002737601287662983, + 0.00539526017382741, + 0.004341123625636101, + 0.00577218784019351, + 0.0039971452206373215, + 0.003934095613658428, + 0.0058584194630384445, + 0.006362373940646648, + 0.004914004355669022, + 0.007175779901444912, + 0.0035455217584967613, + 0.0067634847946465015, + 0.004458435345441103, + 0.007156666833907366, + 0.009394112974405289, + 0.00759668555110693, + 0.0028526796959340572, + 0.0028351980727165937, + 0.0015360736288130283, + 0.002796764252707362, + 0.005400351248681545, + 0.004677158314734697, + 0.004992700181901455, + 0.004788977559655905, + 0.006201781332492828, + 0.00318533880636096, + 0.0038289418444037437, + 0.006226204335689545, + 0.004393070470541716, + 0.00573168508708477, + 0.005720812361687422, + 0.006572634447365999, + 0.001351026352494955, + 0.0006564476061612368, + 0.0008384068496525288, + 0.0009162400383502245, + 0.0006095397984609008, + 0.0008242408512160182, + 0.0011565147433429956, + 0.0010091654257848859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0745200514793396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560310773551464 + ], + "opposite_logprob_tokens": [ + -4.3875603675842285 + ], + "logprob_reward": -0.012560310773551464, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560310773551464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46513208746910095, + 0.13578364253044128, + 0.05452846735715866, + 0.041215069591999054, + 0.12137416005134583, + 0.02554892748594284, + 0.03145620599389076, + 0.051958803087472916, + 0.053868453949689865, + 0.0191341545432806 + ], + "credit_map": [ + 0.02328474447131157, + 0.13368085026741028, + 0.020353931933641434, + 0.03201425075531006, + 0.08158077299594879, + 0.029167648404836655, + 0.03765184432268143, + 0.10739804059267044, + 0.03868393227458, + 0.005523475352674723, + 0.0498046912252903, + 0.008971042931079865, + 0.007151618134230375, + 0.006280971225351095, + 0.011076094582676888, + 0.00829182006418705, + 0.013761329464614391, + 0.009127591736614704, + 0.00701164873316884, + 0.007901923730969429, + 0.004624634515494108, + 0.0026555450167506933, + 0.005261637270450592, + 0.0041841561906039715, + 0.004541260190308094, + 0.0026044202968478203, + 0.009384449571371078, + 0.004112670198082924, + 0.006520713679492474, + 0.0029677492566406727, + 0.008382669650018215, + 0.002701136749237776, + 0.007318200543522835, + 0.00876497570425272, + 0.009640328586101532, + 0.010249288752675056, + 0.0068388208746910095, + 0.037263039499521255, + 0.006739966571331024, + 0.03455953299999237, + 0.00207542791031301, + 0.002486808691173792, + 0.003878346411511302, + 0.0039731357246637344, + 0.0025679897516965866, + 0.0028123578522354364, + 0.004348693415522575, + 0.0034061698243021965, + 0.0035341093316674232, + 0.003046466503292322, + 0.0025476962327957153, + 0.004951901733875275, + 0.003188776783645153, + 0.005414419341832399, + 0.0033924283925443888, + 0.005380408372730017, + 0.009958500042557716, + 0.0039092181250452995, + 0.004732960369437933, + 0.008603250607848167, + 0.00433840649202466, + 0.008045369759202003, + 0.00722731975838542, + 0.005143775139003992, + 0.008626885712146759, + 0.005201346706598997, + 0.005553761031478643, + 0.008729997090995312, + 0.004700787365436554, + 0.006569753400981426, + 0.007225877605378628, + 0.0072600459679961205, + 0.002230081707239151, + 0.0020128998439759016, + 0.0017030241433531046, + 0.002277008956298232, + 0.002183973789215088, + 0.0021795034408569336, + 0.0026684990152716637, + 0.0038791638799011707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.012560310773551464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635316178202629 + ], + "opposite_logprob_tokens": [ + -3.651353120803833 + ], + "logprob_reward": -0.02635316178202629, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635316178202629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19688096642494202, + 0.20088447630405426, + 0.08708693832159042, + 0.04065774381160736, + 0.06113184988498688, + 0.04367237538099289, + 0.1264830231666565, + 0.07654339075088501, + 0.11751928925514221, + 0.04913993179798126 + ], + "credit_map": [ + 0.018746916204690933, + 0.02780989184975624, + 0.022146375849843025, + 0.019958361983299255, + 0.013334563001990318, + 0.01302428264170885, + 0.021995700895786285, + 0.05986488237977028, + 0.04680408537387848, + 0.013809607364237309, + 0.06332425773143768, + 0.013781962916254997, + 0.013890075497329235, + 0.01632714457809925, + 0.01389414444565773, + 0.019053209573030472, + 0.017665956169366837, + 0.015269233845174313, + 0.01146737765520811, + 0.013297990895807743, + 0.00792258232831955, + 0.0048130229115486145, + 0.009031997993588448, + 0.007618782576173544, + 0.0049016582779586315, + 0.003407602198421955, + 0.008163641206920147, + 0.003535811323672533, + 0.006359967868775129, + 0.0048000747337937355, + 0.0053769261576235294, + 0.004112063907086849, + 0.006189621053636074, + 0.004718281794339418, + 0.004943980369716883, + 0.005030129104852676, + 0.004525595810264349, + 0.015417887829244137, + 0.004528638906776905, + 0.015777720138430595, + 0.0040559545159339905, + 0.004734004382044077, + 0.008488458581268787, + 0.007713768165558577, + 0.0030618251767009497, + 0.003642125753685832, + 0.00552896736189723, + 0.00644727423787117, + 0.011051313020288944, + 0.01550739910453558, + 0.007928126491606236, + 0.01789494976401329, + 0.010633697733283043, + 0.01575371064245701, + 0.02325412444770336, + 0.024459712207317352, + 0.011523782275617123, + 0.008250497281551361, + 0.005432278383523226, + 0.008132639341056347, + 0.008498000912368298, + 0.006690040230751038, + 0.01693549007177353, + 0.011080670170485973, + 0.012775519862771034, + 0.01899975724518299, + 0.006363254506140947, + 0.010028409771621227, + 0.024014491587877274, + 0.010153014212846756, + 0.017446843907237053, + 0.01773800700902939, + 0.0063326130621135235, + 0.005888408049941063, + 0.0032242080196738243, + 0.006258185021579266, + 0.005507141351699829, + 0.006575718987733126, + 0.006866323295980692, + 0.008487336337566376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.02635316178202629 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015729695558548 + ], + "opposite_logprob_tokens": [ + -1.7015730142593384 + ], + "logprob_reward": -0.2015729695558548, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015729695558548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47975119948387146, + 0.051204461604356766, + 0.0499705970287323, + 0.0342702642083168, + 0.07373588532209396, + 0.06945732235908508, + 0.08463980257511139, + 0.054842881858348846, + 0.05003095418214798, + 0.05209660157561302 + ], + "credit_map": [ + 0.02854396402835846, + 0.1090783029794693, + 0.025841427966952324, + 0.04035342112183571, + 0.056218940764665604, + 0.04077768698334694, + 0.048354677855968475, + 0.1305827796459198, + 0.005906874313950539, + 0.005196446552872658, + 0.007685249205678701, + 0.003969203680753708, + 0.005598038900643587, + 0.008687393739819527, + 0.00796132255345583, + 0.006199932191520929, + 0.00969460979104042, + 0.00887351669371128, + 0.006506458390504122, + 0.008528480306267738, + 0.004492014180868864, + 0.003010628279298544, + 0.004812841769307852, + 0.004052045289427042, + 0.003596385708078742, + 0.002525793621316552, + 0.007025327533483505, + 0.0026129044126719236, + 0.006059350911527872, + 0.003986192401498556, + 0.005704882089048624, + 0.0027594275306910276, + 0.005780061706900597, + 0.004245790187269449, + 0.0062381490133702755, + 0.005670143291354179, + 0.006039167754352093, + 0.01923478953540325, + 0.0060037351213395596, + 0.0205240435898304, + 0.0064821671694517136, + 0.006884889677166939, + 0.012529059313237667, + 0.011309947818517685, + 0.0071337237022817135, + 0.0059465994127094746, + 0.009452364407479763, + 0.00971856527030468, + 0.006944840773940086, + 0.014474253170192242, + 0.004623610991984606, + 0.010919190011918545, + 0.009784911759197712, + 0.012272690422832966, + 0.012980842962861061, + 0.012639469467103481, + 0.005507562309503555, + 0.006178028881549835, + 0.005213564727455378, + 0.006957538425922394, + 0.009363306686282158, + 0.005132879130542278, + 0.008894055150449276, + 0.007595946546643972, + 0.006402837112545967, + 0.0079007213935256, + 0.0033525496255606413, + 0.005433555692434311, + 0.008566150441765785, + 0.004031318239867687, + 0.006270367186516523, + 0.008073453791439533, + 0.004970456473529339, + 0.007010398432612419, + 0.0034620168153196573, + 0.0058676255866885185, + 0.008653594180941582, + 0.0053941733203828335, + 0.006618884392082691, + 0.010119453072547913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.2015729695558548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780935011804104 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780935011804104, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780935011804104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29469069838523865, + 0.13001470267772675, + 0.05558690056204796, + 0.05331186205148697, + 0.1050601601600647, + 0.15973705053329468, + 0.08832492679357529, + 0.0281030535697937, + 0.030098779127001762, + 0.05507180839776993 + ], + "credit_map": [ + 0.03049878403544426, + 0.08524487167596817, + 0.03556002676486969, + 0.017035426571965218, + 0.0581994503736496, + 0.02600889466702938, + 0.014982691965997219, + 0.02716059796512127, + 0.024895036593079567, + 0.00965056661516428, + 0.034437116235494614, + 0.007804287131875753, + 0.011395771987736225, + 0.012008014135062695, + 0.0155702019110322, + 0.014253722503781319, + 0.010435030795633793, + 0.009455671533942223, + 0.005314039997756481, + 0.010787993669509888, + 0.005374027416110039, + 0.003515782067552209, + 0.005740700755268335, + 0.004963661078363657, + 0.006001809611916542, + 0.0033575515262782574, + 0.01368680503219366, + 0.004729113075882196, + 0.0074643706902861595, + 0.0054109664633870125, + 0.0068545350804924965, + 0.005806716624647379, + 0.008804475888609886, + 0.010128800757229328, + 0.00958425272256136, + 0.008039692416787148, + 0.007790570147335529, + 0.02753741294145584, + 0.008113397285342216, + 0.025061575695872307, + 0.011555666103959084, + 0.010851162485778332, + 0.015232003293931484, + 0.021571919322013855, + 0.023196788504719734, + 0.022227874025702477, + 0.027751171961426735, + 0.027350490912795067, + 0.007956482470035553, + 0.013131109066307545, + 0.00543845072388649, + 0.012296055443584919, + 0.007819157093763351, + 0.011463561095297337, + 0.014611711725592613, + 0.015608415938913822, + 0.003131464822217822, + 0.0033072810620069504, + 0.002378262812271714, + 0.0038924464024603367, + 0.0034479524474591017, + 0.0027713950257748365, + 0.004619057290256023, + 0.004555199295282364, + 0.0034958801697939634, + 0.004146662540733814, + 0.0016160201048478484, + 0.002922828309237957, + 0.00485715176910162, + 0.0030995996203273535, + 0.004413458984345198, + 0.005547181703150272, + 0.005722104571759701, + 0.004640065599232912, + 0.004878876730799675, + 0.008167414925992489, + 0.007799115497618914, + 0.006868139840662479, + 0.007245854940265417, + 0.00975024327635765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.009780935011804104 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal, high-definition digital aesthetic, the film begins with a slow, graceful, doll-like movement, sliding horizontally across the magnificent, ornate facade of a sprawling limestone mansion. Soft, golden morning light casts long shadows across the intricate carvings and towering pillars, rendering every detail with crisp 4K clarity. As the camera continues its steady, uninterrupted path, it perfectly aligns with a spotless, large bay window, revealing a brightly lit, luxurious bathroom. Through the pristine glass, a woman with dark black hair pulled back into a sleek, elegant bun can be seen leaning against a large, ornate mirror. She wears a lustrous, blush-tinged silk robe, the fabric striking in its allure. She meticulously shapes her eyebrows with precise, deliberate movements, her expression intense and unwavering. In her right hand, she holds a pair of pointed metal tweezers, while with her left, she gently pinches the skin above her eyes. The soft morning light from the window and the warm, even glow from the bulb in the mirror-framed dressing table combine to illuminate the scene, creating a focused halo around her face, while the luxurious surroundings of white marble countertops, gleaming gilded fixtures, and crystal accents fade into the soft-focus background, establishing an atmosphere of quiet, personal ritual, and focused perfection.", + "seed": 29973, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_10/index_7152_seed_29973_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.22224142141640185, + "vlm_logprob": -0.22224142141640185, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.22224142141640185, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8471941351890564 + ], + "opposite_logprob_tokens": [ + -0.5971941351890564 + ], + "logprob_reward": -0.8471941351890564, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8471941351890564, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13272427022457123, + 0.12683168053627014, + 0.03388460725545883, + 0.050682276487350464, + 0.2695464789867401, + 0.09668026119470596, + 0.17372387647628784, + 0.07877227663993835, + 0.021024903282523155, + 0.016129299998283386 + ], + "credit_map": [ + 0.010995118878781796, + 0.012627205811440945, + 0.018556205555796623, + 0.016301777213811874, + 0.018471097573637962, + 0.026676282286643982, + 0.012978410348296165, + 0.01611817628145218, + 0.013033583760261536, + 0.011278925463557243, + 0.026753151789307594, + 0.025043997913599014, + 0.011379566043615341, + 0.01188354380428791, + 0.01653221808373928, + 0.010926691815257072, + 0.003149919444695115, + 0.0037681027315557003, + 0.0060022794641554356, + 0.004368157591670752, + 0.0034675858914852142, + 0.0034923634957522154, + 0.0042974757961928844, + 0.005338725168257952, + 0.005608019419014454, + 0.0043898699805140495, + 0.006478275638073683, + 0.005079913884401321, + 0.004174551926553249, + 0.007705810014158487, + 0.00521136773750186, + 0.012034470215439796, + 0.04286932945251465, + 0.008931676857173443, + 0.09823929518461227, + 0.014061001129448414, + 0.012464236468076706, + 0.016272319480776787, + 0.011825704015791416, + 0.06488291919231415, + 0.01817258819937706, + 0.009451823309063911, + 0.011693897657096386, + 0.011140415444970131, + 0.00962910521775484, + 0.007436361629515886, + 0.017074624076485634, + 0.012081438675522804, + 0.01389994565397501, + 0.017718855291604996, + 0.03263690322637558, + 0.013920876197516918, + 0.011471179313957691, + 0.025101253762841225, + 0.014224082231521606, + 0.04475079104304314, + 0.012260415591299534, + 0.008789409883320332, + 0.011438491754233837, + 0.011874745599925518, + 0.007607908919453621, + 0.00719617260619998, + 0.009573129937052727, + 0.010031996294856071, + 0.0028108118567615747, + 0.0026748778764158487, + 0.0025477465242147446, + 0.0035162675194442272, + 0.0019764816388487816, + 0.0023564354050904512, + 0.002694439608603716, + 0.002447843551635742, + 0.002067495370283723, + 0.0025080034974962473, + 0.0012551246909424663, + 0.002130837645381689, + 0.0014750473201274872, + 0.0014029485173523426, + 0.002738239709287882, + 0.0025516024325042963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.8471941351890564 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297035962343216 + ], + "opposite_logprob_tokens": [ + -2.2379703521728516 + ], + "logprob_reward": -0.11297035962343216, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297035962343216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1384049355983734, + 0.03459058701992035, + 0.04006124660372734, + 0.040306732058525085, + 0.3840339183807373, + 0.06409662216901779, + 0.062386151403188705, + 0.08652417361736298, + 0.05375654995441437, + 0.09583905339241028 + ], + "credit_map": [ + 0.012472562491893768, + 0.024185089394450188, + 0.03150744363665581, + 0.016669433563947678, + 0.011850924231112003, + 0.015034455806016922, + 0.015381628647446632, + 0.011303405277431011, + 0.004153556190431118, + 0.003329634666442871, + 0.003944729454815388, + 0.0038555641658604145, + 0.005905667785555124, + 0.0026339495088905096, + 0.0062434677965939045, + 0.0045240214094519615, + 0.0030691104475408792, + 0.003796198172494769, + 0.005161054898053408, + 0.003801353508606553, + 0.004993378184735775, + 0.005601835902780294, + 0.004908356349915266, + 0.008729961700737476, + 0.0039760153740644455, + 0.0032902078237384558, + 0.00510385911911726, + 0.004191135987639427, + 0.004494687542319298, + 0.006078158970922232, + 0.004694866016507149, + 0.008477804251015186, + 0.0593748539686203, + 0.010716011747717857, + 0.14323249459266663, + 0.01893860474228859, + 0.019327998161315918, + 0.020294498652219772, + 0.017519619315862656, + 0.09462986141443253, + 0.011818651109933853, + 0.005971597041934729, + 0.007777136750519276, + 0.004911048337817192, + 0.006239334121346474, + 0.007416109554469585, + 0.00895824283361435, + 0.011004502885043621, + 0.005906590726226568, + 0.005273918621242046, + 0.012292644008994102, + 0.009677338413894176, + 0.004443958401679993, + 0.006188861094415188, + 0.007208499126136303, + 0.011394348926842213, + 0.01476630661636591, + 0.010020274668931961, + 0.011064707301557064, + 0.015048577450215816, + 0.007057316601276398, + 0.00792048778384924, + 0.00897857267409563, + 0.011667937971651554, + 0.007968644611537457, + 0.005718212574720383, + 0.006988601293414831, + 0.008756828494369984, + 0.0045929248444736, + 0.0073232161812484264, + 0.005961707793176174, + 0.006446420680731535, + 0.010637645609676838, + 0.014706009067595005, + 0.007170379161834717, + 0.013928900472819805, + 0.008019665256142616, + 0.008994833566248417, + 0.015348355285823345, + 0.017033284530043602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.11297035962343216 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200118370354176 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200118370354176, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200118370354176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18335817754268646, + 0.09282603859901428, + 0.031016556546092033, + 0.03885457664728165, + 0.12750636041164398, + 0.1598929911851883, + 0.08928285539150238, + 0.10827068239450455, + 0.0915566235780716, + 0.07743512839078903 + ], + "credit_map": [ + 0.018481817096471786, + 0.020554952323436737, + 0.022252801805734634, + 0.020354699343442917, + 0.026756778359413147, + 0.03468397632241249, + 0.02136649191379547, + 0.018906667828559875, + 0.007619346026331186, + 0.007788649294525385, + 0.018731677904725075, + 0.016378086060285568, + 0.009420952759683132, + 0.00886859092861414, + 0.01255547534674406, + 0.011463256552815437, + 0.002422475256025791, + 0.0020621458534151316, + 0.0054817539639770985, + 0.0026611429639160633, + 0.0038819988258183002, + 0.0033242793288081884, + 0.004639471881091595, + 0.006543287541717291, + 0.005041750147938728, + 0.0041971951723098755, + 0.005324725527316332, + 0.004089458379894495, + 0.003792643081396818, + 0.004726957064121962, + 0.003580865915864706, + 0.008100984618067741, + 0.014564394019544125, + 0.0073524597100913525, + 0.03145457059144974, + 0.014556166715919971, + 0.014354612678289413, + 0.009161699563264847, + 0.01639610342681408, + 0.01966635324060917, + 0.014663107693195343, + 0.013085871003568172, + 0.01679036393761635, + 0.01911897398531437, + 0.014342422597110271, + 0.026573702692985535, + 0.022464117035269737, + 0.032854434102773666, + 0.01007710862904787, + 0.00966759491711855, + 0.009254979901015759, + 0.013728341087698936, + 0.009706645272672176, + 0.009061871096491814, + 0.014723003841936588, + 0.013063309714198112, + 0.010086527094244957, + 0.01027583796530962, + 0.011210628785192966, + 0.014861985109746456, + 0.019760798662900925, + 0.007977128028869629, + 0.022649895399808884, + 0.011447879485785961, + 0.009947619400918484, + 0.010356791317462921, + 0.007047539111226797, + 0.010082005523145199, + 0.013736659660935402, + 0.009810137562453747, + 0.01820201799273491, + 0.012373858131468296, + 0.008603295311331749, + 0.00815664418041706, + 0.00599003816023469, + 0.013168816454708576, + 0.007351233158260584, + 0.00798504427075386, + 0.013082226738333702, + 0.013097837567329407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.014200118370354176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270628571510315 + ], + "opposite_logprob_tokens": [ + -2.1270627975463867 + ], + "logprob_reward": -0.1270628571510315, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270628571510315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09409258514642715, + 0.07983416318893433, + 0.048700962215662, + 0.02613547071814537, + 0.3109830617904663, + 0.06537995487451553, + 0.15822289884090424, + 0.13354070484638214, + 0.050158821046352386, + 0.03295138478279114 + ], + "credit_map": [ + 0.007951825857162476, + 0.017476079985499382, + 0.012068232521414757, + 0.014659560285508633, + 0.011061467230319977, + 0.01453774981200695, + 0.008914906531572342, + 0.007422770373523235, + 0.008091035299003124, + 0.006781288888305426, + 0.015581994317471981, + 0.01567472331225872, + 0.006427003536373377, + 0.0073104240000247955, + 0.010717920027673244, + 0.009249775670468807, + 0.004311494529247284, + 0.004934151656925678, + 0.007437908090651035, + 0.004859111271798611, + 0.004454249516129494, + 0.005717856343835592, + 0.007554430048912764, + 0.00943176168948412, + 0.0025960137136280537, + 0.0031847720965743065, + 0.002570047043263912, + 0.0028272056952118874, + 0.002816846128553152, + 0.0038726937491446733, + 0.0030560616869479418, + 0.005211828742176294, + 0.048450663685798645, + 0.008588466793298721, + 0.1184212937951088, + 0.010687761008739471, + 0.009734086692333221, + 0.01930420845746994, + 0.011799641884863377, + 0.08399693667888641, + 0.009625344537198544, + 0.006386942695826292, + 0.006986451800912619, + 0.0056746527552604675, + 0.00833914801478386, + 0.008158133365213871, + 0.011100304313004017, + 0.00910897832363844, + 0.01246146485209465, + 0.014878115616738796, + 0.028413526713848114, + 0.01515447162091732, + 0.011363889090716839, + 0.020947549492120743, + 0.015357987955212593, + 0.03964589536190033, + 0.015671277418732643, + 0.012447957880795002, + 0.02434273064136505, + 0.019638171419501305, + 0.011396314017474651, + 0.01635999232530594, + 0.018691230565309525, + 0.014993016608059406, + 0.007149957120418549, + 0.005165230482816696, + 0.0053770774975419044, + 0.006856691557914019, + 0.004816987551748753, + 0.006710680667310953, + 0.00717425299808383, + 0.006907944101840258, + 0.002765398472547531, + 0.003498814534395933, + 0.001949273282662034, + 0.005394332110881805, + 0.004776549991220236, + 0.0029053923208266497, + 0.0073999143205583096, + 0.004261708818376064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.1270628571510315 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977963674813509 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.00977963674813509, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977963674813509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12335513532161713, + 0.11361977458000183, + 0.03260546922683716, + 0.03443891555070877, + 0.3814404606819153, + 0.07728364318609238, + 0.07933460921049118, + 0.06616903096437454, + 0.03936822712421417, + 0.05238470062613487 + ], + "credit_map": [ + 0.011476606130599976, + 0.014836001209914684, + 0.01850268244743347, + 0.013060693629086018, + 0.01891431212425232, + 0.019239099696278572, + 0.013876384124159813, + 0.013449360616505146, + 0.010351692326366901, + 0.010358434170484543, + 0.022935818880796432, + 0.02415572479367256, + 0.009958581067621708, + 0.010570824146270752, + 0.016433868557214737, + 0.008854825049638748, + 0.002698782831430435, + 0.003232919378206134, + 0.006325316149741411, + 0.004673447459936142, + 0.004164867103099823, + 0.003100368892773986, + 0.00365821341983974, + 0.004751555155962706, + 0.004480644129216671, + 0.003551902947947383, + 0.004785283003002405, + 0.0043158880434930325, + 0.0033031476195901632, + 0.00364531297236681, + 0.003477399470284581, + 0.006879339460283518, + 0.056183502078056335, + 0.012220337055623531, + 0.13716113567352295, + 0.017653295770287514, + 0.01876920275390148, + 0.02432713657617569, + 0.017855411395430565, + 0.09727045893669128, + 0.007792980410158634, + 0.0055453903041779995, + 0.008413170464336872, + 0.006199568044394255, + 0.00565015384927392, + 0.014817962422966957, + 0.00946824811398983, + 0.01939617097377777, + 0.007464176043868065, + 0.008812232874333858, + 0.01243469025939703, + 0.009655904024839401, + 0.004201965406537056, + 0.010644971393048763, + 0.008538618683815002, + 0.01758205145597458, + 0.008790821768343449, + 0.006341850385069847, + 0.007674238178879023, + 0.009617891162633896, + 0.007746774237602949, + 0.008396424353122711, + 0.009626362472772598, + 0.007974671199917793, + 0.00539686344563961, + 0.0070954845286905766, + 0.003857584670186043, + 0.006262526847422123, + 0.004692934453487396, + 0.003524438012391329, + 0.005005352199077606, + 0.003533045295625925, + 0.006065023131668568, + 0.009285889565944672, + 0.006583626847714186, + 0.00848353747278452, + 0.0032292932737618685, + 0.007144487462937832, + 0.006137729622423649, + 0.005455107428133488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.00977963674813509 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal, high-definition digital aesthetic, the film begins with a slow, graceful, doll-like movement, sliding horizontally across the magnificent, ornate facade of a sprawling limestone mansion. Soft, golden morning light casts long shadows across the intricate carvings and towering pillars, rendering every detail with crisp 4K clarity. As the camera continues its steady, uninterrupted path, it perfectly aligns with a spotless, large bay window, revealing a brightly lit, luxurious bathroom. Through the pristine glass, a woman with dark black hair pulled back into a sleek, elegant bun can be seen leaning against a large, ornate mirror. She wears a lustrous, blush-tinged silk robe, the fabric striking in its allure. She meticulously shapes her eyebrows with precise, deliberate movements, her expression intense and unwavering. In her right hand, she holds a pair of pointed metal tweezers, while with her left, she gently pinches the skin above her eyes. The soft morning light from the window and the warm, even glow from the bulb in the mirror-framed dressing table combine to illuminate the scene, creating a focused halo around her face, while the luxurious surroundings of white marble countertops, gleaming gilded fixtures, and crystal accents fade into the soft-focus background, establishing an atmosphere of quiet, personal ritual, and focused perfection.", + "seed": 29974, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_10/index_7152_seed_29974_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1589017110876739, + "vlm_logprob": -0.1589017110876739, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.1589017110876739, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4859861731529236 + ], + "opposite_logprob_tokens": [ + -0.9859861731529236 + ], + "logprob_reward": -0.4859861731529236, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4859861731529236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18887871503829956, + 0.07177148759365082, + 0.07090011239051819, + 0.1242634654045105, + 0.16147027909755707, + 0.0713794082403183, + 0.08125033229589462, + 0.09610375761985779, + 0.07265659421682358, + 0.06132584065198898 + ], + "credit_map": [ + 0.014727943576872349, + 0.021304728463292122, + 0.008483548648655415, + 0.00944609846919775, + 0.02915322035551071, + 0.04103071242570877, + 0.03443552181124687, + 0.03029693476855755, + 0.005867246072739363, + 0.014027675613760948, + 0.002476147376000881, + 0.007774490863084793, + 0.014164429157972336, + 0.00780456280335784, + 0.009973587468266487, + 0.00968335010111332, + 0.004047112073749304, + 0.00892376434057951, + 0.004266213625669479, + 0.009446649812161922, + 0.01426494587212801, + 0.009842416271567345, + 0.01140484120696783, + 0.008704167790710926, + 0.008894858881831169, + 0.019463922828435898, + 0.010972070507705212, + 0.020007971674203873, + 0.02253653295338154, + 0.011062147095799446, + 0.014349631033837795, + 0.016976337879896164, + 0.01227104663848877, + 0.035495027899742126, + 0.02215091697871685, + 0.03978950157761574, + 0.016779733821749687, + 0.011975022032856941, + 0.01057013776153326, + 0.012438897974789143, + 0.007949029095470905, + 0.010019119828939438, + 0.008234567008912563, + 0.012165464460849762, + 0.0072508580051362514, + 0.00592925725504756, + 0.011130879633128643, + 0.00870023388415575, + 0.008686834946274757, + 0.01740252412855625, + 0.007350351195782423, + 0.02331201732158661, + 0.0033658426254987717, + 0.004231704398989677, + 0.010763400234282017, + 0.0061376518569886684, + 0.016526512801647186, + 0.014321235939860344, + 0.013836266472935677, + 0.019014958292245865, + 0.0063873156905174255, + 0.006803086027503014, + 0.009744768030941486, + 0.009469613432884216, + 0.013280294835567474, + 0.012317076325416565, + 0.012556715868413448, + 0.010939037427306175, + 0.004181016702204943, + 0.006748370360583067, + 0.00693588936701417, + 0.005698193795979023, + 0.01125585287809372, + 0.006715054623782635, + 0.009912539273500443, + 0.010412866249680519, + 0.003355979220941663, + 0.005994662642478943, + 0.0060933902859687805, + 0.007585491985082626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.4859861731529236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790153294801712 + ], + "opposite_logprob_tokens": [ + -2.5790152549743652 + ], + "logprob_reward": -0.0790153294801712, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790153294801712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0781782791018486, + 0.091176338493824, + 0.09767915308475494, + 0.16204434633255005, + 0.06179186329245567, + 0.0638708621263504, + 0.1357133686542511, + 0.1526711881160736, + 0.08777374029159546, + 0.06910087168216705 + ], + "credit_map": [ + 0.005318287294358015, + 0.00951403845101595, + 0.0037608591374009848, + 0.004067220259457827, + 0.013265445828437805, + 0.015299985185265541, + 0.01600128598511219, + 0.010951153934001923, + 0.011015746742486954, + 0.01741221733391285, + 0.004252606071531773, + 0.006047964561730623, + 0.017758164554834366, + 0.012069803662598133, + 0.011866227723658085, + 0.010753612034022808, + 0.0057020029053092, + 0.006969282403588295, + 0.00408996781334281, + 0.009144671261310577, + 0.03297417610883713, + 0.014299052767455578, + 0.01710604876279831, + 0.007393948268145323, + 0.009265100583434105, + 0.02469751238822937, + 0.01263708807528019, + 0.03730276599526405, + 0.02432459220290184, + 0.014314869418740273, + 0.016264941543340683, + 0.023237477988004684, + 0.00500947842374444, + 0.010198142379522324, + 0.008128446526825428, + 0.011763302609324455, + 0.009727690368890762, + 0.004586027469485998, + 0.008024667389690876, + 0.004354103002697229, + 0.005630017723888159, + 0.01237013190984726, + 0.0065907323732972145, + 0.010532291606068611, + 0.00589783675968647, + 0.004254295490682125, + 0.01064703706651926, + 0.007948524318635464, + 0.01581590063869953, + 0.03048844262957573, + 0.015123375691473484, + 0.027405403554439545, + 0.006808728910982609, + 0.007771523669362068, + 0.017371229827404022, + 0.014928758144378662, + 0.01880810782313347, + 0.02371261455118656, + 0.020626237615942955, + 0.033126018941402435, + 0.010850435122847557, + 0.011665177531540394, + 0.01730911247432232, + 0.01657349057495594, + 0.017157236114144325, + 0.014359403401613235, + 0.01644367165863514, + 0.012800771743059158, + 0.004849407821893692, + 0.006174643523991108, + 0.007174327038228512, + 0.008814279921352863, + 0.012338003143668175, + 0.007114415988326073, + 0.010813092812895775, + 0.009185158647596836, + 0.00399593124166131, + 0.006061709951609373, + 0.006646395660936832, + 0.01294616237282753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.0790153294801712 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023298904299736023 + ], + "opposite_logprob_tokens": [ + -3.773298978805542 + ], + "logprob_reward": -0.023298904299736023, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023298904299736023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09855090081691742, + 0.03998667001724243, + 0.05763214826583862, + 0.11749760806560516, + 0.09493648260831833, + 0.14017610251903534, + 0.06823904067277908, + 0.17884916067123413, + 0.044451311230659485, + 0.15968061983585358 + ], + "credit_map": [ + 0.008099894970655441, + 0.011477431282401085, + 0.003622278105467558, + 0.00478588230907917, + 0.015373105183243752, + 0.021536104381084442, + 0.017984172329306602, + 0.01567203737795353, + 0.003068708349019289, + 0.006315892096608877, + 0.0016995554324239492, + 0.0032466526608914137, + 0.007871548645198345, + 0.0051017384976148605, + 0.007814288139343262, + 0.004868284333497286, + 0.004357869271188974, + 0.009772672317922115, + 0.003989406395703554, + 0.007761900778859854, + 0.00986071303486824, + 0.008312845602631569, + 0.007582357153296471, + 0.005994379986077547, + 0.008009854704141617, + 0.016538340598344803, + 0.009167637676000595, + 0.020229261368513107, + 0.01944345235824585, + 0.011764331720769405, + 0.011910717934370041, + 0.020433999598026276, + 0.007421696558594704, + 0.011876080185174942, + 0.010241725482046604, + 0.015280287712812424, + 0.016597701236605644, + 0.01344593707472086, + 0.00809517316520214, + 0.011977868154644966, + 0.011159677989780903, + 0.021714311093091965, + 0.015952270478010178, + 0.01903916709125042, + 0.0178554467856884, + 0.012141858227550983, + 0.02319401130080223, + 0.01911935582756996, + 0.005781316198408604, + 0.00939599797129631, + 0.009367531165480614, + 0.007839439436793327, + 0.004606597125530243, + 0.008143821731209755, + 0.008672556839883327, + 0.014431774616241455, + 0.02072305791079998, + 0.031606581062078476, + 0.023039771243929863, + 0.025114089250564575, + 0.013392780907452106, + 0.013301653787493706, + 0.016780700534582138, + 0.034890517592430115, + 0.007755823899060488, + 0.00487677613273263, + 0.006495892535895109, + 0.008443181402981281, + 0.0026668377686291933, + 0.003332632826641202, + 0.0056993719190359116, + 0.005180789157748222, + 0.017474394291639328, + 0.007735210005193949, + 0.018692294135689735, + 0.02105804719030857, + 0.008021341636776924, + 0.018258316442370415, + 0.020121857523918152, + 0.04831913858652115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.023298904299736023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015685737133026 + ], + "opposite_logprob_tokens": [ + -1.701568603515625 + ], + "logprob_reward": -0.2015685737133026, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015685737133026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06755581498146057, + 0.0538528747856617, + 0.08812146633863449, + 0.24273289740085602, + 0.05795716121792793, + 0.21907836198806763, + 0.0821109190583229, + 0.09566270560026169, + 0.06542333215475082, + 0.027504492551088333 + ], + "credit_map": [ + 0.004712285939604044, + 0.00729383947327733, + 0.002139411401003599, + 0.0026255298871546984, + 0.010250876657664776, + 0.015027185901999474, + 0.014766350388526917, + 0.010740332305431366, + 0.004949531052261591, + 0.009396322071552277, + 0.001982208341360092, + 0.004375159740447998, + 0.009855846874415874, + 0.007494655903428793, + 0.006973643321543932, + 0.008825507014989853, + 0.006291334982961416, + 0.01332947053015232, + 0.005185466259717941, + 0.010609383694827557, + 0.014333966188132763, + 0.01602969318628311, + 0.011040102690458298, + 0.011302052065730095, + 0.018550653010606766, + 0.042117513716220856, + 0.02604854106903076, + 0.04671415686607361, + 0.03644314035773277, + 0.0211919117718935, + 0.02392202988266945, + 0.02774495631456375, + 0.004839659668505192, + 0.006292537320405245, + 0.006406004074960947, + 0.010160799138247967, + 0.010479247197508812, + 0.006488760467618704, + 0.006284086965024471, + 0.007006068713963032, + 0.01876177079975605, + 0.038446057587862015, + 0.020415643230080605, + 0.031549613922834396, + 0.02388397417962551, + 0.01971610262989998, + 0.03915511071681976, + 0.027150075882673264, + 0.009391497820615768, + 0.01154773123562336, + 0.008436713367700577, + 0.011783955618739128, + 0.00524222431704402, + 0.008592750877141953, + 0.013786893337965012, + 0.013329152949154377, + 0.012170709669589996, + 0.01849997043609619, + 0.012072484008967876, + 0.015893975272774696, + 0.007324244827032089, + 0.0069778840988874435, + 0.008613310754299164, + 0.01411013025790453, + 0.012156237848103046, + 0.006936676800251007, + 0.009803524240851402, + 0.00887628085911274, + 0.004089322406798601, + 0.006307235453277826, + 0.006902753375470638, + 0.010351301170885563, + 0.0030510700307786465, + 0.0033856870140880346, + 0.003132268087938428, + 0.004645681940019131, + 0.0018100464949384332, + 0.0020865050610154867, + 0.004252123646438122, + 0.0051411110907793045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.2015685737133026 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463957479223609 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.00463957479223609, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463957479223609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1168486475944519, + 0.06330341845750809, + 0.0913810282945633, + 0.228693887591362, + 0.15297622978687286, + 0.05394656956195831, + 0.06017763912677765, + 0.13629788160324097, + 0.037280552089214325, + 0.059094104915857315 + ], + "credit_map": [ + 0.010202188044786453, + 0.011067205108702183, + 0.0039050369523465633, + 0.005729164462536573, + 0.021492063999176025, + 0.028269261121749878, + 0.02012084238231182, + 0.016062883660197258, + 0.006497520953416824, + 0.01465113740414381, + 0.0029805165249854326, + 0.004884501453489065, + 0.00980523880571127, + 0.0081761684268713, + 0.010194468311965466, + 0.006113869603723288, + 0.005735225509852171, + 0.014574627391994, + 0.0047901044599711895, + 0.009887236170470715, + 0.018733955919742584, + 0.01759975403547287, + 0.011816361919045448, + 0.008243759162724018, + 0.01663617230951786, + 0.03524071350693703, + 0.02466513030230999, + 0.03995819017291069, + 0.03531710430979729, + 0.020688379183411598, + 0.02239180915057659, + 0.03379638493061066, + 0.011258102022111416, + 0.02588856965303421, + 0.01874482072889805, + 0.03398731350898743, + 0.01798492670059204, + 0.017756886780261993, + 0.011565369553864002, + 0.015790238976478577, + 0.0057755415327847, + 0.010009055957198143, + 0.005714156199246645, + 0.00802523922175169, + 0.005239560268819332, + 0.003436038037762046, + 0.008941950276494026, + 0.006805026438087225, + 0.008558868430554867, + 0.015264267101883888, + 0.0085964510217309, + 0.009659920819103718, + 0.002974142087623477, + 0.0036430656909942627, + 0.004822053015232086, + 0.006658871192485094, + 0.01683410070836544, + 0.02223033457994461, + 0.0193929523229599, + 0.028253907337784767, + 0.008904787711799145, + 0.01150188222527504, + 0.013115243054926395, + 0.016064690425992012, + 0.007293531205505133, + 0.0037117712199687958, + 0.005373392254114151, + 0.006539374124258757, + 0.002115963725373149, + 0.0037486825603991747, + 0.00417845044285059, + 0.00431938748806715, + 0.008661763742566109, + 0.0065961782820522785, + 0.00807859469205141, + 0.011397537775337696, + 0.003120425622910261, + 0.005783033091574907, + 0.007177589926868677, + 0.008278981782495975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.00463957479223609 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal, high-definition digital aesthetic, the film begins with a slow, graceful, doll-like movement, sliding horizontally across the magnificent, ornate facade of a sprawling limestone mansion. Soft, golden morning light casts long shadows across the intricate carvings and towering pillars, rendering every detail with crisp 4K clarity. As the camera continues its steady, uninterrupted path, it perfectly aligns with a spotless, large bay window, revealing a brightly lit, luxurious bathroom. Through the pristine glass, a woman with dark black hair pulled back into a sleek, elegant bun can be seen leaning against a large, ornate mirror. She wears a lustrous, blush-tinged silk robe, the fabric striking in its allure. She meticulously shapes her eyebrows with precise, deliberate movements, her expression intense and unwavering. In her right hand, she holds a pair of pointed metal tweezers, while with her left, she gently pinches the skin above her eyes. The soft morning light from the window and the warm, even glow from the bulb in the mirror-framed dressing table combine to illuminate the scene, creating a focused halo around her face, while the luxurious surroundings of white marble countertops, gleaming gilded fixtures, and crystal accents fade into the soft-focus background, establishing an atmosphere of quiet, personal ritual, and focused perfection.", + "seed": 29975, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_10/index_7152_seed_29975_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08906244039535523, + "vlm_logprob": -0.08906244039535523, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.08906244039535523, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1680724024772644 + ], + "opposite_logprob_tokens": [ + -2.043072462081909 + ], + "logprob_reward": -0.1680724024772644, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1680724024772644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28466325998306274, + 0.165885329246521, + 0.16044101119041443, + 0.060956697911024094, + 0.09145656228065491, + 0.06409214437007904, + 0.07476143538951874, + 0.030100377276539803, + 0.0528857596218586, + 0.014757471159100533 + ], + "credit_map": [ + 0.01604396291077137, + 0.013342811726033688, + 0.020825158804655075, + 0.01767779141664505, + 0.01962953992187977, + 0.09379909187555313, + 0.021197408437728882, + 0.08214748650789261, + 0.017550870776176453, + 0.016833752393722534, + 0.0171123668551445, + 0.019332831725478172, + 0.03457781299948692, + 0.012932623736560345, + 0.03449821472167969, + 0.013046842999756336, + 0.01806057058274746, + 0.018977008759975433, + 0.03161991760134697, + 0.018938999623060226, + 0.025441262871026993, + 0.01137148030102253, + 0.02148200199007988, + 0.014549761079251766, + 0.005910084582865238, + 0.00785120390355587, + 0.00786818377673626, + 0.009487338364124298, + 0.006927745882421732, + 0.006097579374909401, + 0.008314879611134529, + 0.008499679155647755, + 0.00684822304174304, + 0.0211788397282362, + 0.009454564191401005, + 0.02143477275967598, + 0.005215501878410578, + 0.007087089587002993, + 0.012015881948173046, + 0.00822167843580246, + 0.008016033098101616, + 0.013147935271263123, + 0.00813394133001566, + 0.008204570971429348, + 0.00568182859569788, + 0.009460380300879478, + 0.0055351476185023785, + 0.005912304390221834, + 0.010196697898209095, + 0.009363947436213493, + 0.010703252628445625, + 0.023091062903404236, + 0.003203983884304762, + 0.005419065710157156, + 0.0063712866976857185, + 0.006412136368453503, + 0.002819477114826441, + 0.0019157628994435072, + 0.004237834829837084, + 0.007109728176146746, + 0.0024152372498065233, + 0.0038033919408917427, + 0.003204964566975832, + 0.004593976773321629, + 0.006960016675293446, + 0.0037638128269463778, + 0.006505768280476332, + 0.012053068727254868, + 0.0033529542852193117, + 0.006120034959167242, + 0.005750257056206465, + 0.008379839360713959, + 0.002773376414552331, + 0.0008650458185002208, + 0.0018747210269793868, + 0.002078393241390586, + 0.0010055375751107931, + 0.0016834643902257085, + 0.0016807593638077378, + 0.002796173794195056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.1680724024772644 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029827017337083817 + ], + "opposite_logprob_tokens": [ + -3.529827117919922 + ], + "logprob_reward": -0.029827017337083817, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029827017337083817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3618288040161133, + 0.11924783885478973, + 0.052474234253168106, + 0.06259208172559738, + 0.0916663408279419, + 0.04202713444828987, + 0.03777904063463211, + 0.08205903321504593, + 0.1257907897233963, + 0.024534687399864197 + ], + "credit_map": [ + 0.02214580774307251, + 0.025571659207344055, + 0.026884252205491066, + 0.017014799639582634, + 0.023864375427365303, + 0.12306384742259979, + 0.02845878154039383, + 0.09482529759407043, + 0.008662540465593338, + 0.009702903218567371, + 0.014134022407233715, + 0.0153370201587677, + 0.01992490515112877, + 0.016954341903328896, + 0.018441518768668175, + 0.01609058864414692, + 0.004256315995007753, + 0.008724375627934933, + 0.005270058289170265, + 0.00803995132446289, + 0.0077651310712099075, + 0.0039343154057860374, + 0.008607885800302029, + 0.00587620260193944, + 0.006396832410246134, + 0.008501180447638035, + 0.008909305557608604, + 0.01082150824368, + 0.007443309761583805, + 0.005523626692593098, + 0.009544583968818188, + 0.005451733712106943, + 0.008602257817983627, + 0.015669483691453934, + 0.017911411821842194, + 0.01615527644753456, + 0.0063107870519161224, + 0.006019797641783953, + 0.012141422368586063, + 0.008855911903083324, + 0.0052755787037312984, + 0.007468061055988073, + 0.00528104230761528, + 0.006570635363459587, + 0.003172514261677861, + 0.007654380053281784, + 0.002917136996984482, + 0.0036877840757369995, + 0.0038497010245919228, + 0.003318820847198367, + 0.004382351879030466, + 0.008698923513293266, + 0.003501721192151308, + 0.004271046724170446, + 0.004210100509226322, + 0.0055463784374296665, + 0.012789389118552208, + 0.007837070152163506, + 0.014341166242957115, + 0.02365921251475811, + 0.004584777168929577, + 0.0069573670625686646, + 0.004610850941389799, + 0.007279202342033386, + 0.01926882378757, + 0.010097048245370388, + 0.01534534152597189, + 0.026492154225707054, + 0.008239404298365116, + 0.01582554168999195, + 0.01209008414298296, + 0.018432391807436943, + 0.003097374690696597, + 0.0014290045946836472, + 0.003226448781788349, + 0.004171403124928474, + 0.0020030969753861427, + 0.004664301872253418, + 0.0022812029346823692, + 0.003661854425445199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.029827017337083817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207052998244762 + ], + "opposite_logprob_tokens": [ + -4.264206886291504 + ], + "logprob_reward": -0.014207052998244762, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207052998244762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.114777110517025, + 0.06741605699062347, + 0.1660885512828827, + 0.06782378256320953, + 0.1401965618133545, + 0.05789245292544365, + 0.0484832264482975, + 0.08566927164793015, + 0.1486113965511322, + 0.1030416414141655 + ], + "credit_map": [ + 0.01320781372487545, + 0.018332067877054214, + 0.012253321707248688, + 0.01283217966556549, + 0.0227589663118124, + 0.01151855569332838, + 0.014359680004417896, + 0.009514527395367622, + 0.007525741122663021, + 0.008954169228672981, + 0.011077639646828175, + 0.011280927807092667, + 0.006845526397228241, + 0.006225836928933859, + 0.008232945576310158, + 0.007273269351571798, + 0.02096201851963997, + 0.016644125804305077, + 0.033830832690000534, + 0.015336168929934502, + 0.02197156846523285, + 0.012456035241484642, + 0.025980144739151, + 0.018907666206359863, + 0.005234532058238983, + 0.008773225359618664, + 0.009535448625683784, + 0.008403180167078972, + 0.007905928418040276, + 0.007379369810223579, + 0.012265876866877079, + 0.008326214738190174, + 0.01216952409595251, + 0.014485401101410389, + 0.026476014405488968, + 0.0377391092479229, + 0.008047057315707207, + 0.011266005225479603, + 0.017053574323654175, + 0.012959877029061317, + 0.0049306671135127544, + 0.0076782298274338245, + 0.005710152443498373, + 0.007869274355471134, + 0.007304392289370298, + 0.010409390553832054, + 0.008073985576629639, + 0.00591635936871171, + 0.004885262344032526, + 0.003968851175159216, + 0.006116883363574743, + 0.010053453035652637, + 0.005477121565490961, + 0.0040025245398283005, + 0.008451658301055431, + 0.00552747119218111, + 0.01024018693715334, + 0.006898495834320784, + 0.013361361809074879, + 0.02300255000591278, + 0.005911849904805422, + 0.007845290005207062, + 0.010043720714747906, + 0.008365805260837078, + 0.03144042193889618, + 0.011793018318712711, + 0.020886927843093872, + 0.016618870198726654, + 0.009653028100728989, + 0.019749175757169724, + 0.012559664435684681, + 0.02591029740869999, + 0.022120675072073936, + 0.006431868765503168, + 0.015615181997418404, + 0.017194516956806183, + 0.0067291706800460815, + 0.012642796151340008, + 0.008105698972940445, + 0.014201740734279156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.014207052998244762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557760775089264 + ], + "opposite_logprob_tokens": [ + -1.6005775928497314 + ], + "logprob_reward": -0.22557760775089264, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557760775089264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11370822042226791, + 0.13528481125831604, + 0.10553440451622009, + 0.0930076614022255, + 0.09821441024541855, + 0.11383150517940521, + 0.09672325849533081, + 0.061002135276794434, + 0.12846967577934265, + 0.054223887622356415 + ], + "credit_map": [ + 0.01364313717931509, + 0.016991466283798218, + 0.013460961170494556, + 0.008925547823309898, + 0.013572868891060352, + 0.020668523386120796, + 0.01356262806802988, + 0.012883113697171211, + 0.01888478919863701, + 0.018342144787311554, + 0.019295722246170044, + 0.01886654458940029, + 0.016302110627293587, + 0.01732199266552925, + 0.01435140147805214, + 0.01192012894898653, + 0.009449414908885956, + 0.009212675504386425, + 0.016735130921006203, + 0.009374488145112991, + 0.015525519847869873, + 0.012008517049252987, + 0.01565336063504219, + 0.017575306817889214, + 0.007241003215312958, + 0.012951347045600414, + 0.013524835929274559, + 0.013630259782075882, + 0.012765414081513882, + 0.00667707622051239, + 0.016218163073062897, + 0.00999956950545311, + 0.008474823087453842, + 0.015435016714036465, + 0.009545594453811646, + 0.02307233400642872, + 0.007409875281155109, + 0.008973767049610615, + 0.01378181017935276, + 0.01152119506150484, + 0.011598329991102219, + 0.016465088352560997, + 0.0130824726074934, + 0.016437606886029243, + 0.011024906300008297, + 0.025308430194854736, + 0.010965414345264435, + 0.008949263952672482, + 0.006678359117358923, + 0.0075090909376740456, + 0.010240818373858929, + 0.017580915242433548, + 0.012638979591429234, + 0.012496976181864738, + 0.01666589453816414, + 0.012912234291434288, + 0.011779583059251308, + 0.0058192783035337925, + 0.0092735905200243, + 0.011978044174611568, + 0.00332585908472538, + 0.006547565571963787, + 0.005205774214118719, + 0.007072442676872015, + 0.016715262085199356, + 0.010564258322119713, + 0.014448176138103008, + 0.028314905241131783, + 0.011911495588719845, + 0.01353320199996233, + 0.01682666502892971, + 0.016155729070305824, + 0.009315263479948044, + 0.004676826298236847, + 0.007104639429599047, + 0.008790906518697739, + 0.007975084707140923, + 0.0056862598285079, + 0.005033107474446297, + 0.005641804542392492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.22557760775089264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video feature a slow, graceful, doll-like camera movement sliding horizontally across an ornate limestone mansion facade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a luxurious bathroom with white marble countertops, gilded fixtures, and crystal accents visible through a large bay window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a woman with dark black hair in a sleek bun wearing a blush-tinged silk robe visible leaning against an ornate mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman holding pointed metal tweezers in her right hand and pinching the skin above her eyes with her left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft golden morning light and a warm bulb glow creating a focused halo around the woman's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628121413290501 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007628121413290501, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628121413290501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14063027501106262, + 0.11485621333122253, + 0.05648664012551308, + 0.07236017286777496, + 0.29076236486434937, + 0.052056387066841125, + 0.05356330797076225, + 0.05772171914577484, + 0.09387940168380737, + 0.06768348813056946 + ], + "credit_map": [ + 0.011337269097566605, + 0.009833668358623981, + 0.00932320486754179, + 0.011697528883814812, + 0.011776471510529518, + 0.03769037500023842, + 0.010536296293139458, + 0.03843546286225319, + 0.005931364372372627, + 0.00954444706439972, + 0.014662263914942741, + 0.013857313431799412, + 0.02094312198460102, + 0.015713248401880264, + 0.018747445195913315, + 0.015457015484571457, + 0.004721892531961203, + 0.0058379764668643475, + 0.009153401479125023, + 0.006221958436071873, + 0.00767017574980855, + 0.004937488120049238, + 0.010412285104393959, + 0.007531467359513044, + 0.007537505589425564, + 0.0068215737119317055, + 0.01228786539286375, + 0.009994905441999435, + 0.006732097361236811, + 0.007488024886697531, + 0.012906746007502079, + 0.00859146099537611, + 0.02880733646452427, + 0.04313192889094353, + 0.07018791139125824, + 0.0970015898346901, + 0.010164676234126091, + 0.010189191438257694, + 0.021275538951158524, + 0.010004189796745777, + 0.005792197305709124, + 0.009943545795977116, + 0.006987599655985832, + 0.009243202395737171, + 0.003487116191536188, + 0.007200140506029129, + 0.0056014712899923325, + 0.0038011176511645317, + 0.0057001966051757336, + 0.0056664058938622475, + 0.006130407564342022, + 0.012839598581194878, + 0.005931078922003508, + 0.003970352001488209, + 0.007393676321953535, + 0.005931595806032419, + 0.009576551616191864, + 0.004300628788769245, + 0.009010870940983295, + 0.014151877723634243, + 0.003239481942728162, + 0.007427222095429897, + 0.004020638298243284, + 0.005994455423206091, + 0.011320729739964008, + 0.006167923100292683, + 0.011880097910761833, + 0.022725922986865044, + 0.005224116612225771, + 0.012165132910013199, + 0.009873834438621998, + 0.014521647244691849, + 0.013188806362450123, + 0.005114329047501087, + 0.009846535511314869, + 0.01069248653948307, + 0.007468665484338999, + 0.006198474206030369, + 0.006207720376551151, + 0.008966472931206226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.007628121413290501 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_11/index_tensor([4533, 4533])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_11/index_tensor([4533, 4533])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7f0b26ead80fb9495241c1aae367765d80bdfda5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_11/index_tensor([4533, 4533])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "resolute, locked-off split-horizon shot captures a captivating natural contradiction and the rural charm of a pond-filled landscape. Above the waterline, the surface is a tranquil, unbroken mosaic of distinctive, diamond-shaped water chestnut leaves, their waxy, deep-green surfaces floating in perfect stillness under the soft, golden glow of the late afternoon sun. Beneath this tranquil exterior, the clear, tea-stained water reveals a vibrant world. A powerful, unseen deep current causes the slender, green-brown underwater stems to perform a unified, mesmerizing dance. They sway gracefully yet relentlessly in one direction, bending and straightening in hypnotic, synchronized waves, their feathery underwater roots trailing like silk banners. Sparkling patterns of light created by sunlight filtering through the water's surface, punctuated by the swaying stems and the silty bottom, highlight the contrast between the serene, static world above and the powerful, rhythmic energy pulsating below, creating an atmosphere both profoundly tranquil and vibrant.", + "seed": 32615, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_11/index_4533_seed_32615_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06621594429016113, + "vlm_logprob": -0.06621594429016113, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.06621594429016113, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07293408364057541 + ], + "opposite_logprob_tokens": [ + -2.947934150695801 + ], + "logprob_reward": -0.07293408364057541, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07293408364057541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3074946701526642, + 0.09612179547548294, + 0.11575623601675034, + 0.059145163744688034, + 0.1593957394361496, + 0.028347501531243324, + 0.07978105545043945, + 0.04419582709670067, + 0.061060983687639236, + 0.04870099201798439 + ], + "credit_map": [ + 0.038411542773246765, + 0.037368930876255035, + 0.020034171640872955, + 0.03696464002132416, + 0.03487204760313034, + 0.03394630551338196, + 0.049696799367666245, + 0.056200262159109116, + 0.011401894502341747, + 0.01206732727587223, + 0.008745400235056877, + 0.009235545061528683, + 0.012482946738600731, + 0.008407145738601685, + 0.016866281628608704, + 0.01691526360809803, + 0.016687223687767982, + 0.015613162890076637, + 0.011233034543693066, + 0.017444560304284096, + 0.012314817868173122, + 0.01160998921841383, + 0.014939181506633759, + 0.015914274379611015, + 0.006512362509965897, + 0.004579637199640274, + 0.006273421924561262, + 0.009942722506821156, + 0.0053263562731444836, + 0.005543960724025965, + 0.012183300219476223, + 0.008783407509326935, + 0.017259445041418076, + 0.016493339091539383, + 0.017378568649291992, + 0.028260793536901474, + 0.012376362457871437, + 0.01879166066646576, + 0.027282390743494034, + 0.021553196012973785, + 0.0028785953763872385, + 0.0031034222338348627, + 0.0029892693273723125, + 0.003977406304329634, + 0.003459228901192546, + 0.0026336777955293655, + 0.004631826188415289, + 0.004674077965319157, + 0.01076507568359375, + 0.00622536800801754, + 0.008340259082615376, + 0.012271457351744175, + 0.0112083088606596, + 0.009092144668102264, + 0.011249085888266563, + 0.010629361495375633, + 0.0039878543466329575, + 0.00725591229274869, + 0.0034465582575649023, + 0.005201223772019148, + 0.004485945217311382, + 0.006277252919971943, + 0.006383854430168867, + 0.007157228421419859, + 0.006163784768432379, + 0.006988612003624439, + 0.005206984467804432, + 0.005638573318719864, + 0.013200095854699612, + 0.007328318897634745, + 0.008989984169602394, + 0.007544631604105234, + 0.006791969295591116, + 0.005099763628095388, + 0.0038997025694698095, + 0.005321846809238195, + 0.007531967479735613, + 0.0036523903254419565, + 0.006350151728838682, + 0.01005320344120264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.07293408364057541 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038141824305057526 + ], + "opposite_logprob_tokens": [ + -3.2881417274475098 + ], + "logprob_reward": -0.038141824305057526, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038141824305057526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2819841206073761, + 0.16725744307041168, + 0.0927862599492073, + 0.0490710623562336, + 0.11918515712022781, + 0.15727780759334564, + 0.020253663882613182, + 0.046340953558683395, + 0.01865040510892868, + 0.04719308763742447 + ], + "credit_map": [ + 0.044286154210567474, + 0.03310725465416908, + 0.02361537329852581, + 0.03425884246826172, + 0.042674195021390915, + 0.02837461605668068, + 0.03855450078845024, + 0.03711320459842682, + 0.023609718307852745, + 0.02676439844071865, + 0.013311103917658329, + 0.016731243580579758, + 0.023225141689181328, + 0.020733535289764404, + 0.02471701242029667, + 0.01816529780626297, + 0.015379275195300579, + 0.010038035921752453, + 0.007774027530103922, + 0.011373372748494148, + 0.015504924580454826, + 0.009304660372436047, + 0.01483895257115364, + 0.008573011495172977, + 0.007222162093967199, + 0.004668530076742172, + 0.004164181184023619, + 0.006644588429480791, + 0.0070895967073738575, + 0.005036674905568361, + 0.007312198169529438, + 0.006933137308806181, + 0.01425708457827568, + 0.016130678355693817, + 0.0099479416385293, + 0.01219012401998043, + 0.019229162484407425, + 0.016211383044719696, + 0.017658105120062828, + 0.013560685329139233, + 0.017922576516866684, + 0.022185668349266052, + 0.01298619620501995, + 0.01279798150062561, + 0.035534366965293884, + 0.016474135220050812, + 0.022989287972450256, + 0.016387607902288437, + 0.002406352898105979, + 0.0019262718269601464, + 0.001944280811585486, + 0.0023600533604621887, + 0.0037504772190004587, + 0.0018448579357936978, + 0.00271833804436028, + 0.0033030326012521982, + 0.00462359469383955, + 0.004909965209662914, + 0.004921065643429756, + 0.004873326979577541, + 0.0062969764694571495, + 0.005595863331109285, + 0.0064724707044661045, + 0.008647690527141094, + 0.002373480936512351, + 0.0021827055606991053, + 0.0014013793552294374, + 0.002261265879496932, + 0.003391016274690628, + 0.0016482616774737835, + 0.0024672907311469316, + 0.0029250073712319136, + 0.005077087786048651, + 0.005936320871114731, + 0.0036545582115650177, + 0.004135891795158386, + 0.011701461859047413, + 0.005315559916198254, + 0.0065211947076022625, + 0.004851015284657478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.038141824305057526 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898503541946411 + ], + "opposite_logprob_tokens": [ + -2.5789849758148193 + ], + "logprob_reward": -0.07898503541946411, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898503541946411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33837631344795227, + 0.10962290316820145, + 0.11543641239404678, + 0.0735372006893158, + 0.11733917891979218, + 0.06910690665245056, + 0.06053030118346214, + 0.04528634995222092, + 0.020337674766778946, + 0.05042680352926254 + ], + "credit_map": [ + 0.036696907132864, + 0.035670503973960876, + 0.02880595624446869, + 0.051293302327394485, + 0.027658743783831596, + 0.03238268196582794, + 0.04783754050731659, + 0.07803066074848175, + 0.013644903898239136, + 0.011613004840910435, + 0.012997484765946865, + 0.012795542366802692, + 0.010871603153645992, + 0.009709309786558151, + 0.017907259985804558, + 0.02008378505706787, + 0.019532430917024612, + 0.012254367582499981, + 0.011049652472138405, + 0.01056237705051899, + 0.019224198535084724, + 0.009315717965364456, + 0.0182217825204134, + 0.01527587790042162, + 0.010799786075949669, + 0.009287059307098389, + 0.006435668561607599, + 0.007141477428376675, + 0.01447569951415062, + 0.007854999974370003, + 0.009890344925224781, + 0.007652157451957464, + 0.011535748839378357, + 0.011975263245403767, + 0.010862375609576702, + 0.013481476344168186, + 0.01800554245710373, + 0.010546688921749592, + 0.0220494344830513, + 0.01888263039290905, + 0.007273952942341566, + 0.01012846827507019, + 0.0057708946987986565, + 0.008259264752268791, + 0.011949772015213966, + 0.007050164043903351, + 0.00836606603115797, + 0.010308315977454185, + 0.00702830171212554, + 0.006488395854830742, + 0.005757639184594154, + 0.007688569370657206, + 0.011312877759337425, + 0.005710737779736519, + 0.008333085104823112, + 0.008210686035454273, + 0.004728987347334623, + 0.006773619446903467, + 0.0033880264963954687, + 0.005042418371886015, + 0.008949295617640018, + 0.00471954420208931, + 0.005845379550009966, + 0.005839076358824968, + 0.002143473131582141, + 0.0015836297534406185, + 0.0024777562357485294, + 0.002535803709179163, + 0.0022180962841957808, + 0.002251920523121953, + 0.0032708588987588882, + 0.0038561339024454355, + 0.004995465744286776, + 0.006542261689901352, + 0.0038794141728430986, + 0.004617924336344004, + 0.00989464670419693, + 0.006076897960156202, + 0.007155692670494318, + 0.007264497224241495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.07898503541946411 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062048062682151794 + ], + "opposite_logprob_tokens": [ + -2.8120479583740234 + ], + "logprob_reward": -0.062048062682151794, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062048062682151794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5271719098091125, + 0.06393598765134811, + 0.059419840574264526, + 0.04130105674266815, + 0.05710706114768982, + 0.04308663681149483, + 0.04037068039178848, + 0.06504716724157333, + 0.022864026948809624, + 0.07969565689563751 + ], + "credit_map": [ + 0.07153971493244171, + 0.04969869926571846, + 0.03750757873058319, + 0.059870921075344086, + 0.054121240973472595, + 0.047704800963401794, + 0.0861281231045723, + 0.12060078233480453, + 0.006097853649407625, + 0.007009251043200493, + 0.008045882917940617, + 0.0106985317543149, + 0.003813241608440876, + 0.006164055783301592, + 0.009173582307994366, + 0.012933588586747646, + 0.006936614401638508, + 0.00862069707363844, + 0.005992979742586613, + 0.00879409909248352, + 0.007075360976159573, + 0.005744259804487228, + 0.008540072478353977, + 0.007715755142271519, + 0.005254886578768492, + 0.003896985901519656, + 0.0044723618775606155, + 0.005969975143671036, + 0.0044774748384952545, + 0.004743013996630907, + 0.007218181621283293, + 0.005268172360956669, + 0.005319760646671057, + 0.005776830017566681, + 0.005623389035463333, + 0.008592959493398666, + 0.007340375334024429, + 0.005891367327421904, + 0.010312624275684357, + 0.008249749429523945, + 0.004587613977491856, + 0.005903778597712517, + 0.003646841971203685, + 0.003950200043618679, + 0.009443376213312149, + 0.004575377330183983, + 0.005830095615237951, + 0.00514935003593564, + 0.0040057566948235035, + 0.005766780115664005, + 0.0032751369290053844, + 0.004492474254220724, + 0.007092426531016827, + 0.0049093677662312984, + 0.005294314119964838, + 0.0055344183929264545, + 0.00767246400937438, + 0.007566339336335659, + 0.006370292976498604, + 0.009299259632825851, + 0.011020397767424583, + 0.007258721627295017, + 0.008987905457615852, + 0.006871783174574375, + 0.0026216418482363224, + 0.0022543619852513075, + 0.0024077598936855793, + 0.0026315676514059305, + 0.004529611673206091, + 0.0023231646046042442, + 0.003368817502632737, + 0.0027271006256341934, + 0.009462668560445309, + 0.00760970963165164, + 0.007399726193398237, + 0.009834455326199532, + 0.015382960438728333, + 0.00864198338240385, + 0.01064486987888813, + 0.010719280689954758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.062048062682151794 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897071540355682 + ], + "opposite_logprob_tokens": [ + -2.5789706707000732 + ], + "logprob_reward": -0.07897071540355682, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897071540355682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3564611077308655, + 0.1008596420288086, + 0.0757945328950882, + 0.05835491791367531, + 0.06124446541070938, + 0.03909183293581009, + 0.046489614993333817, + 0.05850140377879143, + 0.04893292486667633, + 0.15426957607269287 + ], + "credit_map": [ + 0.027087092399597168, + 0.03184717893600464, + 0.034023214131593704, + 0.05415027588605881, + 0.01832866296172142, + 0.026877455413341522, + 0.07428347319364548, + 0.08986374735832214, + 0.010235589928925037, + 0.013795924372971058, + 0.01093301735818386, + 0.011995088309049606, + 0.010768307372927666, + 0.009319135919213295, + 0.016314728185534477, + 0.017497848719358444, + 0.009677586145699024, + 0.01109723187983036, + 0.008257023058831692, + 0.00969613902270794, + 0.009166576899588108, + 0.008479425683617592, + 0.01030390989035368, + 0.00911663193255663, + 0.006152878515422344, + 0.0037165097892284393, + 0.007745556533336639, + 0.008005429059267044, + 0.004830053076148033, + 0.005251728463917971, + 0.0092241782695055, + 0.01342858001589775, + 0.006183243822306395, + 0.008306337520480156, + 0.006020134314894676, + 0.008069697767496109, + 0.005465310998260975, + 0.00891435332596302, + 0.008649512194097042, + 0.00963587686419487, + 0.004743556026369333, + 0.003818755503743887, + 0.004056524019688368, + 0.005526451393961906, + 0.006193468812853098, + 0.003581185592338443, + 0.005576894618570805, + 0.005594994407147169, + 0.006018782965838909, + 0.005712537094950676, + 0.004051913041621447, + 0.005769617389887571, + 0.009304378181695938, + 0.0037565683014690876, + 0.005825839936733246, + 0.0060499790124595165, + 0.006752763409167528, + 0.006083407439291477, + 0.006034572143107653, + 0.009080893360078335, + 0.009521105326712132, + 0.006126781925559044, + 0.007772855926305056, + 0.007129025179892778, + 0.005473951809108257, + 0.005709645338356495, + 0.003741280874237418, + 0.004704673308879137, + 0.010472539812326431, + 0.005270005203783512, + 0.006466216873377562, + 0.007094608619809151, + 0.018801722675561905, + 0.0207268875092268, + 0.011340398341417313, + 0.012284349650144577, + 0.035891808569431305, + 0.016582120209932327, + 0.021835869178175926, + 0.016806432977318764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07897071540355682 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "resolute, locked-off split-horizon shot captures a captivating natural contradiction and the rural charm of a pond-filled landscape. Above the waterline, the surface is a tranquil, unbroken mosaic of distinctive, diamond-shaped water chestnut leaves, their waxy, deep-green surfaces floating in perfect stillness under the soft, golden glow of the late afternoon sun. Beneath this tranquil exterior, the clear, tea-stained water reveals a vibrant world. A powerful, unseen deep current causes the slender, green-brown underwater stems to perform a unified, mesmerizing dance. They sway gracefully yet relentlessly in one direction, bending and straightening in hypnotic, synchronized waves, their feathery underwater roots trailing like silk banners. Sparkling patterns of light created by sunlight filtering through the water's surface, punctuated by the swaying stems and the silty bottom, highlight the contrast between the serene, static world above and the powerful, rhythmic energy pulsating below, creating an atmosphere both profoundly tranquil and vibrant.", + "seed": 32616, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_11/index_4533_seed_32616_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07940857224166394, + "vlm_logprob": -0.07940857224166394, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.07940857224166394, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10980453342199326 + ], + "opposite_logprob_tokens": [ + -2.859804630279541 + ], + "logprob_reward": -0.10980453342199326, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.10980453342199326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05391969904303551, + 0.012994996272027493, + 0.019495174288749695, + 0.02737554721534252, + 0.7797471880912781, + 0.021360911428928375, + 0.04918495938181877, + 0.023194726556539536, + 0.007101740688085556, + 0.005624993238598108 + ], + "credit_map": [ + 0.003449730109423399, + 0.0054333023726940155, + 0.01146102324128151, + 0.011532609350979328, + 0.0032271374948322773, + 0.003058273810893297, + 0.007212880067527294, + 0.008544748649001122, + 0.0009399745031259954, + 0.0008701174519956112, + 0.002187450649216771, + 0.002756187692284584, + 0.0015430545900017023, + 0.001108005759306252, + 0.0016620013630017638, + 0.0019282052526250482, + 0.0017473604530096054, + 0.0011064269347116351, + 0.002571063581854105, + 0.0024713235907256603, + 0.0016554836183786392, + 0.0014436474302783608, + 0.004483269993215799, + 0.00401659868657589, + 0.0026454138569533825, + 0.002968061016872525, + 0.004558539018034935, + 0.00468506570905447, + 0.0019594840705394745, + 0.0023750944528728724, + 0.004275646060705185, + 0.003908243961632252, + 0.050711046904325485, + 0.016508016735315323, + 0.27854567766189575, + 0.08733268082141876, + 0.01263706386089325, + 0.028815392404794693, + 0.05614589527249336, + 0.24905145168304443, + 0.0014760427875444293, + 0.0028863344341516495, + 0.00445821788161993, + 0.0040098861791193485, + 0.0017266820650547743, + 0.0016903821378946304, + 0.0025330111384391785, + 0.0025803556200116873, + 0.003911284264177084, + 0.0050937216728925705, + 0.0077990381978452206, + 0.007967785000801086, + 0.005260912235826254, + 0.004762549418956041, + 0.005332404747605324, + 0.00905726756900549, + 0.002476433524861932, + 0.0020212598610669374, + 0.004164789337664843, + 0.002836364321410656, + 0.002416744129732251, + 0.0016501732170581818, + 0.0033075413666665554, + 0.004321421030908823, + 0.0006900648586452007, + 0.0007760012522339821, + 0.0012540207244455814, + 0.0013355828123167157, + 0.000741646159440279, + 0.0005836380296386778, + 0.0010327022755518556, + 0.0006880852160975337, + 0.00048696144949644804, + 0.0006198717746883631, + 0.000671058485750109, + 0.0008749053231440485, + 0.0007483833469450474, + 0.0005509954644367099, + 0.0007759875152260065, + 0.0008968304609879851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.10980453342199326 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009825320914387703 + ], + "opposite_logprob_tokens": [ + -4.634825229644775 + ], + "logprob_reward": -0.009825320914387703, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009825320914387703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14706265926361084, + 0.0960593894124031, + 0.07957977056503296, + 0.0635833665728569, + 0.1692734956741333, + 0.06725601106882095, + 0.06268632411956787, + 0.09370160102844238, + 0.09280300885438919, + 0.1279943883419037 + ], + "credit_map": [ + 0.014611476100981236, + 0.012340273708105087, + 0.021152570843696594, + 0.028012199327349663, + 0.013709704391658306, + 0.015889309346675873, + 0.0198044553399086, + 0.021542677655816078, + 0.006061558611690998, + 0.007715610321611166, + 0.010079939849674702, + 0.017284462228417397, + 0.013632363639771938, + 0.013154465705156326, + 0.008175375871360302, + 0.019955623894929886, + 0.005794287193566561, + 0.005518162157386541, + 0.0071411035023629665, + 0.016618188470602036, + 0.00915087852627039, + 0.006496598478406668, + 0.011189471930265427, + 0.01767108216881752, + 0.005237781908363104, + 0.00613429257646203, + 0.008482313714921474, + 0.009219609200954437, + 0.007298948243260384, + 0.007793877273797989, + 0.00904357060790062, + 0.010372975841164589, + 0.013703994452953339, + 0.01521193515509367, + 0.03440404683351517, + 0.02580190822482109, + 0.015009879134595394, + 0.015325755812227726, + 0.02937369793653488, + 0.020442286506295204, + 0.006702417507767677, + 0.005508363712579012, + 0.011498263105750084, + 0.011055976152420044, + 0.007336669135838747, + 0.004724040161818266, + 0.010929740965366364, + 0.009500543586909771, + 0.006396132986992598, + 0.0037003904581069946, + 0.011101997457444668, + 0.00920081976801157, + 0.003523516468703747, + 0.005490661133080721, + 0.009478607214987278, + 0.01379420142620802, + 0.012395808473229408, + 0.007509226910769939, + 0.01524729747325182, + 0.011873388662934303, + 0.00893884152173996, + 0.008345451205968857, + 0.013101563788950443, + 0.01629003696143627, + 0.012424368411302567, + 0.009429809637367725, + 0.015408255159854889, + 0.011293402872979641, + 0.012324403040111065, + 0.008186988532543182, + 0.010146455839276314, + 0.013589336536824703, + 0.012510125525295734, + 0.011330748908221722, + 0.026037858799099922, + 0.01529658492654562, + 0.015056557022035122, + 0.01913601905107498, + 0.0163485798984766, + 0.012277924455702305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.009825320914387703 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897457480430603 + ], + "opposite_logprob_tokens": [ + -2.578974485397339 + ], + "logprob_reward": -0.07897457480430603, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897457480430603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22018589079380035, + 0.07122651487588882, + 0.10003439337015152, + 0.05693339183926582, + 0.20385858416557312, + 0.13625840842723846, + 0.12139126658439636, + 0.026878831908106804, + 0.025475358590483665, + 0.03775734826922417 + ], + "credit_map": [ + 0.024132048711180687, + 0.030557358637452126, + 0.02712092734873295, + 0.03659458085894585, + 0.02473822794854641, + 0.025039322674274445, + 0.024987392127513885, + 0.0270160511136055, + 0.0051135593093931675, + 0.005487219430506229, + 0.00830847304314375, + 0.016397036612033844, + 0.007062371354550123, + 0.008803588338196278, + 0.009988024830818176, + 0.010066243819892406, + 0.012069127522408962, + 0.005272433161735535, + 0.01297074742615223, + 0.017958348616957664, + 0.00608542887493968, + 0.006429486442357302, + 0.020903075113892555, + 0.01834573969244957, + 0.004480597097426653, + 0.0034996140748262405, + 0.014006220735609531, + 0.006705920677632093, + 0.00555732985958457, + 0.0034814276732504368, + 0.0077458331361413, + 0.0114564448595047, + 0.014090941287577152, + 0.012093039229512215, + 0.06194055452942848, + 0.023009922355413437, + 0.007832534611225128, + 0.012522528879344463, + 0.017721027135849, + 0.05464803799986839, + 0.01038155984133482, + 0.017230354249477386, + 0.03173825517296791, + 0.024589408189058304, + 0.016120078042149544, + 0.011409531347453594, + 0.012389839626848698, + 0.012399380095303059, + 0.008995329961180687, + 0.011804202571511269, + 0.0187599565833807, + 0.020471611991524696, + 0.014400682412087917, + 0.010549696162343025, + 0.014272776432335377, + 0.022137010470032692, + 0.0029328912496566772, + 0.002480503171682358, + 0.004201396368443966, + 0.003286804538220167, + 0.0030490239150822163, + 0.001755320467054844, + 0.005007960367947817, + 0.004164930433034897, + 0.002454078756272793, + 0.0020162807777523994, + 0.006204818841069937, + 0.0038744027260690928, + 0.0019247035961598158, + 0.0017447624122723937, + 0.004580987151712179, + 0.0026753253769129515, + 0.0034413267858326435, + 0.003522435436025262, + 0.006366732530295849, + 0.006597792264074087, + 0.003506526816636324, + 0.004504760727286339, + 0.005663292016834021, + 0.0041544814594089985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.07897457480430603 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038101088255643845 + ], + "opposite_logprob_tokens": [ + -3.2881011962890625 + ], + "logprob_reward": -0.038101088255643845, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038101088255643845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15384437143802643, + 0.06390555948019028, + 0.05072411522269249, + 0.32669922709465027, + 0.25408661365509033, + 0.03958697244524956, + 0.03645898774266243, + 0.03482026606798172, + 0.02529650367796421, + 0.014577397145330906 + ], + "credit_map": [ + 0.012994717806577682, + 0.019608233124017715, + 0.02302561141550541, + 0.026356680318713188, + 0.014320087619125843, + 0.020029330626130104, + 0.014626921154558659, + 0.022882789373397827, + 0.004624225199222565, + 0.005185361951589584, + 0.012347620911896229, + 0.012275101616978645, + 0.0039877076633274555, + 0.004449320957064629, + 0.005522849969565868, + 0.015513367019593716, + 0.004353303927928209, + 0.004759679548442364, + 0.006310755852609873, + 0.008014990016818047, + 0.006234322674572468, + 0.003914170898497105, + 0.009811793453991413, + 0.0073250955902040005, + 0.021926237270236015, + 0.014461799524724483, + 0.1180177852511406, + 0.03334582597017288, + 0.011782200075685978, + 0.012835503555834293, + 0.06807874143123627, + 0.046251147985458374, + 0.01655024290084839, + 0.005742757581174374, + 0.08686026930809021, + 0.02992904558777809, + 0.004410363733768463, + 0.009021705947816372, + 0.02000662498176098, + 0.08156560361385345, + 0.002443333389237523, + 0.0038939774967730045, + 0.009303604252636433, + 0.008716859854757786, + 0.004343992564827204, + 0.002878832397982478, + 0.003647844074293971, + 0.004358526319265366, + 0.003728728275746107, + 0.00386831141076982, + 0.004495990462601185, + 0.005910979583859444, + 0.003978317603468895, + 0.0034581725485622883, + 0.0045427787117660046, + 0.006475712172687054, + 0.004798175301402807, + 0.002016813028603792, + 0.007295322138816118, + 0.0037506702356040478, + 0.0019982841331511736, + 0.0022991241421550512, + 0.004623570945113897, + 0.00803830660879612, + 0.0013541176449507475, + 0.0020214137621223927, + 0.0036812990438193083, + 0.005625435151159763, + 0.0021789895836263895, + 0.002052524359896779, + 0.004158911295235157, + 0.0042238119058310986, + 0.0010455722222104669, + 0.0018319792579859495, + 0.002047705464065075, + 0.0024892818182706833, + 0.0015333984047174454, + 0.001850517001003027, + 0.002005542628467083, + 0.0017734008142724633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.038101088255643845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033734381198883 + ], + "opposite_logprob_tokens": [ + -1.9103373289108276 + ], + "logprob_reward": -0.16033734381198883, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033734381198883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09131006896495819, + 0.04423622041940689, + 0.042603276669979095, + 0.057065509259700775, + 0.42893707752227783, + 0.10602633655071259, + 0.18542830646038055, + 0.020891007035970688, + 0.015920715406537056, + 0.0075814248993992805 + ], + "credit_map": [ + 0.006862482521682978, + 0.00790330208837986, + 0.016649769619107246, + 0.015155644156038761, + 0.007523966953158379, + 0.008382542058825493, + 0.012020283378660679, + 0.01681208796799183, + 0.00428014574572444, + 0.0037774648517370224, + 0.007926027290523052, + 0.007427688222378492, + 0.002766445279121399, + 0.003925791475921869, + 0.0039666881784796715, + 0.010165974497795105, + 0.004798505920916796, + 0.004123672842979431, + 0.004612841177731752, + 0.005453305318951607, + 0.0054854280315339565, + 0.004808707162737846, + 0.0064192283898591995, + 0.006901593878865242, + 0.005190737079828978, + 0.003832494141533971, + 0.011047695763409138, + 0.008419346064329147, + 0.00597940431907773, + 0.0032862171065062284, + 0.007532356306910515, + 0.011777267791330814, + 0.027751820161938667, + 0.009732871316373348, + 0.15241126716136932, + 0.047499943524599075, + 0.00827612355351448, + 0.017883695662021637, + 0.031817514449357986, + 0.13356387615203857, + 0.006765110418200493, + 0.014923028647899628, + 0.023602334782481194, + 0.01984298974275589, + 0.011917955242097378, + 0.00702032633125782, + 0.011185739189386368, + 0.010768859647214413, + 0.014635972678661346, + 0.019418304786086082, + 0.03192363679409027, + 0.030260778963565826, + 0.01838182471692562, + 0.01777372695505619, + 0.01975213922560215, + 0.033281948417425156, + 0.002188432030379772, + 0.0009215642930939794, + 0.0069052549079060555, + 0.0022099658381193876, + 0.0009948263177648187, + 0.0007645270088687539, + 0.0027724222745746374, + 0.004134016577154398, + 0.0012321264948695898, + 0.001694086124189198, + 0.0026131230406463146, + 0.0033381585963070393, + 0.001238343771547079, + 0.0010267577599734068, + 0.0026501005049794912, + 0.0021280194632709026, + 0.0007227313471958041, + 0.0007690572529099882, + 0.0011384451063349843, + 0.0013066204264760017, + 0.0007827025256119668, + 0.0006039703148417175, + 0.0011517424136400223, + 0.001106155919842422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.16033734381198883 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "resolute, locked-off split-horizon shot captures a captivating natural contradiction and the rural charm of a pond-filled landscape. Above the waterline, the surface is a tranquil, unbroken mosaic of distinctive, diamond-shaped water chestnut leaves, their waxy, deep-green surfaces floating in perfect stillness under the soft, golden glow of the late afternoon sun. Beneath this tranquil exterior, the clear, tea-stained water reveals a vibrant world. A powerful, unseen deep current causes the slender, green-brown underwater stems to perform a unified, mesmerizing dance. They sway gracefully yet relentlessly in one direction, bending and straightening in hypnotic, synchronized waves, their feathery underwater roots trailing like silk banners. Sparkling patterns of light created by sunlight filtering through the water's surface, punctuated by the swaying stems and the silty bottom, highlight the contrast between the serene, static world above and the powerful, rhythmic energy pulsating below, creating an atmosphere both profoundly tranquil and vibrant.", + "seed": 32617, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_11/index_4533_seed_32617_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.127836998924613, + "vlm_logprob": -0.127836998924613, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.127836998924613, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12664808332920074 + ], + "opposite_logprob_tokens": [ + -2.251648187637329 + ], + "logprob_reward": -0.12664808332920074, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12664808332920074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.632291316986084, + 0.027194814756512642, + 0.03248917683959007, + 0.015144034288823605, + 0.07877480983734131, + 0.02118024043738842, + 0.04435496777296066, + 0.01201175432652235, + 0.1009274572134018, + 0.03563142567873001 + ], + "credit_map": [ + 0.027255816385149956, + 0.012557721696794033, + 0.017426252365112305, + 0.03305305913090706, + 0.02276979759335518, + 0.036092858761548996, + 0.2654578387737274, + 0.21767795085906982, + 0.0034852398093789816, + 0.00280253728851676, + 0.002438887720927596, + 0.002132402965798974, + 0.0027237085159868, + 0.0030279967468231916, + 0.005548370070755482, + 0.0050356704741716385, + 0.0018781505059450865, + 0.0033141886815428734, + 0.0015922137536108494, + 0.00229031965136528, + 0.0026587177999317646, + 0.0036069094203412533, + 0.008192600682377815, + 0.008956077508628368, + 0.0018570703687146306, + 0.0015892165247350931, + 0.0017418157076463103, + 0.0015113787958398461, + 0.0022820434533059597, + 0.002454727189615369, + 0.0013955133035779, + 0.002312268828973174, + 0.003698338521644473, + 0.004445490427315235, + 0.0030025599990040064, + 0.0031635137274861336, + 0.012469450943171978, + 0.014589322730898857, + 0.017899194732308388, + 0.019506942480802536, + 0.002920233877375722, + 0.002406919142231345, + 0.0032452414743602276, + 0.003763970686122775, + 0.0015207940014079213, + 0.0018301435047760606, + 0.002535519190132618, + 0.002957419026643038, + 0.00319463899359107, + 0.002482138341292739, + 0.006283443886786699, + 0.004888477735221386, + 0.0035140307154506445, + 0.004438003525137901, + 0.008220219984650612, + 0.011334017850458622, + 0.001719186780974269, + 0.0014189941575750709, + 0.0013044006191194057, + 0.0008145925821736455, + 0.0011371441651135683, + 0.0016319185961037874, + 0.001844747574068606, + 0.0021407699678093195, + 0.006854134611785412, + 0.004534918814897537, + 0.007007901091128588, + 0.004951179958879948, + 0.004376268479973078, + 0.006019233260303736, + 0.03390897437930107, + 0.033274851739406586, + 0.001394200255163014, + 0.0009726329008117318, + 0.0011360947974026203, + 0.0015578006859868765, + 0.001851913402788341, + 0.0022088123951107264, + 0.01349684689193964, + 0.013013123534619808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.12664808332920074 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014227032661437988 + ], + "opposite_logprob_tokens": [ + -4.264226913452148 + ], + "logprob_reward": -0.014227032661437988, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014227032661437988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44113433361053467, + 0.09505372494459152, + 0.03937278315424919, + 0.023146262392401695, + 0.13549067080020905, + 0.03477426618337631, + 0.009498793631792068, + 0.017748409882187843, + 0.009266103617846966, + 0.19451461732387543 + ], + "credit_map": [ + 0.02652493678033352, + 0.026827599853277206, + 0.016878832131624222, + 0.0162760429084301, + 0.023366596549749374, + 0.029833657667040825, + 0.1453152596950531, + 0.15611141920089722, + 0.008882525376975536, + 0.005133843049407005, + 0.005785111803561449, + 0.0053262850269675255, + 0.0060435025952756405, + 0.00788310170173645, + 0.029992446303367615, + 0.02600691094994545, + 0.001649474143050611, + 0.0032032637391239405, + 0.0042289188131690025, + 0.004665219224989414, + 0.0031106146052479744, + 0.005428320728242397, + 0.007920301519334316, + 0.009166669100522995, + 0.0020948268938809633, + 0.002723895711824298, + 0.002570122480392456, + 0.0021028972696512938, + 0.003639503149315715, + 0.004033189732581377, + 0.0021000714041292667, + 0.003881756216287613, + 0.00991776492446661, + 0.0093725910410285, + 0.006073229480534792, + 0.005745564587414265, + 0.018403582274913788, + 0.02116742730140686, + 0.03064384125173092, + 0.034166671335697174, + 0.006031165365129709, + 0.0049114045687019825, + 0.004430196713656187, + 0.005148520693182945, + 0.0034246649593114853, + 0.003311420325189829, + 0.003634244902059436, + 0.003882649587467313, + 0.0006049260264262557, + 0.000705055077560246, + 0.0011137545807287097, + 0.001124876900576055, + 0.0010891283163800836, + 0.0018152239499613643, + 0.001177162048406899, + 0.0018686669645830989, + 0.0033023611176759005, + 0.0032583747524768114, + 0.0017225523479282856, + 0.0018426073947921395, + 0.0025427022483199835, + 0.0022016887087374926, + 0.001345178228802979, + 0.0015329456655308604, + 0.001317999791353941, + 0.0009088741498999298, + 0.0010425805812701583, + 0.0010912761790677905, + 0.001035476103425026, + 0.0010760397417470813, + 0.0012765382416546345, + 0.0015173193532973528, + 0.007379302754998207, + 0.003755641169846058, + 0.005737525876611471, + 0.008123896084725857, + 0.007755858823657036, + 0.009915194474160671, + 0.07744241505861282, + 0.0744047686457634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.014227032661437988 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306522756814957 + ], + "opposite_logprob_tokens": [ + -3.168065309524536 + ], + "logprob_reward": -0.04306522756814957, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306522756814957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7056074738502502, + 0.025799332186579704, + 0.04573895037174225, + 0.006628133822232485, + 0.10336709767580032, + 0.01793692447245121, + 0.010006706230342388, + 0.017221638932824135, + 0.012190197594463825, + 0.055503539741039276 + ], + "credit_map": [ + 0.030017131939530373, + 0.013222227804362774, + 0.02046034298837185, + 0.03639901429414749, + 0.02489212527871132, + 0.03620447590947151, + 0.29427316784858704, + 0.2501389980316162, + 0.002019014675170183, + 0.0022647837176918983, + 0.001047389698214829, + 0.0015730467857792974, + 0.002101210178807378, + 0.0017447673017159104, + 0.006796074565500021, + 0.008253045380115509, + 0.003293054411187768, + 0.005029668100178242, + 0.0025385459885001183, + 0.0024508119095116854, + 0.004144392907619476, + 0.0045354957692325115, + 0.010706393048167229, + 0.013040589168667793, + 0.0004040838102810085, + 0.00047718288260512054, + 0.0010409931419417262, + 0.0010955117177218199, + 0.0008964248700067401, + 0.0011525804875418544, + 0.0007465383387170732, + 0.0008148188935592771, + 0.007561190519481897, + 0.006791610270738602, + 0.0054606059566140175, + 0.005175513215363026, + 0.012179831974208355, + 0.013770170509815216, + 0.024657242000102997, + 0.027770934626460075, + 0.002859956817701459, + 0.0028287661261856556, + 0.0013489836128428578, + 0.00239410437643528, + 0.0019552684389054775, + 0.0025978456251323223, + 0.002222157549113035, + 0.0017298422753810883, + 0.0007458268664777279, + 0.0006021159933879972, + 0.001444262103177607, + 0.0014169252244755626, + 0.0011517625534906983, + 0.0016359769506379962, + 0.0013051442801952362, + 0.0017046924913302064, + 0.002518830355256796, + 0.0018712627934291959, + 0.0014524938305839896, + 0.001471889205276966, + 0.002505214884877205, + 0.0028048944659531116, + 0.002345256507396698, + 0.0022517957258969545, + 0.00043306368752382696, + 0.00029843556694686413, + 0.0005287755629979074, + 0.0007218625978566706, + 0.0005227342480793595, + 0.0006992071284912527, + 0.004489690996706486, + 0.004496428184211254, + 0.0018961528548970819, + 0.0016339996363967657, + 0.001733517972752452, + 0.0025208794977515936, + 0.002334835473448038, + 0.0028177874628454447, + 0.021490568295121193, + 0.02107580006122589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.04306522756814957 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347240433096886 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347240433096886, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347240433096886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058066822588443756, + 0.5181919932365417, + 0.252689927816391, + 0.007295574992895126, + 0.02194712683558464, + 0.04720716178417206, + 0.005279242061078548, + 0.03177843242883682, + 0.0046218931674957275, + 0.05292186513543129 + ], + "credit_map": [ + 0.007752411998808384, + 0.005221227649599314, + 0.005956811364740133, + 0.005348312668502331, + 0.010019617155194283, + 0.008590552024543285, + 0.01057722419500351, + 0.004600668791681528, + 0.016800932586193085, + 0.012274887412786484, + 0.015557785518467426, + 0.020400187000632286, + 0.020225653424859047, + 0.025690216571092606, + 0.2036598175764084, + 0.20358248054981232, + 0.01699225604534149, + 0.024453606456518173, + 0.013987550511956215, + 0.011498761363327503, + 0.028799643740057945, + 0.036839988082647324, + 0.05550345405936241, + 0.0646146610379219, + 0.00043143643415533006, + 0.0006443033344112337, + 0.0003846465842798352, + 0.0006819276022724807, + 0.001954018836840987, + 0.0018512486713007092, + 0.0005506892339326441, + 0.0007973046158440411, + 0.0022341511212289333, + 0.0017739603063091636, + 0.0015314915217459202, + 0.0010946700349450111, + 0.0017028864240273833, + 0.0023894852492958307, + 0.005325155332684517, + 0.005895325914025307, + 0.00363584840670228, + 0.006851628422737122, + 0.0028039589524269104, + 0.003275215392932296, + 0.004848198965191841, + 0.009593775495886803, + 0.007658184040337801, + 0.008540351875126362, + 0.00032219375134445727, + 0.00026166377938352525, + 0.0008616921259090304, + 0.0006365465233102441, + 0.0004847019154112786, + 0.0005808600108139217, + 0.0009621858480386436, + 0.001169398077763617, + 0.0019139917567372322, + 0.003241143189370632, + 0.0012355437502264977, + 0.0014579809503629804, + 0.0031828118953853846, + 0.003625797340646386, + 0.007926623336970806, + 0.009194537997245789, + 0.0007564628613181412, + 0.0005318376934155822, + 0.0005322457291185856, + 0.0003720391250681132, + 0.000499094370752573, + 0.0006686871638521552, + 0.0005961621645838022, + 0.0006653638556599617, + 0.0018731217132881284, + 0.0010128030553460121, + 0.0014487635344266891, + 0.002216764260083437, + 0.0020374220330268145, + 0.002742914715781808, + 0.0211593396961689, + 0.020430738106369972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.026347240433096886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288974106311798 + ], + "opposite_logprob_tokens": [ + -1.0538973808288574 + ], + "logprob_reward": -0.4288974106311798, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288974106311798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6454356908798218, + 0.034819766879081726, + 0.012515786103904247, + 0.012363221496343613, + 0.042605772614479065, + 0.006011342629790306, + 0.017571356147527695, + 0.011265221051871777, + 0.06846249103546143, + 0.14894941449165344 + ], + "credit_map": [ + 0.0240625087171793, + 0.011932633817195892, + 0.019420698285102844, + 0.028518881648778915, + 0.029607320204377174, + 0.03394659236073494, + 0.2602902054786682, + 0.2376568466424942, + 0.0010752431116998196, + 0.0011680200695991516, + 0.0013215660583227873, + 0.0015811420744284987, + 0.0012913927203044295, + 0.001676698331721127, + 0.013347034342586994, + 0.01335867214947939, + 0.0017926625441759825, + 0.001139861298725009, + 0.0016843370394781232, + 0.0015857800608500838, + 0.0014891454484313726, + 0.0015767097938805819, + 0.0017335331067442894, + 0.0015137565787881613, + 0.0007136776112020016, + 0.0009013100643642247, + 0.0013460746267810464, + 0.0016281118150800467, + 0.0011986573226749897, + 0.0015960234450176358, + 0.00218965788371861, + 0.00278970948420465, + 0.002238699933513999, + 0.0026220923755317926, + 0.0016246519517153502, + 0.0017336792079731822, + 0.00702694384381175, + 0.00766739109531045, + 0.009280137717723846, + 0.010412178933620453, + 0.0006608205148950219, + 0.00039785628905519843, + 0.0006571677513420582, + 0.0008553543011657894, + 0.0007266282918862998, + 0.0009975660359486938, + 0.0008761583594605327, + 0.0008397910860367119, + 0.0024383242707699537, + 0.001627721474505961, + 0.0023079754319041967, + 0.0017667087959125638, + 0.0020583185832947493, + 0.0033124687615782022, + 0.001682489411905408, + 0.0023773489519953728, + 0.0017750076949596405, + 0.0013492750003933907, + 0.0011289045214653015, + 0.0009371010237373412, + 0.0014609539648517966, + 0.0015988301020115614, + 0.001096096122637391, + 0.0019190524471923709, + 0.005169669631868601, + 0.003136912826448679, + 0.003994811326265335, + 0.00309949042275548, + 0.0026469321455806494, + 0.003978831227868795, + 0.02406921237707138, + 0.02236662246286869, + 0.0060356250032782555, + 0.005099962931126356, + 0.005301860626786947, + 0.00793908629566431, + 0.005331606604158878, + 0.00843333825469017, + 0.05686523765325546, + 0.05394268035888672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.4288974106311798 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "resolute, locked-off split-horizon shot captures a captivating natural contradiction and the rural charm of a pond-filled landscape. Above the waterline, the surface is a tranquil, unbroken mosaic of distinctive, diamond-shaped water chestnut leaves, their waxy, deep-green surfaces floating in perfect stillness under the soft, golden glow of the late afternoon sun. Beneath this tranquil exterior, the clear, tea-stained water reveals a vibrant world. A powerful, unseen deep current causes the slender, green-brown underwater stems to perform a unified, mesmerizing dance. They sway gracefully yet relentlessly in one direction, bending and straightening in hypnotic, synchronized waves, their feathery underwater roots trailing like silk banners. Sparkling patterns of light created by sunlight filtering through the water's surface, punctuated by the swaying stems and the silty bottom, highlight the contrast between the serene, static world above and the powerful, rhythmic energy pulsating below, creating an atmosphere both profoundly tranquil and vibrant.", + "seed": 32618, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_11/index_4533_seed_32618_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07895617932081223, + "vlm_logprob": -0.07895617932081223, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.07895617932081223, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04344938322901726 + ], + "opposite_logprob_tokens": [ + -3.4184494018554688 + ], + "logprob_reward": -0.04344938322901726, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04344938322901726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08945438265800476, + 0.3263864815235138, + 0.07470596581697464, + 0.037709791213274, + 0.06565531343221664, + 0.02002732641994953, + 0.05166298523545265, + 0.11162886023521423, + 0.0985642820596695, + 0.12420453876256943 + ], + "credit_map": [ + 0.007336352486163378, + 0.011750124394893646, + 0.012594000436365604, + 0.019500145688652992, + 0.011914584785699844, + 0.006832853425294161, + 0.011588061228394508, + 0.007938271388411522, + 0.051648370921611786, + 0.07005477696657181, + 0.04285407066345215, + 0.03621552139520645, + 0.0328391008079052, + 0.019576843827962875, + 0.04778221622109413, + 0.02541559748351574, + 0.007682815194129944, + 0.008409549482166767, + 0.009923300705850124, + 0.011759894900023937, + 0.011709108017385006, + 0.00817667692899704, + 0.009820051491260529, + 0.007224577479064465, + 0.011253759264945984, + 0.01384106557816267, + 0.00239648949354887, + 0.002353898948058486, + 0.002600315259769559, + 0.001714419573545456, + 0.0019149837316945195, + 0.0016348641365766525, + 0.020493201911449432, + 0.025537552312016487, + 0.003073187777772546, + 0.0028644718695431948, + 0.0058941347524523735, + 0.002935765776783228, + 0.002538914792239666, + 0.0023180926218628883, + 0.0027602415066212416, + 0.0027719323989003897, + 0.002493031322956085, + 0.002447060775011778, + 0.00308398948982358, + 0.002837573643773794, + 0.0018735091434791684, + 0.0017599877901375294, + 0.007234849967062473, + 0.010109171271324158, + 0.0073055787943303585, + 0.007215745281428099, + 0.005405803676694632, + 0.0038106534630060196, + 0.006277599837630987, + 0.004303584806621075, + 0.011851456947624683, + 0.013920704834163189, + 0.017587076872587204, + 0.01656109094619751, + 0.014529907144606113, + 0.00864445324987173, + 0.016343651339411736, + 0.012190529145300388, + 0.017275094985961914, + 0.0225093774497509, + 0.014081365428864956, + 0.01377247367054224, + 0.00776711106300354, + 0.006354401353746653, + 0.008807308040559292, + 0.007997153326869011, + 0.0189504474401474, + 0.024897459894418716, + 0.018311453983187675, + 0.01602105051279068, + 0.014913626946508884, + 0.008378986269235611, + 0.012623870745301247, + 0.010107649490237236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.04344938322901726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029834769666194916 + ], + "opposite_logprob_tokens": [ + -3.529834747314453 + ], + "logprob_reward": -0.029834769666194916, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029834769666194916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18555313348770142, + 0.17791520059108734, + 0.09258654713630676, + 0.05564619228243828, + 0.11172368377447128, + 0.10937214642763138, + 0.024935321882367134, + 0.030169017612934113, + 0.06235276907682419, + 0.14974600076675415 + ], + "credit_map": [ + 0.021245183423161507, + 0.03132180497050285, + 0.018881656229496002, + 0.02778327278792858, + 0.032319970428943634, + 0.016632044687867165, + 0.02045402303338051, + 0.01691517047584057, + 0.032282616943120956, + 0.04285740107297897, + 0.022789591923356056, + 0.022132741287350655, + 0.015838313847780228, + 0.00875253789126873, + 0.02102387510240078, + 0.012238115072250366, + 0.012189751490950584, + 0.014935133047401905, + 0.00987619161605835, + 0.012279350310564041, + 0.013761700130999088, + 0.008238268084824085, + 0.011472322046756744, + 0.00983382761478424, + 0.012693631462752819, + 0.01571773923933506, + 0.005162164568901062, + 0.005191315896809101, + 0.005900926422327757, + 0.003665320575237274, + 0.0038750243838876486, + 0.003440071363002062, + 0.021434452384710312, + 0.024504872038960457, + 0.011159667745232582, + 0.009569935500621796, + 0.022681469097733498, + 0.009307199157774448, + 0.00720174890011549, + 0.005864342674612999, + 0.018697012215852737, + 0.023535231128335, + 0.01173852477222681, + 0.012939647771418095, + 0.01528274267911911, + 0.008355767466127872, + 0.010428489185869694, + 0.008394725620746613, + 0.0041957576759159565, + 0.005813548341393471, + 0.003170914947986603, + 0.0026663695462048054, + 0.0030626626685261726, + 0.0019986978732049465, + 0.002137475647032261, + 0.0018898961134254932, + 0.0033891082275658846, + 0.0038858787156641483, + 0.0050870818085968494, + 0.005081187002360821, + 0.003617377020418644, + 0.0022116948384791613, + 0.004244914278388023, + 0.002651774790138006, + 0.008707016706466675, + 0.012199760414659977, + 0.005344907287508249, + 0.010817180387675762, + 0.007091662846505642, + 0.0035862259101122618, + 0.007966488599777222, + 0.006639523897320032, + 0.022297699004411697, + 0.030421322211623192, + 0.02001037262380123, + 0.021336905658245087, + 0.014887416735291481, + 0.009530287235975266, + 0.0180966779589653, + 0.013165315613150597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.029834769666194916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634689211845398 + ], + "opposite_logprob_tokens": [ + -3.6513469219207764 + ], + "logprob_reward": -0.02634689211845398, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634689211845398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16702264547348022, + 0.2108927220106125, + 0.13068467378616333, + 0.0573863722383976, + 0.0851684957742691, + 0.07291311025619507, + 0.056332580745220184, + 0.0644114762544632, + 0.09754256904125214, + 0.05764535069465637 + ], + "credit_map": [ + 0.013838231563568115, + 0.021682942286133766, + 0.022766146808862686, + 0.02925047278404236, + 0.023597434163093567, + 0.014452783390879631, + 0.023538878187537193, + 0.01789575256407261, + 0.02909809537231922, + 0.03770096227526665, + 0.028428569436073303, + 0.024022160097956657, + 0.02326330915093422, + 0.018545638769865036, + 0.030594559386372566, + 0.019239434972405434, + 0.01087934710085392, + 0.0149912778288126, + 0.019851040095090866, + 0.02554512955248356, + 0.018406597897410393, + 0.013061702251434326, + 0.015088447369635105, + 0.012861131690442562, + 0.005849358160048723, + 0.008595603518188, + 0.00757258664816618, + 0.00589703768491745, + 0.01018679328262806, + 0.00788107793778181, + 0.005514731165021658, + 0.005889185704290867, + 0.017455268651247025, + 0.022639987990260124, + 0.007711722981184721, + 0.0073362840339541435, + 0.011499844491481781, + 0.008136400021612644, + 0.0057637738063931465, + 0.004625212866812944, + 0.013529707677662373, + 0.018189234659075737, + 0.007294591516256332, + 0.00753509858623147, + 0.007587143685668707, + 0.006490402389317751, + 0.00721703888848424, + 0.005069896578788757, + 0.009867042303085327, + 0.013326114043593407, + 0.00521849375218153, + 0.007198801264166832, + 0.006247979123145342, + 0.004049825482070446, + 0.0059305522590875626, + 0.004493772052228451, + 0.005697343498468399, + 0.007351824548095465, + 0.008393412455916405, + 0.010987724177539349, + 0.008624494075775146, + 0.005027914885431528, + 0.010747639462351799, + 0.007581124547868967, + 0.014493163675069809, + 0.01897105947136879, + 0.010675384663045406, + 0.01824258081614971, + 0.009420100599527359, + 0.005487235728651285, + 0.012718157842755318, + 0.007534886710345745, + 0.008067048154771328, + 0.009510907344520092, + 0.007767581846565008, + 0.009434990584850311, + 0.005940892267972231, + 0.003723128465935588, + 0.006539352238178253, + 0.0066614518873393536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02634689211845398 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306088760495186 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.04306088760495186, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306088760495186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14490799605846405, + 0.18533743917942047, + 0.08977781236171722, + 0.12758107483386993, + 0.047663792967796326, + 0.04856834188103676, + 0.026880823075771332, + 0.09605801850557327, + 0.09979503601789474, + 0.1334296464920044 + ], + "credit_map": [ + 0.010721288621425629, + 0.014822876080870628, + 0.015817994251847267, + 0.04334017261862755, + 0.01479047629982233, + 0.011339319869875908, + 0.017036329954862595, + 0.017039531841874123, + 0.026999900117516518, + 0.03741499409079552, + 0.026606980711221695, + 0.022347694262862206, + 0.015013335272669792, + 0.010458260774612427, + 0.030485764145851135, + 0.01601051539182663, + 0.011003308929502964, + 0.011255085468292236, + 0.011464085429906845, + 0.008905134163796902, + 0.021518532186746597, + 0.008186643011868, + 0.007953813299536705, + 0.00949121080338955, + 0.01933710277080536, + 0.023848939687013626, + 0.013777239248156548, + 0.016343414783477783, + 0.015102888457477093, + 0.012850317172706127, + 0.0146742919459939, + 0.011646886356174946, + 0.006736759562045336, + 0.009112266823649406, + 0.004400122445076704, + 0.0045207804068923, + 0.011305407620966434, + 0.004346445202827454, + 0.003894719062373042, + 0.003347295569255948, + 0.00889140646904707, + 0.011259103193879128, + 0.004606717266142368, + 0.004409419372677803, + 0.008260895498096943, + 0.004123162478208542, + 0.0034907397348433733, + 0.003526900662109256, + 0.003017884213477373, + 0.00390395219437778, + 0.0036511674989014864, + 0.0035968113224953413, + 0.004197541158646345, + 0.00259934039786458, + 0.0027367696166038513, + 0.0031773580703884363, + 0.01533540803939104, + 0.01893637329339981, + 0.010307424701750278, + 0.010482694022357464, + 0.016605213284492493, + 0.007992841303348541, + 0.008266087621450424, + 0.008131973445415497, + 0.014393141493201256, + 0.018074067309498787, + 0.015442263334989548, + 0.01308542862534523, + 0.011307865381240845, + 0.0067079318687319756, + 0.0097650196403265, + 0.011019316501915455, + 0.017801648005843163, + 0.024424046277999878, + 0.019116133451461792, + 0.019693486392498016, + 0.01636260375380516, + 0.009510866366326809, + 0.01454065926373005, + 0.01198019739240408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.04306088760495186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a split-horizon view with a clear waterline separating the surface from the underwater scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are diamond-shaped water chestnut leaves visible on the water surface with waxy, deep-green surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green-brown underwater stems visible swaying in a unified, synchronized wave motion beneath the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by the soft, golden glow of late afternoon sunlight filtering through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera locked-off (stationary) throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520889639854431 + ], + "opposite_logprob_tokens": [ + -1.502089023590088 + ], + "logprob_reward": -0.2520889639854431, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520889639854431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3322676122188568, + 0.04120982438325882, + 0.15056294202804565, + 0.06731384992599487, + 0.14113304018974304, + 0.06986568123102188, + 0.04318520054221153, + 0.03835740685462952, + 0.059043366461992264, + 0.0570610910654068 + ], + "credit_map": [ + 0.05727793648838997, + 0.06785410642623901, + 0.030467184260487556, + 0.03696618229150772, + 0.04738805815577507, + 0.032184988260269165, + 0.034302715212106705, + 0.02582644484937191, + 0.00397503050044179, + 0.0038264738395810127, + 0.007827662862837315, + 0.005729401484131813, + 0.004803715273737907, + 0.003987576346844435, + 0.006525755859911442, + 0.004534208215773106, + 0.027086880058050156, + 0.03558703511953354, + 0.021325066685676575, + 0.014080425724387169, + 0.016814934089779854, + 0.012633374892175198, + 0.011602511629462242, + 0.011432715691626072, + 0.01437171921133995, + 0.017324594780802727, + 0.006194791290909052, + 0.0059381756000220776, + 0.008541126735508442, + 0.005226560402661562, + 0.005669059697538614, + 0.004047824069857597, + 0.03974582254886627, + 0.04721813276410103, + 0.006530527491122484, + 0.0074140699580311775, + 0.019188517704606056, + 0.00816048588603735, + 0.00609193928539753, + 0.006783536169677973, + 0.009435412473976612, + 0.007902915589511395, + 0.007224531844258308, + 0.008612405508756638, + 0.017381887882947922, + 0.007415361702442169, + 0.005826118867844343, + 0.006067049689590931, + 0.008734649047255516, + 0.011074481531977654, + 0.003546083578839898, + 0.005486528854817152, + 0.00521572632715106, + 0.0021792491897940636, + 0.0032317161094397306, + 0.003716766834259033, + 0.004032805096358061, + 0.005369998048990965, + 0.005726569797843695, + 0.005363808944821358, + 0.006391209550201893, + 0.003536252537742257, + 0.003848569467663765, + 0.004088195972144604, + 0.006995049770921469, + 0.009503255598247051, + 0.00925541203469038, + 0.01002356968820095, + 0.006756972521543503, + 0.003885585116222501, + 0.006324770860373974, + 0.006298752501606941, + 0.00611862912774086, + 0.007076351437717676, + 0.007558358833193779, + 0.010751924477517605, + 0.0066513256169855595, + 0.004330446012318134, + 0.008237481117248535, + 0.006336574908345938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.2520889639854431 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_12/index_tensor([6217, 6217])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_12/index_tensor([6217, 6217])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1bc391aae626b1771b70963ba15f272040e94af1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_12/index_tensor([6217, 6217])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in a tense documentary style, a shaky, in-car camera follows a racer from the passenger seat as he speeds along a bumpy dirt mountain road in a rally car. The man, with sharp features and short, messy brown hair drenched in sweat, wears a dark gray professional racing suit stained with dirt and sweat. His expression is one of intense concentration, his lips pursed, his eyes fixed sharply on the winding road ahead. His left hand, clad in a worn black racing glove, controls the steering wheel with a steady, powerful grip, quickly and precisely adjusting the car's position with every bump. His right hand, muscles tensed, his knuckles white with exertion, grips the grip of a modern black tactical rifle. A thin layer of yellow-brown dust coats the rifle's barrel, scope, and stock, which bounces up and down with every violent jolt of the car. The dazzling afternoon sun streamed in at a low angle through the driver's side window, casting sharp shadows that flickered rapidly, lengthened, and deformed across the simple, roll-cage-filled cabin. These shadows swept across the driver's focused face and the cold weapon, and together with the billowing dust outside the car, created a high-risk dynamic and tense atmosphere that was about to explode.", + "seed": 17024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_12/index_6217_seed_17024_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09993719905614853, + "vlm_logprob": -0.09993719905614853, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.09993719905614853, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.042297955602407455 + ], + "opposite_logprob_tokens": [ + -3.542297840118408 + ], + "logprob_reward": -0.042297955602407455, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.042297955602407455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05396918207406998, + 0.325164794921875, + 0.047751035541296005, + 0.049861010164022446, + 0.028987064957618713, + 0.32831811904907227, + 0.03831413388252258, + 0.048024438321590424, + 0.03813990205526352, + 0.04147034138441086 + ], + "credit_map": [ + 0.006518866401165724, + 0.0069657270796597, + 0.005233391188085079, + 0.003922063857316971, + 0.005409063771367073, + 0.006776662543416023, + 0.009597049094736576, + 0.009546355344355106, + 0.03708801791071892, + 0.06093248724937439, + 0.009350527077913284, + 0.02333623543381691, + 0.0554470494389534, + 0.01318929623812437, + 0.11167272180318832, + 0.01414843276143074, + 0.0048065814189612865, + 0.010563565418124199, + 0.0054458570666611195, + 0.0020273649133741856, + 0.003371046856045723, + 0.0034976128954440355, + 0.005063024815171957, + 0.012975976802408695, + 0.00504315784201026, + 0.00464606424793601, + 0.006769627798348665, + 0.004653286654502153, + 0.0065142507664859295, + 0.004864604212343693, + 0.006126367021352053, + 0.01124364510178566, + 0.006729788146913052, + 0.005745365284383297, + 0.004580711014568806, + 0.0018361682305112481, + 0.0032880522776395082, + 0.001969970064237714, + 0.0028229570016264915, + 0.002014052588492632, + 0.13214223086833954, + 0.12398175150156021, + 0.01110914722084999, + 0.011769544333219528, + 0.01892603375017643, + 0.00835818238556385, + 0.013039406388998032, + 0.00899184588342905, + 0.005965924356132746, + 0.008349524810910225, + 0.0031136772595345974, + 0.003358478657901287, + 0.0059968954883515835, + 0.0035696597769856453, + 0.005081757437437773, + 0.0028782105073332787, + 0.01319133210927248, + 0.007614747155457735, + 0.002707553096115589, + 0.00340081425383687, + 0.004663053900003433, + 0.002671879017725587, + 0.010213234461843967, + 0.003561819903552532, + 0.005192126613110304, + 0.006489851046353579, + 0.0024927370250225067, + 0.0030136846471577883, + 0.006926579400897026, + 0.0027560628950595856, + 0.008895418606698513, + 0.0023734401911497116, + 0.005505042150616646, + 0.004701259080320597, + 0.005134867038577795, + 0.0037594453897327185, + 0.007390796672552824, + 0.0037593955639749765, + 0.006870876532047987, + 0.004348657559603453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.042297955602407455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038135167211294174 + ], + "opposite_logprob_tokens": [ + -3.288135051727295 + ], + "logprob_reward": -0.038135167211294174, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038135167211294174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07443305104970932, + 0.5074666142463684, + 0.06133376434445381, + 0.025727292522788048, + 0.008237956091761589, + 0.1484140008687973, + 0.016560133546590805, + 0.07019923627376556, + 0.07202059775590897, + 0.015607320703566074 + ], + "credit_map": [ + 0.010805787518620491, + 0.0090708639472723, + 0.005816961172968149, + 0.005696400068700314, + 0.0062743984162807465, + 0.008717664517462254, + 0.01149192824959755, + 0.016559040173888206, + 0.07541168481111526, + 0.12032034993171692, + 0.015507468953728676, + 0.036843739449977875, + 0.06989150494337082, + 0.01861477829515934, + 0.1510213315486908, + 0.019855760037899017, + 0.006297045852988958, + 0.018199188634753227, + 0.0030847550369799137, + 0.002022291300818324, + 0.004026528913527727, + 0.002638753969222307, + 0.005789655726402998, + 0.019275542348623276, + 0.002215463435277343, + 0.002278505126014352, + 0.0033451239578425884, + 0.002317548729479313, + 0.00504084350541234, + 0.0025981469079852104, + 0.0030902232974767685, + 0.0048414370976388454, + 0.0013825512723997235, + 0.0012407227186486125, + 0.0012766644358634949, + 0.0005757301114499569, + 0.001044982927851379, + 0.0008964602602645755, + 0.000969282235018909, + 0.0008515614317730069, + 0.05439257249236107, + 0.05264488235116005, + 0.0058144619688391685, + 0.006396532524377108, + 0.009992462582886219, + 0.005442477762699127, + 0.008675768040120602, + 0.005054839886724949, + 0.002270323922857642, + 0.0020846142433583736, + 0.0019305105088278651, + 0.0013377043651416898, + 0.002517270389944315, + 0.002048739930614829, + 0.0021467048209160566, + 0.0022242667619138956, + 0.01993316225707531, + 0.010868288576602936, + 0.004498471971601248, + 0.006067183334380388, + 0.005426641553640366, + 0.002420505741611123, + 0.017867783084511757, + 0.003117199521511793, + 0.008347398601472378, + 0.006982884369790554, + 0.007813658565282822, + 0.0059099867939949036, + 0.016256961971521378, + 0.006792712491005659, + 0.014834732748568058, + 0.005082258954644203, + 0.0022843037731945515, + 0.0024030786007642746, + 0.0017731827683746815, + 0.0016034940490499139, + 0.002344179665669799, + 0.001002248958684504, + 0.002484238240867853, + 0.0017125942977145314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.038135167211294174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989970743656158 + ], + "opposite_logprob_tokens": [ + -1.8048996925354004 + ], + "logprob_reward": -0.17989970743656158, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989970743656158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1474016159772873, + 0.16662783920764923, + 0.08297237753868103, + 0.06731623411178589, + 0.044703297317028046, + 0.2722107768058777, + 0.021667953580617905, + 0.08896800130605698, + 0.053656335920095444, + 0.0544755682349205 + ], + "credit_map": [ + 0.018625503405928612, + 0.018917372450232506, + 0.010662239976227283, + 0.010833346284925938, + 0.01464231126010418, + 0.01346865575760603, + 0.035911306738853455, + 0.02434087172150612, + 0.01459795143455267, + 0.024305075407028198, + 0.008353538811206818, + 0.011544096283614635, + 0.032398421317338943, + 0.009038034826517105, + 0.05887597054243088, + 0.007514756638556719, + 0.0072679659351706505, + 0.02337665855884552, + 0.004761523567140102, + 0.00358762638643384, + 0.007049803622066975, + 0.00484510138630867, + 0.007268762681633234, + 0.024814939126372337, + 0.00820219051092863, + 0.00722320843487978, + 0.007705795578658581, + 0.004588205832988024, + 0.008298971690237522, + 0.006863834802061319, + 0.011062576435506344, + 0.013371454551815987, + 0.01121250819414854, + 0.007586895488202572, + 0.008260357193648815, + 0.002902393229305744, + 0.0038789943791925907, + 0.0042713056318461895, + 0.0037537706084549427, + 0.0028370670042932034, + 0.08174550533294678, + 0.08456902951002121, + 0.011433481238782406, + 0.013251133263111115, + 0.022589394822716713, + 0.016519874334335327, + 0.026562346145510674, + 0.015540004707872868, + 0.0031675028149038553, + 0.003024943871423602, + 0.002239012625068426, + 0.0015422927681356668, + 0.004177429247647524, + 0.0013260762207210064, + 0.003833167487755418, + 0.0023575276136398315, + 0.022041432559490204, + 0.012013014405965805, + 0.006426787003874779, + 0.006400415673851967, + 0.011081627570092678, + 0.005344454664736986, + 0.019720105454325676, + 0.005940163973718882, + 0.00981152430176735, + 0.007596874143928289, + 0.006075385957956314, + 0.004819037392735481, + 0.006724145263433456, + 0.0054376935586333275, + 0.007787016220390797, + 0.005404654424637556, + 0.009250374510884285, + 0.011216813698410988, + 0.004759949166327715, + 0.004542882088571787, + 0.008181032724678516, + 0.004668704234063625, + 0.007440749555826187, + 0.004415057133883238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.17989970743656158 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037297248840332 + ], + "opposite_logprob_tokens": [ + -1.9103729724884033 + ], + "logprob_reward": -0.16037297248840332, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037297248840332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07147424668073654, + 0.06649180501699448, + 0.08972285687923431, + 0.026359865441918373, + 0.030398907139897346, + 0.43429169058799744, + 0.03379056230187416, + 0.1323823779821396, + 0.0487791933119297, + 0.06630845367908478 + ], + "credit_map": [ + 0.008603335358202457, + 0.012700607068836689, + 0.005269082263112068, + 0.0068816314451396465, + 0.009030509740114212, + 0.009504339657723904, + 0.012376456521451473, + 0.007108285557478666, + 0.0161652322858572, + 0.021880649030208588, + 0.00450937869027257, + 0.0042812153697013855, + 0.004846869967877865, + 0.005208161659538746, + 0.005991313606500626, + 0.0036089839413762093, + 0.008366569876670837, + 0.023565704002976418, + 0.0059686098247766495, + 0.0032406668178737164, + 0.00713669927790761, + 0.007097430992871523, + 0.00897938385605812, + 0.025367796421051025, + 0.002748608123511076, + 0.0033395537175238132, + 0.0025744575541466475, + 0.0015159845352172852, + 0.00389080960303545, + 0.00290352082811296, + 0.005163513123989105, + 0.004223419353365898, + 0.007758510299026966, + 0.00474730646237731, + 0.006192747037857771, + 0.001937227207235992, + 0.0027210654225200415, + 0.002018847968429327, + 0.0031314347870647907, + 0.0018917698180302978, + 0.16493555903434753, + 0.15930625796318054, + 0.015618976205587387, + 0.016699310392141342, + 0.026602061465382576, + 0.01417490839958191, + 0.022870860993862152, + 0.01408377755433321, + 0.006372984033077955, + 0.004139477387070656, + 0.004944697488099337, + 0.002614735160022974, + 0.005536225624382496, + 0.0023534244392067194, + 0.004737240727990866, + 0.003091780934482813, + 0.03392012417316437, + 0.01746336556971073, + 0.009283519349992275, + 0.011982783675193787, + 0.009563618339598179, + 0.0038484144024550915, + 0.03974765166640282, + 0.006572900805622339, + 0.006289136596024036, + 0.007050565909594297, + 0.005063681863248348, + 0.0036483549047261477, + 0.008119098842144012, + 0.0040052542462944984, + 0.011625217273831367, + 0.0029778850730508566, + 0.010636505670845509, + 0.011067706160247326, + 0.005548417568206787, + 0.005590370856225491, + 0.013291170820593834, + 0.0050216070376336575, + 0.009530367329716682, + 0.005622312892228365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.16037297248840332 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898019254207611 + ], + "opposite_logprob_tokens": [ + -2.5789802074432373 + ], + "logprob_reward": -0.07898019254207611, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898019254207611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020047320052981377, + 0.07280135154724121, + 0.012321055866777897, + 0.009633327834308147, + 0.008026523515582085, + 0.8076694011688232, + 0.006978611461818218, + 0.023942608386278152, + 0.03364640101790428, + 0.004933467134833336 + ], + "credit_map": [ + 0.0022693516220897436, + 0.002026406116783619, + 0.00251975585706532, + 0.0024740498047322035, + 0.0020664981566369534, + 0.0027189054526388645, + 0.0039045840967446566, + 0.002067768247798085, + 0.011665690690279007, + 0.017989853397011757, + 0.002196748973801732, + 0.005218141712248325, + 0.009431234560906887, + 0.0026981330011039972, + 0.020568210631608963, + 0.003033330896869302, + 0.0014222207246348262, + 0.00338707584887743, + 0.0010203743586316705, + 0.0004442649078555405, + 0.0008222544565796852, + 0.0005334193701855838, + 0.0011840704828500748, + 0.0035073745530098677, + 0.0009426275501027703, + 0.0009812569478526711, + 0.0010648368624970317, + 0.000758004782255739, + 0.0015484675532206893, + 0.0008828292484395206, + 0.0010862767230719328, + 0.002369026653468609, + 0.001582892145961523, + 0.0019114518072456121, + 0.0009891472291201353, + 0.0005384072428569198, + 0.0010461211204528809, + 0.0005322391516529024, + 0.0008082900894805789, + 0.0006179736228659749, + 0.32162368297576904, + 0.30484870076179504, + 0.02717856876552105, + 0.030051233246922493, + 0.04616282880306244, + 0.021077074110507965, + 0.03409094735980034, + 0.022636283189058304, + 0.0010463569778949022, + 0.0006742891855537891, + 0.001082713482901454, + 0.0005788493435829878, + 0.001307827653363347, + 0.0005504079163074493, + 0.0009666404221206903, + 0.0007715256069786847, + 0.0071191550232470036, + 0.003826485015451908, + 0.0010807681828737259, + 0.001868786639533937, + 0.0018083369359374046, + 0.0006364525761455297, + 0.006556306034326553, + 0.0010463137878105044, + 0.004918265156447887, + 0.0029581233393400908, + 0.005001633428037167, + 0.0034155379980802536, + 0.005678820889443159, + 0.003058114554733038, + 0.005921665579080582, + 0.002694237744435668, + 0.0007643987773917615, + 0.000652490823995322, + 0.0006553475395776331, + 0.00039274164009839296, + 0.000668566208332777, + 0.00042434330680407584, + 0.0006313288467936218, + 0.0007442494388669729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.07898019254207611 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a tense documentary style, a shaky, in-car camera follows a racer from the passenger seat as he speeds along a bumpy dirt mountain road in a rally car. The man, with sharp features and short, messy brown hair drenched in sweat, wears a dark gray professional racing suit stained with dirt and sweat. His expression is one of intense concentration, his lips pursed, his eyes fixed sharply on the winding road ahead. His left hand, clad in a worn black racing glove, controls the steering wheel with a steady, powerful grip, quickly and precisely adjusting the car's position with every bump. His right hand, muscles tensed, his knuckles white with exertion, grips the grip of a modern black tactical rifle. A thin layer of yellow-brown dust coats the rifle's barrel, scope, and stock, which bounces up and down with every violent jolt of the car. The dazzling afternoon sun streamed in at a low angle through the driver's side window, casting sharp shadows that flickered rapidly, lengthened, and deformed across the simple, roll-cage-filled cabin. These shadows swept across the driver's focused face and the cold weapon, and together with the billowing dust outside the car, created a high-risk dynamic and tense atmosphere that was about to explode.", + "seed": 17025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_12/index_6217_seed_17025_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11985782533884048, + "vlm_logprob": -0.11985782533884048, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.11985782533884048, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.059788741171360016 + ], + "opposite_logprob_tokens": [ + -3.684788703918457 + ], + "logprob_reward": -0.059788741171360016, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.059788741171360016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0984589159488678, + 0.057698555290699005, + 0.06511629372835159, + 0.06371092051267624, + 0.08509935438632965, + 0.06383265554904938, + 0.21098920702934265, + 0.09473540633916855, + 0.15289542078971863, + 0.1074632853269577 + ], + "credit_map": [ + 0.007555178366601467, + 0.011250623501837254, + 0.013075929135084152, + 0.01406355481594801, + 0.013054374605417252, + 0.010905471630394459, + 0.016314908862113953, + 0.012238878756761551, + 0.00310459453612566, + 0.007991088554263115, + 0.00861769262701273, + 0.00893052201718092, + 0.006728965323418379, + 0.011827629990875721, + 0.00508827343583107, + 0.005409786943346262, + 0.002109403256326914, + 0.005156153813004494, + 0.004675934556871653, + 0.00876818411052227, + 0.015450915321707726, + 0.019425129517912865, + 0.0034654976334422827, + 0.00606507807970047, + 0.006123318336904049, + 0.00901755504310131, + 0.007544126361608505, + 0.0065666805021464825, + 0.010871144011616707, + 0.00833207368850708, + 0.006198713555932045, + 0.00905731599777937, + 0.005942399147897959, + 0.013832011260092258, + 0.013269578106701374, + 0.012339483946561813, + 0.010901100933551788, + 0.007284506689757109, + 0.010552994906902313, + 0.010977271012961864, + 0.007822751067578793, + 0.008772949688136578, + 0.00906191673129797, + 0.007958672940731049, + 0.009248915128409863, + 0.006416429299861193, + 0.006488427519798279, + 0.008062591776251793, + 0.015283538959920406, + 0.02738465927541256, + 0.03241216763854027, + 0.027340887114405632, + 0.04554285854101181, + 0.022669652476906776, + 0.02534109726548195, + 0.015014339238405228, + 0.00696905329823494, + 0.010930731892585754, + 0.017330925911664963, + 0.012904494069516659, + 0.015841558575630188, + 0.010500145144760609, + 0.011489528231322765, + 0.008768961764872074, + 0.009806683287024498, + 0.011109622195363045, + 0.012415374629199505, + 0.02734946645796299, + 0.045359592884778976, + 0.027900416404008865, + 0.00987025536596775, + 0.009084020741283894, + 0.009072327986359596, + 0.014484505169093609, + 0.012206817045807838, + 0.015701385214924812, + 0.021746926009655, + 0.012429949827492237, + 0.013181271031498909, + 0.008640103042125702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.059788741171360016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902524620294571 + ], + "opposite_logprob_tokens": [ + -2.5790252685546875 + ], + "logprob_reward": -0.07902524620294571, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902524620294571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12546728551387787, + 0.055293355137109756, + 0.05093597248196602, + 0.09701181948184967, + 0.05168125778436661, + 0.05180603638291359, + 0.201637864112854, + 0.07087135314941406, + 0.2535291612148285, + 0.04176595062017441 + ], + "credit_map": [ + 0.007774311117827892, + 0.016473351046442986, + 0.018980655819177628, + 0.019697438925504684, + 0.017438413575291634, + 0.015693269670009613, + 0.019767452031373978, + 0.009642377495765686, + 0.0024002159480005503, + 0.008743338286876678, + 0.004484768956899643, + 0.004972945433109999, + 0.011751169338822365, + 0.013914567418396473, + 0.003628338687121868, + 0.005398005712777376, + 0.0025106582324951887, + 0.007493156474083662, + 0.005722322966903448, + 0.007554991170763969, + 0.01006392389535904, + 0.006284030620008707, + 0.005657915957272053, + 0.005648971069604158, + 0.007845999673008919, + 0.013056110590696335, + 0.009464188478887081, + 0.009807508438825607, + 0.017956925556063652, + 0.02146073989570141, + 0.0075783818028867245, + 0.009841960854828358, + 0.005622031167149544, + 0.007389105390757322, + 0.006661181338131428, + 0.006543323863297701, + 0.0051053897477686405, + 0.006417450960725546, + 0.005433554295450449, + 0.008509216830134392, + 0.0036390237510204315, + 0.007185693364590406, + 0.005810147617012262, + 0.006897315848618746, + 0.011310742236673832, + 0.007426440715789795, + 0.0059388442896306515, + 0.003597828559577465, + 0.01626131683588028, + 0.0249882061034441, + 0.03873315826058388, + 0.03048861026763916, + 0.03266596794128418, + 0.018297385424375534, + 0.02347419038414955, + 0.01672901213169098, + 0.0046124872751533985, + 0.009919443167746067, + 0.011511996388435364, + 0.008605273440480232, + 0.012045539915561676, + 0.006888682954013348, + 0.01043250784277916, + 0.006855417974293232, + 0.006731375120580196, + 0.020216211676597595, + 0.01402369886636734, + 0.05421658605337143, + 0.08675740659236908, + 0.051394540816545486, + 0.012998857535421848, + 0.0071904598735272884, + 0.002773275366052985, + 0.006011547055095434, + 0.0042341710068285465, + 0.008055546320974827, + 0.008472112938761711, + 0.004808411002159119, + 0.0039282613433897495, + 0.0034826199989765882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.07902524620294571 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015780359506607 + ], + "opposite_logprob_tokens": [ + -1.7015780210494995 + ], + "logprob_reward": -0.2015780359506607, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015780359506607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09373962134122849, + 0.05093596875667572, + 0.09179491549730301, + 0.0413040891289711, + 0.15904483199119568, + 0.08189430087804794, + 0.18700557947158813, + 0.03573348745703697, + 0.10816562175750732, + 0.15038150548934937 + ], + "credit_map": [ + 0.00601680763065815, + 0.01165164727717638, + 0.017656497657299042, + 0.019503658637404442, + 0.012769069522619247, + 0.01146251242607832, + 0.00788639485836029, + 0.006793043576180935, + 0.0031868077348917723, + 0.007686140015721321, + 0.008106403984129429, + 0.008044437505304813, + 0.006593276746571064, + 0.008731847628951073, + 0.0042807189747691154, + 0.004306337330490351, + 0.005093602929264307, + 0.01515150722116232, + 0.011778503656387329, + 0.016161847859621048, + 0.01303481962531805, + 0.009727224707603455, + 0.009934023022651672, + 0.01091338787227869, + 0.003267282387241721, + 0.0056911651045084, + 0.006340016610920429, + 0.007244992069900036, + 0.0045773424208164215, + 0.005257181357592344, + 0.004498600494116545, + 0.004427511710673571, + 0.011798889376223087, + 0.02637403830885887, + 0.020945796743035316, + 0.0194865670055151, + 0.016591155901551247, + 0.019132031127810478, + 0.015643112361431122, + 0.029073255136609077, + 0.00580552825704217, + 0.01416916772723198, + 0.014015455730259418, + 0.012475361116230488, + 0.010772723704576492, + 0.008063027635216713, + 0.007431069388985634, + 0.009161971509456635, + 0.013398819603025913, + 0.02637331187725067, + 0.030179699882864952, + 0.02584988810122013, + 0.03257729858160019, + 0.019055835902690887, + 0.022763315588235855, + 0.016807425767183304, + 0.002627284498885274, + 0.005400463938713074, + 0.00631561828777194, + 0.005069158971309662, + 0.0037895527202636003, + 0.0029897368513047695, + 0.005034395959228277, + 0.0045072780922055244, + 0.005303781479597092, + 0.008429111912846565, + 0.007309590466320515, + 0.023690003901720047, + 0.03252466022968292, + 0.019678834825754166, + 0.005723284557461739, + 0.005506366956979036, + 0.00923354085534811, + 0.025886358693242073, + 0.027364129200577736, + 0.0276611540466547, + 0.016525674611330032, + 0.013539522886276245, + 0.016715383157134056, + 0.013455761596560478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.2015780359506607 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990997433662415 + ], + "opposite_logprob_tokens": [ + -1.8049099445343018 + ], + "logprob_reward": -0.17990997433662415, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990997433662415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09689263254404068, + 0.11228437721729279, + 0.05021011829376221, + 0.09205659478902817, + 0.04753086343407631, + 0.03806062042713165, + 0.1925271451473236, + 0.03189202398061752, + 0.27042192220687866, + 0.0681237056851387 + ], + "credit_map": [ + 0.009941894561052322, + 0.011273943819105625, + 0.012320591136813164, + 0.013355245813727379, + 0.012514897622168064, + 0.010769421234726906, + 0.012779830023646355, + 0.013936811126768589, + 0.007010993082076311, + 0.016127852723002434, + 0.013838590122759342, + 0.015349100343883038, + 0.02121291682124138, + 0.02522652968764305, + 0.0071550761349499226, + 0.006363322492688894, + 0.004430477041751146, + 0.0077378530986607075, + 0.0063704559579491615, + 0.007865224033594131, + 0.0050069489516317844, + 0.006779925897717476, + 0.006072933319956064, + 0.0059463027864694595, + 0.012133948504924774, + 0.009511020965874195, + 0.012623589485883713, + 0.009368376806378365, + 0.014121982268989086, + 0.013825216330587864, + 0.009678704664111137, + 0.010793761350214481, + 0.003476375713944435, + 0.007510658819228411, + 0.007583272643387318, + 0.006757660768926144, + 0.004255873151123524, + 0.005951611325144768, + 0.00438736192882061, + 0.007608047686517239, + 0.0032635850366204977, + 0.006552521605044603, + 0.0058167180977761745, + 0.005281329154968262, + 0.004240767564624548, + 0.003662722883746028, + 0.0037164820823818445, + 0.005526494700461626, + 0.014853591099381447, + 0.0272934939712286, + 0.033157072961330414, + 0.026644324883818626, + 0.02904193289577961, + 0.01888883486390114, + 0.023395495489239693, + 0.019252384081482887, + 0.0024459215346723795, + 0.004950163420289755, + 0.004558566492050886, + 0.004555086605250835, + 0.005482738371938467, + 0.0031365850009024143, + 0.0038691607769578695, + 0.002893805503845215, + 0.02109517902135849, + 0.0206898283213377, + 0.020724918693304062, + 0.04460097849369049, + 0.07962833344936371, + 0.04631452262401581, + 0.017759477719664574, + 0.019608670845627785, + 0.005544016137719154, + 0.011466218158602715, + 0.007632259279489517, + 0.015935663133859634, + 0.00721346028149128, + 0.00631837360560894, + 0.006516743451356888, + 0.007496976759284735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.17990997433662415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898712903261185 + ], + "opposite_logprob_tokens": [ + -2.5789871215820312 + ], + "logprob_reward": -0.07898712903261185, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898712903261185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07590539008378983, + 0.227727472782135, + 0.05731459707021713, + 0.09027432650327682, + 0.06567008793354034, + 0.048758797347545624, + 0.09966245293617249, + 0.10183192789554596, + 0.14524972438812256, + 0.08760521560907364 + ], + "credit_map": [ + 0.005148255731910467, + 0.008916052989661694, + 0.012425394728779793, + 0.015606953762471676, + 0.010816562920808792, + 0.008874881081283092, + 0.007506910245865583, + 0.0066103748977184296, + 0.016187280416488647, + 0.03602483496069908, + 0.026968343183398247, + 0.03975466266274452, + 0.04596831277012825, + 0.026340460404753685, + 0.02052658051252365, + 0.015956975519657135, + 0.006590662058442831, + 0.0061462209559977055, + 0.005358461290597916, + 0.0059822797775268555, + 0.01003868319094181, + 0.010467751882970333, + 0.005902918521314859, + 0.006827615667134523, + 0.007347981911152601, + 0.01283446978777647, + 0.010161059908568859, + 0.012813028879463673, + 0.01635090261697769, + 0.013625496067106724, + 0.0069984267465770245, + 0.010142959654331207, + 0.005747678689658642, + 0.006202608346939087, + 0.012934526428580284, + 0.007926124148070812, + 0.006195181980729103, + 0.008782336488366127, + 0.0063997781835496426, + 0.011481854133307934, + 0.004528324119746685, + 0.006642577704042196, + 0.0043204971589148045, + 0.0064204540103673935, + 0.010585862211883068, + 0.006949007511138916, + 0.005642243195325136, + 0.003669825615361333, + 0.004787576850503683, + 0.01357529778033495, + 0.013912403956055641, + 0.010751149617135525, + 0.02667551301419735, + 0.013779738917946815, + 0.009894434362649918, + 0.006286336574703455, + 0.008502793498337269, + 0.011815733276307583, + 0.014971021562814713, + 0.014669776894152164, + 0.016896674409508705, + 0.010849916376173496, + 0.014578677713871002, + 0.00954732857644558, + 0.01648850552737713, + 0.017402470111846924, + 0.015581421554088593, + 0.02663797326385975, + 0.01799464225769043, + 0.022398583590984344, + 0.012847795151174068, + 0.0158983264118433, + 0.009342389181256294, + 0.010102096013724804, + 0.011931157670915127, + 0.011133583262562752, + 0.011701639741659164, + 0.016039520502090454, + 0.009536685422062874, + 0.00781812984496355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.07898712903261185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a tense documentary style, a shaky, in-car camera follows a racer from the passenger seat as he speeds along a bumpy dirt mountain road in a rally car. The man, with sharp features and short, messy brown hair drenched in sweat, wears a dark gray professional racing suit stained with dirt and sweat. His expression is one of intense concentration, his lips pursed, his eyes fixed sharply on the winding road ahead. His left hand, clad in a worn black racing glove, controls the steering wheel with a steady, powerful grip, quickly and precisely adjusting the car's position with every bump. His right hand, muscles tensed, his knuckles white with exertion, grips the grip of a modern black tactical rifle. A thin layer of yellow-brown dust coats the rifle's barrel, scope, and stock, which bounces up and down with every violent jolt of the car. The dazzling afternoon sun streamed in at a low angle through the driver's side window, casting sharp shadows that flickered rapidly, lengthened, and deformed across the simple, roll-cage-filled cabin. These shadows swept across the driver's focused face and the cold weapon, and together with the billowing dust outside the car, created a high-risk dynamic and tense atmosphere that was about to explode.", + "seed": 17026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_12/index_6217_seed_17026_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06527075450867414, + "vlm_logprob": -0.06527075450867414, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.06527075450867414, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041074611246585846 + ], + "opposite_logprob_tokens": [ + -3.916074514389038 + ], + "logprob_reward": -0.041074611246585846, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.041074611246585846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23980017006397247, + 0.051624272018671036, + 0.06689795106649399, + 0.23208990693092346, + 0.044392406940460205, + 0.06337413191795349, + 0.09224846214056015, + 0.05256175249814987, + 0.12329094111919403, + 0.033719997853040695 + ], + "credit_map": [ + 0.020236928015947342, + 0.0378795750439167, + 0.025742940604686737, + 0.0436384454369545, + 0.03547310829162598, + 0.023194042965769768, + 0.03551466390490532, + 0.018120456486940384, + 0.005695183761417866, + 0.009436064399778843, + 0.004901397041976452, + 0.007909574545919895, + 0.009222676046192646, + 0.005485723726451397, + 0.005178764462471008, + 0.003794892458245158, + 0.0050556473433971405, + 0.010025445371866226, + 0.006396619137376547, + 0.00833058636635542, + 0.008824899792671204, + 0.00839589536190033, + 0.0086295111104846, + 0.01123934518545866, + 0.014752855524420738, + 0.01282420288771391, + 0.013487743213772774, + 0.08723925799131393, + 0.018969155848026276, + 0.04340631887316704, + 0.01646188460290432, + 0.024948475882411003, + 0.003767270129173994, + 0.007479576859623194, + 0.004202717915177345, + 0.005698388908058405, + 0.008788632228970528, + 0.006133688148111105, + 0.004166860599070787, + 0.004155270755290985, + 0.004066722001880407, + 0.009195619262754917, + 0.004900899715721607, + 0.010151875205338001, + 0.01057017408311367, + 0.007477556820958853, + 0.011189655400812626, + 0.005821630358695984, + 0.007852391339838505, + 0.01786002330482006, + 0.010722963139414787, + 0.015718165785074234, + 0.013028280809521675, + 0.015186452306807041, + 0.006200536619871855, + 0.005679645575582981, + 0.005444242153316736, + 0.007611881010234356, + 0.006269625388085842, + 0.008409947156906128, + 0.0053139058873057365, + 0.005318525712937117, + 0.010428192093968391, + 0.0037654319312423468, + 0.009637899696826935, + 0.010196280665695667, + 0.006887819617986679, + 0.011896762065589428, + 0.046444300562143326, + 0.024031473323702812, + 0.007393650710582733, + 0.006802761927247047, + 0.003988036885857582, + 0.005008540116250515, + 0.0030919748824089766, + 0.006272453349083662, + 0.004139834549278021, + 0.004753392655402422, + 0.0033946633338928223, + 0.0030711032450199127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.041074611246585846 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014230440370738506 + ], + "opposite_logprob_tokens": [ + -4.264230251312256 + ], + "logprob_reward": -0.014230440370738506, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014230440370738506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30840864777565, + 0.04353028163313866, + 0.03741902485489845, + 0.07279404997825623, + 0.16770625114440918, + 0.06072197109460831, + 0.05901332199573517, + 0.08056598901748657, + 0.09261522442102432, + 0.07722526788711548 + ], + "credit_map": [ + 0.02074037678539753, + 0.027413194999098778, + 0.019924841821193695, + 0.09094760566949844, + 0.07364514470100403, + 0.02056753635406494, + 0.03778763487935066, + 0.017382286489009857, + 0.003723692148923874, + 0.005901447031646967, + 0.005321149248629808, + 0.006532604806125164, + 0.006348375231027603, + 0.0051118372939527035, + 0.0069888378493487835, + 0.0036023396532982588, + 0.0024232883006334305, + 0.007271965499967337, + 0.0036287487018853426, + 0.004725061822682619, + 0.0037125886883586645, + 0.0048989118076860905, + 0.004618230275809765, + 0.006140229292213917, + 0.006294349208474159, + 0.004710747394710779, + 0.005025202874094248, + 0.021529052406549454, + 0.009855346754193306, + 0.00849275104701519, + 0.008771387860178947, + 0.008115204982459545, + 0.007760854437947273, + 0.01452587265521288, + 0.007508548907935619, + 0.04904957488179207, + 0.05896477773785591, + 0.01466651912778616, + 0.009813575074076653, + 0.005416523665189743, + 0.0030692685395479202, + 0.01431496161967516, + 0.003716881386935711, + 0.006825990043580532, + 0.01039212942123413, + 0.010598808526992798, + 0.007126264274120331, + 0.004677669610828161, + 0.003949470352381468, + 0.014102382585406303, + 0.004856820683926344, + 0.008831321261823177, + 0.0072897677309811115, + 0.011547877453267574, + 0.005217778496444225, + 0.00321790250018239, + 0.00553986756131053, + 0.012285449542105198, + 0.0059665595181286335, + 0.01120397262275219, + 0.010213375091552734, + 0.010545791126787663, + 0.01516646333038807, + 0.009644514881074429, + 0.008231651969254017, + 0.0070690070278942585, + 0.005715049337595701, + 0.012758227996528149, + 0.032345931977033615, + 0.014030558057129383, + 0.008370001800358295, + 0.004094796255230904, + 0.013134473003447056, + 0.013485872186720371, + 0.008208341896533966, + 0.013165338896214962, + 0.010163890197873116, + 0.0068876976147294044, + 0.006646540481597185, + 0.005533120129257441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.014230440370738506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705151736736298 + ], + "opposite_logprob_tokens": [ + -2.127051591873169 + ], + "logprob_reward": -0.12705151736736298, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705151736736298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20465074479579926, + 0.08678514510393143, + 0.055917512625455856, + 0.1605062633752823, + 0.19052670896053314, + 0.08637233823537827, + 0.09089718014001846, + 0.03666040673851967, + 0.05907991901040077, + 0.028603769838809967 + ], + "credit_map": [ + 0.013288083486258984, + 0.024036195129156113, + 0.017251530662178993, + 0.04058009013533592, + 0.054178476333618164, + 0.019876735284924507, + 0.02362436056137085, + 0.0118152666836977, + 0.008623999543488026, + 0.01165956724435091, + 0.006598272826522589, + 0.01062396727502346, + 0.021053588017821312, + 0.009099021553993225, + 0.01009077113121748, + 0.009035957045853138, + 0.0028212282340973616, + 0.010324062779545784, + 0.003972656559199095, + 0.010517139919102192, + 0.007757752668112516, + 0.006755914073437452, + 0.008987504057586193, + 0.004781254567205906, + 0.01013567391782999, + 0.010197707451879978, + 0.008267912082374096, + 0.05630249157547951, + 0.011327633634209633, + 0.034228332340717316, + 0.013196388259530067, + 0.016850123181939125, + 0.008327022194862366, + 0.015842661261558533, + 0.009461717680096626, + 0.053937237709760666, + 0.06416674703359604, + 0.01575549691915512, + 0.014929309487342834, + 0.008106518536806107, + 0.007348454091697931, + 0.01843567192554474, + 0.00629057502374053, + 0.01695512793958187, + 0.00849409680813551, + 0.008054382167756557, + 0.008716202341020107, + 0.012077822349965572, + 0.006499754264950752, + 0.015412641689181328, + 0.00731651159003377, + 0.01891094073653221, + 0.014469952322542667, + 0.015192736871540546, + 0.007397416979074478, + 0.005697227548807859, + 0.0023495422210544348, + 0.008665996603667736, + 0.0020217301789671183, + 0.005952802021056414, + 0.0038524095434695482, + 0.0028380420990288258, + 0.007836572825908661, + 0.0031433093827217817, + 0.009040836244821548, + 0.007125272415578365, + 0.00481368089094758, + 0.008790762163698673, + 0.012592830695211887, + 0.004432216286659241, + 0.006576992105692625, + 0.005707327742129564, + 0.0020242261234670877, + 0.003422022797167301, + 0.0024564098566770554, + 0.0043279025703668594, + 0.0042392173781991005, + 0.006055431440472603, + 0.002300858963280916, + 0.003777700010687113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.12705151736736298 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495937541127205 + ], + "opposite_logprob_tokens": [ + -2.929959297180176 + ], + "logprob_reward": -0.05495937541127205, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495937541127205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3501708507537842, + 0.04236158728599548, + 0.03805726021528244, + 0.07157514244318008, + 0.06866517663002014, + 0.06484004855155945, + 0.16686637699604034, + 0.04664066806435585, + 0.04927685111761093, + 0.10154612362384796 + ], + "credit_map": [ + 0.029368842020630836, + 0.04639511927962303, + 0.024670392274856567, + 0.08972523361444473, + 0.06945148855447769, + 0.02494790032505989, + 0.04262828826904297, + 0.022983549162745476, + 0.004329710733145475, + 0.008034455589950085, + 0.004435833543539047, + 0.006321702618151903, + 0.008589062839746475, + 0.004055941943079233, + 0.003252774477005005, + 0.0033421043772250414, + 0.002404564991593361, + 0.0047263107262551785, + 0.00274113891646266, + 0.005527029279619455, + 0.0072378553450107574, + 0.00534368259832263, + 0.005268649198114872, + 0.004808028694242239, + 0.004605899099260569, + 0.0042635854333639145, + 0.004874182399362326, + 0.02510169707238674, + 0.006456760689616203, + 0.012853036634624004, + 0.005382479168474674, + 0.008037494495511055, + 0.004356146324425936, + 0.01065156515687704, + 0.003379211062565446, + 0.016345521435141563, + 0.01933402381837368, + 0.005376529414206743, + 0.004442891571670771, + 0.004779281094670296, + 0.005847888998687267, + 0.01040046475827694, + 0.004712286870926619, + 0.009067636914551258, + 0.0073633077554404736, + 0.008110865019261837, + 0.011569183319807053, + 0.00776841351762414, + 0.012031704187393188, + 0.027727099135518074, + 0.018128663301467896, + 0.03497292846441269, + 0.024903133511543274, + 0.028399908915162086, + 0.011195564642548561, + 0.009507362730801105, + 0.004627471324056387, + 0.006552441511303186, + 0.005608073435723782, + 0.006286195013672113, + 0.004526857286691666, + 0.006810015998780727, + 0.008253275416791439, + 0.0039763362146914005, + 0.005360088776797056, + 0.0033966246992349625, + 0.0031674362253397703, + 0.006246140692383051, + 0.016845375299453735, + 0.00696090143173933, + 0.003663744544610381, + 0.003636539215222001, + 0.021612178534269333, + 0.015463474206626415, + 0.014637168496847153, + 0.015420826151967049, + 0.014281363226473331, + 0.007561724167317152, + 0.006157646421343088, + 0.0064117335714399815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.05495937541127205 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903782814741135 + ], + "opposite_logprob_tokens": [ + -2.4640378952026367 + ], + "logprob_reward": -0.08903782814741135, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903782814741135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5155183672904968, + 0.06209847703576088, + 0.10038971155881882, + 0.025149047374725342, + 0.030139807611703873, + 0.09175945073366165, + 0.040039706975221634, + 0.05133097618818283, + 0.049125831574201584, + 0.03444865718483925 + ], + "credit_map": [ + 0.03623703494668007, + 0.06806537508964539, + 0.03904249519109726, + 0.10147063434123993, + 0.10549186170101166, + 0.04133538529276848, + 0.08355656266212463, + 0.040319014340639114, + 0.005628284998238087, + 0.006607238668948412, + 0.005652711261063814, + 0.008754365146160126, + 0.01024916023015976, + 0.009449842385947704, + 0.009669672697782516, + 0.006087200250476599, + 0.005862433463335037, + 0.014800007455050945, + 0.006776553113013506, + 0.01879035495221615, + 0.015375870279967785, + 0.01373601146042347, + 0.016281872987747192, + 0.008766609244048595, + 0.0031541704665869474, + 0.0031507944222539663, + 0.002151940017938614, + 0.0040438976138830185, + 0.0036054339725524187, + 0.002069164300337434, + 0.0031849972438067198, + 0.0037886493373662233, + 0.0025134659372270107, + 0.003291290020570159, + 0.002680795965716243, + 0.004619204439222813, + 0.008665027096867561, + 0.0036834452766925097, + 0.002497374312952161, + 0.0021892013028264046, + 0.007019834127277136, + 0.02161591686308384, + 0.007807958871126175, + 0.01145138218998909, + 0.010755282826721668, + 0.012066168710589409, + 0.012742886319756508, + 0.008300027810037136, + 0.002501284470781684, + 0.006091386079788208, + 0.004576215986162424, + 0.010013366118073463, + 0.005909995641559362, + 0.006140473298728466, + 0.0026055469643324614, + 0.0022014379501342773, + 0.003984520211815834, + 0.005939398426562548, + 0.0052750906907022, + 0.007978320121765137, + 0.00639027263969183, + 0.0064652045257389545, + 0.010294188745319843, + 0.005003982689231634, + 0.0031853450927883387, + 0.003784831613302231, + 0.0026979739777743816, + 0.006149023771286011, + 0.018205473199486732, + 0.007137845735996962, + 0.005199411418288946, + 0.0027659316547214985, + 0.005662729497998953, + 0.005657972767949104, + 0.004322603344917297, + 0.006006819661706686, + 0.004715513903647661, + 0.0027309712022542953, + 0.0027871988713741302, + 0.002564848866313696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.08903782814741135 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a tense documentary style, a shaky, in-car camera follows a racer from the passenger seat as he speeds along a bumpy dirt mountain road in a rally car. The man, with sharp features and short, messy brown hair drenched in sweat, wears a dark gray professional racing suit stained with dirt and sweat. His expression is one of intense concentration, his lips pursed, his eyes fixed sharply on the winding road ahead. His left hand, clad in a worn black racing glove, controls the steering wheel with a steady, powerful grip, quickly and precisely adjusting the car's position with every bump. His right hand, muscles tensed, his knuckles white with exertion, grips the grip of a modern black tactical rifle. A thin layer of yellow-brown dust coats the rifle's barrel, scope, and stock, which bounces up and down with every violent jolt of the car. The dazzling afternoon sun streamed in at a low angle through the driver's side window, casting sharp shadows that flickered rapidly, lengthened, and deformed across the simple, roll-cage-filled cabin. These shadows swept across the driver's focused face and the cold weapon, and together with the billowing dust outside the car, created a high-risk dynamic and tense atmosphere that was about to explode.", + "seed": 17027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_12/index_6217_seed_17027_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10072240307927131, + "vlm_logprob": -0.10072240307927131, + "vlm_yes_no_margin": 2.574999952316284, + "vlm_token_credit": -0.10072240307927131, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04786314815282822 + ], + "opposite_logprob_tokens": [ + -4.047863006591797 + ], + "logprob_reward": -0.04786314815282822, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.04786314815282822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06387510895729065, + 0.13284601271152496, + 0.08190927654504776, + 0.09417244791984558, + 0.052887216210365295, + 0.09551921486854553, + 0.27274417877197266, + 0.03773299977183342, + 0.0872732400894165, + 0.0810403898358345 + ], + "credit_map": [ + 0.00567450188100338, + 0.005668595898896456, + 0.007300166878849268, + 0.007407147437334061, + 0.011967259459197521, + 0.005834221839904785, + 0.01095686573535204, + 0.009066340513527393, + 0.01883467473089695, + 0.012206683866679668, + 0.019795505329966545, + 0.014575636014342308, + 0.02372927777469158, + 0.012499118223786354, + 0.01665024645626545, + 0.014554864726960659, + 0.008595624007284641, + 0.008005521260201931, + 0.009724852629005909, + 0.009640948846936226, + 0.009099429473280907, + 0.011162369512021542, + 0.008803101256489754, + 0.016877412796020508, + 0.005693062208592892, + 0.007630384061485529, + 0.007865339517593384, + 0.009757280349731445, + 0.016171608120203018, + 0.009831561706960201, + 0.022386396303772926, + 0.014836805872619152, + 0.005998000968247652, + 0.005583886057138443, + 0.008300459943711758, + 0.011251965537667274, + 0.0070592863485217094, + 0.003920650575309992, + 0.006679926998913288, + 0.0040930346585810184, + 0.010326584801077843, + 0.00997485127300024, + 0.02299896813929081, + 0.012410588562488556, + 0.010536516085267067, + 0.011248696595430374, + 0.007757941260933876, + 0.01026505883783102, + 0.02603362314403057, + 0.011724463663995266, + 0.022734927013516426, + 0.014046979136765003, + 0.030765382573008537, + 0.04079379513859749, + 0.05266880244016647, + 0.07397618144750595, + 0.0037851056549698114, + 0.0030762439128011465, + 0.004322117660194635, + 0.004879937972873449, + 0.007679423782974482, + 0.00459983991459012, + 0.006075494457036257, + 0.003314835485070944, + 0.012663178145885468, + 0.007344490848481655, + 0.010356727056205273, + 0.00966291781514883, + 0.012897439301013947, + 0.0050565325655043125, + 0.016367048025131226, + 0.012924904935061932, + 0.007099112495779991, + 0.006911545526236296, + 0.009725292213261127, + 0.014088769443333149, + 0.012217463925480843, + 0.007196715101599693, + 0.013351033441722393, + 0.010450444184243679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.04786314815282822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430966280400753 + ], + "opposite_logprob_tokens": [ + -3.1680965423583984 + ], + "logprob_reward": -0.0430966280400753, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430966280400753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05940677970647812, + 0.11382436752319336, + 0.07359305024147034, + 0.2288707047700882, + 0.10038075596094131, + 0.11701785773038864, + 0.1299849897623062, + 0.031916625797748566, + 0.07454679906368256, + 0.07045809179544449 + ], + "credit_map": [ + 0.005295174662023783, + 0.006016010884195566, + 0.005615469999611378, + 0.009199674241244793, + 0.014567596837878227, + 0.005554473493248224, + 0.007441661320626736, + 0.005716716405004263, + 0.012033720500767231, + 0.011438812129199505, + 0.017584318295121193, + 0.01564296893775463, + 0.020270271226763725, + 0.011723505333065987, + 0.013627600856125355, + 0.01150316558778286, + 0.01036226935684681, + 0.007898183539509773, + 0.011344361118972301, + 0.011205267161130905, + 0.01135003101080656, + 0.00716130156069994, + 0.00647082319483161, + 0.007800818886607885, + 0.00969111267477274, + 0.008852692320942879, + 0.012739256955683231, + 0.012598088942468166, + 0.027954742312431335, + 0.012902933172881603, + 0.08701510727405548, + 0.05711677670478821, + 0.00699380598962307, + 0.01121080294251442, + 0.013634628616273403, + 0.02228541485965252, + 0.014461146667599678, + 0.010234699584543705, + 0.013006950728595257, + 0.008553304709494114, + 0.014700554311275482, + 0.00965807493776083, + 0.015758994966745377, + 0.011884337291121483, + 0.0198916457593441, + 0.014232194051146507, + 0.014400790445506573, + 0.016491267830133438, + 0.03231603279709816, + 0.008990460075438023, + 0.015406089834868908, + 0.008600694127380848, + 0.008410362526774406, + 0.012952126562595367, + 0.010728281922638416, + 0.032580938190221786, + 0.002759458962827921, + 0.002744445577263832, + 0.003483543172478676, + 0.004990621469914913, + 0.00573300663381815, + 0.0025478280149400234, + 0.005664522759616375, + 0.00399320013821125, + 0.007354219444096088, + 0.006126041989773512, + 0.01132990699261427, + 0.009274418465793133, + 0.00893313717097044, + 0.006351054646074772, + 0.013730132021009922, + 0.01144788321107626, + 0.009496684186160564, + 0.00811050646007061, + 0.010308340191841125, + 0.00994792953133583, + 0.01077493280172348, + 0.0050408192910254, + 0.011005968786776066, + 0.00577290914952755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.0430966280400753 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280273020267487 + ], + "opposite_logprob_tokens": [ + -2.0178027153015137 + ], + "logprob_reward": -0.14280273020267487, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280273020267487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17111670970916748, + 0.10396894067525864, + 0.10107172280550003, + 0.1463303565979004, + 0.08333919942378998, + 0.05942868813872337, + 0.15420691668987274, + 0.09041575342416763, + 0.035503990948200226, + 0.054617732763290405 + ], + "credit_map": [ + 0.028203006833791733, + 0.01551807951182127, + 0.012444528751075268, + 0.03727139160037041, + 0.023392055183649063, + 0.016373934224247932, + 0.020448096096515656, + 0.017465628683567047, + 0.009697859175503254, + 0.0059852744452655315, + 0.01720331981778145, + 0.011115803383290768, + 0.024965085089206696, + 0.006758513860404491, + 0.01785086654126644, + 0.010392224416136742, + 0.013938775286078453, + 0.009262023493647575, + 0.016787990927696228, + 0.012157334014773369, + 0.016411835327744484, + 0.012325333431363106, + 0.00909319706261158, + 0.011095241643488407, + 0.0070641497150063515, + 0.00831766240298748, + 0.010042001493275166, + 0.012028997763991356, + 0.021018443629145622, + 0.031509507447481155, + 0.04128968343138695, + 0.015059909783303738, + 0.011164180003106594, + 0.008221427910029888, + 0.01182803139090538, + 0.01597161591053009, + 0.008645513094961643, + 0.009854961186647415, + 0.009442949667572975, + 0.00821052584797144, + 0.007421639282256365, + 0.004789369646459818, + 0.007008538581430912, + 0.006189302541315556, + 0.012931951321661472, + 0.006002937443554401, + 0.008066247217357159, + 0.007018706761300564, + 0.008487771265208721, + 0.005004126578569412, + 0.012805094942450523, + 0.009698440320789814, + 0.024137916043400764, + 0.01981765404343605, + 0.06695312261581421, + 0.007302805315703154, + 0.010861405171453953, + 0.005935166962444782, + 0.008885559625923634, + 0.017400717362761497, + 0.013763855211436749, + 0.012048736214637756, + 0.013634471222758293, + 0.007885846309363842, + 0.0045396131463348866, + 0.0033294945023953915, + 0.004492452368140221, + 0.0033405504655092955, + 0.007983958348631859, + 0.0030419505201280117, + 0.0050366539508104324, + 0.003739320207387209, + 0.004471343010663986, + 0.005619029048830271, + 0.007622211705893278, + 0.014207504689693451, + 0.007519776001572609, + 0.00530210742726922, + 0.005324328783899546, + 0.004551433492451906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.14280273020267487 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705247104167938 + ], + "opposite_logprob_tokens": [ + -2.1270525455474854 + ], + "logprob_reward": -0.12705247104167938, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705247104167938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08139294385910034, + 0.07709161192178726, + 0.17505940794944763, + 0.05151580274105072, + 0.13761790096759796, + 0.09903577715158463, + 0.15468880534172058, + 0.08921244740486145, + 0.06406840682029724, + 0.07031679898500443 + ], + "credit_map": [ + 0.009463227353990078, + 0.005834746640175581, + 0.00856877863407135, + 0.010647538118064404, + 0.014806997030973434, + 0.009660790674388409, + 0.010178458876907825, + 0.012232416309416294, + 0.009044971317052841, + 0.00911740493029356, + 0.012122203595936298, + 0.009360313415527344, + 0.011643904261291027, + 0.0064077903516590595, + 0.010396952740848064, + 0.008998078294098377, + 0.02393192984163761, + 0.01930193416774273, + 0.021159019321203232, + 0.018314078450202942, + 0.044052764773368835, + 0.02075369469821453, + 0.01612539030611515, + 0.011420626193284988, + 0.0036174480337649584, + 0.006052907556295395, + 0.005700505338609219, + 0.004675277508795261, + 0.010246075689792633, + 0.005675092805176973, + 0.01091386191546917, + 0.00463463831692934, + 0.014456396922469139, + 0.013588571920990944, + 0.022373823449015617, + 0.03218898922204971, + 0.015188709832727909, + 0.012941073626279831, + 0.016346774995326996, + 0.010533586144447327, + 0.012148375622928143, + 0.011548459529876709, + 0.01671813428401947, + 0.013106980361044407, + 0.011297124437987804, + 0.011644478887319565, + 0.011550260707736015, + 0.011021972633898258, + 0.027119241654872894, + 0.005583619233220816, + 0.00928272120654583, + 0.008010747842490673, + 0.012118988670408726, + 0.022878404706716537, + 0.019430069252848625, + 0.050265029072761536, + 0.00732297170907259, + 0.00697407266125083, + 0.010051614604890347, + 0.021966280415654182, + 0.010607926174998283, + 0.00825852807611227, + 0.014660661108791828, + 0.009370402432978153, + 0.012377873994410038, + 0.00448952242732048, + 0.006322871427983046, + 0.005300674121826887, + 0.009733685292303562, + 0.005268601235002279, + 0.009292550384998322, + 0.011282633058726788, + 0.011268996633589268, + 0.008749186061322689, + 0.009796530939638615, + 0.008343455381691456, + 0.011090627871453762, + 0.007102129515260458, + 0.007228562608361244, + 0.006737314164638519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.12705247104167938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from inside a car, specifically from the passenger seat following a driver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement shaky, simulating a handheld or in-car recording?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the driver wearing a dark gray professional racing suit that is stained with dirt and sweat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver holding a modern black tactical rifle in his right hand, which is covered in a thin layer of yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature dazzling afternoon sunlight streaming through the window, casting rapidly flickering and deforming shadows inside the cabin?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279703795909882 + ], + "opposite_logprob_tokens": [ + -2.0177969932556152 + ], + "logprob_reward": -0.14279703795909882, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279703795909882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1083655059337616, + 0.05371811240911484, + 0.13757912814617157, + 0.23093608021736145, + 0.0888536348938942, + 0.12701502442359924, + 0.09425066411495209, + 0.02876007929444313, + 0.05616990849375725, + 0.07435186952352524 + ], + "credit_map": [ + 0.01129141915589571, + 0.008582387119531631, + 0.013871300965547562, + 0.017397500574588776, + 0.023691562935709953, + 0.006842765025794506, + 0.01790677383542061, + 0.008781797252595425, + 0.005616571754217148, + 0.004763822071254253, + 0.006966741289943457, + 0.007084155920892954, + 0.01268764864653349, + 0.003976176027208567, + 0.00873594544827938, + 0.0038870489224791527, + 0.024999799206852913, + 0.013820267282426357, + 0.017617681995034218, + 0.015938865020871162, + 0.03008471429347992, + 0.013380659744143486, + 0.010226343758404255, + 0.011510803364217281, + 0.010191414505243301, + 0.012205823324620724, + 0.012934145517647266, + 0.014996606856584549, + 0.03784587234258652, + 0.023642076179385185, + 0.073702372610569, + 0.04541777819395065, + 0.008645405992865562, + 0.010260969400405884, + 0.012176815420389175, + 0.01901339367032051, + 0.011155214160680771, + 0.00962916575372219, + 0.010777122341096401, + 0.00719554303213954, + 0.013779128901660442, + 0.011763295158743858, + 0.028625691309571266, + 0.017017653211951256, + 0.012857089750468731, + 0.013103920966386795, + 0.015662765130400658, + 0.014205475337803364, + 0.006227114237844944, + 0.004083533305674791, + 0.011546491645276546, + 0.005247291177511215, + 0.01153105404227972, + 0.013309256173670292, + 0.021257901564240456, + 0.02104802057147026, + 0.0038701696321368217, + 0.002145276637747884, + 0.0027042501606047153, + 0.0042710923589766026, + 0.003544378560036421, + 0.00344818364828825, + 0.004301927518099546, + 0.004474800080060959, + 0.013241431675851345, + 0.004073786083608866, + 0.00446715485304594, + 0.00481211906298995, + 0.013365916907787323, + 0.0034676864743232727, + 0.005304927937686443, + 0.007436880376189947, + 0.0072788349352777, + 0.006735020317137241, + 0.01187990140169859, + 0.015688186511397362, + 0.014005028642714024, + 0.0049174875020980835, + 0.00903528556227684, + 0.004812122322618961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.14279703795909882 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_13/index_tensor([2524, 2524])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_13/index_tensor([2524, 2524])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3569ea0c3193e2f40c87d284bba0658ace88499a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_13/index_tensor([2524, 2524])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, slow-motion shot, the camera gently panning to the right, captures two plump, furry giant pandas frolicking in a meadow surrounded by tall, emerald bamboo. Soft afternoon sunlight filters through the bamboo leaves, casting shifting dapples of light on their dense, shaggy black and white fur, making every hair distinct in the slow-motion. One of the pandas takes the lead, clumsily yet adorably flipping its chubby body onto its side, sprawling on the soft grass, its black paws waving slowly in the air. The other panda, with a naive expression and a pigeon-toed gait, lurches in pursuit, then gently bumps into its companion with its massive body, triggering a series of slow, hilarious tumbles. As they roll, green grass leaves and a few withered yellow bamboo leaves are lifted up and float gracefully in the air. The whole picture is filled with an atmosphere of innocence, tranquility and incomparable joy. The main colors are vivid green and pure black and white, embellished by warm golden sunlight, creating a dreamy natural beauty.", + "seed": 1987, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_13/index_2524_seed_1987_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.029749357700347902, + "vlm_logprob": -0.029749357700347902, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.029749357700347902, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008278578519821167 + ], + "opposite_logprob_tokens": [ + -5.3832783699035645 + ], + "logprob_reward": -0.008278578519821167, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.008278578519821167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04573149234056473, + 0.34573692083358765, + 0.024321606382727623, + 0.021014776080846786, + 0.029478561133146286, + 0.07589731365442276, + 0.21133357286453247, + 0.1316547691822052, + 0.05711164325475693, + 0.05771941319108009 + ], + "credit_map": [ + 0.005872800014913082, + 0.004892334342002869, + 0.006076296791434288, + 0.005347349215298891, + 0.006227628327906132, + 0.00886739045381546, + 0.004220227710902691, + 0.004227460827678442, + 0.04690118879079819, + 0.028252240270376205, + 0.032339245080947876, + 0.023107513785362244, + 0.01889924518764019, + 0.10206986963748932, + 0.021572794765233994, + 0.0725947842001915, + 0.0026062061078846455, + 0.0024631728883832693, + 0.003460494801402092, + 0.003114361083135009, + 0.0024926080368459225, + 0.0034392448142170906, + 0.0034793512895703316, + 0.0032661650329828262, + 0.002061333041638136, + 0.0024536666460335255, + 0.00243283249437809, + 0.0021726135164499283, + 0.004585139453411102, + 0.002685257000848651, + 0.0020959724206477404, + 0.002527957782149315, + 0.002640599152073264, + 0.002793604275211692, + 0.006201777141541243, + 0.0033473107032477856, + 0.004060911480337381, + 0.0030327916610985994, + 0.004697106312960386, + 0.00270445947535336, + 0.00843717623502016, + 0.005926948040723801, + 0.009902196936309338, + 0.0066151088103652, + 0.00787561759352684, + 0.012082379311323166, + 0.008050872012972832, + 0.0170070119202137, + 0.05040507763624191, + 0.02494024485349655, + 0.024705737829208374, + 0.0145886754617095, + 0.017359985038638115, + 0.01414570864289999, + 0.04310422018170357, + 0.022083895280957222, + 0.02022266760468483, + 0.014453492127358913, + 0.016539061442017555, + 0.016127757728099823, + 0.014279184862971306, + 0.015085107646882534, + 0.01596011035144329, + 0.018987378105521202, + 0.009846115484833717, + 0.009072513319551945, + 0.0052676014602184296, + 0.007417847402393818, + 0.004602452740073204, + 0.009244555607438087, + 0.006378121208399534, + 0.005282426718622446, + 0.008463289588689804, + 0.003501125378534198, + 0.008434955030679703, + 0.009443497285246849, + 0.005304999183863401, + 0.007761158514767885, + 0.005767738446593285, + 0.009042647667229176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.008278578519821167 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191635608673096 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.018191635608673096, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191635608673096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06452466547489166, + 0.4766536355018616, + 0.013292470015585423, + 0.028321275487542152, + 0.03643225505948067, + 0.01223716326057911, + 0.16381439566612244, + 0.07341646403074265, + 0.10304158180952072, + 0.02826608344912529 + ], + "credit_map": [ + 0.00694514624774456, + 0.006447100546211004, + 0.009591610170900822, + 0.008827499113976955, + 0.010153034701943398, + 0.012063917703926563, + 0.0057688625529408455, + 0.004727493040263653, + 0.06443404406309128, + 0.03989285230636597, + 0.042416539043188095, + 0.030740004032850266, + 0.02676628902554512, + 0.13849566876888275, + 0.028110584244132042, + 0.1057976558804512, + 0.002071358496323228, + 0.002035577315837145, + 0.0020564482547342777, + 0.0010485249804332852, + 0.0013335212133824825, + 0.002146402606740594, + 0.0011262299958616495, + 0.0014744054060429335, + 0.0032334381248801947, + 0.0019090429414063692, + 0.006100540980696678, + 0.0016971280565485358, + 0.002530774800106883, + 0.008161704987287521, + 0.0013698518741875887, + 0.003318789880722761, + 0.003042821539565921, + 0.0038460134528577328, + 0.007589603308588266, + 0.002732916036620736, + 0.006041055079549551, + 0.0032700158189982176, + 0.00712546706199646, + 0.0027843627613037825, + 0.0015826185699552298, + 0.0010155276395380497, + 0.001470169285312295, + 0.0011204190086573362, + 0.0014064981369301677, + 0.0019387745996937156, + 0.0013257167302072048, + 0.0023774392902851105, + 0.03446559980511665, + 0.014976792968809605, + 0.019197838380932808, + 0.012296496890485287, + 0.016106322407722473, + 0.01004757545888424, + 0.03726516664028168, + 0.01945859007537365, + 0.011197684332728386, + 0.014670210890471935, + 0.00808731745928526, + 0.00772158894687891, + 0.007398618850857019, + 0.012630739249289036, + 0.007291256915777922, + 0.004419044125825167, + 0.016982216387987137, + 0.018834799528121948, + 0.01692880503833294, + 0.008719617500901222, + 0.007442430593073368, + 0.01828468032181263, + 0.00906476192176342, + 0.006784270517528057, + 0.004119934514164925, + 0.002949783578515053, + 0.0035372297279536724, + 0.0044817011803388596, + 0.002575399586930871, + 0.00393883278593421, + 0.003958840388804674, + 0.002704357961192727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.018191635608673096 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253676325082779 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.005253676325082779, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253676325082779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03680754825472832, + 0.7001501321792603, + 0.016683438792824745, + 0.05328407138586044, + 0.01841992326080799, + 0.021058598533272743, + 0.05162891000509262, + 0.060690026730298996, + 0.028331879526376724, + 0.012945512309670448 + ], + "credit_map": [ + 0.0031745918095111847, + 0.0035349375102669, + 0.004223786760121584, + 0.005397748202085495, + 0.007173554040491581, + 0.005483885295689106, + 0.004565917421132326, + 0.003253129543736577, + 0.10326844453811646, + 0.06502627581357956, + 0.06869414448738098, + 0.04424595832824707, + 0.036088477820158005, + 0.1874389499425888, + 0.04380249232053757, + 0.15158535540103912, + 0.001256676740013063, + 0.001971543999388814, + 0.0016042961506173015, + 0.0028458184096962214, + 0.0020324047654867172, + 0.0022438124287873507, + 0.002915708115324378, + 0.0018131788820028305, + 0.006287815514951944, + 0.007841061800718307, + 0.011026756837964058, + 0.005259088706225157, + 0.004939195234328508, + 0.009910893626511097, + 0.00260933511890471, + 0.0054099238477647305, + 0.0015125233912840486, + 0.0019881504122167826, + 0.003744702786207199, + 0.0014725727960467339, + 0.002506342250853777, + 0.0016723312437534332, + 0.0035710125230252743, + 0.001952288206666708, + 0.0028268136084079742, + 0.002497965469956398, + 0.002804296789690852, + 0.002191780135035515, + 0.00201904377900064, + 0.002764805220067501, + 0.002031601034104824, + 0.003922292497009039, + 0.011072011664509773, + 0.0073218937031924725, + 0.0053390320390462875, + 0.0038137694355100393, + 0.004050096962600946, + 0.006977356504648924, + 0.007305409759283066, + 0.005749337840825319, + 0.009711464866995811, + 0.0063263350166380405, + 0.00691879540681839, + 0.00858471542596817, + 0.005260880570858717, + 0.0053832801058888435, + 0.007584703620523214, + 0.010919853113591671, + 0.004928364418447018, + 0.0025998339988291264, + 0.005300628021359444, + 0.002863060450181365, + 0.0025449178647249937, + 0.0032807334791868925, + 0.0034765286836773157, + 0.0033378119114786386, + 0.001968009164556861, + 0.0023752169217914343, + 0.0011234882986173034, + 0.0016324830939993262, + 0.0011481508845463395, + 0.0028198331128805876, + 0.0012007468612864614, + 0.000677583971992135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.005253676325082779 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03807377442717552 + ], + "opposite_logprob_tokens": [ + -3.288073778152466 + ], + "logprob_reward": -0.03807377442717552, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03807377442717552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019803540781140327, + 0.27985885739326477, + 0.027037933468818665, + 0.05868307501077652, + 0.03698378801345825, + 0.09674955904483795, + 0.23319868743419647, + 0.1348564326763153, + 0.0630057230591774, + 0.04982244595885277 + ], + "credit_map": [ + 0.003160770982503891, + 0.001930316793732345, + 0.002592663513496518, + 0.0023943467531353235, + 0.002431826665997505, + 0.0026567948516458273, + 0.0024487043265253305, + 0.0021881181746721268, + 0.04106113687157631, + 0.02660308964550495, + 0.029031911864876747, + 0.018750960007309914, + 0.015851221978664398, + 0.07176810503005981, + 0.019123608246445656, + 0.05766880139708519, + 0.00249664019793272, + 0.004036962520331144, + 0.0020938788075000048, + 0.004052985459566116, + 0.003595765447244048, + 0.0042206658981740475, + 0.0034109526313841343, + 0.0031300822738558054, + 0.006414684932678938, + 0.010906940326094627, + 0.009149030782282352, + 0.006669959519058466, + 0.00813896581530571, + 0.008029324002563953, + 0.003993464168161154, + 0.005380709655582905, + 0.0032699555158615112, + 0.0029290886595845222, + 0.01030189823359251, + 0.002605070825666189, + 0.004305648617446423, + 0.0026992426719516516, + 0.0082441670820117, + 0.00262871733866632, + 0.011320224963128567, + 0.00986385066062212, + 0.012778740376234055, + 0.009367912076413631, + 0.01057805772870779, + 0.011020711623132229, + 0.010289482772350311, + 0.02153058350086212, + 0.05802197754383087, + 0.02295716479420662, + 0.021419070661067963, + 0.01568157970905304, + 0.0180499367415905, + 0.01928224042057991, + 0.0507102906703949, + 0.02707643248140812, + 0.02062119171023369, + 0.013606112450361252, + 0.014163612388074398, + 0.019759349524974823, + 0.01441584900021553, + 0.014306893572211266, + 0.01569046452641487, + 0.02229296788573265, + 0.009984754957258701, + 0.007944071665406227, + 0.008447164669632912, + 0.006093727890402079, + 0.006091349758207798, + 0.009727580472826958, + 0.006941374856978655, + 0.007775698788464069, + 0.0069449287839233875, + 0.007595378439873457, + 0.005526574328541756, + 0.00889542791992426, + 0.004647240974009037, + 0.0054964604787528515, + 0.006040187086910009, + 0.0046762507408857346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.03807377442717552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07894912362098694 + ], + "opposite_logprob_tokens": [ + -2.578949213027954 + ], + "logprob_reward": -0.07894912362098694, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894912362098694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029952803626656532, + 0.4840112328529358, + 0.018277298659086227, + 0.06736616045236588, + 0.042323239147663116, + 0.1701761782169342, + 0.052228912711143494, + 0.06448224186897278, + 0.05056145414710045, + 0.0206204392015934 + ], + "credit_map": [ + 0.003589716972783208, + 0.002646457636728883, + 0.005038419738411903, + 0.0029997772071510553, + 0.0022438138257712126, + 0.0076179513707757, + 0.0023340049665421247, + 0.0034826649352908134, + 0.06848299503326416, + 0.042302343994379044, + 0.04536861181259155, + 0.031001441180706024, + 0.02667776308953762, + 0.14113642275333405, + 0.027452055364847183, + 0.10158959776163101, + 0.001723396242596209, + 0.0017344116931781173, + 0.0016794645925983787, + 0.003033542772755027, + 0.001501713995821774, + 0.0024845744483172894, + 0.0030765566043555737, + 0.0030436385422945023, + 0.006874593440443277, + 0.01224592886865139, + 0.012570735067129135, + 0.007414612453430891, + 0.008379687555134296, + 0.010764860548079014, + 0.003619225462898612, + 0.005496521946042776, + 0.004076382610946894, + 0.005752116907387972, + 0.007886572740972042, + 0.005508209113031626, + 0.004720756784081459, + 0.0045334710739552975, + 0.0068060182966291904, + 0.0030397125519812107, + 0.020313134416937828, + 0.019910819828510284, + 0.022412829101085663, + 0.016476552933454514, + 0.0175490602850914, + 0.018713658675551414, + 0.018391020596027374, + 0.03640911355614662, + 0.0078260013833642, + 0.009264353662729263, + 0.006289831828325987, + 0.004149087704718113, + 0.0050075603649020195, + 0.006479666102677584, + 0.004875210579484701, + 0.008337204344570637, + 0.009838934056460857, + 0.011116083711385727, + 0.007474555633962154, + 0.007128970231860876, + 0.007245170883834362, + 0.00930674746632576, + 0.006182852666825056, + 0.006188924890011549, + 0.007608943153172731, + 0.0053075761534273624, + 0.008170167915523052, + 0.007855561561882496, + 0.004672854673117399, + 0.004816431552171707, + 0.006159934215247631, + 0.005969987250864506, + 0.002993663540109992, + 0.002932430012151599, + 0.0031138535123318434, + 0.0031157208140939474, + 0.0020628091879189014, + 0.00228805560618639, + 0.0024394586216658354, + 0.0016744481399655342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.07894912362098694 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, slow-motion shot, the camera gently panning to the right, captures two plump, furry giant pandas frolicking in a meadow surrounded by tall, emerald bamboo. Soft afternoon sunlight filters through the bamboo leaves, casting shifting dapples of light on their dense, shaggy black and white fur, making every hair distinct in the slow-motion. One of the pandas takes the lead, clumsily yet adorably flipping its chubby body onto its side, sprawling on the soft grass, its black paws waving slowly in the air. The other panda, with a naive expression and a pigeon-toed gait, lurches in pursuit, then gently bumps into its companion with its massive body, triggering a series of slow, hilarious tumbles. As they roll, green grass leaves and a few withered yellow bamboo leaves are lifted up and float gracefully in the air. The whole picture is filled with an atmosphere of innocence, tranquility and incomparable joy. The main colors are vivid green and pure black and white, embellished by warm golden sunlight, creating a dreamy natural beauty.", + "seed": 1988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_13/index_2524_seed_1988_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.034477383363991976, + "vlm_logprob": -0.034477383363991976, + "vlm_yes_no_margin": 4.174999952316284, + "vlm_token_credit": -0.034477383363991976, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004770801402628422 + ], + "opposite_logprob_tokens": [ + -6.254770755767822 + ], + "logprob_reward": -0.004770801402628422, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.004770801402628422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08167675882577896, + 0.03237498924136162, + 0.07615524530410767, + 0.07188356667757034, + 0.1130351796746254, + 0.07092812657356262, + 0.181781604886055, + 0.055048681795597076, + 0.1856682449579239, + 0.13144762814044952 + ], + "credit_map": [ + 0.010974450036883354, + 0.007790470961481333, + 0.011895881034433842, + 0.008413895964622498, + 0.010884989984333515, + 0.012113237753510475, + 0.009642653167247772, + 0.009961171075701714, + 0.005136953201144934, + 0.0027322524692863226, + 0.004408749286085367, + 0.004248852375894785, + 0.00327278976328671, + 0.0035172845236957073, + 0.004876900929957628, + 0.004181204829365015, + 0.008142386563122272, + 0.010405602864921093, + 0.00954121071845293, + 0.010636502876877785, + 0.006435448303818703, + 0.01013811957091093, + 0.011686839163303375, + 0.009169126860797405, + 0.007920528762042522, + 0.00945215206593275, + 0.008271067403256893, + 0.010433923453092575, + 0.009504458867013454, + 0.010267273522913456, + 0.008090458810329437, + 0.00794370286166668, + 0.030891990289092064, + 0.010436663404107094, + 0.010230347514152527, + 0.007650857325643301, + 0.00919347908347845, + 0.018959345296025276, + 0.007510669995099306, + 0.018161825835704803, + 0.01189368311315775, + 0.006497831549495459, + 0.005767281167209148, + 0.006989030633121729, + 0.007310532499104738, + 0.008184529840946198, + 0.006656524259597063, + 0.017628710716962814, + 0.034750502556562424, + 0.023173537105321884, + 0.017588334158062935, + 0.014157887548208237, + 0.03047470562160015, + 0.03314702585339546, + 0.01421464141458273, + 0.01427497062832117, + 0.0074299839325249195, + 0.00383181544020772, + 0.005708892829716206, + 0.005882642697542906, + 0.004316725768148899, + 0.012307715602219105, + 0.002847825177013874, + 0.012723078951239586, + 0.04795674607157707, + 0.00704942736774683, + 0.03348406404256821, + 0.007618392817676067, + 0.004822588060051203, + 0.014129363000392914, + 0.006872702855616808, + 0.06373497098684311, + 0.017887160181999207, + 0.005834296811372042, + 0.024898476898670197, + 0.007443822920322418, + 0.011220852844417095, + 0.017198748886585236, + 0.010129585862159729, + 0.03683469444513321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.004770801402628422 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819339208304882 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.01819339208304882, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819339208304882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09559425711631775, + 0.16995100677013397, + 0.07665274292230606, + 0.0353555865585804, + 0.09706377983093262, + 0.14648476243019104, + 0.043183568865060806, + 0.04614584892988205, + 0.20070606470108032, + 0.088862344622612 + ], + "credit_map": [ + 0.01845686323940754, + 0.00953053217381239, + 0.011856064200401306, + 0.009910937398672104, + 0.016276635229587555, + 0.010381429456174374, + 0.008765342645347118, + 0.010416455566883087, + 0.012295092456042767, + 0.0075172907672822475, + 0.008548030629754066, + 0.010095244273543358, + 0.009533362463116646, + 0.009874017909169197, + 0.054944947361946106, + 0.05714302882552147, + 0.009717676788568497, + 0.00677154865115881, + 0.006400931626558304, + 0.006608079187572002, + 0.008075262419879436, + 0.01096570398658514, + 0.01422413345426321, + 0.01388941053301096, + 0.004227499943226576, + 0.004095008596777916, + 0.004738613963127136, + 0.003887148341163993, + 0.004035754594951868, + 0.0045850882306694984, + 0.00406937301158905, + 0.005717102903872728, + 0.014446849003434181, + 0.008634367026388645, + 0.006510479841381311, + 0.009588356129825115, + 0.011810174211859703, + 0.014104558154940605, + 0.007450390141457319, + 0.024518616497516632, + 0.021418748423457146, + 0.015669850632548332, + 0.01488377247005701, + 0.02232213318347931, + 0.01410904061049223, + 0.019820459187030792, + 0.017682354897260666, + 0.02057841047644615, + 0.006627819966524839, + 0.005815617274492979, + 0.005107078235596418, + 0.004938188940286636, + 0.005798015743494034, + 0.006118555087596178, + 0.004114366136491299, + 0.004663929343223572, + 0.005686874035745859, + 0.0044676270335912704, + 0.0049394238740205765, + 0.005266882944852114, + 0.00411219522356987, + 0.012448742054402828, + 0.002362414263188839, + 0.006861689500510693, + 0.05202138051390648, + 0.007585048675537109, + 0.03518247604370117, + 0.007257423363626003, + 0.004923139698803425, + 0.015390638262033463, + 0.009013889357447624, + 0.06933209300041199, + 0.01658274605870247, + 0.006450544111430645, + 0.01396888867020607, + 0.0059529924765229225, + 0.006165517959743738, + 0.014374719932675362, + 0.0067855860106647015, + 0.01858135126531124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.01819339208304882 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738202180713415 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738202180713415, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738202180713415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04343804717063904, + 0.274724543094635, + 0.078594870865345, + 0.02065364271402359, + 0.03238186240196228, + 0.04940014332532883, + 0.11468014866113663, + 0.04114249721169472, + 0.2085627317428589, + 0.13642147183418274 + ], + "credit_map": [ + 0.004590271972119808, + 0.004998070187866688, + 0.006989230867475271, + 0.005781466141343117, + 0.006010651588439941, + 0.005792838055640459, + 0.005144129972904921, + 0.004131386522203684, + 0.021524420008063316, + 0.00731603242456913, + 0.009000658988952637, + 0.013656881637871265, + 0.011416548863053322, + 0.015887048095464706, + 0.07970532774925232, + 0.11621762067079544, + 0.008806254714727402, + 0.010868057608604431, + 0.009891601279377937, + 0.010962429456412792, + 0.006931922864168882, + 0.010135817341506481, + 0.011907479725778103, + 0.009091307409107685, + 0.002373134484514594, + 0.0027600291650742292, + 0.003334511537104845, + 0.002077593933790922, + 0.002352192299440503, + 0.002856892766430974, + 0.001564602367579937, + 0.0033346868585795164, + 0.00405160803347826, + 0.00465402752161026, + 0.0035760439932346344, + 0.004239138215780258, + 0.004882927983999252, + 0.003787006950005889, + 0.0029159430414438248, + 0.004275166895240545, + 0.007110776379704475, + 0.0036068386398255825, + 0.004836260806769133, + 0.0064416625536978245, + 0.006714019924402237, + 0.007361123338341713, + 0.005967272911220789, + 0.00736219109967351, + 0.021466806530952454, + 0.011193497106432915, + 0.00923753622919321, + 0.009833496995270252, + 0.017164340242743492, + 0.02686002291738987, + 0.007941186428070068, + 0.010983262211084366, + 0.005457110237330198, + 0.0035384087823331356, + 0.003227463224902749, + 0.004666632506996393, + 0.0030784360133111477, + 0.008736967109143734, + 0.002465538214892149, + 0.009971939958631992, + 0.052603647112846375, + 0.008898889645934105, + 0.030506959185004234, + 0.008500202558934689, + 0.009818978607654572, + 0.0229784045368433, + 0.010319286957383156, + 0.06493635475635529, + 0.01867060177028179, + 0.007292490452528, + 0.027186177670955658, + 0.007965471595525742, + 0.011307941749691963, + 0.018073230981826782, + 0.008500969037413597, + 0.037424586713314056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.006738202180713415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029780501499772072 + ], + "opposite_logprob_tokens": [ + -3.529780387878418 + ], + "logprob_reward": -0.029780501499772072, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029780501499772072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04644601419568062, + 0.15749435126781464, + 0.020628293976187706, + 0.01623363047838211, + 0.062049757689237595, + 0.04418545961380005, + 0.36138126254081726, + 0.06744307279586792, + 0.023454032838344574, + 0.20068418979644775 + ], + "credit_map": [ + 0.005205913446843624, + 0.004951339215040207, + 0.007313101552426815, + 0.005376741755753756, + 0.005848012864589691, + 0.007225590292364359, + 0.005605401936918497, + 0.004919907543808222, + 0.012384884990751743, + 0.004597282037138939, + 0.005080456845462322, + 0.007749022915959358, + 0.006250525359064341, + 0.00953629519790411, + 0.04584895819425583, + 0.06604690849781036, + 0.0020152893848717213, + 0.0023566074669361115, + 0.001506226253695786, + 0.00333635276183486, + 0.00198019458912313, + 0.0027323223184794188, + 0.0039640916511416435, + 0.0027372066397219896, + 0.0017355575691908598, + 0.002630996285006404, + 0.002166741993278265, + 0.0021971489768475294, + 0.0014244195772334933, + 0.002264619804918766, + 0.001888020895421505, + 0.001926123513840139, + 0.01463595312088728, + 0.006596199236810207, + 0.004827287979424, + 0.00686221569776535, + 0.0064417640678584576, + 0.010259096510708332, + 0.0054504829458892345, + 0.006976755801588297, + 0.005138349253684282, + 0.004781530238687992, + 0.004072077572345734, + 0.004624648485332727, + 0.006978706456720829, + 0.005994573701173067, + 0.005226371809840202, + 0.007369196508079767, + 0.07097989320755005, + 0.043741606175899506, + 0.038706451654434204, + 0.031027041375637054, + 0.05927574262022972, + 0.06111571192741394, + 0.02835182286798954, + 0.02818293496966362, + 0.012783555313944817, + 0.004988181404769421, + 0.004788773600012064, + 0.004083989188075066, + 0.005560928024351597, + 0.009802908636629581, + 0.003527085529640317, + 0.021907638758420944, + 0.0038015423342585564, + 0.002157632727175951, + 0.002484234981238842, + 0.002357180928811431, + 0.0034368871711194515, + 0.0035028669517487288, + 0.002734461100772023, + 0.0029792236164212227, + 0.027579346671700478, + 0.010442952625453472, + 0.0398498997092247, + 0.011585405096411705, + 0.01578061655163765, + 0.027321740984916687, + 0.01241970993578434, + 0.0557045042514801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.029780501499772072 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290401965379715 + ], + "opposite_logprob_tokens": [ + -2.2379040718078613 + ], + "logprob_reward": -0.11290401965379715, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290401965379715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10360567271709442, + 0.13074573874473572, + 0.05403663218021393, + 0.05668874830007553, + 0.04793901741504669, + 0.04924454167485237, + 0.4625202417373657, + 0.052690476179122925, + 0.021515000611543655, + 0.021013975143432617 + ], + "credit_map": [ + 0.015245361253619194, + 0.010470397770404816, + 0.014495612122118473, + 0.01182567235082388, + 0.02026227116584778, + 0.011855375953018665, + 0.010161055251955986, + 0.009289932437241077, + 0.01579180359840393, + 0.003483143402263522, + 0.004956573247909546, + 0.008015272207558155, + 0.005864365957677364, + 0.009941463358700275, + 0.02640155330300331, + 0.05629156902432442, + 0.0081735048443079, + 0.004225133918225765, + 0.005176924169063568, + 0.005494494456797838, + 0.004646200221031904, + 0.006879613269120455, + 0.01073855347931385, + 0.008702205494046211, + 0.008446475490927696, + 0.0074023897759616375, + 0.0057252319529652596, + 0.007810680661350489, + 0.0068728202022612095, + 0.007964035496115685, + 0.005655406042933464, + 0.006811707746237516, + 0.006989831104874611, + 0.005243385210633278, + 0.0041710929945111275, + 0.009640456177294254, + 0.005557569209486246, + 0.006527107208967209, + 0.004094860516488552, + 0.005714712664484978, + 0.004680839367210865, + 0.00658262986689806, + 0.005649261642247438, + 0.005758833605796099, + 0.008870160207152367, + 0.005916471127420664, + 0.006077721249312162, + 0.005708624608814716, + 0.09043975174427032, + 0.05928982421755791, + 0.04904073476791382, + 0.033656347543001175, + 0.07553336769342422, + 0.08424412459135056, + 0.03825553506612778, + 0.03206057474017143, + 0.007444419898092747, + 0.0023211645893752575, + 0.0077377380803227425, + 0.005502152256667614, + 0.0037270833272486925, + 0.009275181218981743, + 0.0024310031440109015, + 0.01425173506140709, + 0.0025479570031166077, + 0.0026231103111058474, + 0.004163775593042374, + 0.0022103346418589354, + 0.00238432758487761, + 0.002448824467137456, + 0.002433197572827339, + 0.0027034739032387733, + 0.003997660707682371, + 0.0011448180302977562, + 0.0027213399298489094, + 0.0015729247825220227, + 0.0018693618476390839, + 0.004237531218677759, + 0.002823663642629981, + 0.002646676730364561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.11290401965379715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, slow-motion shot, the camera gently panning to the right, captures two plump, furry giant pandas frolicking in a meadow surrounded by tall, emerald bamboo. Soft afternoon sunlight filters through the bamboo leaves, casting shifting dapples of light on their dense, shaggy black and white fur, making every hair distinct in the slow-motion. One of the pandas takes the lead, clumsily yet adorably flipping its chubby body onto its side, sprawling on the soft grass, its black paws waving slowly in the air. The other panda, with a naive expression and a pigeon-toed gait, lurches in pursuit, then gently bumps into its companion with its massive body, triggering a series of slow, hilarious tumbles. As they roll, green grass leaves and a few withered yellow bamboo leaves are lifted up and float gracefully in the air. The whole picture is filled with an atmosphere of innocence, tranquility and incomparable joy. The main colors are vivid green and pure black and white, embellished by warm golden sunlight, creating a dreamy natural beauty.", + "seed": 1989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_13/index_2524_seed_1989_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028671437734737994, + "vlm_logprob": -0.028671437734737994, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.028671437734737994, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02017851546406746 + ], + "opposite_logprob_tokens": [ + -4.145178318023682 + ], + "logprob_reward": -0.02017851546406746, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02017851546406746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07751452922821045, + 0.0487203449010849, + 0.046393271535634995, + 0.47329890727996826, + 0.08205554634332657, + 0.06160422042012215, + 0.027548732236027718, + 0.06654343008995056, + 0.05401591956615448, + 0.06230507791042328 + ], + "credit_map": [ + 0.011041992343962193, + 0.009741758927702904, + 0.020528340712189674, + 0.006960212718695402, + 0.007188434712588787, + 0.008480540476739407, + 0.007394195068627596, + 0.006179056596010923, + 0.01042040903121233, + 0.004230735823512077, + 0.010930737480521202, + 0.005643009673804045, + 0.003385975956916809, + 0.0039979093708097935, + 0.004651238676160574, + 0.005460327956825495, + 0.006905974354594946, + 0.005289776250720024, + 0.009350037202239037, + 0.00672629289329052, + 0.0036204501520842314, + 0.0041708266362547874, + 0.006672066170722246, + 0.003657849505543709, + 0.08032748103141785, + 0.012720855884253979, + 0.1840147078037262, + 0.027777785435318947, + 0.016276882961392403, + 0.026428330689668655, + 0.1059826985001564, + 0.019770145416259766, + 0.01773308776319027, + 0.00908129196614027, + 0.008770993910729885, + 0.01733444072306156, + 0.004909384995698929, + 0.005767564754933119, + 0.006001404952257872, + 0.012457380071282387, + 0.012637483887374401, + 0.009945981204509735, + 0.0065747955814003944, + 0.005444008857011795, + 0.008639990352094173, + 0.0076342071406543255, + 0.005544778425246477, + 0.005182977765798569, + 0.004379707854241133, + 0.004888236057013273, + 0.002367991954088211, + 0.003855628427118063, + 0.003853054717183113, + 0.0023395176976919174, + 0.002663417486473918, + 0.0032011771108955145, + 0.011287235654890537, + 0.005564432125538588, + 0.00777315953746438, + 0.005354957655072212, + 0.009114685468375683, + 0.015305422246456146, + 0.006882066838443279, + 0.005261472426354885, + 0.006666562985628843, + 0.010650496929883957, + 0.005440532695502043, + 0.0074935476295650005, + 0.00692789675667882, + 0.008505968376994133, + 0.004243834409862757, + 0.004087081644684076, + 0.006920200772583485, + 0.00973991397768259, + 0.0120231369510293, + 0.00819871760904789, + 0.006112826056778431, + 0.008847083896398544, + 0.005841500125825405, + 0.004621701315045357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.02017851546406746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586160942912102 + ], + "opposite_logprob_tokens": [ + -3.8955862522125244 + ], + "logprob_reward": -0.020586160942912102, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586160942912102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03416655957698822, + 0.036318037658929825, + 0.03260497376322746, + 0.5011049509048462, + 0.04641659930348396, + 0.04349522292613983, + 0.028720488771796227, + 0.04061414673924446, + 0.04487959295511246, + 0.19167949259281158 + ], + "credit_map": [ + 0.005295087117701769, + 0.0033248006366193295, + 0.008372344076633453, + 0.0030165112111717463, + 0.0029808732215315104, + 0.00504566915333271, + 0.0031832794193178415, + 0.0029479938093572855, + 0.008755815215408802, + 0.0038473375607281923, + 0.007203033659607172, + 0.0037463423795998096, + 0.002778009045869112, + 0.002920042024925351, + 0.0035535069182515144, + 0.00351394759491086, + 0.0058655268512666225, + 0.00559177016839385, + 0.006129625719040632, + 0.003911792300641537, + 0.0026489596348255873, + 0.002206098288297653, + 0.0030388147570192814, + 0.0032123872078955173, + 0.09051880240440369, + 0.014024254865944386, + 0.20648379623889923, + 0.026361310854554176, + 0.014993952587246895, + 0.02703092433512211, + 0.10011592507362366, + 0.021575946360826492, + 0.005883739795535803, + 0.005458940751850605, + 0.010462921112775803, + 0.011350046843290329, + 0.0029052570462226868, + 0.0036316027399152517, + 0.0029939827509224415, + 0.0037301075644791126, + 0.008460485376417637, + 0.007532058749347925, + 0.011012712493538857, + 0.004210277460515499, + 0.0028826557099819183, + 0.002971583977341652, + 0.003094715066254139, + 0.0033307294361293316, + 0.004440016578882933, + 0.004848653916269541, + 0.00653111794963479, + 0.0031468658708035946, + 0.003267500316724181, + 0.0024415196385234594, + 0.001704100868664682, + 0.002340710489079356, + 0.007465108763426542, + 0.0029960493557155132, + 0.005803012289106846, + 0.005325333680957556, + 0.0047690048813819885, + 0.005854465998709202, + 0.004618560429662466, + 0.0037826066836714745, + 0.00692576402798295, + 0.005409194156527519, + 0.007869011722505093, + 0.005365451332181692, + 0.0059185940772295, + 0.005989700555801392, + 0.0038636401295661926, + 0.0035382327623665333, + 0.03971599414944649, + 0.01525584701448679, + 0.0197494775056839, + 0.011223714798688889, + 0.040355004370212555, + 0.042546119540929794, + 0.011822265572845936, + 0.01101104449480772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.020586160942912102 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028204445261508226 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.0028204445261508226, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028204445261508226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08128400892019272, + 0.07531663775444031, + 0.11524944752454758, + 0.26928412914276123, + 0.19656424224376678, + 0.05799589678645134, + 0.04057350382208824, + 0.0315643735229969, + 0.05882378667593002, + 0.07334395498037338 + ], + "credit_map": [ + 0.012150518596172333, + 0.009547227993607521, + 0.015149171464145184, + 0.008505641482770443, + 0.007843945175409317, + 0.01067086961120367, + 0.008751906454563141, + 0.008664720691740513, + 0.01268887147307396, + 0.008160320110619068, + 0.011902174912393093, + 0.010768953710794449, + 0.005966123193502426, + 0.0061828396283090115, + 0.008978059515357018, + 0.010669293813407421, + 0.015795154497027397, + 0.011828779242932796, + 0.011083348654210567, + 0.017909225076436996, + 0.01774635538458824, + 0.013815435580909252, + 0.018702026456594467, + 0.008369127288460732, + 0.06963219493627548, + 0.013769092969596386, + 0.09103725105524063, + 0.032484911382198334, + 0.00777282053604722, + 0.01063054520636797, + 0.02725869230926037, + 0.016698626801371574, + 0.038090821355581284, + 0.01764168031513691, + 0.037002433091402054, + 0.06406748294830322, + 0.011756877414882183, + 0.008299170061945915, + 0.009452175348997116, + 0.010253606364130974, + 0.008253186009824276, + 0.008489958941936493, + 0.006989962421357632, + 0.007408573292195797, + 0.006145689636468887, + 0.006258831359446049, + 0.006656738463789225, + 0.00779295340180397, + 0.006386115215718746, + 0.004361619707196951, + 0.009132775478065014, + 0.004162502009421587, + 0.004167293198406696, + 0.006276493426412344, + 0.0025821838062256575, + 0.003504518186673522, + 0.006365880370140076, + 0.002182325581088662, + 0.0040580639615654945, + 0.00439482880756259, + 0.003501826198771596, + 0.004773195832967758, + 0.004084092564880848, + 0.0022041562478989363, + 0.008018452674150467, + 0.006487180013209581, + 0.011298470199108124, + 0.009724252857267857, + 0.006270043086260557, + 0.005321599077433348, + 0.006992002949118614, + 0.004711786285042763, + 0.008782666176557541, + 0.008567585609853268, + 0.009266849607229233, + 0.005138850305229425, + 0.015284840017557144, + 0.014988778159022331, + 0.00533759081736207, + 0.005976802669465542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.0028204445261508226 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029780039563775063 + ], + "opposite_logprob_tokens": [ + -3.529780149459839 + ], + "logprob_reward": -0.029780039563775063, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029780039563775063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04371635243296623, + 0.05276332423090935, + 0.07085848599672318, + 0.3260662853717804, + 0.11429022997617722, + 0.08531691879034042, + 0.048492275178432465, + 0.08834073692560196, + 0.07630008459091187, + 0.09385533630847931 + ], + "credit_map": [ + 0.0064568850211799145, + 0.004721098113805056, + 0.010279674082994461, + 0.0041272202506661415, + 0.003982875496149063, + 0.006158973556011915, + 0.004457328002899885, + 0.003532293252646923, + 0.0068900869227945805, + 0.005055341403931379, + 0.00988728553056717, + 0.006594110280275345, + 0.0048660882748663425, + 0.007976488210260868, + 0.0057072993367910385, + 0.005786621477454901, + 0.010194025933742523, + 0.007924632169306278, + 0.009234512224793434, + 0.00936066173017025, + 0.010482368059456348, + 0.008631110191345215, + 0.009946202859282494, + 0.00508496630936861, + 0.06854502111673355, + 0.010883847251534462, + 0.14353063702583313, + 0.03484414517879486, + 0.006613376550376415, + 0.013154324144124985, + 0.034343164414167404, + 0.014151736162602901, + 0.019839204847812653, + 0.009402109310030937, + 0.021191060543060303, + 0.03883204981684685, + 0.007354223169386387, + 0.005303639452904463, + 0.006250157952308655, + 0.006117779761552811, + 0.01564893312752247, + 0.011202050372958183, + 0.008048835210502148, + 0.009359840303659439, + 0.011610702611505985, + 0.011736095882952213, + 0.010068444535136223, + 0.007642010692507029, + 0.0073855482041835785, + 0.0057113864459097385, + 0.008394106291234493, + 0.005536860320717096, + 0.005872009322047234, + 0.007904003374278545, + 0.002924109809100628, + 0.004764247685670853, + 0.016469022259116173, + 0.006922159343957901, + 0.01084350049495697, + 0.00771148968487978, + 0.010914932936429977, + 0.021495448425412178, + 0.007343083620071411, + 0.006641084793955088, + 0.009864536114037037, + 0.011143491603434086, + 0.007440777495503426, + 0.01102219708263874, + 0.010285818949341774, + 0.013285219669342041, + 0.007313095964491367, + 0.005944947712123394, + 0.01484580710530281, + 0.01819707825779915, + 0.010566744953393936, + 0.010759811848402023, + 0.008266636170446873, + 0.016951052471995354, + 0.006251370068639517, + 0.008016832172870636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.029780039563775063 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999202817678452 + ], + "opposite_logprob_tokens": [ + -2.6949920654296875 + ], + "logprob_reward": -0.06999202817678452, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999202817678452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026561614125967026, + 0.04588588327169418, + 0.04536890238523483, + 0.6014937162399292, + 0.11826231330633163, + 0.02149166725575924, + 0.028034765273332596, + 0.04617627337574959, + 0.03017929010093212, + 0.036545634269714355 + ], + "credit_map": [ + 0.003990317229181528, + 0.002048331778496504, + 0.005603826139122248, + 0.0023662052117288113, + 0.0022846166975796223, + 0.004116810392588377, + 0.003023020923137665, + 0.0031284813303500414, + 0.00904112122952938, + 0.0047757262364029884, + 0.007496912032365799, + 0.005007305648177862, + 0.00577447609975934, + 0.005224442575126886, + 0.004291446413844824, + 0.004274446051567793, + 0.006830166559666395, + 0.005012288223952055, + 0.006018142681568861, + 0.007424296345561743, + 0.0057514747604727745, + 0.00382098532281816, + 0.005695270840078592, + 0.004816272761672735, + 0.12839621305465698, + 0.01966392621397972, + 0.2634357810020447, + 0.042529016733169556, + 0.014125782065093517, + 0.023347200825810432, + 0.08424108475446701, + 0.02575465478003025, + 0.023062465712428093, + 0.010554816573858261, + 0.020335454493761063, + 0.038820285350084305, + 0.007134801242500544, + 0.005196812562644482, + 0.005954372230917215, + 0.007203287910670042, + 0.0047920639626681805, + 0.0037957755848765373, + 0.0029169872868806124, + 0.0023596107494086027, + 0.0020463294349610806, + 0.002044669119641185, + 0.0016434979625046253, + 0.0018927297787740827, + 0.004611277021467686, + 0.0019046873785555363, + 0.0062989741563797, + 0.002701888093724847, + 0.0018492708913981915, + 0.005481235217303038, + 0.0022098126355558634, + 0.002977615687996149, + 0.009032050147652626, + 0.003999566659331322, + 0.005490668583661318, + 0.004414298105984926, + 0.005355929024517536, + 0.009592695161700249, + 0.004674998112022877, + 0.003616063389927149, + 0.004538232460618019, + 0.003023455385118723, + 0.005428998731076717, + 0.0036764144897460938, + 0.0034899218007922173, + 0.0041276938281953335, + 0.003162179607897997, + 0.002732390072196722, + 0.00685630040243268, + 0.0045382184907794, + 0.005202487111091614, + 0.0033763900864869356, + 0.004911099094897509, + 0.005622700788080692, + 0.0031782803125679493, + 0.0028601496014744043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.06999202817678452 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, slow-motion shot, the camera gently panning to the right, captures two plump, furry giant pandas frolicking in a meadow surrounded by tall, emerald bamboo. Soft afternoon sunlight filters through the bamboo leaves, casting shifting dapples of light on their dense, shaggy black and white fur, making every hair distinct in the slow-motion. One of the pandas takes the lead, clumsily yet adorably flipping its chubby body onto its side, sprawling on the soft grass, its black paws waving slowly in the air. The other panda, with a naive expression and a pigeon-toed gait, lurches in pursuit, then gently bumps into its companion with its massive body, triggering a series of slow, hilarious tumbles. As they roll, green grass leaves and a few withered yellow bamboo leaves are lifted up and float gracefully in the air. The whole picture is filled with an atmosphere of innocence, tranquility and incomparable joy. The main colors are vivid green and pure black and white, embellished by warm golden sunlight, creating a dreamy natural beauty.", + "seed": 1990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_13/index_2524_seed_1990_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.017053306754678487, + "vlm_logprob": -0.017053306754678487, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.017053306754678487, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007210186216980219 + ], + "opposite_logprob_tokens": [ + -5.2572102546691895 + ], + "logprob_reward": -0.007210186216980219, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.007210186216980219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18163026869297028, + 0.05460941046476364, + 0.09798604249954224, + 0.06865769624710083, + 0.05645341798663139, + 0.04303407669067383, + 0.060033656656742096, + 0.2340351939201355, + 0.07711091637611389, + 0.12644928693771362 + ], + "credit_map": [ + 0.026789803057909012, + 0.014847341924905777, + 0.03796640411019325, + 0.014776632189750671, + 0.02682660147547722, + 0.02393377386033535, + 0.01773790456354618, + 0.018751809373497963, + 0.007302550598978996, + 0.005986461415886879, + 0.010450048372149467, + 0.008154958486557007, + 0.004074231255799532, + 0.005013701040297747, + 0.00911408755928278, + 0.00451337406411767, + 0.0265069417655468, + 0.007540023885667324, + 0.01137726940214634, + 0.00820168200880289, + 0.008585535921156406, + 0.02073267474770546, + 0.007973474450409412, + 0.007068444974720478, + 0.014220585115253925, + 0.010609536431729794, + 0.0066919149830937386, + 0.00639777397736907, + 0.010087953880429268, + 0.006764980964362621, + 0.009287589229643345, + 0.004597363993525505, + 0.0102679543197155, + 0.0065427604131400585, + 0.008929911069571972, + 0.005770218092948198, + 0.006828281097114086, + 0.00505885249003768, + 0.006646185182034969, + 0.0064092595130205154, + 0.006713206879794598, + 0.003605215810239315, + 0.005588482599705458, + 0.0046086981892585754, + 0.004653940908610821, + 0.00659061037003994, + 0.005511702504009008, + 0.0057622212916612625, + 0.009349075146019459, + 0.006601990200579166, + 0.010592253878712654, + 0.005865832325071096, + 0.004898682702332735, + 0.006370586343109608, + 0.009461820125579834, + 0.006893419660627842, + 0.036755144596099854, + 0.02381223626434803, + 0.027786940336227417, + 0.032690923660993576, + 0.03983404114842415, + 0.021806951612234116, + 0.025773879140615463, + 0.02557509019970894, + 0.011118357069790363, + 0.008558914065361023, + 0.007762677501887083, + 0.010554292239248753, + 0.00794454850256443, + 0.007646626327186823, + 0.005997093860059977, + 0.017528414726257324, + 0.021657487377524376, + 0.016061531379818916, + 0.02683113142848015, + 0.012587590143084526, + 0.012032785452902317, + 0.01044516358524561, + 0.015278797596693039, + 0.011554816737771034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.007210186216980219 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541593052446842 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541593052446842, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541593052446842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16989077627658844, + 0.12944386899471283, + 0.08360041677951813, + 0.024542732164263725, + 0.06292298436164856, + 0.042507290840148926, + 0.17442844808101654, + 0.06370940059423447, + 0.17502149939537048, + 0.07393267005681992 + ], + "credit_map": [ + 0.04670979455113411, + 0.013046132400631905, + 0.041418444365262985, + 0.008106851018965244, + 0.01105834823101759, + 0.015037788078188896, + 0.015506124123930931, + 0.019007286056876183, + 0.027707459405064583, + 0.00781975407153368, + 0.03190877288579941, + 0.009700821712613106, + 0.010210230946540833, + 0.022403661161661148, + 0.008042370900511742, + 0.011650791391730309, + 0.01379802729934454, + 0.007461171131581068, + 0.02380070649087429, + 0.008452055044472218, + 0.005740410648286343, + 0.008601260371506214, + 0.007885288447141647, + 0.00786149688065052, + 0.006045115180313587, + 0.003753118682652712, + 0.00298894289880991, + 0.002245367504656315, + 0.001839166390709579, + 0.0031069829128682613, + 0.0030660731717944145, + 0.0014979649567976594, + 0.011436479166150093, + 0.00901433639228344, + 0.010520130395889282, + 0.00500428257510066, + 0.006937150843441486, + 0.007591628469526768, + 0.005932821426540613, + 0.00648615462705493, + 0.006229627411812544, + 0.00558487931266427, + 0.010277140885591507, + 0.004226505756378174, + 0.004381380509585142, + 0.0036301021464169025, + 0.005281100049614906, + 0.002896558493375778, + 0.05111809819936752, + 0.011953349225223064, + 0.014770965091884136, + 0.04608995467424393, + 0.01078401692211628, + 0.011572877876460552, + 0.013325515203177929, + 0.014813669957220554, + 0.012004119344055653, + 0.007140357978641987, + 0.010324710980057716, + 0.007470417767763138, + 0.005094113759696484, + 0.005764452740550041, + 0.00881938450038433, + 0.007091844920068979, + 0.024323318153619766, + 0.021169032901525497, + 0.029918186366558075, + 0.02200806885957718, + 0.02045535482466221, + 0.011510997079312801, + 0.02481747977435589, + 0.02081906609237194, + 0.01667989231646061, + 0.011343811638653278, + 0.017720641568303108, + 0.005087481811642647, + 0.009044301696121693, + 0.005118603352457285, + 0.0046652997843921185, + 0.0042726341634988785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.012541593052446842 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947155877947807 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947155877947807, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947155877947807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22005073726177216, + 0.08297812193632126, + 0.2038726955652237, + 0.09802697598934174, + 0.04839861020445824, + 0.01663289964199066, + 0.1473066508769989, + 0.07361534237861633, + 0.03291703388094902, + 0.07620096206665039 + ], + "credit_map": [ + 0.045960646122694016, + 0.023115303367376328, + 0.03453496843576431, + 0.02028723992407322, + 0.019430430606007576, + 0.03282158449292183, + 0.02471006102859974, + 0.019190482795238495, + 0.012389307841658592, + 0.007385318633168936, + 0.018760312348604202, + 0.012372594326734543, + 0.006146480794996023, + 0.007502793334424496, + 0.011834748089313507, + 0.006586551200598478, + 0.07522430270910263, + 0.008745593950152397, + 0.0156764667481184, + 0.01028350554406643, + 0.017292959615588188, + 0.05197373777627945, + 0.010879019275307655, + 0.013797101564705372, + 0.02049499936401844, + 0.01785668544471264, + 0.008148042485117912, + 0.01094605028629303, + 0.01203720923513174, + 0.009411091916263103, + 0.012477204203605652, + 0.006655688863247633, + 0.00744242500513792, + 0.007543590851128101, + 0.00832974724471569, + 0.003059870796278119, + 0.0063261594623327255, + 0.004959153942763805, + 0.006369679234921932, + 0.004367979709059, + 0.0024333628825843334, + 0.0018501775339245796, + 0.0034017092548310757, + 0.0012887223856523633, + 0.001721473061479628, + 0.0023093719501048326, + 0.001499929465353489, + 0.002128152409568429, + 0.03416042402386665, + 0.01425089966505766, + 0.014745870605111122, + 0.02733176201581955, + 0.012380789965391159, + 0.01292112935334444, + 0.01808607392013073, + 0.013429705984890461, + 0.014484304003417492, + 0.006434882525354624, + 0.011100176721811295, + 0.007575703784823418, + 0.010831478051841259, + 0.006304352544248104, + 0.010215932503342628, + 0.006668503861874342, + 0.0040669688023626804, + 0.005967461038380861, + 0.0038837434258311987, + 0.004328069742769003, + 0.0031218661461025476, + 0.003558568423613906, + 0.003849830711260438, + 0.004140524659305811, + 0.009876210242509842, + 0.01635744795203209, + 0.007359873969107866, + 0.007932400330901146, + 0.014089247211813927, + 0.0056932829320430756, + 0.0071548703126609325, + 0.00773762259632349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.005947155877947807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02977946028113365 + ], + "opposite_logprob_tokens": [ + -3.5297794342041016 + ], + "logprob_reward": -0.02977946028113365, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02977946028113365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12535184621810913, + 0.06292793154716492, + 0.17523276805877686, + 0.05976646766066551, + 0.10695845633745193, + 0.056065771728754044, + 0.053583621978759766, + 0.2247084677219391, + 0.03698471933603287, + 0.09841995686292648 + ], + "credit_map": [ + 0.020461328327655792, + 0.014495112001895905, + 0.024724112823605537, + 0.013001618906855583, + 0.015896011143922806, + 0.010205866768956184, + 0.015054598450660706, + 0.011513189412653446, + 0.010690980590879917, + 0.005907382350414991, + 0.010952431708574295, + 0.00981555599719286, + 0.004196971654891968, + 0.005386624950915575, + 0.01141588669270277, + 0.004562090616673231, + 0.06054452806711197, + 0.008922208100557327, + 0.01395834144204855, + 0.009446005336940289, + 0.015453563071787357, + 0.044179245829582214, + 0.010973921976983547, + 0.011754954233765602, + 0.012538308277726173, + 0.007106222212314606, + 0.007938757538795471, + 0.0058170524425804615, + 0.007323769386857748, + 0.00712733156979084, + 0.008378557860851288, + 0.003536471864208579, + 0.018134869635105133, + 0.015639014542102814, + 0.017661508172750473, + 0.007363352924585342, + 0.012992003932595253, + 0.010002021677792072, + 0.014681427739560604, + 0.010484255850315094, + 0.008630442433059216, + 0.004983396735042334, + 0.007659444119781256, + 0.003858869895339012, + 0.005917365197092295, + 0.009996567852795124, + 0.006495791487395763, + 0.008523896336555481, + 0.008019324392080307, + 0.007029036525636911, + 0.00869595818221569, + 0.005596932023763657, + 0.006453750189393759, + 0.006567798554897308, + 0.005828630179166794, + 0.005392192397266626, + 0.0408245213329792, + 0.01789304055273533, + 0.029804807156324387, + 0.032499223947525024, + 0.02659178152680397, + 0.021751847118139267, + 0.028004860505461693, + 0.027338366955518723, + 0.006089440546929836, + 0.004481595475226641, + 0.00538221700116992, + 0.005285935942083597, + 0.004415174014866352, + 0.003763406304642558, + 0.004328135866671801, + 0.0032388153485953808, + 0.013251551426947117, + 0.019740471616387367, + 0.015503876842558384, + 0.007954646833240986, + 0.01800086349248886, + 0.0066543701104819775, + 0.008128774352371693, + 0.00918540544807911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.02977946028113365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two plump, furry giant pandas visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a meadow surrounded by tall, emerald bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does one panda clumsily flip onto its side and sprawl on the grass while waving its paws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the second panda lurch in pursuit and bump into the first panda, causing them to tumble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, slow-motion view with the camera gently panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978813834488392 + ], + "opposite_logprob_tokens": [ + -3.5297882556915283 + ], + "logprob_reward": -0.02978813834488392, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978813834488392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2983504831790924, + 0.06533481180667877, + 0.06555794179439545, + 0.07625630497932434, + 0.06836045533418655, + 0.07320745289325714, + 0.18407011032104492, + 0.06777442246675491, + 0.04618781805038452, + 0.05490021035075188 + ], + "credit_map": [ + 0.10021365433931351, + 0.021753555163741112, + 0.050860270857810974, + 0.013544525019824505, + 0.020750435069203377, + 0.05016602575778961, + 0.019448649138212204, + 0.02161337621510029, + 0.010948583483695984, + 0.0068317195400595665, + 0.009690037928521633, + 0.008590016514062881, + 0.005783910863101482, + 0.007293945178389549, + 0.010243027471005917, + 0.005953576881438494, + 0.010093425400555134, + 0.005556230433285236, + 0.014459139667451382, + 0.00715088052675128, + 0.006503396201878786, + 0.009274708107113838, + 0.006698996759951115, + 0.0058211698196828365, + 0.014073722064495087, + 0.010490759275853634, + 0.010741066187620163, + 0.010144446045160294, + 0.0074527328833937645, + 0.008911719545722008, + 0.008047589100897312, + 0.006394266150891781, + 0.00929737277328968, + 0.007870477624237537, + 0.010357029736042023, + 0.007589439861476421, + 0.009134526364505291, + 0.0068813771940767765, + 0.008572685532271862, + 0.008657554164528847, + 0.012166330590844154, + 0.008116716518998146, + 0.009911572560667992, + 0.005888522602617741, + 0.008257091045379639, + 0.010107992216944695, + 0.009280717931687832, + 0.009478511288762093, + 0.05310202017426491, + 0.013264639303088188, + 0.019257089123129845, + 0.04696153476834297, + 0.011052497662603855, + 0.013044293038547039, + 0.011668937280774117, + 0.015719106420874596, + 0.011475020088255405, + 0.007961858995258808, + 0.009072717279195786, + 0.008991128765046597, + 0.0070489211939275265, + 0.006083734333515167, + 0.0085824616253376, + 0.00855858251452446, + 0.007172517478466034, + 0.005114321131259203, + 0.005654439330101013, + 0.00676066754385829, + 0.005047895014286041, + 0.005785685032606125, + 0.004772566258907318, + 0.005879725329577923, + 0.008006301708519459, + 0.007017835974693298, + 0.008157707750797272, + 0.007573376409709454, + 0.005877548363059759, + 0.006038397550582886, + 0.006180903874337673, + 0.0060481359250843525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.02978813834488392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_14/index_tensor([6590, 6590])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_14/index_tensor([6590, 6590])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..783e211d9bd82b1d6505f7f4257b99adf12672b6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_14/index_tensor([6590, 6590])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up shot from a slightly low angle with a static camera focuses on a set of three ornate brass keys suspended from a thick brass ring. Each key, varying in size and complexity, features an intricate Victorian design with delicate filigree and uniquely shaped drill bits, its polished surface showing a subtle patina of age in the crevices. The entire set rotates in a slow, hypnotic, continuous rotation, rhythmically capturing the keys in a single, focused beam of warm golden light from an unseen source. As they turn, sharp, brilliant light glints across their detailed surfaces, momentarily illuminating the exquisite craftsmanship before gracefully fading into deep shadow. The background is a severely out-of-focus, dark, and moody space, suggestive of the interior of an old library, where the blurred forms of leather-bound books on wooden shelves are barely discernible. The shimmering highlights on the keys contrast starkly with the surrounding darkness, and combined with the deliberate, relentless movement, create an atmosphere of mystery and elegance, imbuing these simple objects with a profound sense of importance and an ineffable history.", + "seed": 11792, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_14/index_6590_seed_11792_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.013155198050662875, + "vlm_logprob": -0.013155198050662875, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.013155198050662875, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007843765430152416 + ], + "opposite_logprob_tokens": [ + -5.382843971252441 + ], + "logprob_reward": -0.007843765430152416, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.007843765430152416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17132095992565155, + 0.23712000250816345, + 0.03798398748040199, + 0.06809880584478378, + 0.1225191131234169, + 0.051342520862817764, + 0.09114905446767807, + 0.05587141215801239, + 0.11866162717342377, + 0.04593249410390854 + ], + "credit_map": [ + 0.016770517453551292, + 0.02231965772807598, + 0.022235820069909096, + 0.019051313400268555, + 0.027737552300095558, + 0.01916017197072506, + 0.02282622642815113, + 0.021219702437520027, + 0.02963177300989628, + 0.023686332628130913, + 0.019977467134594917, + 0.017966920509934425, + 0.03422979637980461, + 0.04204356297850609, + 0.04474734142422676, + 0.024836799129843712, + 0.0031588950660079718, + 0.004047478083521128, + 0.003700493136420846, + 0.005055986810475588, + 0.004539396148175001, + 0.003750153351575136, + 0.007294385693967342, + 0.006437200121581554, + 0.007885224185883999, + 0.006898947991430759, + 0.005728772841393948, + 0.005707878153771162, + 0.009458569809794426, + 0.012416684068739414, + 0.011764836497604847, + 0.008237894624471664, + 0.012736487202346325, + 0.012921235524117947, + 0.0119658587500453, + 0.012500900775194168, + 0.02743632160127163, + 0.017632082104682922, + 0.014942439272999763, + 0.012383798137307167, + 0.004491457249969244, + 0.005593629088252783, + 0.0056311506778001785, + 0.0077606020495295525, + 0.007358904927968979, + 0.006111904513090849, + 0.008803210221230984, + 0.005591661669313908, + 0.008082665503025055, + 0.013758158311247826, + 0.007691222708672285, + 0.019010024145245552, + 0.013291561976075172, + 0.009256158024072647, + 0.007974168285727501, + 0.012085099704563618, + 0.005426164250820875, + 0.006268395576626062, + 0.0062276930548250675, + 0.006370211485773325, + 0.008456660434603691, + 0.006182588171213865, + 0.006478788331151009, + 0.010460909456014633, + 0.012417029589414597, + 0.012928081676363945, + 0.0154231833294034, + 0.014371886849403381, + 0.014064610935747623, + 0.015442236326634884, + 0.014001214876770973, + 0.020013384521007538, + 0.005326586309820414, + 0.003196388017386198, + 0.003990387078374624, + 0.0037698810920119286, + 0.004218027461320162, + 0.005271865986287594, + 0.005597949959337711, + 0.014561407268047333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.007843765430152416 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082562506198883 + ], + "opposite_logprob_tokens": [ + -4.141082763671875 + ], + "logprob_reward": -0.016082562506198883, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082562506198883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22353102266788483, + 0.13708773255348206, + 0.06608199328184128, + 0.05055287852883339, + 0.09930732101202011, + 0.06545903533697128, + 0.09956014156341553, + 0.08978681266307831, + 0.045352399349212646, + 0.12328065931797028 + ], + "credit_map": [ + 0.0265691839158535, + 0.0325709767639637, + 0.02314922586083412, + 0.024991096928715706, + 0.031627338379621506, + 0.03120555356144905, + 0.026051873341202736, + 0.027365781366825104, + 0.013724075630307198, + 0.010792768560349941, + 0.012820043601095676, + 0.017230015248060226, + 0.017606675624847412, + 0.01722691021859646, + 0.025693193078041077, + 0.021994061768054962, + 0.007286001928150654, + 0.006400247570127249, + 0.007541599217802286, + 0.005852476693689823, + 0.009389184415340424, + 0.007254797965288162, + 0.012729041278362274, + 0.009628639556467533, + 0.005441422574222088, + 0.005225233733654022, + 0.006008354015648365, + 0.004422742407768965, + 0.008517416194081306, + 0.0074008554220199585, + 0.007270399481058121, + 0.006266455166041851, + 0.010285573080182076, + 0.00935590360313654, + 0.012343569658696651, + 0.008798726834356785, + 0.019704977050423622, + 0.013455929234623909, + 0.01141125988215208, + 0.013951382599771023, + 0.0059845950454473495, + 0.004717570263892412, + 0.0052391835488379, + 0.007227483671158552, + 0.008055735379457474, + 0.00816122442483902, + 0.014385601505637169, + 0.011687636375427246, + 0.010386867448687553, + 0.015181455761194229, + 0.005200429353863001, + 0.017216328531503677, + 0.017018351703882217, + 0.014012555591762066, + 0.008274043910205364, + 0.012270105071365833, + 0.009352423250675201, + 0.010016637854278088, + 0.00771371740847826, + 0.00943305529654026, + 0.01668030582368374, + 0.008563539944589138, + 0.014752077870070934, + 0.013275058008730412, + 0.0037120156921446323, + 0.006339492741972208, + 0.0040615880861878395, + 0.0053670164197683334, + 0.006740076933056116, + 0.005035312380641699, + 0.005903820041567087, + 0.008193076588213444, + 0.012627637013792992, + 0.011570323258638382, + 0.014007450081408024, + 0.013086075894534588, + 0.011658730916678905, + 0.013799969106912613, + 0.014892630279064178, + 0.03163784742355347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.016082562506198883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370029106736183 + ], + "opposite_logprob_tokens": [ + -3.4087002277374268 + ], + "logprob_reward": -0.03370029106736183, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370029106736183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.136577770113945, + 0.04829961061477661, + 0.0608198456466198, + 0.07590162009000778, + 0.10895673930644989, + 0.13762643933296204, + 0.06564453989267349, + 0.09176582098007202, + 0.13621214032173157, + 0.1381954848766327 + ], + "credit_map": [ + 0.016003886237740517, + 0.018253067508339882, + 0.018493030220270157, + 0.01426878571510315, + 0.019506320357322693, + 0.01855989545583725, + 0.011592509225010872, + 0.01990026980638504, + 0.0048162429593503475, + 0.005050854291766882, + 0.0059356847777962685, + 0.005462171044200659, + 0.0075564198195934296, + 0.007967528887093067, + 0.005111541133373976, + 0.006399164441972971, + 0.0054261102341115475, + 0.007179218344390392, + 0.007199211046099663, + 0.0055244602262973785, + 0.008592121303081512, + 0.0057949041947722435, + 0.011040539480745792, + 0.010063275694847107, + 0.007057487964630127, + 0.008229208178818226, + 0.008395710028707981, + 0.008750593289732933, + 0.008076093159615993, + 0.008187687024474144, + 0.013782884925603867, + 0.013421956449747086, + 0.010691148228943348, + 0.010722257196903229, + 0.013183360919356346, + 0.016595570370554924, + 0.011894750408828259, + 0.013170456513762474, + 0.01681414432823658, + 0.015885045751929283, + 0.010776149109005928, + 0.012788071297109127, + 0.014129980467259884, + 0.01576414704322815, + 0.022575756534934044, + 0.013170995749533176, + 0.031982142478227615, + 0.01643918827176094, + 0.005384770222008228, + 0.009004284627735615, + 0.005566473584622145, + 0.013418091461062431, + 0.009147511795163155, + 0.006245806347578764, + 0.005090272985398769, + 0.011787324212491512, + 0.008293255232274532, + 0.010263483040034771, + 0.012166891247034073, + 0.009517383761703968, + 0.01319994404911995, + 0.009700196795165539, + 0.014885355718433857, + 0.013739313930273056, + 0.011547915637493134, + 0.013201318681240082, + 0.01575564593076706, + 0.012828580103814602, + 0.019469743594527245, + 0.015360028482973576, + 0.014738350175321102, + 0.03331054747104645, + 0.011414996348321438, + 0.017014067620038986, + 0.021138694137334824, + 0.022210022434592247, + 0.01645812951028347, + 0.013926612213253975, + 0.01581473834812641, + 0.020218225196003914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.03370029106736183 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950947757810354 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950947757810354, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950947757810354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11187011003494263, + 0.09484346210956573, + 0.05827716737985611, + 0.07076066732406616, + 0.12198885530233383, + 0.19527143239974976, + 0.07037600874900818, + 0.04899195581674576, + 0.09005972743034363, + 0.13756059110164642 + ], + "credit_map": [ + 0.011787873692810535, + 0.016986288130283356, + 0.01338969450443983, + 0.012861096300184727, + 0.016462406143546104, + 0.014830258674919605, + 0.011153344996273518, + 0.014399157837033272, + 0.01027203630656004, + 0.007992364466190338, + 0.009358719922602177, + 0.007657005917280912, + 0.013513150624930859, + 0.01258571445941925, + 0.02048511989414692, + 0.012979364953935146, + 0.005831242073327303, + 0.00670373672619462, + 0.006611404940485954, + 0.00519038550555706, + 0.007530249655246735, + 0.005567293148487806, + 0.011856184341013432, + 0.00898667611181736, + 0.007278075907379389, + 0.008247274905443192, + 0.005705727264285088, + 0.007999783381819725, + 0.010876770131289959, + 0.010675592347979546, + 0.012245926074683666, + 0.007731522433459759, + 0.013313038274645805, + 0.010972253978252411, + 0.010469187051057816, + 0.012056845240294933, + 0.028812149539589882, + 0.019645577296614647, + 0.014292888343334198, + 0.01242692768573761, + 0.01738741062581539, + 0.017066383734345436, + 0.019902698695659637, + 0.020049987360835075, + 0.03185749426484108, + 0.019214892759919167, + 0.04049364849925041, + 0.0292989369481802, + 0.008005623705685139, + 0.008748737163841724, + 0.009149541147053242, + 0.01083025336265564, + 0.009994681924581528, + 0.009325596503913403, + 0.004850469529628754, + 0.009471106342971325, + 0.004991068039089441, + 0.004034092184156179, + 0.0060988194309175014, + 0.004486564081162214, + 0.005994064267724752, + 0.006109706126153469, + 0.008158496581017971, + 0.00911915022879839, + 0.008968304842710495, + 0.01046635303646326, + 0.013000762090086937, + 0.013192988000810146, + 0.00988143589347601, + 0.01092506106942892, + 0.00938486959785223, + 0.014239960350096226, + 0.011953208595514297, + 0.016736755147576332, + 0.021033398807048798, + 0.0215590987354517, + 0.019176898524165154, + 0.01112302765250206, + 0.01974334940314293, + 0.016234859824180603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.005950947757810354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021984234917908907 + ], + "opposite_logprob_tokens": [ + -6.127198219299316 + ], + "logprob_reward": -0.0021984234917908907, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021984234917908907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14627179503440857, + 0.14402709901332855, + 0.09619126468896866, + 0.047311507165431976, + 0.10610731691122055, + 0.059016354382038116, + 0.10942864418029785, + 0.12902547419071198, + 0.05429543927311897, + 0.10832515358924866 + ], + "credit_map": [ + 0.014834748581051826, + 0.014329569414258003, + 0.021664272993803024, + 0.018461966887116432, + 0.01681811735033989, + 0.01745206117630005, + 0.016757085919380188, + 0.02595396898686886, + 0.014849275350570679, + 0.01273613702505827, + 0.02340340055525303, + 0.013545600697398186, + 0.016659138724207878, + 0.01956011727452278, + 0.01461008470505476, + 0.028663333505392075, + 0.012570313178002834, + 0.005916415248066187, + 0.020525364205241203, + 0.0073154945857822895, + 0.0069450801238417625, + 0.011000673286616802, + 0.00797242857515812, + 0.023945491760969162, + 0.00464156549423933, + 0.005474760662764311, + 0.00556171452626586, + 0.004586605820804834, + 0.0075375051237642765, + 0.00522732175886631, + 0.006909431889653206, + 0.007372602354735136, + 0.010684841312468052, + 0.008136446587741375, + 0.01213242020457983, + 0.009048146195709705, + 0.01979667693376541, + 0.012724651955068111, + 0.0097108855843544, + 0.02387324906885624, + 0.005376874003559351, + 0.00562108913436532, + 0.0076758842915296555, + 0.006642628461122513, + 0.009735752828419209, + 0.005737436935305595, + 0.010436855256557465, + 0.007789832539856434, + 0.013225136324763298, + 0.011428195051848888, + 0.012631990015506744, + 0.009001418948173523, + 0.01624985784292221, + 0.010970734059810638, + 0.009288455359637737, + 0.026632852852344513, + 0.012970955111086369, + 0.007602519355714321, + 0.01715885102748871, + 0.0068959686905145645, + 0.013674185611307621, + 0.0158219076693058, + 0.01906766928732395, + 0.03583341836929321, + 0.0047702481970191, + 0.007206418085843325, + 0.0050293998792767525, + 0.007663883734494448, + 0.009019090794026852, + 0.006657105404883623, + 0.006148041691631079, + 0.00780124869197607, + 0.011447817087173462, + 0.007987222634255886, + 0.012144776061177254, + 0.010844766162335873, + 0.01222669892013073, + 0.010919850319623947, + 0.018351072445511818, + 0.024402940645813942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0021984234917908907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a slightly low angle with a static camera focuses on a set of three ornate brass keys suspended from a thick brass ring. Each key, varying in size and complexity, features an intricate Victorian design with delicate filigree and uniquely shaped drill bits, its polished surface showing a subtle patina of age in the crevices. The entire set rotates in a slow, hypnotic, continuous rotation, rhythmically capturing the keys in a single, focused beam of warm golden light from an unseen source. As they turn, sharp, brilliant light glints across their detailed surfaces, momentarily illuminating the exquisite craftsmanship before gracefully fading into deep shadow. The background is a severely out-of-focus, dark, and moody space, suggestive of the interior of an old library, where the blurred forms of leather-bound books on wooden shelves are barely discernible. The shimmering highlights on the keys contrast starkly with the surrounding darkness, and combined with the deliberate, relentless movement, create an atmosphere of mystery and elegance, imbuing these simple objects with a profound sense of importance and an ineffable history.", + "seed": 11793, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_14/index_6590_seed_11793_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02741895182989538, + "vlm_logprob": -0.02741895182989538, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02741895182989538, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011752886697649956 + ], + "opposite_logprob_tokens": [ + -4.886753082275391 + ], + "logprob_reward": -0.011752886697649956, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011752886697649956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01142054982483387, + 0.010894733481109142, + 0.006832648068666458, + 0.006423975341022015, + 0.8911904096603394, + 0.023233534768223763, + 0.016734717413783073, + 0.014540031552314758, + 0.008335907943546772, + 0.010393458418548107 + ], + "credit_map": [ + 0.0010092149022966623, + 0.0007863094797357917, + 0.0016506115207448602, + 0.0005869322922080755, + 0.0009830920025706291, + 0.0037035613786429167, + 0.001017898553982377, + 0.0016829291125759482, + 0.0017587781185284257, + 0.0011603418970480561, + 0.0012603651266545057, + 0.0008433933253400028, + 0.0011845007538795471, + 0.0019092271104454994, + 0.0008895048522390425, + 0.001888622180558741, + 0.0007876208983361721, + 0.0007021230412647128, + 0.0006061085732653737, + 0.0007501475047320127, + 0.0005486792651936412, + 0.0008731816196814179, + 0.0010553550673648715, + 0.0015094317495822906, + 0.000519180204719305, + 0.0010159880621358752, + 0.0009649894200265408, + 0.0007834367570467293, + 0.0006678181234747171, + 0.0007336428388953209, + 0.0007978276116773486, + 0.0009410922648385167, + 0.09458271414041519, + 0.018178945407271385, + 0.33571842312812805, + 0.027195898815989494, + 0.06258682161569595, + 0.23042766749858856, + 0.045644089579582214, + 0.0768558531999588, + 0.002926043001934886, + 0.003308188868686557, + 0.0028234226629137993, + 0.0014789344277232885, + 0.0019525359384715557, + 0.0032883237581700087, + 0.0018876747926697135, + 0.005568408872932196, + 0.0017247044015675783, + 0.002433987334370613, + 0.0020892408210784197, + 0.0013840326573699713, + 0.0017528464086353779, + 0.002807529177516699, + 0.0013115984620526433, + 0.0032307785004377365, + 0.001785574248060584, + 0.0012563858181238174, + 0.0031802786979824305, + 0.0013347833883017302, + 0.001406133291311562, + 0.002083804225549102, + 0.0014119860716164112, + 0.0020810849964618683, + 0.0008719262550584972, + 0.0010388805530965328, + 0.0013181344838812947, + 0.0012619162444025278, + 0.0009350524051114917, + 0.0009547621593810618, + 0.0011695537250488997, + 0.0007856809534132481, + 0.0015334510244429111, + 0.0011825979454442859, + 0.0017944170394912362, + 0.0005717413150705397, + 0.0007855371222831309, + 0.0009832383366301656, + 0.0005973922088742256, + 0.002945082727819681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.011752886697649956 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370810344815254 + ], + "opposite_logprob_tokens": [ + -3.6513707637786865 + ], + "logprob_reward": -0.026370810344815254, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370810344815254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10250765830278397, + 0.10914039611816406, + 0.03514113649725914, + 0.08119263499975204, + 0.32764315605163574, + 0.08487050980329514, + 0.08635768294334412, + 0.0713137686252594, + 0.050750620663166046, + 0.05108235403895378 + ], + "credit_map": [ + 0.01804923824965954, + 0.007365553639829159, + 0.016172515228390694, + 0.00842282921075821, + 0.008863244205713272, + 0.022864937782287598, + 0.00987887941300869, + 0.010890467092394829, + 0.018750041723251343, + 0.010978863574564457, + 0.009439997375011444, + 0.007012287620455027, + 0.009307309053838253, + 0.023321306332945824, + 0.00633960822597146, + 0.023990994319319725, + 0.005446694325655699, + 0.004284735303372145, + 0.003255473216995597, + 0.003231828799471259, + 0.0037788336630910635, + 0.004534719977527857, + 0.0045424881391227245, + 0.0060663684271276, + 0.007053777575492859, + 0.010627681389451027, + 0.010642861016094685, + 0.007625538855791092, + 0.013645794242620468, + 0.00773378973826766, + 0.011193967424333096, + 0.012669232673943043, + 0.04239208996295929, + 0.02902122586965561, + 0.08899875730276108, + 0.01875988021492958, + 0.02272450178861618, + 0.05102161318063736, + 0.027178775519132614, + 0.047546349465847015, + 0.010568712837994099, + 0.008381500840187073, + 0.015665391460061073, + 0.007819079793989658, + 0.007242930121719837, + 0.010422729887068272, + 0.007514840457588434, + 0.017255334183573723, + 0.009565720334649086, + 0.00823132786899805, + 0.012470968998968601, + 0.00781919900327921, + 0.007597790099680424, + 0.014018109068274498, + 0.007771046832203865, + 0.018883535638451576, + 0.007411373779177666, + 0.006217934191226959, + 0.012435161508619785, + 0.009633620269596577, + 0.00791306234896183, + 0.009108681231737137, + 0.010963069275021553, + 0.0076308706775307655, + 0.006097707897424698, + 0.005794696509838104, + 0.011043847538530827, + 0.0036245551891624928, + 0.0050492542795836926, + 0.005947644356638193, + 0.004748908802866936, + 0.008444012142717838, + 0.006337253376841545, + 0.0057188100181519985, + 0.007807323243469, + 0.005685874726623297, + 0.004724021069705486, + 0.005136535502970219, + 0.005387184675782919, + 0.01028535608202219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.026370810344815254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903542906045914 + ], + "opposite_logprob_tokens": [ + -2.4640355110168457 + ], + "logprob_reward": -0.08903542906045914, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903542906045914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01912456564605236, + 0.00990928802639246, + 0.00966245774179697, + 0.02038249745965004, + 0.8712466955184937, + 0.030316654592752457, + 0.012075032107532024, + 0.013525264337658882, + 0.005133436061441898, + 0.008624104782938957 + ], + "credit_map": [ + 0.0015129141975194216, + 0.0015525915659964085, + 0.002641440834850073, + 0.0011011328315362334, + 0.0025391732342541218, + 0.004795089829713106, + 0.002760391915217042, + 0.002221828093752265, + 0.0016799725126475096, + 0.00111387949436903, + 0.0009990574326366186, + 0.0007310811779461801, + 0.0008810636936686933, + 0.00205692439340055, + 0.0005901589174754918, + 0.001857149414718151, + 0.001081847702153027, + 0.0014667724026367068, + 0.0009547538356855512, + 0.0012862386647611856, + 0.001004172838293016, + 0.0013987348647788167, + 0.0011747602839022875, + 0.0012951758690178394, + 0.0025544518139213324, + 0.0022592463064938784, + 0.004717356991022825, + 0.0017803653609007597, + 0.0017570379422977567, + 0.002758400049060583, + 0.0022520129568874836, + 0.002303624525666237, + 0.09257493913173676, + 0.02001204714179039, + 0.3222014904022217, + 0.02667536959052086, + 0.060495566576719284, + 0.2161300629377365, + 0.047144971787929535, + 0.08601220697164536, + 0.003502103965729475, + 0.0058731636963784695, + 0.003483798587694764, + 0.0026806448586285114, + 0.002772391540929675, + 0.003972481470555067, + 0.002344288630411029, + 0.005687778349965811, + 0.0013182872207835317, + 0.0015155150322243571, + 0.001675344305112958, + 0.0012326352298259735, + 0.0012379793915897608, + 0.0018462686566635966, + 0.0012711338931694627, + 0.001977866515517235, + 0.0013376830611377954, + 0.0009816940873861313, + 0.0021094991825520992, + 0.001979211578145623, + 0.0012419658014550805, + 0.0018825953593477607, + 0.0021118689328432083, + 0.0018807437736541033, + 0.0006294547929428518, + 0.0006651839357800782, + 0.0008649108931422234, + 0.00042982172453776, + 0.0004596378421410918, + 0.0006481605232693255, + 0.0004684502782765776, + 0.0009678148780949414, + 0.0011636598501354456, + 0.0010217991657555103, + 0.0013720545684918761, + 0.0006083751795813441, + 0.0007484794477932155, + 0.0008623535395599902, + 0.0006466907798312604, + 0.0022006910294294357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.08903542906045914 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740807089954615 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740807089954615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740807089954615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02761196158826351, + 0.03826681524515152, + 0.015580857172608376, + 0.03863903507590294, + 0.6552053093910217, + 0.058832503855228424, + 0.04556117579340935, + 0.08086765557527542, + 0.021450361236929893, + 0.017984358593821526 + ], + "credit_map": [ + 0.0028296473901718855, + 0.0024781920947134495, + 0.004036053083837032, + 0.0022761209402233362, + 0.003717472543939948, + 0.005970366299152374, + 0.002815305721014738, + 0.0034888037480413914, + 0.005287330597639084, + 0.0031040727626532316, + 0.0034988466650247574, + 0.0021109567023813725, + 0.0035645070020109415, + 0.008949530310928822, + 0.0028558799531310797, + 0.0088956905528903, + 0.0017161433352157474, + 0.001986259827390313, + 0.0016612190520390868, + 0.0019881839398294687, + 0.001380448928102851, + 0.0019400031305849552, + 0.0019436451839283109, + 0.002964954124763608, + 0.0033536397386342287, + 0.004734978079795837, + 0.0048505994491279125, + 0.004586636088788509, + 0.004920898005366325, + 0.0036295063327997923, + 0.005308734718710184, + 0.0072540403343737125, + 0.07065532356500626, + 0.01859174482524395, + 0.2366575449705124, + 0.020377134904265404, + 0.044760867953300476, + 0.15470561385154724, + 0.03822103515267372, + 0.07123599201440811, + 0.008984403684735298, + 0.003916995134204626, + 0.006711233872920275, + 0.003917383961379528, + 0.0038583935238420963, + 0.013247624039649963, + 0.003515163902193308, + 0.01468130573630333, + 0.005012359004467726, + 0.004563349299132824, + 0.006270890589803457, + 0.0037916263099759817, + 0.0042384229600429535, + 0.007925945334136486, + 0.003378329798579216, + 0.01038025226444006, + 0.006489508785307407, + 0.007674032356590033, + 0.00964681152254343, + 0.014239799231290817, + 0.007592396344989538, + 0.010609021410346031, + 0.014384448528289795, + 0.010231631807982922, + 0.002702363533899188, + 0.0018259674543514848, + 0.0039009619504213333, + 0.0016057651955634356, + 0.0022763311862945557, + 0.0037151791620999575, + 0.001720780273899436, + 0.0037030137609690428, + 0.0022819642908871174, + 0.0028012911789119244, + 0.0030960354488343, + 0.0015217900509014726, + 0.002442294964566827, + 0.0017779566114768386, + 0.0014817886985838413, + 0.002581237116828561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.006740807089954615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031948259565979242 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031948259565979242, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031948259565979242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.006333702243864536, + 0.0021673713345080614, + 0.001962672919034958, + 0.005826458800584078, + 0.9380930662155151, + 0.007411484140902758, + 0.019435705617070198, + 0.005306680221110582, + 0.007689241319894791, + 0.005773628130555153 + ], + "credit_map": [ + 0.0007482938235625625, + 0.00041700538713485, + 0.0009487267816439271, + 0.00042898280662484467, + 0.000663105514831841, + 0.0017291076947003603, + 0.0004980784724466503, + 0.0009004002204164863, + 0.00026869477005675435, + 0.00017736840527504683, + 0.00023758718452882022, + 0.00011432338214945048, + 0.00015535221609752625, + 0.0005941835697740316, + 0.00016895195585675538, + 0.0004509093123488128, + 0.00028392879175953567, + 0.0002459787647239864, + 0.00021552611724473536, + 0.00020979299733880907, + 0.00018233062291983515, + 0.0002338658377993852, + 0.0002083459694404155, + 0.0003829032648354769, + 0.0005913974018767476, + 0.0007163416594266891, + 0.0010540146613493562, + 0.0004985234118066728, + 0.0003570739645510912, + 0.0007129808655008674, + 0.0006471656379289925, + 0.0012489601504057646, + 0.10012785345315933, + 0.02031891793012619, + 0.34982284903526306, + 0.02860397659242153, + 0.06496530771255493, + 0.23566913604736328, + 0.050217751413583755, + 0.08836717903614044, + 0.0009077095310203731, + 0.0004299965803511441, + 0.0012790949549525976, + 0.0005103962612338364, + 0.00045327242696657777, + 0.0010407441295683384, + 0.0006073356489650905, + 0.002182932570576668, + 0.0025405092164874077, + 0.0027928545605391264, + 0.002752560656517744, + 0.0016708301845937967, + 0.001304976874962449, + 0.0032202075235545635, + 0.0018304759869351983, + 0.0033232863061130047, + 0.0005884863785468042, + 0.0004075234173797071, + 0.0011765937088057399, + 0.00045957413385622203, + 0.00046851253136992455, + 0.0009544425993226469, + 0.00036638989695347846, + 0.0008851567399688065, + 0.0008426338317804039, + 0.000646962143946439, + 0.0014186843764036894, + 0.000707446422893554, + 0.000604501343332231, + 0.0012873081723228097, + 0.0008981560240499675, + 0.0012835472589358687, + 0.0007369414670392871, + 0.000750935054384172, + 0.0009394966182298958, + 0.00039922911673784256, + 0.00047408920363523066, + 0.0005054274224676192, + 0.00041530453017912805, + 0.0015522034373134375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.0031948259565979242 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a slightly low angle with a static camera focuses on a set of three ornate brass keys suspended from a thick brass ring. Each key, varying in size and complexity, features an intricate Victorian design with delicate filigree and uniquely shaped drill bits, its polished surface showing a subtle patina of age in the crevices. The entire set rotates in a slow, hypnotic, continuous rotation, rhythmically capturing the keys in a single, focused beam of warm golden light from an unseen source. As they turn, sharp, brilliant light glints across their detailed surfaces, momentarily illuminating the exquisite craftsmanship before gracefully fading into deep shadow. The background is a severely out-of-focus, dark, and moody space, suggestive of the interior of an old library, where the blurred forms of leather-bound books on wooden shelves are barely discernible. The shimmering highlights on the keys contrast starkly with the surrounding darkness, and combined with the deliberate, relentless movement, create an atmosphere of mystery and elegance, imbuing these simple objects with a profound sense of importance and an ineffable history.", + "seed": 11794, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_14/index_6590_seed_11794_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019720665458589793, + "vlm_logprob": -0.019720665458589793, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.019720665458589793, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012712513096630573 + ], + "opposite_logprob_tokens": [ + -5.012712478637695 + ], + "logprob_reward": -0.012712513096630573, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012712513096630573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02464674413204193, + 0.04112396016716957, + 0.07091913372278214, + 0.054134778678417206, + 0.5445528030395508, + 0.03498966619372368, + 0.14896851778030396, + 0.016708586364984512, + 0.028099318966269493, + 0.03585643321275711 + ], + "credit_map": [ + 0.0031067482195794582, + 0.0025424198247492313, + 0.003736875718459487, + 0.002414728282019496, + 0.002580096013844013, + 0.0031654450576752424, + 0.003050947794690728, + 0.004049483686685562, + 0.0054184854961931705, + 0.005141811445355415, + 0.0077249398455023766, + 0.007583478931337595, + 0.0035087016876786947, + 0.003853465896099806, + 0.0037260279059410095, + 0.004167047329246998, + 0.008742754347622395, + 0.005732855759561062, + 0.009008506312966347, + 0.006993374787271023, + 0.005635272711515427, + 0.015751276165246964, + 0.006710780318826437, + 0.01234431378543377, + 0.005595668219029903, + 0.00383237237110734, + 0.008837762288749218, + 0.004206476267427206, + 0.005283349193632603, + 0.009792068041861057, + 0.0047698793932795525, + 0.0118172075599432, + 0.03141789510846138, + 0.011916844174265862, + 0.026388410478830338, + 0.007239046972244978, + 0.010608389973640442, + 0.2333265095949173, + 0.011934937909245491, + 0.21172083914279938, + 0.0033708123955875635, + 0.002979877172037959, + 0.004856779705733061, + 0.0038587336894124746, + 0.004964242223650217, + 0.004820156842470169, + 0.00480957655236125, + 0.005329489707946777, + 0.014727989211678505, + 0.0034378012642264366, + 0.05644378066062927, + 0.006903610657900572, + 0.005529060494154692, + 0.04641913250088692, + 0.004675573669373989, + 0.01083157118409872, + 0.002463249024003744, + 0.003893766086548567, + 0.0022405628114938736, + 0.0026599415577948093, + 0.0011449238518252969, + 0.0014420172665268183, + 0.0010117492638528347, + 0.0018523781327530742, + 0.0036133320536464453, + 0.00394321046769619, + 0.0034619118086993694, + 0.006491756532341242, + 0.0024901593569666147, + 0.002666137646883726, + 0.003379080444574356, + 0.0020537315867841244, + 0.004402437247335911, + 0.003478831611573696, + 0.00507597578689456, + 0.005008533131331205, + 0.00412431126460433, + 0.0026241971645504236, + 0.006071379408240318, + 0.005070768762379885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.012712513096630573 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591415464878082 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591415464878082, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591415464878082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06578852981328964, + 0.030603665858507156, + 0.04747001454234123, + 0.4794224500656128, + 0.07621738314628601, + 0.09737012535333633, + 0.0709988996386528, + 0.06981566548347473, + 0.02815469726920128, + 0.034158576279878616 + ], + "credit_map": [ + 0.011685709469020367, + 0.005688875913619995, + 0.014253506436944008, + 0.006663520820438862, + 0.00588225806131959, + 0.007450117263942957, + 0.007848190143704414, + 0.00631635682657361, + 0.003608925500884652, + 0.003697874490171671, + 0.005221785511821508, + 0.0041647725738584995, + 0.0027910617645829916, + 0.003370243590325117, + 0.0027179180178791285, + 0.005031089298427105, + 0.006625744514167309, + 0.004190940875560045, + 0.010835749097168446, + 0.004265681374818087, + 0.0036153439432382584, + 0.008283297531306744, + 0.003151706652715802, + 0.006501554977148771, + 0.047236643731594086, + 0.018563508987426758, + 0.09968861192464828, + 0.02747553586959839, + 0.03797459602355957, + 0.09242802858352661, + 0.027761848643422127, + 0.1282936930656433, + 0.005057059228420258, + 0.002018841216340661, + 0.00723171466961503, + 0.00233090459369123, + 0.0029193696100264788, + 0.029284896329045296, + 0.004205743782222271, + 0.023168858140707016, + 0.01107795163989067, + 0.006569084711372852, + 0.016689250245690346, + 0.008334586396813393, + 0.01084194052964449, + 0.01783127523958683, + 0.00930263102054596, + 0.01672341115772724, + 0.006062587257474661, + 0.0027967931237071753, + 0.022020837292075157, + 0.0043429057113826275, + 0.004497040528804064, + 0.021990258246660233, + 0.003151008626446128, + 0.006137468386441469, + 0.007282100152224302, + 0.009820637293159962, + 0.01063169352710247, + 0.01223041582852602, + 0.007004487328231335, + 0.006796694826334715, + 0.006221491377800703, + 0.009828142821788788, + 0.0031029973179101944, + 0.003990337252616882, + 0.0034888149239122868, + 0.004849335644394159, + 0.002319271909072995, + 0.0029943299014121294, + 0.0028022844344377518, + 0.004607328679412603, + 0.004567614756524563, + 0.0038355509750545025, + 0.004308355040848255, + 0.0048971278592944145, + 0.004095942713320255, + 0.0033573440741747618, + 0.0046372669748961926, + 0.004459376446902752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.020591415464878082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494876950979233 + ], + "opposite_logprob_tokens": [ + -2.9299488067626953 + ], + "logprob_reward": -0.05494876950979233, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494876950979233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02410106174647808, + 0.012402573600411415, + 0.05616215988993645, + 0.2584979832172394, + 0.4368363618850708, + 0.06771253049373627, + 0.0998050719499588, + 0.021723367273807526, + 0.010827908292412758, + 0.011930984444916248 + ], + "credit_map": [ + 0.004434038884937763, + 0.0021420414559543133, + 0.005355828441679478, + 0.0022145439870655537, + 0.0019966913387179375, + 0.002474983222782612, + 0.002591431373730302, + 0.0028915037401020527, + 0.001354249776341021, + 0.0012636302271857858, + 0.002004701877012849, + 0.0012841875432059169, + 0.0009753562626428902, + 0.0018618827452883124, + 0.0016663979040458798, + 0.0019921669736504555, + 0.006718883290886879, + 0.004713431466370821, + 0.014848515391349792, + 0.004659229423850775, + 0.004129990004003048, + 0.008639320731163025, + 0.005071424413472414, + 0.007381363771855831, + 0.025495266541838646, + 0.010252547450363636, + 0.05398079752922058, + 0.01367995236068964, + 0.0242270827293396, + 0.05093464255332947, + 0.016315089538693428, + 0.06361261010169983, + 0.02557852678000927, + 0.011275999248027802, + 0.02134530059993267, + 0.006778115872293711, + 0.00875303614884615, + 0.18231476843357086, + 0.009933819994330406, + 0.1708567887544632, + 0.00708055030554533, + 0.005380151327699423, + 0.01087124366313219, + 0.005931403953582048, + 0.006501436233520508, + 0.014265906997025013, + 0.008574852719902992, + 0.009106985293328762, + 0.010405780747532845, + 0.0022947737015783787, + 0.03919415548443794, + 0.004509942140430212, + 0.002746048616245389, + 0.031303323805332184, + 0.002143171615898609, + 0.007207878399640322, + 0.0029346919618546963, + 0.004166477359831333, + 0.0028706155717372894, + 0.003185831243172288, + 0.002222395036369562, + 0.002258747350424528, + 0.0014253532281145453, + 0.0026592561043798923, + 0.0011353768641129136, + 0.0017575259553268552, + 0.0012361909030005336, + 0.0028600578662008047, + 0.000820305896922946, + 0.0008540750131942332, + 0.0009578599710948765, + 0.0012065168702974916, + 0.0015317872166633606, + 0.0010351230157539248, + 0.001415354199707508, + 0.0015178296016529202, + 0.001074736937880516, + 0.0012701754458248615, + 0.00135138351470232, + 0.002734594978392124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.05494876950979233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736780982464552 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736780982464552, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736780982464552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045621633529663086, + 0.030359115451574326, + 0.0512775294482708, + 0.1825803816318512, + 0.26508525013923645, + 0.10767523944377899, + 0.21008804440498352, + 0.043281130492687225, + 0.026074983179569244, + 0.03795670345425606 + ], + "credit_map": [ + 0.007323185447603464, + 0.00511774281039834, + 0.009752129204571247, + 0.005107077304273844, + 0.004223424941301346, + 0.004669914487749338, + 0.004752080887556076, + 0.004676079377532005, + 0.0029612071812152863, + 0.003706302959471941, + 0.004211004823446274, + 0.0048155346885323524, + 0.0038458893541246653, + 0.0026389288250356913, + 0.0029358817264437675, + 0.005244369152933359, + 0.007134279236197472, + 0.006675575859844685, + 0.008086604997515678, + 0.005057126749306917, + 0.00513356551527977, + 0.007301157806068659, + 0.005743041634559631, + 0.006146176718175411, + 0.017559507861733437, + 0.00803456362336874, + 0.03761789947748184, + 0.010023296810686588, + 0.013223735615611076, + 0.03644644841551781, + 0.00934518314898014, + 0.050329744815826416, + 0.01490423921495676, + 0.0055211856961250305, + 0.016603829339146614, + 0.006375288590788841, + 0.0070390175096690655, + 0.10770554095506668, + 0.011570676229894161, + 0.09536546468734741, + 0.011613903567194939, + 0.010140414349734783, + 0.01661165989935398, + 0.01256831269711256, + 0.011090660467743874, + 0.016380377113819122, + 0.013286336325109005, + 0.01598357781767845, + 0.015825476497411728, + 0.007093018386512995, + 0.06426146626472473, + 0.011091196909546852, + 0.006870313081890345, + 0.07597741484642029, + 0.006304771639406681, + 0.02266441285610199, + 0.005053856875747442, + 0.006084555760025978, + 0.007119344547390938, + 0.007912691682577133, + 0.004337188322097063, + 0.004355792421847582, + 0.0032803507056087255, + 0.005137347616255283, + 0.0033846322912722826, + 0.004481716081500053, + 0.003641054965555668, + 0.005775503348559141, + 0.00200452352873981, + 0.0023763813078403473, + 0.0028056150767952204, + 0.001605556346476078, + 0.006301596760749817, + 0.00302611174993217, + 0.00773634621873498, + 0.004310432355850935, + 0.0037890581879764795, + 0.002337356097996235, + 0.005647087935358286, + 0.004808716010302305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.006736780982464552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613848239183426 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.003613848239183426, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613848239183426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0963304191827774, + 0.06502897292375565, + 0.061417482793331146, + 0.3628945052623749, + 0.08332781493663788, + 0.11387216299772263, + 0.10560683906078339, + 0.02892959490418434, + 0.05571241304278374, + 0.02687973715364933 + ], + "credit_map": [ + 0.012301508337259293, + 0.005611448548734188, + 0.015277424827218056, + 0.008569948375225067, + 0.007171836216002703, + 0.01366173755377531, + 0.01221189834177494, + 0.021524615585803986, + 0.008654074743390083, + 0.0045647528022527695, + 0.011456788517534733, + 0.004273504018783569, + 0.004326048772782087, + 0.011422150768339634, + 0.005039973650127649, + 0.015291686169803143, + 0.008475512266159058, + 0.004755593836307526, + 0.014729147776961327, + 0.004557767882943153, + 0.004847803618758917, + 0.009368833154439926, + 0.0050988937728106976, + 0.009583929553627968, + 0.03458758071064949, + 0.019333798438310623, + 0.07314994186162949, + 0.02360483445227146, + 0.02854265086352825, + 0.06541133671998978, + 0.02732009068131447, + 0.09094428271055222, + 0.006401524413377047, + 0.002552081597968936, + 0.006813539657741785, + 0.002976880641654134, + 0.002907651010900736, + 0.032969534397125244, + 0.003983278293162584, + 0.024723326787352562, + 0.013416687026619911, + 0.006808140315115452, + 0.020083915442228317, + 0.009273014962673187, + 0.011302972212433815, + 0.02437838539481163, + 0.011080625467002392, + 0.01752842217683792, + 0.0057753510773181915, + 0.0036529661156237125, + 0.022549664601683617, + 0.0051249233074486256, + 0.004849686753004789, + 0.043533168733119965, + 0.0031983868684619665, + 0.016922689974308014, + 0.004422488156706095, + 0.006193789187818766, + 0.00404699333012104, + 0.0037366575561463833, + 0.0028323496226221323, + 0.002781998598948121, + 0.0020419969223439693, + 0.0028733222279697657, + 0.006605327595025301, + 0.008646827191114426, + 0.008904838003218174, + 0.015476791188120842, + 0.0036544085014611483, + 0.004318450577557087, + 0.004885783884674311, + 0.0032199847046285868, + 0.003826332511380315, + 0.00330382795073092, + 0.0036238476168364286, + 0.004430751316249371, + 0.0026615746319293976, + 0.0016960168723016977, + 0.003919519018381834, + 0.0034178677015006542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.003613848239183426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a slightly low angle with a static camera focuses on a set of three ornate brass keys suspended from a thick brass ring. Each key, varying in size and complexity, features an intricate Victorian design with delicate filigree and uniquely shaped drill bits, its polished surface showing a subtle patina of age in the crevices. The entire set rotates in a slow, hypnotic, continuous rotation, rhythmically capturing the keys in a single, focused beam of warm golden light from an unseen source. As they turn, sharp, brilliant light glints across their detailed surfaces, momentarily illuminating the exquisite craftsmanship before gracefully fading into deep shadow. The background is a severely out-of-focus, dark, and moody space, suggestive of the interior of an old library, where the blurred forms of leather-bound books on wooden shelves are barely discernible. The shimmering highlights on the keys contrast starkly with the surrounding darkness, and combined with the deliberate, relentless movement, create an atmosphere of mystery and elegance, imbuing these simple objects with a profound sense of importance and an ineffable history.", + "seed": 11795, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_14/index_6590_seed_11795_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024802843760699033, + "vlm_logprob": -0.024802843760699033, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.024802843760699033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00809093564748764 + ], + "opposite_logprob_tokens": [ + -5.258090972900391 + ], + "logprob_reward": -0.00809093564748764, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00809093564748764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1718415468931198, + 0.03827451542019844, + 0.09143754839897156, + 0.2200312465429306, + 0.07589410245418549, + 0.03904658555984497, + 0.05573417618870735, + 0.07488816976547241, + 0.14904405176639557, + 0.0838080644607544 + ], + "credit_map": [ + 0.018791386857628822, + 0.010513495653867722, + 0.016051674261689186, + 0.012194216251373291, + 0.014824810437858105, + 0.055335938930511475, + 0.012143953703343868, + 0.03198608383536339, + 0.003677805420011282, + 0.0036886839661747217, + 0.004728588741272688, + 0.005604766774922609, + 0.005364240147173405, + 0.004616912454366684, + 0.003792812814936042, + 0.006800709757953882, + 0.011236422695219517, + 0.023841621354222298, + 0.015086011961102486, + 0.008592247031629086, + 0.006198822986334562, + 0.010247266851365566, + 0.004110938403755426, + 0.012124219909310341, + 0.012122089974582195, + 0.011848101392388344, + 0.022789884358644485, + 0.019112607464194298, + 0.022908488288521767, + 0.04248085245490074, + 0.02218223549425602, + 0.06658697873353958, + 0.007457171566784382, + 0.005848321132361889, + 0.012976921163499355, + 0.007132743950933218, + 0.007140098139643669, + 0.014163562096655369, + 0.0059402501210570335, + 0.015235042199492455, + 0.004684202838689089, + 0.004791907500475645, + 0.007443016394972801, + 0.0049550035037100315, + 0.00298058963380754, + 0.0037256779614835978, + 0.002880081534385681, + 0.007586104795336723, + 0.007447205483913422, + 0.0071324435994029045, + 0.012696522288024426, + 0.00904545933008194, + 0.004651921801269054, + 0.00645942147821188, + 0.0034802635200321674, + 0.00482094194740057, + 0.008387201465666294, + 0.008335771039128304, + 0.01669255830347538, + 0.008846851997077465, + 0.00863139983266592, + 0.011941643431782722, + 0.0045593115501105785, + 0.007493436336517334, + 0.017927085980772972, + 0.005916589871048927, + 0.041015833616256714, + 0.01167879905551672, + 0.007882416248321533, + 0.01844615675508976, + 0.005087166093289852, + 0.041090015321969986, + 0.015057778917253017, + 0.007058837451040745, + 0.019776687026023865, + 0.007788441609591246, + 0.006640301551669836, + 0.00894912052899599, + 0.009873783215880394, + 0.008663115091621876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.00809093564748764 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635919861495495 + ], + "opposite_logprob_tokens": [ + -3.6513590812683105 + ], + "logprob_reward": -0.02635919861495495, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635919861495495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09868420660495758, + 0.03425681218504906, + 0.037820156663656235, + 0.30683985352516174, + 0.024317465722560883, + 0.02985193394124508, + 0.0350358709692955, + 0.0288202203810215, + 0.3414627015590668, + 0.06291075050830841 + ], + "credit_map": [ + 0.013525526970624924, + 0.006290333345532417, + 0.008955040946602821, + 0.021176956593990326, + 0.006362158805131912, + 0.018555311486124992, + 0.006998253054916859, + 0.016820628196001053, + 0.0027873129583895206, + 0.0027679160702973604, + 0.0045913150534033775, + 0.0045251077972352505, + 0.003958128858357668, + 0.007120730821043253, + 0.0028068486135452986, + 0.005699452478438616, + 0.004896176978945732, + 0.006025956477969885, + 0.004593184217810631, + 0.00448412774130702, + 0.004805658478289843, + 0.004787770565599203, + 0.002826712792739272, + 0.00540057010948658, + 0.017803208902478218, + 0.019683653488755226, + 0.03084980696439743, + 0.031644679605960846, + 0.027177724987268448, + 0.058746013790369034, + 0.03387833014130592, + 0.08705643564462662, + 0.0020723340567201376, + 0.003242908511310816, + 0.0027277732733637094, + 0.003942644689232111, + 0.0033962358720600605, + 0.0030209955293685198, + 0.0025883165653795004, + 0.003326257225126028, + 0.0031509241089224815, + 0.0033252607099711895, + 0.0059535871259868145, + 0.0063522276468575, + 0.003450185526162386, + 0.002379280049353838, + 0.0019322697771713138, + 0.003308199578896165, + 0.003861809615045786, + 0.00435686856508255, + 0.006945601664483547, + 0.005829333793371916, + 0.0041339946910738945, + 0.0034187515266239643, + 0.002403890248388052, + 0.00408562459051609, + 0.003489219816401601, + 0.0031419526785612106, + 0.003670499427244067, + 0.004358103033155203, + 0.004066736903041601, + 0.003780996659770608, + 0.0026952940970659256, + 0.0036174191627651453, + 0.04694226384162903, + 0.012289033271372318, + 0.10834866762161255, + 0.020637191832065582, + 0.014684010297060013, + 0.03041115403175354, + 0.010371102020144463, + 0.0977792888879776, + 0.008531388826668262, + 0.00399143248796463, + 0.009483655914664268, + 0.010387681424617767, + 0.0046383352018892765, + 0.006932384334504604, + 0.009424434043467045, + 0.009521447122097015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.02635919861495495 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897688448429108 + ], + "opposite_logprob_tokens": [ + -2.57897686958313 + ], + "logprob_reward": -0.07897688448429108, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897688448429108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10214310884475708, + 0.054318834096193314, + 0.04855278134346008, + 0.09291355311870575, + 0.05825282260775566, + 0.14093339443206787, + 0.018415501341223717, + 0.08240744471549988, + 0.3199272155761719, + 0.0821354016661644 + ], + "credit_map": [ + 0.0070180087350308895, + 0.004720991477370262, + 0.007578498683869839, + 0.010962065309286118, + 0.006413520313799381, + 0.033705323934555054, + 0.005687236785888672, + 0.02605745755136013, + 0.00546741858124733, + 0.004570456687361002, + 0.009074504487216473, + 0.005459361709654331, + 0.005145538132637739, + 0.012344256974756718, + 0.004295366816222668, + 0.00796192791312933, + 0.005721621215343475, + 0.007947245612740517, + 0.011879167519509792, + 0.0037937345914542675, + 0.005232654977589846, + 0.005839142482727766, + 0.00253561744466424, + 0.00560359051451087, + 0.0072671035304665565, + 0.00836727861315012, + 0.010559520684182644, + 0.012201535515487194, + 0.008710102178156376, + 0.013731579296290874, + 0.012092571705579758, + 0.01998385228216648, + 0.005574855953454971, + 0.007560337893664837, + 0.006300225853919983, + 0.006611324846744537, + 0.005164481233805418, + 0.010340551845729351, + 0.0052841417491436005, + 0.011416896246373653, + 0.01674487069249153, + 0.017364587634801865, + 0.018519766628742218, + 0.016881760209798813, + 0.020223120227456093, + 0.015582987107336521, + 0.013983628712594509, + 0.02163265459239483, + 0.0028730405028909445, + 0.0017122890567407012, + 0.0035590899642556906, + 0.00251942896284163, + 0.0019227623706683517, + 0.0016474025323987007, + 0.0019022285705432296, + 0.00227925693616271, + 0.010325279086828232, + 0.008839288726449013, + 0.010100927203893661, + 0.012360235676169395, + 0.0074318754486739635, + 0.017212631180882454, + 0.006095351651310921, + 0.01004184689372778, + 0.0474228672683239, + 0.010499570518732071, + 0.10454099625349045, + 0.017746232450008392, + 0.010164936073124409, + 0.027586018666625023, + 0.00853645894676447, + 0.09343009442090988, + 0.01662353239953518, + 0.0057129389606416225, + 0.02184535376727581, + 0.009121349081397057, + 0.0050857956521213055, + 0.008518250659108162, + 0.008918961510062218, + 0.006309208460152149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.07897688448429108 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951184779405594 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951184779405594, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951184779405594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1054731011390686, + 0.03316619247198105, + 0.06452286243438721, + 0.10430876165628433, + 0.048371199518442154, + 0.05003499239683151, + 0.0864967405796051, + 0.06422056257724762, + 0.4041709899902344, + 0.039234619587659836 + ], + "credit_map": [ + 0.0111626498401165, + 0.006325810216367245, + 0.008807101286947727, + 0.012025507166981697, + 0.006726244930177927, + 0.024464759975671768, + 0.005187981761991978, + 0.03077305480837822, + 0.0035172230564057827, + 0.0036167120561003685, + 0.0041871825233101845, + 0.0071773952804505825, + 0.0031764835584908724, + 0.0038218542467802763, + 0.0029656030237674713, + 0.00470374058932066, + 0.007923603989183903, + 0.009190874174237251, + 0.011554746888577938, + 0.006928185932338238, + 0.0077773192897439, + 0.006914566271007061, + 0.005298501346260309, + 0.008935069665312767, + 0.006849214434623718, + 0.006477656774222851, + 0.010244988836348057, + 0.010841009207069874, + 0.014245829544961452, + 0.01695661060512066, + 0.012327771633863449, + 0.02636568807065487, + 0.006814599968492985, + 0.006888261996209621, + 0.0051204445771873, + 0.007051797118037939, + 0.004823415074497461, + 0.007524371612817049, + 0.002625179011374712, + 0.007523129228502512, + 0.009362034499645233, + 0.006268475204706192, + 0.005307062063366175, + 0.006687471643090248, + 0.0049733794294297695, + 0.005428636446595192, + 0.004465961363166571, + 0.007541969418525696, + 0.009511878713965416, + 0.008996118791401386, + 0.020010437816381454, + 0.009932318702340126, + 0.010310960933566093, + 0.009484365582466125, + 0.010064979083836079, + 0.008185682818293571, + 0.009140443056821823, + 0.007361371535807848, + 0.013528495095670223, + 0.010970239527523518, + 0.006176251918077469, + 0.008100355044007301, + 0.004516070708632469, + 0.004427340347319841, + 0.06099448353052139, + 0.01171834859997034, + 0.13436955213546753, + 0.020268624648451805, + 0.012322910130023956, + 0.03512457385659218, + 0.01024461630731821, + 0.11912786215543747, + 0.004489417187869549, + 0.004269998986274004, + 0.0063301874324679375, + 0.006007398013025522, + 0.004185250960290432, + 0.006075650919228792, + 0.003933131694793701, + 0.003943585325032473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.005951184779405594 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of three ornate brass keys suspended from a thick brass ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the keys feature intricate Victorian designs with delicate filigree and uniquely shaped drill bits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera static and positioned at a slightly low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the set of keys rotate slowly and continuously in a single beam of warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a severely out-of-focus dark space with barely discernible blurred forms of leather-bound books on wooden shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636015277355909 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004636015277355909, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636015277355909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020988859236240387, + 0.011189015582203865, + 0.11696770042181015, + 0.007193243131041527, + 0.020717019215226173, + 0.009894203394651413, + 0.0050490195862948895, + 0.019714420661330223, + 0.7588523626327515, + 0.02943417616188526 + ], + "credit_map": [ + 0.0024794461205601692, + 0.0012623125221580267, + 0.0021606534719467163, + 0.0023278261069208384, + 0.0017850750591605902, + 0.006161550525575876, + 0.0014865724369883537, + 0.0033254215959459543, + 0.0010222193086519837, + 0.0009707525023259223, + 0.0019481508061289787, + 0.0011806056136265397, + 0.0009936044225469232, + 0.0017282854532822967, + 0.0009974937420338392, + 0.00234790425747633, + 0.01419851090759039, + 0.02949829213321209, + 0.02056865021586418, + 0.011311309412121773, + 0.0073616462759673595, + 0.013075302354991436, + 0.004210359882563353, + 0.016743630170822144, + 0.000772444938775152, + 0.0010491443099454045, + 0.0008093694341368973, + 0.0010741893202066422, + 0.000941791629884392, + 0.0007296256953850389, + 0.0009586832602508366, + 0.0008579944842495024, + 0.0015765136340633035, + 0.0017892172327265143, + 0.0023549676407128572, + 0.0016872142441570759, + 0.0021172550041228533, + 0.004594095051288605, + 0.001806652988307178, + 0.004791103303432465, + 0.001494415570050478, + 0.0013415581779554486, + 0.0014223393518477678, + 0.0021716621704399586, + 0.000646480533760041, + 0.0011043779086321592, + 0.0005522485007531941, + 0.0011611214140430093, + 0.0004761068557854742, + 0.0005867583095096052, + 0.0009672287269495428, + 0.0006127716042101383, + 0.0007264956366270781, + 0.0005175594124011695, + 0.0004398221499286592, + 0.0007222770364023745, + 0.0021218357142060995, + 0.0025320779532194138, + 0.0032523584086447954, + 0.002545046154409647, + 0.0019342158921062946, + 0.0037254055496305227, + 0.0009752234909683466, + 0.002628257730975747, + 0.10619649291038513, + 0.023362623527646065, + 0.2469659447669983, + 0.04126368835568428, + 0.027150386944413185, + 0.06972149014472961, + 0.019178681075572968, + 0.22501301765441895, + 0.005757488775998354, + 0.0018284253310412169, + 0.007509802468121052, + 0.0029819596093147993, + 0.0017063150880858302, + 0.0029968437738716602, + 0.004191502463072538, + 0.0024618373718112707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.004636015277355909 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_15/index_tensor([3450, 3450])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_15/index_tensor([3450, 3450])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..633967e636c5e41ed4ecc1a370029d57cf7d802e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_15/index_tensor([3450, 3450])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot shot from a medium, slightly upward angle, the camera pans steadily to the right, closely following the subject's movements. In the frame, a man in his thirties, with short, dark brown hair and a neatly trimmed beard, leans slightly forward, his expression focused and serious. He wears a dark gray, cotton, long-sleeved Henley-collared T-shirt, the sleeves casually pushed up to his elbows, revealing his muscular forearms. He is in a softly lit bedroom, with the corner of a bed covered in white sheets visible in the background. Steadily and deliberately, the man extends his right arm toward a blue nylon travel bag on the bed. The bag features black canvas handles and a silver metal zipper, half-unzipped. As the camera follows, the thick wad of banknotes clutched in his left hand becomes visible. He then carefully places the wad of bills into the bag's open mouth, his fingers delicately moving as he pushes the money deeper into the bag. The natural morning light filtering in from a side window casts soft shadows on his arm and the surface of the travel bag, highlighting the texture of the nylon and the thickness of the banknotes. The dominant color scheme of the entire scene is the vibrant blue of the travel bag, the dark gray of the man's clothing, and the white of the bed sheet, creating an atmosphere of focus, intimacy, and a touch of tension.", + "seed": 7518, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_15/index_3450_seed_7518_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08047509714961051, + "vlm_logprob": -0.08047509714961051, + "vlm_yes_no_margin": 3.0249999523162843, + "vlm_token_credit": -0.08047509714961051, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12851113080978394 + ], + "opposite_logprob_tokens": [ + -2.1285111904144287 + ], + "logprob_reward": -0.12851113080978394, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12851113080978394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14838457107543945, + 0.09198382496833801, + 0.24753348529338837, + 0.04824976623058319, + 0.09283062815666199, + 0.044807758182287216, + 0.1864812821149826, + 0.10657325387001038, + 0.02277890406548977, + 0.01037651114165783 + ], + "credit_map": [ + 0.018577300012111664, + 0.021881915628910065, + 0.01662633754312992, + 0.014187425374984741, + 0.025652604177594185, + 0.02115447446703911, + 0.015574031509459019, + 0.0147304842248559, + 0.015833526849746704, + 0.014056031592190266, + 0.013091908767819405, + 0.01190956961363554, + 0.012056010775268078, + 0.007185249123722315, + 0.009931348264217377, + 0.00792018510401249, + 0.07020237296819687, + 0.01835015043616295, + 0.029544316232204437, + 0.010247480124235153, + 0.014213806949555874, + 0.08994375169277191, + 0.006340191699564457, + 0.008691417053341866, + 0.006578413303941488, + 0.006272679660469294, + 0.004678360652178526, + 0.006236742250621319, + 0.004621096886694431, + 0.00707053579390049, + 0.00707970978692174, + 0.005712228827178478, + 0.012796874158084393, + 0.03288822993636131, + 0.008246283046901226, + 0.004746200516819954, + 0.0059537203051149845, + 0.020337944850325584, + 0.0033571787644177675, + 0.004504194017499685, + 0.006100213620811701, + 0.004110083449631929, + 0.00814293697476387, + 0.005629379767924547, + 0.006658521015197039, + 0.0036405115388333797, + 0.005945787765085697, + 0.004580322653055191, + 0.0216069333255291, + 0.015213532373309135, + 0.04508113116025925, + 0.04180166870355606, + 0.014428677968680859, + 0.012335354462265968, + 0.02079986035823822, + 0.015214122831821442, + 0.012385745532810688, + 0.010337398387491703, + 0.014193315990269184, + 0.01299351267516613, + 0.01471047941595316, + 0.00976724736392498, + 0.015498948283493519, + 0.01668659970164299, + 0.0028357699047774076, + 0.003192745614796877, + 0.0028047494124621153, + 0.004017142578959465, + 0.002570966724306345, + 0.001888162805698812, + 0.002604932524263859, + 0.0028644362464547157, + 0.0010360992746427655, + 0.0008327445248141885, + 0.0011567495530471206, + 0.0021344125270843506, + 0.0007510359864681959, + 0.0009608916589058936, + 0.0021851619239896536, + 0.0013194158673286438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.12851113080978394 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306705296039581 + ], + "opposite_logprob_tokens": [ + -3.16806697845459 + ], + "logprob_reward": -0.04306705296039581, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306705296039581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10680300742387772, + 0.2451072633266449, + 0.12458048015832901, + 0.0752282589673996, + 0.25232064723968506, + 0.021854370832443237, + 0.039988525211811066, + 0.08214277774095535, + 0.02415040135383606, + 0.02782423608005047 + ], + "credit_map": [ + 0.008411495015025139, + 0.014411168172955513, + 0.011279461905360222, + 0.01828143745660782, + 0.015942001715302467, + 0.008223612792789936, + 0.017645619809627533, + 0.01260821707546711, + 0.0257995817810297, + 0.0443706288933754, + 0.03500756993889809, + 0.04876071214675903, + 0.022576145827770233, + 0.020749041810631752, + 0.028084075078368187, + 0.019759517163038254, + 0.01910620741546154, + 0.022178662940859795, + 0.020287031307816505, + 0.011085248552262783, + 0.015116129070520401, + 0.017658891156315804, + 0.010786118917167187, + 0.00836220383644104, + 0.010587251745164394, + 0.01182290818542242, + 0.008967293426394463, + 0.009221591055393219, + 0.007935194298624992, + 0.009719791822135448, + 0.009544654749333858, + 0.0074295843951404095, + 0.037247925996780396, + 0.09417262673377991, + 0.012783505953848362, + 0.0050236377865076065, + 0.01710555888712406, + 0.07201707363128662, + 0.006058415863662958, + 0.00791194848716259, + 0.002590445103123784, + 0.0029569845646619797, + 0.0035835132002830505, + 0.0023030932061374187, + 0.003467038506641984, + 0.0017640257719904184, + 0.0029104677960276604, + 0.0022788047790527344, + 0.0050834291614592075, + 0.00406463211402297, + 0.006648743990808725, + 0.0064767408184707165, + 0.00666377367451787, + 0.0034826258197426796, + 0.004604186862707138, + 0.0029643967282027006, + 0.007525020744651556, + 0.007994681596755981, + 0.004017670173197985, + 0.0071998354978859425, + 0.020181871950626373, + 0.011447148397564888, + 0.017519472166895866, + 0.006257084663957357, + 0.0036953354720026255, + 0.0033650253899395466, + 0.0035351854749023914, + 0.003473317949101329, + 0.0027314191684126854, + 0.0021651394199579954, + 0.0023281993344426155, + 0.002856781706213951, + 0.002211166312918067, + 0.00279164663515985, + 0.003530077636241913, + 0.005135910119861364, + 0.0028525416273623705, + 0.002981825964525342, + 0.005120810586959124, + 0.003200261387974024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.04306705296039581 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078160256147385 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.011078160256147385, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078160256147385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02023431472480297, + 0.037662606686353683, + 0.22540198266506195, + 0.0171560849994421, + 0.33530208468437195, + 0.03494524955749512, + 0.14059121906757355, + 0.11437328159809113, + 0.048287540674209595, + 0.026045572012662888 + ], + "credit_map": [ + 0.002082687569782138, + 0.0021784498821944, + 0.0023890682496130466, + 0.0023515038192272186, + 0.004512790124863386, + 0.0028317105025053024, + 0.002043378073722124, + 0.0018447263864800334, + 0.0031545478850603104, + 0.0052202926017344, + 0.004880666732788086, + 0.005312291905283928, + 0.00486597279086709, + 0.004976192489266396, + 0.005284999962896109, + 0.0039676399901509285, + 0.06599782407283783, + 0.015096142888069153, + 0.027095794677734375, + 0.009064329788088799, + 0.011847829446196556, + 0.08353160321712494, + 0.004954237025231123, + 0.007814229466021061, + 0.0024326255079358816, + 0.002389325760304928, + 0.0017695670248940587, + 0.002025608904659748, + 0.0018869108753278852, + 0.002274987520650029, + 0.0023575033992528915, + 0.0020195546094328165, + 0.049047376960515976, + 0.12132149934768677, + 0.018611010164022446, + 0.0072552901692688465, + 0.023109814152121544, + 0.09797310084104538, + 0.007418578024953604, + 0.01056542620062828, + 0.004699971526861191, + 0.003468311158940196, + 0.0064689842984080315, + 0.004526429809629917, + 0.004585780669003725, + 0.002967079635709524, + 0.004394156858325005, + 0.0038345346692949533, + 0.017154911532998085, + 0.010953233577311039, + 0.03541428968310356, + 0.03305060416460037, + 0.009643478319048882, + 0.009320135228335857, + 0.013958097435534, + 0.011096474714577198, + 0.010957406833767891, + 0.01187647134065628, + 0.014063668437302113, + 0.01509025041013956, + 0.010510927066206932, + 0.011392579413950443, + 0.023834925144910812, + 0.016647055745124817, + 0.007098707836121321, + 0.005739402025938034, + 0.004974744748324156, + 0.008203214965760708, + 0.006597928237169981, + 0.004074051510542631, + 0.006268261931836605, + 0.0053312270902097225, + 0.0034987174440175295, + 0.0028889132663607597, + 0.004575216677039862, + 0.004017742350697517, + 0.0020362387876957655, + 0.002381744561716914, + 0.003998002968728542, + 0.0026489966548979282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.011078160256147385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186718225479126 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186718225479126, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018186718225479126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05125981569290161, + 0.026151975616812706, + 0.11519733816385269, + 0.03632054105401039, + 0.38680005073547363, + 0.0426311232149601, + 0.17949256300926208, + 0.0863337442278862, + 0.018583128228783607, + 0.05722975358366966 + ], + "credit_map": [ + 0.0041410718113183975, + 0.006088413763791323, + 0.006122359074652195, + 0.0044599659740924835, + 0.009851442649960518, + 0.009579724632203579, + 0.004755885805934668, + 0.006260955240577459, + 0.0028277416713535786, + 0.003527230117470026, + 0.003567259991541505, + 0.0032369280233979225, + 0.00365716265514493, + 0.0016480379272252321, + 0.0048211440443992615, + 0.002866472350433469, + 0.03338826820254326, + 0.007844801992177963, + 0.013765847310423851, + 0.004877122119069099, + 0.006171993911266327, + 0.0420234389603138, + 0.002874432597309351, + 0.004251426085829735, + 0.00527356518432498, + 0.005519495811313391, + 0.005007356870919466, + 0.004512755665928125, + 0.0035483064129948616, + 0.0037784301675856113, + 0.003643370233476162, + 0.005037257447838783, + 0.0552687793970108, + 0.14384815096855164, + 0.02177286520600319, + 0.010485993698239326, + 0.02487403154373169, + 0.10637188702821732, + 0.010803839191794395, + 0.013374517671763897, + 0.005329984240233898, + 0.005124097224324942, + 0.007306524086743593, + 0.004869592841714621, + 0.008498231880366802, + 0.0032473811879754066, + 0.005042518023401499, + 0.0032127939630299807, + 0.02101237140595913, + 0.01333752740174532, + 0.04409933462738991, + 0.04244847223162651, + 0.014646017923951149, + 0.012057718820869923, + 0.018377868458628654, + 0.013513253070414066, + 0.008804341778159142, + 0.007493701763451099, + 0.011100325733423233, + 0.010869041085243225, + 0.008165229111909866, + 0.007807211950421333, + 0.017610877752304077, + 0.014483018778264523, + 0.0026619855780154467, + 0.0024126118514686823, + 0.001756916637532413, + 0.002864585490897298, + 0.0029669038485735655, + 0.0017180440481752157, + 0.002523651346564293, + 0.0016784283798187971, + 0.005920782685279846, + 0.0061425077728927135, + 0.010725016705691814, + 0.010279896669089794, + 0.005945500917732716, + 0.004698952194303274, + 0.007363006006926298, + 0.006154096685349941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.018186718225479126 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153242349624634 + ], + "opposite_logprob_tokens": [ + -1.7015323638916016 + ], + "logprob_reward": -0.20153242349624634, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153242349624634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033189259469509125, + 0.024563884362578392, + 0.17821718752384186, + 0.037702493369579315, + 0.2782335877418518, + 0.050595566630363464, + 0.24485592544078827, + 0.12283696234226227, + 0.028140682727098465, + 0.0016644224524497986 + ], + "credit_map": [ + 0.003726978786289692, + 0.004091371316462755, + 0.0036513335071504116, + 0.004158799536526203, + 0.006292167119681835, + 0.005317697301506996, + 0.0031795103568583727, + 0.002771399449557066, + 0.0023985635489225388, + 0.004032894503325224, + 0.003644460579380393, + 0.0031701659318059683, + 0.0025525109376758337, + 0.002640082035213709, + 0.003184216795489192, + 0.00294098979793489, + 0.05185221508145332, + 0.011819886043667793, + 0.021366644650697708, + 0.0083124078810215, + 0.00881622452288866, + 0.06493894010782242, + 0.004837679211050272, + 0.006273192353546619, + 0.005920280236750841, + 0.005078397691249847, + 0.00403921352699399, + 0.004870842210948467, + 0.0031544792000204325, + 0.005052525550127029, + 0.005866209976375103, + 0.0037205442786216736, + 0.03935510292649269, + 0.10322725772857666, + 0.019065212458372116, + 0.010877644643187523, + 0.017124686390161514, + 0.07052791118621826, + 0.008184539154171944, + 0.009871230460703373, + 0.006567041855305433, + 0.005042396020144224, + 0.009355681017041206, + 0.006598787847906351, + 0.006943955086171627, + 0.003959505818784237, + 0.0070495703257620335, + 0.005078628193587065, + 0.029024917632341385, + 0.01808345504105091, + 0.06213109940290451, + 0.05892639979720116, + 0.01760253496468067, + 0.016178030520677567, + 0.024581890553236008, + 0.018327610567212105, + 0.01212664507329464, + 0.011697900481522083, + 0.016207393258810043, + 0.014910205267369747, + 0.014781336300075054, + 0.011207569390535355, + 0.021797653287649155, + 0.020108262076973915, + 0.0043898895382881165, + 0.003185402136296034, + 0.0029156720265746117, + 0.004759159870445728, + 0.0040111481212079525, + 0.0023931246250867844, + 0.003677844302728772, + 0.0028084435034543276, + 0.0002581056032795459, + 0.00020329128892626613, + 0.0001722820452414453, + 0.000259932508924976, + 0.00014784153609070927, + 0.000130877218907699, + 0.0003244292747694999, + 0.00016766294720582664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.20153242349624634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from a medium, slightly upward angle, the camera pans steadily to the right, closely following the subject's movements. In the frame, a man in his thirties, with short, dark brown hair and a neatly trimmed beard, leans slightly forward, his expression focused and serious. He wears a dark gray, cotton, long-sleeved Henley-collared T-shirt, the sleeves casually pushed up to his elbows, revealing his muscular forearms. He is in a softly lit bedroom, with the corner of a bed covered in white sheets visible in the background. Steadily and deliberately, the man extends his right arm toward a blue nylon travel bag on the bed. The bag features black canvas handles and a silver metal zipper, half-unzipped. As the camera follows, the thick wad of banknotes clutched in his left hand becomes visible. He then carefully places the wad of bills into the bag's open mouth, his fingers delicately moving as he pushes the money deeper into the bag. The natural morning light filtering in from a side window casts soft shadows on his arm and the surface of the travel bag, highlighting the texture of the nylon and the thickness of the banknotes. The dominant color scheme of the entire scene is the vibrant blue of the travel bag, the dark gray of the man's clothing, and the white of the bed sheet, creating an atmosphere of focus, intimacy, and a touch of tension.", + "seed": 7519, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_15/index_3450_seed_7519_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06839551301673055, + "vlm_logprob": -0.06839551301673055, + "vlm_yes_no_margin": 3.4999999523162844, + "vlm_token_credit": -0.06839551301673055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045885372906923294 + ], + "opposite_logprob_tokens": [ + -3.1708853244781494 + ], + "logprob_reward": -0.045885372906923294, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.045885372906923294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2936495244503021, + 0.09539592266082764, + 0.056960225105285645, + 0.06508466601371765, + 0.09050015360116959, + 0.1089935302734375, + 0.06975805759429932, + 0.10240719467401505, + 0.05307135358452797, + 0.0641794204711914 + ], + "credit_map": [ + 0.04536256939172745, + 0.026883982121944427, + 0.08018475770950317, + 0.036390434950590134, + 0.039443787187337875, + 0.023522907868027687, + 0.021977394819259644, + 0.019883684813976288, + 0.013341101817786694, + 0.010519279167056084, + 0.019245797768235207, + 0.010646505281329155, + 0.01323098223656416, + 0.010115290060639381, + 0.009594312869012356, + 0.00870264321565628, + 0.00807100161910057, + 0.007192245218902826, + 0.013084874488413334, + 0.007238800171762705, + 0.006249839439988136, + 0.004433914553374052, + 0.006026044953614473, + 0.004663498140871525, + 0.0085393525660038, + 0.009108435362577438, + 0.010650111362338066, + 0.00754226790741086, + 0.009762330912053585, + 0.00770545331761241, + 0.006260301452130079, + 0.005516410805284977, + 0.006883467547595501, + 0.013516456820070744, + 0.011187896132469177, + 0.013833122327923775, + 0.01370391808450222, + 0.008669869042932987, + 0.013861033134162426, + 0.008844389580190182, + 0.018677417188882828, + 0.016510851681232452, + 0.023341748863458633, + 0.011938586831092834, + 0.011972058564424515, + 0.009366214275360107, + 0.009613888338208199, + 0.007572747766971588, + 0.013092519715428352, + 0.0071737756952643394, + 0.011337117291986942, + 0.008029588498175144, + 0.008970675058662891, + 0.004303009249269962, + 0.008732660673558712, + 0.008118708617985249, + 0.020459435880184174, + 0.01205631997436285, + 0.011256703175604343, + 0.01105036586523056, + 0.013259805738925934, + 0.009140408597886562, + 0.012527118436992168, + 0.012657027691602707, + 0.006735627073794603, + 0.009639696218073368, + 0.005707908887416124, + 0.005309587344527245, + 0.009094706736505032, + 0.006869805045425892, + 0.004275243729352951, + 0.005438776686787605, + 0.010758507065474987, + 0.008801378309726715, + 0.007094801403582096, + 0.006138366647064686, + 0.008097403682768345, + 0.006186492741107941, + 0.008464627899229527, + 0.008637838996946812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.045885372906923294 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201585859060287 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201585859060287, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018201585859060287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23352305591106415, + 0.24116557836532593, + 0.13811995089054108, + 0.05544747784733772, + 0.039023905992507935, + 0.040704090148210526, + 0.06705190241336823, + 0.09644144773483276, + 0.031355101615190506, + 0.057167477905750275 + ], + "credit_map": [ + 0.01714654639363289, + 0.036274105310440063, + 0.03167026862502098, + 0.035617031157016754, + 0.03602635860443115, + 0.022565320134162903, + 0.029872404411435127, + 0.024351028725504875, + 0.028244467452168465, + 0.026451678946614265, + 0.03890382498502731, + 0.029680214822292328, + 0.03679775819182396, + 0.027627019211649895, + 0.030044779181480408, + 0.02341584302484989, + 0.01196566317230463, + 0.025959983468055725, + 0.017165005207061768, + 0.02292470820248127, + 0.019515784457325935, + 0.012919520027935505, + 0.020076967775821686, + 0.007592321839183569, + 0.004529587924480438, + 0.007479759864509106, + 0.009268405847251415, + 0.007690207101404667, + 0.009321660734713078, + 0.005050174426287413, + 0.007031744811683893, + 0.005075937602669001, + 0.004321829881519079, + 0.005595531780272722, + 0.004805764649063349, + 0.006099598947912455, + 0.0046125780791044235, + 0.0036335524637252092, + 0.004424291662871838, + 0.005530759692192078, + 0.005403339862823486, + 0.006288886070251465, + 0.009866074658930302, + 0.005025128368288279, + 0.005023288074880838, + 0.003962484188377857, + 0.002817396307364106, + 0.0023174919188022614, + 0.011159863322973251, + 0.005786214955151081, + 0.011613759212195873, + 0.00843879021704197, + 0.007751579862087965, + 0.004143659956753254, + 0.010192385874688625, + 0.00796564668416977, + 0.018399346619844437, + 0.016541430726647377, + 0.010660210624337196, + 0.008842662908136845, + 0.014580759219825268, + 0.008883810602128506, + 0.009601638652384281, + 0.008931586518883705, + 0.003678030101582408, + 0.004767566919326782, + 0.0037093046121299267, + 0.0029588055331259966, + 0.004775869660079479, + 0.003851387882605195, + 0.0036167714279145002, + 0.003997363615781069, + 0.014898569323122501, + 0.008721655234694481, + 0.00762805063277483, + 0.005339544732123613, + 0.005500298459082842, + 0.005113594233989716, + 0.004465240985155106, + 0.005500528961420059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.018201585859060287 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255692172795534 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255692172795534, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255692172795534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05205565318465233, + 0.059334296733140945, + 0.04986915364861488, + 0.055480796843767166, + 0.20527705550193787, + 0.12306172400712967, + 0.0724719986319542, + 0.08145242929458618, + 0.11458633840084076, + 0.1864105463027954 + ], + "credit_map": [ + 0.006012687925249338, + 0.008335156366229057, + 0.006455859635025263, + 0.007558577228337526, + 0.006507378537207842, + 0.00555576803162694, + 0.005866014864295721, + 0.005764208268374205, + 0.008892456069588661, + 0.004641763400286436, + 0.013726428151130676, + 0.007093670312315226, + 0.007871657609939575, + 0.0056519960053265095, + 0.006260982248932123, + 0.005195339675992727, + 0.0044964211992919445, + 0.004661001730710268, + 0.006556757725775242, + 0.007530570961534977, + 0.0065900967456400394, + 0.006754877045750618, + 0.006675076670944691, + 0.006604352034628391, + 0.00542021356523037, + 0.009657606482505798, + 0.010137947276234627, + 0.005804162006825209, + 0.008909911848604679, + 0.0042287372052669525, + 0.005208526737987995, + 0.006113693583756685, + 0.02551089972257614, + 0.03994935005903244, + 0.031132472679018974, + 0.04422299936413765, + 0.016311608254909515, + 0.017429927363991737, + 0.019194580614566803, + 0.011525215581059456, + 0.023197967559099197, + 0.02355055697262287, + 0.01332769077271223, + 0.01922118104994297, + 0.016977433115243912, + 0.006327469367533922, + 0.012164651416242123, + 0.008294782601296902, + 0.01200211513787508, + 0.007465482223778963, + 0.013055695220828056, + 0.00971629936248064, + 0.009134914726018906, + 0.004648956470191479, + 0.009110508486628532, + 0.007338027935475111, + 0.01604076288640499, + 0.008656278252601624, + 0.011091936379671097, + 0.010481341741979122, + 0.010639197193086147, + 0.006156580988317728, + 0.011266499757766724, + 0.007119831629097462, + 0.015299909748136997, + 0.014849691651761532, + 0.016386263072490692, + 0.016709228977560997, + 0.013615118339657784, + 0.011401617899537086, + 0.013080132193863392, + 0.01324437651783228, + 0.04041889309883118, + 0.03199018910527229, + 0.02002515085041523, + 0.019144859164953232, + 0.01802874729037285, + 0.0191165953874588, + 0.02166660875082016, + 0.016019504517316818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.005255692172795534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579388365149498 + ], + "opposite_logprob_tokens": [ + -3.8955793380737305 + ], + "logprob_reward": -0.020579388365149498, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579388365149498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.151786670088768, + 0.05008532851934433, + 0.08048877120018005, + 0.04568808898329735, + 0.12348518520593643, + 0.09744693338871002, + 0.15724654495716095, + 0.09272933006286621, + 0.03638385236263275, + 0.16465921700000763 + ], + "credit_map": [ + 0.026363365352153778, + 0.012241206131875515, + 0.046456001698970795, + 0.017237212508916855, + 0.013126778416335583, + 0.01282611396163702, + 0.010800319723784924, + 0.012735672295093536, + 0.00636462913826108, + 0.005440150387585163, + 0.010015694424510002, + 0.006136204116046429, + 0.007210664451122284, + 0.0057128374464809895, + 0.0050091189332306385, + 0.004196029156446457, + 0.007626511622220278, + 0.006564427632838488, + 0.009340084157884121, + 0.007894242182374, + 0.010239923372864723, + 0.011939167976379395, + 0.012019244953989983, + 0.014865168370306492, + 0.0035705070476979017, + 0.006189275532960892, + 0.005173972342163324, + 0.00796839501708746, + 0.007572033442556858, + 0.003838608507066965, + 0.006184510886669159, + 0.005190789233893156, + 0.02013561502099037, + 0.018428491428494453, + 0.013735619373619556, + 0.019607551395893097, + 0.010368035174906254, + 0.010406546294689178, + 0.017217181622982025, + 0.01358614768832922, + 0.01569705456495285, + 0.013233581557869911, + 0.013177992776036263, + 0.012011601589620113, + 0.01850016787648201, + 0.006797900889068842, + 0.010618828237056732, + 0.007409802637994289, + 0.026856090873479843, + 0.025577859953045845, + 0.029773669317364693, + 0.0161039587110281, + 0.021855879575014114, + 0.010103359818458557, + 0.016731953248381615, + 0.010243775323033333, + 0.013209331780672073, + 0.011678818613290787, + 0.013514972291886806, + 0.010868767276406288, + 0.015391078777611256, + 0.007621195632964373, + 0.009877204895019531, + 0.01056796032935381, + 0.004296358209103346, + 0.003983973525464535, + 0.004643736872822046, + 0.0060969446785748005, + 0.003634016029536724, + 0.0031905057840049267, + 0.005406668409705162, + 0.005131646059453487, + 0.018381783738732338, + 0.017351394519209862, + 0.016214944422245026, + 0.027889860793948174, + 0.016907118260860443, + 0.01526393461972475, + 0.034490518271923065, + 0.018159661442041397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.020579388365149498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520555257797241 + ], + "opposite_logprob_tokens": [ + -1.5020555257797241 + ], + "logprob_reward": -0.2520555257797241, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520555257797241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09448523819446564, + 0.06308135390281677, + 0.061740193516016006, + 0.13211515545845032, + 0.3293224573135376, + 0.07907459139823914, + 0.07906028628349304, + 0.059689782559871674, + 0.033770062029361725, + 0.06766090542078018 + ], + "credit_map": [ + 0.01643700897693634, + 0.0068640629760921, + 0.0289097111672163, + 0.010357608087360859, + 0.009433142840862274, + 0.0067654759623110294, + 0.008838247507810593, + 0.006879983935505152, + 0.007896369323134422, + 0.004338424652814865, + 0.012753316201269627, + 0.008083399385213852, + 0.00648159347474575, + 0.009563499130308628, + 0.006860405672341585, + 0.007104339078068733, + 0.0077937887981534, + 0.005716534797102213, + 0.011642586439847946, + 0.006813390180468559, + 0.00878528319299221, + 0.0067943399772048, + 0.006689906120300293, + 0.00750436307862401, + 0.011682376265525818, + 0.019162027165293694, + 0.017881575971841812, + 0.023346418514847755, + 0.017212888225913048, + 0.014369556680321693, + 0.017006123438477516, + 0.011454182676970959, + 0.0370158776640892, + 0.05904349684715271, + 0.049087509512901306, + 0.05810707062482834, + 0.03531326726078987, + 0.03823689743876457, + 0.03189752250909805, + 0.020620808005332947, + 0.015690358355641365, + 0.015576737001538277, + 0.014943137764930725, + 0.008335739374160767, + 0.010386127978563309, + 0.004650097340345383, + 0.005827605724334717, + 0.003664785996079445, + 0.012893514707684517, + 0.00931965745985508, + 0.012717359699308872, + 0.0100360456854105, + 0.010677584446966648, + 0.004686631727963686, + 0.009891893714666367, + 0.008837598375976086, + 0.007878579199314117, + 0.009010731242597103, + 0.007057339418679476, + 0.0068504163064062595, + 0.010820063762366772, + 0.005220036953687668, + 0.0063282945193350315, + 0.006524319294840097, + 0.004681414924561977, + 0.006103905383497477, + 0.004034934565424919, + 0.0038358208257704973, + 0.00486072339117527, + 0.0035439697094261646, + 0.0026899559888988733, + 0.0040193344466388226, + 0.006862437818199396, + 0.011388684622943401, + 0.007699029054492712, + 0.006631686817854643, + 0.008251694962382317, + 0.010579765774309635, + 0.00781953614205122, + 0.00842807162553072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.2520555257797241 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from a medium, slightly upward angle, the camera pans steadily to the right, closely following the subject's movements. In the frame, a man in his thirties, with short, dark brown hair and a neatly trimmed beard, leans slightly forward, his expression focused and serious. He wears a dark gray, cotton, long-sleeved Henley-collared T-shirt, the sleeves casually pushed up to his elbows, revealing his muscular forearms. He is in a softly lit bedroom, with the corner of a bed covered in white sheets visible in the background. Steadily and deliberately, the man extends his right arm toward a blue nylon travel bag on the bed. The bag features black canvas handles and a silver metal zipper, half-unzipped. As the camera follows, the thick wad of banknotes clutched in his left hand becomes visible. He then carefully places the wad of bills into the bag's open mouth, his fingers delicately moving as he pushes the money deeper into the bag. The natural morning light filtering in from a side window casts soft shadows on his arm and the surface of the travel bag, highlighting the texture of the nylon and the thickness of the banknotes. The dominant color scheme of the entire scene is the vibrant blue of the travel bag, the dark gray of the man's clothing, and the white of the bed sheet, creating an atmosphere of focus, intimacy, and a touch of tension.", + "seed": 7520, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_15/index_3450_seed_7520_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07338960487395525, + "vlm_logprob": -0.07338960487395525, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.07338960487395525, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05688101425766945 + ], + "opposite_logprob_tokens": [ + -2.9318809509277344 + ], + "logprob_reward": -0.05688101425766945, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05688101425766945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2499939501285553, + 0.10191207379102707, + 0.1233905702829361, + 0.026700707152485847, + 0.04446686431765556, + 0.03926347941160202, + 0.24843646585941315, + 0.06950724869966507, + 0.024583781138062477, + 0.07174484431743622 + ], + "credit_map": [ + 0.02900974452495575, + 0.03626849129796028, + 0.04326384887099266, + 0.031107982620596886, + 0.031527161598205566, + 0.03030383214354515, + 0.0223079863935709, + 0.026204904541373253, + 0.013302912935614586, + 0.01108100451529026, + 0.012481399811804295, + 0.019724711775779724, + 0.012962570413947105, + 0.010283616371452808, + 0.012218400835990906, + 0.009857464581727982, + 0.008921005763113499, + 0.00878722034394741, + 0.009153912775218487, + 0.010219256393611431, + 0.013924367725849152, + 0.021390479058027267, + 0.015663139522075653, + 0.03533119708299637, + 0.0030873888172209263, + 0.003263078397139907, + 0.003816475858911872, + 0.004274371545761824, + 0.0033588288351893425, + 0.003201274899765849, + 0.0027698928024619818, + 0.0029293959960341454, + 0.007151521276682615, + 0.004420426674187183, + 0.00496821291744709, + 0.004147639498114586, + 0.005451277829706669, + 0.004145066253840923, + 0.00771478982642293, + 0.006467931438237429, + 0.005243892315775156, + 0.0026865042746067047, + 0.003860105760395527, + 0.00441593024879694, + 0.0041490220464766026, + 0.003799864323809743, + 0.00909113883972168, + 0.006017021834850311, + 0.018091434612870216, + 0.008630557917058468, + 0.009957092814147472, + 0.005723186302930117, + 0.027161940932273865, + 0.10629074275493622, + 0.009848452173173428, + 0.06273306161165237, + 0.013935606926679611, + 0.00968344695866108, + 0.01051891129463911, + 0.006430007982999086, + 0.009650865569710732, + 0.007405290845781565, + 0.004802651237696409, + 0.007080467883497477, + 0.0026772471610456705, + 0.0033499097917228937, + 0.002962082391604781, + 0.002317391335964203, + 0.005574652925133705, + 0.002641409868374467, + 0.0028611866291612387, + 0.002199901267886162, + 0.008002719841897488, + 0.01751849241554737, + 0.006594942882657051, + 0.005806056782603264, + 0.0162041075527668, + 0.008243000134825706, + 0.004497581161558628, + 0.004877946339547634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.05688101425766945 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586511120200157 + ], + "opposite_logprob_tokens": [ + -3.8955864906311035 + ], + "logprob_reward": -0.020586511120200157, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586511120200157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17830847203731537, + 0.4337652623653412, + 0.06305259466171265, + 0.031863175332546234, + 0.037013452500104904, + 0.029718715697526932, + 0.15355801582336426, + 0.030201006680727005, + 0.017315059900283813, + 0.025204211473464966 + ], + "credit_map": [ + 0.03099052794277668, + 0.025120075792074203, + 0.02566749043762684, + 0.02095850743353367, + 0.022736385464668274, + 0.019153665751218796, + 0.01926058903336525, + 0.014421249739825726, + 0.021228648722171783, + 0.03675004094839096, + 0.021129632368683815, + 0.015375790186226368, + 0.1853111982345581, + 0.1056164875626564, + 0.020206468179821968, + 0.028146998956799507, + 0.006932110991328955, + 0.00652391416952014, + 0.006570411846041679, + 0.005440304521471262, + 0.009021420031785965, + 0.009205112233757973, + 0.007532794028520584, + 0.011826526373624802, + 0.00464582210406661, + 0.0033699474297463894, + 0.005070739425718784, + 0.0036479460541158915, + 0.004371687304228544, + 0.003904851619154215, + 0.003302978118881583, + 0.0035492065362632275, + 0.004224357660859823, + 0.003359678201377392, + 0.004461821634322405, + 0.004772789776325226, + 0.0035755878780037165, + 0.004346213303506374, + 0.0063073341734707355, + 0.0059656728990375996, + 0.005364594049751759, + 0.0018753150943666697, + 0.00338757480494678, + 0.0029970454052090645, + 0.0033719325438141823, + 0.0028342145960778, + 0.005674399435520172, + 0.00421364139765501, + 0.010611938312649727, + 0.006280426401644945, + 0.006268101278692484, + 0.003198196180164814, + 0.01695689558982849, + 0.06548969447612762, + 0.006264256313443184, + 0.03848849982023239, + 0.003767880378291011, + 0.004190483596175909, + 0.003284023841843009, + 0.0030313057359308004, + 0.00600390275940299, + 0.002885544439777732, + 0.002814115257933736, + 0.004223750904202461, + 0.0021203206852078438, + 0.0023056473582983017, + 0.0025725020095705986, + 0.0015522580360993743, + 0.003781635081395507, + 0.001699632266536355, + 0.001419946551322937, + 0.0018631182610988617, + 0.0030698352493345737, + 0.0032674246467649937, + 0.003221618477255106, + 0.0033112489618360996, + 0.00311283185146749, + 0.0024107457138597965, + 0.003254086244851351, + 0.0035564193967729807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.020586511120200157 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629777304828167 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629777304828167, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629777304828167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02646235004067421, + 0.020335093140602112, + 0.02475505694746971, + 0.02259066142141819, + 0.019198430702090263, + 0.021211005747318268, + 0.7972636818885803, + 0.017275532707571983, + 0.03290541470050812, + 0.01800268143415451 + ], + "credit_map": [ + 0.00422364566475153, + 0.0033361490350216627, + 0.004495477769523859, + 0.0023746925871819258, + 0.003232206217944622, + 0.0026824052911251783, + 0.002582928165793419, + 0.0035348469391465187, + 0.0031568880658596754, + 0.0016942594666033983, + 0.00201837788335979, + 0.0025087238755077124, + 0.0045712790451943874, + 0.0022952603176236153, + 0.0018225767416879535, + 0.0022677297238260508, + 0.0028802878223359585, + 0.002239648485556245, + 0.0024439224507659674, + 0.0023296549916267395, + 0.006157726980745792, + 0.003986764699220657, + 0.0025823512114584446, + 0.0021347019355744123, + 0.0017309427494183183, + 0.00324916560202837, + 0.002655932679772377, + 0.0024528794456273317, + 0.0037582421209663153, + 0.0029072852339595556, + 0.0033805735874921083, + 0.0024556422140449286, + 0.002381111728027463, + 0.001645931275561452, + 0.002281315391883254, + 0.0026896893978118896, + 0.002180175855755806, + 0.0019361765589565039, + 0.0033329916186630726, + 0.0027510402724146843, + 0.0030598109588027, + 0.0028298860415816307, + 0.002230573445558548, + 0.0016581732779741287, + 0.0034277199301868677, + 0.002378898672759533, + 0.00231815199367702, + 0.003307791193947196, + 0.05739341676235199, + 0.028008121997117996, + 0.030518095940351486, + 0.012724203057587147, + 0.08766006678342819, + 0.3472737967967987, + 0.029623968526721, + 0.2040620893239975, + 0.0014382506487891078, + 0.0020493739284574986, + 0.0016828968655318022, + 0.0018163180211558938, + 0.0033813335467129946, + 0.0020315006840974092, + 0.0026773237623274326, + 0.0021985373459756374, + 0.0033329532016068697, + 0.004237531218677759, + 0.003047799225896597, + 0.003928130026906729, + 0.005624211393296719, + 0.0029422903899103403, + 0.0040733651258051395, + 0.005719135049730539, + 0.0016865297220647335, + 0.003115747356787324, + 0.0023499722592532635, + 0.0022879159078001976, + 0.0026431167498230934, + 0.0015359355602413416, + 0.00202570972032845, + 0.0023577557876706123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.007629777304828167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794272035360336 + ], + "opposite_logprob_tokens": [ + -3.529794216156006 + ], + "logprob_reward": -0.029794272035360336, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794272035360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08418931066989899, + 0.09615330398082733, + 0.10635911673307419, + 0.02903885208070278, + 0.030131742358207703, + 0.09121502190828323, + 0.3607064485549927, + 0.06712119281291962, + 0.09667538106441498, + 0.03840963914990425 + ], + "credit_map": [ + 0.013604715466499329, + 0.010859423317015171, + 0.010359015315771103, + 0.009786715731024742, + 0.016372734680771828, + 0.009641076438128948, + 0.007016077172011137, + 0.006549557205289602, + 0.012698223814368248, + 0.007275223266333342, + 0.010562642477452755, + 0.009985697455704212, + 0.014576395973563194, + 0.014878200367093086, + 0.007967782206833363, + 0.018209140747785568, + 0.00823769811540842, + 0.007405238226056099, + 0.008831349201500416, + 0.010662040673196316, + 0.007921375334262848, + 0.017191007733345032, + 0.020334461703896523, + 0.02577594853937626, + 0.003175528021529317, + 0.003827539971098304, + 0.0037690140306949615, + 0.0031562333460897207, + 0.004841047804802656, + 0.003926394507288933, + 0.0027585008647292852, + 0.003584593068808317, + 0.004289405886083841, + 0.0035100732930004597, + 0.0026329283136874437, + 0.0034478502348065376, + 0.0034287376329302788, + 0.002964284736663103, + 0.0057371435686945915, + 0.004121319856494665, + 0.009815978817641735, + 0.010688881389796734, + 0.009266572073101997, + 0.0097374701872468, + 0.010920166037976742, + 0.009827596135437489, + 0.017382502555847168, + 0.013575859367847443, + 0.025357503443956375, + 0.013267655856907368, + 0.013996155932545662, + 0.006691477727144957, + 0.03935232385993004, + 0.15505515038967133, + 0.013846736401319504, + 0.09313943982124329, + 0.010547898709774017, + 0.008256080560386181, + 0.00930270366370678, + 0.008581689558923244, + 0.0077017517760396, + 0.006664518732577562, + 0.00630361121147871, + 0.009762930683791637, + 0.011262760497629642, + 0.014988264068961143, + 0.014026479795575142, + 0.009658102877438068, + 0.02070368081331253, + 0.00981765054166317, + 0.007638867478817701, + 0.008579575456678867, + 0.003824780462309718, + 0.004542244598269463, + 0.00724583026021719, + 0.0036768282297998667, + 0.006368489470332861, + 0.003973631653934717, + 0.005112428218126297, + 0.0036654050927609205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.029794272035360336 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25205644965171814 + ], + "opposite_logprob_tokens": [ + -1.5020564794540405 + ], + "logprob_reward": -0.25205644965171814, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25205644965171814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017701059579849243, + 0.031185930594801903, + 0.015490083955228329, + 0.012324785813689232, + 0.011467544361948967, + 0.031241808086633682, + 0.8419851660728455, + 0.023260070011019707, + 0.012524895370006561, + 0.0028187097050249577 + ], + "credit_map": [ + 0.0032314478885382414, + 0.0018302910029888153, + 0.003822093363851309, + 0.0014358508633449674, + 0.002182143507525325, + 0.0020787816029042006, + 0.0012424294836819172, + 0.0018780222162604332, + 0.0013177325017750263, + 0.002226519165560603, + 0.0014728010864928365, + 0.001279085292480886, + 0.011221992783248425, + 0.007206061389297247, + 0.0016424942295998335, + 0.0048192450776696205, + 0.0013152614701539278, + 0.001032396568916738, + 0.001311954576522112, + 0.0017056569922715425, + 0.001382491085678339, + 0.0024719841312617064, + 0.0028555907774716616, + 0.0034147482365369797, + 0.0015575544675812125, + 0.0014298867899924517, + 0.001890302519313991, + 0.0015580481849610806, + 0.00159137649461627, + 0.0014826279366388917, + 0.0013920426135882735, + 0.0014229464577510953, + 0.0014272719854488969, + 0.0013762060552835464, + 0.0014115903759375215, + 0.0016320358263328671, + 0.0013146779965609312, + 0.0013104922836646438, + 0.001790627371519804, + 0.001204642467200756, + 0.0033244534861296415, + 0.0027824370190501213, + 0.002966104308143258, + 0.0035944487899541855, + 0.003177536418661475, + 0.0031602715607732534, + 0.0064023332670331, + 0.00583422277122736, + 0.06088011711835861, + 0.02909882925450802, + 0.03261761739850044, + 0.013199412263929844, + 0.09189800918102264, + 0.36716100573539734, + 0.031176909804344177, + 0.2159532755613327, + 0.0019402108155190945, + 0.003125613322481513, + 0.002594695659354329, + 0.0023849925491958857, + 0.0050219520926475525, + 0.002757846377789974, + 0.0030479752458631992, + 0.0023867834825068712, + 0.0013712486252188683, + 0.0013360804878175259, + 0.0014777039177715778, + 0.0015569464303553104, + 0.00204870174638927, + 0.001160118030384183, + 0.0015896590193733573, + 0.0019844367634505033, + 0.0005174917750991881, + 0.0004799179150722921, + 0.000373106129700318, + 0.00026531171170063317, + 0.0004903406370431185, + 0.00029610013007186353, + 0.00021541686146520078, + 0.00018102471949532628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.25205644965171814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from a medium, slightly upward angle, the camera pans steadily to the right, closely following the subject's movements. In the frame, a man in his thirties, with short, dark brown hair and a neatly trimmed beard, leans slightly forward, his expression focused and serious. He wears a dark gray, cotton, long-sleeved Henley-collared T-shirt, the sleeves casually pushed up to his elbows, revealing his muscular forearms. He is in a softly lit bedroom, with the corner of a bed covered in white sheets visible in the background. Steadily and deliberately, the man extends his right arm toward a blue nylon travel bag on the bed. The bag features black canvas handles and a silver metal zipper, half-unzipped. As the camera follows, the thick wad of banknotes clutched in his left hand becomes visible. He then carefully places the wad of bills into the bag's open mouth, his fingers delicately moving as he pushes the money deeper into the bag. The natural morning light filtering in from a side window casts soft shadows on his arm and the surface of the travel bag, highlighting the texture of the nylon and the thickness of the banknotes. The dominant color scheme of the entire scene is the vibrant blue of the travel bag, the dark gray of the man's clothing, and the white of the bed sheet, creating an atmosphere of focus, intimacy, and a touch of tension.", + "seed": 7521, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_15/index_3450_seed_7521_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026677565556019543, + "vlm_logprob": -0.026677565556019543, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.026677565556019543, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008154788054525852 + ], + "opposite_logprob_tokens": [ + -5.13315486907959 + ], + "logprob_reward": -0.008154788054525852, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008154788054525852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05264100059866905, + 0.027149852365255356, + 0.035381875932216644, + 0.0651269480586052, + 0.030391637235879898, + 0.03581823408603668, + 0.01580822840332985, + 0.3934161961078644, + 0.0907222181558609, + 0.25354379415512085 + ], + "credit_map": [ + 0.008144112303853035, + 0.011853832751512527, + 0.005717060063034296, + 0.005882663186639547, + 0.006934675388038158, + 0.0067411684431135654, + 0.0038888701237738132, + 0.0034786167088896036, + 0.003448059316724539, + 0.005738316103816032, + 0.005193933378905058, + 0.0030597832519561052, + 0.002721988596022129, + 0.0029838604386895895, + 0.001956717111170292, + 0.002047194167971611, + 0.004149443935602903, + 0.006362178362905979, + 0.007138366345316172, + 0.003278964664787054, + 0.004170937463641167, + 0.00375558715313673, + 0.0034334994852542877, + 0.003092898288741708, + 0.007156123872846365, + 0.009789478033781052, + 0.01702052913606167, + 0.00709151104092598, + 0.005508566740900278, + 0.0058055748231709, + 0.005554854869842529, + 0.007200311869382858, + 0.004639938939362764, + 0.006825585849583149, + 0.005810234695672989, + 0.0023557920940220356, + 0.0035660869907587767, + 0.0027879378758370876, + 0.001853905851021409, + 0.002552152844145894, + 0.005454469006508589, + 0.0065301405265927315, + 0.004383230581879616, + 0.002021977212280035, + 0.0073320940136909485, + 0.004481072537600994, + 0.0028082490898668766, + 0.0028070001862943172, + 0.002425712998956442, + 0.0030827282462269068, + 0.0014833384193480015, + 0.0012365903239697218, + 0.0021377753000706434, + 0.0022247519809752703, + 0.001719471998512745, + 0.001497858203947544, + 0.01887667179107666, + 0.020323358476161957, + 0.009486845694482327, + 0.006907259579747915, + 0.1084202453494072, + 0.1467101126909256, + 0.023387786000967026, + 0.05930391699075699, + 0.007638305425643921, + 0.014665556140244007, + 0.005521895829588175, + 0.007326907478272915, + 0.01656057871878147, + 0.015241239219903946, + 0.008928521536290646, + 0.014839217066764832, + 0.02786239981651306, + 0.02548106014728546, + 0.017224105075001717, + 0.028429612517356873, + 0.03395545110106468, + 0.03227117657661438, + 0.05217522755265236, + 0.03614477068185806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.008154788054525852 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079339310526848 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.011079339310526848, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079339310526848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09512822329998016, + 0.07879994809627533, + 0.08251458406448364, + 0.11975567042827606, + 0.033117249608039856, + 0.07191555202007294, + 0.031205078586935997, + 0.1019822210073471, + 0.10626218467950821, + 0.2793193459510803 + ], + "credit_map": [ + 0.017421936616301537, + 0.009991767816245556, + 0.013387247920036316, + 0.013914937153458595, + 0.009631984867155552, + 0.011769919656217098, + 0.010314038023352623, + 0.008696389384567738, + 0.008261332288384438, + 0.014684244990348816, + 0.012695073150098324, + 0.0080744419246912, + 0.007851139642298222, + 0.00900435633957386, + 0.008550598286092281, + 0.00967875961214304, + 0.014744396321475506, + 0.01865321211516857, + 0.01257284265011549, + 0.005506748799234629, + 0.007359736133366823, + 0.0070613534189760685, + 0.006680328398942947, + 0.009935966692864895, + 0.013811145909130573, + 0.018002798780798912, + 0.03436635062098503, + 0.01586233824491501, + 0.01031008455902338, + 0.01058744452893734, + 0.007546104956418276, + 0.00926940143108368, + 0.00615428201854229, + 0.007170054130256176, + 0.004548178054392338, + 0.0030020796693861485, + 0.0036449148319661617, + 0.0030212663114070892, + 0.0025902569759637117, + 0.0029862141236662865, + 0.013845660723745823, + 0.012278079986572266, + 0.009566327556967735, + 0.005144812166690826, + 0.011964022181928158, + 0.00775630958378315, + 0.005556776653975248, + 0.005803566426038742, + 0.005380683112889528, + 0.004829734563827515, + 0.003567140782251954, + 0.002906267764046788, + 0.004209939856082201, + 0.003468731651082635, + 0.0034577578771859407, + 0.003384821815416217, + 0.007466140203177929, + 0.009371732361614704, + 0.004283220041543245, + 0.0024996520951390266, + 0.025134582072496414, + 0.03513399139046669, + 0.005453228484839201, + 0.012639676220715046, + 0.009730738587677479, + 0.011239045299589634, + 0.006549489684402943, + 0.008044467307627201, + 0.010918496176600456, + 0.021448323503136635, + 0.015054846182465553, + 0.02327677235007286, + 0.03137914091348648, + 0.02753453142940998, + 0.020504120737314224, + 0.03291640803217888, + 0.03504770249128342, + 0.0352332666516304, + 0.05754087492823601, + 0.039163291454315186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.011079339310526848 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096688237041235 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096688237041235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096688237041235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06263235211372375, + 0.020989924669265747, + 0.07719264924526215, + 0.06109915301203728, + 0.06555598229169846, + 0.13103485107421875, + 0.03826965391635895, + 0.04872123897075653, + 0.09875056147575378, + 0.3957535922527313 + ], + "credit_map": [ + 0.012913019396364689, + 0.00797908753156662, + 0.008104764856398106, + 0.005815682467073202, + 0.00653929403051734, + 0.008041730150580406, + 0.006675004027783871, + 0.0065637617371976376, + 0.0019477325258776546, + 0.0028791744261980057, + 0.003835688577964902, + 0.002482099924236536, + 0.0018951648380607367, + 0.0024098462890833616, + 0.0025916174054145813, + 0.002948599401861429, + 0.007710461039096117, + 0.012022946029901505, + 0.010463135316967964, + 0.0074870469979941845, + 0.011001021601259708, + 0.009107010439038277, + 0.008342784829437733, + 0.011058252304792404, + 0.005399140063673258, + 0.006736206356436014, + 0.012142330408096313, + 0.0074869319796562195, + 0.0055541242472827435, + 0.005764618515968323, + 0.008437590673565865, + 0.009578211233019829, + 0.009846020489931107, + 0.012524114921689034, + 0.009610552340745926, + 0.005142072215676308, + 0.009693223983049393, + 0.006601917091757059, + 0.005177628248929977, + 0.006960457190871239, + 0.02066989429295063, + 0.020072605460882187, + 0.01429531816393137, + 0.010762332938611507, + 0.020982375368475914, + 0.014577111229300499, + 0.015611412934958935, + 0.014063801616430283, + 0.004706066567450762, + 0.0049451375380158424, + 0.00332397292368114, + 0.0026246574707329273, + 0.0052948445081710815, + 0.006593118421733379, + 0.0048113251104950905, + 0.005970532540231943, + 0.005910091567784548, + 0.007494488265365362, + 0.007370932959020138, + 0.0043286108411848545, + 0.004643742460757494, + 0.005986480507999659, + 0.006439479533582926, + 0.006547414697706699, + 0.012948069721460342, + 0.01607511006295681, + 0.007847784087061882, + 0.007923201657831669, + 0.018114756792783737, + 0.01076130848377943, + 0.01110406219959259, + 0.013976272195577621, + 0.04501062259078026, + 0.034400660544633865, + 0.02926063910126686, + 0.04630190134048462, + 0.05112658068537712, + 0.056826964020729065, + 0.08163221180438995, + 0.05119398608803749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.004096688237041235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977774802595377 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.00977774802595377, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977774802595377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.106584332883358, + 0.04571092501282692, + 0.0360720120370388, + 0.060019589960575104, + 0.034940700978040695, + 0.11938615143299103, + 0.08161681145429611, + 0.18064002692699432, + 0.22422237694263458, + 0.1108071357011795 + ], + "credit_map": [ + 0.015070357359945774, + 0.010433611460030079, + 0.013925295323133469, + 0.014610065147280693, + 0.010213508270680904, + 0.012550592422485352, + 0.01428956724703312, + 0.015491332858800888, + 0.0049575273878872395, + 0.005421601235866547, + 0.008197592571377754, + 0.004525352269411087, + 0.003350894432514906, + 0.005163886584341526, + 0.006575760431587696, + 0.007518311496824026, + 0.003116968786343932, + 0.00347137707285583, + 0.004251462407410145, + 0.0029916383791714907, + 0.004258994944393635, + 0.003857359290122986, + 0.007804972119629383, + 0.006319239269942045, + 0.006229333579540253, + 0.006336513441056013, + 0.011076999828219414, + 0.0068529085256159306, + 0.005219715181738138, + 0.005701236426830292, + 0.008124087937176228, + 0.010478791780769825, + 0.004101042170077562, + 0.0069859763607382774, + 0.0040930029936134815, + 0.003755943849682808, + 0.004963520914316177, + 0.0035880885552614927, + 0.003591579385101795, + 0.0038615453522652388, + 0.026523912325501442, + 0.02190055139362812, + 0.015476283617317677, + 0.007180516608059406, + 0.019139470532536507, + 0.012086303904652596, + 0.007836008444428444, + 0.009243101812899113, + 0.017878936603665352, + 0.016462821513414383, + 0.00758981890976429, + 0.0064313095062971115, + 0.00935957208275795, + 0.008237028494477272, + 0.006700933910906315, + 0.00895639043301344, + 0.008336027152836323, + 0.009240423329174519, + 0.005432013422250748, + 0.0038291062228381634, + 0.04745231941342354, + 0.06491556763648987, + 0.012880543246865273, + 0.028554029762744904, + 0.025378938764333725, + 0.02599363401532173, + 0.019175168126821518, + 0.02168883942067623, + 0.04462131857872009, + 0.03266909718513489, + 0.02934391237795353, + 0.02535146102309227, + 0.009720761328935623, + 0.009653700515627861, + 0.00967799685895443, + 0.01196096371859312, + 0.01866357959806919, + 0.01764223538339138, + 0.02092236839234829, + 0.012565523386001587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.00977774802595377 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a man in his thirties with short dark brown hair and a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray, cotton, long-sleeved Henley-collared T-shirt with sleeves pushed up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a blue nylon travel bag with black canvas handles and a silver metal zipper visible on a bed with white sheets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man extend a thick wad of banknotes from his left hand and place them into the open travel bag?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily to the right while following the subject's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027926415205002 + ], + "opposite_logprob_tokens": [ + -2.3502793312072754 + ], + "logprob_reward": -0.10027926415205002, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027926415205002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07561327517032623, + 0.04496118798851967, + 0.06087025627493858, + 0.05093446373939514, + 0.053913671523332596, + 0.08836816996335983, + 0.05787454918026924, + 0.09754345566034317, + 0.2394394725561142, + 0.23048146069049835 + ], + "credit_map": [ + 0.00904754363000393, + 0.014816085807979107, + 0.008285833522677422, + 0.008117293007671833, + 0.009258649311959743, + 0.009694907814264297, + 0.007837528362870216, + 0.008555435575544834, + 0.005774720571935177, + 0.007113682106137276, + 0.0073312693275511265, + 0.0048948898911476135, + 0.00440138578414917, + 0.005073473323136568, + 0.005500313825905323, + 0.004871451761573553, + 0.005502630025148392, + 0.008419939316809177, + 0.00803932175040245, + 0.004274823237210512, + 0.008387339301407337, + 0.00701858988031745, + 0.009777297265827656, + 0.009450321085751057, + 0.004545565228909254, + 0.009088249877095222, + 0.009465762414038181, + 0.006469379179179668, + 0.004661190789192915, + 0.0045854742638766766, + 0.0064232079312205315, + 0.005695635452866554, + 0.008279276080429554, + 0.010433824732899666, + 0.008127710781991482, + 0.004235136788338423, + 0.007419019937515259, + 0.005854044109582901, + 0.004826050251722336, + 0.004738612566143274, + 0.01573328487575054, + 0.01726522669196129, + 0.013385028578341007, + 0.007245808374136686, + 0.011345026083290577, + 0.009636367671191692, + 0.005722407717257738, + 0.008035019040107727, + 0.010425043292343616, + 0.010987176559865475, + 0.005944058299064636, + 0.005092790815979242, + 0.006020389962941408, + 0.005911606829613447, + 0.007013768423348665, + 0.006479712203145027, + 0.013797213323414326, + 0.027267877012491226, + 0.007254573050886393, + 0.00440995441749692, + 0.010714869014918804, + 0.02098664455115795, + 0.004715201444923878, + 0.008397124707698822, + 0.027107633650302887, + 0.04905232414603233, + 0.01746198907494545, + 0.024351291358470917, + 0.0393906868994236, + 0.03510231152176857, + 0.024557558819651604, + 0.022415680810809135, + 0.026530137285590172, + 0.01837981678545475, + 0.01858978345990181, + 0.027089910581707954, + 0.027551693841814995, + 0.02615504339337349, + 0.054179899394512177, + 0.03200516849756241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.10027926415205002 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_2/index_tensor([7348, 7348])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_2/index_tensor([7348, 7348])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f26570fc2bb1e1741e062dd64487a43e2b319dc0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_2/index_tensor([7348, 7348])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, side-scrolling medium shot, rendered in a fluid 2D animation style, fast-tracks a young elf astride a majestic unicorn through an enchanted forest. The camera moves from left to right, maintaining perfect framing as the two race through dense foliage. The elf, with delicately pointed ears and wide, joyful eyes, leans forward in the saddle, her long, shimmering silver hair flowing dramatically behind her like a silk banner in the wind. Clad in a tunic woven from emerald green leaves and simple brown leather bracers, they clutch the unicorn's thick, flowing mane. The unicorn itself is a creature of pure magic, its fur a pearly white that seems to shimmer, its powerful legs running with swiftness and grace. Its mane and tail are a vibrant array of pastel colors, and its single, spiraling horn emits a soft inner glow. The environment is a symphony of vibrant, saturated color; ancient, gnarled trees with moss-covered bark form a dense canopy, above which hang magical, glowing orbs of light in brilliant shades of pink, teal, and gold, casting a dappled, ethereal glow across the scene. The forest floor is a carpet of soft moss and luminescent fungi, and rapid movement creates a beautiful parallax effect, with foreground ferns and flowers blurring past, enhancing the thrilling sense of speed and creating an atmosphere of pure, unadulterated magical adventure.", + "seed": 30890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_2/index_7348_seed_30890_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11774608460254968, + "vlm_logprob": -0.11774608460254968, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.11774608460254968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02038726396858692 + ], + "opposite_logprob_tokens": [ + -4.020387172698975 + ], + "logprob_reward": -0.02038726396858692, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02038726396858692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21913287043571472, + 0.17196887731552124, + 0.07369915395975113, + 0.08469204604625702, + 0.07089745253324509, + 0.0668843612074852, + 0.11542728543281555, + 0.04661577567458153, + 0.04050504416227341, + 0.11017712205648422 + ], + "credit_map": [ + 0.03185543790459633, + 0.02906780317425728, + 0.029228810220956802, + 0.01998678781092167, + 0.02561381831765175, + 0.0297368336468935, + 0.02811298333108425, + 0.025530394166707993, + 0.025706777349114418, + 0.015228230506181717, + 0.026337651535868645, + 0.013660921715199947, + 0.019727488979697227, + 0.03335559740662575, + 0.01899629645049572, + 0.018955906853079796, + 0.01089173462241888, + 0.014889046549797058, + 0.0066442666575312614, + 0.006280078552663326, + 0.007327612489461899, + 0.006279474589973688, + 0.0068855793215334415, + 0.014501362107694149, + 0.011810391210019588, + 0.009527361020445824, + 0.01741827093064785, + 0.009647274389863014, + 0.009093759581446648, + 0.009209846146404743, + 0.009338375180959702, + 0.0086467694491148, + 0.006753874011337757, + 0.007972842082381248, + 0.007290608715265989, + 0.008674529381096363, + 0.011374207213521004, + 0.007812803611159325, + 0.01045625563710928, + 0.010562335141003132, + 0.019869932904839516, + 0.0069309719838202, + 0.005221784580498934, + 0.004560562781989574, + 0.006604952272027731, + 0.008559802547097206, + 0.007018159609287977, + 0.008118199184536934, + 0.015809113159775734, + 0.011509938165545464, + 0.010460486635565758, + 0.014001673087477684, + 0.01314494013786316, + 0.012545598670840263, + 0.019646765664219856, + 0.018308769911527634, + 0.006629513576626778, + 0.004704004153609276, + 0.0036845169961452484, + 0.005045681726187468, + 0.008249861188232899, + 0.005489815957844257, + 0.00712314248085022, + 0.005689238663762808, + 0.004160638432949781, + 0.0047525460831820965, + 0.006348925177007914, + 0.004632611759006977, + 0.00435195816680789, + 0.004731970839202404, + 0.005260571371763945, + 0.006265822798013687, + 0.017111703753471375, + 0.014180963858962059, + 0.01700778491795063, + 0.013012043200433254, + 0.008058251813054085, + 0.010408081114292145, + 0.01443474367260933, + 0.01596355065703392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02038726396858692 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373163938522339 + ], + "opposite_logprob_tokens": [ + -3.408731698989868 + ], + "logprob_reward": -0.03373163938522339, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373163938522339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15351158380508423, + 0.11413852870464325, + 0.1048772931098938, + 0.11015500128269196, + 0.056192293763160706, + 0.04523551091551781, + 0.08835562318563461, + 0.2012210488319397, + 0.07986100763082504, + 0.04645203426480293 + ], + "credit_map": [ + 0.019649585708975792, + 0.02707804925739765, + 0.016981130465865135, + 0.015830274671316147, + 0.015101088210940361, + 0.01928480714559555, + 0.021860115230083466, + 0.01772654429078102, + 0.010053581558167934, + 0.009775050915777683, + 0.02802707441151142, + 0.0096919359639287, + 0.009413144551217556, + 0.02412954531610012, + 0.014118230901658535, + 0.00892997719347477, + 0.011586494743824005, + 0.02559026889503002, + 0.010394143871963024, + 0.008861321024596691, + 0.008678345009684563, + 0.007715101353824139, + 0.00998622365295887, + 0.022065402939915657, + 0.014907093718647957, + 0.009034574031829834, + 0.01917489804327488, + 0.010967036709189415, + 0.011299125850200653, + 0.017460163682699203, + 0.011877995915710926, + 0.015434112399816513, + 0.004287552088499069, + 0.006087161134928465, + 0.005481029860675335, + 0.006163529120385647, + 0.0093556372448802, + 0.0065541863441467285, + 0.011901828460395336, + 0.006361372768878937, + 0.008407292887568474, + 0.007677758578211069, + 0.003504703985527158, + 0.00466916523873806, + 0.005513771902769804, + 0.003866911632940173, + 0.004677749238908291, + 0.006918162107467651, + 0.008020752109587193, + 0.0096496706828475, + 0.008730418048799038, + 0.012235448695719242, + 0.01273522898554802, + 0.009452369064092636, + 0.01733756624162197, + 0.01019418053328991, + 0.03457898646593094, + 0.021826786920428276, + 0.029430273920297623, + 0.02259862795472145, + 0.02027500793337822, + 0.023274686187505722, + 0.02306206524372101, + 0.0261746384203434, + 0.006655971985310316, + 0.006703669670969248, + 0.010565311647951603, + 0.011756893247365952, + 0.014479185454547405, + 0.008226975798606873, + 0.0125748785212636, + 0.008898133412003517, + 0.007427632343024015, + 0.00587723171338439, + 0.006364061497151852, + 0.006795062683522701, + 0.0035845304373651743, + 0.004407040774822235, + 0.0057819015346467495, + 0.006214572116732597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.03373163938522339 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015149556566029787 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015149556566029787, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015149556566029787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1757282167673111, + 0.2560725510120392, + 0.09569235146045685, + 0.08478501439094543, + 0.10360708087682724, + 0.04427199438214302, + 0.06345158815383911, + 0.05531452223658562, + 0.037218693643808365, + 0.08385804295539856 + ], + "credit_map": [ + 0.019910799339413643, + 0.028898391872644424, + 0.01892714574933052, + 0.016416026279330254, + 0.020170750096440315, + 0.022430887445807457, + 0.027427196502685547, + 0.021546998992562294, + 0.022630134597420692, + 0.01624181866645813, + 0.08405733853578568, + 0.02150540053844452, + 0.016594024375081062, + 0.047730255872011185, + 0.030158251523971558, + 0.017155319452285767, + 0.010742377489805222, + 0.012387555092573166, + 0.010839051567018032, + 0.009939475916326046, + 0.02215523272752762, + 0.010212576016783714, + 0.007851293310523033, + 0.011564779095351696, + 0.010480520315468311, + 0.009570740163326263, + 0.01510943565517664, + 0.008253573440015316, + 0.010108661837875843, + 0.01048478577286005, + 0.008042320609092712, + 0.012734971009194851, + 0.007842963561415672, + 0.015506413765251637, + 0.008002073504030704, + 0.01598040573298931, + 0.01646668277680874, + 0.013351649977266788, + 0.0148620018735528, + 0.011594876646995544, + 0.005710390862077475, + 0.005602309480309486, + 0.004759275354444981, + 0.005293881054967642, + 0.0060205827467143536, + 0.0059069073759019375, + 0.005639281589537859, + 0.005339359864592552, + 0.007764345500618219, + 0.009212520904839039, + 0.008512841537594795, + 0.008281187154352665, + 0.006326238624751568, + 0.007534443400800228, + 0.007943568751215935, + 0.007876435294747353, + 0.008551866747438908, + 0.00441486993804574, + 0.007979781366884708, + 0.007836178876459599, + 0.0064612217247486115, + 0.0059706950560212135, + 0.007704768795520067, + 0.006395129486918449, + 0.0029462971724569798, + 0.0035688444040715694, + 0.004182864911854267, + 0.005895079579204321, + 0.007029026746749878, + 0.003810859750956297, + 0.005635027773678303, + 0.004150690045207739, + 0.011167599819600582, + 0.009483194909989834, + 0.013755181804299355, + 0.00838060025125742, + 0.008807143196463585, + 0.009100580587983131, + 0.011273911222815514, + 0.01188982930034399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0015149556566029787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778928942978382 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778928942978382, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778928942978382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19704298675060272, + 0.04136030003428459, + 0.20330002903938293, + 0.06563477963209152, + 0.10326644778251648, + 0.0961260199546814, + 0.04922518879175186, + 0.06642208993434906, + 0.05463116616010666, + 0.12299107015132904 + ], + "credit_map": [ + 0.025301916524767876, + 0.03223723545670509, + 0.022909408435225487, + 0.016892628744244576, + 0.02499683015048504, + 0.025427663698792458, + 0.02719554677605629, + 0.022081729024648666, + 0.006032788660377264, + 0.0036329650320112705, + 0.00581583334133029, + 0.005687584634870291, + 0.006302523426711559, + 0.004808987025171518, + 0.004416582174599171, + 0.004663033410906792, + 0.017224358394742012, + 0.051317960023880005, + 0.022539740428328514, + 0.014359947293996811, + 0.031318433582782745, + 0.02487936243414879, + 0.01603059470653534, + 0.02562960982322693, + 0.007454934064298868, + 0.008719397708773613, + 0.012399337254464626, + 0.007133635692298412, + 0.007742215413600206, + 0.00689761433750391, + 0.006472610402852297, + 0.008815030567348003, + 0.009281350299715996, + 0.012852673418819904, + 0.006684232037514448, + 0.015097658149898052, + 0.018092036247253418, + 0.010936991311609745, + 0.018002824857831, + 0.012318666093051434, + 0.01239139772951603, + 0.010487206280231476, + 0.010094868019223213, + 0.010232838802039623, + 0.016978368163108826, + 0.013373891822993755, + 0.010268365032970905, + 0.0122990682721138, + 0.006427729967981577, + 0.00873961765319109, + 0.004923089873045683, + 0.006007623393088579, + 0.005533025600016117, + 0.005703823175281286, + 0.006293809972703457, + 0.005596463102847338, + 0.007648797240108252, + 0.006443313788622618, + 0.005512481555342674, + 0.0072221155278384686, + 0.014126204885542393, + 0.007994841784238815, + 0.010905432514846325, + 0.006568895652890205, + 0.0045221829786896706, + 0.004342246800661087, + 0.006414584815502167, + 0.008453783579170704, + 0.01039501465857029, + 0.006058814469724894, + 0.008359162136912346, + 0.0060853660106658936, + 0.018521932885050774, + 0.01282423920929432, + 0.017062989994883537, + 0.014211111702024937, + 0.01539713516831398, + 0.013876114971935749, + 0.016274284571409225, + 0.014823243953287601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.009778928942978382 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233176350593567 + ], + "opposite_logprob_tokens": [ + -0.8983176350593567 + ], + "logprob_reward": -0.5233176350593567, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233176350593567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10935170948505402, + 0.0545828714966774, + 0.06595735251903534, + 0.04871802031993866, + 0.054874151945114136, + 0.08532791584730148, + 0.03397538512945175, + 0.050168491899967194, + 0.13034465909004211, + 0.3666994869709015 + ], + "credit_map": [ + 0.013856618665158749, + 0.013582772575318813, + 0.01438581570982933, + 0.011330820620059967, + 0.012701110914349556, + 0.01398205291479826, + 0.015642277896404266, + 0.013870240189135075, + 0.005340843927115202, + 0.004081640392541885, + 0.007989929988980293, + 0.004674353636801243, + 0.006438052747398615, + 0.013752189464867115, + 0.00746935186907649, + 0.004836509004235268, + 0.007476778235286474, + 0.016132529824972153, + 0.005619551055133343, + 0.00593670504167676, + 0.010642638429999352, + 0.005303315818309784, + 0.005186932161450386, + 0.009658905677497387, + 0.007311413995921612, + 0.006818488705903292, + 0.006108985282480717, + 0.005296742543578148, + 0.0066992430947721004, + 0.006690520327538252, + 0.005351888947188854, + 0.00444073649123311, + 0.0053560021333396435, + 0.006420508027076721, + 0.005811768118292093, + 0.00812210887670517, + 0.007459741551429033, + 0.004934418480843306, + 0.008866121992468834, + 0.00790348369628191, + 0.027560822665691376, + 0.007420451380312443, + 0.007331805769354105, + 0.0059137470088899136, + 0.008684452623128891, + 0.008219811134040356, + 0.0076851435005664825, + 0.01251168828457594, + 0.003367240307852626, + 0.0037943345960229635, + 0.00451132794842124, + 0.003530084853991866, + 0.004458417184650898, + 0.0027767166029661894, + 0.005879208445549011, + 0.0056580533273518085, + 0.009119097143411636, + 0.005302915815263987, + 0.005929805804044008, + 0.005090442020446062, + 0.00820231158286333, + 0.004763773176819086, + 0.0061272550374269485, + 0.005632889922708273, + 0.011914113536477089, + 0.01247689500451088, + 0.01636826992034912, + 0.019872672855854034, + 0.01958010531961918, + 0.0135328508913517, + 0.02103528007864952, + 0.015564467757940292, + 0.056680794805288315, + 0.04231113940477371, + 0.06041749194264412, + 0.03898392617702484, + 0.02995133027434349, + 0.03727482259273529, + 0.04938012361526489, + 0.05169985443353653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.5233176350593567 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, side-scrolling medium shot, rendered in a fluid 2D animation style, fast-tracks a young elf astride a majestic unicorn through an enchanted forest. The camera moves from left to right, maintaining perfect framing as the two race through dense foliage. The elf, with delicately pointed ears and wide, joyful eyes, leans forward in the saddle, her long, shimmering silver hair flowing dramatically behind her like a silk banner in the wind. Clad in a tunic woven from emerald green leaves and simple brown leather bracers, they clutch the unicorn's thick, flowing mane. The unicorn itself is a creature of pure magic, its fur a pearly white that seems to shimmer, its powerful legs running with swiftness and grace. Its mane and tail are a vibrant array of pastel colors, and its single, spiraling horn emits a soft inner glow. The environment is a symphony of vibrant, saturated color; ancient, gnarled trees with moss-covered bark form a dense canopy, above which hang magical, glowing orbs of light in brilliant shades of pink, teal, and gold, casting a dappled, ethereal glow across the scene. The forest floor is a carpet of soft moss and luminescent fungi, and rapid movement creates a beautiful parallax effect, with foreground ferns and flowers blurring past, enhancing the thrilling sense of speed and creating an atmosphere of pure, unadulterated magical adventure.", + "seed": 30891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_2/index_7348_seed_30891_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1547511408571154, + "vlm_logprob": -0.1547511408571154, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.1547511408571154, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026898376643657684 + ], + "opposite_logprob_tokens": [ + -3.7768983840942383 + ], + "logprob_reward": -0.026898376643657684, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026898376643657684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18581894040107727, + 0.08724116533994675, + 0.039502087980508804, + 0.21131958067417145, + 0.05135910585522652, + 0.05079914629459381, + 0.1929093301296234, + 0.040223609656095505, + 0.08631455898284912, + 0.054512497037649155 + ], + "credit_map": [ + 0.027695871889591217, + 0.014720435254275799, + 0.04249285161495209, + 0.011695710010826588, + 0.017746340483427048, + 0.045574087649583817, + 0.009787672199308872, + 0.016105977818369865, + 0.01446146983653307, + 0.006715397350490093, + 0.01360689289867878, + 0.00936480425298214, + 0.010646690614521503, + 0.013189617544412613, + 0.007691659964621067, + 0.011564632877707481, + 0.006406362634152174, + 0.004369979724287987, + 0.0062817600555717945, + 0.004228834994137287, + 0.004700439982116222, + 0.0050743138417601585, + 0.004657141398638487, + 0.003783258143812418, + 0.02157261036336422, + 0.019664905965328217, + 0.034961868077516556, + 0.020783085376024246, + 0.0460696704685688, + 0.028347469866275787, + 0.017351679503917694, + 0.022568294778466225, + 0.006300076376646757, + 0.004530218429863453, + 0.0067588635720312595, + 0.004349276889115572, + 0.010995380580425262, + 0.008063847199082375, + 0.0038373039569705725, + 0.006524136755615473, + 0.013700775802135468, + 0.0035644210875034332, + 0.010859576053917408, + 0.006588608492165804, + 0.005481648724526167, + 0.0029708140064030886, + 0.0042057279497385025, + 0.003427572548389435, + 0.021668093279004097, + 0.008321727626025677, + 0.06268329918384552, + 0.011164464987814426, + 0.04648634046316147, + 0.016675928607583046, + 0.011640212498605251, + 0.014269264414906502, + 0.004669645335525274, + 0.0033968789502978325, + 0.008087512105703354, + 0.003711237106472254, + 0.006095551885664463, + 0.006441149860620499, + 0.0034515189472585917, + 0.00437011756002903, + 0.010376610793173313, + 0.005428002681583166, + 0.02281181327998638, + 0.009787552058696747, + 0.012850583530962467, + 0.011176778934895992, + 0.008491646498441696, + 0.005391573067754507, + 0.005750820506364107, + 0.004879744257777929, + 0.010712262243032455, + 0.005362288095057011, + 0.0074576628394424915, + 0.008533291518688202, + 0.005516901612281799, + 0.006299525499343872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.026898376643657684 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038135167211294174 + ], + "opposite_logprob_tokens": [ + -3.288135051727295 + ], + "logprob_reward": -0.038135167211294174, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038135167211294174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09438982605934143, + 0.08389371633529663, + 0.050897542387247086, + 0.0728551521897316, + 0.03724662587046623, + 0.036172304302453995, + 0.5151430368423462, + 0.024902218952775, + 0.04976700246334076, + 0.03473254665732384 + ], + "credit_map": [ + 0.018111687153577805, + 0.007130348589271307, + 0.015895647928118706, + 0.010008228942751884, + 0.012383753433823586, + 0.0168031994253397, + 0.006611124612390995, + 0.007445828523486853, + 0.008407574146986008, + 0.006713011767715216, + 0.02479742467403412, + 0.006623469293117523, + 0.01335747167468071, + 0.011239754967391491, + 0.005818769335746765, + 0.006936242338269949, + 0.0047060828655958176, + 0.0044283573515713215, + 0.006134081166237593, + 0.0035961144603788853, + 0.012560156174004078, + 0.010292666032910347, + 0.004167220089584589, + 0.005012865643948317, + 0.007151039317250252, + 0.006950219627469778, + 0.009565556421875954, + 0.00675618601962924, + 0.014590620994567871, + 0.012774751521646976, + 0.0058647701516747475, + 0.00920200813561678, + 0.0037359667476266623, + 0.0022642859257757664, + 0.004483877681195736, + 0.002794884843751788, + 0.009089348837733269, + 0.009690196253359318, + 0.002099723555147648, + 0.003088343655690551, + 0.005432681646198034, + 0.0022817549761384726, + 0.007673063315451145, + 0.0029037189669907093, + 0.011714044958353043, + 0.0027222810313105583, + 0.0018891878426074982, + 0.0015555697027593851, + 0.07612492144107819, + 0.018327325582504272, + 0.19198641180992126, + 0.02607107162475586, + 0.0974874421954155, + 0.0494413897395134, + 0.021348299458622932, + 0.03435617312788963, + 0.0036917799152433872, + 0.0014595569809898734, + 0.004460531752556562, + 0.001795208896510303, + 0.002633151365444064, + 0.006730422377586365, + 0.0012987289810553193, + 0.0028328392654657364, + 0.006113311741501093, + 0.0023800416383892298, + 0.015354892238974571, + 0.0038491208106279373, + 0.00747689139097929, + 0.008752455934882164, + 0.0032074032351374626, + 0.0026328819803893566, + 0.004292195662856102, + 0.002674625487998128, + 0.006730322260409594, + 0.0034570691641420126, + 0.004068411886692047, + 0.0064582498744130135, + 0.003855463583022356, + 0.0031962101347744465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.038135167211294174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821395406499505 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.002821395406499505, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821395406499505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06744547933340073, + 0.04603230580687523, + 0.13000674545764923, + 0.11832573264837265, + 0.023387586697936058, + 0.05431751161813736, + 0.37053176760673523, + 0.0689568966627121, + 0.08341098576784134, + 0.03758501634001732 + ], + "credit_map": [ + 0.010336042381823063, + 0.00869496539235115, + 0.009654730558395386, + 0.0083118611946702, + 0.0076585388742387295, + 0.008894641883671284, + 0.007014585193246603, + 0.006880119908601046, + 0.006299211643636227, + 0.004253890365362167, + 0.009666419588029385, + 0.004621544387191534, + 0.006052185781300068, + 0.007994997315108776, + 0.003730118740350008, + 0.0034139351919293404, + 0.014300300739705563, + 0.017016224563121796, + 0.015047607012093067, + 0.013318143784999847, + 0.01685013435781002, + 0.026360902935266495, + 0.016168266534805298, + 0.010945172980427742, + 0.011380569078028202, + 0.01192792970687151, + 0.014806767925620079, + 0.010780605487525463, + 0.023073986172676086, + 0.01879376359283924, + 0.010333042591810226, + 0.017229067161679268, + 0.0033015443477779627, + 0.0017195874825119972, + 0.0031892883125692606, + 0.0017031605821102858, + 0.0051547870971262455, + 0.00412269402295351, + 0.0013910738052800298, + 0.0028054506983608007, + 0.004055315628647804, + 0.0033753460738807917, + 0.010213024914264679, + 0.008387768641114235, + 0.019690323621034622, + 0.0034082040656358004, + 0.003454536898061633, + 0.0017329917754977942, + 0.053666237741708755, + 0.014293051324784756, + 0.13315460085868835, + 0.019450029358267784, + 0.08044083416461945, + 0.02545115165412426, + 0.017630193382501602, + 0.026445655152201653, + 0.011249578557908535, + 0.003912726417183876, + 0.0102263567969203, + 0.004780914168804884, + 0.0069434852339327335, + 0.020828623324632645, + 0.0041350917890667915, + 0.0068801166489720345, + 0.009962458163499832, + 0.003670803736895323, + 0.026419105008244514, + 0.007310677319765091, + 0.012918705120682716, + 0.01361940149217844, + 0.005104268901050091, + 0.004405569285154343, + 0.004957532975822687, + 0.0033589424565434456, + 0.007324445992708206, + 0.0037391488440334797, + 0.004287888761609793, + 0.005350144580006599, + 0.003994491882622242, + 0.004572420381009579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.002821395406499505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538767419755459 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.012538767419755459, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538767419755459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1160711720585823, + 0.11743514984846115, + 0.1291445940732956, + 0.11114165186882019, + 0.038257043808698654, + 0.02336232177913189, + 0.37591353058815, + 0.02209537848830223, + 0.030293714255094528, + 0.03628545254468918 + ], + "credit_map": [ + 0.021180538460612297, + 0.008920565247535706, + 0.02570832334458828, + 0.0086909094825387, + 0.01188165508210659, + 0.023603860288858414, + 0.007154058665037155, + 0.008931262418627739, + 0.011490347795188427, + 0.009836236014962196, + 0.0326850563287735, + 0.010417606681585312, + 0.018600907176733017, + 0.01640983670949936, + 0.007813642732799053, + 0.010181508958339691, + 0.01309875026345253, + 0.00685517443343997, + 0.013554712757468224, + 0.006765794940292835, + 0.03690866008400917, + 0.03397345170378685, + 0.008284213952720165, + 0.009703843854367733, + 0.011164716444909573, + 0.011553090997040272, + 0.014326359145343304, + 0.010628961026668549, + 0.022567326202988625, + 0.015451410785317421, + 0.010731898248195648, + 0.0147178890183568, + 0.005382383707910776, + 0.0023916198406368494, + 0.006517757661640644, + 0.002932786475867033, + 0.0070582227781414986, + 0.007035588845610619, + 0.0026191910728812218, + 0.004319493658840656, + 0.003909375984221697, + 0.0020405463874340057, + 0.0045839399099349976, + 0.0018837093375623226, + 0.003739449195563793, + 0.0027391957119107246, + 0.0018409090116620064, + 0.0026251967065036297, + 0.05830853804945946, + 0.01288041565567255, + 0.1356886625289917, + 0.017333868891000748, + 0.08494162559509277, + 0.02658235654234886, + 0.012001565657556057, + 0.02817649580538273, + 0.0029038588982075453, + 0.0014247120125219226, + 0.004483417142182589, + 0.0015997481532394886, + 0.00293932412751019, + 0.004409376531839371, + 0.0021735001355409622, + 0.0021614409051835537, + 0.003088052850216627, + 0.0025657135993242264, + 0.005309321917593479, + 0.0028279051184654236, + 0.00430363928899169, + 0.007377639412879944, + 0.0028840210288763046, + 0.001937424298375845, + 0.004621993750333786, + 0.0033850716426968575, + 0.0073547521606087685, + 0.0037122026551514864, + 0.00420158077031374, + 0.005617828574031591, + 0.004038775339722633, + 0.003353251377120614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.012538767419755459 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933619976043701 + ], + "opposite_logprob_tokens": [ + -0.6933619976043701 + ], + "logprob_reward": -0.6933619976043701, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933619976043701, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09699804335832596, + 0.16043037176132202, + 0.15431949496269226, + 0.1681092530488968, + 0.0516904778778553, + 0.07756155729293823, + 0.07073168456554413, + 0.025541147217154503, + 0.16853591799736023, + 0.026082079857587814 + ], + "credit_map": [ + 0.014519356191158295, + 0.01113288663327694, + 0.013708509504795074, + 0.011702493764460087, + 0.008232747204601765, + 0.017030734568834305, + 0.01000905316323042, + 0.0106622613966465, + 0.019792679697275162, + 0.007323855999857187, + 0.04230460524559021, + 0.012377085164189339, + 0.023844506591558456, + 0.019658207893371582, + 0.011442003771662712, + 0.02368743158876896, + 0.01774589903652668, + 0.020403381437063217, + 0.0208699069917202, + 0.016995741054415703, + 0.01664809137582779, + 0.02704661153256893, + 0.01895781233906746, + 0.015652058646082878, + 0.016616391018033028, + 0.014201049692928791, + 0.020972605794668198, + 0.013802398927509785, + 0.036518387496471405, + 0.028019534423947334, + 0.012644397094845772, + 0.02533450350165367, + 0.006605216767638922, + 0.004122512880712748, + 0.007476077880710363, + 0.004267251584678888, + 0.01081263367086649, + 0.008039354346692562, + 0.0043207798153162, + 0.006046651396900415, + 0.009959070011973381, + 0.005852414295077324, + 0.01443664263933897, + 0.016849635168910027, + 0.009154744446277618, + 0.008873366750776768, + 0.006022043526172638, + 0.006413635797798634, + 0.011188600212335587, + 0.003951275255531073, + 0.013603897765278816, + 0.0033853596542030573, + 0.01772482879459858, + 0.011007077060639858, + 0.004985863342881203, + 0.004884787835180759, + 0.0032051282469183207, + 0.0023645972833037376, + 0.0045792353339493275, + 0.0024252322036772966, + 0.0028853381518274546, + 0.0052603064104914665, + 0.0025505078956484795, + 0.0022708019241690636, + 0.02053661085665226, + 0.011702079325914383, + 0.038823094218969345, + 0.019330520182847977, + 0.02505190670490265, + 0.02059101313352585, + 0.018877942115068436, + 0.013622768223285675, + 0.0054966299794614315, + 0.001802058657631278, + 0.004261263180524111, + 0.002002965658903122, + 0.0025093997828662395, + 0.003629248356446624, + 0.0027048802003264427, + 0.003675634739920497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.6933619976043701 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "dynamic, side-scrolling medium shot, rendered in a fluid 2D animation style, fast-tracks a young elf astride a majestic unicorn through an enchanted forest. The camera moves from left to right, maintaining perfect framing as the two race through dense foliage. The elf, with delicately pointed ears and wide, joyful eyes, leans forward in the saddle, her long, shimmering silver hair flowing dramatically behind her like a silk banner in the wind. Clad in a tunic woven from emerald green leaves and simple brown leather bracers, they clutch the unicorn's thick, flowing mane. The unicorn itself is a creature of pure magic, its fur a pearly white that seems to shimmer, its powerful legs running with swiftness and grace. Its mane and tail are a vibrant array of pastel colors, and its single, spiraling horn emits a soft inner glow. The environment is a symphony of vibrant, saturated color; ancient, gnarled trees with moss-covered bark form a dense canopy, above which hang magical, glowing orbs of light in brilliant shades of pink, teal, and gold, casting a dappled, ethereal glow across the scene. The forest floor is a carpet of soft moss and luminescent fungi, and rapid movement creates a beautiful parallax effect, with foreground ferns and flowers blurring past, enhancing the thrilling sense of speed and creating an atmosphere of pure, unadulterated magical adventure.", + "seed": 30892, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_2/index_7348_seed_30892_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10674127899110317, + "vlm_logprob": -0.10674127899110317, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.10674127899110317, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020666619762778282 + ], + "opposite_logprob_tokens": [ + -4.020666599273682 + ], + "logprob_reward": -0.020666619762778282, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.020666619762778282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19579263031482697, + 0.041198424994945526, + 0.04466885328292847, + 0.09218043833971024, + 0.25339463353157043, + 0.03823750093579292, + 0.043233610689640045, + 0.056250039488077164, + 0.039374202489852905, + 0.19566965103149414 + ], + "credit_map": [ + 0.03642585873603821, + 0.02541293576359749, + 0.017276788130402565, + 0.016294313594698906, + 0.021511930972337723, + 0.02992783673107624, + 0.02320084162056446, + 0.025742148980498314, + 0.005339359864592552, + 0.00374132813885808, + 0.004843864589929581, + 0.004033676348626614, + 0.0057710618712008, + 0.004326558671891689, + 0.007220925763249397, + 0.005921653006225824, + 0.0033841365948319435, + 0.0023895122576504946, + 0.0038066573906689882, + 0.0027844123542308807, + 0.004524580202996731, + 0.005059681832790375, + 0.010534033179283142, + 0.012185840867459774, + 0.015020682476460934, + 0.015326241962611675, + 0.007950197905302048, + 0.008051738142967224, + 0.012398581020534039, + 0.01323152706027031, + 0.009169720113277435, + 0.011031757108867168, + 0.09016069769859314, + 0.07741308212280273, + 0.013384418562054634, + 0.012438418343663216, + 0.01563461869955063, + 0.014680696651339531, + 0.01159533392637968, + 0.018087392672896385, + 0.004855378530919552, + 0.0053144111298024654, + 0.0039426414296031, + 0.006666143890470266, + 0.0038754865527153015, + 0.0036884741857647896, + 0.004882052540779114, + 0.005012915004044771, + 0.004675534553825855, + 0.004652891308069229, + 0.005418167915195227, + 0.005145968869328499, + 0.00972861610352993, + 0.004847186151891947, + 0.004772139713168144, + 0.003993107471615076, + 0.006936392281204462, + 0.011843805201351643, + 0.005870304070413113, + 0.006128182169049978, + 0.00636643823236227, + 0.006134635303169489, + 0.0047511449083685875, + 0.008219145238399506, + 0.004998878110200167, + 0.004784841556102037, + 0.00629826495423913, + 0.0037950854748487473, + 0.004097058437764645, + 0.004092680290341377, + 0.005795527249574661, + 0.0055118706077337265, + 0.014095256105065346, + 0.016343673691153526, + 0.012722847051918507, + 0.03404822200536728, + 0.02969878911972046, + 0.04766608774662018, + 0.023572906851768494, + 0.017521874979138374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.020666619762778282 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361985132098198 + ], + "opposite_logprob_tokens": [ + -3.6513619422912598 + ], + "logprob_reward": -0.026361985132098198, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361985132098198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1669854074716568, + 0.11312044411897659, + 0.07220849394798279, + 0.06780682504177094, + 0.15704889595508575, + 0.07569784671068192, + 0.05783194303512573, + 0.08266760408878326, + 0.08873888850212097, + 0.11789362877607346 + ], + "credit_map": [ + 0.025308620184659958, + 0.019273502752184868, + 0.026053281500935555, + 0.015476701781153679, + 0.02161586470901966, + 0.018511472269892693, + 0.020602261647582054, + 0.02014370821416378, + 0.014126974157989025, + 0.01198946125805378, + 0.017213033512234688, + 0.009801208972930908, + 0.014688028953969479, + 0.013350714929401875, + 0.013632213696837425, + 0.018318811431527138, + 0.006051435135304928, + 0.004885328002274036, + 0.005518526770174503, + 0.0044357734732329845, + 0.007332495413720608, + 0.011446340009570122, + 0.013871180824935436, + 0.018667414784431458, + 0.008418912999331951, + 0.007153937593102455, + 0.005711844190955162, + 0.007309319917112589, + 0.010107840411365032, + 0.007308612111955881, + 0.010254127904772758, + 0.01154223084449768, + 0.05103408917784691, + 0.036413680762052536, + 0.015521908178925514, + 0.006335522513836622, + 0.015946172177791595, + 0.009829146787524223, + 0.007643379271030426, + 0.014325002208352089, + 0.010099503211677074, + 0.012080674059689045, + 0.008599556051194668, + 0.012188836932182312, + 0.008319562301039696, + 0.007054021582007408, + 0.009358322247862816, + 0.00799737311899662, + 0.014693167991936207, + 0.006730619817972183, + 0.0065555935725569725, + 0.0050537013448774815, + 0.005148503463715315, + 0.008112762123346329, + 0.004959671758115292, + 0.006577921565622091, + 0.011025975458323956, + 0.01709304191172123, + 0.009294440969824791, + 0.008151169866323471, + 0.006908442825078964, + 0.010441558435559273, + 0.007662069518119097, + 0.012090903706848621, + 0.012828117236495018, + 0.008418329991400242, + 0.007855450734496117, + 0.010663666762411594, + 0.012305695563554764, + 0.010509965941309929, + 0.015865754336118698, + 0.010291903279721737, + 0.01190642174333334, + 0.01007082685828209, + 0.007018400821834803, + 0.019753480330109596, + 0.017640547826886177, + 0.03134772181510925, + 0.008614148944616318, + 0.011542081832885742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.026361985132098198 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013375869020819664 + ], + "opposite_logprob_tokens": [ + -6.62633752822876 + ], + "logprob_reward": -0.0013375869020819664, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013375869020819664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14220410585403442, + 0.037684306502342224, + 0.0672842338681221, + 0.06358896195888519, + 0.3932211399078369, + 0.07580782473087311, + 0.0864759311079979, + 0.03135381266474724, + 0.02961055003106594, + 0.07276912778615952 + ], + "credit_map": [ + 0.02484390325844288, + 0.01874002441763878, + 0.012621055357158184, + 0.010708739049732685, + 0.015309318900108337, + 0.025450367480516434, + 0.01445284578949213, + 0.020077846944332123, + 0.004079420585185289, + 0.0037102100905030966, + 0.003944865427911282, + 0.004297764040529728, + 0.00398635771125555, + 0.0067931851372122765, + 0.006269738543778658, + 0.004602766130119562, + 0.004924905486404896, + 0.005682267714291811, + 0.0033551689703017473, + 0.005191373638808727, + 0.006368428003042936, + 0.008225436322391033, + 0.011003928259015083, + 0.02253272384405136, + 0.013934032991528511, + 0.008213344030082226, + 0.006487051956355572, + 0.005302226170897484, + 0.007796312216669321, + 0.004976435098797083, + 0.007815921679139137, + 0.009063635021448135, + 0.1378968060016632, + 0.1283848136663437, + 0.023645862936973572, + 0.016544533893465996, + 0.026516113430261612, + 0.021411623805761337, + 0.015871165320277214, + 0.022950245067477226, + 0.010122662410140038, + 0.005757926031947136, + 0.007184577640146017, + 0.014262073673307896, + 0.007318489253520966, + 0.00630340538918972, + 0.01374227087944746, + 0.011116420850157738, + 0.01899927668273449, + 0.010432897135615349, + 0.009475871920585632, + 0.00671448465436697, + 0.014384589157998562, + 0.0071744476445019245, + 0.008326773531734943, + 0.01096759270876646, + 0.0037996512837707996, + 0.006069997325539589, + 0.0036283521912992, + 0.0037238565273582935, + 0.0033360410016030073, + 0.0031709589529782534, + 0.0035481280647218227, + 0.004076826851814985, + 0.004031185060739517, + 0.005472650285810232, + 0.00395579170435667, + 0.00430084764957428, + 0.0026340314652770758, + 0.002739086514338851, + 0.003099371213465929, + 0.003377585904672742, + 0.005955659784376621, + 0.006506298668682575, + 0.005131132435053587, + 0.011862426996231079, + 0.010385471396148205, + 0.016672996804118156, + 0.008537964895367622, + 0.0077171726152300835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.0013375869020819664 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071558110415936 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071558110415936, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071558110415936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21163657307624817, + 0.08935555070638657, + 0.06608141213655472, + 0.12135031819343567, + 0.07017306238412857, + 0.04559740051627159, + 0.08723922818899155, + 0.0727323666214943, + 0.07044146955013275, + 0.16539259254932404 + ], + "credit_map": [ + 0.039839617908000946, + 0.02569088526070118, + 0.021609900519251823, + 0.0148885203525424, + 0.02293936349451542, + 0.037046853452920914, + 0.021676065400242805, + 0.027945352718234062, + 0.01215369626879692, + 0.009122887626290321, + 0.007983560673892498, + 0.007835334166884422, + 0.010056337341666222, + 0.020722458139061928, + 0.008821227587759495, + 0.012660046108067036, + 0.005573345348238945, + 0.007014373317360878, + 0.003282577032223344, + 0.00610523484647274, + 0.005196158308535814, + 0.005202010273933411, + 0.015686577185988426, + 0.01802114024758339, + 0.014341628178954124, + 0.010617854073643684, + 0.0084606371819973, + 0.008181341923773289, + 0.014868210069835186, + 0.018222855404019356, + 0.020947866141796112, + 0.025709932669997215, + 0.011932003311812878, + 0.023798953741788864, + 0.003909124061465263, + 0.0037501694168895483, + 0.005413061007857323, + 0.005529158748686314, + 0.003808752167969942, + 0.012031836435198784, + 0.00600601127371192, + 0.005414627958089113, + 0.004798729903995991, + 0.006697488948702812, + 0.007005929946899414, + 0.003854123642668128, + 0.0066032023169100285, + 0.005217287223786116, + 0.015378213487565517, + 0.009928231127560139, + 0.007748379837721586, + 0.008893960155546665, + 0.01348335575312376, + 0.009336242452263832, + 0.0143280616030097, + 0.00814278144389391, + 0.008788465522229671, + 0.023067103698849678, + 0.006660969462245703, + 0.005727275274693966, + 0.005705190356820822, + 0.006476008333265781, + 0.004827241413295269, + 0.011480111628770828, + 0.009987885132431984, + 0.00796337891370058, + 0.007504013366997242, + 0.012439790181815624, + 0.008373637683689594, + 0.004502798896282911, + 0.010939444415271282, + 0.008730522356927395, + 0.01146671548485756, + 0.012699104845523834, + 0.008604403585195541, + 0.03897494450211525, + 0.01616089791059494, + 0.052673354744911194, + 0.012162914499640465, + 0.0126502625644207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.011071558110415936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742686450481415 + ], + "opposite_logprob_tokens": [ + -0.9742686748504639 + ], + "logprob_reward": -0.4742686450481415, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742686450481415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10296981781721115, + 0.031998950988054276, + 0.07860378175973892, + 0.051325201988220215, + 0.4819183945655823, + 0.04843498021364212, + 0.0560242235660553, + 0.03798218071460724, + 0.06254841387271881, + 0.04819408059120178 + ], + "credit_map": [ + 0.012973503209650517, + 0.011533566750586033, + 0.011196400970220566, + 0.009833789430558681, + 0.01764247938990593, + 0.016194812953472137, + 0.013137733563780785, + 0.010457533411681652, + 0.004800525028258562, + 0.0036391678731888533, + 0.0035687144845724106, + 0.003234842559322715, + 0.004528528545051813, + 0.004440269432961941, + 0.003588307648897171, + 0.004198594484478235, + 0.005541376769542694, + 0.005431261844933033, + 0.0044299280270934105, + 0.005002111196517944, + 0.0058181798085570335, + 0.007962480187416077, + 0.02447570487856865, + 0.019942736253142357, + 0.0056761340238153934, + 0.0060407803393900394, + 0.004661460407078266, + 0.005907166283577681, + 0.00795073714107275, + 0.00554380239918828, + 0.005935097578912973, + 0.009610023349523544, + 0.1740560233592987, + 0.15894486010074615, + 0.026802908629179, + 0.01999799720942974, + 0.028539609163999557, + 0.024809427559375763, + 0.01865006610751152, + 0.03011750802397728, + 0.007199807558208704, + 0.005181432235985994, + 0.004734031390398741, + 0.008210097439587116, + 0.006009304895997047, + 0.00451672775670886, + 0.0070037152618169785, + 0.005579867400228977, + 0.007812174502760172, + 0.00610214751213789, + 0.005496655125170946, + 0.007298202719539404, + 0.005741377826780081, + 0.00871710479259491, + 0.00855854619294405, + 0.00629801582545042, + 0.005782504100352526, + 0.01200750470161438, + 0.003986448980867863, + 0.0026766336522996426, + 0.002542592119425535, + 0.003337981877848506, + 0.002438809722661972, + 0.005209704861044884, + 0.009346164762973785, + 0.007417188957333565, + 0.00804216880351305, + 0.009508018381893635, + 0.007447864860296249, + 0.0049959574826061726, + 0.009056705981492996, + 0.006734343711286783, + 0.004103942308574915, + 0.004324015229940414, + 0.0029623243026435375, + 0.0077886818908154964, + 0.008815652690827847, + 0.008801526390016079, + 0.004845843650400639, + 0.00655209319666028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.4742686450481415 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, side-scrolling medium shot, rendered in a fluid 2D animation style, fast-tracks a young elf astride a majestic unicorn through an enchanted forest. The camera moves from left to right, maintaining perfect framing as the two race through dense foliage. The elf, with delicately pointed ears and wide, joyful eyes, leans forward in the saddle, her long, shimmering silver hair flowing dramatically behind her like a silk banner in the wind. Clad in a tunic woven from emerald green leaves and simple brown leather bracers, they clutch the unicorn's thick, flowing mane. The unicorn itself is a creature of pure magic, its fur a pearly white that seems to shimmer, its powerful legs running with swiftness and grace. Its mane and tail are a vibrant array of pastel colors, and its single, spiraling horn emits a soft inner glow. The environment is a symphony of vibrant, saturated color; ancient, gnarled trees with moss-covered bark form a dense canopy, above which hang magical, glowing orbs of light in brilliant shades of pink, teal, and gold, casting a dappled, ethereal glow across the scene. The forest floor is a carpet of soft moss and luminescent fungi, and rapid movement creates a beautiful parallax effect, with foreground ferns and flowers blurring past, enhancing the thrilling sense of speed and creating an atmosphere of pure, unadulterated magical adventure.", + "seed": 30893, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_2/index_7348_seed_30893_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13202744459267707, + "vlm_logprob": -0.13202744459267707, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.13202744459267707, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025204651057720184 + ], + "opposite_logprob_tokens": [ + -3.900204658508301 + ], + "logprob_reward": -0.025204651057720184, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025204651057720184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1643231213092804, + 0.1296498328447342, + 0.039606742560863495, + 0.14796534180641174, + 0.1151139959692955, + 0.03654550388455391, + 0.042316243052482605, + 0.2307765632867813, + 0.040196169167757034, + 0.053506482392549515 + ], + "credit_map": [ + 0.020401373505592346, + 0.017618326470255852, + 0.012627290561795235, + 0.02076343633234501, + 0.03216709941625595, + 0.022140787914395332, + 0.022529080510139465, + 0.016075728461146355, + 0.01747303456068039, + 0.019474128261208534, + 0.009708947502076626, + 0.019243622198700905, + 0.014421365223824978, + 0.02030133455991745, + 0.01104120071977377, + 0.017986197024583817, + 0.005545645020902157, + 0.005377981346100569, + 0.0037831629160791636, + 0.0066018844954669476, + 0.004204548895359039, + 0.004448866005986929, + 0.003626137739047408, + 0.006018516607582569, + 0.013829363510012627, + 0.04391246289014816, + 0.010087217204272747, + 0.022502515465021133, + 0.010875830426812172, + 0.02191704325377941, + 0.014734133146703243, + 0.01010676845908165, + 0.03618757799267769, + 0.007036047987639904, + 0.004889458883553743, + 0.0322691835463047, + 0.0059973206371068954, + 0.013535624369978905, + 0.008883836679160595, + 0.006314950063824654, + 0.0049545918591320515, + 0.006416907999664545, + 0.003783032763749361, + 0.004916962701827288, + 0.002833204809576273, + 0.0044248951599001884, + 0.005133758764714003, + 0.004082150291651487, + 0.005900464486330748, + 0.003493092255666852, + 0.005580566823482513, + 0.007576651871204376, + 0.004726777784526348, + 0.0038090755697339773, + 0.00648586405441165, + 0.004743747878819704, + 0.04647141695022583, + 0.043275944888591766, + 0.020199615508317947, + 0.02979023940861225, + 0.019137883558869362, + 0.018546098843216896, + 0.02167578600347042, + 0.03167957812547684, + 0.00531170517206192, + 0.005201239138841629, + 0.003913159016519785, + 0.0058371564373373985, + 0.0035918850917369127, + 0.006079590413719416, + 0.004947143606841564, + 0.0053142886608839035, + 0.006522650830447674, + 0.007945551536977291, + 0.005345936398953199, + 0.007859201170504093, + 0.005222537089139223, + 0.005601207260042429, + 0.009671535342931747, + 0.005337863694876432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.025204651057720184 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309285804629326 + ], + "opposite_logprob_tokens": [ + -3.168092966079712 + ], + "logprob_reward": -0.04309285804629326, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309285804629326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19871561229228973, + 0.10547253489494324, + 0.023973383009433746, + 0.07562549412250519, + 0.31848692893981934, + 0.09098174422979355, + 0.03227341175079346, + 0.07800028473138809, + 0.01722373068332672, + 0.05924684926867485 + ], + "credit_map": [ + 0.028119102120399475, + 0.02509039267897606, + 0.01983899064362049, + 0.027115024626255035, + 0.031822942197322845, + 0.023222099989652634, + 0.0175912007689476, + 0.02591584250330925, + 0.016020091250538826, + 0.01203977782279253, + 0.009729518555104733, + 0.011585591360926628, + 0.01165839284658432, + 0.02001972869038582, + 0.007821375504136086, + 0.01659805327653885, + 0.0034924389328807592, + 0.0027235376182943583, + 0.0023712608963251114, + 0.003776612225919962, + 0.0024689435958862305, + 0.0029093066696077585, + 0.0024832971394062042, + 0.0037479877937585115, + 0.011941616423428059, + 0.00833118986338377, + 0.008263534866273403, + 0.016610750928521156, + 0.006161881145089865, + 0.011860768310725689, + 0.006829812191426754, + 0.005625947378575802, + 0.11579274386167526, + 0.01307642925530672, + 0.008779686875641346, + 0.09612955898046494, + 0.01120053231716156, + 0.03857474774122238, + 0.02314835973083973, + 0.011784866452217102, + 0.015623467974364758, + 0.03380201384425163, + 0.008411780931055546, + 0.005270806141197681, + 0.009023169055581093, + 0.007000968791544437, + 0.003927784971892834, + 0.007921750657260418, + 0.005288159474730492, + 0.0036922143772244453, + 0.004643085412681103, + 0.0037839366123080254, + 0.003692865138873458, + 0.004196548834443092, + 0.003039336996152997, + 0.003937266301363707, + 0.01302663516253233, + 0.013181040994822979, + 0.00884042214602232, + 0.008475917391479015, + 0.004953124560415745, + 0.008495199494063854, + 0.0066545503214001656, + 0.014373399317264557, + 0.0021494023967534304, + 0.0023430297151207924, + 0.0018104659393429756, + 0.0019858244340866804, + 0.002273386809974909, + 0.0023330466356128454, + 0.002438824623823166, + 0.0018897510599344969, + 0.007273725233972073, + 0.008183910511434078, + 0.007870234549045563, + 0.006178080104291439, + 0.006413206458091736, + 0.006695553660392761, + 0.009477144107222557, + 0.00715499697253108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.04309285804629326 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015163839561864734 + ], + "opposite_logprob_tokens": [ + -6.501516342163086 + ], + "logprob_reward": -0.0015163839561864734, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015163839561864734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2943018674850464, + 0.05570667237043381, + 0.033665336668491364, + 0.1690727174282074, + 0.06253405660390854, + 0.09516800194978714, + 0.0828414261341095, + 0.11539709568023682, + 0.0279651191085577, + 0.06334764510393143 + ], + "credit_map": [ + 0.04381484165787697, + 0.03195825591683388, + 0.02364029549062252, + 0.03585539013147354, + 0.06835050880908966, + 0.02872168831527233, + 0.03394860774278641, + 0.02801230177283287, + 0.0070334142073988914, + 0.00684457877650857, + 0.005253124050796032, + 0.009416530840098858, + 0.006816253997385502, + 0.008541170507669449, + 0.004937129095196724, + 0.006864472292363644, + 0.005040549673140049, + 0.0030324431136250496, + 0.002582674380391836, + 0.00440961541607976, + 0.0067818970419466496, + 0.003917313180863857, + 0.00280326628126204, + 0.005097582004964352, + 0.015167387202382088, + 0.04361294209957123, + 0.011224058456718922, + 0.03274506330490112, + 0.019627399742603302, + 0.023852121084928513, + 0.01188402809202671, + 0.010959722101688385, + 0.019002269953489304, + 0.0031453431583940983, + 0.002427904400974512, + 0.02117358148097992, + 0.003550933441147208, + 0.004371258895844221, + 0.005908309016376734, + 0.0029544588178396225, + 0.020548351109027863, + 0.03640908747911453, + 0.007888480089604855, + 0.004461419302970171, + 0.007137438748031855, + 0.006340836174786091, + 0.003770631505176425, + 0.008611761964857578, + 0.01236318051815033, + 0.009451604448258877, + 0.006423426326364279, + 0.01157711073756218, + 0.010113604366779327, + 0.012060444802045822, + 0.011531873606145382, + 0.009320186451077461, + 0.017700595781207085, + 0.011470003053545952, + 0.012327468954026699, + 0.012020621448755264, + 0.015000768005847931, + 0.014080578461289406, + 0.009276537224650383, + 0.02352052368223667, + 0.003839806653559208, + 0.004176572430878878, + 0.002601020270958543, + 0.0033303287345916033, + 0.003858691779896617, + 0.0031776006799191236, + 0.003215085482224822, + 0.0037660158704966307, + 0.01088014431297779, + 0.007153086364269257, + 0.009668804705142975, + 0.009153619408607483, + 0.005053801462054253, + 0.00750884972512722, + 0.0068697938695549965, + 0.007059547118842602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0015163839561864734 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193654991686344 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193654991686344, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193654991686344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4043378531932831, + 0.035196032375097275, + 0.01647077314555645, + 0.07662762701511383, + 0.21995913982391357, + 0.06460195779800415, + 0.043569400906562805, + 0.07881224900484085, + 0.019551865756511688, + 0.04087304696440697 + ], + "credit_map": [ + 0.060379207134246826, + 0.04749617725610733, + 0.03309357166290283, + 0.04748687893152237, + 0.10393961519002914, + 0.0404973030090332, + 0.03322100266814232, + 0.038224101066589355, + 0.004769525956362486, + 0.0037009369116276503, + 0.00296273292042315, + 0.004981903824955225, + 0.00318732182495296, + 0.006174844224005938, + 0.003731412347406149, + 0.005687353201210499, + 0.003056966233998537, + 0.0019050571136176586, + 0.0014495407231152058, + 0.002100487472489476, + 0.0014761625789105892, + 0.0029765223152935505, + 0.0012184022925794125, + 0.002287635812535882, + 0.012599010020494461, + 0.016378432512283325, + 0.005670006852596998, + 0.007316941861063242, + 0.008179835975170135, + 0.015019028447568417, + 0.005452806130051613, + 0.006011569406837225, + 0.07764336466789246, + 0.009381226263940334, + 0.0058530173264443874, + 0.07322777807712555, + 0.007561706937849522, + 0.02054179459810257, + 0.018147755414247513, + 0.007602491416037083, + 0.011299734935164452, + 0.019941553473472595, + 0.008358665741980076, + 0.003850087057799101, + 0.005385871510952711, + 0.005849326960742474, + 0.004071588162332773, + 0.005845129955559969, + 0.005189091432839632, + 0.0044471509754657745, + 0.005045049358159304, + 0.007093240972608328, + 0.005219123791903257, + 0.003571502398699522, + 0.008532133884727955, + 0.004472111817449331, + 0.011023260653018951, + 0.01008086372166872, + 0.007027598563581705, + 0.010303939692676067, + 0.009658637456595898, + 0.01021228451281786, + 0.008519523777067661, + 0.011986143887043, + 0.0029501942917704582, + 0.002266587922349572, + 0.001757810590788722, + 0.002411442343145609, + 0.001937695313245058, + 0.0031833138782531023, + 0.002200803253799677, + 0.0028440181631594896, + 0.005347155034542084, + 0.006320569198578596, + 0.00364860100671649, + 0.003436587518081069, + 0.004264127928763628, + 0.004773837514221668, + 0.007548956666141748, + 0.005533213261514902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.014193654991686344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a side-scrolling medium shot featuring a young elf riding a unicorn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the elf wearing a tunic woven from emerald green leaves and brown leather bracers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the unicorn's fur pearly white with a mane and tail in vibrant pastel colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the unicorn's single spiraling horn emit a soft inner glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move from left to right while maintaining perfect framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576129674911499 + ], + "opposite_logprob_tokens": [ + -0.826129674911499 + ], + "logprob_reward": -0.576129674911499, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576129674911499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2474696785211563, + 0.05893106758594513, + 0.03809095546603203, + 0.1198498085141182, + 0.1890484094619751, + 0.09809929132461548, + 0.07367322593927383, + 0.038618750870227814, + 0.0549875944852829, + 0.08123117685317993 + ], + "credit_map": [ + 0.03023615851998329, + 0.03284810110926628, + 0.017546409741044044, + 0.030920693650841713, + 0.05173778533935547, + 0.022927509620785713, + 0.035090599209070206, + 0.026162398979067802, + 0.007920387201011181, + 0.008972023613750935, + 0.004629206378012896, + 0.009808777831494808, + 0.005962715949863195, + 0.007239038124680519, + 0.006424010265618563, + 0.00797491054981947, + 0.006730496417731047, + 0.004328236449509859, + 0.002774958498775959, + 0.007418316323310137, + 0.003483446780592203, + 0.004050360526889563, + 0.005027967970818281, + 0.004277173429727554, + 0.012284133583307266, + 0.038797345012426376, + 0.0071968683041632175, + 0.014932427555322647, + 0.009685841389000416, + 0.02131536602973938, + 0.008783295750617981, + 0.006854526698589325, + 0.04185067489743233, + 0.008482512086629868, + 0.006621206644922495, + 0.07027097046375275, + 0.010150093585252762, + 0.025963053107261658, + 0.017317185178399086, + 0.008392705582082272, + 0.019232604652643204, + 0.03272338584065437, + 0.00872943177819252, + 0.006638731341809034, + 0.007938322611153126, + 0.007440115790814161, + 0.0071978699415922165, + 0.008198839612305164, + 0.009052962996065617, + 0.007070847321301699, + 0.008733254857361317, + 0.012303397059440613, + 0.009413636289536953, + 0.005310201086103916, + 0.01402114424854517, + 0.007767775561660528, + 0.006460193544626236, + 0.004804939962923527, + 0.0040233866311609745, + 0.004555938299745321, + 0.00392569275572896, + 0.004755535628646612, + 0.004612421616911888, + 0.00548064149916172, + 0.006771078798919916, + 0.008838365785777569, + 0.004245735239237547, + 0.006838641595095396, + 0.00682523287832737, + 0.005811311770230532, + 0.007200280204415321, + 0.008456951007246971, + 0.01195899210870266, + 0.010416746139526367, + 0.012143404223024845, + 0.01056359801441431, + 0.00632233452051878, + 0.00978010892868042, + 0.00963131058961153, + 0.010414674878120422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.576129674911499 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_3/index_tensor([182, 182])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_3/index_tensor([182, 182])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2f9abbb20cbaad569b3e682a2e2672ad742d29d2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_3/index_tensor([182, 182])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, extreme wide shot slowly advances, establishing a sprawling estate at dusk. A grand, imposing Gothic Revival mansion, constructed of dark, weathered stone and crowned with a soaring spire, rises atop a gentle hill, silhouetted against a deep indigo and violet-purple sky, the last fiery orange glow of sunset streaming along the horizon. Far below, a lone vintage luxury car, its sleek black body reflecting the fading twilight, steadily and deliberately moves down a long, winding gravel driveway. Its powerful twin headlights cut through the low-lying mist clinging to the manicured grounds, streaking sharp white beams across the gnarled, ancient trees lining the road. As the vehicle navigates the winding curves, its steady progress is the only movement in the silent landscape, its destination the distant, warmly lit windows of the mansion its only destination, creating an air of elegant isolation and a powerful atmosphere of quiet mystery.", + "seed": 31395, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_3/index_182_seed_31395_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15411190409213305, + "vlm_logprob": -0.15411190409213305, + "vlm_yes_no_margin": 2.750000011920929, + "vlm_token_credit": -0.15411190409213305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4432503581047058 + ], + "opposite_logprob_tokens": [ + -1.1932504177093506 + ], + "logprob_reward": -0.4432503581047058, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.4432503581047058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16252529621124268, + 0.03725910559296608, + 0.08415132015943527, + 0.10253389179706573, + 0.1338251382112503, + 0.07950741797685623, + 0.3449092209339142, + 0.010341818444430828, + 0.021756455302238464, + 0.023190289735794067 + ], + "credit_map": [ + 0.01767146773636341, + 0.013429985381662846, + 0.025562530383467674, + 0.020596204325556755, + 0.014812580309808254, + 0.014959678053855896, + 0.02992265671491623, + 0.025570185855031013, + 0.0036316905170679092, + 0.003507727524265647, + 0.005472245160490274, + 0.005524880718439817, + 0.005333796609193087, + 0.003627537749707699, + 0.00490781432017684, + 0.005253411363810301, + 0.009907681494951248, + 0.010816114954650402, + 0.011082470417022705, + 0.013286562636494637, + 0.007505950052291155, + 0.00566056976094842, + 0.013198227621614933, + 0.01269374880939722, + 0.010874035768210888, + 0.017006125301122665, + 0.011679861694574356, + 0.018081335350871086, + 0.010590571910142899, + 0.009563782252371311, + 0.010968499816954136, + 0.013769669458270073, + 0.017056778073310852, + 0.016756778582930565, + 0.013751567341387272, + 0.018149467185139656, + 0.01737966760993004, + 0.011181311681866646, + 0.02279817871749401, + 0.016751395538449287, + 0.00886625237762928, + 0.00538154412060976, + 0.010976742021739483, + 0.012890988029539585, + 0.008515769615769386, + 0.00828677136451006, + 0.013348561711609364, + 0.011240788735449314, + 0.04084035009145737, + 0.02715544030070305, + 0.0526026114821434, + 0.05587073415517807, + 0.022937260568141937, + 0.037080589681863785, + 0.05149300768971443, + 0.05692922696471214, + 0.0007892217836342752, + 0.0005691410042345524, + 0.0013851255644112825, + 0.0021181206684559584, + 0.0006806206074543297, + 0.0006475321133621037, + 0.002558897016569972, + 0.0015931593952700496, + 0.0036154931876808405, + 0.001342381234280765, + 0.0036189528182148933, + 0.0028524231165647507, + 0.00116634089499712, + 0.002401818288490176, + 0.002892718417569995, + 0.003866327926516533, + 0.0020798665937036276, + 0.0012941131135448813, + 0.0037434271071106195, + 0.004411313682794571, + 0.0010620934190228581, + 0.002047695452347398, + 0.004976583644747734, + 0.003575196722522378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.4432503581047058 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22563953697681427 + ], + "opposite_logprob_tokens": [ + -1.6006395816802979 + ], + "logprob_reward": -0.22563953697681427, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22563953697681427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16714167594909668, + 0.04699058458209038, + 0.041556477546691895, + 0.09018922597169876, + 0.336346834897995, + 0.026471571996808052, + 0.08300872892141342, + 0.02476510964334011, + 0.048090819269418716, + 0.1354389786720276 + ], + "credit_map": [ + 0.017060788348317146, + 0.014502705074846745, + 0.03154227137565613, + 0.027011355385184288, + 0.011147469282150269, + 0.01158386655151844, + 0.022981826215982437, + 0.031311385333538055, + 0.0054898993112146854, + 0.00287307845428586, + 0.007668065372854471, + 0.0064317905344069, + 0.004873292520642281, + 0.005557919852435589, + 0.007382663432508707, + 0.006713876500725746, + 0.003762745065614581, + 0.0051002600230276585, + 0.0042161718010902405, + 0.00515465484932065, + 0.0071580782532691956, + 0.003455106634646654, + 0.0073799388483166695, + 0.005329522304236889, + 0.010783350095152855, + 0.01846124231815338, + 0.009317643940448761, + 0.010384595952928066, + 0.011395849287509918, + 0.007221963722258806, + 0.007998429238796234, + 0.014626157470047474, + 0.039088040590286255, + 0.053128670901060104, + 0.029688436537981033, + 0.04427587240934372, + 0.040510378777980804, + 0.030376359820365906, + 0.05225476250052452, + 0.04702429473400116, + 0.0029404049273580313, + 0.002460590098053217, + 0.003194552380591631, + 0.003808887442573905, + 0.003276042640209198, + 0.003018653020262718, + 0.004245263524353504, + 0.0035271800588816404, + 0.009935569949448109, + 0.007004790939390659, + 0.013808144256472588, + 0.011789338663220406, + 0.005778837949037552, + 0.00806192122399807, + 0.014427627436816692, + 0.01220250315964222, + 0.002343865344300866, + 0.0015389272011816502, + 0.0038995135109871626, + 0.005288602318614721, + 0.001616980298422277, + 0.002468031831085682, + 0.003727307543158531, + 0.0038818796165287495, + 0.005959263537079096, + 0.0026798027101904154, + 0.008044400252401829, + 0.009890425018966198, + 0.0025779239367693663, + 0.005048183258622885, + 0.006476144772022963, + 0.007414680439978838, + 0.013755333609879017, + 0.008077537640929222, + 0.031654730439186096, + 0.026637397706508636, + 0.006721453741192818, + 0.011624012142419815, + 0.016921870410442352, + 0.02004665695130825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.22563953697681427 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036361785605549812 + ], + "opposite_logprob_tokens": [ + -5.628636360168457 + ], + "logprob_reward": -0.0036361785605549812, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036361785605549812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.261080265045166, + 0.06505788862705231, + 0.042887426912784576, + 0.045625049620866776, + 0.3524363040924072, + 0.03481176495552063, + 0.06865434348583221, + 0.06360342353582382, + 0.03514806926250458, + 0.030695509165525436 + ], + "credit_map": [ + 0.03461940214037895, + 0.02412971667945385, + 0.03589831665158272, + 0.04253256693482399, + 0.0306532122194767, + 0.02127663977444172, + 0.02934889681637287, + 0.04262150451540947, + 0.00846293568611145, + 0.010125299915671349, + 0.008184745907783508, + 0.008800501003861427, + 0.007688689511269331, + 0.005104340612888336, + 0.007710735779255629, + 0.00898064486682415, + 0.0049414378590881824, + 0.003989837598055601, + 0.0061681815423071384, + 0.008793790824711323, + 0.0032730710227042437, + 0.002889340277761221, + 0.00579865975305438, + 0.007033109199255705, + 0.005276130978018045, + 0.00838868785649538, + 0.004249989986419678, + 0.006619843654334545, + 0.005004105623811483, + 0.004523551091551781, + 0.005094383377581835, + 0.006468359846621752, + 0.03853027522563934, + 0.05985519289970398, + 0.03137288615107536, + 0.04931572452187538, + 0.04416753724217415, + 0.03512606769800186, + 0.04413338378071785, + 0.04993520677089691, + 0.003124970942735672, + 0.003087775083258748, + 0.0035313095431774855, + 0.004765699617564678, + 0.005055728368461132, + 0.0034268333110958338, + 0.007346850819885731, + 0.004472595173865557, + 0.006775503978133202, + 0.006574349943548441, + 0.010682320222258568, + 0.011646361090242863, + 0.009861037135124207, + 0.006525017321109772, + 0.008374091237783432, + 0.008215661160647869, + 0.005230132024735212, + 0.006703071296215057, + 0.007077385671436787, + 0.01152236107736826, + 0.006418758071959019, + 0.00613382738083601, + 0.012150228954851627, + 0.008367658592760563, + 0.0027132269460707903, + 0.0036721406504511833, + 0.0070154485292732716, + 0.006683608517050743, + 0.0032811718992888927, + 0.002693122485652566, + 0.004668702371418476, + 0.004420647397637367, + 0.002438775496557355, + 0.002111569046974182, + 0.0066247559152543545, + 0.005195837002247572, + 0.0028473692946135998, + 0.0023899071384221315, + 0.005120603367686272, + 0.003966693766415119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.0036361785605549812 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043069109320640564 + ], + "opposite_logprob_tokens": [ + -3.1680691242218018 + ], + "logprob_reward": -0.043069109320640564, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043069109320640564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14617037773132324, + 0.05343310162425041, + 0.03784257918596268, + 0.035723697394132614, + 0.640735387802124, + 0.012230107560753822, + 0.020280903205275536, + 0.015188300982117653, + 0.010369447991251945, + 0.02802601456642151 + ], + "credit_map": [ + 0.015354975126683712, + 0.020538531243801117, + 0.020622683688998222, + 0.02550903521478176, + 0.015877394005656242, + 0.013240859843790531, + 0.016741063445806503, + 0.018285851925611496, + 0.005465222056955099, + 0.005845446605235338, + 0.009034305810928345, + 0.0064816223457455635, + 0.00466341245919466, + 0.004662238992750645, + 0.00876239687204361, + 0.008518459275364876, + 0.0030732012819498777, + 0.0038146348670125008, + 0.005378770641982555, + 0.0060276915319263935, + 0.005591649562120438, + 0.0029154131188988686, + 0.0059515247121453285, + 0.0050896937027573586, + 0.003511295421048999, + 0.005800758954137564, + 0.0037467163056135178, + 0.006618931889533997, + 0.00321760680526495, + 0.0029424161184579134, + 0.004644451662898064, + 0.0052415188401937485, + 0.07381793111562729, + 0.08929303288459778, + 0.05672864988446236, + 0.09819507598876953, + 0.08031722158193588, + 0.05242022126913071, + 0.10062854737043381, + 0.08933473378419876, + 0.001085543306544423, + 0.0010463513899594545, + 0.0021656949538737535, + 0.0020603619050234556, + 0.0016836612485349178, + 0.000977407325990498, + 0.001858850009739399, + 0.0013522370718419552, + 0.0015369978500530124, + 0.0015601830091327429, + 0.002815610496327281, + 0.0036966288462281227, + 0.00228559086099267, + 0.002075496129691601, + 0.0035687729250639677, + 0.0027416234370321035, + 0.0019450305262580514, + 0.0014946567825973034, + 0.002081900369375944, + 0.002406562212854624, + 0.000929059402551502, + 0.0017825034447014332, + 0.0024839353282004595, + 0.0020646536722779274, + 0.0013058813055977225, + 0.0008889027521945536, + 0.0016461351187899709, + 0.0016953806625679135, + 0.0008238063310272992, + 0.0009543198975734413, + 0.0014901173999533057, + 0.0015649046981707215, + 0.0030202537309378386, + 0.0021324437111616135, + 0.0034215697087347507, + 0.003961123060435057, + 0.002113038208335638, + 0.002989334985613823, + 0.006381835322827101, + 0.004006417002528906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.043069109320640564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549643374979496 + ], + "opposite_logprob_tokens": [ + -2.929964303970337 + ], + "logprob_reward": -0.0549643374979496, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549643374979496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5354705452919006, + 0.15274102985858917, + 0.02192632667720318, + 0.018171152099967003, + 0.13018453121185303, + 0.020260123535990715, + 0.03528453782200813, + 0.03626750037074089, + 0.02842279151082039, + 0.02127152681350708 + ], + "credit_map": [ + 0.06682384759187698, + 0.07366430014371872, + 0.0760471299290657, + 0.07081577181816101, + 0.05154713988304138, + 0.04852892458438873, + 0.06411266326904297, + 0.08393070846796036, + 0.015175036154687405, + 0.016419323161244392, + 0.023918410763144493, + 0.02027234062552452, + 0.01454851496964693, + 0.01348000206053257, + 0.026425855234265327, + 0.022501543164253235, + 0.002368568442761898, + 0.002077138749882579, + 0.003369170008227229, + 0.004128443542867899, + 0.00213692057877779, + 0.0014144539600238204, + 0.0028079990297555923, + 0.0036236324813216925, + 0.0019854262936860323, + 0.0027221967466175556, + 0.0020955728832632303, + 0.0027217890601605177, + 0.002023137640208006, + 0.001676528831012547, + 0.0023385598324239254, + 0.002607941161841154, + 0.013295026496052742, + 0.013169830664992332, + 0.01206851564347744, + 0.018050890415906906, + 0.013304885476827621, + 0.008987884037196636, + 0.03366516903042793, + 0.017642328515648842, + 0.001787340035662055, + 0.0020893123000860214, + 0.0017123189754784107, + 0.003474821336567402, + 0.002607520669698715, + 0.0020459641236811876, + 0.0038060734514147043, + 0.002736772643402219, + 0.0031177655328065157, + 0.002899304497987032, + 0.005222846753895283, + 0.006255864165723324, + 0.0033311103470623493, + 0.0032879537902772427, + 0.006615460850298405, + 0.004554232582449913, + 0.003798206802457571, + 0.0030314831528812647, + 0.005687188822776079, + 0.006553162820637226, + 0.002394207753241062, + 0.004734494257718325, + 0.005005865357816219, + 0.005062890239059925, + 0.003472021082416177, + 0.002974178409203887, + 0.004877125844359398, + 0.004900227766484022, + 0.0025133758317679167, + 0.002424761652946472, + 0.003415921004489064, + 0.00384517852216959, + 0.0018398801330477, + 0.0015260973013937473, + 0.00351505633443594, + 0.003993290942162275, + 0.0013609692687168717, + 0.0019637125078588724, + 0.003812903305515647, + 0.003259616205468774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0549643374979496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, extreme wide shot slowly advances, establishing a sprawling estate at dusk. A grand, imposing Gothic Revival mansion, constructed of dark, weathered stone and crowned with a soaring spire, rises atop a gentle hill, silhouetted against a deep indigo and violet-purple sky, the last fiery orange glow of sunset streaming along the horizon. Far below, a lone vintage luxury car, its sleek black body reflecting the fading twilight, steadily and deliberately moves down a long, winding gravel driveway. Its powerful twin headlights cut through the low-lying mist clinging to the manicured grounds, streaking sharp white beams across the gnarled, ancient trees lining the road. As the vehicle navigates the winding curves, its steady progress is the only movement in the silent landscape, its destination the distant, warmly lit windows of the mansion its only destination, creating an air of elegant isolation and a powerful atmosphere of quiet mystery.", + "seed": 31396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_3/index_182_seed_31396_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08389490656554699, + "vlm_logprob": -0.08389490656554699, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.08389490656554699, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1320556402206421 + ], + "opposite_logprob_tokens": [ + -2.6320557594299316 + ], + "logprob_reward": -0.1320556402206421, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.1320556402206421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.500789999961853, + 0.1947459876537323, + 0.04847563058137894, + 0.012677300721406937, + 0.012315778061747551, + 0.011532265692949295, + 0.011463632807135582, + 0.0496920570731163, + 0.15479138493537903, + 0.003515914548188448 + ], + "credit_map": [ + 0.04229004308581352, + 0.031059425324201584, + 0.18278926610946655, + 0.022274846211075783, + 0.006701160222291946, + 0.012511747889220715, + 0.01744847185909748, + 0.18571506440639496, + 0.010772692039608955, + 0.00804655347019434, + 0.07467004656791687, + 0.007936536334455013, + 0.0027485177852213383, + 0.005329592619091272, + 0.007087663747370243, + 0.07815439254045486, + 0.0034970410633832216, + 0.0026759356260299683, + 0.014377213083207607, + 0.0032889889553189278, + 0.002616513753309846, + 0.0017533347709104419, + 0.0032861395739018917, + 0.01698046177625656, + 0.001060982933267951, + 0.0007007929380051792, + 0.0023761573247611523, + 0.002273246645927429, + 0.001256741932593286, + 0.00071776716504246, + 0.0023520435206592083, + 0.0019395683193579316, + 0.0016766428016126156, + 0.0013317022239789367, + 0.0015740537783131003, + 0.0017902692779898643, + 0.0013312690425664186, + 0.0008539801347069442, + 0.002170610474422574, + 0.0015872500371187925, + 0.0010191031033173203, + 0.0005991688231006265, + 0.0015326872235164046, + 0.002224290044978261, + 0.000999911455437541, + 0.000565261289011687, + 0.002853471552953124, + 0.0017383723752573133, + 0.0014466920401901007, + 0.0008925616275519133, + 0.001539971912279725, + 0.0023537357337772846, + 0.0011649865191429853, + 0.0008713786373846233, + 0.001475182711146772, + 0.0017191238002851605, + 0.004236346576362848, + 0.0026225498877465725, + 0.009423271752893925, + 0.016135824844241142, + 0.0015773990890011191, + 0.002171810483559966, + 0.006031589116901159, + 0.007493266835808754, + 0.03890965133905411, + 0.022757619619369507, + 0.017232898622751236, + 0.023152336478233337, + 0.008916020393371582, + 0.010295569896697998, + 0.015836386010050774, + 0.017690900713205338, + 0.00035116152139380574, + 0.00031526514794677496, + 0.00042149616638198495, + 0.0008234716369770467, + 0.00021155233844183385, + 0.0002632435644045472, + 0.0006141290650703013, + 0.0005155953695066273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.1320556402206421 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559672594070435 + ], + "opposite_logprob_tokens": [ + -1.6005966663360596 + ], + "logprob_reward": -0.22559672594070435, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559672594070435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8664577007293701, + 0.05259004235267639, + 0.03356516361236572, + 0.0047054472379386425, + 0.004674971103668213, + 0.011755414307117462, + 0.006108131259679794, + 0.010409202426671982, + 0.004315723665058613, + 0.0054181767627596855 + ], + "credit_map": [ + 0.051965925842523575, + 0.04442410171031952, + 0.3359173536300659, + 0.035551927983760834, + 0.01388746127486229, + 0.01967320404946804, + 0.027044227346777916, + 0.3379935026168823, + 0.0031534205190837383, + 0.0021121185272932053, + 0.019078223034739494, + 0.0020464444532990456, + 0.0007725959876552224, + 0.0016483968356624246, + 0.002597937360405922, + 0.021180903539061546, + 0.0018578211311250925, + 0.0021185444202274084, + 0.010532970540225506, + 0.0018148146336898208, + 0.0010542531963437796, + 0.001241919002495706, + 0.002584443660452962, + 0.012360393069684505, + 0.0005437782965600491, + 0.00041253017843700945, + 0.0007095468463376164, + 0.00061478913994506, + 0.0004319651343394071, + 0.0003091053804382682, + 0.0008229631348513067, + 0.0008607687777839601, + 0.00039873196510598063, + 0.0002985378378070891, + 0.0007539037615060806, + 0.0005261826445348561, + 0.0005727892857976258, + 0.0003538183809723705, + 0.0008004651172086596, + 0.0009705417905934155, + 0.0014568495098501444, + 0.0006090353126637638, + 0.002173530636355281, + 0.0019290470518171787, + 0.000880945532117039, + 0.0008665774948894978, + 0.0021064665634185076, + 0.0017329617403447628, + 0.0005756390164606273, + 0.00034560952917672694, + 0.0007750540971755981, + 0.001143086701631546, + 0.0006524851196445525, + 0.0005122775328345597, + 0.0012605225201696157, + 0.0008434561314061284, + 0.0007599773816764355, + 0.0005032002227380872, + 0.0012782845878973603, + 0.0025536397006362677, + 0.0005137561238370836, + 0.0004530051664914936, + 0.0028001698665320873, + 0.0015471683582291007, + 0.0007122877286747098, + 0.00032583012944087386, + 0.0003975426370743662, + 0.0006395255913957953, + 0.000392733869375661, + 0.00043989805271849036, + 0.0008016243809834123, + 0.0006062811589799821, + 0.0006545937503688037, + 0.0005356293404474854, + 0.0006587494863197207, + 0.00110585056245327, + 0.0005009096348658204, + 0.0004688097396865487, + 0.0008210336090996861, + 0.0006726001156494021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.22559672594070435 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007647404447197914 + ], + "opposite_logprob_tokens": [ + -4.882647514343262 + ], + "logprob_reward": -0.007647404447197914, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647404447197914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5605947375297546, + 0.27438071370124817, + 0.022912869229912758, + 0.012232773005962372, + 0.009847532026469707, + 0.010898401029407978, + 0.011561782099306583, + 0.018962394446134567, + 0.03176191821694374, + 0.0468469001352787 + ], + "credit_map": [ + 0.03528209775686264, + 0.02539031021296978, + 0.21708090603351593, + 0.023965662345290184, + 0.008631343953311443, + 0.014644983224570751, + 0.01800517365336418, + 0.2175942361354828, + 0.015382909215986729, + 0.016912853345274925, + 0.10138160735368729, + 0.01401606760919094, + 0.006050794385373592, + 0.007276458665728569, + 0.010330281220376492, + 0.10302972048521042, + 0.0030732653103768826, + 0.002382214181125164, + 0.003770554205402732, + 0.0028716002125293016, + 0.0014501186087727547, + 0.0022039879113435745, + 0.0028983892407268286, + 0.004262738395482302, + 0.0012949530500918627, + 0.0009891435038298368, + 0.0016131108859553933, + 0.0015268336283043027, + 0.0008242788026109338, + 0.0009294748306274414, + 0.0027495098765939474, + 0.0023054692428559065, + 0.0014991848729550838, + 0.0008307677926495671, + 0.0015055827097967267, + 0.0013747139601036906, + 0.0009606393286958337, + 0.0007385966600850224, + 0.0016055272426456213, + 0.0013325208565220237, + 0.0010965006658807397, + 0.0005685169599018991, + 0.0021479479037225246, + 0.0018712766468524933, + 0.000864634697791189, + 0.000712180626578629, + 0.0019705710001289845, + 0.0016667717136442661, + 0.0011536821257323027, + 0.0006276652566157281, + 0.0016486849635839462, + 0.0024220822378993034, + 0.0013551420997828245, + 0.0009395064553245902, + 0.0019629509188234806, + 0.0014520669355988503, + 0.002776453970000148, + 0.0014354502782225609, + 0.001971915829926729, + 0.0034245350398123264, + 0.0010696776444092393, + 0.00130016531329602, + 0.004522250033915043, + 0.002461945405229926, + 0.004558758344501257, + 0.0030721838120371103, + 0.0047796256840229034, + 0.0074623641557991505, + 0.002221273025497794, + 0.001548711210489273, + 0.004186347592622042, + 0.00393265625461936, + 0.004276076797395945, + 0.0038979453966021538, + 0.004398274701088667, + 0.005767208989709616, + 0.00606163963675499, + 0.005740099586546421, + 0.010781020857393742, + 0.005924636498093605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.007647404447197914 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607493683695793 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.01607493683695793, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607493683695793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7928986549377441, + 0.1449010819196701, + 0.01109018363058567, + 0.0028137920890003443, + 0.005358980502933264, + 0.007394363172352314, + 0.001977215288206935, + 0.00911327637732029, + 0.021559039130806923, + 0.0028933293651789427 + ], + "credit_map": [ + 0.04741249233484268, + 0.029933936893939972, + 0.3123430907726288, + 0.030814073979854584, + 0.009000887162983418, + 0.019635288044810295, + 0.024509191513061523, + 0.3192497491836548, + 0.0076458812691271305, + 0.004866156727075577, + 0.05490906164050102, + 0.0063145458698272705, + 0.0020294052083045244, + 0.004134608898311853, + 0.006367931608110666, + 0.05863349139690399, + 0.0008070077165029943, + 0.0006734680500812829, + 0.0032843847293406725, + 0.0006899205618537962, + 0.000390902190702036, + 0.0006027485360391438, + 0.0009631243883632123, + 0.0036786277778446674, + 0.00026693433756008744, + 0.00014720464241690934, + 0.0004085026739630848, + 0.00042264306102879345, + 0.000211629769182764, + 0.00032428489066660404, + 0.0005957061657682061, + 0.0004368867084849626, + 0.0007984915864653885, + 0.0004566448915284127, + 0.0009008417255245149, + 0.0006174136651679873, + 0.0004390453395899385, + 0.0003532829286996275, + 0.0011025018757209182, + 0.0006907585193403065, + 0.0009425841853953898, + 0.00043802661821246147, + 0.0012734896736219525, + 0.0012814127840101719, + 0.0005176134291104972, + 0.00047305767657235265, + 0.0013959683710709214, + 0.001072210492566228, + 0.0003407160402275622, + 0.0001653730869293213, + 0.00021605803340207785, + 0.00026478321524336934, + 0.00023144019360188395, + 0.00020801408390980214, + 0.00030395854264497757, + 0.00024687210679985583, + 0.0009983173804357648, + 0.00046227595885284245, + 0.0012859898852184415, + 0.0020436118356883526, + 0.0005432805046439171, + 0.0005905237048864365, + 0.0019388520158827305, + 0.0012504241894930601, + 0.004963370971381664, + 0.0028627633582800627, + 0.0027970082592219114, + 0.003717577550560236, + 0.001352054299786687, + 0.0015510341618210077, + 0.00189951341599226, + 0.002415715018287301, + 0.00025152848684228957, + 0.00020306429360061884, + 0.00037158402847126126, + 0.0005089526530355215, + 0.0003051407402381301, + 0.0002746924292296171, + 0.0005596479750238359, + 0.0004187188169453293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01607493683695793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038099825382232666 + ], + "opposite_logprob_tokens": [ + -3.288099765777588 + ], + "logprob_reward": -0.038099825382232666, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038099825382232666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8897257447242737, + 0.06462088227272034, + 0.009171565063297749, + 0.002523698378354311, + 0.002376983407884836, + 0.003690241137519479, + 0.0024609421379864216, + 0.006668377667665482, + 0.014642561785876751, + 0.004118945449590683 + ], + "credit_map": [ + 0.05016596242785454, + 0.03418842703104019, + 0.34965193271636963, + 0.03524499014019966, + 0.010556760244071484, + 0.021695580333471298, + 0.029025636613368988, + 0.35919657349586487, + 0.004840654321014881, + 0.004651266150176525, + 0.022678207606077194, + 0.0026521245017647743, + 0.0008765361271798611, + 0.0018956154817715287, + 0.002469210885465145, + 0.02455727569758892, + 0.0015608816174790263, + 0.0013323371531441808, + 0.0013606268912553787, + 0.0009599036420695484, + 0.000575995771214366, + 0.0005957346875220537, + 0.0013113670283928514, + 0.0014747199602425098, + 0.00021360142272897065, + 0.00018905136676039547, + 0.00044657656690105796, + 0.0003142415080219507, + 0.0001654243387747556, + 0.00017748455866239965, + 0.0005364112439565361, + 0.00048090791096910834, + 0.0002669570385478437, + 0.00015033825184218585, + 0.0004465820675250143, + 0.00040814775275066495, + 0.0001721469161566347, + 0.00020498631056398153, + 0.00038800222682766616, + 0.0003398230765014887, + 0.0005183181492611766, + 0.00023399294877890497, + 0.0005456202779896557, + 0.0005900158430449665, + 0.00028926736558787525, + 0.00030878587858751416, + 0.0006826204480603337, + 0.0005216209101490676, + 0.0002632600662764162, + 0.0001415753795299679, + 0.00041398670873604715, + 0.0005562398000620306, + 0.00021868653129786253, + 0.00022064853692427278, + 0.00031277345260605216, + 0.0003337720991112292, + 0.0004548051510937512, + 0.00031751373899169266, + 0.0008108257316052914, + 0.0016585092525929213, + 0.00024168359232135117, + 0.0003249784931540489, + 0.0018750823801383376, + 0.000984980957582593, + 0.0037877392023801804, + 0.002156621776521206, + 0.0013357815332710743, + 0.001865399070084095, + 0.0008249637321569026, + 0.0009230978321284056, + 0.0019964457023888826, + 0.0017525141593068838, + 0.0004317303537391126, + 0.00032391544664278626, + 0.0005292956484481692, + 0.001154341734945774, + 0.0002730980049818754, + 0.0003038996073883027, + 0.0005683686467818916, + 0.0005342965596355498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.038099825382232666 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, extreme wide shot slowly advances, establishing a sprawling estate at dusk. A grand, imposing Gothic Revival mansion, constructed of dark, weathered stone and crowned with a soaring spire, rises atop a gentle hill, silhouetted against a deep indigo and violet-purple sky, the last fiery orange glow of sunset streaming along the horizon. Far below, a lone vintage luxury car, its sleek black body reflecting the fading twilight, steadily and deliberately moves down a long, winding gravel driveway. Its powerful twin headlights cut through the low-lying mist clinging to the manicured grounds, streaking sharp white beams across the gnarled, ancient trees lining the road. As the vehicle navigates the winding curves, its steady progress is the only movement in the silent landscape, its destination the distant, warmly lit windows of the mansion its only destination, creating an air of elegant isolation and a powerful atmosphere of quiet mystery.", + "seed": 31397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_3/index_182_seed_31397_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1424046216532588, + "vlm_logprob": -0.1424046216532588, + "vlm_yes_no_margin": 2.499999976158142, + "vlm_token_credit": -0.1424046216532588, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17272575199604034 + ], + "opposite_logprob_tokens": [ + -2.1727256774902344 + ], + "logprob_reward": -0.17272575199604034, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.17272575199604034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27576225996017456, + 0.08088839054107666, + 0.048434752970933914, + 0.20165632665157318, + 0.11223416775465012, + 0.08421418070793152, + 0.04051104187965393, + 0.030377158895134926, + 0.04898893088102341, + 0.07693276554346085 + ], + "credit_map": [ + 0.03963695093989372, + 0.03539421409368515, + 0.045816946774721146, + 0.06193002313375473, + 0.028622465208172798, + 0.02044042944908142, + 0.019764728844165802, + 0.02415650524199009, + 0.011064664460718632, + 0.008260353468358517, + 0.013841151259839535, + 0.015666792169213295, + 0.008275678381323814, + 0.00718688266351819, + 0.0073971678502857685, + 0.00919571053236723, + 0.0052757724188268185, + 0.003707225201651454, + 0.008829123340547085, + 0.010256413370370865, + 0.004423755686730146, + 0.0057871052995324135, + 0.006148674990981817, + 0.004006687086075544, + 0.04293585196137428, + 0.017932996153831482, + 0.03672536835074425, + 0.03772333264350891, + 0.012955378741025925, + 0.015845144167542458, + 0.017157573252916336, + 0.02038070559501648, + 0.02925034612417221, + 0.009667408652603626, + 0.01824735663831234, + 0.016705965623259544, + 0.010323021560907364, + 0.008114321157336235, + 0.010098112747073174, + 0.009827651083469391, + 0.014980372041463852, + 0.006726715248078108, + 0.01259663701057434, + 0.014231834560632706, + 0.008963064290583134, + 0.009760735556483269, + 0.008856489323079586, + 0.008098342455923557, + 0.004168673884123564, + 0.002997042378410697, + 0.005731312558054924, + 0.007194988429546356, + 0.004707556217908859, + 0.006129242014139891, + 0.006019744090735912, + 0.0035624869633466005, + 0.002854906488209963, + 0.002798709785565734, + 0.005955727770924568, + 0.004159295465797186, + 0.004119235090911388, + 0.002947608008980751, + 0.004890525713562965, + 0.002651155460625887, + 0.005214706528931856, + 0.002965420251712203, + 0.011227930895984173, + 0.009070401079952717, + 0.002716808347031474, + 0.004437181632965803, + 0.009665925987064838, + 0.0036905622109770775, + 0.006047590635716915, + 0.004370477516204119, + 0.012469119392335415, + 0.014146063476800919, + 0.0064493813551962376, + 0.007669462822377682, + 0.01846817135810852, + 0.007312504108995199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.17272575199604034 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38712409138679504 + ], + "opposite_logprob_tokens": [ + -1.1371240615844727 + ], + "logprob_reward": -0.38712409138679504, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38712409138679504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21844348311424255, + 0.05421939119696617, + 0.09108949452638626, + 0.039485663175582886, + 0.060945577919483185, + 0.13629235327243805, + 0.033991724252700806, + 0.08850973099470139, + 0.05803058296442032, + 0.2189919799566269 + ], + "credit_map": [ + 0.03082609735429287, + 0.02926388569176197, + 0.0318201519548893, + 0.027696315199136734, + 0.02927556075155735, + 0.027903787791728973, + 0.019833553582429886, + 0.02182411588728428, + 0.006919837091118097, + 0.006889555137604475, + 0.008699088357388973, + 0.011084082536399364, + 0.005639744456857443, + 0.005166142247617245, + 0.004286450333893299, + 0.005534488707780838, + 0.013353868387639523, + 0.014146831817924976, + 0.014071870595216751, + 0.0182925034314394, + 0.009826282039284706, + 0.009328710846602917, + 0.005585540551692247, + 0.006483880802989006, + 0.004622874781489372, + 0.004479055292904377, + 0.006310244556516409, + 0.00816319789737463, + 0.0045238700695335865, + 0.00277367583476007, + 0.005331332329660654, + 0.0032814119476825, + 0.011650478467345238, + 0.0070426808670163155, + 0.008271974511444569, + 0.008186583407223225, + 0.005892585963010788, + 0.00559858325868845, + 0.006418995559215546, + 0.007883695885539055, + 0.025254113599658012, + 0.010013241320848465, + 0.020832844078540802, + 0.022309117019176483, + 0.010301753878593445, + 0.013034312054514885, + 0.01944688893854618, + 0.015100075863301754, + 0.004719059448689222, + 0.003682713722810149, + 0.0044179982505738735, + 0.005133615341037512, + 0.0046127657406032085, + 0.0034416026901453733, + 0.005021583288908005, + 0.0029623855371028185, + 0.008825398981571198, + 0.007887828163802624, + 0.013783654198050499, + 0.011545766144990921, + 0.01236665714532137, + 0.010021435096859932, + 0.016020091250538826, + 0.008058897219598293, + 0.006359752733260393, + 0.003786396235227585, + 0.010758258402347565, + 0.008617088198661804, + 0.0042749252170324326, + 0.006423428654670715, + 0.013138407841324806, + 0.004672324750572443, + 0.02011922560632229, + 0.01309297140687704, + 0.03387507051229477, + 0.04132780805230141, + 0.021675484254956245, + 0.019666597247123718, + 0.04362739250063896, + 0.02560742013156414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.38712409138679504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011100324802100658 + ], + "opposite_logprob_tokens": [ + -4.5111002922058105 + ], + "logprob_reward": -0.011100324802100658, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011100324802100658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29526957869529724, + 0.11610142141580582, + 0.09590143710374832, + 0.09811688959598541, + 0.09830153733491898, + 0.08915647119283676, + 0.03527432680130005, + 0.029080023989081383, + 0.027417579665780067, + 0.11538065969944 + ], + "credit_map": [ + 0.03168970346450806, + 0.03699750825762749, + 0.04917243868112564, + 0.05846477299928665, + 0.03023185208439827, + 0.038582704961299896, + 0.020314255729317665, + 0.029816387221217155, + 0.014177381061017513, + 0.019567305222153664, + 0.017056817188858986, + 0.02096622623503208, + 0.01230695005506277, + 0.00922252144664526, + 0.010120698250830173, + 0.012683534063398838, + 0.009405468590557575, + 0.012502764351665974, + 0.016773439943790436, + 0.016213005408644676, + 0.00947028212249279, + 0.0071578798815608025, + 0.01330669317394495, + 0.01107191015034914, + 0.013708576560020447, + 0.010050139389932156, + 0.01836089789867401, + 0.02154839225113392, + 0.00911880936473608, + 0.008686259388923645, + 0.008612427860498428, + 0.008031393401324749, + 0.021836554631590843, + 0.008375028148293495, + 0.01419998612254858, + 0.012697347439825535, + 0.011577463708817959, + 0.008547956123948097, + 0.010470720939338207, + 0.010596483014523983, + 0.013108422979712486, + 0.008332140743732452, + 0.012265986762940884, + 0.014297401532530785, + 0.012537413276731968, + 0.01074447762221098, + 0.009915083646774292, + 0.007955556735396385, + 0.005361984018236399, + 0.003821436082944274, + 0.00490969093516469, + 0.00445136334747076, + 0.0050705415196716785, + 0.0036297128535807133, + 0.003656084183603525, + 0.004373516421765089, + 0.0033659327309578657, + 0.002969523426145315, + 0.005071632098406553, + 0.004391870461404324, + 0.004514231346547604, + 0.0025967040564864874, + 0.004038630053400993, + 0.0021315019112080336, + 0.002478434704244137, + 0.0024676586035639048, + 0.003978915512561798, + 0.0036265719681978226, + 0.0029193228110671043, + 0.005391859449446201, + 0.0042696925811469555, + 0.0022851263638585806, + 0.008636769838631153, + 0.005679894704371691, + 0.02010670304298401, + 0.02225535176694393, + 0.007710047531872988, + 0.012569110840559006, + 0.028662288561463356, + 0.009760508313775063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.011100324802100658 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206576153635979 + ], + "opposite_logprob_tokens": [ + -2.812065839767456 + ], + "logprob_reward": -0.06206576153635979, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206576153635979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1539888083934784, + 0.07765456289052963, + 0.15324346721172333, + 0.0592934750020504, + 0.062140777707099915, + 0.2732899487018585, + 0.019635558128356934, + 0.018334176391363144, + 0.032210495322942734, + 0.15020866692066193 + ], + "credit_map": [ + 0.01243620365858078, + 0.018049979582428932, + 0.021607985720038414, + 0.023230068385601044, + 0.022317837923765182, + 0.027515023946762085, + 0.013741901144385338, + 0.015089835040271282, + 0.009319924749433994, + 0.011884385719895363, + 0.011140309274196625, + 0.014376015402376652, + 0.007313824724406004, + 0.009251211769878864, + 0.005942715331912041, + 0.0084261829033494, + 0.01544943917542696, + 0.019675705581903458, + 0.026824409142136574, + 0.027204327285289764, + 0.01644422672688961, + 0.012245286256074905, + 0.01891126111149788, + 0.016488825902342796, + 0.011627416126430035, + 0.004782464820891619, + 0.00932741817086935, + 0.008319455198943615, + 0.006008157506585121, + 0.007146499119699001, + 0.004377729259431362, + 0.007704338524490595, + 0.013491110876202583, + 0.005173849873244762, + 0.010856609791517258, + 0.01183369755744934, + 0.004492856562137604, + 0.003906728234142065, + 0.00781018752604723, + 0.004575742408633232, + 0.046068158000707626, + 0.02187054231762886, + 0.04250922426581383, + 0.04799584671854973, + 0.01933612860739231, + 0.027516746893525124, + 0.03893064334988594, + 0.029062671586871147, + 0.001720731146633625, + 0.0017236907733604312, + 0.002412973903119564, + 0.002468626480549574, + 0.0028798612765967846, + 0.003347043413668871, + 0.0030695388559252024, + 0.0020130935590714216, + 0.002124458085745573, + 0.0014198493445292115, + 0.00343769951723516, + 0.0025206063874065876, + 0.0020144530571997166, + 0.0015817648964002728, + 0.003329607890918851, + 0.0019057398894801736, + 0.0025028763338923454, + 0.002052688505500555, + 0.006605855189263821, + 0.0053273956291377544, + 0.0025649978779256344, + 0.00381529051810503, + 0.006808963604271412, + 0.002532427664846182, + 0.010294593870639801, + 0.00837419368326664, + 0.028174329549074173, + 0.024139126762747765, + 0.010925604030489922, + 0.014968032948672771, + 0.03651101514697075, + 0.016821784898638725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.06206576153635979 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900717854499817 + ], + "opposite_logprob_tokens": [ + -2.579007148742676 + ], + "logprob_reward": -0.07900717854499817, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900717854499817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17090585827827454, + 0.1430404931306839, + 0.1482923924922943, + 0.055406443774700165, + 0.11207947134971619, + 0.25417694449424744, + 0.03441929817199707, + 0.01771220937371254, + 0.02430873177945614, + 0.03965822234749794 + ], + "credit_map": [ + 0.026109274476766586, + 0.023134548217058182, + 0.026392707601189613, + 0.02459774538874626, + 0.016426948830485344, + 0.02754918485879898, + 0.012362329289317131, + 0.014333114959299564, + 0.01731998845934868, + 0.020356839522719383, + 0.020650701597332954, + 0.024368146434426308, + 0.012441632337868214, + 0.018437078222632408, + 0.012920049950480461, + 0.01654604636132717, + 0.021679498255252838, + 0.02296408638358116, + 0.0210444126278162, + 0.023348085582256317, + 0.011288071982562542, + 0.01666027307510376, + 0.016682084649801254, + 0.014625867828726768, + 0.009755183011293411, + 0.003911513835191727, + 0.008673149161040783, + 0.008907917886972427, + 0.004956778138875961, + 0.008472496643662453, + 0.0035988243762403727, + 0.007130573038011789, + 0.022539207711815834, + 0.01516919769346714, + 0.018083924427628517, + 0.020493289455771446, + 0.010617366060614586, + 0.006546288728713989, + 0.0106087327003479, + 0.008021451532840729, + 0.04252868890762329, + 0.021468769758939743, + 0.03965633362531662, + 0.04555997997522354, + 0.0228052269667387, + 0.022482773289084435, + 0.03222523257136345, + 0.02744991146028042, + 0.005784826818853617, + 0.002505068201571703, + 0.005086258053779602, + 0.005428191740065813, + 0.0031358867418020964, + 0.002914030570536852, + 0.0042138900607824326, + 0.005351142026484013, + 0.0019921690691262484, + 0.001321924151852727, + 0.0038392087444663048, + 0.003408509772270918, + 0.0015603734645992517, + 0.0015512091340497136, + 0.002614083467051387, + 0.0014247291255742311, + 0.002495658118277788, + 0.0015555860009044409, + 0.005350649356842041, + 0.004256672691553831, + 0.0015314153861254454, + 0.0035963496193289757, + 0.003385731251910329, + 0.0021366674918681383, + 0.0038696881383657455, + 0.002843397669494152, + 0.007956419140100479, + 0.005973822437226772, + 0.003342657582834363, + 0.006523219868540764, + 0.0047569177113473415, + 0.004392096772789955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.07900717854499817 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, extreme wide shot slowly advances, establishing a sprawling estate at dusk. A grand, imposing Gothic Revival mansion, constructed of dark, weathered stone and crowned with a soaring spire, rises atop a gentle hill, silhouetted against a deep indigo and violet-purple sky, the last fiery orange glow of sunset streaming along the horizon. Far below, a lone vintage luxury car, its sleek black body reflecting the fading twilight, steadily and deliberately moves down a long, winding gravel driveway. Its powerful twin headlights cut through the low-lying mist clinging to the manicured grounds, streaking sharp white beams across the gnarled, ancient trees lining the road. As the vehicle navigates the winding curves, its steady progress is the only movement in the silent landscape, its destination the distant, warmly lit windows of the mansion its only destination, creating an air of elegant isolation and a powerful atmosphere of quiet mystery.", + "seed": 31398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_3/index_182_seed_31398_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13709588069468737, + "vlm_logprob": -0.13709588069468737, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.13709588069468737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06880434602499008 + ], + "opposite_logprob_tokens": [ + -4.3188042640686035 + ], + "logprob_reward": -0.06880434602499008, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.06880434602499008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08385911583900452, + 0.09666291624307632, + 0.06831259280443192, + 0.04323079064488411, + 0.04059068486094475, + 0.15038716793060303, + 0.21537844836711884, + 0.23616375029087067, + 0.03978870436549187, + 0.025625837966799736 + ], + "credit_map": [ + 0.008169254288077354, + 0.012389623560011387, + 0.012088696472346783, + 0.0181624423712492, + 0.005620912648737431, + 0.004476303234696388, + 0.011319568380713463, + 0.011632310226559639, + 0.004919259808957577, + 0.007608108688145876, + 0.015601596795022488, + 0.019259203225374222, + 0.01216820441186428, + 0.005772475618869066, + 0.013072590343654156, + 0.018261481076478958, + 0.004949535708874464, + 0.006441229488700628, + 0.008555559441447258, + 0.012999992817640305, + 0.009456126019358635, + 0.003494859905913472, + 0.010727869346737862, + 0.011687422171235085, + 0.002296402584761381, + 0.0038040366489440203, + 0.006411262787878513, + 0.007268537301570177, + 0.00500348350033164, + 0.0031027316581457853, + 0.006326170172542334, + 0.009018165990710258, + 0.002605935325846076, + 0.004898828454315662, + 0.00632226699963212, + 0.008201023563742638, + 0.003168355207890272, + 0.0029249980580061674, + 0.005900531075894833, + 0.0065687475726008415, + 0.008591845631599426, + 0.012669206596910954, + 0.02462230622768402, + 0.02274218201637268, + 0.014323526993393898, + 0.008570374920964241, + 0.030150843784213066, + 0.02871686965227127, + 0.010407809168100357, + 0.011225622147321701, + 0.029897011816501617, + 0.0341908223927021, + 0.015846731141209602, + 0.01065870001912117, + 0.06868371367454529, + 0.03446805104613304, + 0.013181121088564396, + 0.01649780571460724, + 0.035049136728048325, + 0.039913371205329895, + 0.031919144093990326, + 0.018634231761097908, + 0.04006655886769295, + 0.04090239480137825, + 0.0025404240004718304, + 0.003942490555346012, + 0.0059310803189873695, + 0.005579962860792875, + 0.004692353308200836, + 0.0023539375979453325, + 0.00825339462608099, + 0.006495061330497265, + 0.0016737752594053745, + 0.0026134494692087173, + 0.0028217281214892864, + 0.003568410873413086, + 0.0026472732424736023, + 0.0030234039295464754, + 0.004755896050482988, + 0.004521900322288275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.06880434602499008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28136202692985535 + ], + "opposite_logprob_tokens": [ + -1.4063620567321777 + ], + "logprob_reward": -0.28136202692985535, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28136202692985535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12333337217569351, + 0.12226037681102753, + 0.038509733974933624, + 0.10654708743095398, + 0.06800927966833115, + 0.07504630833864212, + 0.15694338083267212, + 0.0722358226776123, + 0.1051253229379654, + 0.13198933005332947 + ], + "credit_map": [ + 0.010768572799861431, + 0.013416651636362076, + 0.0165448896586895, + 0.02414007857441902, + 0.017346780747175217, + 0.008441456593573093, + 0.01439263578504324, + 0.018282311037182808, + 0.008488510735332966, + 0.01023868191987276, + 0.01413190457969904, + 0.021011628210544586, + 0.01742088980972767, + 0.006910869851708412, + 0.022772423923015594, + 0.021285466849803925, + 0.0037212171591818333, + 0.005424363072961569, + 0.006497015245258808, + 0.005826985463500023, + 0.003844457445666194, + 0.0024483902379870415, + 0.00480232248082757, + 0.005944982171058655, + 0.00505300285294652, + 0.0075125726871192455, + 0.011940796859562397, + 0.020923787727952003, + 0.0165457371622324, + 0.006148029118776321, + 0.018157999962568283, + 0.020265160128474236, + 0.004962748847901821, + 0.003785715438425541, + 0.010576090775430202, + 0.015487090684473515, + 0.005579876247793436, + 0.004923438187688589, + 0.010684219188988209, + 0.01201009564101696, + 0.004852960351854563, + 0.005066247191280127, + 0.011695211753249168, + 0.018252400681376457, + 0.00683404179289937, + 0.00497607234865427, + 0.011530970223248005, + 0.011838400736451149, + 0.00785756390541792, + 0.008639371022582054, + 0.021964779123663902, + 0.03226755931973457, + 0.016975896432995796, + 0.0102395536378026, + 0.03463708609342575, + 0.024361560121178627, + 0.0032514501363039017, + 0.004374097567051649, + 0.01010441966354847, + 0.008510718122124672, + 0.005511126480996609, + 0.004728843457996845, + 0.021746287122368813, + 0.014008883386850357, + 0.006243028212338686, + 0.008546489290893078, + 0.010312489233911037, + 0.016559157520532608, + 0.013580767437815666, + 0.0070228311233222485, + 0.024684416130185127, + 0.01817614585161209, + 0.007897386327385902, + 0.016916494816541672, + 0.01208502147346735, + 0.019983036443591118, + 0.01637500524520874, + 0.010473991744220257, + 0.02969168871641159, + 0.018566707149147987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.28136202692985535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556307949125767 + ], + "opposite_logprob_tokens": [ + -4.387556076049805 + ], + "logprob_reward": -0.012556307949125767, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556307949125767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16132105886936188, + 0.1647481471300125, + 0.028970761224627495, + 0.04655223712325096, + 0.06719791144132614, + 0.1130739375948906, + 0.15754248201847076, + 0.09614235907793045, + 0.04103734716773033, + 0.12341377139091492 + ], + "credit_map": [ + 0.018821999430656433, + 0.02107951045036316, + 0.02601475454866886, + 0.02752836048603058, + 0.011535809375345707, + 0.008334311656653881, + 0.02077912539243698, + 0.027227196842432022, + 0.008778928779065609, + 0.017070328816771507, + 0.02374211698770523, + 0.0301558468490839, + 0.019399354234337807, + 0.01007983461022377, + 0.026740338653326035, + 0.02878139726817608, + 0.0024973072577267885, + 0.003974242601543665, + 0.005047440528869629, + 0.004019807558506727, + 0.0024109315127134323, + 0.002480422845110297, + 0.004563915077596903, + 0.0039766947738826275, + 0.003183895256370306, + 0.004829919897019863, + 0.005786890164017677, + 0.0071331169456243515, + 0.005578002892434597, + 0.0034073563292622566, + 0.008183394558727741, + 0.00844966247677803, + 0.005169225856661797, + 0.007774019613862038, + 0.010999429039657116, + 0.010575502179563046, + 0.005268882494419813, + 0.004884046036750078, + 0.009949883446097374, + 0.012576920911669731, + 0.006086082197725773, + 0.007997977547347546, + 0.017245391383767128, + 0.01808297634124756, + 0.009320730343461037, + 0.0072388192638754845, + 0.024908168241381645, + 0.02219378389418125, + 0.006859095301479101, + 0.006865871138870716, + 0.02195786125957966, + 0.036498408764600754, + 0.011015611700713634, + 0.00647869985550642, + 0.04166329652070999, + 0.026203639805316925, + 0.004689847119152546, + 0.005809240508824587, + 0.015080137178301811, + 0.015214996412396431, + 0.01142712403088808, + 0.006219280418008566, + 0.018919331952929497, + 0.018782395869493484, + 0.0026822322979569435, + 0.0033947841729968786, + 0.004230675287544727, + 0.006361466832458973, + 0.003873943816870451, + 0.0035828365944325924, + 0.010317154228687286, + 0.006594253238290548, + 0.006651429925113916, + 0.012868234887719154, + 0.01496090367436409, + 0.017623402178287506, + 0.012787245213985443, + 0.00782325305044651, + 0.02736462838947773, + 0.023334674537181854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.012556307949125767 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428273320198059 + ], + "opposite_logprob_tokens": [ + -2.017827272415161 + ], + "logprob_reward": -0.1428273320198059, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428273320198059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11821085214614868, + 0.20021489262580872, + 0.10368283838033676, + 0.028846412897109985, + 0.05612979084253311, + 0.1286156177520752, + 0.1524992734193802, + 0.13940434157848358, + 0.04064313322305679, + 0.0317528136074543 + ], + "credit_map": [ + 0.010702565312385559, + 0.017973238602280617, + 0.01566099189221859, + 0.021207749843597412, + 0.008777331560850143, + 0.005891634151339531, + 0.013563623651862144, + 0.024433735758066177, + 0.01188425999134779, + 0.023229673504829407, + 0.0269149336963892, + 0.036675821989774704, + 0.024846216663718224, + 0.012301194481551647, + 0.029258595779538155, + 0.03510420769453049, + 0.008619341067969799, + 0.007067727390676737, + 0.012167085893452168, + 0.017239388078451157, + 0.017936306074261665, + 0.007190713658928871, + 0.018753405660390854, + 0.014708886854350567, + 0.002956181764602661, + 0.0022685539443045855, + 0.003917648456990719, + 0.005008051171898842, + 0.0020414183381944895, + 0.002520639216527343, + 0.005449103657156229, + 0.004684820305556059, + 0.003376150503754616, + 0.004187172744423151, + 0.010444916784763336, + 0.009249485097825527, + 0.003003884805366397, + 0.0037972829304635525, + 0.00997032318264246, + 0.0121005829423666, + 0.004984302446246147, + 0.007936554029583931, + 0.019570346921682358, + 0.02357340045273304, + 0.009904284030199051, + 0.0080387182533741, + 0.02991543710231781, + 0.02469259686768055, + 0.008401093073189259, + 0.006958883721381426, + 0.021580155938863754, + 0.0291131678968668, + 0.013748177327215672, + 0.007071240339428186, + 0.04253631830215454, + 0.02309025079011917, + 0.0068399799056351185, + 0.008677289821207523, + 0.021839085966348648, + 0.020238535478711128, + 0.016849616542458534, + 0.009038983844220638, + 0.028487732633948326, + 0.027433117851614952, + 0.002688134554773569, + 0.0031895674765110016, + 0.005771937780082226, + 0.006747828796505928, + 0.002944645006209612, + 0.0036237009335309267, + 0.009121724404394627, + 0.006555598229169846, + 0.0017252092948183417, + 0.002529992489144206, + 0.004011055920273066, + 0.004995471797883511, + 0.003526893211528659, + 0.0027231608983129263, + 0.00637661712244153, + 0.0058644176460802555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.1428273320198059 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle, extreme wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly advancing towards the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grand Gothic Revival mansion made of dark, weathered stone with a soaring spire visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone vintage luxury car with a sleek black body moving down a long, winding gravel driveway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the car's twin headlights emitting sharp white beams cutting through low-lying mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992939054965973 + ], + "opposite_logprob_tokens": [ + -1.8049293756484985 + ], + "logprob_reward": -0.17992939054965973, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992939054965973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24044418334960938, + 0.14812035858631134, + 0.0789530947804451, + 0.09028176963329315, + 0.06188496574759483, + 0.10015162825584412, + 0.13219788670539856, + 0.06375792622566223, + 0.043908022344112396, + 0.040300145745277405 + ], + "credit_map": [ + 0.022042790427803993, + 0.028565894812345505, + 0.0394066721200943, + 0.060371994972229004, + 0.015657445415854454, + 0.010082077234983444, + 0.02140101231634617, + 0.04291630536317825, + 0.008069757372140884, + 0.014856308698654175, + 0.021692734211683273, + 0.029642360284924507, + 0.0185337346047163, + 0.008917417377233505, + 0.01855621114373207, + 0.02785182185471058, + 0.005230272654443979, + 0.010389791801571846, + 0.009973818436264992, + 0.01566026546061039, + 0.007406267337501049, + 0.005915699992328882, + 0.012690999545156956, + 0.011685987003147602, + 0.007230144925415516, + 0.007431489881128073, + 0.012489404529333115, + 0.011966445483267307, + 0.011295546777546406, + 0.007114201784133911, + 0.01730993390083313, + 0.015444609336555004, + 0.0037927357479929924, + 0.004797022324055433, + 0.010947153903543949, + 0.010675434954464436, + 0.0035506233107298613, + 0.005217327736318111, + 0.011560534127056599, + 0.011344133876264095, + 0.00400995509698987, + 0.006024971138685942, + 0.015428462065756321, + 0.018076324835419655, + 0.00833973754197359, + 0.006235279142856598, + 0.0225223358720541, + 0.01951456256210804, + 0.00741969607770443, + 0.00690336711704731, + 0.020726051181554794, + 0.020793618634343147, + 0.010597619228065014, + 0.005205879453569651, + 0.040117621421813965, + 0.020434018224477768, + 0.0029573494102805853, + 0.003363218856975436, + 0.010290230624377728, + 0.011092329397797585, + 0.006393528543412685, + 0.004279808606952429, + 0.01242731511592865, + 0.012954149395227432, + 0.003006467828527093, + 0.0030301790684461594, + 0.0064145284704864025, + 0.006531126331537962, + 0.0036481234710663557, + 0.0048892563208937645, + 0.008915003389120102, + 0.007473335135728121, + 0.002269433345645666, + 0.00269122002646327, + 0.0062958779744803905, + 0.0064185429364442825, + 0.004011237528175116, + 0.0042754351161420345, + 0.0074889990501105785, + 0.0068493993021547794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.17992939054965973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_4/index_tensor([2173, 2173])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_4/index_tensor([2173, 2173])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bbdae0802c4cc61dbe00bb19751bbae44af8587f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_4/index_tensor([2173, 2173])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly upward perspective, initially static, focuses on a crow perched on a bare, cracked, gray branch. Its inky black feathers gleam like a metallic sheen in the soft, diffuse light of an overcast sky. Its sharp, alert black eyes survey its surroundings, its body tense, as if poised for a sudden explosion. The background is a pale, gloomy, soft-focus sky, a blurry, milky white canvas. Suddenly, it pushes away from the branch, decisively and swiftly spreading its massive black wings. With a powerful downward flap, its entire body instantly soars into the air. The camera then rapidly tilts upward, capturing its outstretched wings as they sweep across the sky in a soft-focus frame, transforming into an elegant yet swift black arc before disappearing from view. The entire process, captured in minimalist black and gray tones, creates a dramatic tension of loneliness, awe, and primal vitality.", + "seed": 27637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_4/index_2173_seed_27637_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009268021886236965, + "vlm_logprob": -0.009268021886236965, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.009268021886236965, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008889740332961082 + ], + "opposite_logprob_tokens": [ + -5.633889675140381 + ], + "logprob_reward": -0.008889740332961082, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.008889740332961082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08150368928909302, + 0.07132904976606369, + 0.15938016772270203, + 0.041798580437898636, + 0.07264900952577591, + 0.044366415590047836, + 0.09022966772317886, + 0.12265465408563614, + 0.06141160801053047, + 0.2546771466732025 + ], + "credit_map": [ + 0.010717370547354221, + 0.007300720550119877, + 0.0129298847168684, + 0.00928157102316618, + 0.005040144547820091, + 0.013077105395495892, + 0.008562857285141945, + 0.014594040811061859, + 0.008286484517157078, + 0.006446328945457935, + 0.01885977014899254, + 0.007912000641226768, + 0.004320546053349972, + 0.008953348733484745, + 0.005132029764354229, + 0.011418534442782402, + 0.02113999053835869, + 0.012035232037305832, + 0.0220049861818552, + 0.014176758006215096, + 0.009952607564628124, + 0.02480938285589218, + 0.022738637402653694, + 0.032522574067115784, + 0.005889833439141512, + 0.003551766276359558, + 0.007689826190471649, + 0.003043057397007942, + 0.004298827610909939, + 0.007933417335152626, + 0.0031069908291101456, + 0.006284860894083977, + 0.007840828038752079, + 0.008867857046425343, + 0.007108480203896761, + 0.008885555900633335, + 0.008357376791536808, + 0.014106052927672863, + 0.008145892061293125, + 0.00933696050196886, + 0.00478829862549901, + 0.0043772244825959206, + 0.007071624509990215, + 0.003588557941839099, + 0.006110109388828278, + 0.009265906177461147, + 0.004600712563842535, + 0.0045639825984835625, + 0.008494239300489426, + 0.011715412139892578, + 0.011021995916962624, + 0.008391673676669598, + 0.013202046044170856, + 0.014728178270161152, + 0.010041183792054653, + 0.012634940445423126, + 0.009553429670631886, + 0.02367585524916649, + 0.020577384158968925, + 0.014170719310641289, + 0.019815126433968544, + 0.01553887315094471, + 0.009958048351109028, + 0.009365227073431015, + 0.0052550784312188625, + 0.00855787843465805, + 0.008764521218836308, + 0.01184788253158331, + 0.006332606077194214, + 0.005199831444770098, + 0.00960477814078331, + 0.005849027074873447, + 0.015753252431750298, + 0.03419976308941841, + 0.024090351536870003, + 0.061787210404872894, + 0.02850615419447422, + 0.021802110597491264, + 0.0326586551964283, + 0.03587964549660683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.008889740332961082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014241958037018776 + ], + "opposite_logprob_tokens": [ + -4.264242172241211 + ], + "logprob_reward": -0.014241958037018776, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014241958037018776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22299540042877197, + 0.06956196576356888, + 0.057365063577890396, + 0.1251637041568756, + 0.05403422191739082, + 0.10087643563747406, + 0.07767533510923386, + 0.12532564997673035, + 0.06071556732058525, + 0.1062866598367691 + ], + "credit_map": [ + 0.02970796450972557, + 0.014415883459150791, + 0.033173464238643646, + 0.01910518668591976, + 0.012829488143324852, + 0.04332365095615387, + 0.021842092275619507, + 0.048597682267427444, + 0.009273514151573181, + 0.00602963799610734, + 0.01556301862001419, + 0.006676585879176855, + 0.003469242947176099, + 0.009355605579912663, + 0.007208768278360367, + 0.011985598132014275, + 0.008014854975044727, + 0.004213525913655758, + 0.008366326801478863, + 0.0037718869280070066, + 0.003830290399491787, + 0.009996882639825344, + 0.006054130848497152, + 0.013117166236042976, + 0.014949393458664417, + 0.011494636535644531, + 0.024693438783288002, + 0.011320036835968494, + 0.014014671556651592, + 0.020174510776996613, + 0.008841356262564659, + 0.01967565529048443, + 0.003821159712970257, + 0.006072119809687138, + 0.006875704508274794, + 0.007884870283305645, + 0.007592296693474054, + 0.006792624481022358, + 0.007919461466372013, + 0.007075982168316841, + 0.008205541409552097, + 0.011548604816198349, + 0.011109652929008007, + 0.012571873143315315, + 0.012707257643342018, + 0.023098450154066086, + 0.010020459070801735, + 0.011614597402513027, + 0.008221165277063847, + 0.011144622229039669, + 0.007852760143578053, + 0.004576005507260561, + 0.01086209062486887, + 0.012494826689362526, + 0.004844054579734802, + 0.017679810523986816, + 0.010788759216666222, + 0.025965549051761627, + 0.022751955315470695, + 0.013466463424265385, + 0.01773090846836567, + 0.013886029832065105, + 0.010305446572601795, + 0.010430538095533848, + 0.005982179660350084, + 0.00669461116194725, + 0.007749846205115318, + 0.010760662145912647, + 0.0102932034060359, + 0.00530757661908865, + 0.008295558393001556, + 0.005631932523101568, + 0.006377711426466703, + 0.01855110563337803, + 0.011704081669449806, + 0.021965298801660538, + 0.011571984738111496, + 0.007956829853355885, + 0.013821779750287533, + 0.014337868429720402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.014241958037018776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011870964663103223 + ], + "opposite_logprob_tokens": [ + -6.751187324523926 + ], + "logprob_reward": -0.0011870964663103223, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011870964663103223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05988670885562897, + 0.06491135060787201, + 0.06251654028892517, + 0.029303845018148422, + 0.05307593196630478, + 0.2369692027568817, + 0.04807554930448532, + 0.14770253002643585, + 0.04661359265446663, + 0.25094473361968994 + ], + "credit_map": [ + 0.007100602611899376, + 0.00653193099424243, + 0.013598974794149399, + 0.008362042717635632, + 0.0029060360975563526, + 0.008275077678263187, + 0.005094079300761223, + 0.008017963729798794, + 0.00837925635278225, + 0.0051833391189575195, + 0.01039168145507574, + 0.007421415764838457, + 0.0037920186296105385, + 0.008934401907026768, + 0.009262478910386562, + 0.011546756140887737, + 0.008643238805234432, + 0.004776760004460812, + 0.008182700723409653, + 0.004703998100012541, + 0.003781720297411084, + 0.010952603071928024, + 0.007061302661895752, + 0.014414215460419655, + 0.004850477911531925, + 0.0024068232160061598, + 0.004748494364321232, + 0.0017210922669619322, + 0.002460562391206622, + 0.0056253341026604176, + 0.001622420153580606, + 0.005868639796972275, + 0.0065218498930335045, + 0.005258859600871801, + 0.007423587143421173, + 0.005673879291862249, + 0.005723859183490276, + 0.008335993625223637, + 0.006780665367841721, + 0.0073572383262217045, + 0.020799972116947174, + 0.030874663963913918, + 0.030963316559791565, + 0.019185801967978477, + 0.023200858384370804, + 0.06634802371263504, + 0.018987474963068962, + 0.026609094813466072, + 0.005503032356500626, + 0.006470841821283102, + 0.005248609930276871, + 0.003721275832504034, + 0.007712498772889376, + 0.0078069730661809444, + 0.006295677740126848, + 0.005316637456417084, + 0.012339525856077671, + 0.033321529626846313, + 0.025453269481658936, + 0.018077302724123, + 0.018084384500980377, + 0.01439546886831522, + 0.0129327941685915, + 0.01309825200587511, + 0.004084256011992693, + 0.005859754979610443, + 0.006196481175720692, + 0.010079491883516312, + 0.004829317796975374, + 0.00478884344920516, + 0.0062786913476884365, + 0.004496757406741381, + 0.01530435774475336, + 0.05206172168254852, + 0.025529926642775536, + 0.04844999313354492, + 0.0291665717959404, + 0.019578201696276665, + 0.029157958924770355, + 0.03169599920511246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.0011870964663103223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069892793893814 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.016069892793893814, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069892793893814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09484802186489105, + 0.03043697215616703, + 0.012944518588483334, + 0.03284946084022522, + 0.028499828651547432, + 0.1749831587076187, + 0.0584443174302578, + 0.3549301028251648, + 0.04553432762622833, + 0.16652929782867432 + ], + "credit_map": [ + 0.012948687188327312, + 0.007976608350872993, + 0.013124432414770126, + 0.01020197756588459, + 0.005162368528544903, + 0.01388676930218935, + 0.012419466860592365, + 0.019127700477838516, + 0.003915946464985609, + 0.002371425973251462, + 0.0034886389039456844, + 0.0036037268582731485, + 0.0018710995791479945, + 0.005703880451619625, + 0.004477109760046005, + 0.005005143117159605, + 0.001726012909784913, + 0.0013071232242509723, + 0.0016995060723274946, + 0.0009995995787903666, + 0.0010125271510332823, + 0.0026766015216708183, + 0.0011241650208830833, + 0.0023989833425730467, + 0.004655219614505768, + 0.0034594470635056496, + 0.004152318928390741, + 0.001899140072055161, + 0.0032520147506147623, + 0.006449112202972174, + 0.003612296422943473, + 0.005369910504668951, + 0.0026652796659618616, + 0.0032314613927155733, + 0.003522769082337618, + 0.0027554109692573547, + 0.004950998816639185, + 0.004077252931892872, + 0.004366403911262751, + 0.002930249320343137, + 0.012043128721415997, + 0.01836625300347805, + 0.019720928743481636, + 0.02059406042098999, + 0.022649236023426056, + 0.04440966993570328, + 0.019082047045230865, + 0.018117839470505714, + 0.00505648460239172, + 0.008111882954835892, + 0.008197196759283543, + 0.0055811116471886635, + 0.007983430288732052, + 0.007584922481328249, + 0.004683398641645908, + 0.011245884001255035, + 0.03553410619497299, + 0.08507734537124634, + 0.06265546381473541, + 0.03941981866955757, + 0.04812069237232208, + 0.032877109944820404, + 0.023109406232833862, + 0.028136156499385834, + 0.0037493088748306036, + 0.005897595547139645, + 0.006092709954828024, + 0.008498565293848515, + 0.007151789031922817, + 0.00394477928057313, + 0.005784470122307539, + 0.004415109753608704, + 0.011392669752240181, + 0.028218740597367287, + 0.02366223931312561, + 0.03939235582947731, + 0.017045052722096443, + 0.014299053698778152, + 0.01678100973367691, + 0.01573816128075123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.016069892793893814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059514218010008335 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.0059514218010008335, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059514218010008335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20833365619182587, + 0.03292117267847061, + 0.02795466035604477, + 0.02713296189904213, + 0.03408297896385193, + 0.06911972165107727, + 0.10528041422367096, + 0.2623324990272522, + 0.12060905992984772, + 0.11223286390304565 + ], + "credit_map": [ + 0.026906397193670273, + 0.016204599291086197, + 0.023411650210618973, + 0.021385252475738525, + 0.011850996874272823, + 0.04202582687139511, + 0.022982053458690643, + 0.04356689006090164, + 0.004421876277774572, + 0.002145881997421384, + 0.005552525166422129, + 0.0018994410056620836, + 0.002068121684715152, + 0.0072031342424452305, + 0.0022074957378208637, + 0.007422696799039841, + 0.003776495112106204, + 0.002505925018340349, + 0.005130047909915447, + 0.0024551579263061285, + 0.0015795856015756726, + 0.00479131517931819, + 0.0034756886307150126, + 0.004240445792675018, + 0.004247656092047691, + 0.0028739082626998425, + 0.005897912196815014, + 0.0021689177956432104, + 0.0025829540099948645, + 0.003603913588449359, + 0.002243263181298971, + 0.003514437237754464, + 0.004324044566601515, + 0.003664022544398904, + 0.005392435006797314, + 0.0040838224813342094, + 0.003542480291798711, + 0.004956196062266827, + 0.0035130942706018686, + 0.004606882110238075, + 0.006120066624134779, + 0.0057265921495854855, + 0.0094932084903121, + 0.007307550869882107, + 0.011436163447797298, + 0.01607181504368782, + 0.00634424714371562, + 0.006620078347623348, + 0.012621767818927765, + 0.014068017713725567, + 0.011866318061947823, + 0.006850196048617363, + 0.018643494695425034, + 0.015827149152755737, + 0.009713280014693737, + 0.01569019816815853, + 0.023589838296175003, + 0.053021643310785294, + 0.05099630728363991, + 0.026389816775918007, + 0.03433436155319214, + 0.03255494311451912, + 0.02175627462565899, + 0.019689321517944336, + 0.010405216366052628, + 0.013745700940489769, + 0.016906462609767914, + 0.02875315584242344, + 0.01404908299446106, + 0.00903310813009739, + 0.016382811591029167, + 0.011333527974784374, + 0.010560081340372562, + 0.013607512228190899, + 0.024595333263278008, + 0.017029501497745514, + 0.01220667827874422, + 0.008169393055140972, + 0.01354847103357315, + 0.012515888549387455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.0059514218010008335 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward perspective, initially static, focuses on a crow perched on a bare, cracked, gray branch. Its inky black feathers gleam like a metallic sheen in the soft, diffuse light of an overcast sky. Its sharp, alert black eyes survey its surroundings, its body tense, as if poised for a sudden explosion. The background is a pale, gloomy, soft-focus sky, a blurry, milky white canvas. Suddenly, it pushes away from the branch, decisively and swiftly spreading its massive black wings. With a powerful downward flap, its entire body instantly soars into the air. The camera then rapidly tilts upward, capturing its outstretched wings as they sweep across the sky in a soft-focus frame, transforming into an elegant yet swift black arc before disappearing from view. The entire process, captured in minimalist black and gray tones, creates a dramatic tension of loneliness, awe, and primal vitality.", + "seed": 27638, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_4/index_2173_seed_27638_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.031227430142462254, + "vlm_logprob": -0.031227430142462254, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.031227430142462254, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011222454719245434 + ], + "opposite_logprob_tokens": [ + -5.2612223625183105 + ], + "logprob_reward": -0.011222454719245434, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.011222454719245434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3809199333190918, + 0.02588428370654583, + 0.12554778158664703, + 0.02272641845047474, + 0.06052297726273537, + 0.05626891925930977, + 0.12728627026081085, + 0.11146175861358643, + 0.03261983394622803, + 0.05676178261637688 + ], + "credit_map": [ + 0.06830917298793793, + 0.028467226773500443, + 0.04628613963723183, + 0.01776539720594883, + 0.02854706533253193, + 0.05514902621507645, + 0.022218605503439903, + 0.11417730897665024, + 0.003337682457640767, + 0.003045144723728299, + 0.003845510073006153, + 0.0034449915401637554, + 0.002441778779029846, + 0.003359077963978052, + 0.002330201445147395, + 0.004079898353666067, + 0.023433132097125053, + 0.009148801676928997, + 0.010355525650084019, + 0.01084631122648716, + 0.008210712112486362, + 0.018482819199562073, + 0.009381631389260292, + 0.03568883612751961, + 0.003967452794313431, + 0.0022324954625219107, + 0.0018699831562116742, + 0.002450100611895323, + 0.0024505278561264277, + 0.003223168896511197, + 0.001749102957546711, + 0.004783586598932743, + 0.007114522624760866, + 0.0081642372533679, + 0.008098366670310497, + 0.0046571907587349415, + 0.007896495051681995, + 0.008187612518668175, + 0.006544180680066347, + 0.009860377758741379, + 0.009958285838365555, + 0.004075365606695414, + 0.006347594782710075, + 0.005141349975019693, + 0.003618017304688692, + 0.006822418887168169, + 0.007829706184566021, + 0.012476178817451, + 0.01743125542998314, + 0.015111591666936874, + 0.016065329313278198, + 0.020469902083277702, + 0.010207108221948147, + 0.014968073926866055, + 0.015144522301852703, + 0.017888491973280907, + 0.012836014851927757, + 0.015680262818932533, + 0.012389807030558586, + 0.015347670763731003, + 0.01267580222338438, + 0.014769316650927067, + 0.011401611380279064, + 0.016361277550458908, + 0.0036584713961929083, + 0.003771995659917593, + 0.004413078539073467, + 0.005462825298309326, + 0.004785821307450533, + 0.0038024773821234703, + 0.003947740886360407, + 0.0027774248737841845, + 0.007258787285536528, + 0.01319399569183588, + 0.0070032039657235146, + 0.007662962190806866, + 0.0072554172948002815, + 0.0038291008677333593, + 0.004731683526188135, + 0.005826635751873255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.011222454719245434 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009810565039515495 + ], + "opposite_logprob_tokens": [ + -4.634810447692871 + ], + "logprob_reward": -0.009810565039515495, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009810565039515495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2055748552083969, + 0.18921025097370148, + 0.048103056848049164, + 0.033171605318784714, + 0.1617625206708908, + 0.05321824550628662, + 0.10434888303279877, + 0.06582342088222504, + 0.030277561396360397, + 0.1085095927119255 + ], + "credit_map": [ + 0.028809653595089912, + 0.024173956364393234, + 0.03979166969656944, + 0.017325062304735184, + 0.016511661931872368, + 0.02507132478058338, + 0.015207829885184765, + 0.0386837013065815, + 0.026222195476293564, + 0.015799803659319878, + 0.029557205736637115, + 0.0204368457198143, + 0.014251668937504292, + 0.02445444092154503, + 0.011290164664387703, + 0.04719793424010277, + 0.011123073287308216, + 0.004882697947323322, + 0.004163755569607019, + 0.0036553218960762024, + 0.004128033760935068, + 0.007390395272523165, + 0.0030724855605512857, + 0.009687294252216816, + 0.004950511269271374, + 0.0026978079695254564, + 0.004324148874729872, + 0.00398100633174181, + 0.002565230941399932, + 0.004164827521890402, + 0.0035731669049710035, + 0.006914906669408083, + 0.019023984670639038, + 0.017274877056479454, + 0.018231842666864395, + 0.009974858723580837, + 0.028549300506711006, + 0.02314523048698902, + 0.010142359882593155, + 0.03542006388306618, + 0.007397803012281656, + 0.005107061006128788, + 0.009587901644408703, + 0.005480612628161907, + 0.004021350294351578, + 0.005229777656495571, + 0.008222542703151703, + 0.0081712007522583, + 0.018084844574332237, + 0.012114417739212513, + 0.01250044722110033, + 0.013595392927527428, + 0.009046811610460281, + 0.012577748857438564, + 0.010736804455518723, + 0.015692410990595818, + 0.008159042336046696, + 0.009247541427612305, + 0.008631289936602116, + 0.010194128379225731, + 0.0069302781485021114, + 0.008206184953451157, + 0.0063136569224298, + 0.008141294121742249, + 0.002989127766340971, + 0.0038482178933918476, + 0.004575778730213642, + 0.004776631947606802, + 0.003426400711759925, + 0.002453728811815381, + 0.004792649298906326, + 0.0034150267019867897, + 0.011487692594528198, + 0.021021079272031784, + 0.015978461131453514, + 0.017906585708260536, + 0.014829764142632484, + 0.008109918795526028, + 0.008754136972129345, + 0.010421950370073318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009810565039515495 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013397298753261566 + ], + "opposite_logprob_tokens": [ + -6.626339912414551 + ], + "logprob_reward": -0.0013397298753261566, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013397298753261566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2631891071796417, + 0.0535939484834671, + 0.09615124762058258, + 0.031226027756929398, + 0.12126781046390533, + 0.05419335514307022, + 0.11618280410766602, + 0.12167435884475708, + 0.04408663883805275, + 0.0984347015619278 + ], + "credit_map": [ + 0.03677191585302353, + 0.019384007900953293, + 0.040131185203790665, + 0.015627918764948845, + 0.022313950583338737, + 0.03278335928916931, + 0.019399244338274002, + 0.07677754014730453, + 0.00807989202439785, + 0.00476073008030653, + 0.00710277259349823, + 0.0036438829265534878, + 0.0033270339481532574, + 0.006971352733671665, + 0.0042563858442008495, + 0.01545189693570137, + 0.01827247440814972, + 0.006574898958206177, + 0.00778562156483531, + 0.006547823082655668, + 0.006855457555502653, + 0.014263958670198917, + 0.005794962402433157, + 0.030056055635213852, + 0.005550852511078119, + 0.0031228449661284685, + 0.004291831981390715, + 0.003021832322701812, + 0.0023027476854622364, + 0.004533774219453335, + 0.002140162279829383, + 0.006261979695409536, + 0.013119434006512165, + 0.013316105119884014, + 0.014610533602535725, + 0.009116352535784245, + 0.012969027273356915, + 0.0184901375323534, + 0.009640616364777088, + 0.030005602166056633, + 0.0088353818282485, + 0.004724899772554636, + 0.007309205364435911, + 0.006851376965641975, + 0.004290407057851553, + 0.005677109584212303, + 0.008425810374319553, + 0.008079164661467075, + 0.02076854184269905, + 0.014357514679431915, + 0.014470728114247322, + 0.01862238347530365, + 0.008754179812967777, + 0.011182757094502449, + 0.01269186194986105, + 0.015334837138652802, + 0.014617472887039185, + 0.018895769491791725, + 0.012897352688014507, + 0.014827619306743145, + 0.014362191781401634, + 0.015586975030601025, + 0.012555049732327461, + 0.017931926995515823, + 0.0063927615992724895, + 0.008026419207453728, + 0.0059770396910607815, + 0.006776667665690184, + 0.005633694585412741, + 0.0032669701613485813, + 0.004483573604375124, + 0.0035295095294713974, + 0.014099765568971634, + 0.019720518961548805, + 0.012734450399875641, + 0.014583556912839413, + 0.017357096076011658, + 0.005948964972048998, + 0.00843552965670824, + 0.005554819945245981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0013397298753261566 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270282119512558 + ], + "opposite_logprob_tokens": [ + -2.127028226852417 + ], + "logprob_reward": -0.1270282119512558, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270282119512558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11343690752983093, + 0.08659633249044418, + 0.21056105196475983, + 0.03785421699285507, + 0.08908434957265854, + 0.07127325981855392, + 0.1608457863330841, + 0.053768664598464966, + 0.05056910961866379, + 0.12601038813591003 + ], + "credit_map": [ + 0.01435464434325695, + 0.010873698629438877, + 0.017963148653507233, + 0.012530209496617317, + 0.009647944942116737, + 0.012031125836074352, + 0.008972838521003723, + 0.02706330269575119, + 0.015178130939602852, + 0.008571139536798, + 0.009306665509939194, + 0.006418115925043821, + 0.007721779402345419, + 0.013425412587821484, + 0.005475619342178106, + 0.02049946039915085, + 0.041377175599336624, + 0.01628083735704422, + 0.01625494286417961, + 0.017924673855304718, + 0.013088392093777657, + 0.032440923154354095, + 0.01276890467852354, + 0.06042521446943283, + 0.0048447889275848866, + 0.004830698482692242, + 0.004341014660894871, + 0.003343335585668683, + 0.003910321742296219, + 0.005120463203638792, + 0.0035829092375934124, + 0.007880686782300472, + 0.011457378976047039, + 0.008903464302420616, + 0.007801542058587074, + 0.00517963757738471, + 0.00795834232121706, + 0.014814650639891624, + 0.004674976691603661, + 0.028294362127780914, + 0.012151803821325302, + 0.00881450530141592, + 0.011123638600111008, + 0.006660935003310442, + 0.005226874724030495, + 0.008602242916822433, + 0.008386360481381416, + 0.010306904092431068, + 0.02392483316361904, + 0.017417242750525475, + 0.01843075267970562, + 0.017561640590429306, + 0.021150454878807068, + 0.028538845479488373, + 0.016103470697999, + 0.017718538641929626, + 0.0074661667458713055, + 0.008627048693597317, + 0.005318128503859043, + 0.00772430095821619, + 0.0072423722594976425, + 0.007819859310984612, + 0.004261234775185585, + 0.005309551488608122, + 0.008586253970861435, + 0.008255488239228725, + 0.005450275726616383, + 0.007989639416337013, + 0.006338542327284813, + 0.002994536655023694, + 0.006262287497520447, + 0.004692086949944496, + 0.012101789936423302, + 0.02740411087870598, + 0.01659916713833809, + 0.016695883125066757, + 0.020114408805966377, + 0.010793509893119335, + 0.010619716718792915, + 0.01168180163949728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 9 + ], + "selected_reward": -0.1270282119512558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736189126968384 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736189126968384, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736189126968384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22691617906093597, + 0.06917648017406464, + 0.03901319205760956, + 0.04244751110672951, + 0.13547037541866302, + 0.036783091723918915, + 0.11507760733366013, + 0.06429702043533325, + 0.0409744530916214, + 0.2298441082239151 + ], + "credit_map": [ + 0.0400579497218132, + 0.015918051823973656, + 0.024873139336705208, + 0.012183012440800667, + 0.01730484701693058, + 0.03378189727663994, + 0.011606188490986824, + 0.07119109481573105, + 0.010924591682851315, + 0.005421068519353867, + 0.007490098010748625, + 0.009974134154617786, + 0.004219328053295612, + 0.009602501057088375, + 0.003926875535398722, + 0.017617886886000633, + 0.005777448881417513, + 0.003669548546895385, + 0.004049357026815414, + 0.004160786978900433, + 0.0033793491311371326, + 0.005016669165343046, + 0.0033843875862658024, + 0.009575645439326763, + 0.006814491935074329, + 0.0033953767269849777, + 0.004701824858784676, + 0.003430767683312297, + 0.003389298915863037, + 0.005776784848421812, + 0.003723540110513568, + 0.01121542602777481, + 0.016500290483236313, + 0.018130553886294365, + 0.01468590646982193, + 0.009848633781075478, + 0.012275912798941135, + 0.019880017265677452, + 0.017896147444844246, + 0.02625291794538498, + 0.006902570370584726, + 0.0031902047339826822, + 0.005367453675717115, + 0.004320536740124226, + 0.0028796005062758923, + 0.004098035860806704, + 0.005449802149087191, + 0.004574890714138746, + 0.02257208526134491, + 0.011389744468033314, + 0.01521595474332571, + 0.01455674134194851, + 0.01028816681355238, + 0.014114378951489925, + 0.010477878153324127, + 0.016462653875350952, + 0.015306960791349411, + 0.015167823061347008, + 0.006606494076550007, + 0.006680944003164768, + 0.005474864039570093, + 0.006450086832046509, + 0.0038893527816981077, + 0.0047204988077282906, + 0.005803469568490982, + 0.007642021868377924, + 0.004917875397950411, + 0.005553579423576593, + 0.0053854756988584995, + 0.005116206593811512, + 0.0035570799373090267, + 0.0029987443704158068, + 0.03502083197236061, + 0.02759958803653717, + 0.03368741646409035, + 0.05465347319841385, + 0.025777315720915794, + 0.017022863030433655, + 0.021017853170633316, + 0.01506475917994976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.006736189126968384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward perspective, initially static, focuses on a crow perched on a bare, cracked, gray branch. Its inky black feathers gleam like a metallic sheen in the soft, diffuse light of an overcast sky. Its sharp, alert black eyes survey its surroundings, its body tense, as if poised for a sudden explosion. The background is a pale, gloomy, soft-focus sky, a blurry, milky white canvas. Suddenly, it pushes away from the branch, decisively and swiftly spreading its massive black wings. With a powerful downward flap, its entire body instantly soars into the air. The camera then rapidly tilts upward, capturing its outstretched wings as they sweep across the sky in a soft-focus frame, transforming into an elegant yet swift black arc before disappearing from view. The entire process, captured in minimalist black and gray tones, creates a dramatic tension of loneliness, awe, and primal vitality.", + "seed": 27639, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_4/index_2173_seed_27639_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.008691758173517883, + "vlm_logprob": -0.008691758173517883, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.008691758173517883, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00983441062271595 + ], + "opposite_logprob_tokens": [ + -6.134834289550781 + ], + "logprob_reward": -0.00983441062271595, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.00983441062271595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05682962387800217, + 0.04348758980631828, + 0.0646989643573761, + 0.03692910447716713, + 0.0427798256278038, + 0.07165117561817169, + 0.18751060962677002, + 0.13812081515789032, + 0.11425167322158813, + 0.24374064803123474 + ], + "credit_map": [ + 0.007310737855732441, + 0.00595643138512969, + 0.01228346023708582, + 0.004623889923095703, + 0.003222377272322774, + 0.010375574231147766, + 0.002632236573845148, + 0.01042491290718317, + 0.007763110101222992, + 0.0024871279019862413, + 0.0037226402200758457, + 0.003528028028085828, + 0.0016201641410589218, + 0.008480512537062168, + 0.00302324746735394, + 0.012862752191722393, + 0.008923701010644436, + 0.007415121886879206, + 0.008985036052763462, + 0.00449136458337307, + 0.010401015169918537, + 0.010583538562059402, + 0.004069445189088583, + 0.00982972513884306, + 0.0054367356933653355, + 0.006505713332444429, + 0.0032984165009111166, + 0.002848482457920909, + 0.004245552234351635, + 0.005130931735038757, + 0.0026522469706833363, + 0.0068110255524516106, + 0.005006063263863325, + 0.005087139084935188, + 0.006369445007294416, + 0.005031587556004524, + 0.004637794569134712, + 0.007374454289674759, + 0.0034267196897417307, + 0.005846616346389055, + 0.007904418744146824, + 0.008055875077843666, + 0.009972603991627693, + 0.006797199137508869, + 0.007298928685486317, + 0.009733730927109718, + 0.010057657957077026, + 0.011830754578113556, + 0.029836906120181084, + 0.03871919959783554, + 0.022611523047089577, + 0.009717365726828575, + 0.03824006766080856, + 0.018802853301167488, + 0.013388392515480518, + 0.016194291412830353, + 0.009956417605280876, + 0.0198762696236372, + 0.016833599656820297, + 0.015265892259776592, + 0.015498204156756401, + 0.022038986906409264, + 0.016217270866036415, + 0.022434158250689507, + 0.006469754967838526, + 0.04338637739419937, + 0.006362645886838436, + 0.0065414817072451115, + 0.024252772331237793, + 0.011705329641699791, + 0.005478094797581434, + 0.010055203922092915, + 0.024232205003499985, + 0.01846208982169628, + 0.09569885581731796, + 0.02263747714459896, + 0.040173377841711044, + 0.018284911289811134, + 0.014161082915961742, + 0.010090633295476437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.00983441062271595 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008664513006806374 + ], + "opposite_logprob_tokens": [ + -4.758664608001709 + ], + "logprob_reward": -0.008664513006806374, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008664513006806374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0504976287484169, + 0.055997319519519806, + 0.1230592206120491, + 0.05399462953209877, + 0.03480800613760948, + 0.07458055764436722, + 0.12191280722618103, + 0.05561927333474159, + 0.21800628304481506, + 0.21152426302433014 + ], + "credit_map": [ + 0.006450035143643618, + 0.004543424118310213, + 0.005841022823005915, + 0.003832933260127902, + 0.0025607943534851074, + 0.008573705330491066, + 0.003225244814530015, + 0.015470471233129501, + 0.009627598337829113, + 0.0032951384782791138, + 0.006019323132932186, + 0.009914310649037361, + 0.0022071930579841137, + 0.009224134497344494, + 0.0041711959056556225, + 0.011538425460457802, + 0.016256416216492653, + 0.01255960576236248, + 0.014485899358987808, + 0.008811380714178085, + 0.021233869716525078, + 0.021971266716718674, + 0.006066717207431793, + 0.021674059331417084, + 0.009693654254078865, + 0.007620878051966429, + 0.004500807262957096, + 0.003073977306485176, + 0.006423512008041143, + 0.01236729882657528, + 0.003599779913201928, + 0.0067147244699299335, + 0.0036928243935108185, + 0.003285396611317992, + 0.005246405955404043, + 0.0037147642578929663, + 0.005444753915071487, + 0.005095779895782471, + 0.003303425619378686, + 0.005024655256420374, + 0.00815872848033905, + 0.010689744725823402, + 0.0110096400603652, + 0.009146591648459435, + 0.006298601161688566, + 0.009924683719873428, + 0.006438110023736954, + 0.012914462946355343, + 0.011312803253531456, + 0.0393088161945343, + 0.00820552371442318, + 0.004583690781146288, + 0.038806453347206116, + 0.007059535011649132, + 0.005801598075777292, + 0.006834385450929403, + 0.004839559085667133, + 0.01421933900564909, + 0.00604994548484683, + 0.005583812482655048, + 0.011165372096002102, + 0.004934023134410381, + 0.003558187047019601, + 0.005269037559628487, + 0.009481810964643955, + 0.09491388499736786, + 0.005888143088668585, + 0.011026917025446892, + 0.05779317393898964, + 0.022667882964015007, + 0.007881972007453442, + 0.008352501317858696, + 0.020897237583994865, + 0.017306825146079063, + 0.08593994379043579, + 0.021950198337435722, + 0.027571257203817368, + 0.014618447981774807, + 0.014830928295850754, + 0.008409421890974045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.008664513006806374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013429442187771201 + ], + "opposite_logprob_tokens": [ + -6.6263427734375 + ], + "logprob_reward": -0.0013429442187771201, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013429442187771201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029739592224359512, + 0.03130777180194855, + 0.024713486433029175, + 0.07397866994142532, + 0.12310176342725754, + 0.0801256000995636, + 0.1852625459432602, + 0.10883961617946625, + 0.18449972569942474, + 0.15843115746974945 + ], + "credit_map": [ + 0.004033998586237431, + 0.0033063979353755713, + 0.0032320234458893538, + 0.003428790718317032, + 0.001626966055482626, + 0.0051087308675050735, + 0.001867657178081572, + 0.007135030347853899, + 0.0048800320364534855, + 0.002103727078065276, + 0.0026184082962572575, + 0.0044785900972783566, + 0.0016608134610578418, + 0.006039068568497896, + 0.0019713432993739843, + 0.00755578838288784, + 0.0033610432874411345, + 0.0024362425319850445, + 0.0043266951106488705, + 0.0020663663744926453, + 0.0021800990216434, + 0.004469574894756079, + 0.0011585006723180413, + 0.004714966751635075, + 0.009218401275575161, + 0.015519645065069199, + 0.00877588614821434, + 0.0045343865640461445, + 0.0069773816503584385, + 0.010709092952311039, + 0.00516646821051836, + 0.013077421113848686, + 0.012815574184060097, + 0.010061499662697315, + 0.014300201088190079, + 0.019194522872567177, + 0.01625797338783741, + 0.02032107673585415, + 0.009268797002732754, + 0.0208821389824152, + 0.009442447684705257, + 0.008984079584479332, + 0.010951845906674862, + 0.007383336778730154, + 0.008291414007544518, + 0.014488189481198788, + 0.010030262172222137, + 0.010554030537605286, + 0.012444584630429745, + 0.07003387808799744, + 0.008126684464514256, + 0.004589504096657038, + 0.06932241469621658, + 0.00796063058078289, + 0.00665367441251874, + 0.00613120524212718, + 0.009989368729293346, + 0.023347020149230957, + 0.012637147679924965, + 0.011288035660982132, + 0.019560135900974274, + 0.01278342679142952, + 0.01006925106048584, + 0.009165233001112938, + 0.006301873829215765, + 0.07334200292825699, + 0.004802666138857603, + 0.007772132288664579, + 0.06767093390226364, + 0.011705789715051651, + 0.006216878071427345, + 0.006687452085316181, + 0.01426731701940298, + 0.013118180446326733, + 0.060360103845596313, + 0.014243584126234055, + 0.020052572712302208, + 0.014509972184896469, + 0.014141835272312164, + 0.0077376109547913074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.0013429442187771201 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074859648942947 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011074859648942947, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074859648942947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0651218518614769, + 0.03788938373327255, + 0.01663336344063282, + 0.05833686888217926, + 0.07510249316692352, + 0.07235937565565109, + 0.12145965546369553, + 0.1528429239988327, + 0.2621515393257141, + 0.13810254633426666 + ], + "credit_map": [ + 0.00923633947968483, + 0.006061161402612925, + 0.010822233743965626, + 0.012518303468823433, + 0.0031426269561052322, + 0.010739150457084179, + 0.004462504293769598, + 0.008139532059431076, + 0.005577784962952137, + 0.0033956870902329683, + 0.003398379310965538, + 0.00810337159782648, + 0.0023602175060659647, + 0.0064293635077774525, + 0.003214366966858506, + 0.005410211160778999, + 0.002273882506415248, + 0.0013318877900019288, + 0.0021533621475100517, + 0.0016867389203980565, + 0.0017556853126734495, + 0.002956608310341835, + 0.0008258564048446715, + 0.0036493432708084583, + 0.009831659495830536, + 0.010346832685172558, + 0.004637540318071842, + 0.002949146321043372, + 0.005445299670100212, + 0.010276119224727154, + 0.00420595146715641, + 0.010644320398569107, + 0.007645191624760628, + 0.006071838084608316, + 0.009694461710751057, + 0.009408229030668736, + 0.008930359967052937, + 0.01164017803966999, + 0.007808674592524767, + 0.013903560116887093, + 0.009420856833457947, + 0.009822623804211617, + 0.01167282648384571, + 0.008303303271532059, + 0.007192681543529034, + 0.010952435433864594, + 0.005223355256021023, + 0.009771294891834259, + 0.022271880879998207, + 0.02278001420199871, + 0.017117643728852272, + 0.01012815348803997, + 0.0218012947589159, + 0.008999500423669815, + 0.009045854210853577, + 0.009315311908721924, + 0.012733133509755135, + 0.03238939493894577, + 0.02030329406261444, + 0.016934795305132866, + 0.027434371411800385, + 0.01707109622657299, + 0.011612899601459503, + 0.01436393428593874, + 0.011222999542951584, + 0.06299855560064316, + 0.018327713012695312, + 0.01715564914047718, + 0.04156922549009323, + 0.08266807347536087, + 0.017250165343284607, + 0.010959165170788765, + 0.014436346478760242, + 0.012011419050395489, + 0.056322790682315826, + 0.016752837225794792, + 0.012878927402198315, + 0.00971272960305214, + 0.008073775097727776, + 0.007913729175925255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.011074859648942947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542063370347023 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542063370347023, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542063370347023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02795761078596115, + 0.04244682192802429, + 0.01621331088244915, + 0.039439138025045395, + 0.025394098833203316, + 0.11874178797006607, + 0.0984107255935669, + 0.12922927737236023, + 0.32089707255363464, + 0.18127018213272095 + ], + "credit_map": [ + 0.003978132735937834, + 0.001966634998098016, + 0.004787241108715534, + 0.004287092015147209, + 0.0018763677217066288, + 0.005196818616241217, + 0.0017629940994083881, + 0.0041023255325853825, + 0.005851447116583586, + 0.003920906223356724, + 0.006903661880642176, + 0.006251711398363113, + 0.0020761017221957445, + 0.007940640673041344, + 0.0030692811124026775, + 0.006433069705963135, + 0.0019236673833802342, + 0.0017795962048694491, + 0.0028445147909224033, + 0.0014182024169713259, + 0.0024902948644012213, + 0.0022795354016125202, + 0.0011497170198708773, + 0.002327781170606613, + 0.00647903373464942, + 0.007754412014037371, + 0.0031890100799500942, + 0.0027450716588646173, + 0.002879898063838482, + 0.007431395351886749, + 0.002710035303607583, + 0.006250275298953056, + 0.0032826669048517942, + 0.0026378200855106115, + 0.0040799882262945175, + 0.0029412703588604927, + 0.0022626689169555902, + 0.004186706151813269, + 0.0020308196544647217, + 0.003972156438976526, + 0.016135768964886665, + 0.02150525152683258, + 0.0196507778018713, + 0.013152357190847397, + 0.010350275784730911, + 0.014767658896744251, + 0.008194005116820335, + 0.014985675923526287, + 0.011378024704754353, + 0.026956891641020775, + 0.007459641434252262, + 0.00528362113982439, + 0.031243449077010155, + 0.006387940142303705, + 0.00496557867154479, + 0.00473557785153389, + 0.01494093518704176, + 0.020324446260929108, + 0.01694473624229431, + 0.013880860060453415, + 0.019488828256726265, + 0.016557326540350914, + 0.010690078139305115, + 0.016402067616581917, + 0.014974365942180157, + 0.1400647759437561, + 0.007388433441519737, + 0.015105833299458027, + 0.06069711595773697, + 0.06327696889638901, + 0.007287739310413599, + 0.012101821601390839, + 0.01821262761950493, + 0.012929143384099007, + 0.07086516916751862, + 0.01840459369122982, + 0.026307664811611176, + 0.015897447243332863, + 0.009757552295923233, + 0.008895965293049812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.012542063370347023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward perspective, initially static, focuses on a crow perched on a bare, cracked, gray branch. Its inky black feathers gleam like a metallic sheen in the soft, diffuse light of an overcast sky. Its sharp, alert black eyes survey its surroundings, its body tense, as if poised for a sudden explosion. The background is a pale, gloomy, soft-focus sky, a blurry, milky white canvas. Suddenly, it pushes away from the branch, decisively and swiftly spreading its massive black wings. With a powerful downward flap, its entire body instantly soars into the air. The camera then rapidly tilts upward, capturing its outstretched wings as they sweep across the sky in a soft-focus frame, transforming into an elegant yet swift black arc before disappearing from view. The entire process, captured in minimalist black and gray tones, creates a dramatic tension of loneliness, awe, and primal vitality.", + "seed": 27640, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_4/index_2173_seed_27640_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.00986575852148235, + "vlm_logprob": -0.00986575852148235, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.00986575852148235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009293628856539726 + ], + "opposite_logprob_tokens": [ + -5.384293556213379 + ], + "logprob_reward": -0.009293628856539726, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.009293628856539726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10323143005371094, + 0.09398973733186722, + 0.0683417022228241, + 0.07323653995990753, + 0.10718832910060883, + 0.07460031658411026, + 0.07255437225103378, + 0.18168184161186218, + 0.10500374436378479, + 0.12017197906970978 + ], + "credit_map": [ + 0.017875466495752335, + 0.009881456382572651, + 0.016357239335775375, + 0.009820427745580673, + 0.011250227689743042, + 0.012029877863824368, + 0.005596954375505447, + 0.020419778302311897, + 0.02258702553808689, + 0.00870460644364357, + 0.017406927421689034, + 0.006092860829085112, + 0.007760572712868452, + 0.015832189470529556, + 0.004305082838982344, + 0.011300470679998398, + 0.009167521260678768, + 0.005714617669582367, + 0.017173705622553825, + 0.006117607932537794, + 0.004635793622583151, + 0.008740120567381382, + 0.003609550418332219, + 0.013182786293327808, + 0.009318850934505463, + 0.00781736709177494, + 0.008517664857208729, + 0.008314663544297218, + 0.005728280637413263, + 0.008328163996338844, + 0.004800673574209213, + 0.020410878583788872, + 0.005742030218243599, + 0.007233981508761644, + 0.01817636750638485, + 0.0134765999391675, + 0.009617123752832413, + 0.01179504208266735, + 0.006669978611171246, + 0.034477200359106064, + 0.009322010912001133, + 0.012788393534719944, + 0.007411802653223276, + 0.007469280157238245, + 0.011092827655375004, + 0.010701674036681652, + 0.004907367285341024, + 0.010906961746513844, + 0.010879093781113625, + 0.008864014409482479, + 0.01050986535847187, + 0.007587475702166557, + 0.006616535130888224, + 0.009499303996562958, + 0.009116443805396557, + 0.009481635875999928, + 0.018554696813225746, + 0.06237348914146423, + 0.0661325454711914, + 0.009306983090937138, + 0.005911848973482847, + 0.0050679827108979225, + 0.0047447956167161465, + 0.009589487686753273, + 0.0087298434227705, + 0.015629829838871956, + 0.013548262417316437, + 0.027540702372789383, + 0.009284255094826221, + 0.00848229881376028, + 0.01193760335445404, + 0.009850951842963696, + 0.014395958743989468, + 0.014935852959752083, + 0.01548527367413044, + 0.010216210037469864, + 0.0197156872600317, + 0.01548284012824297, + 0.02053917944431305, + 0.009400970302522182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.009293628856539726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008666876703500748 + ], + "opposite_logprob_tokens": [ + -4.7586669921875 + ], + "logprob_reward": -0.008666876703500748, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008666876703500748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09038062393665314, + 0.06838475167751312, + 0.22244811058044434, + 0.06194008141756058, + 0.09650973230600357, + 0.042665380984544754, + 0.0342436246573925, + 0.081352099776268, + 0.028836991637945175, + 0.27323853969573975 + ], + "credit_map": [ + 0.013050056993961334, + 0.010001459158957005, + 0.01527228020131588, + 0.009206095710396767, + 0.00872219167649746, + 0.01080939918756485, + 0.005215565208345652, + 0.01810358092188835, + 0.010890156961977482, + 0.007354624103754759, + 0.011403022333979607, + 0.006226025056093931, + 0.008298158645629883, + 0.010336236096918583, + 0.004037601873278618, + 0.009838934987783432, + 0.05535160005092621, + 0.02121155709028244, + 0.02321719564497471, + 0.013275619596242905, + 0.009970090351998806, + 0.03254310041666031, + 0.0073369755409657955, + 0.059542011469602585, + 0.0097498819231987, + 0.008080961182713509, + 0.006568859331309795, + 0.005366652738302946, + 0.004318488761782646, + 0.007651215419173241, + 0.003297198796644807, + 0.016906827688217163, + 0.008526559919118881, + 0.00925473589450121, + 0.009046992287039757, + 0.01427173987030983, + 0.013576431199908257, + 0.010790474712848663, + 0.00810221303254366, + 0.022940602153539658, + 0.005680640693753958, + 0.0076602972112596035, + 0.0045255753211677074, + 0.004203151911497116, + 0.006194581743329763, + 0.006066294386982918, + 0.0026421761140227318, + 0.005692669656127691, + 0.004897845443338156, + 0.0037405123002827168, + 0.003719407832249999, + 0.0037551329005509615, + 0.003554446855559945, + 0.004229401238262653, + 0.005055603571236134, + 0.0052912794053554535, + 0.007296797353774309, + 0.022586259990930557, + 0.025477496907114983, + 0.007009592838585377, + 0.0045559764839708805, + 0.004429588094353676, + 0.0037361939903348684, + 0.006260203197598457, + 0.0030359807424247265, + 0.003098748391494155, + 0.005022092256695032, + 0.0031013928819447756, + 0.004698379430919886, + 0.002934020245447755, + 0.003170013427734375, + 0.0037763675209134817, + 0.02512798272073269, + 0.03183305263519287, + 0.03946478292346001, + 0.034434787929058075, + 0.06425781548023224, + 0.02422553114593029, + 0.03505152463912964, + 0.01884307712316513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.008666876703500748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013405631761997938 + ], + "opposite_logprob_tokens": [ + -6.626340389251709 + ], + "logprob_reward": -0.0013405631761997938, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013405631761997938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052415188401937485, + 0.060772839933633804, + 0.061976004391908646, + 0.05550968647003174, + 0.13534772396087646, + 0.10434871912002563, + 0.07614237070083618, + 0.07852538675069809, + 0.07511294633150101, + 0.29984912276268005 + ], + "credit_map": [ + 0.007165156304836273, + 0.004717860370874405, + 0.009111497551202774, + 0.0043830531649291515, + 0.003528245957568288, + 0.006468566600233316, + 0.0025715494994074106, + 0.014469261281192303, + 0.009471823461353779, + 0.007561282254755497, + 0.013380905613303185, + 0.004769760649651289, + 0.005561338271945715, + 0.008730738423764706, + 0.003007060382515192, + 0.008289928548038006, + 0.014605947770178318, + 0.006755861919373274, + 0.0063481866382062435, + 0.004207083489745855, + 0.004506173077970743, + 0.00855363067239523, + 0.002472246764227748, + 0.01452687755227089, + 0.0060091689229011536, + 0.008431605994701385, + 0.0067804669961333275, + 0.008623149245977402, + 0.005209201481193304, + 0.0067216684110462666, + 0.004012704826891422, + 0.009721722453832626, + 0.011368515901267529, + 0.014589781872928143, + 0.014965660870075226, + 0.01838618516921997, + 0.015450208447873592, + 0.015565469861030579, + 0.010976441204547882, + 0.03404545783996582, + 0.011702099815011024, + 0.011914847418665886, + 0.021153148263692856, + 0.012751692906022072, + 0.011238443665206432, + 0.012793726287782192, + 0.01016486156731844, + 0.012629898265004158, + 0.01287220511585474, + 0.009606427513062954, + 0.010845949873328209, + 0.006715742871165276, + 0.00670749181881547, + 0.00975753366947174, + 0.007005560211837292, + 0.012631461024284363, + 0.0072998409159481525, + 0.018260572105646133, + 0.019173022359609604, + 0.005647108890116215, + 0.006172087974846363, + 0.0077422610484063625, + 0.005706701893359423, + 0.00852379109710455, + 0.008737211115658283, + 0.009128383360803127, + 0.012431928887963295, + 0.008445951156318188, + 0.01015422958880663, + 0.006148949731141329, + 0.008883776143193245, + 0.011182515881955624, + 0.025985652580857277, + 0.029096372425556183, + 0.03606643155217171, + 0.03321971371769905, + 0.081791952252388, + 0.022187944501638412, + 0.05604821443557739, + 0.015452837571501732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.0013405631761997938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286443203687668 + ], + "opposite_logprob_tokens": [ + -3.7732863426208496 + ], + "logprob_reward": -0.023286443203687668, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286443203687668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027012299746274948, + 0.026119783520698547, + 0.044721487909555435, + 0.025540325790643692, + 0.045331619679927826, + 0.03467034921050072, + 0.030702097341418266, + 0.1341114342212677, + 0.10079892724752426, + 0.5309917330741882 + ], + "credit_map": [ + 0.004531729966402054, + 0.003994498401880264, + 0.003925170283764601, + 0.0022584218531847, + 0.0018996447324752808, + 0.0029941750690340996, + 0.0012944394256919622, + 0.006114219315350056, + 0.005363303702324629, + 0.0034215592313557863, + 0.005301218945533037, + 0.002276849700137973, + 0.0016019047470763326, + 0.0034800604917109013, + 0.0011123975273221731, + 0.0035624864976853132, + 0.008664213120937347, + 0.003555286442860961, + 0.007362631149590015, + 0.003231558483093977, + 0.0016236701048910618, + 0.005678684916347265, + 0.002230030717328191, + 0.01237540878355503, + 0.0030723954550921917, + 0.002994195092469454, + 0.003400749061256647, + 0.0019476303132250905, + 0.0014542334247380495, + 0.0032547456212341785, + 0.001438312465324998, + 0.007978062145411968, + 0.0022158208303153515, + 0.00269379373639822, + 0.003941407892853022, + 0.0071530649438500404, + 0.003984822425991297, + 0.004853594116866589, + 0.002829346340149641, + 0.017659766599535942, + 0.0048377737402915955, + 0.0063139814883470535, + 0.0037375108804553747, + 0.0033574418630450964, + 0.004786169622093439, + 0.004928397014737129, + 0.0019504318479448557, + 0.004758641589432955, + 0.006820840295404196, + 0.003441567998379469, + 0.0036640663165599108, + 0.0025694118812680244, + 0.0025755148380994797, + 0.003238277742639184, + 0.003436954692006111, + 0.004955460783094168, + 0.017580470070242882, + 0.03817949444055557, + 0.04359942674636841, + 0.00838074367493391, + 0.005463140085339546, + 0.006422929000109434, + 0.004922136198729277, + 0.009563080035150051, + 0.013318806886672974, + 0.01398228295147419, + 0.01987607032060623, + 0.014324361458420753, + 0.012559006921947002, + 0.007609691470861435, + 0.009180374443531036, + 0.009948314167559147, + 0.04266566038131714, + 0.05819006264209747, + 0.07000650465488434, + 0.06090160459280014, + 0.1436668485403061, + 0.036310337483882904, + 0.08479262888431549, + 0.034458015114068985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.023286443203687668 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a close-up of a crow perched on a bare, cracked, gray branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have inky black feathers with a metallic sheen and sharp black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow push off the branch and spread its massive black wings to soar upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera rapidly tilt upward to capture the crow's wings sweeping across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in minimalist black and gray tones with a soft-focus background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741280667483807 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.006741280667483807, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741280667483807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10841414332389832, + 0.027879800647497177, + 0.17044202983379364, + 0.03486733138561249, + 0.04234616085886955, + 0.08929750323295593, + 0.07093483954668045, + 0.2311481237411499, + 0.04877202585339546, + 0.17589810490608215 + ], + "credit_map": [ + 0.018295686691999435, + 0.01000523567199707, + 0.016530422493815422, + 0.008530043065547943, + 0.004236158914864063, + 0.014280037954449654, + 0.0038892794400453568, + 0.032647278159856796, + 0.004138409160077572, + 0.002642653416842222, + 0.005956416949629784, + 0.0025186811108142138, + 0.0023710436653345823, + 0.003270598128437996, + 0.0015482380986213684, + 0.005433760117739439, + 0.041803937405347824, + 0.012892434373497963, + 0.016114862635731697, + 0.01089001726359129, + 0.006010021548718214, + 0.025319905951619148, + 0.006139243021607399, + 0.05127159506082535, + 0.005105690564960241, + 0.003938968759030104, + 0.0038103971164673567, + 0.004001286346465349, + 0.0027920398861169815, + 0.003939478192478418, + 0.001982834655791521, + 0.00929663609713316, + 0.003940454684197903, + 0.0039930157363414764, + 0.003758747363463044, + 0.006328300107270479, + 0.004636105615645647, + 0.004960541147738695, + 0.0031977505423128605, + 0.011531246826052666, + 0.011813275516033173, + 0.01020385418087244, + 0.011198385618627071, + 0.007826728746294975, + 0.009438584558665752, + 0.012663657777011395, + 0.00773437088355422, + 0.01841863989830017, + 0.011768000200390816, + 0.008620360866189003, + 0.009492579847574234, + 0.007309275679290295, + 0.007260789629071951, + 0.009572350420057774, + 0.007092443760484457, + 0.00981904473155737, + 0.016480186954140663, + 0.07823682576417923, + 0.08310778439044952, + 0.011901244521141052, + 0.008314638398587704, + 0.010987496003508568, + 0.008639253675937653, + 0.013480685651302338, + 0.006209328770637512, + 0.007191504817456007, + 0.007512028329074383, + 0.010980747640132904, + 0.003961692098528147, + 0.0025879002641886473, + 0.006891048513352871, + 0.003437774023041129, + 0.013604840263724327, + 0.019376499578356743, + 0.027574991807341576, + 0.024350767955183983, + 0.040845952928066254, + 0.01300019957125187, + 0.02671392820775509, + 0.010430927388370037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.006741280667483807 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_5/index_tensor([8576, 8576])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_5/index_tensor([8576, 8576])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f848ddef425c61aefe50aa050ea1b59bed6a1367 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_5/index_tensor([8576, 8576])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static wide-angle shot captures a remarkably tranquil body of water, its surface a calm, tea-colored, light brown liquid, suggestive of suspended sediment at the pond's bottom. Gentle, almost imperceptible ripples drift across the surface, capturing the warm, diffuse glow of the afternoon sun and creating a subtle, shimmering texture. In the lower right corner of the frame, two weathered rocks interrupt the water's calm, creating a rugged contrast. The larger rock in the foreground, a solid dark gray, its uneven, damp surface partially covered in vibrant green moss, sits heavily in its semi-submerged position, serving as a solid anchor for the composition. Just behind it, a smaller, similarly textured rock creates a sense of depth. The camera remains perfectly still, allowing the subtle friction of water against stone to be the only subtle movement in a scene dominated by a warm, earthy palette, evoking an atmosphere of deep, contemplative tranquility and undisturbed natural peace.", + "seed": 29598, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_5/index_8576_seed_29598_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15955201382748782, + "vlm_logprob": -0.15955201382748782, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.15955201382748782, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12517417967319489 + ], + "opposite_logprob_tokens": [ + -2.3751742839813232 + ], + "logprob_reward": -0.12517417967319489, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12517417967319489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0705709382891655, + 0.2935768663883209, + 0.019819121807813644, + 0.007305392529815435, + 0.31297504901885986, + 0.11915673315525055, + 0.041864361613988876, + 0.06506694853305817, + 0.03851793706417084, + 0.031146634370088577 + ], + "credit_map": [ + 0.009324540384113789, + 0.010513518936932087, + 0.007377460598945618, + 0.004998081363737583, + 0.012975514866411686, + 0.016290662810206413, + 0.0044345553033053875, + 0.0046566021628677845, + 0.009780151769518852, + 0.12757864594459534, + 0.004821532405912876, + 0.003966137301176786, + 0.015670115128159523, + 0.1249549463391304, + 0.0027499233838170767, + 0.004055417142808437, + 0.002035995479673147, + 0.004488030448555946, + 0.002104982966557145, + 0.0013719693524762988, + 0.0029756955336779356, + 0.0046193948946893215, + 0.001263159909285605, + 0.0009598920005373657, + 0.000809977762401104, + 0.0010725902393460274, + 0.0010013715364038944, + 0.0005389780271798372, + 0.0011489744065329432, + 0.0014112386852502823, + 0.000693502661306411, + 0.0006287590367719531, + 0.012891946360468864, + 0.13565827906131744, + 0.009517358615994453, + 0.007700987160205841, + 0.055538859218358994, + 0.08182866871356964, + 0.005127383396029472, + 0.004711557179689407, + 0.003952978644520044, + 0.009434902109205723, + 0.0030600738245993853, + 0.0023406087420880795, + 0.052301835268735886, + 0.044119007885456085, + 0.0020031335297971964, + 0.0019442031625658274, + 0.0031504821963608265, + 0.004672209732234478, + 0.002172897569835186, + 0.0023080362007021904, + 0.014314665459096432, + 0.01140211895108223, + 0.002505967393517494, + 0.0013379895826801658, + 0.0026446592528373003, + 0.02722793072462082, + 0.0012805956648662686, + 0.0012174647999927402, + 0.02660980634391308, + 0.0035175757948309183, + 0.0014992542564868927, + 0.0010696683311834931, + 0.0013679750263690948, + 0.015400355681777, + 0.001190122333355248, + 0.0008826832636259496, + 0.002470439299941063, + 0.015682056546211243, + 0.0007888989639468491, + 0.0007354050176218152, + 0.0013430026592686772, + 0.010614085011184216, + 0.001003991812467575, + 0.0007531047449447215, + 0.013364005833864212, + 0.003067078534513712, + 0.0004611291515175253, + 0.0005402363603934646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.12517417967319489 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328402161598206 + ], + "opposite_logprob_tokens": [ + -0.7578402161598206 + ], + "logprob_reward": -0.6328402161598206, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328402161598206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05216320604085922, + 0.3383353352546692, + 0.022766750305891037, + 0.008982724510133266, + 0.07848962396383286, + 0.08413343876600266, + 0.04169977456331253, + 0.11565744876861572, + 0.10372177511453629, + 0.15404996275901794 + ], + "credit_map": [ + 0.004825112409889698, + 0.01268253568559885, + 0.003595735179260373, + 0.0029844327364116907, + 0.015918059274554253, + 0.007740525528788567, + 0.0022732699289917946, + 0.002143535064533353, + 0.010500837117433548, + 0.14934349060058594, + 0.004930765368044376, + 0.003994086757302284, + 0.019025787711143494, + 0.1423036754131317, + 0.0034806050825864077, + 0.004756104201078415, + 0.0031175625044852495, + 0.004312530159950256, + 0.0029082249384373426, + 0.0018232642905786633, + 0.0032696989364922047, + 0.004450365900993347, + 0.001610766164958477, + 0.0012743385741487145, + 0.0009488115319982171, + 0.0012485668994486332, + 0.001259439974091947, + 0.0007713448721915483, + 0.0018713375320658088, + 0.0014985608868300915, + 0.0008480745018459857, + 0.0005365882534533739, + 0.003735405160114169, + 0.03323618322610855, + 0.002422141144052148, + 0.0022355548571795225, + 0.014812747947871685, + 0.019010167568922043, + 0.0014716596342623234, + 0.001565762679092586, + 0.0032539265230298042, + 0.006837767548859119, + 0.0026671835221350193, + 0.0017606545006856322, + 0.03594832122325897, + 0.030248120427131653, + 0.001998772146180272, + 0.0014186907792463899, + 0.0035126630682498217, + 0.004214808810502291, + 0.0029540937393903732, + 0.0022685183212161064, + 0.011477502062916756, + 0.012789957225322723, + 0.0029233742970973253, + 0.0015588597161695361, + 0.003204124514013529, + 0.04023957625031471, + 0.0018845346057787538, + 0.0013383813202381134, + 0.05274707078933716, + 0.013154094107449055, + 0.001776456949301064, + 0.001313217799179256, + 0.010126186534762383, + 0.031664419919252396, + 0.007987414486706257, + 0.006403209641575813, + 0.01803032122552395, + 0.017686745151877403, + 0.0059343306347727776, + 0.005889144726097584, + 0.005110130179673433, + 0.05985293164849281, + 0.00359381758607924, + 0.0027621581684798002, + 0.06726958602666855, + 0.009902995079755783, + 0.0027640406042337418, + 0.0027943048626184464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.6328402161598206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410274276509881 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.00410274276509881, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410274276509881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07218852639198303, + 0.4450434744358063, + 0.018743781372904778, + 0.011656830087304115, + 0.15716996788978577, + 0.04427095502614975, + 0.008130249567329884, + 0.0884869396686554, + 0.041736967861652374, + 0.11257226765155792 + ], + "credit_map": [ + 0.00945145171135664, + 0.012349717319011688, + 0.007367527112364769, + 0.005065761506557465, + 0.012270666658878326, + 0.01652519777417183, + 0.005369945429265499, + 0.0037882623728364706, + 0.01387840323150158, + 0.19504857063293457, + 0.006918045226484537, + 0.0058564599603414536, + 0.02426282688975334, + 0.18858776986598969, + 0.004703625570982695, + 0.005787809379398823, + 0.001349473255686462, + 0.0025750482454895973, + 0.0014470623573288321, + 0.0010108078131452203, + 0.006058974657207727, + 0.004697494674474001, + 0.0007495680474676192, + 0.0008553536608815193, + 0.0014082312118262053, + 0.0018074971158057451, + 0.0014921199763193727, + 0.0010412196861580014, + 0.0020216091070324183, + 0.0025515651796013117, + 0.0006314748898148537, + 0.0007031136192381382, + 0.007420441601425409, + 0.06567398458719254, + 0.005669974721968174, + 0.004708445630967617, + 0.031224289909005165, + 0.03541335090994835, + 0.004125847015529871, + 0.0029336651787161827, + 0.0021717301569879055, + 0.005841237027198076, + 0.0016160878585651517, + 0.0015019546262919903, + 0.018229695037007332, + 0.012560061179101467, + 0.0013312628725543618, + 0.0010189312743023038, + 0.0012271783780306578, + 0.0013303853338584304, + 0.0010896071325987577, + 0.0006802140851505101, + 0.001141566433943808, + 0.001410284312441945, + 0.0009055830887518823, + 0.00034543106448836625, + 0.002511525759473443, + 0.03149472922086716, + 0.0016430788673460484, + 0.0010942243970930576, + 0.04008883237838745, + 0.009732517413794994, + 0.0012406916357576847, + 0.0006813466898165643, + 0.0012933183461427689, + 0.010495751164853573, + 0.001396532286889851, + 0.0012760100653395057, + 0.007557119708508253, + 0.017581278458237648, + 0.0011266663204878569, + 0.0010102966334670782, + 0.0033696922473609447, + 0.045624990016222, + 0.0020048904698342085, + 0.001997942803427577, + 0.05049167573451996, + 0.006955363787710667, + 0.0011771736899390817, + 0.0009505588095635176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.00410274276509881 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370444104075432 + ], + "opposite_logprob_tokens": [ + -3.4087045192718506 + ], + "logprob_reward": -0.03370444104075432, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370444104075432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0447152815759182, + 0.17329491674900055, + 0.031010396778583527, + 0.010299336165189743, + 0.07888457179069519, + 0.4524960219860077, + 0.008572136983275414, + 0.07128330320119858, + 0.07555444538593292, + 0.05388959124684334 + ], + "credit_map": [ + 0.005457001738250256, + 0.008437979966402054, + 0.0068353405222296715, + 0.0041830819100141525, + 0.007078584283590317, + 0.007642944809049368, + 0.0025335883256047964, + 0.002546759322285652, + 0.005289190914481878, + 0.07264293730258942, + 0.0028598764911293983, + 0.00387941999360919, + 0.009512442164123058, + 0.07427038252353668, + 0.002277567284181714, + 0.0025631077587604523, + 0.002520443871617317, + 0.004062476102262735, + 0.0032871211878955364, + 0.0029173530638217926, + 0.007064892444759607, + 0.00773812597617507, + 0.002049434231594205, + 0.001370552578009665, + 0.0009189798147417605, + 0.0015955546405166388, + 0.0013582012616097927, + 0.001198407611809671, + 0.0020627868361771107, + 0.0012936177663505077, + 0.0011600716970860958, + 0.0007117163040675223, + 0.0032629494089633226, + 0.03422698378562927, + 0.001814955729059875, + 0.0023382622748613358, + 0.015717599540948868, + 0.01873374916613102, + 0.0013547735288739204, + 0.0014352976577356458, + 0.01904524862766266, + 0.03666539490222931, + 0.02464115619659424, + 0.012440881691873074, + 0.17429529130458832, + 0.1532345861196518, + 0.018858565017580986, + 0.013314907439053059, + 0.0006508114165626466, + 0.0023809457197785378, + 0.001055423985235393, + 0.000580166291911155, + 0.0010687008034437895, + 0.0018380212131887674, + 0.0005655501736328006, + 0.00043251720489934087, + 0.005495366640388966, + 0.02344965748488903, + 0.0029109951574355364, + 0.0022436303552240133, + 0.0085090771317482, + 0.02296665497124195, + 0.0033023536670953035, + 0.00240556662902236, + 0.0045600454322993755, + 0.010580133646726608, + 0.004789438098669052, + 0.0031615965999662876, + 0.027360713109374046, + 0.018027810379862785, + 0.0034764907322824, + 0.0035982185509055853, + 0.002166916849091649, + 0.019094478338956833, + 0.002019134582951665, + 0.0013361989986151457, + 0.02274959720671177, + 0.0043549672700464725, + 0.0010417649755254388, + 0.0011265343055129051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.03370444104075432 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019384894985705614 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019384894985705614, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019384894985705614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04702821373939514, + 0.3679460287094116, + 0.027517210692167282, + 0.01335460226982832, + 0.10024415701627731, + 0.11942316591739655, + 0.01244689803570509, + 0.10122483968734741, + 0.03292662277817726, + 0.17788822948932648 + ], + "credit_map": [ + 0.005640207789838314, + 0.007681285031139851, + 0.004966227803379297, + 0.0038905867841094732, + 0.008355755358934402, + 0.009569613263010979, + 0.0038028559647500515, + 0.003121687564998865, + 0.011376303620636463, + 0.1614389270544052, + 0.005542303901165724, + 0.004927898291498423, + 0.02003919705748558, + 0.15605434775352478, + 0.0037384917959570885, + 0.004828630480915308, + 0.0016102994559332728, + 0.003414752194657922, + 0.0017760874470695853, + 0.001339406124316156, + 0.00903030950576067, + 0.007856306619942188, + 0.0014147300971671939, + 0.0010753219248726964, + 0.0014886828139424324, + 0.001965628005564213, + 0.0017347864340990782, + 0.0013655475340783596, + 0.002238266170024872, + 0.002820282708853483, + 0.0009514998528175056, + 0.0007899099146015942, + 0.006601887755095959, + 0.03569710627198219, + 0.006092703435570002, + 0.00609058840200305, + 0.014406244270503521, + 0.02287205494940281, + 0.00514624360948801, + 0.0033373390324413776, + 0.003901841351762414, + 0.008506338112056255, + 0.003930622711777687, + 0.0017266682116314769, + 0.051244527101516724, + 0.045830272138118744, + 0.002355782315135002, + 0.001927131088450551, + 0.0013781874440610409, + 0.002543893177062273, + 0.00168606115039438, + 0.0009750822209753096, + 0.002156901406124234, + 0.0019498441834002733, + 0.0011144462041556835, + 0.0006424837047234178, + 0.008383807726204395, + 0.009630806744098663, + 0.007017931435257196, + 0.0042426325380802155, + 0.03444948419928551, + 0.028662817552685738, + 0.005448577459901571, + 0.0033887908793985844, + 0.0013462910428643227, + 0.010197709314525127, + 0.0011542150750756264, + 0.0010263168951496482, + 0.0038765498902648687, + 0.0135949170216918, + 0.0007851891568861902, + 0.0009454381070099771, + 0.005587195046246052, + 0.06987703591585159, + 0.003709128824993968, + 0.003462081542238593, + 0.07891087234020233, + 0.01159337256103754, + 0.0024624131619930267, + 0.0022861394099891186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.0019384894985705614 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide-angle shot captures a remarkably tranquil body of water, its surface a calm, tea-colored, light brown liquid, suggestive of suspended sediment at the pond's bottom. Gentle, almost imperceptible ripples drift across the surface, capturing the warm, diffuse glow of the afternoon sun and creating a subtle, shimmering texture. In the lower right corner of the frame, two weathered rocks interrupt the water's calm, creating a rugged contrast. The larger rock in the foreground, a solid dark gray, its uneven, damp surface partially covered in vibrant green moss, sits heavily in its semi-submerged position, serving as a solid anchor for the composition. Just behind it, a smaller, similarly textured rock creates a sense of depth. The camera remains perfectly still, allowing the subtle friction of water against stone to be the only subtle movement in a scene dominated by a warm, earthy palette, evoking an atmosphere of deep, contemplative tranquility and undisturbed natural peace.", + "seed": 29599, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_5/index_8576_seed_29599_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21466834405437113, + "vlm_logprob": -0.21466834405437113, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.21466834405437113, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1410759836435318 + ], + "opposite_logprob_tokens": [ + -2.26607608795166 + ], + "logprob_reward": -0.1410759836435318, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1410759836435318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11634133011102676, + 0.10412756353616714, + 0.2882763743400574, + 0.03899559751152992, + 0.16754218935966492, + 0.08324454724788666, + 0.12757787108421326, + 0.026093091815710068, + 0.017271412536501884, + 0.030530013144016266 + ], + "credit_map": [ + 0.007378836162388325, + 0.00882540363818407, + 0.01383276004344225, + 0.02964925579726696, + 0.018702317029237747, + 0.016730835661292076, + 0.01348165050148964, + 0.0077402638271451, + 0.006397986318916082, + 0.012798321433365345, + 0.014074500650167465, + 0.024855948984622955, + 0.015675533562898636, + 0.012332570739090443, + 0.009687511250376701, + 0.008305189199745655, + 0.01590096205472946, + 0.011052594520151615, + 0.10132798552513123, + 0.012425202876329422, + 0.024993108585476875, + 0.1100655347108841, + 0.006683683488518, + 0.005827291868627071, + 0.0028824477922171354, + 0.003207283327355981, + 0.005108975805342197, + 0.00956691987812519, + 0.0048880064859986305, + 0.0050918529741466045, + 0.004970721434801817, + 0.0032793839927762747, + 0.010987521149218082, + 0.01790405809879303, + 0.018282856792211533, + 0.025835677981376648, + 0.024443821981549263, + 0.030254704877734184, + 0.019371068105101585, + 0.020462462678551674, + 0.0047757625579833984, + 0.006239268928766251, + 0.00915172602981329, + 0.013274068012833595, + 0.011840512044727802, + 0.01490771770477295, + 0.011438083834946156, + 0.01161740068346262, + 0.007601687218993902, + 0.014241951517760754, + 0.02582850679755211, + 0.01407131738960743, + 0.017938027158379555, + 0.028573019430041313, + 0.012851833365857601, + 0.006471526809036732, + 0.0017943503335118294, + 0.0026901320088654757, + 0.003441395703703165, + 0.004055520053952932, + 0.00395899498835206, + 0.0038268635980784893, + 0.0029492531903088093, + 0.003376577515155077, + 0.0012955106794834137, + 0.0025254145730286837, + 0.0027302487287670374, + 0.0024098199792206287, + 0.0029694095719605684, + 0.0026788890827447176, + 0.0015262194210663438, + 0.0011358977062627673, + 0.0018879282288253307, + 0.0031699954997748137, + 0.007517063058912754, + 0.003134175669401884, + 0.0031916683074086905, + 0.0072021144442260265, + 0.0020393168088048697, + 0.002387748099863529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.1410759836435318 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983590602874756 + ], + "opposite_logprob_tokens": [ + -0.5233590602874756 + ], + "logprob_reward": -0.8983590602874756, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983590602874756, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2634779214859009, + 0.030000140890479088, + 0.24724426865577698, + 0.01994754932820797, + 0.32276248931884766, + 0.022335562855005264, + 0.02517400123178959, + 0.016311410814523697, + 0.017805680632591248, + 0.0349409393966198 + ], + "credit_map": [ + 0.015578522346913815, + 0.02113422192633152, + 0.03793787956237793, + 0.03669613599777222, + 0.04098423197865486, + 0.0523628406226635, + 0.026312993839383125, + 0.03247109055519104, + 0.001917892717756331, + 0.002560898894444108, + 0.004139287862926722, + 0.006303867790848017, + 0.004427590873092413, + 0.003337838687002659, + 0.003722352208569646, + 0.0035904129035770893, + 0.0138887669891119, + 0.010065337643027306, + 0.08460294455289841, + 0.012949855998158455, + 0.020370008423924446, + 0.09352746605873108, + 0.0069480775855481625, + 0.004891808610409498, + 0.0016034779837355018, + 0.002709164982661605, + 0.002816367195919156, + 0.003152267076075077, + 0.0036219735629856586, + 0.0027545748744159937, + 0.0020628338679671288, + 0.0012268904829397798, + 0.022856703028082848, + 0.03577249497175217, + 0.03526747599244118, + 0.06223082169890404, + 0.04548616707324982, + 0.05054625868797302, + 0.03367176279425621, + 0.036930836737155914, + 0.0018795952200889587, + 0.0027417957317084074, + 0.002931013936176896, + 0.0036258711479604244, + 0.0039539341814816, + 0.0032222040463238955, + 0.0019978571217507124, + 0.0019832919351756573, + 0.002192932413890958, + 0.0027260081842541695, + 0.00277406326495111, + 0.00347298476845026, + 0.004410397727042437, + 0.003961736336350441, + 0.003463736502453685, + 0.002172142034396529, + 0.0009842792060226202, + 0.0014386451803147793, + 0.0028039116878062487, + 0.0021386879961937666, + 0.0019465211080387235, + 0.0035443229135125875, + 0.0014167471090331674, + 0.002038294682279229, + 0.0011372683802619576, + 0.002041787374764681, + 0.002947930246591568, + 0.0022231093607842922, + 0.002607962815091014, + 0.002941949525848031, + 0.0018460057908669114, + 0.0020596664398908615, + 0.0028290930204093456, + 0.003515881486237049, + 0.007258584722876549, + 0.0053313118405640125, + 0.0035042583476752043, + 0.006417445372790098, + 0.0028095534071326256, + 0.0032748111989349127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.8983590602874756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964694079011679 + ], + "opposite_logprob_tokens": [ + -5.130964756011963 + ], + "logprob_reward": -0.005964694079011679, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964694079011679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12534818053245544, + 0.06299518793821335, + 0.052816975861787796, + 0.19682615995407104, + 0.17251260578632355, + 0.1014375239610672, + 0.030568446964025497, + 0.057077351957559586, + 0.10106636583805084, + 0.09935113042593002 + ], + "credit_map": [ + 0.009326721541583538, + 0.015779288485646248, + 0.014804257079958916, + 0.019621064886450768, + 0.02378508262336254, + 0.022381329908967018, + 0.012051990255713463, + 0.007598449010401964, + 0.0037315129302442074, + 0.0068129380233585835, + 0.007606717757880688, + 0.01153458934277296, + 0.01062869094312191, + 0.009816726669669151, + 0.006370718590915203, + 0.006493291351944208, + 0.003752009244635701, + 0.007422704715281725, + 0.00640486553311348, + 0.004891807213425636, + 0.01719946600496769, + 0.007276751566678286, + 0.0033338358625769615, + 0.0025355375837534666, + 0.012863615527749062, + 0.017918091267347336, + 0.01944582723081112, + 0.03461218997836113, + 0.028755126520991325, + 0.03163246810436249, + 0.02347302809357643, + 0.028125816956162453, + 0.011494536884129047, + 0.017951341345906258, + 0.019134454429149628, + 0.028786763548851013, + 0.026990188285708427, + 0.0303824283182621, + 0.01925065740942955, + 0.018522238358855247, + 0.006270548794418573, + 0.008436706848442554, + 0.010318426415324211, + 0.01629633828997612, + 0.015294712968170643, + 0.01774757169187069, + 0.013165101408958435, + 0.013908117078244686, + 0.002250144723802805, + 0.004892433527857065, + 0.004823110532015562, + 0.003381991060450673, + 0.006991971749812365, + 0.005139561370015144, + 0.0017099587712436914, + 0.0013792766258120537, + 0.0031681875698268414, + 0.004227727651596069, + 0.01685461960732937, + 0.003328423947095871, + 0.007775210775434971, + 0.01786637306213379, + 0.0024667575489729643, + 0.0013900548219680786, + 0.006875723600387573, + 0.009414897300302982, + 0.013305764645338058, + 0.013952765613794327, + 0.014329524710774422, + 0.0192402433604002, + 0.011747215874493122, + 0.01220023538917303, + 0.006091201677918434, + 0.005493644624948502, + 0.030634406954050064, + 0.00806751474738121, + 0.008753125555813313, + 0.030453970655798912, + 0.004977147560566664, + 0.00488011259585619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.005964694079011679 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301932960748672 + ], + "opposite_logprob_tokens": [ + -3.773301839828491 + ], + "logprob_reward": -0.023301932960748672, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301932960748672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3237239420413971, + 0.06355325877666473, + 0.06397164613008499, + 0.0778522789478302, + 0.27342507243156433, + 0.1002093181014061, + 0.013022341765463352, + 0.017282215878367424, + 0.01027175784111023, + 0.056688204407691956 + ], + "credit_map": [ + 0.01913304813206196, + 0.02353730984032154, + 0.05306202545762062, + 0.04500366747379303, + 0.04525052011013031, + 0.06507349014282227, + 0.03186783939599991, + 0.040796034038066864, + 0.0038704893086105585, + 0.00568721117451787, + 0.008111750707030296, + 0.019318392500281334, + 0.007176677696406841, + 0.006626836955547333, + 0.00843728519976139, + 0.004324611742049456, + 0.0035699037835001945, + 0.003670673817396164, + 0.018549464643001556, + 0.005813676863908768, + 0.008459855802357197, + 0.018500514328479767, + 0.003387470031157136, + 0.0020200847648084164, + 0.00474170409142971, + 0.006696230266243219, + 0.008308392949402332, + 0.013358590193092823, + 0.01160221453756094, + 0.01313785556703806, + 0.009145345538854599, + 0.010861938819289207, + 0.019691212102770805, + 0.03465823084115982, + 0.030881691724061966, + 0.0572030246257782, + 0.04010642319917679, + 0.03960825502872467, + 0.027240924537181854, + 0.024035286158323288, + 0.006308774463832378, + 0.009290400892496109, + 0.009634077548980713, + 0.0156713780015707, + 0.01727357693016529, + 0.016888156533241272, + 0.011991694569587708, + 0.013151242397725582, + 0.0009398680413141847, + 0.0009851439390331507, + 0.0025318085681647062, + 0.001635029911994934, + 0.0018873409135267138, + 0.0026816113386303186, + 0.0014251554384827614, + 0.0009363816352561116, + 0.001027419464662671, + 0.0016649917233735323, + 0.002084707608446479, + 0.0031319879926741123, + 0.002246266230940819, + 0.0027225518133491278, + 0.001946280593983829, + 0.0024580087047070265, + 0.0007546297274529934, + 0.0014225445920601487, + 0.0014285158831626177, + 0.0014449426671490073, + 0.0019610931631177664, + 0.001511830952949822, + 0.0009114636923186481, + 0.0008367362897843122, + 0.003046564059332013, + 0.0037952468264847994, + 0.017355557531118393, + 0.003980949986726046, + 0.00538224820047617, + 0.01781335473060608, + 0.0032627200707793236, + 0.002051558345556259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.023301932960748672 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640049301087856 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.004640049301087856, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640049301087856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11349760740995407, + 0.10007303953170776, + 0.07509850710630417, + 0.07939726859331131, + 0.12677112221717834, + 0.0836077630519867, + 0.018964247778058052, + 0.030934929847717285, + 0.09948941320180893, + 0.27216607332229614 + ], + "credit_map": [ + 0.00803249143064022, + 0.011925405822694302, + 0.011797559447586536, + 0.015920234844088554, + 0.02264522947371006, + 0.022338101640343666, + 0.009712453931570053, + 0.011126140132546425, + 0.005567565094679594, + 0.009239411912858486, + 0.012209744192659855, + 0.02721971459686756, + 0.01310505997389555, + 0.011728945188224316, + 0.012473550625145435, + 0.008529048413038254, + 0.00396491726860404, + 0.005561651661992073, + 0.019486846402287483, + 0.0061993226408958435, + 0.011745858937501907, + 0.021298065781593323, + 0.004560679197311401, + 0.0022811745293438435, + 0.005265159998089075, + 0.008518766611814499, + 0.00813822541385889, + 0.01196298561990261, + 0.012568352743983269, + 0.014361806213855743, + 0.008966365829110146, + 0.009615608491003513, + 0.008449187502264977, + 0.012331186793744564, + 0.015049622394144535, + 0.02161809429526329, + 0.018844671547412872, + 0.022646354511380196, + 0.015501514077186584, + 0.012330502271652222, + 0.005490487907081842, + 0.0076676891185343266, + 0.009266898967325687, + 0.012665900401771069, + 0.012842024676501751, + 0.012328262440860271, + 0.011248908936977386, + 0.012097599916160107, + 0.0017536305822432041, + 0.0023271534591913223, + 0.001969162840396166, + 0.003936831373721361, + 0.0025165702681988478, + 0.0023708355147391558, + 0.0024844445288181305, + 0.0016056210733950138, + 0.0018219451885670424, + 0.0026447996497154236, + 0.0043879966251552105, + 0.005095260683447123, + 0.004400815349072218, + 0.005114846397191286, + 0.003120890585705638, + 0.004348380491137505, + 0.006218187976628542, + 0.008226575329899788, + 0.019849034026265144, + 0.012091905809938908, + 0.010180951096117496, + 0.022959116846323013, + 0.009056257084012032, + 0.010907387360930443, + 0.015045489184558392, + 0.01429114118218422, + 0.08474170416593552, + 0.017518233507871628, + 0.027119487524032593, + 0.09090139716863632, + 0.013136880472302437, + 0.00941174104809761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.004640049301087856 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide-angle shot captures a remarkably tranquil body of water, its surface a calm, tea-colored, light brown liquid, suggestive of suspended sediment at the pond's bottom. Gentle, almost imperceptible ripples drift across the surface, capturing the warm, diffuse glow of the afternoon sun and creating a subtle, shimmering texture. In the lower right corner of the frame, two weathered rocks interrupt the water's calm, creating a rugged contrast. The larger rock in the foreground, a solid dark gray, its uneven, damp surface partially covered in vibrant green moss, sits heavily in its semi-submerged position, serving as a solid anchor for the composition. Just behind it, a smaller, similarly textured rock creates a sense of depth. The camera remains perfectly still, allowing the subtle friction of water against stone to be the only subtle movement in a scene dominated by a warm, earthy palette, evoking an atmosphere of deep, contemplative tranquility and undisturbed natural peace.", + "seed": 29600, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_5/index_8576_seed_29600_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1301151880295947, + "vlm_logprob": -0.1301151880295947, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.1301151880295947, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10265742987394333 + ], + "opposite_logprob_tokens": [ + -2.7276573181152344 + ], + "logprob_reward": -0.10265742987394333, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.10265742987394333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39114677906036377, + 0.06338664889335632, + 0.051225218921899796, + 0.06344310194253922, + 0.06259550899267197, + 0.09728667140007019, + 0.05281327664852142, + 0.08373727649450302, + 0.09559297561645508, + 0.03877251595258713 + ], + "credit_map": [ + 0.045228712260723114, + 0.05733216553926468, + 0.051516540348529816, + 0.04423441365361214, + 0.043352603912353516, + 0.046583276242017746, + 0.049434736371040344, + 0.05346433445811272, + 0.007308221887797117, + 0.007837618701159954, + 0.007929484359920025, + 0.008956503123044968, + 0.007996851578354836, + 0.00744427926838398, + 0.007461930625140667, + 0.00845175888389349, + 0.006355840712785721, + 0.008407567627727985, + 0.005625711753964424, + 0.007206154987215996, + 0.005143847782164812, + 0.005786803085356951, + 0.006175699178129435, + 0.006523598451167345, + 0.007603536359965801, + 0.010205870494246483, + 0.007858945056796074, + 0.0085424380376935, + 0.00647076265886426, + 0.006763827055692673, + 0.008015924133360386, + 0.0079817995429039, + 0.006847146432846785, + 0.007695164531469345, + 0.007671106141060591, + 0.009391279891133308, + 0.005765779875218868, + 0.009051582776010036, + 0.009346053935587406, + 0.0068274000659585, + 0.016710009425878525, + 0.01872442290186882, + 0.011270293034613132, + 0.011327792890369892, + 0.009090476669371128, + 0.010533247143030167, + 0.010610043071210384, + 0.009020392782986164, + 0.005894818343222141, + 0.0066587058827281, + 0.006315259728580713, + 0.007038558833301067, + 0.004816667176783085, + 0.006471114698797464, + 0.007319532334804535, + 0.008298621512949467, + 0.015585175715386868, + 0.01421519834548235, + 0.009838245809078217, + 0.007743213325738907, + 0.007987244985997677, + 0.008753855712711811, + 0.00953582115471363, + 0.010078531689941883, + 0.01165076531469822, + 0.025155754759907722, + 0.011266406625509262, + 0.007150167133659124, + 0.014074116945266724, + 0.008501395583152771, + 0.0076157874427735806, + 0.010178590193390846, + 0.005527903325855732, + 0.00714893126860261, + 0.004544883966445923, + 0.004515263717621565, + 0.0044836667366325855, + 0.004930864553898573, + 0.003733317833393812, + 0.0038876840844750404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.10265742987394333 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233617424964905 + ], + "opposite_logprob_tokens": [ + -0.8983617424964905 + ], + "logprob_reward": -0.5233617424964905, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233617424964905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3056137263774872, + 0.058118075132369995, + 0.04629817232489586, + 0.1182766705751419, + 0.09003003686666489, + 0.07906993478536606, + 0.09157050400972366, + 0.049328599125146866, + 0.08782659471035004, + 0.07386772334575653 + ], + "credit_map": [ + 0.040171802043914795, + 0.03572537750005722, + 0.034275468438863754, + 0.04983990639448166, + 0.028674565255641937, + 0.0313192643225193, + 0.04770161211490631, + 0.03790569305419922, + 0.0075882417149841785, + 0.010746769607067108, + 0.006283440627157688, + 0.006431711837649345, + 0.007187108509242535, + 0.005920616909861565, + 0.006830702070146799, + 0.007129478733986616, + 0.007443211041390896, + 0.007468574680387974, + 0.006115502677857876, + 0.005725783761590719, + 0.004268301650881767, + 0.006314233876764774, + 0.004679394885897636, + 0.004283163696527481, + 0.014531374908983707, + 0.021601131185889244, + 0.014017194509506226, + 0.01759890466928482, + 0.011804693378508091, + 0.012749944813549519, + 0.01377780456095934, + 0.012195603922009468, + 0.013119038194417953, + 0.01655600219964981, + 0.009989503771066666, + 0.011269429698586464, + 0.008760028518736362, + 0.011676009744405746, + 0.011150212958455086, + 0.007509796880185604, + 0.009685590863227844, + 0.012915405444800854, + 0.00882349070161581, + 0.010912397876381874, + 0.00810408964753151, + 0.01113851647824049, + 0.007632782217115164, + 0.009857648983597755, + 0.013193010352551937, + 0.012890883721411228, + 0.011079560965299606, + 0.012224863283336163, + 0.008286255411803722, + 0.011355480179190636, + 0.0110761858522892, + 0.011464255861938, + 0.0049472651444375515, + 0.0072961170226335526, + 0.005924704484641552, + 0.005436364561319351, + 0.004897185601294041, + 0.006434353068470955, + 0.005529309157282114, + 0.008863294497132301, + 0.00887245498597622, + 0.021928302943706512, + 0.011026452295482159, + 0.007267566863447428, + 0.012048804201185703, + 0.008665218017995358, + 0.008175750263035297, + 0.009842028841376305, + 0.007943020202219486, + 0.016857966780662537, + 0.008673574775457382, + 0.009046859107911587, + 0.007880348712205887, + 0.008184214122593403, + 0.008568172343075275, + 0.006713559851050377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.5233617424964905 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017244244227185845 + ], + "opposite_logprob_tokens": [ + -6.3767242431640625 + ], + "logprob_reward": -0.0017244244227185845, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017244244227185845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19886751472949982, + 0.12230025231838226, + 0.09942606091499329, + 0.07856260240077972, + 0.10617688298225403, + 0.05111183971166611, + 0.0741267278790474, + 0.07089456915855408, + 0.11996538937091827, + 0.07856814563274384 + ], + "credit_map": [ + 0.02700486220419407, + 0.03782743588089943, + 0.021140070632100105, + 0.02235483005642891, + 0.026550544425845146, + 0.022202011197805405, + 0.024203835055232048, + 0.01758393831551075, + 0.02451389469206333, + 0.017002427950501442, + 0.014387388713657856, + 0.014594792388379574, + 0.012286370620131493, + 0.009943337179720402, + 0.01603875681757927, + 0.013533291406929493, + 0.01672814041376114, + 0.012837073765695095, + 0.012285112403333187, + 0.013914327137172222, + 0.007844142615795135, + 0.009970244020223618, + 0.014094661921262741, + 0.011752369813621044, + 0.011190391145646572, + 0.013957781717181206, + 0.009724033065140247, + 0.00977361761033535, + 0.007118547800928354, + 0.007574927061796188, + 0.009724740870296955, + 0.009498567320406437, + 0.014684070833027363, + 0.0158856138586998, + 0.012887354008853436, + 0.015908557921648026, + 0.009282813407480717, + 0.010961625725030899, + 0.014856498688459396, + 0.011710354126989841, + 0.008764887228608131, + 0.009983543306589127, + 0.005906982347369194, + 0.005867796018719673, + 0.005065972451120615, + 0.004986110609024763, + 0.006078139878809452, + 0.004458412993699312, + 0.00818677432835102, + 0.009532934054732323, + 0.00999384094029665, + 0.009881912730634212, + 0.007530676666647196, + 0.00888130720704794, + 0.01044399943202734, + 0.00967528484761715, + 0.01047136913985014, + 0.008669192902743816, + 0.00810978002846241, + 0.01003036554902792, + 0.007069279905408621, + 0.008592265658080578, + 0.00825748685747385, + 0.00969482772052288, + 0.016372065991163254, + 0.031938232481479645, + 0.014129365794360638, + 0.008039717562496662, + 0.01666693016886711, + 0.011169858276844025, + 0.010046985931694508, + 0.011602243408560753, + 0.008212040178477764, + 0.01851075515151024, + 0.009059623815119267, + 0.012893065810203552, + 0.008521715179085732, + 0.007916810922324657, + 0.007193015422672033, + 0.006261130329221487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0017244244227185845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195616081357002 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195616081357002, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195616081357002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3134993314743042, + 0.113203264772892, + 0.09138774871826172, + 0.04829924926161766, + 0.1813802570104599, + 0.07297974824905396, + 0.040071871131658554, + 0.06433270126581192, + 0.02454974688589573, + 0.05029609799385071 + ], + "credit_map": [ + 0.043181903660297394, + 0.03658956661820412, + 0.04128199443221092, + 0.033312276005744934, + 0.03082728572189808, + 0.03098836913704872, + 0.048028744757175446, + 0.049289196729660034, + 0.01519338320940733, + 0.01119681354612112, + 0.014001691713929176, + 0.012250523082911968, + 0.011477462016046047, + 0.013885541819036007, + 0.018780698999762535, + 0.01641714945435524, + 0.013068040832877159, + 0.011082097887992859, + 0.011955412104725838, + 0.009267374873161316, + 0.008951099589467049, + 0.011739957146346569, + 0.012684301473200321, + 0.012639464810490608, + 0.007265088148415089, + 0.007180154789239168, + 0.005486238282173872, + 0.005415063817054033, + 0.005018648691475391, + 0.0049192458391189575, + 0.006100660189986229, + 0.006914146710187197, + 0.024286722764372826, + 0.0256753358989954, + 0.022628873586654663, + 0.024860698729753494, + 0.016385992988944054, + 0.01660662516951561, + 0.03045095130801201, + 0.020485054701566696, + 0.007574176881462336, + 0.008234236389398575, + 0.010169651359319687, + 0.009956136345863342, + 0.006853861268609762, + 0.007529654540121555, + 0.011838707141578197, + 0.0108233243227005, + 0.004272808786481619, + 0.004381165374070406, + 0.004802398849278688, + 0.006243785843253136, + 0.00387822394259274, + 0.004476706963032484, + 0.006614389829337597, + 0.005402388516813517, + 0.006405287887901068, + 0.006529652047902346, + 0.008181310258805752, + 0.00775153748691082, + 0.006290195044130087, + 0.007681928109377623, + 0.011713500134646893, + 0.009779292158782482, + 0.0033892132341861725, + 0.004758425988256931, + 0.0028164528775960207, + 0.0019337278790771961, + 0.0028869174420833588, + 0.002309161704033613, + 0.0035441219806671143, + 0.002911725314334035, + 0.004879375919699669, + 0.007481166161596775, + 0.005748150870203972, + 0.006514722015708685, + 0.005237571429461241, + 0.006082662846893072, + 0.0076839434914290905, + 0.00666850758716464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.018195616081357002 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636727273464203 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636727273464203, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636727273464203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3295793831348419, + 0.09837016463279724, + 0.07784020900726318, + 0.09303603321313858, + 0.08246511965990067, + 0.055196523666381836, + 0.03525441512465477, + 0.059633396565914154, + 0.06547559797763824, + 0.1031491756439209 + ], + "credit_map": [ + 0.05739937722682953, + 0.039440907537937164, + 0.036003123968839645, + 0.038109391927719116, + 0.03264033794403076, + 0.034127384424209595, + 0.031053386628627777, + 0.06080547720193863, + 0.020027965307235718, + 0.015911970287561417, + 0.010308357886970043, + 0.010428457520902157, + 0.010432898998260498, + 0.010367375798523426, + 0.009091421030461788, + 0.011801725253462791, + 0.015230010263621807, + 0.009214555844664574, + 0.009133591316640377, + 0.008370799012482166, + 0.006328977644443512, + 0.008649093098938465, + 0.007678133435547352, + 0.013235049322247505, + 0.016389207914471626, + 0.01400848850607872, + 0.01011392567306757, + 0.010902319103479385, + 0.007840322330594063, + 0.011654095724225044, + 0.009855353273451328, + 0.012272323481738567, + 0.012204140424728394, + 0.010073410347104073, + 0.00819996278733015, + 0.009367679245769978, + 0.008001203648746014, + 0.00956265814602375, + 0.01101029384881258, + 0.014045764692127705, + 0.005645219702273607, + 0.006437338888645172, + 0.0062466831877827644, + 0.008530990220606327, + 0.005597986746579409, + 0.004977606236934662, + 0.007767826784402132, + 0.00999287236481905, + 0.005085780750960112, + 0.005098439287394285, + 0.003976854961365461, + 0.004216169938445091, + 0.0032276250422000885, + 0.005317978095263243, + 0.004210642073303461, + 0.00412092637270689, + 0.007682088762521744, + 0.0068715400993824005, + 0.008067981339991093, + 0.0063304896466434, + 0.004860639106482267, + 0.007067280355840921, + 0.008653511293232441, + 0.010099860839545727, + 0.008613165467977524, + 0.017933910712599754, + 0.008049391210079193, + 0.004838262218981981, + 0.008482145145535469, + 0.006952227093279362, + 0.005890193860977888, + 0.0047163027338683605, + 0.01722518727183342, + 0.018253492191433907, + 0.011680922470986843, + 0.009610794484615326, + 0.011110466904938221, + 0.012329639866948128, + 0.011010272428393364, + 0.011928405612707138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.004636727273464203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide-angle shot captures a remarkably tranquil body of water, its surface a calm, tea-colored, light brown liquid, suggestive of suspended sediment at the pond's bottom. Gentle, almost imperceptible ripples drift across the surface, capturing the warm, diffuse glow of the afternoon sun and creating a subtle, shimmering texture. In the lower right corner of the frame, two weathered rocks interrupt the water's calm, creating a rugged contrast. The larger rock in the foreground, a solid dark gray, its uneven, damp surface partially covered in vibrant green moss, sits heavily in its semi-submerged position, serving as a solid anchor for the composition. Just behind it, a smaller, similarly textured rock creates a sense of depth. The camera remains perfectly still, allowing the subtle friction of water against stone to be the only subtle movement in a scene dominated by a warm, earthy palette, evoking an atmosphere of deep, contemplative tranquility and undisturbed natural peace.", + "seed": 29601, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_5/index_8576_seed_29601_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.26143514290452, + "vlm_logprob": -0.26143514290452, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.26143514290452, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3226329982280731 + ], + "opposite_logprob_tokens": [ + -1.4476330280303955 + ], + "logprob_reward": -0.3226329982280731, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3226329982280731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42445579171180725, + 0.18255557119846344, + 0.05316316336393356, + 0.025035005062818527, + 0.02689508907496929, + 0.07821578532457352, + 0.10299084335565567, + 0.016329141333699226, + 0.012701600790023804, + 0.07765798270702362 + ], + "credit_map": [ + 0.014454390853643417, + 0.028783323243260384, + 0.02068066969513893, + 0.017606323584914207, + 0.14938849210739136, + 0.15773028135299683, + 0.017613019794225693, + 0.01819927804172039, + 0.006973089184612036, + 0.06162768974900246, + 0.00852396059781313, + 0.008270256221294403, + 0.0681128278374672, + 0.014063438400626183, + 0.007572369184345007, + 0.00741193164139986, + 0.005877471994608641, + 0.011482656933367252, + 0.009860683232545853, + 0.004261698108166456, + 0.005515962839126587, + 0.00832564476877451, + 0.003988322336226702, + 0.0038507224526256323, + 0.0022172308526933193, + 0.004358514677733183, + 0.004289194475859404, + 0.0026939602103084326, + 0.004627482034265995, + 0.0025319948326796293, + 0.0022902712225914, + 0.0020263579208403826, + 0.00182925199624151, + 0.005724722519516945, + 0.0031044299248605967, + 0.0033285461831837893, + 0.005855970550328493, + 0.0030386075377464294, + 0.001911233295686543, + 0.0021023284643888474, + 0.003161319065839052, + 0.024933531880378723, + 0.006662785075604916, + 0.003439817577600479, + 0.028939463198184967, + 0.006058239843696356, + 0.002097110729664564, + 0.002923517255112529, + 0.0067059691064059734, + 0.03196335583925247, + 0.009788463823497295, + 0.006689753383398056, + 0.013281762599945068, + 0.024288063868880272, + 0.0037623385433107615, + 0.006511135026812553, + 0.0011451822938397527, + 0.003963683731853962, + 0.0025076253805309534, + 0.0011714877327904105, + 0.003643725998699665, + 0.0020252452231943607, + 0.000857895181979984, + 0.0010142970131710172, + 0.0011216939892619848, + 0.002234780928120017, + 0.0016448750393465161, + 0.001073078834451735, + 0.0016141097294166684, + 0.002344683278352022, + 0.0013075125170871615, + 0.0013608665904030204, + 0.003142599482089281, + 0.0226063821464777, + 0.006476701237261295, + 0.0032852240838110447, + 0.02795531414449215, + 0.008282516151666641, + 0.0024213846772909164, + 0.0034878647420555353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.3226329982280731 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983911871910095 + ], + "opposite_logprob_tokens": [ + -0.5233911871910095 + ], + "logprob_reward": -0.8983911871910095, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983911871910095, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3049171566963196, + 0.1158454641699791, + 0.041758161038160324, + 0.031762607395648956, + 0.00819657277315855, + 0.10290305316448212, + 0.05376899614930153, + 0.015373555943369865, + 0.02866067923605442, + 0.29681381583213806 + ], + "credit_map": [ + 0.00830009300261736, + 0.013253321871161461, + 0.009185106493532658, + 0.005796545185148716, + 0.12818674743175507, + 0.13096317648887634, + 0.0050123282708227634, + 0.004219822585582733, + 0.007168213836848736, + 0.04257216304540634, + 0.010693544521927834, + 0.004832741338759661, + 0.03190312534570694, + 0.011584775522351265, + 0.004176488611847162, + 0.0029144107829779387, + 0.00504134688526392, + 0.006284430623054504, + 0.008341752924025059, + 0.004126104526221752, + 0.005213264375925064, + 0.005346615333110094, + 0.003528314409777522, + 0.0038763307966291904, + 0.0028125112876296043, + 0.00658704387024045, + 0.005246875341981649, + 0.0026085611898452044, + 0.0062253437936306, + 0.0030318780336529016, + 0.0025240378454327583, + 0.0027263567317277193, + 0.0008252789848484099, + 0.0011304040672257543, + 0.0013151818420737982, + 0.0011294418945908546, + 0.0009666409459896386, + 0.0010292899096384645, + 0.0008790628635324538, + 0.0009212716249749064, + 0.003924370277673006, + 0.037225957959890366, + 0.00811994168907404, + 0.005061005242168903, + 0.02756812982261181, + 0.012392047792673111, + 0.0034893767442554235, + 0.005122223868966103, + 0.004152732435613871, + 0.010925132781267166, + 0.008663216605782509, + 0.005680202040821314, + 0.005414465442299843, + 0.010541773401200771, + 0.0037724694702774286, + 0.004619003739207983, + 0.0011785547249019146, + 0.002179775619879365, + 0.0027004689909517765, + 0.0015074402326717973, + 0.003111705183982849, + 0.0018203104846179485, + 0.0014499673852697015, + 0.0014253329718485475, + 0.0023150306660681963, + 0.004758401773869991, + 0.0063020698726177216, + 0.0023511494509875774, + 0.0047833602875471115, + 0.004096719436347485, + 0.001766202854923904, + 0.0022877443116158247, + 0.009708398953080177, + 0.11144352704286575, + 0.020530810579657555, + 0.00901844259351492, + 0.11525919288396835, + 0.01729416660964489, + 0.005940204951912165, + 0.007619073614478111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.8983911871910095 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870428517460823 + ], + "opposite_logprob_tokens": [ + -3.0487043857574463 + ], + "logprob_reward": -0.04870428517460823, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870428517460823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10229876637458801, + 0.20984719693660736, + 0.014095705933868885, + 0.017948487773537636, + 0.014287585392594337, + 0.10395977646112442, + 0.02059650421142578, + 0.043901700526475906, + 0.013085479848086834, + 0.459978848695755 + ], + "credit_map": [ + 0.006923386361449957, + 0.018085051327943802, + 0.01486810203641653, + 0.00735766114667058, + 0.025044972077012062, + 0.014180712401866913, + 0.00634805066511035, + 0.00949081964790821, + 0.00518806092441082, + 0.08114349097013474, + 0.004721245728433132, + 0.0053135124035179615, + 0.0901246964931488, + 0.014096248894929886, + 0.004079778678715229, + 0.005180145613849163, + 0.0015533259138464928, + 0.002293100580573082, + 0.002540961606428027, + 0.0016167493304237723, + 0.0018105898052453995, + 0.001962872687727213, + 0.0011476613581180573, + 0.0011704432545229793, + 0.0008455214556306601, + 0.005975386127829552, + 0.0013677573297172785, + 0.001063890173099935, + 0.004946848377585411, + 0.0019316505640745163, + 0.0009040742297656834, + 0.0009133587009273469, + 0.0013080924982205033, + 0.002353294752538204, + 0.0022417542058974504, + 0.0014642855385318398, + 0.002551761921495199, + 0.0017269124509766698, + 0.0014463469851762056, + 0.0011951365740969777, + 0.0036652320995926857, + 0.01301466766744852, + 0.008190912194550037, + 0.005921980366110802, + 0.03755037114024162, + 0.026012832298874855, + 0.005751927848905325, + 0.0038518484216183424, + 0.0012682529632002115, + 0.004667803645133972, + 0.0025261463597416878, + 0.002330317161977291, + 0.002386834705248475, + 0.004045296926051378, + 0.0018711462616920471, + 0.001500704325735569, + 0.0026463589165359735, + 0.012321284972131252, + 0.005088174249976873, + 0.003537188982591033, + 0.007871024310588837, + 0.0078949686139822, + 0.0020764649379998446, + 0.002466234378516674, + 0.0011495467042550445, + 0.0019046012312173843, + 0.003210377646610141, + 0.00145347963552922, + 0.001537240226753056, + 0.0015488805947825313, + 0.000976086943410337, + 0.0013052655849605799, + 0.016028469428420067, + 0.18464887142181396, + 0.03244609013199806, + 0.011059274896979332, + 0.17499376833438873, + 0.02423429675400257, + 0.00754539342597127, + 0.009022656828165054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.04870428517460823 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029815100133419037 + ], + "opposite_logprob_tokens": [ + -3.529815196990967 + ], + "logprob_reward": -0.029815100133419037, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815100133419037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09767463803291321, + 0.2595168948173523, + 0.02821352705359459, + 0.016098057851195335, + 0.018889153376221657, + 0.13298460841178894, + 0.07153821736574173, + 0.012472322210669518, + 0.004493372049182653, + 0.3581192195415497 + ], + "credit_map": [ + 0.005131785292178392, + 0.009974570944905281, + 0.012863961048424244, + 0.0069147865287959576, + 0.024121345952153206, + 0.021952500566840172, + 0.00881492905318737, + 0.007900760509073734, + 0.006476570852100849, + 0.10330643504858017, + 0.0050560543313622475, + 0.0032515558414161205, + 0.11748064309358597, + 0.018140077590942383, + 0.003136242972686887, + 0.0026693083345890045, + 0.0017608770867809653, + 0.008552493527531624, + 0.002461415948346257, + 0.0021913473028689623, + 0.006093076430261135, + 0.0038264591712504625, + 0.0015771035104990005, + 0.0017507515149191022, + 0.001532151596620679, + 0.003353406209498644, + 0.0020748130045831203, + 0.0018603649223223329, + 0.00225904886610806, + 0.0023588070180267096, + 0.0016087584663182497, + 0.0010507070692256093, + 0.0012229949934408069, + 0.004847947042435408, + 0.00221083452925086, + 0.001357046770863235, + 0.0039788102731108665, + 0.0031348527409136295, + 0.0007096643676050007, + 0.0014270051615312696, + 0.003385100280866027, + 0.05070221424102783, + 0.005200437735766172, + 0.002685191109776497, + 0.05781745910644531, + 0.009242597036063671, + 0.002225054893642664, + 0.0017265513306483626, + 0.004792878869920969, + 0.01981036551296711, + 0.008397958241403103, + 0.005450564902275801, + 0.008784583769738674, + 0.01603476144373417, + 0.002650410169735551, + 0.00561669422313571, + 0.0006253168685361743, + 0.004055758938193321, + 0.0007861974299885333, + 0.0008298088796436787, + 0.003023782977834344, + 0.002144760452210903, + 0.00044578261440619826, + 0.0005609142826870084, + 0.00037829834036529064, + 0.000773277715779841, + 0.000797289190813899, + 0.0004027751856483519, + 0.0006782796117477119, + 0.0007379950257018209, + 0.00033620992326177657, + 0.00038924714317545295, + 0.010589070618152618, + 0.14723406732082367, + 0.021796418353915215, + 0.00799630582332611, + 0.1420779973268509, + 0.017692390829324722, + 0.004943512380123138, + 0.005789457820355892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.029815100133419037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water surface a calm, tea-colored, light brown liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are two weathered rocks visible in the lower right corner of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the larger foreground rock dark gray with vibrant green moss on its uneven, damp surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632143795490265 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632143795490265, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632143795490265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4821295738220215, + 0.07996976375579834, + 0.04885067790746689, + 0.03031778149306774, + 0.0547872930765152, + 0.0962076410651207, + 0.058177463710308075, + 0.040436457842588425, + 0.05567468702793121, + 0.0534486398100853 + ], + "credit_map": [ + 0.02336851879954338, + 0.04343205690383911, + 0.04007120057940483, + 0.04920482262969017, + 0.1479032188653946, + 0.12416144460439682, + 0.029375875368714333, + 0.02461245283484459, + 0.007687342818826437, + 0.017147744074463844, + 0.010491281747817993, + 0.008962227962911129, + 0.012908659875392914, + 0.011496935971081257, + 0.005004734266549349, + 0.006270838435739279, + 0.004756906535476446, + 0.006223652046173811, + 0.008606163784861565, + 0.008937809616327286, + 0.006379438564181328, + 0.005424215458333492, + 0.0046607814729213715, + 0.0038617118261754513, + 0.002019015606492758, + 0.00561669934540987, + 0.0037351008504629135, + 0.0036521954461932182, + 0.005525118671357632, + 0.003037206595763564, + 0.004011886194348335, + 0.0027205580845475197, + 0.004732358269393444, + 0.007959899492561817, + 0.007141661364585161, + 0.00906527042388916, + 0.007138872519135475, + 0.006719445809721947, + 0.006796336732804775, + 0.005233450327068567, + 0.0028303347062319517, + 0.03690045326948166, + 0.003431130899116397, + 0.0028461141046136618, + 0.03961790353059769, + 0.00548218609765172, + 0.0031485024373978376, + 0.001951020210981369, + 0.0040979222394526005, + 0.010750272311270237, + 0.009288016706705093, + 0.008998867124319077, + 0.005493070464581251, + 0.01069812010973692, + 0.0036722749937325716, + 0.0051789190620183945, + 0.0025234485510736704, + 0.00961406622081995, + 0.006053722929209471, + 0.004881878849118948, + 0.0071316491812467575, + 0.004861102439463139, + 0.0025526403915137053, + 0.002817950677126646, + 0.004496677312999964, + 0.010221322998404503, + 0.008459809236228466, + 0.005504443775862455, + 0.009366847574710846, + 0.011025759391486645, + 0.0029854930471628904, + 0.0036143327597528696, + 0.004874060396105051, + 0.009332187473773956, + 0.009852192364633083, + 0.005714677274227142, + 0.010185843333601952, + 0.006779979448765516, + 0.0029359583277255297, + 0.0037737409584224224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.007632143795490265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_6/index_tensor([2275, 2275])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_6/index_tensor([2275, 2275])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dc1d412b8b3e9a8ca25a870f04f0ef1f6bf40378 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_6/index_tensor([2275, 2275])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with an extreme close-up, the camera frozen, focused on a single, intact dandelion. In the soft, golden afternoon sunlight, its perfect sphere of countless delicate white hairs appears remarkably sacred, the light filtering through, giving each hair a soft halo around its edges. The background blurs into a soft patch of green and gold. A gentle breeze sweeps across the dandelion, causing the entire ball to tremble slightly. Dozens of mature seeds begin to loosen, then, carefully lifted by the wind, detach themselves from the dark brown receptacle and begin their journey into the distance. At this moment, the camera, in a seamless transition, precisely shifts focus from the mother plant to one of the newly departed seeds, and begins a smooth tracking motion to the right. This seed, with its dark brown spindle-shaped seeds and radiating light umbrella-shaped pappus, gracefully rotates and undulates in the air. Its white pappus appears translucent in the light. It is carried steadily and calmly by the air currents, floating towards the depths of the blurred, sun-drenched meadow in the distance. The whole picture is filled with a poetic atmosphere of tranquility, hope and continuation of life, creating an ultimate sense of fantasy.", + "seed": 32829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_6/index_2275_seed_32829_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027072652406059206, + "vlm_logprob": -0.027072652406059206, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.027072652406059206, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011313453316688538 + ], + "opposite_logprob_tokens": [ + -4.636313438415527 + ], + "logprob_reward": -0.011313453316688538, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011313453316688538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09321274608373642, + 0.3500271439552307, + 0.20155543088912964, + 0.07689584791660309, + 0.08765947818756104, + 0.041422512382268906, + 0.04103473946452141, + 0.016890432685613632, + 0.029271086677908897, + 0.06203048676252365 + ], + "credit_map": [ + 0.009581085294485092, + 0.019358651712536812, + 0.008895468898117542, + 0.012488513253629208, + 0.013728895224630833, + 0.011171900667250156, + 0.008253362029790878, + 0.00973487738519907, + 0.012650581076741219, + 0.010588421486318111, + 0.007462549023330212, + 0.008847116492688656, + 0.014761538244783878, + 0.1486072987318039, + 0.00910437386482954, + 0.138005331158638, + 0.010913465172052383, + 0.015985732898116112, + 0.012828316539525986, + 0.012109627015888691, + 0.013077293522655964, + 0.052793826907873154, + 0.01940273679792881, + 0.06444443017244339, + 0.01113703940063715, + 0.00848880223929882, + 0.010134424082934856, + 0.009161663241684437, + 0.00654870318248868, + 0.011152485385537148, + 0.008818622678518295, + 0.011454113759100437, + 0.013929075561463833, + 0.0072906967252492905, + 0.011025204323232174, + 0.007182480301707983, + 0.00667128199711442, + 0.02076834626495838, + 0.008387823589146137, + 0.012404578737914562, + 0.004229956306517124, + 0.002541616326197982, + 0.007848654873669147, + 0.004586081951856613, + 0.0034273117780685425, + 0.005982944276183844, + 0.004354155156761408, + 0.008451797068119049, + 0.002743855118751526, + 0.0021413166541606188, + 0.0034247192088514566, + 0.003427265677601099, + 0.004419299773871899, + 0.01469956710934639, + 0.003702989313751459, + 0.006475729867815971, + 0.00126210181042552, + 0.0020012061577290297, + 0.001940957736223936, + 0.00198812922462821, + 0.002707104431465268, + 0.001798943500034511, + 0.0027727894484996796, + 0.0024192016571760178, + 0.0024402220733463764, + 0.0037336056120693684, + 0.00298710516653955, + 0.002864109817892313, + 0.006635101046413183, + 0.0037561720237135887, + 0.0029956151265650988, + 0.0038591583725064993, + 0.004265402909368277, + 0.009922612458467484, + 0.005920193623751402, + 0.0066898236982524395, + 0.011004732921719551, + 0.01458999328315258, + 0.004510386846959591, + 0.005127349868416786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.011313453316688538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812966123223305 + ], + "opposite_logprob_tokens": [ + -3.2881295680999756 + ], + "logprob_reward": -0.03812966123223305, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812966123223305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05602102354168892, + 0.46105217933654785, + 0.09612719714641571, + 0.05655330419540405, + 0.12149995565414429, + 0.055140163749456406, + 0.025087326765060425, + 0.03476577624678612, + 0.03610176965594292, + 0.05765125900506973 + ], + "credit_map": [ + 0.006084608845412731, + 0.013393267057836056, + 0.0041201491840183735, + 0.006185482256114483, + 0.008142028003931046, + 0.007298233453184366, + 0.004102938808500767, + 0.00669432245194912, + 0.011355714872479439, + 0.012989075854420662, + 0.005057159811258316, + 0.006485870573669672, + 0.02015908807516098, + 0.20537614822387695, + 0.008922700770199299, + 0.19070649147033691, + 0.008874018676578999, + 0.02355378121137619, + 0.01058602798730135, + 0.015913454815745354, + 0.007539794780313969, + 0.009544024243950844, + 0.010461484082043171, + 0.009654616005718708, + 0.0042797415517270565, + 0.00589788518846035, + 0.004127717111259699, + 0.005008483771234751, + 0.007178688421845436, + 0.011111919768154621, + 0.007194129750132561, + 0.011754742823541164, + 0.021001247689127922, + 0.012625551782548428, + 0.015260957181453705, + 0.0145801967009902, + 0.008093774318695068, + 0.02357647195458412, + 0.011543212458491325, + 0.014818553812801838, + 0.005802251864224672, + 0.0026242807507514954, + 0.012716894038021564, + 0.004911242052912712, + 0.004140640143305063, + 0.007923666387796402, + 0.009404990822076797, + 0.007616201415657997, + 0.001662430469878018, + 0.0018772862385958433, + 0.002352295909076929, + 0.0033151551615446806, + 0.004159562289714813, + 0.004159686621278524, + 0.0050696455873548985, + 0.0024912660010159016, + 0.002471894957125187, + 0.002583068562671542, + 0.003195906290784478, + 0.0033114065881818533, + 0.006103897001594305, + 0.006974731106311083, + 0.006849031895399094, + 0.0032758433371782303, + 0.002311056014150381, + 0.005004566162824631, + 0.002778600202873349, + 0.003254551440477371, + 0.008932865224778652, + 0.006611264310777187, + 0.004730578511953354, + 0.0024782882537692785, + 0.0038437144830822945, + 0.009177510626614094, + 0.004686508793383837, + 0.006603668909519911, + 0.009593694470822811, + 0.014745602384209633, + 0.004539497196674347, + 0.004461066797375679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.03812966123223305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196122996509075 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014196122996509075, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196122996509075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04906412586569786, + 0.2105930745601654, + 0.25868821144104004, + 0.08965884149074554, + 0.15270213782787323, + 0.09263002127408981, + 0.03936724737286568, + 0.04123478755354881, + 0.05294104292988777, + 0.013120538555085659 + ], + "credit_map": [ + 0.004387409891933203, + 0.008011466823518276, + 0.004873300436884165, + 0.0091885169968009, + 0.0059455991722643375, + 0.005083219613879919, + 0.006444490049034357, + 0.005130116827785969, + 0.005986519157886505, + 0.006044040434062481, + 0.0036063329316675663, + 0.005235429387539625, + 0.010636085644364357, + 0.09075570851564407, + 0.005447182338684797, + 0.08288175612688065, + 0.008844404481351376, + 0.016783151775598526, + 0.00881294347345829, + 0.010419352911412716, + 0.020777925848960876, + 0.07743533700704575, + 0.02232293412089348, + 0.09329215437173843, + 0.006702682469040155, + 0.0049918033182621, + 0.008413394913077354, + 0.008345991373062134, + 0.0078052943572402, + 0.022024406120181084, + 0.008408503606915474, + 0.02296675555408001, + 0.02036028541624546, + 0.013097953982651234, + 0.021982794627547264, + 0.015191659331321716, + 0.011268213391304016, + 0.02948628179728985, + 0.021464386954903603, + 0.019850559532642365, + 0.010356420651078224, + 0.007525716442614794, + 0.010251360014081001, + 0.010683853179216385, + 0.008219095878303051, + 0.013136460445821285, + 0.01341856550425291, + 0.019038543105125427, + 0.0030169752426445484, + 0.0035908306017518044, + 0.0036233433056622744, + 0.008202839642763138, + 0.0065255737863481045, + 0.00581212667748332, + 0.0060554975643754005, + 0.0025400593876838684, + 0.0034731237683445215, + 0.003435281803831458, + 0.004438774194568396, + 0.003574573900550604, + 0.006838785018771887, + 0.007674692198634148, + 0.005699703935533762, + 0.0060998485423624516, + 0.00408161198720336, + 0.006042193155735731, + 0.0049971421249210835, + 0.005013840738683939, + 0.010765372775495052, + 0.011795788072049618, + 0.0047086807899177074, + 0.005536411423236132, + 0.0007826238870620728, + 0.0018781579565256834, + 0.0013840551255270839, + 0.0017723874188959599, + 0.002287859097123146, + 0.0028052704874426126, + 0.0010636398801580071, + 0.0011465431889519095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.014196122996509075 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000947743654251 + ], + "opposite_logprob_tokens": [ + -2.695009469985962 + ], + "logprob_reward": -0.07000947743654251, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000947743654251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033614546060562134, + 0.172673299908638, + 0.1903679519891739, + 0.14419929683208466, + 0.060530222952365875, + 0.06527566909790039, + 0.1335054486989975, + 0.057035893201828, + 0.10851705074310303, + 0.03428060561418533 + ], + "credit_map": [ + 0.003608075203374028, + 0.0028274981305003166, + 0.003509075613692403, + 0.004338615573942661, + 0.003915064036846161, + 0.004777287598699331, + 0.0048296405002474785, + 0.005809288006275892, + 0.004607336595654488, + 0.005864105187356472, + 0.002719848183915019, + 0.003494193544611335, + 0.008130819536745548, + 0.07451534271240234, + 0.0038619867991656065, + 0.06947965919971466, + 0.0066618663258850574, + 0.008065923117101192, + 0.006837593857198954, + 0.008405081927776337, + 0.014760049991309643, + 0.05869651585817337, + 0.01816956140100956, + 0.0687713548541069, + 0.01488038431853056, + 0.013167796656489372, + 0.016103418543934822, + 0.01827227883040905, + 0.016692573204636574, + 0.02269335836172104, + 0.017909113317728043, + 0.024480367079377174, + 0.007982325740158558, + 0.005198258440941572, + 0.006850441452115774, + 0.011347426101565361, + 0.005104063078761101, + 0.007873139344155788, + 0.008268539793789387, + 0.007906024344265461, + 0.007279199548065662, + 0.0038116152863949537, + 0.015704331919550896, + 0.006574718747287989, + 0.00524882273748517, + 0.009452072903513908, + 0.007758839521557093, + 0.009446063078939915, + 0.009890285320580006, + 0.0083279088139534, + 0.011443449184298515, + 0.018652841448783875, + 0.014253775589168072, + 0.0424315519630909, + 0.010232478380203247, + 0.01827314868569374, + 0.004601562395691872, + 0.008519766852259636, + 0.005932812113314867, + 0.0062735723331570625, + 0.010304653085768223, + 0.007672193460166454, + 0.01068936474621296, + 0.0030419707763940096, + 0.010538850910961628, + 0.01235712319612503, + 0.009074347093701363, + 0.009750837460160255, + 0.018428068608045578, + 0.02602432295680046, + 0.011280250735580921, + 0.01106324978172779, + 0.0023218721617013216, + 0.004414173774421215, + 0.0032625612802803516, + 0.004512470215559006, + 0.004034380428493023, + 0.010501406155526638, + 0.0024122721515595913, + 0.0028214699123054743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.07000947743654251 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017145470483228564 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017145470483228564, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017145470483228564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02726803347468376, + 0.23368887603282928, + 0.24930712580680847, + 0.046749867498874664, + 0.05028200149536133, + 0.16961468756198883, + 0.036224063485860825, + 0.06503096222877502, + 0.03717338666319847, + 0.08466100692749023 + ], + "credit_map": [ + 0.0022600037045776844, + 0.0033095020335167646, + 0.0022519687190651894, + 0.006221598945558071, + 0.0030857811216264963, + 0.0033027559984475374, + 0.003361587645485997, + 0.0034748350735753775, + 0.005682395305484533, + 0.006207618862390518, + 0.002994115697219968, + 0.0036725373938679695, + 0.010063321329653263, + 0.10436531156301498, + 0.004675470292568207, + 0.09602810442447662, + 0.008238309994339943, + 0.009181508794426918, + 0.005671326071023941, + 0.007212646305561066, + 0.01801920495927334, + 0.08231000602245331, + 0.0200577974319458, + 0.09861632436513901, + 0.0033157130237668753, + 0.004068868700414896, + 0.0027790889143943787, + 0.0025809372309595346, + 0.0062071229331195354, + 0.011759081855416298, + 0.005268807522952557, + 0.010770251974463463, + 0.008019416593015194, + 0.004485886078327894, + 0.004654952324926853, + 0.005825465079396963, + 0.004300386179238558, + 0.010432375594973564, + 0.006154537200927734, + 0.006408985238522291, + 0.02266491949558258, + 0.009114665910601616, + 0.04577942192554474, + 0.013189607299864292, + 0.009214771911501884, + 0.027293110266327858, + 0.018070662394165993, + 0.024287518113851547, + 0.002789747202768922, + 0.0036783828400075436, + 0.004085727501660585, + 0.008777225390076637, + 0.0058941892348229885, + 0.004663830157369375, + 0.004199513234198093, + 0.002135448157787323, + 0.005953813437372446, + 0.012660699896514416, + 0.008778625167906284, + 0.008759879507124424, + 0.009233134798705578, + 0.007630184292793274, + 0.006877087987959385, + 0.0051375338807702065, + 0.0034109046682715416, + 0.004106818698346615, + 0.005160116590559483, + 0.0038426462560892105, + 0.005980851128697395, + 0.008606446906924248, + 0.003769264090806246, + 0.0022963397204875946, + 0.007567403372377157, + 0.008537300862371922, + 0.010687831789255142, + 0.011809862218797207, + 0.004728296305984259, + 0.027688806876540184, + 0.005412094760686159, + 0.00822941493242979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.0017145470483228564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an extreme close-up, the camera frozen, focused on a single, intact dandelion. In the soft, golden afternoon sunlight, its perfect sphere of countless delicate white hairs appears remarkably sacred, the light filtering through, giving each hair a soft halo around its edges. The background blurs into a soft patch of green and gold. A gentle breeze sweeps across the dandelion, causing the entire ball to tremble slightly. Dozens of mature seeds begin to loosen, then, carefully lifted by the wind, detach themselves from the dark brown receptacle and begin their journey into the distance. At this moment, the camera, in a seamless transition, precisely shifts focus from the mother plant to one of the newly departed seeds, and begins a smooth tracking motion to the right. This seed, with its dark brown spindle-shaped seeds and radiating light umbrella-shaped pappus, gracefully rotates and undulates in the air. Its white pappus appears translucent in the light. It is carried steadily and calmly by the air currents, floating towards the depths of the blurred, sun-drenched meadow in the distance. The whole picture is filled with a poetic atmosphere of tranquility, hope and continuation of life, creating an ultimate sense of fantasy.", + "seed": 32830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_6/index_2275_seed_32830_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04660987977404148, + "vlm_logprob": -0.04660987977404148, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.04660987977404148, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013732128776609898 + ], + "opposite_logprob_tokens": [ + -4.388731956481934 + ], + "logprob_reward": -0.013732128776609898, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013732128776609898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0540640763938427, + 0.08356443792581558, + 0.2089061290025711, + 0.05867435038089752, + 0.20251913368701935, + 0.04162265732884407, + 0.04484141990542412, + 0.11587626487016678, + 0.1585448682308197, + 0.03138666972517967 + ], + "credit_map": [ + 0.0062818368896842, + 0.005282743833959103, + 0.006565704010426998, + 0.004847047850489616, + 0.005839962977916002, + 0.00752576207742095, + 0.006926220841705799, + 0.010794798843562603, + 0.010745164006948471, + 0.004515381529927254, + 0.010857382789254189, + 0.005518683698028326, + 0.0070471749641001225, + 0.01486422773450613, + 0.009403559379279613, + 0.02061286009848118, + 0.028055503964424133, + 0.041209958493709564, + 0.021141568198800087, + 0.01792246662080288, + 0.026084378361701965, + 0.024207768961787224, + 0.018059730529785156, + 0.03222475200891495, + 0.006840521469712257, + 0.00682988204061985, + 0.006638381630182266, + 0.0061859977431595325, + 0.009034735150635242, + 0.00926950667053461, + 0.0075541152618825436, + 0.0063212099485099316, + 0.0666164979338646, + 0.08780606091022491, + 0.007377925794571638, + 0.007314362097531557, + 0.007725730538368225, + 0.014035993255674839, + 0.00645027169957757, + 0.005192293785512447, + 0.0054387361742556095, + 0.006769248284399509, + 0.005825035274028778, + 0.004587535280734301, + 0.006564782932400703, + 0.005349947139620781, + 0.0040643177926540375, + 0.003023054450750351, + 0.005475467536598444, + 0.004615775775164366, + 0.005862439516931772, + 0.005262166261672974, + 0.0049921064637601376, + 0.006574912928044796, + 0.006088453345000744, + 0.005970098078250885, + 0.016735892742872238, + 0.021542109549045563, + 0.021339308470487595, + 0.014296330511569977, + 0.00753667950630188, + 0.010305982083082199, + 0.012625462375581264, + 0.011494501493871212, + 0.015054532326757908, + 0.021087396889925003, + 0.03064587339758873, + 0.019569987431168556, + 0.012268227525055408, + 0.019792435690760612, + 0.01960587315261364, + 0.020520532503724098, + 0.003077695146203041, + 0.004552746191620827, + 0.004549970850348473, + 0.0046279169619083405, + 0.0029252134263515472, + 0.00267581595107913, + 0.005492896307259798, + 0.0034844123292714357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.013732128776609898 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003645583987236 + ], + "opposite_logprob_tokens": [ + -2.3503644466400146 + ], + "logprob_reward": -0.1003645583987236, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003645583987236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07943131774663925, + 0.160959854722023, + 0.10998512804508209, + 0.12380512803792953, + 0.24971818923950195, + 0.03240643069148064, + 0.06002949923276901, + 0.06622439622879028, + 0.1047239601612091, + 0.012716089375317097 + ], + "credit_map": [ + 0.009412501938641071, + 0.005191063042730093, + 0.010930373333394527, + 0.006881645880639553, + 0.0068752346560359, + 0.015542193315923214, + 0.008456909097731113, + 0.016141405329108238, + 0.02082289569079876, + 0.009447979740798473, + 0.02100103534758091, + 0.012603197246789932, + 0.01587487943470478, + 0.029589489102363586, + 0.021822011098265648, + 0.02979835867881775, + 0.009942295029759407, + 0.01718819886445999, + 0.014135550707578659, + 0.011287647299468517, + 0.01582172140479088, + 0.013553417287766933, + 0.010909744538366795, + 0.017146551981568336, + 0.01726207695901394, + 0.014103665947914124, + 0.009656015783548355, + 0.01577495038509369, + 0.019514011219143867, + 0.01597483642399311, + 0.01836901158094406, + 0.013150554150342941, + 0.06888236850500107, + 0.08765920251607895, + 0.014186265878379345, + 0.017020922154188156, + 0.011021688580513, + 0.020755214616656303, + 0.016102513298392296, + 0.014090018346905708, + 0.0040431455709040165, + 0.004033191595226526, + 0.0030053192749619484, + 0.0036952125374227762, + 0.006071502808481455, + 0.004195402376353741, + 0.0028490866534411907, + 0.004513570573180914, + 0.011705839075148106, + 0.015094535425305367, + 0.006160602439194918, + 0.007811469957232475, + 0.004453487694263458, + 0.004329122602939606, + 0.005669137462973595, + 0.004805303178727627, + 0.009340970776975155, + 0.009489350952208042, + 0.010847772471606731, + 0.008198563940823078, + 0.005519410129636526, + 0.008487618528306484, + 0.00704539567232132, + 0.0072953058406710625, + 0.014350265264511108, + 0.013206899166107178, + 0.020911889150738716, + 0.013037402182817459, + 0.007987280376255512, + 0.011901200748980045, + 0.011479958891868591, + 0.011849066242575645, + 0.0011521890992298722, + 0.0015096981078386307, + 0.0021044984459877014, + 0.001908586360514164, + 0.0019011294934898615, + 0.0012455105315893888, + 0.0017588890623301268, + 0.0011355887399986386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.1003645583987236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019417019793763757 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019417019793763757, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019417019793763757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04767342656850815, + 0.043354328721761703, + 0.22650457918643951, + 0.046699222177267075, + 0.13171833753585815, + 0.10367859899997711, + 0.04319826513528824, + 0.13014717400074005, + 0.21117444336414337, + 0.015851670876145363 + ], + "credit_map": [ + 0.007680351845920086, + 0.0029719755984842777, + 0.00533231720328331, + 0.003976929932832718, + 0.005562799517065287, + 0.006453932262957096, + 0.00822332501411438, + 0.007471794728189707, + 0.005323098506778479, + 0.0026081607211381197, + 0.005277566611766815, + 0.0038656059186905622, + 0.00510045001283288, + 0.006180753465741873, + 0.007014667149633169, + 0.007984023541212082, + 0.023396186530590057, + 0.0351918563246727, + 0.02580999955534935, + 0.020697224885225296, + 0.029991431161761284, + 0.03108280897140503, + 0.02151491679251194, + 0.03882014378905296, + 0.005496683064848185, + 0.005901100113987923, + 0.005024423822760582, + 0.005124873481690884, + 0.004533342085778713, + 0.007116764318197966, + 0.005989293567836285, + 0.007512737065553665, + 0.04024264216423035, + 0.049960240721702576, + 0.0055260309018194675, + 0.009774675592780113, + 0.005357746034860611, + 0.008509999141097069, + 0.006485984660685062, + 0.005861018784344196, + 0.016015997156500816, + 0.02303430810570717, + 0.01149025745689869, + 0.01606942154467106, + 0.013494069688022137, + 0.009778209030628204, + 0.0072629135102033615, + 0.006533417850732803, + 0.0036072004586458206, + 0.004677505698055029, + 0.005735547281801701, + 0.005967488046735525, + 0.005560646299272776, + 0.005401571746915579, + 0.0055587091483175755, + 0.0066895922645926476, + 0.015259201638400555, + 0.01467197760939598, + 0.023170752450823784, + 0.016883770003914833, + 0.009550927206873894, + 0.01881380006670952, + 0.011948115192353725, + 0.019848626106977463, + 0.03768311068415642, + 0.0336628220975399, + 0.03737443685531616, + 0.023033466190099716, + 0.014056364074349403, + 0.021367672830820084, + 0.02204008214175701, + 0.02195647545158863, + 0.0018438195111230016, + 0.0026478348299860954, + 0.0018971515819430351, + 0.0022586036939173937, + 0.0016795045230537653, + 0.001507594482973218, + 0.002591238822788, + 0.0014259214513003826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.0019417019793763757 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291562765836716 + ], + "opposite_logprob_tokens": [ + -2.237915515899658 + ], + "logprob_reward": -0.11291562765836716, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291562765836716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029390573501586914, + 0.009089591912925243, + 0.10209006071090698, + 0.015279233455657959, + 0.33516252040863037, + 0.07154498994350433, + 0.030756644904613495, + 0.1345922350883484, + 0.263311505317688, + 0.008782683871686459 + ], + "credit_map": [ + 0.004917473532259464, + 0.0022634880151599646, + 0.0028888543602079153, + 0.003309677354991436, + 0.00411146879196167, + 0.0031086853705346584, + 0.005899107549339533, + 0.0028918166644871235, + 0.001231675036251545, + 0.0007270702044479549, + 0.0011433926410973072, + 0.0009645365062169731, + 0.0009554560529068112, + 0.0013225519796833396, + 0.0011948943138122559, + 0.001550013548694551, + 0.010074624791741371, + 0.016621505841612816, + 0.012180655263364315, + 0.010133475065231323, + 0.015019125305116177, + 0.013399006798863411, + 0.010262084193527699, + 0.01439957320690155, + 0.0021970027592033148, + 0.0013846592046320438, + 0.0019323138985782862, + 0.0016950778663158417, + 0.0014260902535170317, + 0.002700903220102191, + 0.0017048935405910015, + 0.002238290151581168, + 0.11254440993070602, + 0.14442284405231476, + 0.011857296340167522, + 0.012474720366299152, + 0.012272670865058899, + 0.022652797400951385, + 0.010709457099437714, + 0.00822830107063055, + 0.012954906560480595, + 0.018870387226343155, + 0.007461304310709238, + 0.00913112424314022, + 0.006158788688480854, + 0.007164338603615761, + 0.004812575876712799, + 0.004991559777408838, + 0.0027920391876250505, + 0.0035651344805955887, + 0.0032410193234682083, + 0.004457020666450262, + 0.003908342681825161, + 0.003084855619817972, + 0.004262414760887623, + 0.005445815157145262, + 0.019019251689314842, + 0.024363946169614792, + 0.02557450346648693, + 0.01702755317091942, + 0.009090825915336609, + 0.01331383641809225, + 0.013680308125913143, + 0.012521974742412567, + 0.03529847040772438, + 0.03131990507245064, + 0.05147169530391693, + 0.03196070343255997, + 0.019590305164456367, + 0.030641142278909683, + 0.030940096825361252, + 0.032089173793792725, + 0.0012764499988406897, + 0.0019703449215739965, + 0.0013964201789349318, + 0.0010671026539057493, + 0.000534990627784282, + 0.0007568799774162471, + 0.0011113383807241917, + 0.000669156142976135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.11291562765836716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095382057130337 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095382057130337, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095382057130337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03428854048252106, + 0.11335282027721405, + 0.14507301151752472, + 0.04480262100696564, + 0.34248119592666626, + 0.05750863626599312, + 0.03708226606249809, + 0.06390822678804398, + 0.04332103580236435, + 0.11818163096904755 + ], + "credit_map": [ + 0.004842023830860853, + 0.0029213265515863895, + 0.004214330576360226, + 0.002892535412684083, + 0.004173872526735067, + 0.006072907242923975, + 0.003962370101362467, + 0.005209173541516066, + 0.014565826393663883, + 0.006995127070695162, + 0.015082130208611488, + 0.008627022616565228, + 0.010822593234479427, + 0.02321980521082878, + 0.01378476433455944, + 0.020255545154213905, + 0.014069897122681141, + 0.017390869557857513, + 0.01660924032330513, + 0.015126526355743408, + 0.018810607492923737, + 0.021871479228138924, + 0.013727460987865925, + 0.02746693603694439, + 0.005896138027310371, + 0.003597974544391036, + 0.005756323225796223, + 0.0051632528193295, + 0.005753092467784882, + 0.008322104811668396, + 0.004689319524914026, + 0.005624414421617985, + 0.11395237594842911, + 0.1475452184677124, + 0.01199974026530981, + 0.012937739491462708, + 0.012493695132434368, + 0.02326689101755619, + 0.01125966664403677, + 0.009025883860886097, + 0.009836168959736824, + 0.012735644355416298, + 0.007175307720899582, + 0.006058598402887583, + 0.004902281798422337, + 0.008506862446665764, + 0.0029703755863010883, + 0.005323396995663643, + 0.004688797052949667, + 0.005222877021878958, + 0.004184789024293423, + 0.004717628937214613, + 0.0044269561767578125, + 0.0043197679333388805, + 0.0047368877567350864, + 0.004784560762345791, + 0.014707785099744797, + 0.015739833936095238, + 0.005113399121910334, + 0.005120931193232536, + 0.0036106377374380827, + 0.005258807446807623, + 0.005899603944271803, + 0.008457229472696781, + 0.012119406834244728, + 0.012751413509249687, + 0.005270273890346289, + 0.0035939584486186504, + 0.0016371967503800988, + 0.0024144980125129223, + 0.003082762937992811, + 0.0024515266995877028, + 0.023837558925151825, + 0.02592465654015541, + 0.022140715271234512, + 0.010768103413283825, + 0.003929299768060446, + 0.007648642640560865, + 0.009384074248373508, + 0.014548582956194878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.004095382057130337 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an extreme close-up, the camera frozen, focused on a single, intact dandelion. In the soft, golden afternoon sunlight, its perfect sphere of countless delicate white hairs appears remarkably sacred, the light filtering through, giving each hair a soft halo around its edges. The background blurs into a soft patch of green and gold. A gentle breeze sweeps across the dandelion, causing the entire ball to tremble slightly. Dozens of mature seeds begin to loosen, then, carefully lifted by the wind, detach themselves from the dark brown receptacle and begin their journey into the distance. At this moment, the camera, in a seamless transition, precisely shifts focus from the mother plant to one of the newly departed seeds, and begins a smooth tracking motion to the right. This seed, with its dark brown spindle-shaped seeds and radiating light umbrella-shaped pappus, gracefully rotates and undulates in the air. Its white pappus appears translucent in the light. It is carried steadily and calmly by the air currents, floating towards the depths of the blurred, sun-drenched meadow in the distance. The whole picture is filled with a poetic atmosphere of tranquility, hope and continuation of life, creating an ultimate sense of fantasy.", + "seed": 32831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_6/index_2275_seed_32831_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019581071962602437, + "vlm_logprob": -0.019581071962602437, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.019581071962602437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013034261763095856 + ], + "opposite_logprob_tokens": [ + -4.513034343719482 + ], + "logprob_reward": -0.013034261763095856, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013034261763095856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03214152529835701, + 0.03713798522949219, + 0.050293199717998505, + 0.06285262107849121, + 0.12979115545749664, + 0.09519203752279282, + 0.12189969420433044, + 0.03142555430531502, + 0.06007913500070572, + 0.37918710708618164 + ], + "credit_map": [ + 0.0035007258411496878, + 0.00408188859000802, + 0.0044202059507369995, + 0.004323222208768129, + 0.004934630822390318, + 0.0030685511883348227, + 0.0037642568349838257, + 0.004048042465001345, + 0.0020721752662211657, + 0.005540990270674229, + 0.0018998186569660902, + 0.003626563586294651, + 0.0033032314386218786, + 0.00922730565071106, + 0.0031143403612077236, + 0.00835355743765831, + 0.006089651957154274, + 0.007844265550374985, + 0.0037780930288136005, + 0.005679482594132423, + 0.0069785406813025475, + 0.007478604093194008, + 0.005910069681704044, + 0.00653449259698391, + 0.005042309407144785, + 0.012842698022723198, + 0.007340657990425825, + 0.010044334456324577, + 0.008033164776861668, + 0.007375306449830532, + 0.007484345231205225, + 0.004689805209636688, + 0.016168832778930664, + 0.029260002076625824, + 0.012773429043591022, + 0.010233075357973576, + 0.016732869669795036, + 0.015820929780602455, + 0.014735706150531769, + 0.014066317118704319, + 0.009628597646951675, + 0.008899962529540062, + 0.008661932311952114, + 0.008682884275913239, + 0.015088408254086971, + 0.013886692933738232, + 0.016621319577097893, + 0.013722239062190056, + 0.009720881469547749, + 0.023277781903743744, + 0.012979182414710522, + 0.014958028681576252, + 0.006640495732426643, + 0.015566226094961166, + 0.006130796391516924, + 0.03262630105018616, + 0.003491688519716263, + 0.004354399628937244, + 0.004414529073983431, + 0.0030762769747525454, + 0.0032911726739257574, + 0.006359892897307873, + 0.003266350831836462, + 0.0031712462659925222, + 0.00383575726300478, + 0.01602245308458805, + 0.003942802082747221, + 0.004993070848286152, + 0.003737138118594885, + 0.017714623361825943, + 0.004849468357861042, + 0.004983821883797646, + 0.019358037039637566, + 0.16177120804786682, + 0.008718458004295826, + 0.03289249911904335, + 0.040792692452669144, + 0.0982951819896698, + 0.0069727906957268715, + 0.010386237874627113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.013034261763095856 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499965697526932 + ], + "opposite_logprob_tokens": [ + -2.929999589920044 + ], + "logprob_reward": -0.05499965697526932, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499965697526932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06455317884683609, + 0.058133382350206375, + 0.09389787167310715, + 0.03127521649003029, + 0.10688208043575287, + 0.12388220429420471, + 0.07472418248653412, + 0.0815911516547203, + 0.176935613155365, + 0.1881251186132431 + ], + "credit_map": [ + 0.005933860316872597, + 0.0077131655998528, + 0.004315569065511227, + 0.011028366163372993, + 0.009861847385764122, + 0.0065206796862185, + 0.011279942467808723, + 0.007899747230112553, + 0.0038708189968019724, + 0.007509480230510235, + 0.0037363902665674686, + 0.006675467360764742, + 0.005212768446654081, + 0.012476401403546333, + 0.004627414047718048, + 0.014024639502167702, + 0.008629491552710533, + 0.018626773729920387, + 0.01022183895111084, + 0.011514774523675442, + 0.010843484662473202, + 0.010952519252896309, + 0.013014932163059711, + 0.010094056837260723, + 0.003456435166299343, + 0.004989776760339737, + 0.0043911682441830635, + 0.004228219855576754, + 0.003329192055389285, + 0.0044321781024336815, + 0.0032537118531763554, + 0.0031945358496159315, + 0.009349983185529709, + 0.02275470830500126, + 0.011290258727967739, + 0.010956566780805588, + 0.012868886813521385, + 0.013493906706571579, + 0.012082912027835846, + 0.014084859751164913, + 0.015493965707719326, + 0.013586207292973995, + 0.012634322978556156, + 0.01165588665753603, + 0.011273874901235104, + 0.018027793616056442, + 0.02290184050798416, + 0.018308311700820923, + 0.006428451277315617, + 0.013759263791143894, + 0.009640838950872421, + 0.007253176067024469, + 0.009312789887189865, + 0.009382783435285091, + 0.008294993080198765, + 0.010651888325810432, + 0.009298414923250675, + 0.010293117724359035, + 0.011800250969827175, + 0.007177996914833784, + 0.008645563386380672, + 0.013338418677449226, + 0.011941999197006226, + 0.0090953903272748, + 0.009900347329676151, + 0.05438409000635147, + 0.009712575934827328, + 0.01570575311779976, + 0.012114760465919971, + 0.05256975442171097, + 0.011475244536995888, + 0.011073080822825432, + 0.013606362976133823, + 0.06740475445985794, + 0.01754610426723957, + 0.02076655812561512, + 0.024264203384518623, + 0.028070703148841858, + 0.007342681288719177, + 0.009123766794800758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.05499965697526932 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010449193650856614 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010449193650856614, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010449193650856614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04633080214262009, + 0.048625145107507706, + 0.025074895471334457, + 0.02359292469918728, + 0.06708566099405289, + 0.08640103787183762, + 0.218439519405365, + 0.18394675850868225, + 0.1290932297706604, + 0.17141002416610718 + ], + "credit_map": [ + 0.0056468648836016655, + 0.006655368022620678, + 0.005346784833818674, + 0.005283800885081291, + 0.006944681517779827, + 0.006887077819555998, + 0.0047080921940505505, + 0.004858129192143679, + 0.003579846816137433, + 0.004595960956066847, + 0.0034660985693335533, + 0.007127754390239716, + 0.003325199242681265, + 0.011147447861731052, + 0.004346285481005907, + 0.011036553420126438, + 0.001888817991130054, + 0.004324376583099365, + 0.002328821923583746, + 0.003477212507277727, + 0.003154583042487502, + 0.004211318679153919, + 0.0030618507880717516, + 0.0026279159355908632, + 0.0021279549691826105, + 0.004181030206382275, + 0.003005087608471513, + 0.003212462179362774, + 0.0026402981020510197, + 0.0035170249175280333, + 0.001830964582040906, + 0.003078102134168148, + 0.010638243518769741, + 0.0070570376701653, + 0.008623063564300537, + 0.008479070849716663, + 0.008939827792346478, + 0.00919849332422018, + 0.0066114263609051704, + 0.007538497913628817, + 0.009867488406598568, + 0.011013172566890717, + 0.009473951533436775, + 0.006724820472300053, + 0.011731671169400215, + 0.012367112562060356, + 0.013173059560358524, + 0.012049760669469833, + 0.016417203471064568, + 0.039725471287965775, + 0.020177332684397697, + 0.025013839825987816, + 0.01354195922613144, + 0.028774017468094826, + 0.014186201617121696, + 0.06060350313782692, + 0.01381191611289978, + 0.025938730686903, + 0.01931227557361126, + 0.020703891292214394, + 0.021258193999528885, + 0.042895663529634476, + 0.024890940636396408, + 0.01513515505939722, + 0.01072706189006567, + 0.0284698698669672, + 0.008833231404423714, + 0.015782218426465988, + 0.022285016253590584, + 0.01591394655406475, + 0.015582958236336708, + 0.011498936451971531, + 0.012250639498233795, + 0.03421050310134888, + 0.010238115675747395, + 0.026710540056228638, + 0.03066711686551571, + 0.019907962530851364, + 0.018442686647176743, + 0.018982453271746635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0010449193650856614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026334816589951515 + ], + "opposite_logprob_tokens": [ + -3.651334762573242 + ], + "logprob_reward": -0.026334816589951515, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026334816589951515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02529432624578476, + 0.036242082715034485, + 0.01753806509077549, + 0.023542387410998344, + 0.13301433622837067, + 0.06616353243589401, + 0.07359759509563446, + 0.0375148244202137, + 0.3851853907108307, + 0.2019074261188507 + ], + "credit_map": [ + 0.002878724131733179, + 0.004026507493108511, + 0.0022384587209671736, + 0.0027436893433332443, + 0.0043073175475001335, + 0.003655402222648263, + 0.0026331604458391666, + 0.002811065409332514, + 0.002671387977898121, + 0.0028869209345430136, + 0.001929603866301477, + 0.002866205759346485, + 0.0018662342336028814, + 0.010489393025636673, + 0.0015974992420524359, + 0.011934836395084858, + 0.0012479573488235474, + 0.002623162232339382, + 0.0020514933858066797, + 0.003354626474902034, + 0.0019871192052960396, + 0.0021746521815657616, + 0.0022706654854118824, + 0.0018283899407833815, + 0.0022612635511904955, + 0.004583127796649933, + 0.0025789774954319, + 0.0032577503006905317, + 0.0030075630638748407, + 0.003075401997193694, + 0.002209892263635993, + 0.002568410709500313, + 0.009713332168757915, + 0.02720475196838379, + 0.013941056095063686, + 0.01666080765426159, + 0.012490167282521725, + 0.017841245979070663, + 0.0172817874699831, + 0.017881188541650772, + 0.007867978885769844, + 0.007665163837373257, + 0.008258730173110962, + 0.006407986395061016, + 0.010323859751224518, + 0.009710771031677723, + 0.006650707684457302, + 0.009278339333832264, + 0.005813161376863718, + 0.009392695501446724, + 0.009403349831700325, + 0.007453067693859339, + 0.008877472952008247, + 0.009837834164500237, + 0.007438333705067635, + 0.015381676144897938, + 0.005271029658615589, + 0.006107843481004238, + 0.004552754573523998, + 0.004024348687380552, + 0.0036473439540714025, + 0.006407002918422222, + 0.0036523453891277313, + 0.003852157387882471, + 0.028879817575216293, + 0.12127935141324997, + 0.020704444497823715, + 0.022089852020144463, + 0.026155736297369003, + 0.1291952282190323, + 0.01419451367110014, + 0.022686433047056198, + 0.01059806253761053, + 0.07993195950984955, + 0.013215983286499977, + 0.017438672482967377, + 0.020982014015316963, + 0.04622386023402214, + 0.007797955069690943, + 0.005718924105167389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.026334816589951515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024917051196098328 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024917051196098328, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024917051196098328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15559349954128265, + 0.04204532131552696, + 0.02906959503889084, + 0.027377992868423462, + 0.05714857950806618, + 0.09542369842529297, + 0.027421580627560616, + 0.03753114864230156, + 0.22983112931251526, + 0.2985574007034302 + ], + "credit_map": [ + 0.02454604208469391, + 0.014181512407958508, + 0.014648704789578915, + 0.015545722097158432, + 0.014054525643587112, + 0.02760661393404007, + 0.018162989988923073, + 0.02684740349650383, + 0.0062692295759916306, + 0.004238907713443041, + 0.004917575046420097, + 0.005942059215158224, + 0.004240299109369516, + 0.007510773837566376, + 0.00445169722661376, + 0.004474782384932041, + 0.0031060294713824987, + 0.004872896242886782, + 0.002816226799041033, + 0.002606677357107401, + 0.003002103418111801, + 0.004453111439943314, + 0.002490508835762739, + 0.005722044967114925, + 0.002693003509193659, + 0.004808891098946333, + 0.0029054919723421335, + 0.00321459979750216, + 0.0031963279470801353, + 0.0036662688944488764, + 0.003915620036423206, + 0.0029777921736240387, + 0.007105041760951281, + 0.011150963604450226, + 0.0056043281219899654, + 0.004258628934621811, + 0.007288061082363129, + 0.007479130756109953, + 0.008091446943581104, + 0.006170980632305145, + 0.015468401834368706, + 0.012172211892902851, + 0.011329025961458683, + 0.009616594761610031, + 0.012232719920575619, + 0.01286184974014759, + 0.010667477734386921, + 0.011075424961745739, + 0.003077887697145343, + 0.0035005027893930674, + 0.003762173466384411, + 0.004001901485025883, + 0.002883262000977993, + 0.003349600825458765, + 0.00315541191957891, + 0.003690843703225255, + 0.00315313832834363, + 0.0042197853326797485, + 0.0045708236284554005, + 0.00398568669334054, + 0.004003438167273998, + 0.007597592193633318, + 0.006056671030819416, + 0.003944013733416796, + 0.018722325563430786, + 0.06347470730543137, + 0.018220439553260803, + 0.030881185084581375, + 0.03440287336707115, + 0.028893006965517998, + 0.015292027965188026, + 0.01994459517300129, + 0.016719292849302292, + 0.11868031322956085, + 0.013215241022408009, + 0.024720698595046997, + 0.03259219974279404, + 0.07086408138275146, + 0.010344711132347584, + 0.011420891620218754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0024917051196098328 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an extreme close-up, the camera frozen, focused on a single, intact dandelion. In the soft, golden afternoon sunlight, its perfect sphere of countless delicate white hairs appears remarkably sacred, the light filtering through, giving each hair a soft halo around its edges. The background blurs into a soft patch of green and gold. A gentle breeze sweeps across the dandelion, causing the entire ball to tremble slightly. Dozens of mature seeds begin to loosen, then, carefully lifted by the wind, detach themselves from the dark brown receptacle and begin their journey into the distance. At this moment, the camera, in a seamless transition, precisely shifts focus from the mother plant to one of the newly departed seeds, and begins a smooth tracking motion to the right. This seed, with its dark brown spindle-shaped seeds and radiating light umbrella-shaped pappus, gracefully rotates and undulates in the air. Its white pappus appears translucent in the light. It is carried steadily and calmly by the air currents, floating towards the depths of the blurred, sun-drenched meadow in the distance. The whole picture is filled with a poetic atmosphere of tranquility, hope and continuation of life, creating an ultimate sense of fantasy.", + "seed": 32832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_6/index_2275_seed_32832_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024717477057129143, + "vlm_logprob": -0.024717477057129143, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.024717477057129143, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005443867761641741 + ], + "opposite_logprob_tokens": [ + -5.505444049835205 + ], + "logprob_reward": -0.005443867761641741, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005443867761641741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04302345961332321, + 0.02116028033196926, + 0.3451588749885559, + 0.041002705693244934, + 0.07949977368116379, + 0.024733494967222214, + 0.344743937253952, + 0.027287982404232025, + 0.05861871317028999, + 0.01477082259953022 + ], + "credit_map": [ + 0.005560922436416149, + 0.004568819887936115, + 0.00656779995188117, + 0.0035316720604896545, + 0.003504754975438118, + 0.008591574616730213, + 0.00446484237909317, + 0.006233074702322483, + 0.003476629266515374, + 0.0018894757376983762, + 0.0027279385831207037, + 0.002547062234953046, + 0.0019900647457689047, + 0.003948951605707407, + 0.0021973373368382454, + 0.0023828197736293077, + 0.05425610765814781, + 0.020070983096957207, + 0.03998946398496628, + 0.02096598595380783, + 0.021816402673721313, + 0.1106049194931984, + 0.019167693331837654, + 0.05828734114766121, + 0.005058625712990761, + 0.005839409306645393, + 0.003831737209111452, + 0.003778229933232069, + 0.004561335779726505, + 0.00827957782894373, + 0.00419554952532053, + 0.005458236206322908, + 0.007929232902824879, + 0.016997262835502625, + 0.006535861175507307, + 0.009670869447290897, + 0.009261091239750385, + 0.011615484952926636, + 0.010557614266872406, + 0.006932360585778952, + 0.0029674398247152567, + 0.003768801921978593, + 0.00313760107383132, + 0.003480197163298726, + 0.0019433811539784074, + 0.002288239309564233, + 0.0033068242482841015, + 0.003841011319309473, + 0.02187497541308403, + 0.08255703747272491, + 0.05763532593846321, + 0.03254903107881546, + 0.05761620029807091, + 0.034834228456020355, + 0.036668404936790466, + 0.02100871130824089, + 0.0021602739579975605, + 0.003508942201733589, + 0.002331844065338373, + 0.0020170514471828938, + 0.0032130805775523186, + 0.00877652782946825, + 0.0030087691266089678, + 0.0022714934311807156, + 0.00493540009483695, + 0.004235492087900639, + 0.0030838162638247013, + 0.0056908330880105495, + 0.01618887297809124, + 0.014259099960327148, + 0.007250273134559393, + 0.0029749227687716484, + 0.001855193404480815, + 0.0022842951584607363, + 0.0011262597981840372, + 0.002000976586714387, + 0.002105425577610731, + 0.0025141043588519096, + 0.0019735300447791815, + 0.000911037321202457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.005443867761641741 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029827827587723732 + ], + "opposite_logprob_tokens": [ + -3.529827833175659 + ], + "logprob_reward": -0.029827827587723732, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029827827587723732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020571669563651085, + 0.026814918965101242, + 0.041522156447172165, + 0.03673914447426796, + 0.1220845952630043, + 0.09735757857561111, + 0.4634416401386261, + 0.0385013073682785, + 0.1437782198190689, + 0.009188787080347538 + ], + "credit_map": [ + 0.003283172380179167, + 0.001900009112432599, + 0.002332218922674656, + 0.0017151852371171117, + 0.0018100660527125, + 0.005029633175581694, + 0.002182598924264312, + 0.0023187852930277586, + 0.0049706376157701015, + 0.0034244046546518803, + 0.003661521477624774, + 0.0026272537652403116, + 0.0030605902429670095, + 0.0037027625367045403, + 0.002789337420836091, + 0.0025784100871533155, + 0.005354124121367931, + 0.003674518782645464, + 0.0041836644522845745, + 0.0057473196648061275, + 0.004169099498540163, + 0.008393658325076103, + 0.004742027726024389, + 0.005257740616798401, + 0.0034298875834792852, + 0.0047487011179327965, + 0.003328109858557582, + 0.0030788492877036333, + 0.0045259571634233, + 0.007880527526140213, + 0.005126082804054022, + 0.0046210261061787605, + 0.008749463595449924, + 0.036932509392499924, + 0.006726618856191635, + 0.00862538069486618, + 0.012358295731246471, + 0.012990837916731834, + 0.018942460417747498, + 0.016759023070335388, + 0.012898864224553108, + 0.02835189923644066, + 0.011012844741344452, + 0.012133367359638214, + 0.006316516548395157, + 0.007826910354197025, + 0.010927511379122734, + 0.007889657281339169, + 0.032121770083904266, + 0.13025441765785217, + 0.08286240696907043, + 0.038347505033016205, + 0.06830905377864838, + 0.04375377297401428, + 0.04230106621980667, + 0.025491634383797646, + 0.0035765119828283787, + 0.004380690399557352, + 0.004396497271955013, + 0.006147569511085749, + 0.005267185624688864, + 0.004058083053678274, + 0.007739542052149773, + 0.0029352258425205946, + 0.0052103339694440365, + 0.007763795554637909, + 0.004959926474839449, + 0.010318626649677753, + 0.050314925611019135, + 0.04586353898048401, + 0.013487459160387516, + 0.0058596087619662285, + 0.0008227479411289096, + 0.0009937690338119864, + 0.0008282916969619691, + 0.001996520673856139, + 0.0012625143863260746, + 0.0014036856591701508, + 0.0011832860764116049, + 0.0006979706231504679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.029827827587723732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095857031643391 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004095857031643391, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095857031643391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022712552919983864, + 0.011273863725364208, + 0.0768275037407875, + 0.07557054609060287, + 0.17755195498466492, + 0.0934319868683815, + 0.4444985091686249, + 0.06794780492782593, + 0.018604492768645287, + 0.011580699123442173 + ], + "credit_map": [ + 0.00450350483879447, + 0.002027250826358795, + 0.003178187645971775, + 0.0019743817392736673, + 0.0015202939976006746, + 0.004080112092196941, + 0.002226679353043437, + 0.0032021424267441034, + 0.0013325968757271767, + 0.000840395106934011, + 0.001620063791051507, + 0.0016791030066087842, + 0.001087841228581965, + 0.0021497036796063185, + 0.0014363436494022608, + 0.0011278161546215415, + 0.011873266659677029, + 0.005482742562890053, + 0.009015246294438839, + 0.005528394132852554, + 0.006812828592956066, + 0.02054288052022457, + 0.007466704584658146, + 0.01010543666779995, + 0.008316515944898129, + 0.011874875985085964, + 0.007137315347790718, + 0.006477696821093559, + 0.007662617135792971, + 0.017048833891749382, + 0.006029042415320873, + 0.011023644357919693, + 0.01838424988090992, + 0.026906631886959076, + 0.016897426918148994, + 0.031030574813485146, + 0.020271044224500656, + 0.028119295835494995, + 0.02050592005252838, + 0.01543680764734745, + 0.011751615442335606, + 0.03471579775214195, + 0.008096479810774326, + 0.00934026762843132, + 0.005903217475861311, + 0.005646431352943182, + 0.007491959258913994, + 0.010486215353012085, + 0.037528473883867264, + 0.1578778326511383, + 0.09973298758268356, + 0.028988966718316078, + 0.04378027841448784, + 0.032271966338157654, + 0.022416554391384125, + 0.021901462227106094, + 0.005305835511535406, + 0.00934222899377346, + 0.006270769517868757, + 0.007649521343410015, + 0.007606841158121824, + 0.016279302537441254, + 0.010261035524308681, + 0.005232269410043955, + 0.0023285094648599625, + 0.0014502181438729167, + 0.0011949763866141438, + 0.0015519890002906322, + 0.004003855865448713, + 0.004464203491806984, + 0.002432050183415413, + 0.0011786905815824866, + 0.0012107169022783637, + 0.001495976815931499, + 0.0009614045266062021, + 0.0018686865223571658, + 0.0018164142966270447, + 0.0023433160968124866, + 0.0012440942227840424, + 0.000640089507214725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.004095857031643391 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896212488412857 + ], + "opposite_logprob_tokens": [ + -2.5789620876312256 + ], + "logprob_reward": -0.07896212488412857, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896212488412857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038574837148189545, + 0.017860669642686844, + 0.0484241284430027, + 0.23252084851264954, + 0.24137473106384277, + 0.0788155347108841, + 0.2291029989719391, + 0.07166178524494171, + 0.027960501611232758, + 0.013703959994018078 + ], + "credit_map": [ + 0.009100457653403282, + 0.0030246786773204803, + 0.005072230938822031, + 0.002934408141300082, + 0.0023522907868027687, + 0.007745750714093447, + 0.002863505855202675, + 0.0054815164767205715, + 0.0035382811911404133, + 0.0015017542755231261, + 0.003588420106098056, + 0.0017237977590411901, + 0.0013443282805383205, + 0.0024340266827493906, + 0.0017839939100667834, + 0.0019460663897916675, + 0.008082622662186623, + 0.0033002474810928106, + 0.005272259470075369, + 0.0047241137363016605, + 0.0031723645515739918, + 0.012998134829103947, + 0.005112479440867901, + 0.005761909298598766, + 0.03302311897277832, + 0.02498645894229412, + 0.02044083923101425, + 0.01716538332402706, + 0.0248477254062891, + 0.049858566373586655, + 0.02752341330051422, + 0.03467535227537155, + 0.022186201065778732, + 0.06919115781784058, + 0.02290908806025982, + 0.030058184638619423, + 0.015341449528932571, + 0.023424645885825157, + 0.030152106657624245, + 0.028111882507801056, + 0.013718364760279655, + 0.021387428045272827, + 0.010700973682105541, + 0.008805617690086365, + 0.005871056579053402, + 0.005179890431463718, + 0.007719254586845636, + 0.005432948004454374, + 0.018829014152288437, + 0.08226422965526581, + 0.05087784677743912, + 0.014139745384454727, + 0.024413948878645897, + 0.01697840914130211, + 0.010171111673116684, + 0.011428704485297203, + 0.005881036631762981, + 0.009448579512536526, + 0.008170168846845627, + 0.007536841090768576, + 0.008828764781355858, + 0.014266419224441051, + 0.011397169902920723, + 0.0061328052543103695, + 0.003643778385594487, + 0.0027248908299952745, + 0.002253324491903186, + 0.0038367377128452063, + 0.004291507415473461, + 0.004983656108379364, + 0.004006140399724245, + 0.002220469294115901, + 0.0016361739253625274, + 0.0019499497720971704, + 0.0013054781593382359, + 0.0019056799355894327, + 0.0019783475436270237, + 0.0018321285024285316, + 0.0019603543914854527, + 0.001135847414843738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.07896212488412857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of a single, intact dandelion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of seeds detach from the dark brown receptacle and float away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera shift focus from the mother plant to a specific departing seed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth tracking motion to the right after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred, sun-drenched meadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257708020508289 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257708020508289, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257708020508289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057727549225091934, + 0.037099212408065796, + 0.036876287311315536, + 0.05117466300725937, + 0.029018176719546318, + 0.03929774835705757, + 0.30967122316360474, + 0.08833402395248413, + 0.23307588696479797, + 0.1177251785993576 + ], + "credit_map": [ + 0.010912910103797913, + 0.0039442311972379684, + 0.009304477833211422, + 0.0037444857880473137, + 0.003312711836770177, + 0.01339676883071661, + 0.006334848701953888, + 0.0067771184258162975, + 0.007301047444343567, + 0.0027796155773103237, + 0.004707288928329945, + 0.00434066541492939, + 0.0029019254725426435, + 0.007847461849451065, + 0.00328278006054461, + 0.003938427194952965, + 0.004290390759706497, + 0.0031033942941576242, + 0.0036506159231066704, + 0.005085339769721031, + 0.003855113871395588, + 0.007502870634198189, + 0.0035533772315829992, + 0.0058351848274469376, + 0.005306759849190712, + 0.007272852119058371, + 0.004562932997941971, + 0.004937134683132172, + 0.00537931127473712, + 0.009383885189890862, + 0.007453013677150011, + 0.006878775544464588, + 0.0026698485016822815, + 0.004109017085283995, + 0.0031312101054936647, + 0.003280541393905878, + 0.003886354388669133, + 0.003942355513572693, + 0.005596295464783907, + 0.0024025533348321915, + 0.005073337350040674, + 0.003935632761567831, + 0.004926757421344519, + 0.006206397898495197, + 0.004494979977607727, + 0.004927807487547398, + 0.00648768013343215, + 0.003245154395699501, + 0.024939777329564095, + 0.11869945377111435, + 0.07431849837303162, + 0.020062465220689774, + 0.024443546310067177, + 0.02131379395723343, + 0.01114293560385704, + 0.014750741422176361, + 0.010734678246080875, + 0.01101652067154646, + 0.015781650319695473, + 0.013098349794745445, + 0.010373697616159916, + 0.01042983029037714, + 0.010501660406589508, + 0.006397634744644165, + 0.010344967246055603, + 0.01227113138884306, + 0.01079490128904581, + 0.021588187664747238, + 0.06433552503585815, + 0.06509654968976974, + 0.03504393249750137, + 0.013600679114460945, + 0.009347904473543167, + 0.013314741663634777, + 0.007603820879012346, + 0.016891123726963997, + 0.02124643512070179, + 0.03248044475913048, + 0.009479016065597534, + 0.007361685857176781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.005257708020508289 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_tensor([6448, 6448])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_tensor([6448, 6448])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f0638dae018e34de8ade6d737d38eb0150cc5ebe --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_tensor([6448, 6448])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle close-up slowly pans out onto a stack of clear, nested glass bowls, placed on a dark, polished wooden coffee table. The largest bowl at the bottom is filled with fluffy, buttered popcorn, its irregular golden kernels gleaming in the light. Stacked within this is a series of progressively smaller, empty glass bowls, their clean, curved surfaces reflecting and distorting the scene in miniature. The only illumination comes from the realistic, flickering, cool blue glow of an off-screen television, casting long, dancing shadows across the table and creating ever-shifting highlights on the glass borders and textured popcorn. This single light source, combined with subtle camera movements, creates a dynamic atmosphere of cozy anticipation, perfectly capturing the intimacy and focused energy of an impending movie night. The background is intentionally kept dark and out of focus to draw everyone's attention to the enticing snacks.", + "seed": 14522, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_6448_seed_14522_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.042356252856552604, + "vlm_logprob": -0.042356252856552604, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.042356252856552604, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006921718828380108 + ], + "opposite_logprob_tokens": [ + -5.506921768188477 + ], + "logprob_reward": -0.006921718828380108, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006921718828380108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0562051385641098, + 0.14923621714115143, + 0.1339833289384842, + 0.13309244811534882, + 0.23300091922283173, + 0.06486184149980545, + 0.03013288415968418, + 0.11654642224311829, + 0.05317879095673561, + 0.029761984944343567 + ], + "credit_map": [ + 0.008008052594959736, + 0.008416300639510155, + 0.006728773470968008, + 0.009042938239872456, + 0.007704540155827999, + 0.00422471622005105, + 0.006225946359336376, + 0.005853870417922735, + 0.017766935750842094, + 0.009685203433036804, + 0.030110174790024757, + 0.04560506343841553, + 0.011879337951540947, + 0.008825153112411499, + 0.01053929515182972, + 0.01482505351305008, + 0.01216008048504591, + 0.03932114318013191, + 0.035237811505794525, + 0.013731861487030983, + 0.010250904597342014, + 0.006138897035270929, + 0.01047627441585064, + 0.006666367873549461, + 0.013684054836630821, + 0.019786357879638672, + 0.01777881756424904, + 0.020478878170251846, + 0.019696366041898727, + 0.009795334190130234, + 0.023194987326860428, + 0.008677659556269646, + 0.03681802377104759, + 0.037202876061201096, + 0.030246412381529808, + 0.02725556306540966, + 0.03843706473708153, + 0.014277678914368153, + 0.03128611668944359, + 0.01747717894613743, + 0.008042513392865658, + 0.006649306043982506, + 0.0054964483715593815, + 0.012736611068248749, + 0.009098958224058151, + 0.008382996544241905, + 0.009010527282953262, + 0.005444476380944252, + 0.0041620866395533085, + 0.004278437700122595, + 0.003860309487208724, + 0.003423915710300207, + 0.0039276075549423695, + 0.0036084961611777544, + 0.0027781471144407988, + 0.004093884956091642, + 0.009589864872395992, + 0.013897374272346497, + 0.013521186076104641, + 0.012725021690130234, + 0.01340342964977026, + 0.017901429906487465, + 0.0172100979834795, + 0.01829802244901657, + 0.0054430910386145115, + 0.008992211893200874, + 0.004777275025844574, + 0.006286504678428173, + 0.006533184088766575, + 0.00902514811605215, + 0.005366202909499407, + 0.006755174603313208, + 0.004186051897704601, + 0.002104251878336072, + 0.0032403592485934496, + 0.0024347477592527866, + 0.0030025066807866096, + 0.00355339702218771, + 0.005457416642457247, + 0.005783252418041229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.006921718828380108 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090303771197796 + ], + "opposite_logprob_tokens": [ + -4.511090278625488 + ], + "logprob_reward": -0.011090303771197796, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090303771197796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24654236435890198, + 0.20339301228523254, + 0.17749741673469543, + 0.0787469744682312, + 0.11011020839214325, + 0.056240130215883255, + 0.021667657420039177, + 0.050101399421691895, + 0.03088059090077877, + 0.024820275604724884 + ], + "credit_map": [ + 0.032975539565086365, + 0.02550993114709854, + 0.033107656985521317, + 0.025316953659057617, + 0.031041273847222328, + 0.019419481977820396, + 0.03268228843808174, + 0.04648924618959427, + 0.043376997113227844, + 0.02021920494735241, + 0.037669841200113297, + 0.020517995581030846, + 0.029730260372161865, + 0.011047213338315487, + 0.02600063756108284, + 0.014830859377980232, + 0.022619301453232765, + 0.027090584859251976, + 0.023196104913949966, + 0.03673684224486351, + 0.014036925509572029, + 0.010264988988637924, + 0.034545861184597015, + 0.009006794542074203, + 0.007689537480473518, + 0.014077442698180676, + 0.011014485731720924, + 0.011439730413258076, + 0.011906531639397144, + 0.005272347014397383, + 0.012674436904489994, + 0.004672458861023188, + 0.018556928262114525, + 0.017048189416527748, + 0.014257064089179039, + 0.01336398534476757, + 0.016866713762283325, + 0.006564231589436531, + 0.015216735191643238, + 0.008236357010900974, + 0.006894620601087809, + 0.008527315221726894, + 0.006087541580200195, + 0.006277137901633978, + 0.008680380880832672, + 0.008612615056335926, + 0.006051046308130026, + 0.005109475925564766, + 0.0024688350968062878, + 0.0027210500556975603, + 0.0028545560780912638, + 0.003334851237013936, + 0.002707532374188304, + 0.0023060787934809923, + 0.002853973302990198, + 0.002420780947431922, + 0.004610233940184116, + 0.004480208735913038, + 0.008662229403853416, + 0.007576903328299522, + 0.0069325705990195274, + 0.005563524551689625, + 0.007289334665983915, + 0.004986391868442297, + 0.005205722991377115, + 0.004420362412929535, + 0.003022152464836836, + 0.004408857319504023, + 0.00326768821105361, + 0.003740173066034913, + 0.0032959759701043367, + 0.0035196568351238966, + 0.0034830542281270027, + 0.0019896638114005327, + 0.002426173770800233, + 0.0022252765484154224, + 0.002588556846603751, + 0.00350837130099535, + 0.004320083651691675, + 0.004279096145182848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011090303771197796 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798955351114273 + ], + "opposite_logprob_tokens": [ + -1.8048955202102661 + ], + "logprob_reward": -0.1798955351114273, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798955351114273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1447131633758545, + 0.11147496849298477, + 0.15641717612743378, + 0.16155797243118286, + 0.05701533704996109, + 0.15426023304462433, + 0.04400409385561943, + 0.020038925111293793, + 0.10379736870527267, + 0.04672076180577278 + ], + "credit_map": [ + 0.012130583636462688, + 0.011763869784772396, + 0.015245093032717705, + 0.014943737536668777, + 0.021408485248684883, + 0.01813851110637188, + 0.025327103212475777, + 0.025755779817700386, + 0.01051949243992567, + 0.007819050922989845, + 0.019122889265418053, + 0.03225288912653923, + 0.010267375037074089, + 0.010343646630644798, + 0.013771728612482548, + 0.007377901580184698, + 0.020145535469055176, + 0.03253664821386337, + 0.01716609112918377, + 0.02379043959081173, + 0.0158270001411438, + 0.012747425585985184, + 0.022594671696424484, + 0.01160937175154686, + 0.016686830669641495, + 0.03240473195910454, + 0.021465014666318893, + 0.023268429562449455, + 0.023521555587649345, + 0.008891922421753407, + 0.025998616591095924, + 0.009320872835814953, + 0.008934425190091133, + 0.006818998139351606, + 0.00944617111235857, + 0.006132251117378473, + 0.006402335595339537, + 0.005237447563558817, + 0.006822753231972456, + 0.007220952771604061, + 0.018803732469677925, + 0.02365504391491413, + 0.014460823498666286, + 0.019427912309765816, + 0.027925854548811913, + 0.022824224084615707, + 0.016884295269846916, + 0.010278350673615932, + 0.005027946084737778, + 0.0048212711699306965, + 0.00577244209125638, + 0.006013174075633287, + 0.0072539797984063625, + 0.006376238539814949, + 0.00480609992519021, + 0.003932942170649767, + 0.0016430820105597377, + 0.0019672273192554712, + 0.0034042918123304844, + 0.0028177027124911547, + 0.0020757168531417847, + 0.003946272190660238, + 0.002184347016736865, + 0.002000286942347884, + 0.009468608535826206, + 0.013431839644908905, + 0.009114653803408146, + 0.01272681262344122, + 0.017514536157250404, + 0.01322726346552372, + 0.012221157550811768, + 0.016092505306005478, + 0.004084710963070393, + 0.0035232289228588343, + 0.004825813230127096, + 0.0051969438791275024, + 0.00812575314193964, + 0.005444865208119154, + 0.006917819380760193, + 0.008601626381278038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.1798955351114273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09453027695417404, + 0.13608455657958984, + 0.11841696500778198, + 0.13232214748859406, + 0.42165154218673706, + 0.032778941094875336, + 0.02058728039264679, + 0.009827107191085815, + 0.028130395337939262, + 0.005670720245689154 + ], + "credit_map": [ + 0.018924664705991745, + 0.008177142590284348, + 0.027442367747426033, + 0.012831579893827438, + 0.009623143821954727, + 0.005233666393905878, + 0.006860391702502966, + 0.005437325686216354, + 0.011104981414973736, + 0.007741108071058989, + 0.028767062351107597, + 0.04683841019868851, + 0.007497692015022039, + 0.0075387535616755486, + 0.00939149409532547, + 0.01720506325364113, + 0.014009762555360794, + 0.031366076320409775, + 0.024440249428153038, + 0.012759353965520859, + 0.011541031301021576, + 0.006650516763329506, + 0.011078388430178165, + 0.006571598816663027, + 0.014216049574315548, + 0.01611297018826008, + 0.017674729228019714, + 0.020142555236816406, + 0.019644832238554955, + 0.01011530589312315, + 0.025841258466243744, + 0.008574455045163631, + 0.06915763020515442, + 0.06400719285011292, + 0.055009011179208755, + 0.049001533538103104, + 0.06861436367034912, + 0.02634466253221035, + 0.05885156989097595, + 0.030665606260299683, + 0.0040595196187496185, + 0.0034851995296776295, + 0.003030921332538128, + 0.006015351507812738, + 0.004473336506634951, + 0.004331432748585939, + 0.004430732689797878, + 0.002952447859570384, + 0.002218707464635372, + 0.002055593766272068, + 0.0024209623225033283, + 0.0031505608931183815, + 0.0024470787029713392, + 0.002872834913432598, + 0.002187018748372793, + 0.0032345266081392765, + 0.00134415237698704, + 0.0012027411721646786, + 0.0015895608812570572, + 0.0011787110706791282, + 0.001146078109741211, + 0.0013820868916809559, + 0.0009735701605677605, + 0.0010102068772539496, + 0.0029931406024843454, + 0.004147569183260202, + 0.00227882107719779, + 0.002635691547766328, + 0.004932316951453686, + 0.00371186644770205, + 0.002858654363080859, + 0.004572336561977863, + 0.0006036831182427704, + 0.0005283745704218745, + 0.0006024784524925053, + 0.0008273312705568969, + 0.0007222710410133004, + 0.0006707849679514766, + 0.0008804029203020036, + 0.0008353937300853431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.009780817665159702 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092888906598091 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092888906598091, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092888906598091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17504261434078217, + 0.0981920063495636, + 0.07119728624820709, + 0.1851469874382019, + 0.16742904484272003, + 0.043168794363737106, + 0.05086430162191391, + 0.058364685624837875, + 0.1118626669049263, + 0.038731612265110016 + ], + "credit_map": [ + 0.030728043988347054, + 0.014364580623805523, + 0.04997468367218971, + 0.025733692571520805, + 0.015346523374319077, + 0.010750286281108856, + 0.0156601183116436, + 0.012484690174460411, + 0.011219438165426254, + 0.007469217758625746, + 0.013878407888114452, + 0.019558299332857132, + 0.01141451857984066, + 0.008738820441067219, + 0.01652885414659977, + 0.00938444770872593, + 0.007383921183645725, + 0.011011704802513123, + 0.01259308960288763, + 0.01087870728224516, + 0.00906183198094368, + 0.006640882696956396, + 0.007826021872460842, + 0.005801131948828697, + 0.01851605251431465, + 0.03044680692255497, + 0.02891136147081852, + 0.029651813209056854, + 0.027295297011733055, + 0.011358973570168018, + 0.028047246858477592, + 0.010919447056949139, + 0.032595958560705185, + 0.022719010710716248, + 0.01851614937186241, + 0.017965538427233696, + 0.024767152965068817, + 0.012414318509399891, + 0.02567577175796032, + 0.012775137089192867, + 0.005069124978035688, + 0.005773543380200863, + 0.00620525237172842, + 0.004870154429227114, + 0.0067197117023169994, + 0.005155783612281084, + 0.004822570830583572, + 0.004552649799734354, + 0.0051534054800868034, + 0.006399549078196287, + 0.00706679280847311, + 0.00660323491320014, + 0.009065549820661545, + 0.006296203006058931, + 0.005286114290356636, + 0.004993455018848181, + 0.004514819942414761, + 0.008093533106148243, + 0.00711455661803484, + 0.007084137760102749, + 0.0061936448328197, + 0.007225065492093563, + 0.009204654954373837, + 0.008934271521866322, + 0.011294335126876831, + 0.013672081753611565, + 0.01186555065214634, + 0.012081339955329895, + 0.02082611620426178, + 0.01910237781703472, + 0.010723656974732876, + 0.012297213077545166, + 0.007904795929789543, + 0.0031236836221069098, + 0.004808446392416954, + 0.0037698105443269014, + 0.003556008916348219, + 0.005953518208116293, + 0.004122172482311726, + 0.0054931784979999065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.004092888906598091 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up slowly pans out onto a stack of clear, nested glass bowls, placed on a dark, polished wooden coffee table. The largest bowl at the bottom is filled with fluffy, buttered popcorn, its irregular golden kernels gleaming in the light. Stacked within this is a series of progressively smaller, empty glass bowls, their clean, curved surfaces reflecting and distorting the scene in miniature. The only illumination comes from the realistic, flickering, cool blue glow of an off-screen television, casting long, dancing shadows across the table and creating ever-shifting highlights on the glass borders and textured popcorn. This single light source, combined with subtle camera movements, creates a dynamic atmosphere of cozy anticipation, perfectly capturing the intimacy and focused energy of an impending movie night. The background is intentionally kept dark and out of focus to draw everyone's attention to the enticing snacks.", + "seed": 14523, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_6448_seed_14523_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02447678381577134, + "vlm_logprob": -0.02447678381577134, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.02447678381577134, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007293029688298702 + ], + "opposite_logprob_tokens": [ + -5.507293224334717 + ], + "logprob_reward": -0.007293029688298702, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.007293029688298702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3106420040130615, + 0.0985090583562851, + 0.11972694098949432, + 0.07152388989925385, + 0.14435546100139618, + 0.11497137695550919, + 0.04909569025039673, + 0.04423050209879875, + 0.03298425301909447, + 0.013960786163806915 + ], + "credit_map": [ + 0.03859461843967438, + 0.030443178489804268, + 0.10191277414560318, + 0.03822016343474388, + 0.019222162663936615, + 0.013918377459049225, + 0.02897053025662899, + 0.03936021402478218, + 0.008003092370927334, + 0.010485814884305, + 0.014926347881555557, + 0.018507832661271095, + 0.012800496071577072, + 0.008895402774214745, + 0.014647445641458035, + 0.010242625139653683, + 0.016235580667853355, + 0.007930542342364788, + 0.029290977865457535, + 0.013774711638689041, + 0.011745821684598923, + 0.006649831309914589, + 0.010763938538730145, + 0.0233355313539505, + 0.010326661169528961, + 0.00826717633754015, + 0.010621171444654465, + 0.007754393853247166, + 0.009848986752331257, + 0.005772427190095186, + 0.012947418726980686, + 0.005985649768263102, + 0.024659745395183563, + 0.01599068008363247, + 0.01501932367682457, + 0.018670212477445602, + 0.025067100301384926, + 0.0069077759981155396, + 0.026432303711771965, + 0.011608315631747246, + 0.026889672502875328, + 0.009251401759684086, + 0.029996763914823532, + 0.00987540278583765, + 0.010764055885374546, + 0.011698738671839237, + 0.010979616083204746, + 0.005515730939805508, + 0.0069379606284201145, + 0.004924192558974028, + 0.0055444510653615, + 0.005596712231636047, + 0.005418416578322649, + 0.006174629088491201, + 0.006947206798940897, + 0.007552118971943855, + 0.009466893970966339, + 0.0034256121143698692, + 0.00337643432430923, + 0.003149024909362197, + 0.004901300650089979, + 0.0036884001456201077, + 0.005190242547541857, + 0.011032593436539173, + 0.005086001008749008, + 0.0035942154936492443, + 0.0028223651461303234, + 0.004756447393447161, + 0.0050277565605938435, + 0.0027670131530612707, + 0.005428525619208813, + 0.0035019274801015854, + 0.002501334762200713, + 0.0012617132160812616, + 0.001467752968892455, + 0.0013106840196996927, + 0.0018913528183475137, + 0.0014116381062194705, + 0.0017248103395104408, + 0.0023915008641779423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.007293029688298702 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647376671433449 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647376671433449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647376671433449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14103038609027863, + 0.0943174734711647, + 0.3350425958633423, + 0.07797141373157501, + 0.1408456563949585, + 0.03711685538291931, + 0.07758321613073349, + 0.03761023283004761, + 0.03861803561449051, + 0.019864095374941826 + ], + "credit_map": [ + 0.011304779909551144, + 0.016121525317430496, + 0.039098385721445084, + 0.018711913377046585, + 0.009162621572613716, + 0.006463455501943827, + 0.014129964634776115, + 0.02603774517774582, + 0.008742249570786953, + 0.010919800959527493, + 0.011980610899627209, + 0.01636422984302044, + 0.011980429291725159, + 0.010299348272383213, + 0.013286874629557133, + 0.010743930004537106, + 0.05111958086490631, + 0.036192379891872406, + 0.07246407121419907, + 0.050205446779727936, + 0.03888712823390961, + 0.018086597323417664, + 0.03713540360331535, + 0.030952004715800285, + 0.012364424765110016, + 0.007492649834603071, + 0.011213398538529873, + 0.008164872415363789, + 0.006718790158629417, + 0.0074847666546702385, + 0.010484827682375908, + 0.014047684147953987, + 0.022402510046958923, + 0.018660781905055046, + 0.015170762315392494, + 0.015357047319412231, + 0.022265009582042694, + 0.01412653923034668, + 0.019069427624344826, + 0.013793570920825005, + 0.008953174576163292, + 0.0038146136794239283, + 0.009370354935526848, + 0.0037334405351430178, + 0.003834002185612917, + 0.0024794198106974363, + 0.003279606346040964, + 0.0016522436635568738, + 0.010949765332043171, + 0.007924971170723438, + 0.008402687497437, + 0.010619029402732849, + 0.0092670489102602, + 0.010197451338171959, + 0.00899630505591631, + 0.011225958354771137, + 0.005611273925751448, + 0.004589304327964783, + 0.0036081187427043915, + 0.004133324604481459, + 0.007871626876294613, + 0.0028828957583755255, + 0.005656751804053783, + 0.0032569377217441797, + 0.00482649402692914, + 0.003916867543011904, + 0.002608442911878228, + 0.00505124730989337, + 0.006988703738898039, + 0.002175688510760665, + 0.008026078343391418, + 0.005024515558034182, + 0.0028335738461464643, + 0.0016814626287668943, + 0.0019363336032256484, + 0.00247684633359313, + 0.003538214834406972, + 0.001184385735541582, + 0.004113166127353907, + 0.0021001130808144808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.008647376671433449 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890466570854187 + ], + "opposite_logprob_tokens": [ + -2.4640467166900635 + ], + "logprob_reward": -0.0890466570854187, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890466570854187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09361148625612259, + 0.1614762246608734, + 0.1478165090084076, + 0.12287280708551407, + 0.10669334977865219, + 0.1041094958782196, + 0.10632534325122833, + 0.04097696393728256, + 0.06987888365983963, + 0.046238914132118225 + ], + "credit_map": [ + 0.008310261182487011, + 0.012939754873514175, + 0.016992229968309402, + 0.013467728160321712, + 0.009312556125223637, + 0.010818245820701122, + 0.011750118806958199, + 0.01002059131860733, + 0.011183678172528744, + 0.017928175628185272, + 0.01714015007019043, + 0.024730270728468895, + 0.03125154599547386, + 0.0187864787876606, + 0.026211382821202278, + 0.01424454152584076, + 0.019379502162337303, + 0.014508930034935474, + 0.023618903011083603, + 0.028169060125947, + 0.022922338917851448, + 0.007366174831986427, + 0.012875661253929138, + 0.01897594705224037, + 0.01852099783718586, + 0.01162108313292265, + 0.016638614237308502, + 0.015424727462232113, + 0.011081647127866745, + 0.011618860065937042, + 0.014062819071114063, + 0.02390405908226967, + 0.016173364594578743, + 0.01297737192362547, + 0.009483998641371727, + 0.01123750302940607, + 0.014407267794013023, + 0.013738630339503288, + 0.014143412932753563, + 0.014531798660755157, + 0.03077729605138302, + 0.006308480631560087, + 0.03491165488958359, + 0.009920359589159489, + 0.006246010307222605, + 0.004731637891381979, + 0.006586459930986166, + 0.004627590533345938, + 0.016118871048092842, + 0.009771229699254036, + 0.014844700694084167, + 0.012539968825876713, + 0.013306699693202972, + 0.01204718928784132, + 0.013564391992986202, + 0.01413229200989008, + 0.007188495248556137, + 0.0036870744079351425, + 0.005148875992745161, + 0.004945465829223394, + 0.0057265800423920155, + 0.003870392916724086, + 0.005492348689585924, + 0.00491773197427392, + 0.008509524166584015, + 0.007531359326094389, + 0.004866448231041431, + 0.009762391448020935, + 0.013492214493453503, + 0.004293345846235752, + 0.014306992292404175, + 0.007116606924682856, + 0.006399576086550951, + 0.0039741965010762215, + 0.004231305327266455, + 0.006354338489472866, + 0.006633137818425894, + 0.0024948017671704292, + 0.010441227816045284, + 0.005710331257432699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.0890466570854187 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063792631030083 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.016063792631030083, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063792631030083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23580828309059143, + 0.12383871525526047, + 0.1258484125137329, + 0.0999542847275734, + 0.16246718168258667, + 0.19194257259368896, + 0.023158203810453415, + 0.01396925374865532, + 0.012321598827838898, + 0.01069151982665062 + ], + "credit_map": [ + 0.04381583631038666, + 0.0267648883163929, + 0.05580862984061241, + 0.03671108931303024, + 0.016668103635311127, + 0.012629971839487553, + 0.02837175875902176, + 0.015038028359413147, + 0.010513952933251858, + 0.01350435335189104, + 0.016786932945251465, + 0.024795453995466232, + 0.01589570939540863, + 0.011440892703831196, + 0.017610015347599983, + 0.01329141017049551, + 0.016743894666433334, + 0.011675172485411167, + 0.015151653438806534, + 0.018959494307637215, + 0.014683717861771584, + 0.00850636325776577, + 0.016070229932665825, + 0.024057889357209206, + 0.012328850105404854, + 0.011995777487754822, + 0.01307349931448698, + 0.011117933318018913, + 0.013269854709506035, + 0.010479089803993702, + 0.018718555569648743, + 0.008970735594630241, + 0.029705887660384178, + 0.01608494110405445, + 0.020154790952801704, + 0.020816465839743614, + 0.01972641609609127, + 0.014957613311707973, + 0.02439441904425621, + 0.01662665233016014, + 0.05002589896321297, + 0.015282434411346912, + 0.05542319267988205, + 0.017129231244325638, + 0.01586286537349224, + 0.014090655371546745, + 0.016521669924259186, + 0.007606633473187685, + 0.0038395544979721308, + 0.0017481637187302113, + 0.001669951481744647, + 0.003783985273912549, + 0.003116344101727009, + 0.0027338843792676926, + 0.002712845802307129, + 0.0035534766502678394, + 0.0027400997933000326, + 0.0011900155805051327, + 0.0014093597419559956, + 0.001154187018983066, + 0.001725015346892178, + 0.001153929391875863, + 0.0017141873249784112, + 0.00288246083073318, + 0.0016633603954687715, + 0.0012843430740758777, + 0.001433784025721252, + 0.0019975192844867706, + 0.0018076511332765222, + 0.0012970955576747656, + 0.0016161852981895208, + 0.0012216607574373484, + 0.0015318177174776793, + 0.0011616869596764445, + 0.0012004844611510634, + 0.00102761120069772, + 0.001570273656398058, + 0.0007718586712144315, + 0.002257994143292308, + 0.0011697927257046103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.016063792631030083 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013330630026757717 + ], + "opposite_logprob_tokens": [ + -6.626333236694336 + ], + "logprob_reward": -0.0013330630026757717, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013330630026757717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22917234897613525, + 0.09002702683210373, + 0.10653699189424515, + 0.101282998919487, + 0.10427390784025192, + 0.11025353521108627, + 0.05067019164562225, + 0.07091501355171204, + 0.08261118829250336, + 0.05425674095749855 + ], + "credit_map": [ + 0.036897532641887665, + 0.021810470148921013, + 0.06402371823787689, + 0.031923335045576096, + 0.015164307318627834, + 0.011889792047441006, + 0.023979971185326576, + 0.02348323166370392, + 0.0076979054138064384, + 0.00956352986395359, + 0.013061787001788616, + 0.014508758671581745, + 0.014753890223801136, + 0.007676113396883011, + 0.01450197771191597, + 0.008263061754405499, + 0.016972118988633156, + 0.00976458191871643, + 0.01595066301524639, + 0.017571359872817993, + 0.017427561804652214, + 0.007634466048330069, + 0.007812330033630133, + 0.013403905555605888, + 0.014406601898372173, + 0.009831933304667473, + 0.015046710148453712, + 0.012585312128067017, + 0.010050048120319843, + 0.008627044036984444, + 0.015345868654549122, + 0.015389477834105492, + 0.016706367954611778, + 0.009685668163001537, + 0.017879832535982132, + 0.012763261795043945, + 0.013202346861362457, + 0.011220598593354225, + 0.010562182404100895, + 0.012253651395440102, + 0.03129559010267258, + 0.00860202219337225, + 0.03316386044025421, + 0.009113886393606663, + 0.008556660264730453, + 0.006157486699521542, + 0.007390367332845926, + 0.00597366550937295, + 0.008150678128004074, + 0.004210396204143763, + 0.008605594746768475, + 0.010797400958836079, + 0.004363880958408117, + 0.004357138182967901, + 0.005229963455349207, + 0.004955141339451075, + 0.01214528176933527, + 0.0068059395998716354, + 0.009037861600518227, + 0.007688954006880522, + 0.009614448063075542, + 0.006660090293735266, + 0.009525533765554428, + 0.009436908178031445, + 0.011334221810102463, + 0.009089743718504906, + 0.005972249899059534, + 0.011784778907895088, + 0.014569181948900223, + 0.005094702821224928, + 0.015981284901499748, + 0.008785020560026169, + 0.009696315973997116, + 0.004904735367745161, + 0.005433166399598122, + 0.007197514642030001, + 0.005732683464884758, + 0.0050033158622682095, + 0.010625171475112438, + 0.005663839168846607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0013330630026757717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up slowly pans out onto a stack of clear, nested glass bowls, placed on a dark, polished wooden coffee table. The largest bowl at the bottom is filled with fluffy, buttered popcorn, its irregular golden kernels gleaming in the light. Stacked within this is a series of progressively smaller, empty glass bowls, their clean, curved surfaces reflecting and distorting the scene in miniature. The only illumination comes from the realistic, flickering, cool blue glow of an off-screen television, casting long, dancing shadows across the table and creating ever-shifting highlights on the glass borders and textured popcorn. This single light source, combined with subtle camera movements, creates a dynamic atmosphere of cozy anticipation, perfectly capturing the intimacy and focused energy of an impending movie night. The background is intentionally kept dark and out of focus to draw everyone's attention to the enticing snacks.", + "seed": 14524, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_6448_seed_14524_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09204270448535681, + "vlm_logprob": -0.09204270448535681, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.09204270448535681, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011216678656637669 + ], + "opposite_logprob_tokens": [ + -5.136216640472412 + ], + "logprob_reward": -0.011216678656637669, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.011216678656637669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3440057337284088, + 0.04030174762010574, + 0.0642063245177269, + 0.050468720495700836, + 0.10868216305971146, + 0.08993962407112122, + 0.13504129648208618, + 0.05660263076424599, + 0.04305671527981758, + 0.06769502907991409 + ], + "credit_map": [ + 0.025669118389487267, + 0.03910071775317192, + 0.04809577018022537, + 0.08473162353038788, + 0.03062349371612072, + 0.026215389370918274, + 0.034107837826013565, + 0.055461786687374115, + 0.004455581773072481, + 0.0038895781617611647, + 0.006098441779613495, + 0.005435842089354992, + 0.004995189607143402, + 0.00563438143581152, + 0.004898937419056892, + 0.004893796052783728, + 0.006632589735090733, + 0.007577225100249052, + 0.008128723129630089, + 0.01278656255453825, + 0.007006795611232519, + 0.00832391157746315, + 0.006706096697598696, + 0.007044415455311537, + 0.005348526407033205, + 0.004735791590064764, + 0.011059652082622051, + 0.0041573611088097095, + 0.008209607563912868, + 0.004221875686198473, + 0.007371831219643354, + 0.005364073906093836, + 0.010662720538675785, + 0.007089060731232166, + 0.015329273417592049, + 0.014758125878870487, + 0.01628994010388851, + 0.01197060290724039, + 0.019245225936174393, + 0.01333721075206995, + 0.01230931468307972, + 0.00839950330555439, + 0.012040085159242153, + 0.010730575770139694, + 0.012887807562947273, + 0.01492246799170971, + 0.009623396210372448, + 0.009026473388075829, + 0.020148085430264473, + 0.010135140269994736, + 0.024521155282855034, + 0.011779178865253925, + 0.007877497002482414, + 0.037631772458553314, + 0.011326811276376247, + 0.011621656827628613, + 0.00726793659850955, + 0.003792369272559881, + 0.009794275276362896, + 0.00420206505805254, + 0.004705163184553385, + 0.010541073977947235, + 0.007637981325387955, + 0.008661768399178982, + 0.0039123534224927425, + 0.003173242788761854, + 0.007266767788678408, + 0.004051178228110075, + 0.004085990134626627, + 0.009157766588032246, + 0.004993468523025513, + 0.0064159478060901165, + 0.00779800396412611, + 0.006491438951343298, + 0.010168751701712608, + 0.005364443641155958, + 0.008541012182831764, + 0.009008495137095451, + 0.010994743555784225, + 0.0093281464651227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.011216678656637669 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964338313788176 + ], + "opposite_logprob_tokens": [ + -5.130964279174805 + ], + "logprob_reward": -0.005964338313788176, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964338313788176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35691991448402405, + 0.08402874320745468, + 0.02715691737830639, + 0.053166355937719345, + 0.11233575642108917, + 0.09264808893203735, + 0.03413083031773567, + 0.07458748668432236, + 0.10284600406885147, + 0.062179967761039734 + ], + "credit_map": [ + 0.021415837109088898, + 0.04126916453242302, + 0.041944216936826706, + 0.09502483159303665, + 0.03425603732466698, + 0.032825883477926254, + 0.04346098750829697, + 0.046722959727048874, + 0.007882214151322842, + 0.009050412103533745, + 0.00967242382466793, + 0.01287856139242649, + 0.014715254306793213, + 0.008005536161363125, + 0.013101407326757908, + 0.00872293021529913, + 0.0033680095802992582, + 0.0030217343010008335, + 0.0038244985044002533, + 0.002874807920306921, + 0.002340727485716343, + 0.0032921084202826023, + 0.004029222298413515, + 0.004405807238072157, + 0.006579446606338024, + 0.004365057684481144, + 0.011966001242399216, + 0.005241075064986944, + 0.010588309727609158, + 0.004161788150668144, + 0.00523744011297822, + 0.005027235485613346, + 0.01144018117338419, + 0.009185624308884144, + 0.012688563205301762, + 0.013402322307229042, + 0.018064482137560844, + 0.013459139503538609, + 0.01958976499736309, + 0.01450567040592432, + 0.010094665922224522, + 0.00993508379906416, + 0.012749334797263145, + 0.010051755234599113, + 0.014417281374335289, + 0.012824783101677895, + 0.01183517836034298, + 0.01073999609798193, + 0.003332983935251832, + 0.003408942837268114, + 0.0038079041987657547, + 0.00319324410520494, + 0.006522524636238813, + 0.004485039971768856, + 0.006433094851672649, + 0.002947096712887287, + 0.005502012092620134, + 0.003801124868914485, + 0.008166409097611904, + 0.004992898087948561, + 0.007734507787972689, + 0.022485017776489258, + 0.007982729934155941, + 0.013922778889536858, + 0.009372136555612087, + 0.006146002095192671, + 0.01732126995921135, + 0.0063858297653496265, + 0.006593399215489626, + 0.03258240222930908, + 0.009311437606811523, + 0.015133517794311047, + 0.005712741985917091, + 0.004622966982424259, + 0.010448841378092766, + 0.006319655571132898, + 0.004276474937796593, + 0.011137108318507671, + 0.011316210962831974, + 0.008345959708094597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.005964338313788176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038018465042114 + ], + "opposite_logprob_tokens": [ + -1.9103801250457764 + ], + "logprob_reward": -0.16038018465042114, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038018465042114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08869301527738571, + 0.06306595355272293, + 0.05925997719168663, + 0.06234915181994438, + 0.3816055357456207, + 0.22388900816440582, + 0.020656200125813484, + 0.015842558816075325, + 0.04452178627252579, + 0.04011676833033562 + ], + "credit_map": [ + 0.006252171006053686, + 0.008184346370398998, + 0.007980897091329098, + 0.00966340396553278, + 0.009843808598816395, + 0.011503172107040882, + 0.017346946522593498, + 0.01791827380657196, + 0.006391062866896391, + 0.0062716444954276085, + 0.00796332024037838, + 0.009958184324204922, + 0.010387432761490345, + 0.007145368494093418, + 0.007521381136029959, + 0.007427564822137356, + 0.00824097078293562, + 0.005116437561810017, + 0.008133339695632458, + 0.005294950678944588, + 0.004742825403809547, + 0.006931917276233435, + 0.008861864916980267, + 0.011937672272324562, + 0.008290400728583336, + 0.005017408169806004, + 0.015512925572693348, + 0.006151091773062944, + 0.012510743923485279, + 0.005528981331735849, + 0.004906998947262764, + 0.004430601838976145, + 0.03611765056848526, + 0.0304995309561491, + 0.04866330698132515, + 0.04363580048084259, + 0.06717045605182648, + 0.04487687721848488, + 0.06382845342159271, + 0.04681343585252762, + 0.016763951629400253, + 0.02606147527694702, + 0.029600638896226883, + 0.026429414749145508, + 0.03465721011161804, + 0.022950049489736557, + 0.03547373786568642, + 0.031952522695064545, + 0.002123570768162608, + 0.0018924538744613528, + 0.003083863528445363, + 0.0013921195641160011, + 0.0028584818355739117, + 0.0048901853151619434, + 0.0028719326946884394, + 0.0015435926616191864, + 0.001625441131182015, + 0.001523389364592731, + 0.0019914009608328342, + 0.0016770680667832494, + 0.0016534797614440322, + 0.0023140215780586004, + 0.003346498589962721, + 0.0017112590139731765, + 0.0035959973465651274, + 0.0027468400076031685, + 0.007333952002227306, + 0.0030284347012639046, + 0.0034405270125716925, + 0.013225073926150799, + 0.004716619849205017, + 0.006434341426938772, + 0.00417803879827261, + 0.002712093060836196, + 0.0071364338509738445, + 0.003935457207262516, + 0.0032043082173913717, + 0.006701441016048193, + 0.007055049762129784, + 0.005193945020437241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.16038018465042114 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813171148300171 + ], + "opposite_logprob_tokens": [ + -1.406317114830017 + ], + "logprob_reward": -0.2813171148300171, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813171148300171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12747015058994293, + 0.04151024669408798, + 0.03456681966781616, + 0.06464136391878128, + 0.09466452896595001, + 0.0652535930275917, + 0.27440664172172546, + 0.11358650773763657, + 0.0429520383477211, + 0.14094802737236023 + ], + "credit_map": [ + 0.011563656851649284, + 0.012968833558261395, + 0.023498035967350006, + 0.029065925627946854, + 0.011248065158724785, + 0.011920785531401634, + 0.01252283900976181, + 0.014682010747492313, + 0.004508817568421364, + 0.004010878503322601, + 0.008740667253732681, + 0.004462751559913158, + 0.005644716788083315, + 0.006675302982330322, + 0.0038275900296866894, + 0.00363952131010592, + 0.0036746917758136988, + 0.004067402333021164, + 0.004449833650141954, + 0.005955622997134924, + 0.003005027072504163, + 0.004541635978966951, + 0.004434517119079828, + 0.004438087344169617, + 0.006900492124259472, + 0.007637914270162582, + 0.01099770050495863, + 0.007390228565782309, + 0.00664520775899291, + 0.009801320731639862, + 0.00733454804867506, + 0.007933948189020157, + 0.009716669097542763, + 0.007990849204361439, + 0.013028291054069996, + 0.014061945490539074, + 0.013597209937870502, + 0.010219485498964787, + 0.013950172811746597, + 0.012099911458790302, + 0.01243889331817627, + 0.005798955913633108, + 0.009302942082285881, + 0.0061432295478880405, + 0.005842825397849083, + 0.010599945671856403, + 0.007508212234824896, + 0.007618588395416737, + 0.039816565811634064, + 0.019949940964579582, + 0.05053935572504997, + 0.027128249406814575, + 0.014204803854227066, + 0.07919801771640778, + 0.01918879523873329, + 0.024380935356020927, + 0.015053058974444866, + 0.007559892255812883, + 0.020810557529330254, + 0.011594266630709171, + 0.006209320854395628, + 0.024518195539712906, + 0.009623179212212563, + 0.018218034878373146, + 0.004077498335391283, + 0.002922856016084552, + 0.007445082534104586, + 0.004120859317481518, + 0.0035815387964248657, + 0.00963672436773777, + 0.005058302078396082, + 0.006109176203608513, + 0.0181263517588377, + 0.01080765575170517, + 0.01812608912587166, + 0.01531529612839222, + 0.010349404998123646, + 0.016352036967873573, + 0.03336343914270401, + 0.018507756292819977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.2813171148300171 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335205975919962 + ], + "opposite_logprob_tokens": [ + -6.626335144042969 + ], + "logprob_reward": -0.001335205975919962, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335205975919962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2259792983531952, + 0.09346212446689606, + 0.04202631488442421, + 0.03518911078572273, + 0.2214636355638504, + 0.08886460959911346, + 0.0992090106010437, + 0.05236159637570381, + 0.04003758355975151, + 0.10140663385391235 + ], + "credit_map": [ + 0.01901145465672016, + 0.023209841921925545, + 0.030146513134241104, + 0.04030168056488037, + 0.023087153211236, + 0.032091639935970306, + 0.022907335311174393, + 0.03522367402911186, + 0.011147932149469852, + 0.009173261933028698, + 0.01430515106767416, + 0.011533308774232864, + 0.020794371142983437, + 0.008507343009114265, + 0.009339510463178158, + 0.008661244064569473, + 0.0049347118474543095, + 0.004549930337816477, + 0.005110679194331169, + 0.005209422670304775, + 0.006039158906787634, + 0.005817161872982979, + 0.005495128687471151, + 0.004870121832937002, + 0.004227130673825741, + 0.0032677992712706327, + 0.004894694313406944, + 0.005472096614539623, + 0.0054932269267737865, + 0.002733855973929167, + 0.004846150055527687, + 0.00425415625795722, + 0.02497563324868679, + 0.01713819056749344, + 0.038070473819971085, + 0.03418516740202904, + 0.030627775937318802, + 0.02077168971300125, + 0.030253037810325623, + 0.025441663339734077, + 0.01412236038595438, + 0.00960573460906744, + 0.011804664507508278, + 0.0056364573538303375, + 0.012537533417344093, + 0.013839388266205788, + 0.009618345648050308, + 0.011700130999088287, + 0.013823604211211205, + 0.007498701568692923, + 0.018546003848314285, + 0.008317667990922928, + 0.006036071572452784, + 0.030214985832571983, + 0.007130448240786791, + 0.007641524076461792, + 0.004670301917940378, + 0.0041163344867527485, + 0.007498051971197128, + 0.004522719886153936, + 0.006127160508185625, + 0.008051665499806404, + 0.009406575001776218, + 0.007968787103891373, + 0.004014814272522926, + 0.0021626101806759834, + 0.007890558801591396, + 0.003716125851497054, + 0.002341810381039977, + 0.01123854424804449, + 0.003539034863933921, + 0.005134084261953831, + 0.012373123317956924, + 0.009664040990173817, + 0.01127071212977171, + 0.010206039063632488, + 0.010438154451549053, + 0.013909834437072277, + 0.02338297665119171, + 0.010161753743886948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.001335205975919962 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up slowly pans out onto a stack of clear, nested glass bowls, placed on a dark, polished wooden coffee table. The largest bowl at the bottom is filled with fluffy, buttered popcorn, its irregular golden kernels gleaming in the light. Stacked within this is a series of progressively smaller, empty glass bowls, their clean, curved surfaces reflecting and distorting the scene in miniature. The only illumination comes from the realistic, flickering, cool blue glow of an off-screen television, casting long, dancing shadows across the table and creating ever-shifting highlights on the glass borders and textured popcorn. This single light source, combined with subtle camera movements, creates a dynamic atmosphere of cozy anticipation, perfectly capturing the intimacy and focused energy of an impending movie night. The background is intentionally kept dark and out of focus to draw everyone's attention to the enticing snacks.", + "seed": 14525, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_7/index_6448_seed_14525_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026883113419171423, + "vlm_logprob": -0.026883113419171423, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.026883113419171423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007840217091143131 + ], + "opposite_logprob_tokens": [ + -5.382840156555176 + ], + "logprob_reward": -0.007840217091143131, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.007840217091143131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13763371109962463, + 0.1274566799402237, + 0.05183861777186394, + 0.0806473046541214, + 0.06659253686666489, + 0.1840183585882187, + 0.11008559167385101, + 0.060221511870622635, + 0.12829752266407013, + 0.053208161145448685 + ], + "credit_map": [ + 0.022432705387473106, + 0.015681393444538116, + 0.01846981607377529, + 0.01771213300526142, + 0.01108997780829668, + 0.01946971006691456, + 0.011516944505274296, + 0.021261023357510567, + 0.016793709248304367, + 0.01783379353582859, + 0.018365567550063133, + 0.021757271140813828, + 0.009905549697577953, + 0.01034074742347002, + 0.01825697161257267, + 0.014203062281012535, + 0.005100794602185488, + 0.004343807697296143, + 0.009239139035344124, + 0.014687225222587585, + 0.003769796574488282, + 0.0037143228109925985, + 0.004002976231276989, + 0.006980556063354015, + 0.013041699305176735, + 0.012304035946726799, + 0.010865822434425354, + 0.011352099478244781, + 0.011620490811765194, + 0.005886491853743792, + 0.009724188596010208, + 0.005852478556334972, + 0.007028898224234581, + 0.007725407835096121, + 0.006601349916309118, + 0.007427306845784187, + 0.011373431421816349, + 0.0085314204916358, + 0.009989002719521523, + 0.007915723137557507, + 0.014593728817999363, + 0.015017488971352577, + 0.059950511902570724, + 0.02926904335618019, + 0.008989222347736359, + 0.0070925126783549786, + 0.01414454448968172, + 0.034961309283971786, + 0.009531001560389996, + 0.02268344536423683, + 0.013244976289570332, + 0.02921590395271778, + 0.008057210594415665, + 0.006528015714138746, + 0.010583683848381042, + 0.01024135947227478, + 0.007286432199180126, + 0.0072318739257752895, + 0.0071694375947117805, + 0.008066173642873764, + 0.005535769276320934, + 0.00851009227335453, + 0.005549927707761526, + 0.010871805250644684, + 0.015600345097482204, + 0.009771310724318027, + 0.01360282115638256, + 0.017895594239234924, + 0.012118013575673103, + 0.017929963767528534, + 0.016515664756298065, + 0.02486380748450756, + 0.005038257222622633, + 0.0052652740851044655, + 0.006803648546338081, + 0.005879672709852457, + 0.010333743877708912, + 0.005063511896878481, + 0.009035778231918812, + 0.0057882750406861305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.007840217091143131 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561016716063023 + ], + "opposite_logprob_tokens": [ + -4.387560844421387 + ], + "logprob_reward": -0.012561016716063023, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561016716063023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18261532485485077, + 0.21873287856578827, + 0.11516518890857697, + 0.01991642452776432, + 0.06343860924243927, + 0.03983982279896736, + 0.04941520839929581, + 0.03722308576107025, + 0.2409837543964386, + 0.03266968950629234 + ], + "credit_map": [ + 0.026090729981660843, + 0.026501424610614777, + 0.02482711523771286, + 0.03152063116431236, + 0.012621712870895863, + 0.017235901206731796, + 0.013288086280226707, + 0.030529731884598732, + 0.04570535570383072, + 0.03197347745299339, + 0.02811076119542122, + 0.043728336691856384, + 0.013953998684883118, + 0.011328521184623241, + 0.02527793124318123, + 0.018654482439160347, + 0.019322119653224945, + 0.026627514511346817, + 0.011965682730078697, + 0.016474636271595955, + 0.015732798725366592, + 0.008264319971203804, + 0.008695032447576523, + 0.008083083666861057, + 0.0023603562731295824, + 0.0032260827720165253, + 0.0032181052956730127, + 0.003527042455971241, + 0.002125700004398823, + 0.0010234765941277146, + 0.0027797501534223557, + 0.0016559096984565258, + 0.0080970898270607, + 0.006586312782019377, + 0.008151018992066383, + 0.0080402921885252, + 0.00824157614260912, + 0.005543955136090517, + 0.00856900867074728, + 0.010209353640675545, + 0.003521831938996911, + 0.00518420897424221, + 0.007948354817926884, + 0.006641086656600237, + 0.00340887694619596, + 0.0028185690753161907, + 0.00487836729735136, + 0.005438526626676321, + 0.004978113807737827, + 0.01071134489029646, + 0.006500001531094313, + 0.008945118635892868, + 0.004444281104952097, + 0.004144946578890085, + 0.004867475014179945, + 0.0048239273019135, + 0.005695154424756765, + 0.0039618415758013725, + 0.005798012018203735, + 0.005395022686570883, + 0.002653982024639845, + 0.003774333745241165, + 0.003702920163050294, + 0.006241816561669111, + 0.02848336286842823, + 0.02615218237042427, + 0.031152881681919098, + 0.023741813376545906, + 0.03941396623849869, + 0.02268846333026886, + 0.02849001996219158, + 0.04086107388138771, + 0.002766399411484599, + 0.003188733709976077, + 0.003906656987965107, + 0.0036107490304857492, + 0.007072398904711008, + 0.003126550931483507, + 0.005609765183180571, + 0.0033884355798363686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.012561016716063023 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003770768642426 + ], + "opposite_logprob_tokens": [ + -2.695037603378296 + ], + "logprob_reward": -0.07003770768642426, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003770768642426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1810835301876068, + 0.11038472503423691, + 0.03825577348470688, + 0.04227352514863014, + 0.08995238691568375, + 0.2634582221508026, + 0.095076784491539, + 0.07176072895526886, + 0.08558309823274612, + 0.022171223536133766 + ], + "credit_map": [ + 0.019210919737815857, + 0.024003181606531143, + 0.01959497667849064, + 0.030312273651361465, + 0.018541907891631126, + 0.02606828138232231, + 0.018772384151816368, + 0.02457961067557335, + 0.014756212010979652, + 0.010552060790359974, + 0.02866864763200283, + 0.013542376458644867, + 0.00838417001068592, + 0.008696234785020351, + 0.009871560148894787, + 0.0159134604036808, + 0.004602779168635607, + 0.004864869639277458, + 0.004598035477101803, + 0.005113354418426752, + 0.005048531573265791, + 0.004695490468293428, + 0.005367823410779238, + 0.003964888863265514, + 0.006229361053556204, + 0.005868078675121069, + 0.0043844315223395824, + 0.00571284256875515, + 0.006289665587246418, + 0.0036703837104141712, + 0.005398530047386885, + 0.004720229189842939, + 0.007281510159373283, + 0.010367565788328648, + 0.00924751628190279, + 0.010616128332912922, + 0.010573547333478928, + 0.013011466711759567, + 0.018075289204716682, + 0.010779366828501225, + 0.020517613738775253, + 0.011216101236641407, + 0.10251762717962265, + 0.037954431027173996, + 0.01010102964937687, + 0.010947954840958118, + 0.011449266225099564, + 0.05875419080257416, + 0.012675357051193714, + 0.013860676437616348, + 0.013142049312591553, + 0.014016998000442982, + 0.009318266063928604, + 0.011907260864973068, + 0.007598516996949911, + 0.012557651847600937, + 0.011589085683226585, + 0.004724119324237108, + 0.009645629674196243, + 0.008782347664237022, + 0.005967515986412764, + 0.013120646588504314, + 0.005783972796052694, + 0.012147408910095692, + 0.010088171809911728, + 0.009812511503696442, + 0.01049228198826313, + 0.00890524685382843, + 0.01353537105023861, + 0.009232665412127972, + 0.009940369054675102, + 0.013576480560004711, + 0.002394328825175762, + 0.0028766803443431854, + 0.0029852630104869604, + 0.0020168491173535585, + 0.003641796298325062, + 0.00257217139005661, + 0.0030821161344647408, + 0.002602019812911749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.07003770768642426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305734857916832 + ], + "opposite_logprob_tokens": [ + -3.168057441711426 + ], + "logprob_reward": -0.04305734857916832, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305734857916832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0671243816614151, + 0.08033279329538345, + 0.06213553622364998, + 0.04882023483514786, + 0.07280726730823517, + 0.0455334298312664, + 0.10673202574253082, + 0.11984561383724213, + 0.3101102113723755, + 0.08655858784914017 + ], + "credit_map": [ + 0.00841822475194931, + 0.006642650347203016, + 0.008734127506613731, + 0.01463822741061449, + 0.005074868444353342, + 0.007260608021169901, + 0.007203570567071438, + 0.00915210135281086, + 0.013656835071742535, + 0.008786209858953953, + 0.013353352434933186, + 0.012552210129797459, + 0.0066224089823663235, + 0.007382445968687534, + 0.007417719811201096, + 0.010561599396169186, + 0.0062681338749825954, + 0.005739912856370211, + 0.008588731288909912, + 0.015547258779406548, + 0.00558479642495513, + 0.007010715082287788, + 0.006832002196460962, + 0.006563973147422075, + 0.008239777758717537, + 0.006464625708758831, + 0.005857148207724094, + 0.007584474049508572, + 0.006327827461063862, + 0.0029265466146171093, + 0.005492419935762882, + 0.0059274123050272465, + 0.006618782877922058, + 0.007970183156430721, + 0.00881438609212637, + 0.007785334251821041, + 0.013653119094669819, + 0.007365211844444275, + 0.009964393451809883, + 0.010635853745043278, + 0.0051317373290658, + 0.006435713265091181, + 0.006392079405486584, + 0.008649401366710663, + 0.004869929980486631, + 0.0038344019558280706, + 0.005782731808722019, + 0.004437430761754513, + 0.009277930483222008, + 0.024152852594852448, + 0.01232620608061552, + 0.02736477553844452, + 0.008918825536966324, + 0.00447652256116271, + 0.012076123617589474, + 0.00813878420740366, + 0.01790802553296089, + 0.01512007974088192, + 0.020313307642936707, + 0.017851902171969414, + 0.012748927809298038, + 0.010659153573215008, + 0.010885469615459442, + 0.014358730986714363, + 0.0340183787047863, + 0.0268507432192564, + 0.034217674285173416, + 0.04613788425922394, + 0.025382012128829956, + 0.04262816160917282, + 0.04248284921050072, + 0.05839249864220619, + 0.014948368072509766, + 0.013182197697460651, + 0.010924947448074818, + 0.012640457600355148, + 0.0069519104436039925, + 0.010470743291079998, + 0.005979860667139292, + 0.01146009098738432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.04305734857916832 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle close-up of a stack of clear, nested glass bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the largest bottom bowl filled with fluffy, buttered popcorn with irregular golden kernels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a realistic, flickering, cool blue glow from an off-screen television?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan out from the initial close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and out of focus to draw attention to the bowls and popcorn?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009192770230583847 + ], + "opposite_logprob_tokens": [ + -7.000919342041016 + ], + "logprob_reward": -0.0009192770230583847, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009192770230583847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12636864185333252, + 0.14408551156520844, + 0.06562434881925583, + 0.04682434722781181, + 0.18673482537269592, + 0.07251778244972229, + 0.1104583740234375, + 0.05688682943582535, + 0.1576504111289978, + 0.032848894596099854 + ], + "credit_map": [ + 0.012720976956188679, + 0.01379104983061552, + 0.015543659217655659, + 0.02069925144314766, + 0.014481606893241405, + 0.018192315474152565, + 0.01569106988608837, + 0.01524871215224266, + 0.019225597381591797, + 0.010448519140481949, + 0.04213244467973709, + 0.018211128190159798, + 0.008531250059604645, + 0.010732254013419151, + 0.011755506508052349, + 0.023048831149935722, + 0.005708408076316118, + 0.0077679771929979324, + 0.00916721485555172, + 0.01120650302618742, + 0.010417697951197624, + 0.00672199996188283, + 0.008115566335618496, + 0.006518983282148838, + 0.005705389194190502, + 0.005153119098395109, + 0.005122266244143248, + 0.0073226457461714745, + 0.005695028230547905, + 0.006283150985836983, + 0.005540919955819845, + 0.006001831963658333, + 0.0168868787586689, + 0.017971832305192947, + 0.023932067677378654, + 0.022138774394989014, + 0.01595960557460785, + 0.028234602883458138, + 0.02474663220345974, + 0.03686446323990822, + 0.005910870153456926, + 0.004192953463643789, + 0.023730065673589706, + 0.010016251355409622, + 0.004208340309560299, + 0.0045393286272883415, + 0.0051921964623034, + 0.014727786183357239, + 0.00965138804167509, + 0.02456016093492508, + 0.012822043150663376, + 0.025108490139245987, + 0.0084357475861907, + 0.008610479533672333, + 0.011717512272298336, + 0.009552564471960068, + 0.008139818906784058, + 0.0062846895307302475, + 0.00791157130151987, + 0.007394644431769848, + 0.007543017156422138, + 0.0061814808286726475, + 0.0062439278699457645, + 0.007187683135271072, + 0.017982566729187965, + 0.0184435173869133, + 0.019477391615509987, + 0.018100280314683914, + 0.023007988929748535, + 0.017125379294157028, + 0.021261373534798622, + 0.022251926362514496, + 0.0040366752073168755, + 0.005909458734095097, + 0.0043586562387645245, + 0.0049840291030704975, + 0.003232671646401286, + 0.003116603009402752, + 0.003696476574987173, + 0.0035143280401825905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.0009192770230583847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_8/index_tensor([2970, 2970])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_8/index_tensor([2970, 2970])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ea3b98f98a19d08e2b384c64ac7c6e19c7d26eff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_8/index_tensor([2970, 2970])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking close-up tracking shot, held at a slightly upward angle, locks onto the \"head\" of a massive Tyrannosaurus rex. The entire image is imbued with Hirohiko Araki's signature aesthetic: sharp, hard lines and dazzling, saturated colors. The mutant plant's skin, instead of its natural green, is an eerie blend of acid green and electric purple, covered in sculpted geometric angles and deep grooves. Knife-sharp magenta spikes jut from its ridges. As the camera steadily tracks, a bud-like structure at the plant's apex begins to pulsate violently and rhythmically, emanating an ominous golden glow from within. Suddenly, it violently explodes outward, opening into a terrifying, jagged \"maw\" and emitting a silent yet visually shattering roar. A visible circular shock wave composed of pure energy gushed out from its core. The shock wave was translucent cyan, with unstable white arcs flashing at the edges. It expanded outward at an astonishing speed, distorting and refracting the air wherever it passed, instantly blurring the equally colorful surreal sky and earth in the background, creating an atmosphere full of power, extreme tension, and grotesque splendor.", + "seed": 17711, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_8/index_2970_seed_17711_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.014465975202620029, + "vlm_logprob": -0.014465975202620029, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.014465975202620029, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008691104128956795 + ], + "opposite_logprob_tokens": [ + -5.008691310882568 + ], + "logprob_reward": -0.008691104128956795, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008691104128956795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010599545203149319, + 0.006248687393963337, + 0.0166227538138628, + 0.007591304834932089, + 0.007100547663867474, + 0.00972647126764059, + 0.9362188577651978, + 0.002028942806646228, + 0.0021620909683406353, + 0.001700839027762413 + ], + "credit_map": [ + 0.0013052059803158045, + 0.001006870879791677, + 0.0009064217447303236, + 0.0013476043241098523, + 0.0018571425462141633, + 0.001682882197201252, + 0.0010741441510617733, + 0.0014192722737789154, + 0.0008252075640484691, + 0.00050435058074072, + 0.0005763312801718712, + 0.0005637538270093501, + 0.0006796331726945937, + 0.0010234230430796742, + 0.0005409642471931875, + 0.0015350234461948276, + 0.002671619178727269, + 0.0012907549971714616, + 0.0016619795933365822, + 0.0013486293610185385, + 0.002607084345072508, + 0.0022685707081109285, + 0.002365813823416829, + 0.0024083002936095, + 0.0009188430849462748, + 0.0010204805294051766, + 0.0006666500121355057, + 0.0006233021849766374, + 0.0013100227806717157, + 0.0010202793637290597, + 0.0011436091735959053, + 0.0008881170069798827, + 0.0015032439259812236, + 0.0009006888139992952, + 0.0006980982725508511, + 0.0006121320766396821, + 0.000728876912035048, + 0.0008982259896583855, + 0.0008196699200198054, + 0.0009396111709065735, + 0.0015741420211270452, + 0.0010465944651514292, + 0.0015804246068000793, + 0.0007236211095005274, + 0.0008948796312324703, + 0.0010297655826434493, + 0.0008728115353733301, + 0.0020042327232658863, + 0.4069305956363678, + 0.03337451443076134, + 0.39788323640823364, + 0.0276882853358984, + 0.020586587488651276, + 0.018584350124001503, + 0.01688394322991371, + 0.014287314377725124, + 0.0004513606836553663, + 0.00030007559689693153, + 0.00026997350505553186, + 0.0002063446881948039, + 0.00020861238590441644, + 0.000152798427734524, + 0.00024242501240223646, + 0.00019735225941985846, + 0.0003284357371740043, + 0.00045962087460793555, + 0.0002923356951214373, + 0.00023735762806609273, + 0.00017638588906265795, + 0.00019237626111134887, + 0.0002741791831795126, + 0.0002013995690504089, + 0.0002772658772300929, + 0.00032066256972029805, + 0.0002247372904093936, + 0.00019015565339941531, + 0.00014455824566539377, + 0.00017363006190862507, + 0.00018252017616759986, + 0.00018730912415776402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.008691104128956795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004107254557311535 + ], + "opposite_logprob_tokens": [ + -5.504107475280762 + ], + "logprob_reward": -0.004107254557311535, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004107254557311535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011044134385883808, + 0.00568972947075963, + 0.010023522190749645, + 0.019728660583496094, + 0.018223755061626434, + 0.00610017403960228, + 0.9170448184013367, + 0.0048834532499313354, + 0.004265538416802883, + 0.002996231894940138 + ], + "credit_map": [ + 0.0010930090211331844, + 0.0010904542868956923, + 0.0008989240741357207, + 0.0009754009079188108, + 0.0016972864978015423, + 0.0018247062107548118, + 0.0011066480074077845, + 0.002357705030590296, + 0.0006600117776542902, + 0.000861606327816844, + 0.0005907372105866671, + 0.0007524676620960236, + 0.0005440078093670309, + 0.0007633562781848013, + 0.0004948726273141801, + 0.0010226693702861667, + 0.0018347251461818814, + 0.000735822191927582, + 0.0009445360046811402, + 0.000689889770001173, + 0.0014884232077747583, + 0.0014902137918397784, + 0.001307354774326086, + 0.0015325576532632113, + 0.0029200187418609858, + 0.00400402769446373, + 0.002236755099147558, + 0.002104158280417323, + 0.002752886386588216, + 0.0025718563701957464, + 0.00184546853415668, + 0.001293489127419889, + 0.003019957570359111, + 0.003809985239058733, + 0.0017833791207522154, + 0.0014705921057611704, + 0.001965656178072095, + 0.0020562130957841873, + 0.0019962675869464874, + 0.0021217039320617914, + 0.0008271417464129627, + 0.0006246984703466296, + 0.0009225946851074696, + 0.0004654577060136944, + 0.0009353671921417117, + 0.000576139020267874, + 0.0009953113039955497, + 0.0007534638862125576, + 0.39876917004585266, + 0.0327720120549202, + 0.39004749059677124, + 0.027174726128578186, + 0.01999507285654545, + 0.018028561025857925, + 0.016443980857729912, + 0.013813858851790428, + 0.0012951560784131289, + 0.000649356166832149, + 0.0006722731632180512, + 0.0005249793757684529, + 0.00041470909491181374, + 0.00041386348311789334, + 0.0004725440521724522, + 0.00044057201012037694, + 0.0007743163732811809, + 0.0009433124796487391, + 0.0006038404535502195, + 0.0004514180473051965, + 0.0002900288673117757, + 0.00035352137638255954, + 0.0005255901487544179, + 0.00032351058325730264, + 0.0006755889626219869, + 0.0007166448049247265, + 0.00043015569099225104, + 0.0002378718345426023, + 0.00023063572007231414, + 0.00019424495985731483, + 0.00032757973531261086, + 0.0001835103757912293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.004107254557311535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606437936425209 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.01606437936425209, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606437936425209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024169858545064926, + 0.007924250327050686, + 0.010311613790690899, + 0.017001088708639145, + 0.01553741842508316, + 0.031547002494335175, + 0.8808220624923706, + 0.004968302324414253, + 0.0037868854124099016, + 0.003931475803256035 + ], + "credit_map": [ + 0.002293142955750227, + 0.0017198559362441301, + 0.001121973735280335, + 0.0018662188667804003, + 0.007513428572565317, + 0.0056384471245110035, + 0.001599423703737557, + 0.002417367184534669, + 0.001053853309713304, + 0.0008567468612454832, + 0.0009408333571627736, + 0.0010860568145290017, + 0.0009373491047881544, + 0.0009957989677786827, + 0.0009173043654300272, + 0.001136306906118989, + 0.0013336340198293328, + 0.0008877591462805867, + 0.0010101328371092677, + 0.0009665123652666807, + 0.0015888758935034275, + 0.0013221022672951221, + 0.001569295534864068, + 0.001633300562389195, + 0.0018458025297150016, + 0.00185479037463665, + 0.001543670892715454, + 0.0013512232108041644, + 0.003061925061047077, + 0.002134302631020546, + 0.0036465309094637632, + 0.0015628418186679482, + 0.0022107716649770737, + 0.002522388705983758, + 0.001609194208867848, + 0.0015586448134854436, + 0.0018596663139760494, + 0.0018475042888894677, + 0.002036321209743619, + 0.0018929283833131194, + 0.004593233577907085, + 0.003096339525654912, + 0.004982964135706425, + 0.002025448717176914, + 0.00409400649368763, + 0.003646194702014327, + 0.003998373169451952, + 0.0051104421727359295, + 0.38305994868278503, + 0.03148118034005165, + 0.37426936626434326, + 0.026112891733646393, + 0.019218899309635162, + 0.017402295023202896, + 0.015871519222855568, + 0.013406015001237392, + 0.0010216115042567253, + 0.0007729814387857914, + 0.0007258123368956149, + 0.0004892683355137706, + 0.0005665495409630239, + 0.000421784381614998, + 0.0005555549287237227, + 0.0004147398576606065, + 0.0004629505565389991, + 0.0006643745000474155, + 0.00046211283188313246, + 0.00048523617442697287, + 0.0003616517933551222, + 0.0003206704568583518, + 0.0006763244746252894, + 0.00035356456646695733, + 0.0007738308049738407, + 0.000849151227157563, + 0.0004615506622940302, + 0.0003900330630131066, + 0.0003631633589975536, + 0.00030599028104916215, + 0.0003482699685264379, + 0.0004394861462060362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.01606437936425209 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779755026102066 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779755026102066, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779755026102066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022275546565651894, + 0.0052763125859200954, + 0.01024580281227827, + 0.007743644528090954, + 0.014914213679730892, + 0.011127928271889687, + 0.8681264519691467, + 0.015323915518820286, + 0.015712518244981766, + 0.029253657907247543 + ], + "credit_map": [ + 0.0019903937354683876, + 0.0013349727960303426, + 0.0015208728145807981, + 0.0013519596541300416, + 0.0038928755093365908, + 0.004449130967259407, + 0.0017101247794926167, + 0.006025218404829502, + 0.0007410800317302346, + 0.0006723428959958255, + 0.0005762429791502655, + 0.0005653043044731021, + 0.0005646583158522844, + 0.0006424114690162241, + 0.0006169301923364401, + 0.0008973433286882937, + 0.0008554858504794538, + 0.0009441323927603662, + 0.0010259823175147176, + 0.0010719492565840483, + 0.001735706813633442, + 0.0013571851886808872, + 0.0017867331625893712, + 0.0014686293434351683, + 0.001459938008338213, + 0.0007320812437683344, + 0.0007997936336323619, + 0.0006698995130136609, + 0.001306536141782999, + 0.0011656126007437706, + 0.000947632419411093, + 0.0006621521315537393, + 0.0024707999546080828, + 0.001591988024301827, + 0.0009944667108356953, + 0.0015387050807476044, + 0.001252490095794201, + 0.0024424600414931774, + 0.0017925163265317678, + 0.002830790588632226, + 0.00179011351428926, + 0.0009186493116430938, + 0.002708080690354109, + 0.0006173621513880789, + 0.0013211016776040196, + 0.0009741916437633336, + 0.0013810554519295692, + 0.0014173752861097455, + 0.37722811102867126, + 0.03090664930641651, + 0.36917784810066223, + 0.025714771822094917, + 0.019103912636637688, + 0.017264142632484436, + 0.01554163359105587, + 0.01318943127989769, + 0.004517194349318743, + 0.002416884293779731, + 0.0023104280699044466, + 0.0014774035662412643, + 0.001140523818321526, + 0.0009599864133633673, + 0.00129497644957155, + 0.0012065217597410083, + 0.0030942391604185104, + 0.003238447941839695, + 0.0022151898592710495, + 0.0018771719187498093, + 0.0013188960729166865, + 0.001086618984118104, + 0.001687048701569438, + 0.001194908283650875, + 0.007212869822978973, + 0.00669736647978425, + 0.004238899331539869, + 0.0026515380013734102, + 0.001875378075055778, + 0.0024128940422087908, + 0.0020795653108507395, + 0.002085147425532341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.009779755026102066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368738293647766 + ], + "opposite_logprob_tokens": [ + -3.4086873531341553 + ], + "logprob_reward": -0.03368738293647766, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368738293647766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04495900869369507, + 0.010230491869151592, + 0.00894829723984003, + 0.006324945017695427, + 0.013571703806519508, + 0.009869788773357868, + 0.8904852867126465, + 0.004807175602763891, + 0.005867385305464268, + 0.004935939330607653 + ], + "credit_map": [ + 0.0032226338516920805, + 0.002856477862223983, + 0.0020600156858563423, + 0.0030231287237256765, + 0.015516864135861397, + 0.011508317664265633, + 0.0030526926275342703, + 0.0037188755813986063, + 0.0009721025708131492, + 0.0011796883773058653, + 0.001624089083634317, + 0.0012555633438751101, + 0.0013549710856750607, + 0.001223599654622376, + 0.0009031205554492772, + 0.0017173561500385404, + 0.002390265464782715, + 0.0012424214510247111, + 0.0008763992809690535, + 0.0006594702135771513, + 0.0009049827349372208, + 0.001147998496890068, + 0.000818168162368238, + 0.0009085905039682984, + 0.0007187052979134023, + 0.0007181463879533112, + 0.0007670619525015354, + 0.0005701061454601586, + 0.0010798961156979203, + 0.0008422677638009191, + 0.0008916856022551656, + 0.0007370751118287444, + 0.0027503652963787317, + 0.0018053126987069845, + 0.0010584621923044324, + 0.0008853707695379853, + 0.0012175990268588066, + 0.0018831840716302395, + 0.0011776128085330129, + 0.0027937961276620626, + 0.0014703708002343774, + 0.0009625231032259762, + 0.0017557275714352727, + 0.0006063809269107878, + 0.0013798531144857407, + 0.000996043672785163, + 0.0013676349772140384, + 0.0013312530936673284, + 0.3866557776927948, + 0.03181580826640129, + 0.37851226329803467, + 0.02636018581688404, + 0.019650043919682503, + 0.01773313619196415, + 0.016119465231895447, + 0.013638554140925407, + 0.000969262735452503, + 0.0007656445959582925, + 0.0006409056368283927, + 0.0004299969004932791, + 0.0005226414068602026, + 0.0003622989170253277, + 0.0006884682225063443, + 0.00042795672197826207, + 0.0010171463945880532, + 0.0012233472662046552, + 0.0007460786146111786, + 0.0006788627943024039, + 0.0004637225647456944, + 0.00044197728857398033, + 0.0007644236320629716, + 0.0005318261682987213, + 0.0008405608241446316, + 0.0009018997079692781, + 0.00064309366280213, + 0.00038365251384675503, + 0.00046385478344745934, + 0.0005154117825441062, + 0.0006751660257577896, + 0.0005122995935380459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.03368738293647766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up tracking shot, held at a slightly upward angle, locks onto the \"head\" of a massive Tyrannosaurus rex. The entire image is imbued with Hirohiko Araki's signature aesthetic: sharp, hard lines and dazzling, saturated colors. The mutant plant's skin, instead of its natural green, is an eerie blend of acid green and electric purple, covered in sculpted geometric angles and deep grooves. Knife-sharp magenta spikes jut from its ridges. As the camera steadily tracks, a bud-like structure at the plant's apex begins to pulsate violently and rhythmically, emanating an ominous golden glow from within. Suddenly, it violently explodes outward, opening into a terrifying, jagged \"maw\" and emitting a silent yet visually shattering roar. A visible circular shock wave composed of pure energy gushed out from its core. The shock wave was translucent cyan, with unstable white arcs flashing at the edges. It expanded outward at an astonishing speed, distorting and refracting the air wherever it passed, instantly blurring the equally colorful surreal sky and earth in the background, creating an atmosphere full of power, extreme tension, and grotesque splendor.", + "seed": 17712, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_8/index_2970_seed_17712_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01745761288329959, + "vlm_logprob": -0.01745761288329959, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.01745761288329959, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010840100236237049 + ], + "opposite_logprob_tokens": [ + -4.885839939117432 + ], + "logprob_reward": -0.010840100236237049, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010840100236237049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09063755720853806, + 0.06724261492490768, + 0.04950455203652382, + 0.0422210767865181, + 0.05470860376954079, + 0.06909798085689545, + 0.11150099337100983, + 0.2640853822231293, + 0.17019084095954895, + 0.08081039041280746 + ], + "credit_map": [ + 0.009205454029142857, + 0.008660638704895973, + 0.006455786991864443, + 0.008269436657428741, + 0.008799521252512932, + 0.013066799379885197, + 0.0213530994951725, + 0.014826817438006401, + 0.007474708836525679, + 0.007332219276577234, + 0.007258010096848011, + 0.007364280521869659, + 0.006641752552241087, + 0.01060516107827425, + 0.009963331744074821, + 0.010603150352835655, + 0.009050910361111164, + 0.004584067966789007, + 0.00496922992169857, + 0.003988989163190126, + 0.004058977589011192, + 0.005454085301607847, + 0.009061616845428944, + 0.00833667442202568, + 0.006167669780552387, + 0.0035169317852705717, + 0.0037623276002705097, + 0.004265793599188328, + 0.003690983634442091, + 0.006022599991410971, + 0.008209586143493652, + 0.0065851821564137936, + 0.007676054257899523, + 0.004858490079641342, + 0.005199795588850975, + 0.006379546131938696, + 0.005584106780588627, + 0.008746711537241936, + 0.005241697188466787, + 0.011022204533219337, + 0.009530819021165371, + 0.006600998342037201, + 0.009863892570137978, + 0.007537251804023981, + 0.008494372479617596, + 0.006707214284688234, + 0.012773540802299976, + 0.007589892949908972, + 0.01477783266454935, + 0.01990940421819687, + 0.01164582371711731, + 0.010178587399423122, + 0.016069183126091957, + 0.011716863140463829, + 0.009464576840400696, + 0.017738718539476395, + 0.03719498589634895, + 0.03293181583285332, + 0.028596241027116776, + 0.02721838466823101, + 0.03703036159276962, + 0.03411538153886795, + 0.03814975172281265, + 0.028848446905612946, + 0.023341672495007515, + 0.020219793543219566, + 0.01335817202925682, + 0.013356652110815048, + 0.03353768214583397, + 0.020890826359391212, + 0.027221737429499626, + 0.018264302983880043, + 0.011932030320167542, + 0.00937385018914938, + 0.014132125303149223, + 0.006582486908882856, + 0.009024089202284813, + 0.008541368879377842, + 0.012530487030744553, + 0.008693955838680267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.010840100236237049 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003210867755115032 + ], + "opposite_logprob_tokens": [ + -5.75321102142334 + ], + "logprob_reward": -0.003210867755115032, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003210867755115032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13810840249061584, + 0.08246178925037384, + 0.078407421708107, + 0.07706237584352493, + 0.12443310767412186, + 0.048821087926626205, + 0.07491553574800491, + 0.15821106731891632, + 0.1049816682934761, + 0.11259754002094269 + ], + "credit_map": [ + 0.013825531117618084, + 0.013457776978611946, + 0.011833013966679573, + 0.011318904347717762, + 0.011494187638163567, + 0.02530168928205967, + 0.021615423262119293, + 0.02926190197467804, + 0.00844424869865179, + 0.008872557431459427, + 0.008108110167086124, + 0.008652567863464355, + 0.007375879213213921, + 0.011671720072627068, + 0.01135820709168911, + 0.017978504300117493, + 0.007857000455260277, + 0.009305370040237904, + 0.007145450916141272, + 0.007735929451882839, + 0.007739534601569176, + 0.008407840505242348, + 0.017996279522776604, + 0.012220022268593311, + 0.008603690192103386, + 0.007989516481757164, + 0.0075164008885622025, + 0.007659888826310635, + 0.007035945076495409, + 0.010571986436843872, + 0.017230365425348282, + 0.010454579256474972, + 0.011474927887320518, + 0.014312892220914364, + 0.011183531954884529, + 0.013724221847951412, + 0.020186563953757286, + 0.02308976650238037, + 0.015108861960470676, + 0.015352347865700722, + 0.005728581920266151, + 0.004514792934060097, + 0.005770625080913305, + 0.008409637026488781, + 0.005749623756855726, + 0.004772994667291641, + 0.009311742149293423, + 0.004563090857118368, + 0.00951048731803894, + 0.010126408189535141, + 0.00678975461050868, + 0.009068481624126434, + 0.012529497034847736, + 0.00864191260188818, + 0.007209179922938347, + 0.011039819568395615, + 0.019761240109801292, + 0.019022751599550247, + 0.01809503138065338, + 0.01748429425060749, + 0.02282826416194439, + 0.020307078957557678, + 0.022772273048758507, + 0.01794014312326908, + 0.007921571843326092, + 0.012584183365106583, + 0.008086967281997204, + 0.007025017403066158, + 0.020517805591225624, + 0.011594222858548164, + 0.02625262551009655, + 0.010999282822012901, + 0.01534369308501482, + 0.012320077046751976, + 0.015966160222887993, + 0.00909511186182499, + 0.013445807620882988, + 0.011486195959150791, + 0.020880533382296562, + 0.01405996922403574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.003210867755115032 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195887371897697 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195887371897697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195887371897697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10673888027667999, + 0.0651417002081871, + 0.041960086673498154, + 0.05438769608736038, + 0.039064984768629074, + 0.0709957554936409, + 0.04825442656874657, + 0.15896445512771606, + 0.21210740506649017, + 0.2023846060037613 + ], + "credit_map": [ + 0.007279229816049337, + 0.012519143521785736, + 0.007038720417767763, + 0.011315925978124142, + 0.008571775630116463, + 0.013904350809752941, + 0.028133802115917206, + 0.017975937575101852, + 0.005583708640187979, + 0.006459069438278675, + 0.006518505048006773, + 0.005323640536516905, + 0.006896946579217911, + 0.009922860190272331, + 0.009786379523575306, + 0.014650598168373108, + 0.004672853741794825, + 0.004365318454802036, + 0.004427551757544279, + 0.005263967905193567, + 0.0038853378500789404, + 0.003554797498509288, + 0.008951393887400627, + 0.006838869769126177, + 0.0055454629473388195, + 0.004637329839169979, + 0.005281672812998295, + 0.007009550929069519, + 0.005546479020267725, + 0.004538798704743385, + 0.014437279663980007, + 0.007391128223389387, + 0.006263194605708122, + 0.003655360545963049, + 0.0036833372432738543, + 0.004335951060056686, + 0.004638586193323135, + 0.006732010282576084, + 0.004021421540528536, + 0.005735128186643124, + 0.008951670490205288, + 0.008125865831971169, + 0.00980469398200512, + 0.00745085533708334, + 0.011125963181257248, + 0.007809005212038755, + 0.009672549553215504, + 0.008055157028138638, + 0.006539905909448862, + 0.0069893961772322655, + 0.005145584233105183, + 0.004585383925586939, + 0.007349872961640358, + 0.004659629426896572, + 0.006146048195660114, + 0.006838610861450434, + 0.020534273236989975, + 0.012845627963542938, + 0.020738648250699043, + 0.014847339130938053, + 0.018341386690735817, + 0.020374389365315437, + 0.026783233508467674, + 0.02449955977499485, + 0.01813809759914875, + 0.029039163142442703, + 0.017325986176729202, + 0.016531316563487053, + 0.04490291699767113, + 0.022753452882170677, + 0.041943952441215515, + 0.021472537890076637, + 0.028999874368309975, + 0.02406558394432068, + 0.026422761380672455, + 0.016098488122224808, + 0.02945183776319027, + 0.020356835797429085, + 0.03742838278412819, + 0.019560862332582474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.014195887371897697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095263313502073 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.004095263313502073, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095263313502073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06296319514513016, + 0.04550439491868019, + 0.029341334477066994, + 0.051045458763837814, + 0.045980054885149, + 0.05984766408801079, + 0.06641832739114761, + 0.21458032727241516, + 0.16176345944404602, + 0.2625557780265808 + ], + "credit_map": [ + 0.005670018028467894, + 0.009521872736513615, + 0.005803843028843403, + 0.009230929426848888, + 0.007545230444520712, + 0.006451541092246771, + 0.008796312846243382, + 0.009943450801074505, + 0.004691441543400288, + 0.004277809523046017, + 0.004319362808018923, + 0.0035709047224372625, + 0.004064788110554218, + 0.008366403169929981, + 0.004125699400901794, + 0.012087992392480373, + 0.002700568176805973, + 0.0033956882543861866, + 0.0031938792672008276, + 0.0036483369767665863, + 0.003073644358664751, + 0.0033195135183632374, + 0.005138692446053028, + 0.004871014505624771, + 0.006451790686696768, + 0.00438734283670783, + 0.004746382124722004, + 0.005514658056199551, + 0.004714497365057468, + 0.006063963286578655, + 0.011672901920974255, + 0.007493922486901283, + 0.004312761593610048, + 0.005509585607796907, + 0.004717545118182898, + 0.005604696460068226, + 0.006204006262123585, + 0.006992686074227095, + 0.006539345718920231, + 0.006099429912865162, + 0.007795960642397404, + 0.006214027758687735, + 0.009455092251300812, + 0.005831556860357523, + 0.00875020306557417, + 0.005805950611829758, + 0.009405141696333885, + 0.0065897381864488125, + 0.009165219962596893, + 0.00938487984240055, + 0.006651670206338167, + 0.0069628022611141205, + 0.010169325396418571, + 0.0067228595726192, + 0.007943025790154934, + 0.009418550878763199, + 0.02484813518822193, + 0.01658702827990055, + 0.02562509849667549, + 0.017645716667175293, + 0.022844674065709114, + 0.023343754932284355, + 0.04207045957446098, + 0.04161546379327774, + 0.019995002076029778, + 0.02297673560678959, + 0.014638479799032211, + 0.012921691872179508, + 0.03237251192331314, + 0.01762550137937069, + 0.024070333689451218, + 0.017163224518299103, + 0.039328694343566895, + 0.05308612808585167, + 0.036119941622018814, + 0.021312426775693893, + 0.021429739892482758, + 0.029299957677721977, + 0.03511682152748108, + 0.02686207741498947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.004095263313502073 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054945945739746094 + ], + "opposite_logprob_tokens": [ + -2.929945945739746 + ], + "logprob_reward": -0.054945945739746094, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054945945739746094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0890112966299057, + 0.07248154282569885, + 0.03040865994989872, + 0.05279138311743736, + 0.031848035752773285, + 0.03978217393159866, + 0.05988413840532303, + 0.21103248000144958, + 0.30106326937675476, + 0.11169704794883728 + ], + "credit_map": [ + 0.009292110800743103, + 0.009249858558177948, + 0.006903314497321844, + 0.00851708184927702, + 0.008943913504481316, + 0.010058346204459667, + 0.01657954603433609, + 0.019467128440737724, + 0.006202866788953543, + 0.008945338428020477, + 0.00470355711877346, + 0.008442147634923458, + 0.008755814284086227, + 0.005530731752514839, + 0.01854284107685089, + 0.011358252726495266, + 0.006982420105487108, + 0.0034103975631296635, + 0.0029060787055641413, + 0.0030063509475439787, + 0.0019361973972991109, + 0.003126817988231778, + 0.003837907686829567, + 0.005202488508075476, + 0.004725050646811724, + 0.00596973579376936, + 0.00594692025333643, + 0.007068149745464325, + 0.005773832090198994, + 0.006379690952599049, + 0.008885873481631279, + 0.00804213061928749, + 0.003273044480010867, + 0.0037647488061338663, + 0.002782106399536133, + 0.0030394478235393763, + 0.00434842798858881, + 0.0049124909564852715, + 0.0036769723519682884, + 0.006050795782357454, + 0.005823684856295586, + 0.004096006974577904, + 0.004849822726100683, + 0.004344655200839043, + 0.005760561674833298, + 0.004373899661004543, + 0.0061521404422819614, + 0.004381401930004358, + 0.006463914178311825, + 0.009003814309835434, + 0.006867120508104563, + 0.0057829031720757484, + 0.0078029814176261425, + 0.008166289888322353, + 0.006940382067114115, + 0.00885673426091671, + 0.02495761774480343, + 0.016055552288889885, + 0.01867312751710415, + 0.017489183694124222, + 0.02319222502410412, + 0.02408190257847309, + 0.048052456229925156, + 0.038530416786670685, + 0.03396247699856758, + 0.037935882806777954, + 0.026851339265704155, + 0.026232359930872917, + 0.05939670279622078, + 0.029919331893324852, + 0.055339280515909195, + 0.03142591193318367, + 0.01962173357605934, + 0.014157108031213284, + 0.018388543277978897, + 0.009075564332306385, + 0.012596562504768372, + 0.01315128430724144, + 0.013170263729989529, + 0.011535978876054287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.054945945739746094 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up tracking shot, held at a slightly upward angle, locks onto the \"head\" of a massive Tyrannosaurus rex. The entire image is imbued with Hirohiko Araki's signature aesthetic: sharp, hard lines and dazzling, saturated colors. The mutant plant's skin, instead of its natural green, is an eerie blend of acid green and electric purple, covered in sculpted geometric angles and deep grooves. Knife-sharp magenta spikes jut from its ridges. As the camera steadily tracks, a bud-like structure at the plant's apex begins to pulsate violently and rhythmically, emanating an ominous golden glow from within. Suddenly, it violently explodes outward, opening into a terrifying, jagged \"maw\" and emitting a silent yet visually shattering roar. A visible circular shock wave composed of pure energy gushed out from its core. The shock wave was translucent cyan, with unstable white arcs flashing at the edges. It expanded outward at an astonishing speed, distorting and refracting the air wherever it passed, instantly blurring the equally colorful surreal sky and earth in the background, creating an atmosphere full of power, extreme tension, and grotesque splendor.", + "seed": 17713, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_8/index_2970_seed_17713_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02831659922376275, + "vlm_logprob": -0.02831659922376275, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02831659922376275, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01575475186109543 + ], + "opposite_logprob_tokens": [ + -4.390754699707031 + ], + "logprob_reward": -0.01575475186109543, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01575475186109543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0610801987349987, + 0.057539477944374084, + 0.060706667602062225, + 0.07299795746803284, + 0.12239830940961838, + 0.0728076696395874, + 0.26349151134490967, + 0.1709810197353363, + 0.07697907835245132, + 0.04101808741688728 + ], + "credit_map": [ + 0.007317154668271542, + 0.007378709502518177, + 0.01018740888684988, + 0.006640592124313116, + 0.010031894780695438, + 0.006568447221070528, + 0.006259929854422808, + 0.006696063559502363, + 0.005541754420846701, + 0.009442150592803955, + 0.00715182488784194, + 0.010026663541793823, + 0.008015422150492668, + 0.005305747501552105, + 0.007696757558733225, + 0.0043591586872935295, + 0.005595685914158821, + 0.004544941708445549, + 0.006133285351097584, + 0.018247930333018303, + 0.005164599511772394, + 0.006998696830123663, + 0.008193095214664936, + 0.005828438326716423, + 0.006946225184947252, + 0.0058031389489769936, + 0.005940805654972792, + 0.011394654400646687, + 0.008256479166448116, + 0.00993668008595705, + 0.012525040656328201, + 0.01219493430107832, + 0.020101984962821007, + 0.011528225615620613, + 0.01606658101081848, + 0.009913232177495956, + 0.010906209237873554, + 0.014883110299706459, + 0.011851110495626926, + 0.027147851884365082, + 0.009894182905554771, + 0.011072883382439613, + 0.008980665355920792, + 0.009769364260137081, + 0.008064917288720608, + 0.007943251170217991, + 0.008599920198321342, + 0.008482490666210651, + 0.05501479655504227, + 0.019124014303088188, + 0.023488635197281837, + 0.048448964953422546, + 0.02223155088722706, + 0.018325623124837875, + 0.02240072563290596, + 0.05445721000432968, + 0.022051982581615448, + 0.024405736476182938, + 0.01615089178085327, + 0.023959079757332802, + 0.023697959259152412, + 0.0170715544372797, + 0.01863970048725605, + 0.025004126131534576, + 0.010776069946587086, + 0.012224542908370495, + 0.007979001849889755, + 0.010125848464667797, + 0.011049646884202957, + 0.006015478633344173, + 0.007430731318891048, + 0.011377755552530289, + 0.006452272180467844, + 0.007287106476724148, + 0.004846387077122927, + 0.005292192101478577, + 0.004307401832193136, + 0.003479276318103075, + 0.003330034902319312, + 0.006023413967341185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.01575475186109543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004655356053262949 + ], + "opposite_logprob_tokens": [ + -5.379655361175537 + ], + "logprob_reward": -0.004655356053262949, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004655356053262949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07957763969898224, + 0.06722775846719742, + 0.1269473284482956, + 0.08048856258392334, + 0.07563157379627228, + 0.1172088086605072, + 0.15736855566501617, + 0.08493754267692566, + 0.052953340113162994, + 0.15765886008739471 + ], + "credit_map": [ + 0.009362773038446903, + 0.007941192016005516, + 0.009688107296824455, + 0.009490832686424255, + 0.014673225581645966, + 0.008980754762887955, + 0.008067050948739052, + 0.011373708955943584, + 0.008294682949781418, + 0.0072603956796228886, + 0.006006539799273014, + 0.014531910419464111, + 0.006800467148423195, + 0.007621214725077152, + 0.009122268296778202, + 0.007590287830680609, + 0.012560971081256866, + 0.00951610691845417, + 0.013707360252737999, + 0.043289072811603546, + 0.006401076912879944, + 0.01627855934202671, + 0.01767849735915661, + 0.007515686564147472, + 0.009222481399774551, + 0.008901724591851234, + 0.01062684878706932, + 0.014543481171131134, + 0.008860129863023758, + 0.0074930619448423386, + 0.011782742105424404, + 0.009058102034032345, + 0.01289951242506504, + 0.0074640903621912, + 0.009676930494606495, + 0.008930369280278683, + 0.0059812357649207115, + 0.009261908940970898, + 0.006499205715954304, + 0.014918322674930096, + 0.026286311447620392, + 0.01774044893682003, + 0.01422076765447855, + 0.01179653499275446, + 0.01025665644556284, + 0.009793003089725971, + 0.009961765259504318, + 0.01715332269668579, + 0.0287482887506485, + 0.014184996485710144, + 0.013571847230196, + 0.03872673213481903, + 0.014722773805260658, + 0.010076161473989487, + 0.015354517847299576, + 0.021983256563544273, + 0.01287867035716772, + 0.011450630612671375, + 0.008139620535075665, + 0.011240694671869278, + 0.008559166453778744, + 0.007002305705100298, + 0.009027116931974888, + 0.01663934625685215, + 0.0077147092670202255, + 0.007323646917939186, + 0.006829261314123869, + 0.009316064417362213, + 0.004982073325663805, + 0.003088561352342367, + 0.004569486249238253, + 0.009129542857408524, + 0.02597133256494999, + 0.029537120833992958, + 0.022358257323503494, + 0.0169050432741642, + 0.01590503565967083, + 0.009698725305497646, + 0.010997597128152847, + 0.02628575637936592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.004655356053262949 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160688366740942 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160688366740942, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160688366740942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10667793452739716, + 0.074205182492733, + 0.12288226187229156, + 0.246736079454422, + 0.07762259989976883, + 0.09659846127033234, + 0.08025208860635757, + 0.0896826684474945, + 0.055839620530605316, + 0.049503106623888016 + ], + "credit_map": [ + 0.010505071841180325, + 0.012339146807789803, + 0.013229826465249062, + 0.014112825505435467, + 0.01239411998540163, + 0.009036376141011715, + 0.014327998273074627, + 0.020732566714286804, + 0.00893755815923214, + 0.01005154475569725, + 0.006665543653070927, + 0.011724187061190605, + 0.010811683721840382, + 0.008295937441289425, + 0.009915810078382492, + 0.007802919484674931, + 0.01847601681947708, + 0.00824347510933876, + 0.012666178867220879, + 0.035078730434179306, + 0.006807285360991955, + 0.019638581201434135, + 0.01457280945032835, + 0.007399191614240408, + 0.021169528365135193, + 0.020236315205693245, + 0.01822539232671261, + 0.04554009810090065, + 0.02621847204864025, + 0.02654619701206684, + 0.04472627490758896, + 0.04407379403710365, + 0.010431683622300625, + 0.009563528932631016, + 0.008240872994065285, + 0.007954100146889687, + 0.010276706889271736, + 0.008078599348664284, + 0.010056808590888977, + 0.013020298443734646, + 0.016405710950493813, + 0.012733635492622852, + 0.009756100364029408, + 0.011394255794584751, + 0.011029851622879505, + 0.012045139446854591, + 0.010386568494141102, + 0.012847207486629486, + 0.009898178279399872, + 0.009851539507508278, + 0.012340818531811237, + 0.012014485895633698, + 0.008836707100272179, + 0.007512566167861223, + 0.006930376403033733, + 0.012867415323853493, + 0.013695196248590946, + 0.007617514114826918, + 0.007401354610919952, + 0.010050129145383835, + 0.014232040382921696, + 0.01082538440823555, + 0.009898020885884762, + 0.015963027253746986, + 0.00782855600118637, + 0.006814881227910519, + 0.006833525374531746, + 0.010180823504924774, + 0.008501038886606693, + 0.00415677297860384, + 0.005922278389334679, + 0.005601748824119568, + 0.007743387017399073, + 0.0073709432035684586, + 0.0071780080907046795, + 0.00687285466119647, + 0.005114567466080189, + 0.00371956336311996, + 0.004084146115928888, + 0.007419635076075792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0160688366740942 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204738840460777 + ], + "opposite_logprob_tokens": [ + -2.8120474815368652 + ], + "logprob_reward": -0.06204738840460777, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204738840460777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08928419649600983, + 0.05490749701857567, + 0.3274649977684021, + 0.10181674361228943, + 0.044695474207401276, + 0.04631615802645683, + 0.07046554982662201, + 0.10759180039167404, + 0.10087022930383682, + 0.0565873384475708 + ], + "credit_map": [ + 0.012513402849435806, + 0.014262888580560684, + 0.010685916990041733, + 0.010155065916478634, + 0.012773239985108376, + 0.007470330689102411, + 0.00818769633769989, + 0.013235659338533878, + 0.005123419221490622, + 0.006086325738579035, + 0.005567850545048714, + 0.013049122877418995, + 0.005725295282900333, + 0.007106907665729523, + 0.007707993965595961, + 0.0045405845157802105, + 0.03148100525140762, + 0.02525499276816845, + 0.03114388696849346, + 0.12298987805843353, + 0.009550133720040321, + 0.04280871897935867, + 0.04855337738990784, + 0.015682995319366455, + 0.009217964485287666, + 0.008918077684938908, + 0.007862472906708717, + 0.021555503830313683, + 0.010894207283854485, + 0.015157245099544525, + 0.013527899980545044, + 0.01468337420374155, + 0.007543876767158508, + 0.003949228208512068, + 0.005154299084097147, + 0.0036239868495613337, + 0.0035031242296099663, + 0.005248368252068758, + 0.004544790368527174, + 0.01112779974937439, + 0.008160186931490898, + 0.007848509587347507, + 0.005280580371618271, + 0.005800953600555658, + 0.004275356885045767, + 0.004405339248478413, + 0.003863824065774679, + 0.006681409198790789, + 0.007622034288942814, + 0.008967921137809753, + 0.008457145653665066, + 0.014424044638872147, + 0.007227074354887009, + 0.007397675886750221, + 0.0070977224968373775, + 0.009271928109228611, + 0.021858325228095055, + 0.015061273239552975, + 0.010695564560592175, + 0.012608085758984089, + 0.012445232830941677, + 0.00939115323126316, + 0.009552490897476673, + 0.015979668125510216, + 0.012966789305210114, + 0.01387707144021988, + 0.013339603319764137, + 0.020265348255634308, + 0.010831781663000584, + 0.006669262424111366, + 0.01184110064059496, + 0.011079274117946625, + 0.0068770828656852245, + 0.006462005898356438, + 0.006392764393240213, + 0.00666685588657856, + 0.006442169193178415, + 0.005087966565042734, + 0.00493167620152235, + 0.013726822100579739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.06204738840460777 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430566631257534 + ], + "opposite_logprob_tokens": [ + -3.1680567264556885 + ], + "logprob_reward": -0.0430566631257534, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430566631257534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1820807158946991, + 0.11128824204206467, + 0.10718291252851486, + 0.04977584630250931, + 0.05858139321208, + 0.09670250117778778, + 0.24854348599910736, + 0.039743728935718536, + 0.0827162116765976, + 0.023384977132081985 + ], + "credit_map": [ + 0.0192483551800251, + 0.022087058052420616, + 0.026253459975123405, + 0.021621881052851677, + 0.03419102728366852, + 0.01620842143893242, + 0.023308878764510155, + 0.019161639735102654, + 0.01431050430983305, + 0.010144319385290146, + 0.012794377282261848, + 0.02125260978937149, + 0.013541768305003643, + 0.015412472188472748, + 0.012700612656772137, + 0.011131574399769306, + 0.011504827998578548, + 0.008074074052274227, + 0.008961415849626064, + 0.035508230328559875, + 0.004638675134629011, + 0.013512049801647663, + 0.01532942708581686, + 0.00965422298759222, + 0.007139801047742367, + 0.0046050106175243855, + 0.0055795758962631226, + 0.006231147330254316, + 0.005489772651344538, + 0.005973179824650288, + 0.00870714895427227, + 0.0060502151027321815, + 0.008961619809269905, + 0.009146999567747116, + 0.006510902661830187, + 0.010550500825047493, + 0.006813628599047661, + 0.0032920415978878736, + 0.0056800441816449165, + 0.007625658065080643, + 0.020658154040575027, + 0.013691036030650139, + 0.012068606913089752, + 0.0096796415746212, + 0.007436818443238735, + 0.007748557720333338, + 0.008213812485337257, + 0.017205869778990746, + 0.05714768171310425, + 0.027386832982301712, + 0.019840290769934654, + 0.025027241557836533, + 0.016622265800833702, + 0.018543217331171036, + 0.017336087301373482, + 0.06663985550403595, + 0.005594679154455662, + 0.006491898559033871, + 0.005302643869072199, + 0.00637265807017684, + 0.004061580169945955, + 0.00422841589897871, + 0.003182044019922614, + 0.004509808495640755, + 0.011466896161437035, + 0.01097468938678503, + 0.00879125576466322, + 0.008870182558894157, + 0.01156945526599884, + 0.008853279054164886, + 0.008889543823897839, + 0.013300907798111439, + 0.0029959913808852434, + 0.0024402476847171783, + 0.0023753440473228693, + 0.0024320357479155064, + 0.002891738433390856, + 0.0020974797662347555, + 0.003328184364363551, + 0.004823956172913313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0430566631257534 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up tracking shot, held at a slightly upward angle, locks onto the \"head\" of a massive Tyrannosaurus rex. The entire image is imbued with Hirohiko Araki's signature aesthetic: sharp, hard lines and dazzling, saturated colors. The mutant plant's skin, instead of its natural green, is an eerie blend of acid green and electric purple, covered in sculpted geometric angles and deep grooves. Knife-sharp magenta spikes jut from its ridges. As the camera steadily tracks, a bud-like structure at the plant's apex begins to pulsate violently and rhythmically, emanating an ominous golden glow from within. Suddenly, it violently explodes outward, opening into a terrifying, jagged \"maw\" and emitting a silent yet visually shattering roar. A visible circular shock wave composed of pure energy gushed out from its core. The shock wave was translucent cyan, with unstable white arcs flashing at the edges. It expanded outward at an astonishing speed, distorting and refracting the air wherever it passed, instantly blurring the equally colorful surreal sky and earth in the background, creating an atmosphere full of power, extreme tension, and grotesque splendor.", + "seed": 17714, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_8/index_2970_seed_17714_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01564853861927986, + "vlm_logprob": -0.01564853861927986, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.01564853861927986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015937911346554756 + ], + "opposite_logprob_tokens": [ + -4.265937805175781 + ], + "logprob_reward": -0.015937911346554756, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015937911346554756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11953691393136978, + 0.31863006949424744, + 0.05896647274494171, + 0.04300156980752945, + 0.037988681346178055, + 0.03825859725475311, + 0.051079653203487396, + 0.17060618102550507, + 0.0785590335726738, + 0.08337284624576569 + ], + "credit_map": [ + 0.015032479539513588, + 0.013998707756400108, + 0.014510276727378368, + 0.011861930601298809, + 0.012801133096218109, + 0.01757843792438507, + 0.019235841929912567, + 0.014518099837005138, + 0.02626725658774376, + 0.020253175869584084, + 0.04112032800912857, + 0.024855369701981544, + 0.03924163803458214, + 0.052878327667713165, + 0.049536872655153275, + 0.0644771158695221, + 0.0063958908431231976, + 0.006498470902442932, + 0.006102433893829584, + 0.005808653309941292, + 0.008708116598427296, + 0.009715882129967213, + 0.006500416900962591, + 0.009236607700586319, + 0.0038212821818888187, + 0.004633957985788584, + 0.005568103399127722, + 0.004925800487399101, + 0.005716999527066946, + 0.005890043452382088, + 0.005959860049188137, + 0.006485522259026766, + 0.006378234829753637, + 0.002447892911732197, + 0.0035542279947549105, + 0.0036112493835389614, + 0.004036354832351208, + 0.007592903915792704, + 0.00562258530408144, + 0.0047452314756810665, + 0.004155553411692381, + 0.0038649458438158035, + 0.004675352014601231, + 0.00514138163998723, + 0.0046483492478728294, + 0.004804523661732674, + 0.0049990382976830006, + 0.005969453603029251, + 0.008377699181437492, + 0.0033872253261506557, + 0.008543841540813446, + 0.005412305239588022, + 0.005101324524730444, + 0.0073525863699615, + 0.005963391624391079, + 0.006941280327737331, + 0.021519819274544716, + 0.018153395503759384, + 0.022515801712870598, + 0.018582865595817566, + 0.017720136791467667, + 0.028748296201229095, + 0.018889913335442543, + 0.02447594702243805, + 0.012193459086120129, + 0.006545704789459705, + 0.012185351923108101, + 0.005223492160439491, + 0.007538722362369299, + 0.012788253836333752, + 0.00539874704554677, + 0.016685297712683678, + 0.015109077095985413, + 0.010700312443077564, + 0.012599373236298561, + 0.00752486614510417, + 0.008256502449512482, + 0.011384296230971813, + 0.007511564530432224, + 0.010286848060786724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.015937911346554756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041108159348368645 + ], + "opposite_logprob_tokens": [ + -5.504110813140869 + ], + "logprob_reward": -0.0041108159348368645, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041108159348368645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3072476387023926, + 0.1345730721950531, + 0.1623488813638687, + 0.058763571083545685, + 0.030202800408005714, + 0.020900584757328033, + 0.036799076944589615, + 0.06664891541004181, + 0.11651400476694107, + 0.06600141525268555 + ], + "credit_map": [ + 0.03351222723722458, + 0.038855843245983124, + 0.027072738856077194, + 0.023970702663064003, + 0.05267413705587387, + 0.040665153414011, + 0.057412393391132355, + 0.03308442607522011, + 0.01092711091041565, + 0.008532102219760418, + 0.020000996068120003, + 0.008338496088981628, + 0.015694670379161835, + 0.02521830052137375, + 0.014290385879576206, + 0.03157101944088936, + 0.014858017675578594, + 0.025038275867700577, + 0.016237642616033554, + 0.014800628647208214, + 0.02229270152747631, + 0.026818541809916496, + 0.021542152389883995, + 0.020760923624038696, + 0.005852739792317152, + 0.005501064471900463, + 0.005685271695256233, + 0.00705529423430562, + 0.008012580685317516, + 0.008193986490368843, + 0.010653280653059483, + 0.007809358183294535, + 0.0025842601899057627, + 0.002862381050363183, + 0.003394604893401265, + 0.0051997085101902485, + 0.0043572718277573586, + 0.0031935812439769506, + 0.004437311086803675, + 0.004173681139945984, + 0.0023432252928614616, + 0.0019122667144984007, + 0.0028162330854684114, + 0.002400655997917056, + 0.0023843401577323675, + 0.0033173798583447933, + 0.0022674433421343565, + 0.0034590403083711863, + 0.005319274961948395, + 0.003202612278982997, + 0.005376843735575676, + 0.0034463019110262394, + 0.006424652878195047, + 0.0056393262930214405, + 0.002606433816254139, + 0.0047836326994001865, + 0.0068934764713048935, + 0.0038678369019180536, + 0.01032314170151949, + 0.00520463939756155, + 0.0033907543402165174, + 0.01460024993866682, + 0.004268560092896223, + 0.018100257962942123, + 0.01685221493244171, + 0.006705829408019781, + 0.015881899744272232, + 0.005079073365777731, + 0.0064807250164449215, + 0.033291760832071304, + 0.005108747631311417, + 0.02711375802755356, + 0.01281076017767191, + 0.00520970206707716, + 0.010756483301520348, + 0.0071341064758598804, + 0.006528103724122047, + 0.008805186487734318, + 0.006570528261363506, + 0.008186550810933113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0041108159348368645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028190179727971554 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.0028190179727971554, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028190179727971554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17774426937103271, + 0.09664050489664078, + 0.060560643672943115, + 0.07644295692443848, + 0.03987531736493111, + 0.03787023201584816, + 0.10401608049869537, + 0.11356760561466217, + 0.20152227580547333, + 0.09176008403301239 + ], + "credit_map": [ + 0.01705038920044899, + 0.020201578736305237, + 0.010202538222074509, + 0.017182737588882446, + 0.0253340732306242, + 0.033350225538015366, + 0.03379128500819206, + 0.0206314530223608, + 0.011191957630217075, + 0.008327578194439411, + 0.016622597351670265, + 0.007871847599744797, + 0.008821268565952778, + 0.019881442189216614, + 0.008870857767760754, + 0.015052955597639084, + 0.006551746279001236, + 0.00561343552544713, + 0.00958336889743805, + 0.007612826302647591, + 0.005721608176827431, + 0.008634856902062893, + 0.008050868287682533, + 0.008791936561465263, + 0.00750374561175704, + 0.006380683742463589, + 0.009835788980126381, + 0.008128823712468147, + 0.009468551725149155, + 0.009635189548134804, + 0.011858219280838966, + 0.013631951063871384, + 0.005884155631065369, + 0.003604218363761902, + 0.004572045058012009, + 0.004311831202358007, + 0.004259615670889616, + 0.005563688464462757, + 0.006656961981207132, + 0.005022803787142038, + 0.0033989606890827417, + 0.004284202121198177, + 0.004002868197858334, + 0.00630697887390852, + 0.003536426229402423, + 0.003942425362765789, + 0.007064135745167732, + 0.005334236193448305, + 0.01101711392402649, + 0.011369545944035053, + 0.00907029490917921, + 0.015911614522337914, + 0.017037998884916306, + 0.010783685371279716, + 0.01620236597955227, + 0.012623464688658714, + 0.022190315648913383, + 0.008872190490365028, + 0.016412068158388138, + 0.010386957786977291, + 0.008524219505488873, + 0.023105943575501442, + 0.01024541910737753, + 0.013830495066940784, + 0.025492500513792038, + 0.017257193103432655, + 0.02494538389146328, + 0.014751452952623367, + 0.025040514767169952, + 0.04369566962122917, + 0.01503470167517662, + 0.03530488163232803, + 0.016028625890612602, + 0.01052490808069706, + 0.011772779747843742, + 0.009110063314437866, + 0.012264751829206944, + 0.010148497298359871, + 0.008706343360245228, + 0.013204120099544525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0028190179727971554 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067332289181649685 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.0067332289181649685, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067332289181649685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06806973367929459, + 0.1032271757721901, + 0.04582958668470383, + 0.02637675404548645, + 0.01860867626965046, + 0.022334197536110878, + 0.03330819681286812, + 0.08701762557029724, + 0.2663956582546234, + 0.32883238792419434 + ], + "credit_map": [ + 0.008469315245747566, + 0.007838962599635124, + 0.007413988467305899, + 0.005381754599511623, + 0.011985134333372116, + 0.008306387811899185, + 0.009001987986266613, + 0.009672198444604874, + 0.008649354800581932, + 0.0059202020056545734, + 0.013146226294338703, + 0.006615275051444769, + 0.012577071785926819, + 0.019276367500424385, + 0.01354996208101511, + 0.023492708802223206, + 0.004505061078816652, + 0.004432715009897947, + 0.0054486882872879505, + 0.004695889074355364, + 0.005427531432360411, + 0.008636110462248325, + 0.006276832893490791, + 0.0064067537896335125, + 0.0034662059042602777, + 0.002094101859256625, + 0.0037037977017462254, + 0.002681535203009844, + 0.003085553180426359, + 0.0033650794066488743, + 0.003916249144822359, + 0.004064228851348162, + 0.002450350672006607, + 0.0016594183398410678, + 0.002206709235906601, + 0.001962461741641164, + 0.002212545368820429, + 0.0028373426757752895, + 0.002390502952039242, + 0.002889343537390232, + 0.0022960177157074213, + 0.0019079355988651514, + 0.0027853352949023247, + 0.003058945992961526, + 0.0026533007621765137, + 0.0027445070445537567, + 0.003431214252486825, + 0.003456938313320279, + 0.005159085616469383, + 0.002583759371191263, + 0.005352040287107229, + 0.0022939899936318398, + 0.0048889415338635445, + 0.00537393381819129, + 0.0032087441068142653, + 0.0044476971961557865, + 0.021728692576289177, + 0.005491956602782011, + 0.009983750991523266, + 0.004359730985015631, + 0.004905513487756252, + 0.027066625654697418, + 0.003317924914881587, + 0.010163418017327785, + 0.04333043843507767, + 0.012128227390348911, + 0.03806016594171524, + 0.01074732281267643, + 0.011132600717246532, + 0.07737705856561661, + 0.01329265721142292, + 0.060327187180519104, + 0.05545515939593315, + 0.04338326305150986, + 0.04464658349752426, + 0.02495509199798107, + 0.0360165610909462, + 0.050942156463861465, + 0.028503093868494034, + 0.044930484145879745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.0067332289181649685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up tracking shot of a massive Tyrannosaurus rex head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the T-Rex skin colored with an eerie blend of acid green and electric purple featuring geometric angles and magenta spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a bud-like structure at the apex pulsate violently and emit an ominous golden glow before exploding into a jagged maw?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a translucent cyan circular shock wave with unstable white arcs emitted from the core, distorting the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held at a slightly upward angle while steadily tracking the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864171892404556 + ], + "opposite_logprob_tokens": [ + -3.0486416816711426 + ], + "logprob_reward": -0.04864171892404556, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864171892404556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09112807363271713, + 0.2360362708568573, + 0.0628533586859703, + 0.04061310365796089, + 0.042040932923555374, + 0.03022138960659504, + 0.04046855866909027, + 0.12143144756555557, + 0.2578805685043335, + 0.07732624560594559 + ], + "credit_map": [ + 0.010481361299753189, + 0.014619935303926468, + 0.008776932023465633, + 0.00810333527624607, + 0.01558699645102024, + 0.01182504277676344, + 0.013574321754276752, + 0.008160156197845936, + 0.022875728085637093, + 0.016966886818408966, + 0.030979584902524948, + 0.021711168810725212, + 0.03552393615245819, + 0.032688893377780914, + 0.03942154720425606, + 0.035868529230356216, + 0.006708748172968626, + 0.006270501762628555, + 0.00564459478482604, + 0.006722545251250267, + 0.008862135000526905, + 0.0099129443988204, + 0.009501438587903976, + 0.009230446070432663, + 0.0032086814753711224, + 0.003658080007880926, + 0.005545203574001789, + 0.004495651926845312, + 0.005851412657648325, + 0.0049044606275856495, + 0.007247396279126406, + 0.005702218506485224, + 0.007354286033660173, + 0.0030259322375059128, + 0.005446566268801689, + 0.003183247521519661, + 0.004687194246798754, + 0.008034721948206425, + 0.003990569151937962, + 0.0063184150494635105, + 0.003412740072235465, + 0.0030306687112897635, + 0.0045293355360627174, + 0.0029512248001992702, + 0.003238114295527339, + 0.0041616433300077915, + 0.003729385556653142, + 0.005168275907635689, + 0.005404860246926546, + 0.0032017859630286694, + 0.006167197600007057, + 0.003967781085520983, + 0.005230701994150877, + 0.004393735900521278, + 0.0062377192080020905, + 0.00586477667093277, + 0.01825234852731228, + 0.006155672017484903, + 0.017070002853870392, + 0.006274412851780653, + 0.00690426817163825, + 0.03305664658546448, + 0.00523901404812932, + 0.02847908064723015, + 0.032419051975011826, + 0.014773682691156864, + 0.022358378395438194, + 0.010496959090232849, + 0.011165719479322433, + 0.08402366191148758, + 0.010643589310348034, + 0.07199952751398087, + 0.011438642628490925, + 0.007488855626434088, + 0.01110068242996931, + 0.006585877388715744, + 0.007218037731945515, + 0.010365934111177921, + 0.007625301368534565, + 0.01550291571766138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.04864171892404556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_9/index_tensor([1, 1])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_9/index_tensor([1, 1])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..68e027a5d6b1d16cc1ff1a6765fc14d6a960c07b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_9/index_tensor([1, 1])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, upward-focused shot draws us into a boundless, pure Klein blue abyss, devoid of visible rocks or seabed, dominated by absolute silence and emptiness. At the center of the frame, a massive branching jellyfish hangs suspended within it, a complex organism composed of countless luminous units. Its translucent, chain-like structures emanate a soft, inner glow of white and pale cyan, contrasting sharply against the deep blue backdrop. A silent, Abstract Expressionist explosion of life unfolds at its core: Shot in stunning slow motion, countless shimmering golden and silver microparticles, like awakened cosmic dust, gracefully and slowly erupt and disperse from its gelatinous mass. These particles, far from a chaotic mess, form spiraling, dreamlike galaxies that dance and diffuse through the still waters, creating a vast, dynamic canvas filled with the miracle of life. The brilliant light emitted by the jellyfish itself and the particles is the only light source, dyeing the surrounding Klein blue water with a mysterious halo. The whole scene is filled with the solemnity, tranquility and surreal dreamlike feeling of the birth of the universe.", + "seed": 33405, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_9/index_1_seed_33405_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.21665366017259657, + "vlm_logprob": -0.21665366017259657, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.21665366017259657, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018829645588994026 + ], + "opposite_logprob_tokens": [ + -4.018829822540283 + ], + "logprob_reward": -0.018829645588994026, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018829645588994026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050157330930233, + 0.7903767228126526, + 0.008231857791543007, + 0.008299878798425198, + 0.010246428661048412, + 0.028255542740225792, + 0.04173387587070465, + 0.033215444535017014, + 0.02203342318534851, + 0.007449539378285408 + ], + "credit_map": [ + 0.007388560101389885, + 0.005459357984364033, + 0.008854439482092857, + 0.00318476022221148, + 0.006064512766897678, + 0.00496219377964735, + 0.004017557017505169, + 0.010225947946310043, + 0.04536648467183113, + 0.031333088874816895, + 0.31098443269729614, + 0.0284541305154562, + 0.03419877961277962, + 0.025316640734672546, + 0.026192069053649902, + 0.2885311245918274, + 0.0009615860763005912, + 0.0010193586349487305, + 0.0007918913615867496, + 0.0009265469270758331, + 0.0006495655979961157, + 0.0011426417622715235, + 0.0015542445471510291, + 0.0011860223021358252, + 0.0008374914759770036, + 0.0010043655056506395, + 0.0011579507263377309, + 0.0013278040569275618, + 0.0011270211543887854, + 0.0009626420214772224, + 0.0010592624312266707, + 0.0008233410771936178, + 0.0008379940409213305, + 0.0012181279016658664, + 0.0012432225048542023, + 0.0011559645645320415, + 0.0022229603491723537, + 0.0012542264303192496, + 0.0015545047353953123, + 0.0007594291237182915, + 0.002250629011541605, + 0.003104366594925523, + 0.0041618263348937035, + 0.0042969826608896255, + 0.005385609343647957, + 0.0018231818685308099, + 0.0030553359538316727, + 0.004177611321210861, + 0.0031360872089862823, + 0.0022059697657823563, + 0.013860994018614292, + 0.002405957318842411, + 0.0025569775607436895, + 0.003672759048640728, + 0.002386346459388733, + 0.011508787050843239, + 0.00368467066437006, + 0.003433976788073778, + 0.00414005434140563, + 0.0040397741831839085, + 0.005752039607614279, + 0.004125027917325497, + 0.0032956586219370365, + 0.004744241014122963, + 0.0017191666411235929, + 0.002025222172960639, + 0.002166767604649067, + 0.00253611383959651, + 0.005302716046571732, + 0.0028774356469511986, + 0.002409549430012703, + 0.0029964528512209654, + 0.0009952998952940106, + 0.0009039185242727399, + 0.0012107979273423553, + 0.0006802844000048935, + 0.000902806525118649, + 0.0009817329701036215, + 0.0008542377036064863, + 0.0009204618399962783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.018829645588994026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032027876004576683 + ], + "opposite_logprob_tokens": [ + -5.75320291519165 + ], + "logprob_reward": -0.0032027876004576683, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032027876004576683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3401499390602112, + 0.2862168550491333, + 0.05459858104586601, + 0.08304408192634583, + 0.08602549880743027, + 0.019498737528920174, + 0.04005700722336769, + 0.026041854172945023, + 0.042184729129076004, + 0.02218266762793064 + ], + "credit_map": [ + 0.04024537280201912, + 0.04695361852645874, + 0.06577010452747345, + 0.027247129008173943, + 0.04344639554619789, + 0.03670656681060791, + 0.036347851157188416, + 0.04343292862176895, + 0.017079971730709076, + 0.018581219017505646, + 0.10372930765151978, + 0.010666104033589363, + 0.023058783262968063, + 0.009291986003518105, + 0.01215534657239914, + 0.09165416657924652, + 0.004584154114127159, + 0.007299186661839485, + 0.00553900608792901, + 0.00691395066678524, + 0.0060211974196136, + 0.007481001317501068, + 0.01039914507418871, + 0.006360948551446199, + 0.006887795403599739, + 0.005956413224339485, + 0.008987372741103172, + 0.02335839346051216, + 0.016040680930018425, + 0.007243439555168152, + 0.008361954241991043, + 0.006208046339452267, + 0.009118441492319107, + 0.011141709052026272, + 0.010160312056541443, + 0.01047493051737547, + 0.014439097605645657, + 0.009843803010880947, + 0.01355317048728466, + 0.007294042501598597, + 0.0018039005808532238, + 0.002539010252803564, + 0.0022496210876852274, + 0.0026442669332027435, + 0.004256661050021648, + 0.0015981055330485106, + 0.002217861358076334, + 0.0021893128287047148, + 0.003370360005646944, + 0.0026875766925513744, + 0.006313324440270662, + 0.004345025401562452, + 0.005008117761462927, + 0.004472987726330757, + 0.003179130842909217, + 0.01068048644810915, + 0.0027479121927171946, + 0.0024802430998533964, + 0.002862125402316451, + 0.0031655223574489355, + 0.0042861527763307095, + 0.00322162127122283, + 0.00297378352843225, + 0.004304497502744198, + 0.004381082020699978, + 0.004494448658078909, + 0.0038707221392542124, + 0.005914576351642609, + 0.00478335889056325, + 0.005897196475416422, + 0.004959553945809603, + 0.00788379367440939, + 0.003102908842265606, + 0.0025655387435108423, + 0.001974443905055523, + 0.002034184755757451, + 0.00441775843501091, + 0.0026383623480796814, + 0.0027646925300359726, + 0.002684779930859804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0032027876004576683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193875541910529 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193875541910529, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193875541910529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053836118429899216, + 0.5733813643455505, + 0.03154066950082779, + 0.07594428956508636, + 0.017394812777638435, + 0.033689677715301514, + 0.012370684184134007, + 0.02315550111234188, + 0.07306891679763794, + 0.1056179329752922 + ], + "credit_map": [ + 0.005705240182578564, + 0.0053663854487240314, + 0.008546090684831142, + 0.004466951824724674, + 0.005543714389204979, + 0.006859941873699427, + 0.005281548947095871, + 0.012066247873008251, + 0.03372432291507721, + 0.023930104449391365, + 0.22440774738788605, + 0.020689891651272774, + 0.02650492824614048, + 0.018637314438819885, + 0.01835593767464161, + 0.20713114738464355, + 0.00380090344697237, + 0.004064054228365421, + 0.0031567062251269817, + 0.003941250033676624, + 0.0030525638721883297, + 0.005953946150839329, + 0.004136263392865658, + 0.0034349861089140177, + 0.00740198465064168, + 0.01015450619161129, + 0.01097958441823721, + 0.012905115261673927, + 0.011574994772672653, + 0.008502508513629436, + 0.00875419843941927, + 0.005671400111168623, + 0.0021007454488426447, + 0.0015593934804201126, + 0.0031173124443739653, + 0.0011529262410476804, + 0.0026342899072915316, + 0.0015057658310979605, + 0.003208388574421406, + 0.0021159911993891, + 0.0033226858358830214, + 0.004207679536193609, + 0.004078313242644072, + 0.0045335725881159306, + 0.005326236132532358, + 0.003864023368805647, + 0.004552563652396202, + 0.003804604522883892, + 0.0010889599798247218, + 0.0012504813494160771, + 0.0017541164997965097, + 0.001438872772268951, + 0.002453040797263384, + 0.0014538749819621444, + 0.0013654979411512613, + 0.0015658398624509573, + 0.0024858361575752497, + 0.0027847143355757, + 0.0032078269869089127, + 0.0028264729771763086, + 0.00492177065461874, + 0.0027032482903450727, + 0.0022036340087652206, + 0.0020219972357153893, + 0.005620055366307497, + 0.00833994708955288, + 0.006298467982560396, + 0.0092870332300663, + 0.01682048849761486, + 0.009549042209982872, + 0.007757378742098808, + 0.009396503679454327, + 0.01971844583749771, + 0.011700289323925972, + 0.010497025214135647, + 0.008243095129728317, + 0.01449628733098507, + 0.012069624848663807, + 0.0093802809715271, + 0.019512895494699478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.003193875541910529 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096450749784708 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096450749784708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096450749784708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05712992325425148, + 0.7045735120773315, + 0.016770480200648308, + 0.02172124572098255, + 0.020390434190630913, + 0.07791713625192642, + 0.03382241353392601, + 0.015741966664791107, + 0.031033145263791084, + 0.02089974842965603 + ], + "credit_map": [ + 0.006259366869926453, + 0.006997895427048206, + 0.009526113979518414, + 0.004459352698177099, + 0.00562969408929348, + 0.00539981946349144, + 0.00543109979480505, + 0.01342658419162035, + 0.03939570114016533, + 0.017081595957279205, + 0.288519024848938, + 0.02602340467274189, + 0.012632890604436398, + 0.023263251408934593, + 0.02254962921142578, + 0.2751080393791199, + 0.0022317753173410892, + 0.002266752067953348, + 0.0017622929299250245, + 0.0018441590946167707, + 0.0015495439292863011, + 0.00311891152523458, + 0.0020939249079674482, + 0.0019031212432309985, + 0.002491004765033722, + 0.002686877269297838, + 0.0030986617784947157, + 0.004571095108985901, + 0.003536934033036232, + 0.0014838783536106348, + 0.002431262284517288, + 0.0014215335249900818, + 0.0027050154749304056, + 0.0022724359296262264, + 0.0027432620991021395, + 0.0022822446189820766, + 0.002870025346055627, + 0.0026269208174198866, + 0.002535846084356308, + 0.0023546824231743813, + 0.007068189326673746, + 0.008964392356574535, + 0.011767522431910038, + 0.010535082779824734, + 0.011167030781507492, + 0.010015285573899746, + 0.010276483371853828, + 0.00812315009534359, + 0.0024408602621406317, + 0.001805658102966845, + 0.007347290404140949, + 0.002412269590422511, + 0.0032828915864229202, + 0.003857111558318138, + 0.002224953146651387, + 0.010451378300786018, + 0.0014555150410160422, + 0.0015423438744619489, + 0.0021564888302236795, + 0.001647946541197598, + 0.0030771775636821985, + 0.001999492524191737, + 0.0014482431579381227, + 0.002414759248495102, + 0.0022238106466829777, + 0.0034480339381843805, + 0.002806704258546233, + 0.004521941300481558, + 0.007405316457152367, + 0.00362255796790123, + 0.0038584908470511436, + 0.0031462893821299076, + 0.0037409693468362093, + 0.002161228796467185, + 0.0024587216321378946, + 0.0018975820858031511, + 0.0019921446219086647, + 0.002694252645596862, + 0.002358598867431283, + 0.0035962488036602736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.004096450749784708 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.053945541381836 + ], + "opposite_logprob_tokens": [ + -0.42894554138183594 + ], + "logprob_reward": -1.053945541381836, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.053945541381836, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.041461676359176636, + 0.6589003205299377, + 0.012799945659935474, + 0.02289043739438057, + 0.02256087213754654, + 0.035190731287002563, + 0.027723481878638268, + 0.02269485406577587, + 0.0363183468580246, + 0.1194593608379364 + ], + "credit_map": [ + 0.0064038182608783245, + 0.0029821840580552816, + 0.00788120087236166, + 0.0037423535250127316, + 0.0037569550331681967, + 0.003899374045431614, + 0.005817197263240814, + 0.006978584453463554, + 0.03772341087460518, + 0.0257035531103611, + 0.2600920796394348, + 0.023652048781514168, + 0.027897385880351067, + 0.021141238510608673, + 0.02161926031112671, + 0.24107126891613007, + 0.0014029624871909618, + 0.001988450763747096, + 0.0014167592162266374, + 0.0016995689366012812, + 0.001181130763143301, + 0.0022129786666482687, + 0.0014984754379838705, + 0.0013996162451803684, + 0.0029593801591545343, + 0.0032178990077227354, + 0.00337283406406641, + 0.0042922040447592735, + 0.0024804852437227964, + 0.0023573637008666992, + 0.0026585557498037815, + 0.001551713445223868, + 0.0027817143127322197, + 0.002715206239372492, + 0.0035478186327964067, + 0.0025414018891751766, + 0.0035880114883184433, + 0.0028326839674264193, + 0.002079611411318183, + 0.0024744204711169004, + 0.002951395930722356, + 0.0036758410278707743, + 0.005905950907617807, + 0.004514757078140974, + 0.005611875094473362, + 0.0028054516296833754, + 0.004253509920090437, + 0.005471942480653524, + 0.0019484772346913815, + 0.0016231345944106579, + 0.007642300333827734, + 0.0019863913767039776, + 0.004997336305677891, + 0.0029613524675369263, + 0.0017218373250216246, + 0.004842646885663271, + 0.0027011127676814795, + 0.002536270534619689, + 0.002727571874856949, + 0.002454417757689953, + 0.003811236470937729, + 0.002806831616908312, + 0.002342594088986516, + 0.0033148149959743023, + 0.0033068053890019655, + 0.004166834522038698, + 0.003658419707790017, + 0.0045323544181883335, + 0.006960326340049505, + 0.0045609320513904095, + 0.003949097357690334, + 0.005183570086956024, + 0.025119392201304436, + 0.014535987749695778, + 0.012493723072111607, + 0.008775214664638042, + 0.012665675953030586, + 0.014024795033037663, + 0.011869545094668865, + 0.019975002855062485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -1.053945541381836 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, upward-focused shot draws us into a boundless, pure Klein blue abyss, devoid of visible rocks or seabed, dominated by absolute silence and emptiness. At the center of the frame, a massive branching jellyfish hangs suspended within it, a complex organism composed of countless luminous units. Its translucent, chain-like structures emanate a soft, inner glow of white and pale cyan, contrasting sharply against the deep blue backdrop. A silent, Abstract Expressionist explosion of life unfolds at its core: Shot in stunning slow motion, countless shimmering golden and silver microparticles, like awakened cosmic dust, gracefully and slowly erupt and disperse from its gelatinous mass. These particles, far from a chaotic mess, form spiraling, dreamlike galaxies that dance and diffuse through the still waters, creating a vast, dynamic canvas filled with the miracle of life. The brilliant light emitted by the jellyfish itself and the particles is the only light source, dyeing the surrounding Klein blue water with a mysterious halo. The whole scene is filled with the solemnity, tranquility and surreal dreamlike feeling of the birth of the universe.", + "seed": 33406, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_9/index_1_seed_33406_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21550059448927641, + "vlm_logprob": -0.21550059448927641, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.21550059448927641, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014930173754692078 + ], + "opposite_logprob_tokens": [ + -4.264930248260498 + ], + "logprob_reward": -0.014930173754692078, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014930173754692078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12008661776781082, + 0.12053507566452026, + 0.0860128402709961, + 0.3010585904121399, + 0.04061301797628403, + 0.04387223720550537, + 0.1001276895403862, + 0.06505833566188812, + 0.0988745167851448, + 0.023761065676808357 + ], + "credit_map": [ + 0.008227163925766945, + 0.00772159593179822, + 0.013029948808252811, + 0.013395080342888832, + 0.013342387974262238, + 0.022752871736884117, + 0.013471301645040512, + 0.028146272525191307, + 0.011955210007727146, + 0.011912639252841473, + 0.008765652775764465, + 0.01679806038737297, + 0.013575717806816101, + 0.022404098883271217, + 0.010675792582333088, + 0.024447904899716377, + 0.0042639560997486115, + 0.0037660105153918266, + 0.0034672028850764036, + 0.004687885753810406, + 0.0041092499159276485, + 0.029917431995272636, + 0.004440522752702236, + 0.03136058151721954, + 0.01859559305012226, + 0.022976789623498917, + 0.027072150260210037, + 0.026242271065711975, + 0.042845360934734344, + 0.050089359283447266, + 0.06874469667673111, + 0.044492386281490326, + 0.002907183952629566, + 0.0036477912217378616, + 0.003630975028499961, + 0.003909318242222071, + 0.0054392400197684765, + 0.008763762190937996, + 0.0037107972893863916, + 0.008603949099779129, + 0.003917928319424391, + 0.0033719574566930532, + 0.004955758340656757, + 0.005575892981141806, + 0.009214784950017929, + 0.005345512181520462, + 0.0060658883303403854, + 0.005424512550234795, + 0.009306928142905235, + 0.00973188504576683, + 0.013966381549835205, + 0.011162694543600082, + 0.01882746070623398, + 0.014084546826779842, + 0.009405822493135929, + 0.013641966506838799, + 0.005982606206089258, + 0.009636313654482365, + 0.0076637715101242065, + 0.0063291629776358604, + 0.01314078364521265, + 0.008226577192544937, + 0.006134419236332178, + 0.00794469565153122, + 0.009286780841648579, + 0.009374934248626232, + 0.006944591645151377, + 0.011955702677369118, + 0.01717960275709629, + 0.01277529913932085, + 0.017284194007515907, + 0.014073412865400314, + 0.003027112688869238, + 0.00245771836489439, + 0.0028362846933305264, + 0.0031899476889520884, + 0.0033691003918647766, + 0.002749219536781311, + 0.00323601090349257, + 0.0028956725727766752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.014930173754692078 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017227584030479193 + ], + "opposite_logprob_tokens": [ + -6.376722812652588 + ], + "logprob_reward": -0.0017227584030479193, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017227584030479193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1314787119626999, + 0.245727077126503, + 0.14749614894390106, + 0.07876165956258774, + 0.10311540216207504, + 0.041284579783678055, + 0.1191527247428894, + 0.04386959597468376, + 0.050681229680776596, + 0.03843289241194725 + ], + "credit_map": [ + 0.008921115659177303, + 0.007840934209525585, + 0.013702968135476112, + 0.011692334897816181, + 0.011305414140224457, + 0.028638796880841255, + 0.009811094030737877, + 0.03956606239080429, + 0.028163691982626915, + 0.023402683436870575, + 0.019986530765891075, + 0.035723548382520676, + 0.021955056115984917, + 0.04495486244559288, + 0.02162957563996315, + 0.04991114139556885, + 0.010687406174838543, + 0.011959471739828587, + 0.010147259570658207, + 0.018961099907755852, + 0.011447693221271038, + 0.033456671983003616, + 0.016811247915029526, + 0.03402528911828995, + 0.005001444835215807, + 0.008345232345163822, + 0.008553976193070412, + 0.008675014600157738, + 0.009160544723272324, + 0.011958721093833447, + 0.015362156555056572, + 0.011704566888511181, + 0.006999680772423744, + 0.007209849543869495, + 0.00885720457881689, + 0.01096152700483799, + 0.007428938522934914, + 0.026716547086834908, + 0.009088367223739624, + 0.02585328370332718, + 0.0033947655465453863, + 0.004616111982613802, + 0.004366722423583269, + 0.00675747636705637, + 0.007344275712966919, + 0.004809905309230089, + 0.005756913218647242, + 0.00423840805888176, + 0.009227602742612362, + 0.011202273890376091, + 0.01363204512745142, + 0.012579655274748802, + 0.024096108973026276, + 0.019355574622750282, + 0.011901232413947582, + 0.01715823821723461, + 0.003613733686506748, + 0.005732131656259298, + 0.004578161519020796, + 0.004940889310091734, + 0.01042521558701992, + 0.005639716982841492, + 0.0037753537762910128, + 0.005164396949112415, + 0.005256438162177801, + 0.006190864834934473, + 0.004129040986299515, + 0.007984824478626251, + 0.005724652670323849, + 0.006428392603993416, + 0.006402631290256977, + 0.008564382791519165, + 0.0036922006402164698, + 0.0036928809713572264, + 0.0030810399912297726, + 0.003973434213548899, + 0.005637536756694317, + 0.005558240693062544, + 0.005744235124439001, + 0.007053321227431297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0017227584030479193 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028208012226969004 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028208012226969004, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028208012226969004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08229613304138184, + 0.05429168790578842, + 0.04730677604675293, + 0.18785354495048523, + 0.06708875298500061, + 0.1126638650894165, + 0.1499287337064743, + 0.10209682583808899, + 0.15048815310001373, + 0.045985519886016846 + ], + "credit_map": [ + 0.008783905766904354, + 0.005922222509980202, + 0.007184249814599752, + 0.01277607586234808, + 0.009071681648492813, + 0.01266274880617857, + 0.009016926400363445, + 0.016878316178917885, + 0.004795825108885765, + 0.004813699051737785, + 0.0048472392372787, + 0.008892684243619442, + 0.007053067442029715, + 0.008872908540070057, + 0.005278521683067083, + 0.009737745858728886, + 0.0036316995974630117, + 0.003425024449825287, + 0.003278984222561121, + 0.0062356991693377495, + 0.004078769590705633, + 0.01006187591701746, + 0.005715577397495508, + 0.010879145003855228, + 0.01755722425878048, + 0.019421637058258057, + 0.022310134023427963, + 0.02386939711868763, + 0.01739649660885334, + 0.0316147617995739, + 0.022026170045137405, + 0.0336577370762825, + 0.004548520315438509, + 0.007622557692229748, + 0.006660704966634512, + 0.006288111675530672, + 0.006162784993648529, + 0.01636822707951069, + 0.005273581948131323, + 0.014164263382554054, + 0.009534907527267933, + 0.013055972754955292, + 0.010886524803936481, + 0.025964055210351944, + 0.014061025343835354, + 0.011932261288166046, + 0.014990459196269512, + 0.01223866268992424, + 0.010183430276811123, + 0.01316550187766552, + 0.01844041794538498, + 0.019350511953234673, + 0.029771074652671814, + 0.023371621966362, + 0.01134770829230547, + 0.024298472329974174, + 0.006917630787938833, + 0.012085674330592155, + 0.006936721969395876, + 0.009352714754641056, + 0.03004927933216095, + 0.014741945080459118, + 0.009134545922279358, + 0.012878317385911942, + 0.018264397978782654, + 0.01673656329512596, + 0.01156611368060112, + 0.021104246377944946, + 0.016887813806533813, + 0.019791344180703163, + 0.019093412905931473, + 0.02704426273703575, + 0.004094318486750126, + 0.0058580925688147545, + 0.006617418024688959, + 0.004839895758777857, + 0.007855404168367386, + 0.005401487462222576, + 0.005230701994150877, + 0.006088206544518471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.0028208012226969004 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095857031643391 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004095857031643391, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095857031643391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10943838208913803, + 0.057309091091156006, + 0.11149030178785324, + 0.1268414556980133, + 0.13880214095115662, + 0.04744555801153183, + 0.1186891496181488, + 0.07854665070772171, + 0.14018094539642334, + 0.07125639915466309 + ], + "credit_map": [ + 0.007964582182466984, + 0.008890501223504543, + 0.011771638877689838, + 0.013277595862746239, + 0.01596161350607872, + 0.015009874477982521, + 0.016940254718065262, + 0.019622305408120155, + 0.005548951216042042, + 0.004041895270347595, + 0.005080792121589184, + 0.004992302972823381, + 0.006898506544530392, + 0.011762149631977081, + 0.00450580520555377, + 0.014478683471679688, + 0.00854209903627634, + 0.008226333186030388, + 0.007932757027447224, + 0.011325767263770103, + 0.007987186312675476, + 0.028139753267169, + 0.010135280899703503, + 0.029201114550232887, + 0.011742720380425453, + 0.011719711124897003, + 0.01595444604754448, + 0.01335467305034399, + 0.012668883427977562, + 0.022721782326698303, + 0.01613287255167961, + 0.022546350955963135, + 0.009158944711089134, + 0.016455894336104393, + 0.0105125792324543, + 0.01350371539592743, + 0.012595885433256626, + 0.03496783599257469, + 0.00946133490651846, + 0.03214593231678009, + 0.004537325352430344, + 0.004762217868119478, + 0.00545013602823019, + 0.006504308432340622, + 0.008095945231616497, + 0.005483483429998159, + 0.007290679961442947, + 0.005321457050740719, + 0.00779743492603302, + 0.012681670486927032, + 0.013751832768321037, + 0.01395847275853157, + 0.024162573739886284, + 0.018544135615229607, + 0.009903638623654842, + 0.01788938045501709, + 0.0064292228780686855, + 0.007898417301476002, + 0.013259942643344402, + 0.009101007133722305, + 0.013554180040955544, + 0.010386692360043526, + 0.007637208793312311, + 0.010279963724315166, + 0.019206639379262924, + 0.01203568559139967, + 0.010720789432525635, + 0.013560706749558449, + 0.018817488104104996, + 0.019956547766923904, + 0.015102716162800789, + 0.03078034520149231, + 0.006634613499045372, + 0.010123726911842823, + 0.006252800114452839, + 0.007934785448014736, + 0.013949239626526833, + 0.00909424014389515, + 0.008677192963659763, + 0.008589794859290123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.004095857031643391 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539333820343018 + ], + "opposite_logprob_tokens": [ + -0.42893341183662415 + ], + "logprob_reward": -1.0539333820343018, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539333820343018, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2477918267250061, + 0.09011302888393402, + 0.15327419340610504, + 0.17032375931739807, + 0.0685076117515564, + 0.06815933436155319, + 0.05645781755447388, + 0.038305748254060745, + 0.07806135714054108, + 0.029005372896790504 + ], + "credit_map": [ + 0.0188861433416605, + 0.017410721629858017, + 0.026089252904057503, + 0.024162866175174713, + 0.030271507799625397, + 0.046464819461107254, + 0.02155090868473053, + 0.0629555881023407, + 0.007962636649608612, + 0.008371111005544662, + 0.0061102332547307014, + 0.010502700693905354, + 0.009728115051984787, + 0.018973523750901222, + 0.007428806275129318, + 0.021035900339484215, + 0.010841228999197483, + 0.009779136627912521, + 0.007770984899252653, + 0.013090207241475582, + 0.010514682158827782, + 0.04421169310808182, + 0.01241703238338232, + 0.04464924335479736, + 0.01187213510274887, + 0.016027770936489105, + 0.022484153509140015, + 0.02557573840022087, + 0.01737925410270691, + 0.027745598927140236, + 0.022474611178040504, + 0.026764506474137306, + 0.005649313796311617, + 0.006642454769462347, + 0.008155936375260353, + 0.009194862097501755, + 0.006580795627087355, + 0.01283169724047184, + 0.007505924440920353, + 0.011946634389460087, + 0.006452963687479496, + 0.009294372983276844, + 0.00753480801358819, + 0.010930592194199562, + 0.009318900294601917, + 0.007894853129982948, + 0.008927674032747746, + 0.007805165834724903, + 0.004941948689520359, + 0.007478541694581509, + 0.006668217480182648, + 0.006892388686537743, + 0.008904819376766682, + 0.007406488060951233, + 0.007227250840514898, + 0.006938163191080093, + 0.002850712276995182, + 0.004348117392510176, + 0.004945400636643171, + 0.0039267772808671, + 0.008893412537872791, + 0.00511810090392828, + 0.0034384673926979303, + 0.004784759134054184, + 0.01113867200911045, + 0.006173519883304834, + 0.005714758764952421, + 0.006093702744692564, + 0.008881289511919022, + 0.011603426188230515, + 0.007609638385474682, + 0.020846351981163025, + 0.003279661061242223, + 0.002676015254110098, + 0.0028568434063345194, + 0.00383287132717669, + 0.0040872120298445225, + 0.003916630521416664, + 0.0037010444793850183, + 0.00465509295463562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -1.0539333820343018 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, upward-focused shot draws us into a boundless, pure Klein blue abyss, devoid of visible rocks or seabed, dominated by absolute silence and emptiness. At the center of the frame, a massive branching jellyfish hangs suspended within it, a complex organism composed of countless luminous units. Its translucent, chain-like structures emanate a soft, inner glow of white and pale cyan, contrasting sharply against the deep blue backdrop. A silent, Abstract Expressionist explosion of life unfolds at its core: Shot in stunning slow motion, countless shimmering golden and silver microparticles, like awakened cosmic dust, gracefully and slowly erupt and disperse from its gelatinous mass. These particles, far from a chaotic mess, form spiraling, dreamlike galaxies that dance and diffuse through the still waters, creating a vast, dynamic canvas filled with the miracle of life. The brilliant light emitted by the jellyfish itself and the particles is the only light source, dyeing the surrounding Klein blue water with a mysterious halo. The whole scene is filled with the solemnity, tranquility and surreal dreamlike feeling of the birth of the universe.", + "seed": 33407, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_9/index_1_seed_33407_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20040021762251853, + "vlm_logprob": -0.20040021762251853, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.20040021762251853, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008257297798991203 + ], + "opposite_logprob_tokens": [ + -5.0082573890686035 + ], + "logprob_reward": -0.008257297798991203, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008257297798991203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20636974275112152, + 0.09192800521850586, + 0.07082850486040115, + 0.06276167929172516, + 0.07087373733520508, + 0.2598212659358978, + 0.02539656311273575, + 0.08323928713798523, + 0.07226147502660751, + 0.056519780308008194 + ], + "credit_map": [ + 0.015514940023422241, + 0.017235809937119484, + 0.04811382293701172, + 0.01703227497637272, + 0.01881064660847187, + 0.015081395395100117, + 0.022849716246128082, + 0.051731135696172714, + 0.005797206424176693, + 0.012489426881074905, + 0.013256029225885868, + 0.011686003766953945, + 0.008693250827491283, + 0.01242272648960352, + 0.013326249085366726, + 0.014257112517952919, + 0.004850425757467747, + 0.010299228131771088, + 0.010771188884973526, + 0.00809727143496275, + 0.010114832781255245, + 0.007025261875241995, + 0.009563312865793705, + 0.010106977075338364, + 0.006253108847886324, + 0.006078848149627447, + 0.007408006116747856, + 0.006339636165648699, + 0.007382380776107311, + 0.009002196602523327, + 0.010997200384736061, + 0.00930030271410942, + 0.005510107614099979, + 0.009355170652270317, + 0.007163028232753277, + 0.011540533043444157, + 0.009390778839588165, + 0.006162532605230808, + 0.012522817589342594, + 0.009228772483766079, + 0.01647530123591423, + 0.0854315385222435, + 0.01180683821439743, + 0.013290222734212875, + 0.0405929759144783, + 0.01431956049054861, + 0.017893048003315926, + 0.06001177802681923, + 0.003086177632212639, + 0.0024442635476589203, + 0.0032532576005905867, + 0.0033928563352674246, + 0.0029817214235663414, + 0.0032226659823209047, + 0.004505331162363291, + 0.0025102875661104918, + 0.010220468044281006, + 0.008376041427254677, + 0.011057405732572079, + 0.013215224258601665, + 0.008783453144133091, + 0.009298298507928848, + 0.01238727755844593, + 0.00990111380815506, + 0.009939971379935741, + 0.010528278537094593, + 0.011689803563058376, + 0.008030499331653118, + 0.009992823004722595, + 0.0056719062849879265, + 0.011540073901414871, + 0.004868127405643463, + 0.014826202765107155, + 0.003232394577935338, + 0.011921918019652367, + 0.00519764656201005, + 0.004426232073456049, + 0.004137441981583834, + 0.00809092354029417, + 0.004687019623816013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.008257297798991203 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028259125538170338 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028259125538170338, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028259125538170338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10678209364414215, + 0.07606727629899979, + 0.13098616898059845, + 0.05615924298763275, + 0.0699695274233818, + 0.12331822514533997, + 0.14076510071754456, + 0.07147268205881119, + 0.1107136532664299, + 0.11376605182886124 + ], + "credit_map": [ + 0.009990458376705647, + 0.011620357632637024, + 0.01467969361692667, + 0.011162647977471352, + 0.01528916321694851, + 0.013079195283353329, + 0.012046131305396557, + 0.018914442509412766, + 0.005977586843073368, + 0.008666745387017727, + 0.012424432672560215, + 0.008086062036454678, + 0.009489323012530804, + 0.01065176259726286, + 0.00705103762447834, + 0.01372032891958952, + 0.012272180058062077, + 0.03328859806060791, + 0.01276805903762579, + 0.011644396930932999, + 0.024151841178536415, + 0.009953505359590054, + 0.011204631999135017, + 0.015702970325946808, + 0.009136329405009747, + 0.005967705510556698, + 0.005329736974090338, + 0.005402261391282082, + 0.006831357255578041, + 0.008234748616814613, + 0.007911182940006256, + 0.0073459213599562645, + 0.006020062603056431, + 0.009004074148833752, + 0.007204925175756216, + 0.0122779356315732, + 0.011050212197005749, + 0.00674982788041234, + 0.00925015565007925, + 0.008412332274019718, + 0.00887630321085453, + 0.03211049363017082, + 0.009646068327128887, + 0.01048412173986435, + 0.018544018268585205, + 0.00855262391269207, + 0.013121718540787697, + 0.02198287472128868, + 0.011089878156781197, + 0.017662344500422478, + 0.021229565143585205, + 0.017370134592056274, + 0.0183203537017107, + 0.009922255761921406, + 0.031798526644706726, + 0.013372039422392845, + 0.00919037964195013, + 0.007522130385041237, + 0.00984811782836914, + 0.008989815600216389, + 0.008509690873324871, + 0.00677550258114934, + 0.012111236341297626, + 0.008525813929736614, + 0.0164570901542902, + 0.010277364403009415, + 0.01875162310898304, + 0.012322230264544487, + 0.013951078988611698, + 0.011266331188380718, + 0.015411743894219398, + 0.012276188470423222, + 0.017918743193149567, + 0.01276245154440403, + 0.01536695659160614, + 0.009666197001934052, + 0.014632459729909897, + 0.013517792336642742, + 0.019121814519166946, + 0.010779629461467266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.0028259125538170338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535941787064075 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012535941787064075, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535941787064075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.071529820561409, + 0.052911654114723206, + 0.12149102985858917, + 0.03650840371847153, + 0.036080095916986465, + 0.42681819200515747, + 0.03674713149666786, + 0.07840366661548615, + 0.047466207295656204, + 0.09204371273517609 + ], + "credit_map": [ + 0.006635709200054407, + 0.010844224132597446, + 0.009355047717690468, + 0.008512139320373535, + 0.009263069368898869, + 0.007848955690860748, + 0.009634897112846375, + 0.009435780346393585, + 0.003980980720371008, + 0.0059960330836474895, + 0.00837346538901329, + 0.005793025251477957, + 0.006687132641673088, + 0.007770764175802469, + 0.006674191914498806, + 0.007636059541255236, + 0.009692343883216381, + 0.01438046433031559, + 0.01723310351371765, + 0.013706858269870281, + 0.017240548506379128, + 0.018395772203803062, + 0.015846218913793564, + 0.014995735138654709, + 0.005393762607127428, + 0.003534839255735278, + 0.003449117299169302, + 0.004095619544386864, + 0.004223146941512823, + 0.0032271770760416985, + 0.006715049967169762, + 0.005869689863175154, + 0.002417500363662839, + 0.005334824323654175, + 0.0029910034500062466, + 0.004897783976048231, + 0.004685650113970041, + 0.0033146697096526623, + 0.00667489692568779, + 0.0057637677527964115, + 0.028841866180300713, + 0.1382906138896942, + 0.020850159227848053, + 0.019475383684039116, + 0.06496243178844452, + 0.024060046300292015, + 0.030674947425723076, + 0.09966275840997696, + 0.0029672663658857346, + 0.006413439754396677, + 0.005061788484454155, + 0.004001070745289326, + 0.004919532686471939, + 0.005409743171185255, + 0.004405491519719362, + 0.003568800166249275, + 0.010643108747899532, + 0.01167233195155859, + 0.010533954948186874, + 0.010000737383961678, + 0.011819199658930302, + 0.006901378743350506, + 0.01030020508915186, + 0.006532751023769379, + 0.009628355503082275, + 0.006352482363581657, + 0.009066901169717312, + 0.003520707134157419, + 0.0055374871008098125, + 0.0039467099122703075, + 0.005227156449109316, + 0.004186409059911966, + 0.01612468622624874, + 0.008029150776565075, + 0.012409351766109467, + 0.008963060565292835, + 0.011424385011196136, + 0.009722993709146976, + 0.015665914863348007, + 0.00970416609197855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.012535941787064075 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095975775271654 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004095975775271654, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095975775271654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08995208144187927, + 0.12780168652534485, + 0.12011940032243729, + 0.06428994983434677, + 0.05038481205701828, + 0.13334225118160248, + 0.07981571555137634, + 0.11347272992134094, + 0.1168651282787323, + 0.10395623743534088 + ], + "credit_map": [ + 0.00898913573473692, + 0.010961812920868397, + 0.012488879263401031, + 0.00940872821956873, + 0.010308780707418919, + 0.008408088237047195, + 0.012037502601742744, + 0.017349164932966232, + 0.010403215885162354, + 0.013544579036533833, + 0.020913787186145782, + 0.015222828835248947, + 0.015078955329954624, + 0.01736123487353325, + 0.013362371362745762, + 0.021914729848504066, + 0.01149419043213129, + 0.01837543584406376, + 0.014376536943018436, + 0.013737034983932972, + 0.020784642547369003, + 0.011734268628060818, + 0.012894710525870323, + 0.01672258786857128, + 0.008435331284999847, + 0.00598871149122715, + 0.007732481230050325, + 0.00797309260815382, + 0.007201972883194685, + 0.005005793180316687, + 0.013015377335250378, + 0.008937200531363487, + 0.003934476524591446, + 0.007121041417121887, + 0.004951196722686291, + 0.006182776764035225, + 0.009109416045248508, + 0.0050910599529743195, + 0.006859336048364639, + 0.007135514635592699, + 0.00932250265032053, + 0.03867194056510925, + 0.00862039066851139, + 0.00768709322437644, + 0.019001958891749382, + 0.007936718873679638, + 0.011251825839281082, + 0.030849825590848923, + 0.004981381818652153, + 0.01155681163072586, + 0.009391596540808678, + 0.009946070611476898, + 0.012206817045807838, + 0.006745056249201298, + 0.016618603840470314, + 0.008369388058781624, + 0.016927292570471764, + 0.012923836708068848, + 0.013429855927824974, + 0.01880914717912674, + 0.015041051432490349, + 0.009734374471008778, + 0.017212824895977974, + 0.009394356980919838, + 0.017967160791158676, + 0.013472063466906548, + 0.016380632296204567, + 0.009775933809578419, + 0.015034387819468975, + 0.012384012341499329, + 0.019151976332068443, + 0.012698975391685963, + 0.018433179706335068, + 0.010691783390939236, + 0.013888237997889519, + 0.00891765858978033, + 0.013330375775694847, + 0.011366194114089012, + 0.017709387466311455, + 0.009619427844882011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.004095975775271654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742859601974487 + ], + "opposite_logprob_tokens": [ + -0.47428593039512634 + ], + "logprob_reward": -0.9742859601974487, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742859601974487, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1552218645811081, + 0.06688252091407776, + 0.12358725070953369, + 0.04493752121925354, + 0.06228649616241455, + 0.21247521042823792, + 0.0768740326166153, + 0.05189567059278488, + 0.10549543052911758, + 0.10034406185150146 + ], + "credit_map": [ + 0.009718847461044788, + 0.010654719546437263, + 0.035180915147066116, + 0.016227325424551964, + 0.012153934687376022, + 0.013537129387259483, + 0.02473735623061657, + 0.03301163390278816, + 0.004995627794414759, + 0.008389553986489773, + 0.009128286503255367, + 0.007562797516584396, + 0.010105449706315994, + 0.009280800819396973, + 0.008845597505569458, + 0.008574404753744602, + 0.01164623349905014, + 0.029448693618178368, + 0.01667303405702114, + 0.011562570929527283, + 0.018742965534329414, + 0.010236786678433418, + 0.010054638609290123, + 0.01522232498973608, + 0.004085686523467302, + 0.00440340768545866, + 0.0055854301899671555, + 0.005148841999471188, + 0.004964854102581739, + 0.005976565647870302, + 0.008629330433905125, + 0.00614340091124177, + 0.005501792766153812, + 0.00814063847064972, + 0.00660580163821578, + 0.00807578768581152, + 0.007335068192332983, + 0.007086186204105616, + 0.010711034759879112, + 0.008830184116959572, + 0.015296422876417637, + 0.0657946914434433, + 0.011695553548634052, + 0.012154292315244675, + 0.031700994819402695, + 0.012762106955051422, + 0.015345923602581024, + 0.04772520437836647, + 0.005219468846917152, + 0.016717055812478065, + 0.013989534229040146, + 0.008379815146327019, + 0.011207562871277332, + 0.006095296237617731, + 0.009077735245227814, + 0.00618756515905261, + 0.007198668550699949, + 0.005594133399426937, + 0.008478185161948204, + 0.007062958087772131, + 0.0049589467234909534, + 0.005813660100102425, + 0.0070693292655050755, + 0.005719789303839207, + 0.014333641156554222, + 0.009095552377402782, + 0.01482605841010809, + 0.011879809200763702, + 0.017647653818130493, + 0.010857602581381798, + 0.017929434776306152, + 0.008925684727728367, + 0.023761345073580742, + 0.008435939438641071, + 0.0183575339615345, + 0.008840598165988922, + 0.008879314176738262, + 0.00967981293797493, + 0.014157247729599476, + 0.008232267573475838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.9742859601974487 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, upward-focused shot draws us into a boundless, pure Klein blue abyss, devoid of visible rocks or seabed, dominated by absolute silence and emptiness. At the center of the frame, a massive branching jellyfish hangs suspended within it, a complex organism composed of countless luminous units. Its translucent, chain-like structures emanate a soft, inner glow of white and pale cyan, contrasting sharply against the deep blue backdrop. A silent, Abstract Expressionist explosion of life unfolds at its core: Shot in stunning slow motion, countless shimmering golden and silver microparticles, like awakened cosmic dust, gracefully and slowly erupt and disperse from its gelatinous mass. These particles, far from a chaotic mess, form spiraling, dreamlike galaxies that dance and diffuse through the still waters, creating a vast, dynamic canvas filled with the miracle of life. The brilliant light emitted by the jellyfish itself and the particles is the only light source, dyeing the surrounding Klein blue water with a mysterious halo. The whole scene is filled with the solemnity, tranquility and surreal dreamlike feeling of the birth of the universe.", + "seed": 33408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000067/group_9/index_1_seed_33408_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.21603925414383413, + "vlm_logprob": -0.21603925414383413, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.21603925414383413, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013360025361180305 + ], + "opposite_logprob_tokens": [ + -4.388360023498535 + ], + "logprob_reward": -0.013360025361180305, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013360025361180305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30024203658103943, + 0.07220885902643204, + 0.052172280848026276, + 0.12029603123664856, + 0.05657990649342537, + 0.03225470334291458, + 0.2535506784915924, + 0.027519427239894867, + 0.030711032450199127, + 0.054465118795633316 + ], + "credit_map": [ + 0.025904064998030663, + 0.03634677827358246, + 0.014946778304874897, + 0.020261093974113464, + 0.10141443461179733, + 0.01785375364124775, + 0.013049639761447906, + 0.07046543806791306, + 0.009865343570709229, + 0.007228875532746315, + 0.012630275450646877, + 0.009277119301259518, + 0.007462008390575647, + 0.0058365026488900185, + 0.008999337442219257, + 0.01090939436107874, + 0.007393622305244207, + 0.00442101713269949, + 0.010888135991990566, + 0.005472676362842321, + 0.0063302647322416306, + 0.006369518581777811, + 0.004256037995219231, + 0.007041007746011019, + 0.022509774193167686, + 0.02780519798398018, + 0.006316248793154955, + 0.012219328433275223, + 0.017172707244753838, + 0.007583309896290302, + 0.006429374683648348, + 0.020260069519281387, + 0.006241121795028448, + 0.009939484298229218, + 0.006136568728834391, + 0.007130299229174852, + 0.007579102646559477, + 0.006308670621365309, + 0.005736451130360365, + 0.007508204784244299, + 0.0026220993604511023, + 0.004863424226641655, + 0.0023398795165121555, + 0.0037610488943755627, + 0.003805963322520256, + 0.003254637587815523, + 0.007021609228104353, + 0.00458604097366333, + 0.024202274158596992, + 0.019915781915187836, + 0.016049573197960854, + 0.01853213459253311, + 0.06396357715129852, + 0.02389991655945778, + 0.032726727426052094, + 0.05426066741347313, + 0.0037960028275847435, + 0.00431640213355422, + 0.0035872412845492363, + 0.003165655070915818, + 0.003306726459413767, + 0.003611645195633173, + 0.0028814715333282948, + 0.0028542792424559593, + 0.0033591545652598143, + 0.004361120983958244, + 0.0045990231446921825, + 0.003537873737514019, + 0.00414600595831871, + 0.00329924956895411, + 0.004288850352168083, + 0.0031197508797049522, + 0.006820407230407, + 0.008322486653923988, + 0.005960284732282162, + 0.004885176662355661, + 0.0065492601133883, + 0.007299985736608505, + 0.0068321931175887585, + 0.0077953217551112175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.013360025361180305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046440837904810905 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.0046440837904810905, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046440837904810905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5258907675743103, + 0.02699693851172924, + 0.05413242429494858, + 0.06786922365427017, + 0.0477478951215744, + 0.05708707123994827, + 0.07993888854980469, + 0.01728806085884571, + 0.05207090824842453, + 0.07097781449556351 + ], + "credit_map": [ + 0.04476270079612732, + 0.06391660124063492, + 0.01717947982251644, + 0.03200196847319603, + 0.17839550971984863, + 0.02792336419224739, + 0.02218182012438774, + 0.139529287815094, + 0.002571774646639824, + 0.002966593485325575, + 0.0033382319379597902, + 0.004041105043143034, + 0.0033258518669754267, + 0.0025032206904143095, + 0.004604853689670563, + 0.0036453085485845804, + 0.003698585322126746, + 0.005761748179793358, + 0.006131251808255911, + 0.0058907088823616505, + 0.009150671772658825, + 0.0112976199015975, + 0.005717317573726177, + 0.006484519690275192, + 0.008830602280795574, + 0.01037883572280407, + 0.007982647977769375, + 0.011397277936339378, + 0.008600126020610332, + 0.006352765951305628, + 0.005623727571219206, + 0.008703243918716908, + 0.0052205598913133144, + 0.00870231632143259, + 0.0033384195994585752, + 0.005781902000308037, + 0.007108354941010475, + 0.005394714418798685, + 0.006757085677236319, + 0.005444540176540613, + 0.004948463756591082, + 0.006315888371318579, + 0.005203445442020893, + 0.011475156992673874, + 0.007424125913530588, + 0.008218567818403244, + 0.007903954945504665, + 0.005597470793873072, + 0.007247858680784702, + 0.007630467414855957, + 0.005009160842746496, + 0.006981083657592535, + 0.020646410062909126, + 0.008176523260772228, + 0.00809632707387209, + 0.016151057556271553, + 0.0019435222493484616, + 0.0018988188821822405, + 0.0014700975734740496, + 0.001964310649782419, + 0.0028313014190644026, + 0.002154827583581209, + 0.0026526579167693853, + 0.0023725249338895082, + 0.0046708290465176105, + 0.008002142421901226, + 0.006615249440073967, + 0.0060446178540587425, + 0.008962739259004593, + 0.0058900946751236916, + 0.006999213248491287, + 0.004886021837592125, + 0.011771570891141891, + 0.007909540086984634, + 0.006511290092021227, + 0.005589893553406, + 0.00762697821483016, + 0.009448427706956863, + 0.00647710170596838, + 0.015643011778593063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0046440837904810905 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463898153975606 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463898153975606, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463898153975606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4119836688041687, + 0.04594993218779564, + 0.04962507635354996, + 0.14439459145069122, + 0.058974724262952805, + 0.1446661502122879, + 0.042460039258003235, + 0.019160090014338493, + 0.029623663052916527, + 0.0531620979309082 + ], + "credit_map": [ + 0.03543435409665108, + 0.04975096881389618, + 0.016548113897442818, + 0.02455267310142517, + 0.13908669352531433, + 0.023068981245160103, + 0.017040226608514786, + 0.10650164633989334, + 0.005282169673591852, + 0.004627867601811886, + 0.009516608901321888, + 0.005559436045587063, + 0.004074709489941597, + 0.004401703830808401, + 0.0052984547801315784, + 0.007188975345343351, + 0.004091453738510609, + 0.005112672224640846, + 0.006672185845673084, + 0.007594312075525522, + 0.008264387026429176, + 0.0067663006484508514, + 0.005143960006535053, + 0.005979797802865505, + 0.023664236068725586, + 0.04165009409189224, + 0.008676082827150822, + 0.017821935936808586, + 0.018630318343639374, + 0.010239269584417343, + 0.008407543413341045, + 0.015305093489587307, + 0.006228267215192318, + 0.0093425577506423, + 0.004893071949481964, + 0.007016716990619898, + 0.0094135832041502, + 0.008380348794162273, + 0.0061549595557153225, + 0.007545210421085358, + 0.02051411010324955, + 0.013547278940677643, + 0.011825366877019405, + 0.016024593263864517, + 0.040280405431985855, + 0.01228481438010931, + 0.01914820447564125, + 0.011041364632546902, + 0.004487297497689724, + 0.006153062451630831, + 0.0046312143094837666, + 0.005686463322490454, + 0.004970618523657322, + 0.0062796794809401035, + 0.005217810161411762, + 0.005033891648054123, + 0.0016873970162123442, + 0.002893463708460331, + 0.0024587158113718033, + 0.002484166994690895, + 0.0027444083243608475, + 0.003327438607811928, + 0.00202527386136353, + 0.0015392234781756997, + 0.002641688333824277, + 0.004858296364545822, + 0.003469489049166441, + 0.003035388421267271, + 0.005010961089283228, + 0.0034195375628769398, + 0.0039463345892727375, + 0.0032419636845588684, + 0.006156882271170616, + 0.007092084735631943, + 0.006907422561198473, + 0.006105336826294661, + 0.006071816198527813, + 0.005985788535326719, + 0.006406335160136223, + 0.008436428382992744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.00463898153975606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036168177612125874 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036168177612125874, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036168177612125874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14198489487171173, + 0.05255753546953201, + 0.05602659285068512, + 0.16635753214359283, + 0.13858941197395325, + 0.12106910347938538, + 0.058659400790929794, + 0.04221062734723091, + 0.07351169735193253, + 0.14903317391872406 + ], + "credit_map": [ + 0.013491139747202396, + 0.013673131354153156, + 0.010225540027022362, + 0.012009740807116032, + 0.039869386702775955, + 0.012176411226391792, + 0.0073549626395106316, + 0.03318459168076515, + 0.006392201408743858, + 0.004923837259411812, + 0.007993949577212334, + 0.00672060577198863, + 0.005871776957064867, + 0.006431271322071552, + 0.00719395512714982, + 0.00702994130551815, + 0.004861699882894754, + 0.005263043101876974, + 0.005287281237542629, + 0.008137545548379421, + 0.008916940540075302, + 0.008643866516649723, + 0.007241661194711924, + 0.007674556691199541, + 0.029913440346717834, + 0.037208691239356995, + 0.017784735187888145, + 0.019376177340745926, + 0.01865674927830696, + 0.013338851742446423, + 0.009619668126106262, + 0.02045922726392746, + 0.014643605798482895, + 0.023591049015522003, + 0.016711710020899773, + 0.017203234136104584, + 0.018031705170869827, + 0.019010402262210846, + 0.01725882850587368, + 0.012138892896473408, + 0.015551025979220867, + 0.008952477015554905, + 0.009082785807549953, + 0.010306826792657375, + 0.034102097153663635, + 0.011356601491570473, + 0.021371591836214066, + 0.010345710441470146, + 0.005400235764682293, + 0.004774890840053558, + 0.005452703218907118, + 0.0059944214299321175, + 0.009120174683630466, + 0.006571274250745773, + 0.010365523397922516, + 0.010980184189975262, + 0.003912265412509441, + 0.00452386261895299, + 0.004160902462899685, + 0.005669896025210619, + 0.008178346790373325, + 0.0051596625708043575, + 0.0062671685591340065, + 0.004338528495281935, + 0.007922885939478874, + 0.01142649445682764, + 0.0070287855342030525, + 0.008604631759226322, + 0.011407102458178997, + 0.007864577695727348, + 0.01216974575072527, + 0.007087480276823044, + 0.015406126156449318, + 0.01647384837269783, + 0.013332509435713291, + 0.019195372238755226, + 0.017632488161325455, + 0.019949302077293396, + 0.015335572883486748, + 0.031707968562841415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.0036168177612125874 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a boundless Klein blue abyss without visible rocks or seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive branching jellyfish with translucent, chain-like structures visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless shimmering golden and silver microparticles erupt and disperse from the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the microparticles form spiraling patterns that dance and diffuse through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot in slow motion with the jellyfish and particles as the only light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539363622665405 + ], + "opposite_logprob_tokens": [ + -0.4289363622665405 + ], + "logprob_reward": -1.0539363622665405, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539363622665405, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6644738912582397, + 0.02664756216108799, + 0.02692483365535736, + 0.0984615758061409, + 0.02297455258667469, + 0.022267749533057213, + 0.020337576046586037, + 0.029377005994319916, + 0.016856323927640915, + 0.07167893648147583 + ], + "credit_map": [ + 0.05540226772427559, + 0.08358623087406158, + 0.023743590340018272, + 0.03880847617983818, + 0.2309752255678177, + 0.03726475313305855, + 0.026639657095074654, + 0.16805368661880493, + 0.0036887903697788715, + 0.0033944074530154467, + 0.004495244938880205, + 0.0029102854896336794, + 0.002601374639198184, + 0.0019990974105894566, + 0.003270574612542987, + 0.0042877900414168835, + 0.0020115431398153305, + 0.002500350121408701, + 0.003228087443858385, + 0.0035113149788230658, + 0.004554863087832928, + 0.004318284336477518, + 0.0031271378975361586, + 0.0036732538137584925, + 0.017179317772388458, + 0.024897387251257896, + 0.00611974997445941, + 0.009313230402767658, + 0.015600868500769138, + 0.005723205395042896, + 0.004862478468567133, + 0.01476533617824316, + 0.002811507787555456, + 0.004588541109114885, + 0.002340510254725814, + 0.0036153185646981, + 0.0032790806144475937, + 0.001625783508643508, + 0.0025208438746631145, + 0.0021929668728262186, + 0.0032467509154230356, + 0.0020433191675692797, + 0.0014077083906158805, + 0.001865455531515181, + 0.006403908599168062, + 0.002719088224694133, + 0.0027355882339179516, + 0.0018459312850609422, + 0.001942634698934853, + 0.0022536145988851786, + 0.0020048513542860746, + 0.002637552795931697, + 0.003738369792699814, + 0.0021013475488871336, + 0.0024964024778455496, + 0.003162801032885909, + 0.004249378573149443, + 0.004753049463033676, + 0.0033478373661637306, + 0.003918184898793697, + 0.0036016625817865133, + 0.003793086390942335, + 0.003105232957750559, + 0.0026085730642080307, + 0.0021751411259174347, + 0.0025363683234900236, + 0.0026737882290035486, + 0.0016398527659475803, + 0.0022935608867555857, + 0.0019609236624091864, + 0.0020048152655363083, + 0.00157187320291996, + 0.009110293351113796, + 0.008439763449132442, + 0.00779998442158103, + 0.007786326110363007, + 0.008756140246987343, + 0.008086886256933212, + 0.007855009287595749, + 0.013844538480043411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -1.0539363622665405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_0/index_tensor([306, 306])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_0/index_tensor([306, 306])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6ebc40a1c30596365f210e3fd469947d268d72e0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_0/index_tensor([306, 306])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely wide shot, positioned at a low angle, moves slowly forward along the center of a vast, empty urban canyon, creating a sense of steady, relentless movement. Towering skyscrapers of dark, reflective glass and polished steel rise on either side, their smooth, minimalist facades piercing a clear, late-afternoon sky. The black asphalt, still damp and shimmering from a recent rain, stretches like a dark, empty river. In the exact center of this vast emptiness, a lone, distant figure walks directly away from the camera in a steady, unwavering rhythm. This figure is a sharply defined individual, clad in a long, dark coat, their form framed by the golden light between the buildings. They advance with a deliberate gait, each step measured and consistent, hands in their pockets, plunging deep into the urban void. The low-hanging sun casts long, dramatic shadows that stretch across the sidewalk, creating a stark visual interplay of warm, bright light and cool, deep shadows. The atmosphere is one of profound solitude and awe, a cinematic and contemplative one, where the sheer size of the quiet modern metropolis dwarfs the single, enduring human element within it.", + "seed": 30904, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_0/index_306_seed_30904_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010368426819331944, + "vlm_logprob": -0.010368426819331944, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.010368426819331944, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02085985615849495 + ], + "opposite_logprob_tokens": [ + -4.520859718322754 + ], + "logprob_reward": -0.02085985615849495, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.02085985615849495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34891706705093384, + 0.08923953026533127, + 0.05515193194150925, + 0.02660573460161686, + 0.025063946843147278, + 0.024130599573254585, + 0.28261181712150574, + 0.06490053236484528, + 0.05452334135770798, + 0.028855562210083008 + ], + "credit_map": [ + 0.044107936322689056, + 0.027394430711865425, + 0.06161314621567726, + 0.0640338584780693, + 0.018198544159531593, + 0.040073711425065994, + 0.03866935148835182, + 0.05482605844736099, + 0.011591458693146706, + 0.004905456677079201, + 0.016536811366677284, + 0.01939355581998825, + 0.004554610699415207, + 0.006988275330513716, + 0.011100344359874725, + 0.01416900847107172, + 0.005858118180185556, + 0.00501842238008976, + 0.008862776681780815, + 0.010039028711616993, + 0.00401741499081254, + 0.00543642183765769, + 0.00616214657202363, + 0.00975759793072939, + 0.0037011154927313328, + 0.0023301015608012676, + 0.003312820801511407, + 0.003901195479556918, + 0.002237041713669896, + 0.0024386346340179443, + 0.0036007233429700136, + 0.005084099248051643, + 0.003679127898067236, + 0.001689269789494574, + 0.003991445992141962, + 0.004428173881024122, + 0.001784966909326613, + 0.0023866784758865833, + 0.0026984750293195248, + 0.004405805841088295, + 0.0032556902151554823, + 0.0021739257499575615, + 0.0045238700695335865, + 0.0031921030022203922, + 0.0019390609813854098, + 0.002751290099695325, + 0.002661163220182061, + 0.003633493324741721, + 0.05344775691628456, + 0.015736442059278488, + 0.0510115772485733, + 0.04268960282206535, + 0.02103731594979763, + 0.030847150832414627, + 0.02170577459037304, + 0.046136174350976944, + 0.006582318339496851, + 0.004218734335154295, + 0.01346446294337511, + 0.00688951788470149, + 0.0038556600920856, + 0.010089969262480736, + 0.005986639764159918, + 0.013813230209052563, + 0.006202130112797022, + 0.004676347598433495, + 0.009509087540209293, + 0.007537518162280321, + 0.003340460592880845, + 0.005294686648994684, + 0.005132440011948347, + 0.012830664403736591, + 0.0029383953660726547, + 0.001959038432687521, + 0.006506768055260181, + 0.004469048697501421, + 0.002337090903893113, + 0.0018142054323107004, + 0.003926674835383892, + 0.004904339089989662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.02085985615849495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067444778978824615 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.0067444778978824615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067444778978824615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11950431019067764, + 0.06090231612324715, + 0.03566621616482735, + 0.03460516408085823, + 0.035997889935970306, + 0.042907118797302246, + 0.4705452024936676, + 0.1335223764181137, + 0.04366055876016617, + 0.0226888544857502 + ], + "credit_map": [ + 0.013879715465009212, + 0.010392209514975548, + 0.016834525391459465, + 0.02000158093869686, + 0.0072256457060575485, + 0.01129490789026022, + 0.016985369846224785, + 0.022890346124768257, + 0.008780176751315594, + 0.004734409973025322, + 0.012985356152057648, + 0.010521524585783482, + 0.0038278785068541765, + 0.005141737870872021, + 0.007378073409199715, + 0.007533159106969833, + 0.005060022231191397, + 0.0025732717476785183, + 0.006499957758933306, + 0.007370768114924431, + 0.0023843201342970133, + 0.002458718605339527, + 0.0029859303031116724, + 0.006333229597657919, + 0.0051124426536262035, + 0.002860108157619834, + 0.0048987469635903835, + 0.005287099163979292, + 0.0023913676850497723, + 0.0031085058581084013, + 0.0031210528686642647, + 0.007825843058526516, + 0.00609759334474802, + 0.003001750446856022, + 0.0062038227915763855, + 0.004638108890503645, + 0.002538849599659443, + 0.0036502669099718332, + 0.004318175371736288, + 0.005549321882426739, + 0.007757428102195263, + 0.004317863378673792, + 0.007016507443040609, + 0.0060475957579910755, + 0.002809103112667799, + 0.0041345227509737015, + 0.003513880306854844, + 0.007310217712074518, + 0.09570490568876266, + 0.023426024243235588, + 0.08429769426584244, + 0.06341122835874557, + 0.03454424440860748, + 0.05267525464296341, + 0.035691238939762115, + 0.08079461008310318, + 0.015740199014544487, + 0.011617384850978851, + 0.030352039262652397, + 0.013980455696582794, + 0.00690031610429287, + 0.016620147973299026, + 0.012721001170575619, + 0.02559085004031658, + 0.006300569046288729, + 0.0034177328925579786, + 0.007725980598479509, + 0.004778540227562189, + 0.0032751040998846292, + 0.00392942363396287, + 0.00451991893351078, + 0.009713287465274334, + 0.002052870811894536, + 0.0014492844929918647, + 0.005237052217125893, + 0.003679947927594185, + 0.0011245046043768525, + 0.0022074736189097166, + 0.0020588261540979147, + 0.004878893028944731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0067444778978824615 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017147850012406707 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017147850012406707, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017147850012406707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41731876134872437, + 0.04953482747077942, + 0.03799181804060936, + 0.047337424010038376, + 0.07594264298677444, + 0.0730460062623024, + 0.07001703977584839, + 0.09160557389259338, + 0.06229583919048309, + 0.0749099925160408 + ], + "credit_map": [ + 0.045010704547166824, + 0.04254445433616638, + 0.06836766749620438, + 0.0788874551653862, + 0.03235843777656555, + 0.056554991751909256, + 0.03576122969388962, + 0.05783385783433914, + 0.00578333018347621, + 0.005007739644497633, + 0.00800345465540886, + 0.009475264698266983, + 0.004351115319877863, + 0.006376746576279402, + 0.006162310019135475, + 0.004374873358756304, + 0.0051893251948058605, + 0.004159429110586643, + 0.006208841223269701, + 0.006165061146020889, + 0.0032485274132341146, + 0.004815235733985901, + 0.0037220760714262724, + 0.004483326803892851, + 0.005410202778875828, + 0.004379017278552055, + 0.006794915068894625, + 0.008165821433067322, + 0.0035573493223637342, + 0.005357769783586264, + 0.0049677882343530655, + 0.008704563602805138, + 0.01544320210814476, + 0.006486911326646805, + 0.011403031647205353, + 0.010614322498440742, + 0.004324364010244608, + 0.00696823513135314, + 0.008357911370694637, + 0.012344668619334698, + 0.010093447752296925, + 0.005504248663783073, + 0.012342557311058044, + 0.009284492582082748, + 0.004532978404313326, + 0.009737045504152775, + 0.010104524902999401, + 0.011446715332567692, + 0.015173042193055153, + 0.005181161221116781, + 0.012131557799875736, + 0.00675800209864974, + 0.0038398702163249254, + 0.00994967482984066, + 0.005143273621797562, + 0.011840464547276497, + 0.010753929615020752, + 0.00764383003115654, + 0.016892673447728157, + 0.010451452806591988, + 0.0064459070563316345, + 0.013238484039902687, + 0.010947151109576225, + 0.015232149511575699, + 0.007671990431845188, + 0.005222367122769356, + 0.009929606691002846, + 0.006845380179584026, + 0.00475089019164443, + 0.00929079856723547, + 0.0072709303349256516, + 0.011313880793750286, + 0.007242941297590733, + 0.00460424367338419, + 0.015868529677391052, + 0.01212960947304964, + 0.0051323287189006805, + 0.006424209102988243, + 0.005325596779584885, + 0.0181825403124094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0017147850012406707 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019384894985705614 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019384894985705614, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019384894985705614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16467347741127014, + 0.07866904884576797, + 0.037147995084524155, + 0.05736410990357399, + 0.23036931455135345, + 0.2799071669578552, + 0.024550937116146088, + 0.03287629783153534, + 0.05599597468972206, + 0.03844570368528366 + ], + "credit_map": [ + 0.01912176050245762, + 0.016107672825455666, + 0.024553030729293823, + 0.027944499626755714, + 0.016168592497706413, + 0.020870743319392204, + 0.0168721042573452, + 0.023035069927573204, + 0.008824930526316166, + 0.006995490752160549, + 0.013927181251347065, + 0.015115471556782722, + 0.005882528144866228, + 0.009309063665568829, + 0.009969170205295086, + 0.008645211346447468, + 0.0038209983613342047, + 0.003535165684297681, + 0.005108133889734745, + 0.006916032172739506, + 0.0033709274139255285, + 0.004174120258539915, + 0.0057079424150288105, + 0.004514674656093121, + 0.007117102388292551, + 0.00608433410525322, + 0.007288359105587006, + 0.009858658537268639, + 0.0048882970586419106, + 0.007159318309277296, + 0.008137887343764305, + 0.006830155849456787, + 0.03843307122588158, + 0.016903851181268692, + 0.036323875188827515, + 0.03265098109841347, + 0.013847940601408482, + 0.02257459983229637, + 0.033040259033441544, + 0.03659473732113838, + 0.036112744361162186, + 0.021838083863258362, + 0.04443530738353729, + 0.03805341199040413, + 0.019680941477417946, + 0.03829073905944824, + 0.04187111556529999, + 0.039624810218811035, + 0.0035142337437719107, + 0.001629565958864987, + 0.003858471056446433, + 0.004339941777288914, + 0.0017760137561708689, + 0.0033622211776673794, + 0.002680595964193344, + 0.003389893099665642, + 0.0033172001130878925, + 0.0015871908981353045, + 0.00880645401775837, + 0.004246434196829796, + 0.0014905304415151477, + 0.0034444292541593313, + 0.0030488488264381886, + 0.006935210432857275, + 0.007818507961928844, + 0.005573844071477652, + 0.010453147813677788, + 0.00823898147791624, + 0.003964231815189123, + 0.004746501334011555, + 0.006151013541966677, + 0.00904974713921547, + 0.00386517820879817, + 0.0022095604799687862, + 0.00873589888215065, + 0.006515789777040482, + 0.002800379414111376, + 0.002792827319353819, + 0.003331185784190893, + 0.008194881491363049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0019384894985705614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584525540471077 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.020584525540471077, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584525540471077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18308696150779724, + 0.06880168616771698, + 0.06060837209224701, + 0.04827849194407463, + 0.07332739979028702, + 0.1056050956249237, + 0.26679474115371704, + 0.057414568960666656, + 0.0520162470638752, + 0.08406645059585571 + ], + "credit_map": [ + 0.023152723908424377, + 0.019905762746930122, + 0.028016548603773117, + 0.029383843764662743, + 0.011660512536764145, + 0.024481818079948425, + 0.023220889270305634, + 0.023264866322278976, + 0.007406169548630714, + 0.0058313338086009026, + 0.010998477227985859, + 0.01038274634629488, + 0.005602049175649881, + 0.005571024026721716, + 0.011071178130805492, + 0.01193870697170496, + 0.007117742672562599, + 0.005719827022403479, + 0.008453494869172573, + 0.011273106560111046, + 0.005072652827948332, + 0.0042862314730882645, + 0.012340910732746124, + 0.006344406399875879, + 0.005800062324851751, + 0.004848511889576912, + 0.007118972018361092, + 0.007304387167096138, + 0.00404318468645215, + 0.005477588158100843, + 0.005940091796219349, + 0.007745692040771246, + 0.010518384166061878, + 0.006523982156068087, + 0.01010588277131319, + 0.012283021584153175, + 0.006067808251827955, + 0.009547066874802113, + 0.00872064009308815, + 0.009560604579746723, + 0.01069286372512579, + 0.007753733079880476, + 0.018144385889172554, + 0.015996932983398438, + 0.007678698748350143, + 0.011860167607665062, + 0.018438374623656273, + 0.015039928257465363, + 0.05670035257935524, + 0.012624661438167095, + 0.04836717247962952, + 0.034949321299791336, + 0.017746958881616592, + 0.030166450887918472, + 0.019744135439395905, + 0.046495676040649414, + 0.005636806134134531, + 0.00497200945392251, + 0.013012436218559742, + 0.008110800758004189, + 0.003163603600114584, + 0.0050851223058998585, + 0.007336553186178207, + 0.010097228921949863, + 0.00694034481421113, + 0.00612995820119977, + 0.009945026598870754, + 0.007901093922555447, + 0.003517749020829797, + 0.005217272788286209, + 0.003968861885368824, + 0.008395937271416187, + 0.0121017349883914, + 0.007789008319377899, + 0.01482448447495699, + 0.013079184107482433, + 0.006693216972053051, + 0.008400160819292068, + 0.007093713153153658, + 0.01408493984490633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.020584525540471077 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely wide shot, positioned at a low angle, moves slowly forward along the center of a vast, empty urban canyon, creating a sense of steady, relentless movement. Towering skyscrapers of dark, reflective glass and polished steel rise on either side, their smooth, minimalist facades piercing a clear, late-afternoon sky. The black asphalt, still damp and shimmering from a recent rain, stretches like a dark, empty river. In the exact center of this vast emptiness, a lone, distant figure walks directly away from the camera in a steady, unwavering rhythm. This figure is a sharply defined individual, clad in a long, dark coat, their form framed by the golden light between the buildings. They advance with a deliberate gait, each step measured and consistent, hands in their pockets, plunging deep into the urban void. The low-hanging sun casts long, dramatic shadows that stretch across the sidewalk, creating a stark visual interplay of warm, bright light and cool, deep shadows. The atmosphere is one of profound solitude and awe, a cinematic and contemplative one, where the sheer size of the quiet modern metropolis dwarfs the single, enduring human element within it.", + "seed": 30905, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_0/index_306_seed_30905_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03536109197884798, + "vlm_logprob": -0.03536109197884798, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.03536109197884798, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1261146068572998 + ], + "opposite_logprob_tokens": [ + -2.3761146068573 + ], + "logprob_reward": -0.1261146068572998, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1261146068572998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2750445008277893, + 0.18237312138080597, + 0.1550372689962387, + 0.07966946810483932, + 0.03157252445816994, + 0.02990015409886837, + 0.08002287894487381, + 0.07027018070220947, + 0.06585931032896042, + 0.03025064989924431 + ], + "credit_map": [ + 0.07132255285978317, + 0.02580009028315544, + 0.05265170335769653, + 0.039272770285606384, + 0.020952647551894188, + 0.015464761294424534, + 0.027130044996738434, + 0.02244991436600685, + 0.036933936178684235, + 0.01399686373770237, + 0.03553326800465584, + 0.03590927645564079, + 0.0152001166716218, + 0.009565132670104504, + 0.019701063632965088, + 0.015533465892076492, + 0.019669976085424423, + 0.01042969897389412, + 0.027567889541387558, + 0.029824301600456238, + 0.017471641302108765, + 0.005425962619483471, + 0.03466615080833435, + 0.009981646202504635, + 0.01906542479991913, + 0.008707966655492783, + 0.014474738389253616, + 0.012653703801333904, + 0.006152935326099396, + 0.0024643265642225742, + 0.011615460738539696, + 0.004534913692623377, + 0.004904367495328188, + 0.002546006580814719, + 0.0048692491836845875, + 0.004020985681563616, + 0.0029881612863391638, + 0.003519589314237237, + 0.005564473569393158, + 0.003159689949825406, + 0.005173865705728531, + 0.003793047508224845, + 0.005112295038998127, + 0.0056831552647054195, + 0.002344604814425111, + 0.0013578209327533841, + 0.004476645961403847, + 0.0019587180577218533, + 0.010180705226957798, + 0.009633811190724373, + 0.014913416467607021, + 0.012333638966083527, + 0.007734593003988266, + 0.005577062722295523, + 0.01120440848171711, + 0.008445244282484055, + 0.016080260276794434, + 0.006390295922756195, + 0.012773850001394749, + 0.010181139223277569, + 0.007147347554564476, + 0.004160713404417038, + 0.007425673305988312, + 0.006110901944339275, + 0.010462171398103237, + 0.01261009182780981, + 0.010861456394195557, + 0.009450419805943966, + 0.009172426536679268, + 0.0024107675999403, + 0.007324814330786467, + 0.00356715964153409, + 0.003492159303277731, + 0.002421884797513485, + 0.005108789075165987, + 0.006054166238754988, + 0.002978317905217409, + 0.0011631130473688245, + 0.0073032500222325325, + 0.0017289675306528807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.1261146068572998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060753107070923 + ], + "opposite_logprob_tokens": [ + -3.8956074714660645 + ], + "logprob_reward": -0.02060753107070923, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060753107070923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21244776248931885, + 0.10164415091276169, + 0.24969595670700073, + 0.0884343683719635, + 0.08145338296890259, + 0.07136549055576324, + 0.05658448114991188, + 0.0807061493396759, + 0.038183849304914474, + 0.01948440819978714 + ], + "credit_map": [ + 0.035174135118722916, + 0.026522155851125717, + 0.02983608841896057, + 0.03346128389239311, + 0.019275281578302383, + 0.010354163125157356, + 0.034421999007463455, + 0.023402657359838486, + 0.021231111139059067, + 0.006485339254140854, + 0.018832003697752953, + 0.01717834733426571, + 0.011690516956150532, + 0.004625776782631874, + 0.014021058566868305, + 0.007579985540360212, + 0.046952273696660995, + 0.027099041268229485, + 0.03938997536897659, + 0.037922583520412445, + 0.038767773658037186, + 0.008890114724636078, + 0.0378275066614151, + 0.012846698053181171, + 0.012793086469173431, + 0.005425571464002132, + 0.015021836385130882, + 0.011977647431194782, + 0.010091415606439114, + 0.004967593587934971, + 0.02221253141760826, + 0.005944682285189629, + 0.009864747524261475, + 0.00895834993571043, + 0.014141756109893322, + 0.010334246791899204, + 0.009342002682387829, + 0.005489839240908623, + 0.017226001247763634, + 0.006096438970416784, + 0.0153121929615736, + 0.008714483119547367, + 0.012636154890060425, + 0.01288241520524025, + 0.005624351091682911, + 0.002601864282041788, + 0.009575840085744858, + 0.004018183797597885, + 0.007266046479344368, + 0.005585226230323315, + 0.010123512707650661, + 0.007741543930023909, + 0.0045563834719359875, + 0.0029557934030890465, + 0.01406740490347147, + 0.0042885662987828255, + 0.02094004862010479, + 0.01018444262444973, + 0.0141312750056386, + 0.007641514763236046, + 0.01101143192499876, + 0.004131936933845282, + 0.008163363672792912, + 0.004502138588577509, + 0.006875007413327694, + 0.0035127706360071898, + 0.0060728066600859165, + 0.00593723077327013, + 0.005190401338040829, + 0.0018401342676952481, + 0.005341232288628817, + 0.0034142667427659035, + 0.003998015075922012, + 0.002506759949028492, + 0.002969206776469946, + 0.0028998421039432287, + 0.0015606553060933948, + 0.0007613330380991101, + 0.002708136336877942, + 0.0020804591476917267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02060753107070923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002203062642365694 + ], + "opposite_logprob_tokens": [ + -6.127202987670898 + ], + "logprob_reward": -0.002203062642365694, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002203062642365694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30108314752578735, + 0.24093732237815857, + 0.09603020548820496, + 0.06707993894815445, + 0.04843008890748024, + 0.06516388803720474, + 0.040096480399370193, + 0.05159907042980194, + 0.052087653428316116, + 0.03749225288629532 + ], + "credit_map": [ + 0.07472826540470123, + 0.044506557285785675, + 0.05084507539868355, + 0.03299294412136078, + 0.0299528818577528, + 0.02099856175482273, + 0.030942274257540703, + 0.01611660234630108, + 0.03151632100343704, + 0.020824342966079712, + 0.03857364133000374, + 0.041960712522268295, + 0.04449501261115074, + 0.012222684919834137, + 0.034264642745256424, + 0.017079951241612434, + 0.02593591995537281, + 0.006923029199242592, + 0.019279519096016884, + 0.011326786130666733, + 0.011215203441679478, + 0.0035683345049619675, + 0.012800676748156548, + 0.004980742931365967, + 0.02090592123568058, + 0.005978419445455074, + 0.012580649927258492, + 0.007951529696583748, + 0.004781668074429035, + 0.0024650588165968657, + 0.008188963867723942, + 0.004227722063660622, + 0.00791491661220789, + 0.004502444993704557, + 0.007375464309006929, + 0.007667488418519497, + 0.004445044789463282, + 0.006267677526921034, + 0.004687891807407141, + 0.005569157190620899, + 0.011454882100224495, + 0.00872350949794054, + 0.010084929876029491, + 0.008755790069699287, + 0.010124715976417065, + 0.0034391104709357023, + 0.009419791400432587, + 0.003161156550049782, + 0.004285781644284725, + 0.0032605056185275316, + 0.007248938549309969, + 0.007864592596888542, + 0.0038429833948612213, + 0.002442268654704094, + 0.007680546026676893, + 0.0034708636812865734, + 0.00972827710211277, + 0.0054228221997618675, + 0.008587980642914772, + 0.007131506688892841, + 0.0071474527940154076, + 0.003155231010168791, + 0.005935339257121086, + 0.0044904593378305435, + 0.006927808281034231, + 0.005974516738206148, + 0.007785946596413851, + 0.008636781945824623, + 0.008685785345733166, + 0.0036556690465658903, + 0.006129345390945673, + 0.004291802179068327, + 0.005511281546205282, + 0.0032020886428654194, + 0.0062128049321472645, + 0.007351103704422712, + 0.003909165505319834, + 0.001879892428405583, + 0.0062028816901147366, + 0.0032230361830443144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.002203062642365694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001512932125478983 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.001512932125478983, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001512932125478983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14541666209697723, + 0.15642428398132324, + 0.14066524803638458, + 0.31300243735313416, + 0.034789152443408966, + 0.029799509793519974, + 0.021717695519328117, + 0.054182760417461395, + 0.021433761343359947, + 0.0825684666633606 + ], + "credit_map": [ + 0.017721854150295258, + 0.017123043537139893, + 0.020934974774718285, + 0.022694816812872887, + 0.013430694118142128, + 0.005744532216340303, + 0.034657884389162064, + 0.0131088737398386, + 0.023291846737265587, + 0.015172390267252922, + 0.02684478648006916, + 0.027723805978894234, + 0.023236950859427452, + 0.005093601066619158, + 0.026142476126551628, + 0.008918434381484985, + 0.03210698440670967, + 0.015381640754640102, + 0.025000162422657013, + 0.015981312841176987, + 0.023759741336107254, + 0.006226849276572466, + 0.014262106269598007, + 0.007946467027068138, + 0.06276194006204605, + 0.025845184922218323, + 0.058342717587947845, + 0.047320976853370667, + 0.018686633557081223, + 0.010720544494688511, + 0.07064683735370636, + 0.018677612766623497, + 0.00642324099317193, + 0.0029710347298532724, + 0.005630175583064556, + 0.004336687736213207, + 0.0032204564195126295, + 0.002407406223937869, + 0.007161679212003946, + 0.002638468751683831, + 0.004232550505548716, + 0.003294579219073057, + 0.005029040854424238, + 0.0058535137213766575, + 0.0025123090017586946, + 0.001967804739251733, + 0.004262520465999842, + 0.002647190121933818, + 0.0026994687505066395, + 0.002711157314479351, + 0.003825010033324361, + 0.0036726545076817274, + 0.0023329046089202166, + 0.0016920024063438177, + 0.0028997589834034443, + 0.0018847394967451692, + 0.00711696594953537, + 0.0052702585235238075, + 0.008036859333515167, + 0.007969725877046585, + 0.010364514775574207, + 0.004597765393555164, + 0.006255850195884705, + 0.004570817109197378, + 0.004994840361177921, + 0.0022256500087678432, + 0.003843187354505062, + 0.003273082198575139, + 0.002302431268617511, + 0.000946871645282954, + 0.002479914575815201, + 0.0013677830575034022, + 0.009587569162249565, + 0.008732100948691368, + 0.013296015560626984, + 0.015127171762287617, + 0.009972990490496159, + 0.003258924465626478, + 0.015283387154340744, + 0.007310309913009405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.001512932125478983 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026367327198386192 + ], + "opposite_logprob_tokens": [ + -3.651367425918579 + ], + "logprob_reward": -0.026367327198386192, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026367327198386192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15352602303028107, + 0.2310894876718521, + 0.0674428790807724, + 0.15854735672473907, + 0.07398375123739243, + 0.059220656752586365, + 0.03127599135041237, + 0.07045517861843109, + 0.09107652306556702, + 0.06338206678628922 + ], + "credit_map": [ + 0.03070925921201706, + 0.013387013226747513, + 0.02529193088412285, + 0.016896450892090797, + 0.01568012498319149, + 0.00834568403661251, + 0.027996119111776352, + 0.01521945558488369, + 0.04153700917959213, + 0.016765832901000977, + 0.04128791391849518, + 0.04808352142572403, + 0.025278771296143532, + 0.007627627346664667, + 0.032887451350688934, + 0.01762138493359089, + 0.009926690720021725, + 0.007011376786977053, + 0.010203653946518898, + 0.013852844946086407, + 0.006457145791500807, + 0.0034406622871756554, + 0.01154984813183546, + 0.005000665783882141, + 0.03786865249276161, + 0.012991059571504593, + 0.02840949408710003, + 0.02658930793404579, + 0.00981844775378704, + 0.005250341724604368, + 0.02458975650370121, + 0.01303030364215374, + 0.01565115712583065, + 0.005165292881429195, + 0.013385342434048653, + 0.008853726089000702, + 0.0074121905490756035, + 0.005119647365063429, + 0.01404022891074419, + 0.00435616634786129, + 0.015389998443424702, + 0.005592516157776117, + 0.010387396439909935, + 0.009237482212483883, + 0.005887393839657307, + 0.002980519086122513, + 0.00583714060485363, + 0.00390821835026145, + 0.003491051960736513, + 0.0029904667753726244, + 0.005377680994570255, + 0.005046929698437452, + 0.0033850972540676594, + 0.0018185963854193687, + 0.00667726993560791, + 0.002488902769982815, + 0.012292684055864811, + 0.006519179791212082, + 0.011120147071778774, + 0.009869785979390144, + 0.011288723908364773, + 0.005352320149540901, + 0.007922055199742317, + 0.006090290378779173, + 0.014400867745280266, + 0.012889808975160122, + 0.013552061282098293, + 0.012703284621238708, + 0.01643703505396843, + 0.00498217111453414, + 0.01042797788977623, + 0.0056833322159945965, + 0.007734251674264669, + 0.0058592273853719234, + 0.010409589856863022, + 0.012035258114337921, + 0.007258264347910881, + 0.0025746014434844255, + 0.012644252739846706, + 0.0048666223883628845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.026367327198386192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely wide shot, positioned at a low angle, moves slowly forward along the center of a vast, empty urban canyon, creating a sense of steady, relentless movement. Towering skyscrapers of dark, reflective glass and polished steel rise on either side, their smooth, minimalist facades piercing a clear, late-afternoon sky. The black asphalt, still damp and shimmering from a recent rain, stretches like a dark, empty river. In the exact center of this vast emptiness, a lone, distant figure walks directly away from the camera in a steady, unwavering rhythm. This figure is a sharply defined individual, clad in a long, dark coat, their form framed by the golden light between the buildings. They advance with a deliberate gait, each step measured and consistent, hands in their pockets, plunging deep into the urban void. The low-hanging sun casts long, dramatic shadows that stretch across the sidewalk, creating a stark visual interplay of warm, bright light and cool, deep shadows. The atmosphere is one of profound solitude and awe, a cinematic and contemplative one, where the sheer size of the quiet modern metropolis dwarfs the single, enduring human element within it.", + "seed": 30906, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_0/index_306_seed_30906_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011419289163313806, + "vlm_logprob": -0.011419289163313806, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.011419289163313806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01728377304971218 + ], + "opposite_logprob_tokens": [ + -4.767283916473389 + ], + "logprob_reward": -0.01728377304971218, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01728377304971218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24139970541000366, + 0.05058443918824196, + 0.08977000415325165, + 0.07932215183973312, + 0.024719148874282837, + 0.035470277070999146, + 0.10654044896364212, + 0.08831430971622467, + 0.1504003256559372, + 0.1334792673587799 + ], + "credit_map": [ + 0.04555222764611244, + 0.025561444461345673, + 0.034147847443819046, + 0.021599285304546356, + 0.0224525798112154, + 0.03183266520500183, + 0.027620628476142883, + 0.03263303264975548, + 0.010549348779022694, + 0.003912035841494799, + 0.005598132964223623, + 0.0053423624485731125, + 0.0025992749724537134, + 0.009715215303003788, + 0.004534540232270956, + 0.008333532139658928, + 0.012781353667378426, + 0.0072700586169958115, + 0.011476067826151848, + 0.013737483881413937, + 0.005884004291146994, + 0.013200566172599792, + 0.011586572043597698, + 0.013833899050951004, + 0.013989328406751156, + 0.0061933184042572975, + 0.009439083747565746, + 0.00926266610622406, + 0.004963726736605167, + 0.01472418662160635, + 0.009131165221333504, + 0.01161867380142212, + 0.003220705548301339, + 0.0017504898132756352, + 0.0026133826468139887, + 0.003923529759049416, + 0.0016106375260278583, + 0.00437416136264801, + 0.0029744890052825212, + 0.004251754842698574, + 0.005895146634429693, + 0.003240956226363778, + 0.003869435051456094, + 0.004245589952915907, + 0.0032955885399132967, + 0.0055414773523807526, + 0.004018011037260294, + 0.005364067852497101, + 0.02840549685060978, + 0.008245070464909077, + 0.008402048610150814, + 0.008016237057745457, + 0.009988057427108288, + 0.016593053936958313, + 0.008445276878774166, + 0.018445201218128204, + 0.013388902880251408, + 0.007677390240132809, + 0.008433924987912178, + 0.0066390857100486755, + 0.014636139385402203, + 0.013518218882381916, + 0.00978771224617958, + 0.014232932589948177, + 0.04666567221283913, + 0.012046507559716702, + 0.01029854267835617, + 0.007549645844846964, + 0.009046143852174282, + 0.022549256682395935, + 0.012586639262735844, + 0.029657913371920586, + 0.02141406573355198, + 0.012465535663068295, + 0.010297859087586403, + 0.009316534735262394, + 0.021696705371141434, + 0.022604864090681076, + 0.01303471066057682, + 0.02264898270368576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.01728377304971218 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204702340066433 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204702340066433, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204702340066433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20006363093852997, + 0.06926631182432175, + 0.05370468273758888, + 0.04460936412215233, + 0.07799385488033295, + 0.07061347365379333, + 0.1025388166308403, + 0.1781550794839859, + 0.02805311232805252, + 0.1750016212463379 + ], + "credit_map": [ + 0.022565141320228577, + 0.02304692566394806, + 0.022186318412423134, + 0.021817831322550774, + 0.018951481208205223, + 0.04545726999640465, + 0.0169009268283844, + 0.029137734323740005, + 0.009420418180525303, + 0.007382448296993971, + 0.006555423140525818, + 0.006093294359743595, + 0.010482181794941425, + 0.010491127148270607, + 0.009288525208830833, + 0.009552890434861183, + 0.010441717691719532, + 0.005102250259369612, + 0.005990301258862019, + 0.006788307800889015, + 0.0037559892516583204, + 0.007077615708112717, + 0.007639588788151741, + 0.006908910349011421, + 0.00639605987817049, + 0.0030713602900505066, + 0.006187706254422665, + 0.0071450998075306416, + 0.0025105164386332035, + 0.006442702375352383, + 0.006121602840721607, + 0.006734314374625683, + 0.012003045529127121, + 0.007713701110333204, + 0.00516689894720912, + 0.008718515746295452, + 0.011659775860607624, + 0.014447716996073723, + 0.006598527077585459, + 0.011685673147439957, + 0.011324761435389519, + 0.005826468113809824, + 0.009410789236426353, + 0.011002936400473118, + 0.0044010295532643795, + 0.00915544293820858, + 0.00858160201460123, + 0.010910439305007458, + 0.010581518523395061, + 0.009369106963276863, + 0.007344281766563654, + 0.011937239207327366, + 0.018446315079927444, + 0.01639723964035511, + 0.012850243598222733, + 0.015612874180078506, + 0.018113890662789345, + 0.014401153661310673, + 0.010892348363995552, + 0.018909262493252754, + 0.02550029382109642, + 0.01733512058854103, + 0.040986619889736176, + 0.032016392797231674, + 0.006963531486690044, + 0.002419827040284872, + 0.0024262014776468277, + 0.0017049703747034073, + 0.0017323349602520466, + 0.004398303106427193, + 0.002686038613319397, + 0.005721905268728733, + 0.018264619633555412, + 0.023282449692487717, + 0.01516001671552658, + 0.013441823422908783, + 0.03971756994724274, + 0.01924354024231434, + 0.022731764242053032, + 0.023159833624958992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.014204702340066433 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009230881696566939 + ], + "opposite_logprob_tokens": [ + -7.000923156738281 + ], + "logprob_reward": -0.0009230881696566939, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009230881696566939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1802745759487152, + 0.08813236653804779, + 0.08621779084205627, + 0.0957469493150711, + 0.03732374683022499, + 0.033914897590875626, + 0.042314041405916214, + 0.17914840579032898, + 0.08061482012271881, + 0.1763123720884323 + ], + "credit_map": [ + 0.03189442306756973, + 0.016682378947734833, + 0.020652668550610542, + 0.020481929183006287, + 0.019192473962903023, + 0.028314638882875443, + 0.01677675172686577, + 0.02627931907773018, + 0.018622972071170807, + 0.008226479403674603, + 0.008173748850822449, + 0.010929879732429981, + 0.00823049247264862, + 0.012862258590757847, + 0.006656741723418236, + 0.01442980207502842, + 0.014851766638457775, + 0.006632396951317787, + 0.011725463904440403, + 0.011213845573365688, + 0.005719044711440802, + 0.012335390783846378, + 0.010472751222550869, + 0.013267135247588158, + 0.017022503539919853, + 0.0072950622998178005, + 0.012455782853066921, + 0.011082001961767673, + 0.005802284460514784, + 0.01624135486781597, + 0.01126991305500269, + 0.014578056521713734, + 0.004299553111195564, + 0.004006640519946814, + 0.0028193381149321795, + 0.0034164173994213343, + 0.006900191772729158, + 0.005444801412522793, + 0.004456720780581236, + 0.005980086512863636, + 0.005054596345871687, + 0.0031840552110224962, + 0.0028141657821834087, + 0.0026966645382344723, + 0.005542541388422251, + 0.005700747016817331, + 0.00365038332529366, + 0.00527174910530448, + 0.009957647882401943, + 0.0034094704315066338, + 0.0026683404576033354, + 0.0034523645881563425, + 0.003998561762273312, + 0.007163898553699255, + 0.004495406989008188, + 0.007168355397880077, + 0.019113702699542046, + 0.015798022970557213, + 0.010530916042625904, + 0.016824664548039436, + 0.029342567548155785, + 0.01677972637116909, + 0.038130443543195724, + 0.0326283760368824, + 0.027402274310588837, + 0.006308502517640591, + 0.0056358142755925655, + 0.003299688920378685, + 0.004997360520064831, + 0.011537754908204079, + 0.005728440824896097, + 0.015704991295933723, + 0.021541070193052292, + 0.020276395604014397, + 0.015049131587147713, + 0.012839962728321552, + 0.03610193729400635, + 0.025182513520121574, + 0.019920997321605682, + 0.02540038339793682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.0009230881696566939 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040924143977463245 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.0040924143977463245, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040924143977463245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.070573590695858, + 0.07819639146327972, + 0.025695588439702988, + 0.03920407593250275, + 0.0386740043759346, + 0.0390133410692215, + 0.24150530993938446, + 0.28674596548080444, + 0.02616124041378498, + 0.15423054993152618 + ], + "credit_map": [ + 0.007777854334563017, + 0.008114943280816078, + 0.007183725014328957, + 0.009788831695914268, + 0.009684178046882153, + 0.008568129502236843, + 0.009768450632691383, + 0.009687473997473717, + 0.015975745394825935, + 0.008976446464657784, + 0.006579803768545389, + 0.0056550041772425175, + 0.011621934361755848, + 0.010054849088191986, + 0.008120001293718815, + 0.011212605983018875, + 0.0034619311336427927, + 0.0029704831540584564, + 0.002704658778384328, + 0.005061319563537836, + 0.0024558573495596647, + 0.0031013761181384325, + 0.0027041665744036436, + 0.003235796233639121, + 0.005915475077927113, + 0.00324522377923131, + 0.005017407704144716, + 0.005129087250679731, + 0.0028210836462676525, + 0.006310411263257265, + 0.00487686600536108, + 0.0058885193429887295, + 0.004469878040254116, + 0.00410150783136487, + 0.002906481036916375, + 0.0046803951263427734, + 0.007103100419044495, + 0.0051696449518203735, + 0.004284246824681759, + 0.005958750378340483, + 0.004671961069107056, + 0.004706748761236668, + 0.002815620042383671, + 0.0034120511263608932, + 0.007658322807401419, + 0.0055329082533717155, + 0.004495738539844751, + 0.00571999279782176, + 0.023302512243390083, + 0.02811298705637455, + 0.010683441534638405, + 0.01870115101337433, + 0.05445152893662453, + 0.03383816406130791, + 0.03368477523326874, + 0.03873074799776077, + 0.029333481565117836, + 0.025693019852042198, + 0.01607648842036724, + 0.02736484818160534, + 0.04639621451497078, + 0.027210162952542305, + 0.06275603175163269, + 0.05191570520401001, + 0.00769102293998003, + 0.0022177945356816053, + 0.0019483062205836177, + 0.0013915757881477475, + 0.0014785637613385916, + 0.004188794642686844, + 0.002110494300723076, + 0.005134688690304756, + 0.01647702232003212, + 0.02075575664639473, + 0.013011954724788666, + 0.011692553758621216, + 0.03460834547877312, + 0.017657482996582985, + 0.01897827722132206, + 0.021049149334430695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.0040924143977463245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592467859387398 + ], + "opposite_logprob_tokens": [ + -3.895592451095581 + ], + "logprob_reward": -0.020592467859387398, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592467859387398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06761925667524338, + 0.0617152638733387, + 0.01396030094474554, + 0.056443557143211365, + 0.020375743508338928, + 0.02813495323061943, + 0.33130133152008057, + 0.2970544099807739, + 0.06568818539381027, + 0.05770695209503174 + ], + "credit_map": [ + 0.008975563570857048, + 0.006291392259299755, + 0.006871322635561228, + 0.007665219716727734, + 0.008903160691261292, + 0.009136948734521866, + 0.01020190492272377, + 0.009573747403919697, + 0.015723299235105515, + 0.005960556678473949, + 0.006281325127929449, + 0.005269793327897787, + 0.0036220138426870108, + 0.010093999095261097, + 0.004124061204493046, + 0.010640215128660202, + 0.00166791002266109, + 0.0012762792175635695, + 0.0021593119017779827, + 0.0020560231059789658, + 0.0008973857038654387, + 0.0017643250757828355, + 0.0017707645893096924, + 0.0023683023173362017, + 0.009917636401951313, + 0.004649150185286999, + 0.007190995384007692, + 0.007444597315043211, + 0.0036991231609135866, + 0.008364380337297916, + 0.00683149928227067, + 0.00834618415683508, + 0.0033353781327605247, + 0.0015999542083591223, + 0.0014752265997231007, + 0.002550613833591342, + 0.0019356627017259598, + 0.004832063801586628, + 0.0016874686116352677, + 0.0029593762010335922, + 0.005068667698651552, + 0.002260538749396801, + 0.0033003194257616997, + 0.003493081545457244, + 0.0024569660890847445, + 0.004346813540905714, + 0.002799087902531028, + 0.00440948037430644, + 0.0321631096303463, + 0.03700093552470207, + 0.01687995158135891, + 0.025755513459444046, + 0.07710365206003189, + 0.037025220692157745, + 0.05075602978467941, + 0.05461692810058594, + 0.04078163951635361, + 0.02662002295255661, + 0.016236813738942146, + 0.02601737156510353, + 0.04940135404467583, + 0.02862568385899067, + 0.05922207608819008, + 0.05014946684241295, + 0.021308142691850662, + 0.005320793017745018, + 0.004135873634368181, + 0.0029092191252857447, + 0.0038116327486932278, + 0.010242843069136143, + 0.004882016219198704, + 0.013077668845653534, + 0.008956982754170895, + 0.005104704760015011, + 0.00680211465805769, + 0.0042113750241696835, + 0.00798441655933857, + 0.008381051942706108, + 0.006603209767490625, + 0.00966309942305088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.020592467859387398 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely wide shot, positioned at a low angle, moves slowly forward along the center of a vast, empty urban canyon, creating a sense of steady, relentless movement. Towering skyscrapers of dark, reflective glass and polished steel rise on either side, their smooth, minimalist facades piercing a clear, late-afternoon sky. The black asphalt, still damp and shimmering from a recent rain, stretches like a dark, empty river. In the exact center of this vast emptiness, a lone, distant figure walks directly away from the camera in a steady, unwavering rhythm. This figure is a sharply defined individual, clad in a long, dark coat, their form framed by the golden light between the buildings. They advance with a deliberate gait, each step measured and consistent, hands in their pockets, plunging deep into the urban void. The low-hanging sun casts long, dramatic shadows that stretch across the sidewalk, creating a stark visual interplay of warm, bright light and cool, deep shadows. The atmosphere is one of profound solitude and awe, a cinematic and contemplative one, where the sheer size of the quiet modern metropolis dwarfs the single, enduring human element within it.", + "seed": 30907, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_0/index_306_seed_30907_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010678122437093407, + "vlm_logprob": -0.010678122437093407, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.010678122437093407, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022837484255433083 + ], + "opposite_logprob_tokens": [ + -4.7728376388549805 + ], + "logprob_reward": -0.022837484255433083, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.022837484255433083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17553842067718506, + 0.04299037531018257, + 0.061482835561037064, + 0.10393589735031128, + 0.04962536320090294, + 0.29841411113739014, + 0.05991429463028908, + 0.06499015539884567, + 0.08809874206781387, + 0.05500981584191322 + ], + "credit_map": [ + 0.027400797232985497, + 0.014805872924625874, + 0.037129536271095276, + 0.024664688855409622, + 0.014348977245390415, + 0.014516151510179043, + 0.010925186797976494, + 0.031747203320264816, + 0.004670949187129736, + 0.004354486241936684, + 0.007709768135100603, + 0.005808759480714798, + 0.0030733984895050526, + 0.004950202535837889, + 0.004295474383980036, + 0.008127336390316486, + 0.005211937241256237, + 0.0042594242841005325, + 0.0099636884406209, + 0.005585946142673492, + 0.008054031059145927, + 0.009581266902387142, + 0.00438010785728693, + 0.014446431770920753, + 0.020488601177930832, + 0.011681470088660717, + 0.017837870866060257, + 0.01165054738521576, + 0.007452068384736776, + 0.008622679859399796, + 0.008486866019666195, + 0.01771578937768936, + 0.005094117950648069, + 0.00572313042357564, + 0.0065154400654137135, + 0.0073015266098082066, + 0.005345318932086229, + 0.0053194258362054825, + 0.007554452400654554, + 0.006771949119865894, + 0.031909964978694916, + 0.0329570434987545, + 0.04318642243742943, + 0.03422067314386368, + 0.03887588903307915, + 0.025317339226603508, + 0.04011351242661476, + 0.05183327570557594, + 0.007674823515117168, + 0.004939408972859383, + 0.012413633987307549, + 0.007097110617905855, + 0.004679726902395487, + 0.005048959981650114, + 0.008887935429811478, + 0.009172698482871056, + 0.009873838163912296, + 0.0065021938644349575, + 0.008562726899981499, + 0.007002699188888073, + 0.009185786359012127, + 0.007198419887572527, + 0.006935196928679943, + 0.009729299694299698, + 0.010218001902103424, + 0.009084941819310188, + 0.012386421672999859, + 0.009969700127840042, + 0.007333424407988787, + 0.011706034652888775, + 0.013726357370615005, + 0.013673859648406506, + 0.007228991482406855, + 0.0052949427627027035, + 0.009090789593756199, + 0.006272937171161175, + 0.006095788441598415, + 0.006174467038363218, + 0.006060041021555662, + 0.008791857399046421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.022837484255433083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067492141388356686 + ], + "opposite_logprob_tokens": [ + -5.006749153137207 + ], + "logprob_reward": -0.0067492141388356686, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067492141388356686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3232142925262451, + 0.059307761490345, + 0.06505157798528671, + 0.09277647733688354, + 0.03467301279306412, + 0.05735477805137634, + 0.07916594296693802, + 0.0894155278801918, + 0.162573903799057, + 0.03646663948893547 + ], + "credit_map": [ + 0.03662959113717079, + 0.014595502987504005, + 0.1142975240945816, + 0.029137155041098595, + 0.012704623863101006, + 0.023708434775471687, + 0.02009700983762741, + 0.07204446941614151, + 0.0074368990026414394, + 0.0055573610588908195, + 0.013083447702229023, + 0.005452336277812719, + 0.006676018703728914, + 0.0062990738078951836, + 0.005127438809722662, + 0.009675189852714539, + 0.01387086883187294, + 0.009745659306645393, + 0.010729465633630753, + 0.00751138012856245, + 0.004049916286021471, + 0.005625319667160511, + 0.0038017407059669495, + 0.009717235341668129, + 0.01778387650847435, + 0.006411049049347639, + 0.01865718513727188, + 0.009431513957679272, + 0.006599077023565769, + 0.007147669792175293, + 0.010554041713476181, + 0.016192076727747917, + 0.003518348094075918, + 0.003690500510856509, + 0.00454404903575778, + 0.004910007584840059, + 0.003251437097787857, + 0.00435057794675231, + 0.005056073423475027, + 0.0053520179353654385, + 0.0073221721686422825, + 0.005824641324579716, + 0.010812494903802872, + 0.008926477283239365, + 0.004713964648544788, + 0.004888162016868591, + 0.0038398639298975468, + 0.011027003638446331, + 0.00879369955509901, + 0.006367038935422897, + 0.011607597582042217, + 0.009256205521523952, + 0.008718360215425491, + 0.010411986149847507, + 0.01375217642635107, + 0.010258883237838745, + 0.017087744548916817, + 0.008160070516169071, + 0.011044098995625973, + 0.010998593643307686, + 0.008976908400654793, + 0.0068464879877865314, + 0.01324327103793621, + 0.013058355078101158, + 0.02136233262717724, + 0.020786229521036148, + 0.025950808078050613, + 0.016800768673419952, + 0.010767304338514805, + 0.01788625866174698, + 0.016506986692547798, + 0.03251323848962784, + 0.0051130130887031555, + 0.003252103691920638, + 0.005828821565955877, + 0.0037884076591581106, + 0.0025418992154300213, + 0.0037750883493572474, + 0.005108592566102743, + 0.007058717776089907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0067492141388356686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024921807926148176 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.0024921807926148176, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024921807926148176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3230275511741638, + 0.07815209776163101, + 0.06425343453884125, + 0.2101958692073822, + 0.04613325372338295, + 0.03559710085391998, + 0.07102453708648682, + 0.038860343396663666, + 0.042480822652578354, + 0.09027494490146637 + ], + "credit_map": [ + 0.04818892478942871, + 0.026199938729405403, + 0.07295066863298416, + 0.03374839201569557, + 0.016200631856918335, + 0.027996400371193886, + 0.017524125054478645, + 0.08021847903728485, + 0.008240512572228909, + 0.006839767098426819, + 0.016453854739665985, + 0.009477414190769196, + 0.0057039642706513405, + 0.007502239663153887, + 0.005108656361699104, + 0.01882569119334221, + 0.009830087423324585, + 0.007546107750386, + 0.011904341168701649, + 0.009841610677540302, + 0.004330561496317387, + 0.005231327377259731, + 0.006433438044041395, + 0.009135963395237923, + 0.03536117821931839, + 0.02036743424832821, + 0.03630036488175392, + 0.023529812693595886, + 0.014738786965608597, + 0.0215243361890316, + 0.013709847815334797, + 0.04466411471366882, + 0.006035314407199621, + 0.006446830928325653, + 0.006227932404726744, + 0.005679423920810223, + 0.005478032398968935, + 0.004973096307367086, + 0.004519501235336065, + 0.0067731221206486225, + 0.0038514561019837856, + 0.003348695347085595, + 0.005831384100019932, + 0.005691810976713896, + 0.002907373709604144, + 0.0031602263916283846, + 0.004152772482484579, + 0.006653385236859322, + 0.011630832217633724, + 0.005641634576022625, + 0.01236763410270214, + 0.006851543672382832, + 0.011174925602972507, + 0.005761799868196249, + 0.005613897927105427, + 0.0119822658598423, + 0.005269265733659267, + 0.00470577972009778, + 0.005384964868426323, + 0.003713799174875021, + 0.003837952855974436, + 0.0037828139029443264, + 0.0049776048399508, + 0.007188158575445414, + 0.004448080901056528, + 0.003990475554019213, + 0.006968135479837656, + 0.005224919877946377, + 0.004877220373600721, + 0.0036389315500855446, + 0.004875826183706522, + 0.008457235060632229, + 0.012610146775841713, + 0.012949183583259583, + 0.012561571784317493, + 0.011264394968748093, + 0.010005481541156769, + 0.010544050484895706, + 0.007542774081230164, + 0.012797342613339424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0024921807926148176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007180972606875002 + ], + "opposite_logprob_tokens": [ + -7.250718116760254 + ], + "logprob_reward": -0.0007180972606875002, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007180972606875002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06241298094391823, + 0.03659025952219963, + 0.026858074590563774, + 0.4249225854873657, + 0.11454879492521286, + 0.06944870948791504, + 0.05043558031320572, + 0.06088080629706383, + 0.1147986650466919, + 0.039103586226701736 + ], + "credit_map": [ + 0.00932404212653637, + 0.005469773896038532, + 0.010722736828029156, + 0.008488856256008148, + 0.004350847564637661, + 0.006365278270095587, + 0.007599499076604843, + 0.010091951116919518, + 0.00449297484010458, + 0.0031594771426171064, + 0.00636059558019042, + 0.004149351269006729, + 0.0031806074548512697, + 0.004719968885183334, + 0.004277809988707304, + 0.0062494738958776, + 0.0036373722832649946, + 0.003091694787144661, + 0.00533019844442606, + 0.0037892635446041822, + 0.0014990256167948246, + 0.0027484495658427477, + 0.002170186722651124, + 0.004591884091496468, + 0.0752350315451622, + 0.04198543727397919, + 0.08010581880807877, + 0.042185574769973755, + 0.023021770641207695, + 0.041004255414009094, + 0.032676346600055695, + 0.08870835602283478, + 0.010022652335464954, + 0.008141295053064823, + 0.0181563850492239, + 0.015600540675222874, + 0.016128886491060257, + 0.009740161709487438, + 0.023481318727135658, + 0.013277556747198105, + 0.006327129900455475, + 0.007989641278982162, + 0.010142787359654903, + 0.010083241388201714, + 0.008421218954026699, + 0.006053868681192398, + 0.010365782305598259, + 0.010065040551126003, + 0.006487113423645496, + 0.004101941362023354, + 0.010474562644958496, + 0.005214259028434753, + 0.008171639405190945, + 0.0046090311370790005, + 0.0035688597708940506, + 0.0078081707470119, + 0.006512601859867573, + 0.0053745233453810215, + 0.010532844811677933, + 0.006852925755083561, + 0.005434545688331127, + 0.005792373791337013, + 0.00833144225180149, + 0.012049549259245396, + 0.015096801333129406, + 0.01269128080457449, + 0.016985056921839714, + 0.011961781419813633, + 0.008039281703531742, + 0.012879443354904652, + 0.013036261312663555, + 0.024108754470944405, + 0.005041912663727999, + 0.0031758337281644344, + 0.006422844715416431, + 0.0036101406440138817, + 0.003277689451351762, + 0.004777159541845322, + 0.005907011684030294, + 0.006890993099659681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.0007180972606875002 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot extremely wide and positioned at a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly forward along the center of a vast, empty urban canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there towering skyscrapers with dark, reflective glass and polished steel facades on either side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a lone, distant figure visible walking directly away from the camera in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the figure clad in a long, dark coat with hands in their pockets?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020593635737895966 + ], + "opposite_logprob_tokens": [ + -3.8955936431884766 + ], + "logprob_reward": -0.020593635737895966, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020593635737895966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15840573608875275, + 0.04037884250283241, + 0.03145960718393326, + 0.12527096271514893, + 0.02564285323023796, + 0.11910691857337952, + 0.052967704832553864, + 0.15860360860824585, + 0.18710482120513916, + 0.10105890780687332 + ], + "credit_map": [ + 0.02202470228075981, + 0.014345058239996433, + 0.028587214648723602, + 0.01514382753521204, + 0.007728226482868195, + 0.02155008167028427, + 0.010424593463540077, + 0.03860202059149742, + 0.005163353867828846, + 0.0034357556141912937, + 0.0062940651550889015, + 0.005099820904433727, + 0.004095450509339571, + 0.005443313159048557, + 0.004739878233522177, + 0.006107202265411615, + 0.0036517742555588484, + 0.0032276823185384274, + 0.005539146717637777, + 0.0033926849719136953, + 0.002753810491412878, + 0.0034469859674572945, + 0.003053734777495265, + 0.006393788382411003, + 0.02317105606198311, + 0.01325466949492693, + 0.02013762854039669, + 0.012537071481347084, + 0.006871524266898632, + 0.012685470283031464, + 0.007923776283860207, + 0.02868976630270481, + 0.0029570101760327816, + 0.002740241354331374, + 0.004297284409403801, + 0.0032194925006479025, + 0.0018667214317247272, + 0.0024484319146722555, + 0.002852826612070203, + 0.0052608465775847435, + 0.01127411425113678, + 0.011850426904857159, + 0.01724320650100708, + 0.015042691491544247, + 0.011836962774395943, + 0.011600728146731853, + 0.02147921361029148, + 0.01877957209944725, + 0.008148926310241222, + 0.004645132925361395, + 0.00922324601560831, + 0.006498347036540508, + 0.00723946513608098, + 0.0048362258821725845, + 0.0039972905069589615, + 0.00837907288223505, + 0.017457351088523865, + 0.012739035300910473, + 0.029267987236380577, + 0.015673339366912842, + 0.008403653278946877, + 0.020235208794474602, + 0.01268268283456564, + 0.04214435815811157, + 0.02417665906250477, + 0.013509840704500675, + 0.03079458512365818, + 0.01875172182917595, + 0.013096301816403866, + 0.019974060356616974, + 0.01754164882004261, + 0.04926000162959099, + 0.007815130986273289, + 0.006789283361285925, + 0.01698254607617855, + 0.01402113027870655, + 0.011850234121084213, + 0.011558948084712029, + 0.011895716190338135, + 0.02014591544866562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.020593635737895966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_1/index_tensor([2018, 2018])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_1/index_tensor([2018, 2018])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e0f6de17fb140b479a5caae84facb08897dc30bc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_1/index_tensor([2018, 2018])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot, imbued with subtle film grain for a cinematic quality, captures exhausted patients in a sterile hospital ward. With damp, matted hair plastered to their pale foreheads, the patients struggle to sit. They wear loose, light-blue, short-sleeved hospital gowns that hang limply from their weakened bodies. Their arms tremble noticeably, and a look of effort etched on their faces, their palms pressed deliberately against the mattress. As their torsos slowly rise, the wrinkled, thin, white hospital sheets, creased from restless sleep, slide down their chests and pool around their waists. The bed is a standard metal-framed model; in the slightly out-of-focus background, a metal IV rack stands sentry. The entire scene is illuminated by the cool, flat, unforgiving light of overhead fluorescent lamps, casting soft but sharp shadows that eliminate any sense of warmth and create a somber and deeply personal atmosphere filled with quiet struggle and vulnerability.", + "seed": 30897, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_1/index_2018_seed_30897_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19020993784070014, + "vlm_logprob": -0.19020993784070014, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.19020993784070014, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03961043804883957 + ], + "opposite_logprob_tokens": [ + -3.5396103858947754 + ], + "logprob_reward": -0.03961043804883957, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03961043804883957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18577465415000916, + 0.07208391278982162, + 0.22578975558280945, + 0.06036337465047836, + 0.11455610394477844, + 0.05498960241675377, + 0.18885205686092377, + 0.0443730428814888, + 0.027090849354863167, + 0.02612663432955742 + ], + "credit_map": [ + 0.016380634158849716, + 0.024685585871338844, + 0.039498358964920044, + 0.03251786157488823, + 0.021998563781380653, + 0.011912673711776733, + 0.020794803276658058, + 0.017986193299293518, + 0.008030232042074203, + 0.008825250901281834, + 0.012068246491253376, + 0.010061684064567089, + 0.010664960369467735, + 0.006903542205691338, + 0.008812814950942993, + 0.006717189680784941, + 0.023204363882541656, + 0.025443060323596, + 0.025943327695131302, + 0.035830412060022354, + 0.0334491990506649, + 0.03441188111901283, + 0.029588812962174416, + 0.017918717116117477, + 0.012522125616669655, + 0.006010778248310089, + 0.00751689076423645, + 0.0066288793459534645, + 0.009015467949211597, + 0.008954250253736973, + 0.005528089124709368, + 0.004186900798231363, + 0.02182815968990326, + 0.012967060320079327, + 0.012448348104953766, + 0.01274034846574068, + 0.01611144281923771, + 0.02028101123869419, + 0.012006031349301338, + 0.006173721980303526, + 0.006154338363558054, + 0.00721011171117425, + 0.0032996961381286383, + 0.005506129935383797, + 0.010699931532144547, + 0.010401234962046146, + 0.005699326284229755, + 0.00601884163916111, + 0.018841814249753952, + 0.03654706850647926, + 0.011559820733964443, + 0.012606707401573658, + 0.047859515994787216, + 0.029855241999030113, + 0.014656283892691135, + 0.016925612464547157, + 0.004303845576941967, + 0.0058167316019535065, + 0.004329056944698095, + 0.005704817362129688, + 0.007977264001965523, + 0.006020938511937857, + 0.0037121239583939314, + 0.006508271209895611, + 0.003019990399479866, + 0.0024017756804823875, + 0.0025413462426513433, + 0.003437838051468134, + 0.004242532420903444, + 0.002802007831633091, + 0.005043770186603069, + 0.0036015913356095552, + 0.003945906646549702, + 0.003479571547359228, + 0.003613315289840102, + 0.0029621487483382225, + 0.003246098989620805, + 0.003244700375944376, + 0.0023749645333737135, + 0.0032599312253296375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.03961043804883957 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062072597444057465 + ], + "opposite_logprob_tokens": [ + -2.812072515487671 + ], + "logprob_reward": -0.062072597444057465, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062072597444057465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24574361741542816, + 0.08368036150932312, + 0.13556039333343506, + 0.05546165630221367, + 0.0610281378030777, + 0.09547170996665955, + 0.07408875972032547, + 0.1251784861087799, + 0.03534531220793724, + 0.08844156563282013 + ], + "credit_map": [ + 0.020549485459923744, + 0.02898142673075199, + 0.04876183345913887, + 0.0484662763774395, + 0.026585375890135765, + 0.018839845433831215, + 0.03127790987491608, + 0.02228146232664585, + 0.007796822115778923, + 0.01056454423815012, + 0.010810024105012417, + 0.012882675975561142, + 0.01507287286221981, + 0.009408988058567047, + 0.010374480858445168, + 0.006769946310669184, + 0.011690696701407433, + 0.015560541301965714, + 0.01785152778029442, + 0.022468077018857002, + 0.018598631024360657, + 0.021552691236138344, + 0.01633843034505844, + 0.011499802581965923, + 0.008564223535358906, + 0.005235554184764624, + 0.006644519045948982, + 0.007073221728205681, + 0.007321590092033148, + 0.009621788747608662, + 0.007032629568129778, + 0.003968131262809038, + 0.008952941745519638, + 0.006696057040244341, + 0.00556667847558856, + 0.005899474024772644, + 0.00769914872944355, + 0.010857650078833103, + 0.006923199165612459, + 0.008432986214756966, + 0.012125534936785698, + 0.0147517966106534, + 0.004797678906470537, + 0.00805562175810337, + 0.024455586448311806, + 0.014982596039772034, + 0.006824510172009468, + 0.009478382766246796, + 0.00688977912068367, + 0.013253085315227509, + 0.0044355071149766445, + 0.004615211393684149, + 0.018713802099227905, + 0.01487748697400093, + 0.004981090314686298, + 0.006322793196886778, + 0.008955798111855984, + 0.015438315458595753, + 0.006937256548553705, + 0.016426989808678627, + 0.022524341940879822, + 0.011150825768709183, + 0.009971845895051956, + 0.03377310559153557, + 0.0046032266691327095, + 0.003536238567903638, + 0.0034537012688815594, + 0.0035940343514084816, + 0.0057239774614572525, + 0.003320522839203477, + 0.006464305799454451, + 0.004649306647479534, + 0.011697197332978249, + 0.015399743802845478, + 0.010132214054465294, + 0.006331432145088911, + 0.021865298971533775, + 0.010513434186577797, + 0.006902969907969236, + 0.0055992743000388145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.062072597444057465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293096095323563 + ], + "opposite_logprob_tokens": [ + -2.2379310131073 + ], + "logprob_reward": -0.11293096095323563, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293096095323563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15315976738929749, + 0.09041769802570343, + 0.15866225957870483, + 0.05998452007770538, + 0.11830559372901917, + 0.09352951496839523, + 0.07241974025964737, + 0.07129489630460739, + 0.06497077643871307, + 0.11725520342588425 + ], + "credit_map": [ + 0.0184865053743124, + 0.018438901752233505, + 0.03103548474609852, + 0.02212073653936386, + 0.01773589476943016, + 0.015497813932597637, + 0.016387570649385452, + 0.01345685962587595, + 0.010059705004096031, + 0.0108738262206316, + 0.013620752841234207, + 0.012265559285879135, + 0.016233088448643684, + 0.008517035283148289, + 0.011170608922839165, + 0.007677116896957159, + 0.015353059396147728, + 0.018200566992163658, + 0.01681879162788391, + 0.02796805277466774, + 0.020961713045835495, + 0.02479378692805767, + 0.020877322182059288, + 0.01368896197527647, + 0.005136619787663221, + 0.005378411151468754, + 0.005012203473597765, + 0.00865076296031475, + 0.007108128163963556, + 0.011789200827479362, + 0.010195054113864899, + 0.00671414053067565, + 0.01576387695968151, + 0.013859436847269535, + 0.009582608006894588, + 0.013215181417763233, + 0.017364490777254105, + 0.02363864704966545, + 0.011949420906603336, + 0.012931938283145428, + 0.009381081908941269, + 0.014700002036988735, + 0.0067908503115177155, + 0.009101051837205887, + 0.0165198165923357, + 0.01848571002483368, + 0.009580839425325394, + 0.008970159105956554, + 0.007919005118310452, + 0.015737304463982582, + 0.004229740705341101, + 0.004768501501530409, + 0.017080826684832573, + 0.008519220165908337, + 0.005528571549803019, + 0.008636566810309887, + 0.00532400980591774, + 0.0088824313133955, + 0.0038241660222411156, + 0.008641934022307396, + 0.01478324644267559, + 0.008443044498562813, + 0.005709442310035229, + 0.015686623752117157, + 0.009338908828794956, + 0.008826641365885735, + 0.007342033553868532, + 0.008344982750713825, + 0.007785686291754246, + 0.007793327793478966, + 0.008314045146107674, + 0.00722515070810914, + 0.014754760079085827, + 0.020657699555158615, + 0.012078349478542805, + 0.009478366002440453, + 0.02943735010921955, + 0.014941229484975338, + 0.009370798245072365, + 0.006536649074405432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.11293096095323563 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933678388595581 + ], + "opposite_logprob_tokens": [ + -0.6933678388595581 + ], + "logprob_reward": -0.6933678388595581, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933678388595581, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22765888273715973, + 0.08717898279428482, + 0.07372840493917465, + 0.056412436068058014, + 0.047668516635894775, + 0.061654895544052124, + 0.24404487013816833, + 0.09973981976509094, + 0.032117653638124466, + 0.06979548931121826 + ], + "credit_map": [ + 0.024896275252103806, + 0.023934846743941307, + 0.031236033886671066, + 0.03870941326022148, + 0.03144008666276932, + 0.030911345034837723, + 0.024852246046066284, + 0.021678637713193893, + 0.0074157994240522385, + 0.00821699108928442, + 0.010078276507556438, + 0.013150474987924099, + 0.022742176428437233, + 0.009846141561865807, + 0.00920021254569292, + 0.00652891444042325, + 0.005428323056548834, + 0.009982146322727203, + 0.008790351450443268, + 0.011228437535464764, + 0.011401922442018986, + 0.0118406405672431, + 0.009563025087118149, + 0.0054935566149652, + 0.01279545109719038, + 0.0068587674759328365, + 0.00833121221512556, + 0.00597882317379117, + 0.007829678244888783, + 0.006348340306431055, + 0.00432592025026679, + 0.003944246098399162, + 0.008486736565828323, + 0.005686021875590086, + 0.004812262021005154, + 0.0051804156973958015, + 0.006464008707553148, + 0.009884378872811794, + 0.003866256680339575, + 0.0032884348183870316, + 0.005194745026528835, + 0.006867080926895142, + 0.004102213308215141, + 0.004619439598172903, + 0.011856524273753166, + 0.019112246111035347, + 0.00547184469178319, + 0.004430803004652262, + 0.02974587120115757, + 0.05713526904582977, + 0.01293298602104187, + 0.01998167484998703, + 0.05723407119512558, + 0.027259809896349907, + 0.019490040838718414, + 0.020265156403183937, + 0.010404706932604313, + 0.012279881164431572, + 0.010833182372152805, + 0.012134038843214512, + 0.01612365059554577, + 0.01504820678383112, + 0.0067615327425301075, + 0.016154618933796883, + 0.005637794267386198, + 0.0038386310916393995, + 0.003939641173928976, + 0.0030547466594725847, + 0.004121868405491114, + 0.004061662591993809, + 0.004008846823126078, + 0.003454463556408882, + 0.011695349588990211, + 0.009722158312797546, + 0.008205162361264229, + 0.005385798402130604, + 0.011197746731340885, + 0.010672552511096, + 0.005822957493364811, + 0.0070937625132501125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.6933678388595581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306785389780998 + ], + "opposite_logprob_tokens": [ + -3.1680679321289062 + ], + "logprob_reward": -0.04306785389780998, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306785389780998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07217478007078171, + 0.051037684082984924, + 0.03704502433538437, + 0.024172531440854073, + 0.13426826894283295, + 0.03989546000957489, + 0.5771070122718811, + 0.02012460120022297, + 0.01696804352104664, + 0.02720659412443638 + ], + "credit_map": [ + 0.010507773607969284, + 0.006434558890759945, + 0.012333672493696213, + 0.008838570676743984, + 0.006941863801330328, + 0.01195768266916275, + 0.007648967206478119, + 0.007511686068028212, + 0.005782339256256819, + 0.005212043412029743, + 0.006138677708804607, + 0.006494581233710051, + 0.010363053530454636, + 0.008996192365884781, + 0.004534739535301924, + 0.003516058437526226, + 0.0030332128517329693, + 0.004922352731227875, + 0.00429873913526535, + 0.005618659779429436, + 0.006484366022050381, + 0.0058248694986104965, + 0.004548531491309404, + 0.002314291661605239, + 0.0028416463173925877, + 0.0029482506215572357, + 0.002831254852935672, + 0.0037973758298903704, + 0.0028058933094143867, + 0.003499323036521673, + 0.0030163221526890993, + 0.0024324674159288406, + 0.02382477931678295, + 0.016538133844733238, + 0.013050093315541744, + 0.01283044833689928, + 0.017445314675569534, + 0.030682971701025963, + 0.012520825490355492, + 0.007375708781182766, + 0.0033607706427574158, + 0.004928205627948046, + 0.00308468216098845, + 0.004528782796114683, + 0.00887612346559763, + 0.008616194128990173, + 0.003288817126303911, + 0.0032118866220116615, + 0.06832916289567947, + 0.13416007161140442, + 0.030907051637768745, + 0.04627201333642006, + 0.1367770880460739, + 0.06626302748918533, + 0.04691013693809509, + 0.04748845100402832, + 0.002712005516514182, + 0.003680497407913208, + 0.0015069146174937487, + 0.0017331726849079132, + 0.0031911463011056185, + 0.0029384493827819824, + 0.0019394810078665614, + 0.0024229346308857203, + 0.0028989834245294333, + 0.001957664033398032, + 0.0011714824941009283, + 0.0013735478278249502, + 0.0035571991465985775, + 0.0015422720462083817, + 0.0017795319436118007, + 0.002687362488359213, + 0.0032200003042817116, + 0.0041989050805568695, + 0.003867083229124546, + 0.0026761454064399004, + 0.00469609722495079, + 0.003693429986014962, + 0.0027529115322977304, + 0.002102021360769868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.04306785389780998 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot, imbued with subtle film grain for a cinematic quality, captures exhausted patients in a sterile hospital ward. With damp, matted hair plastered to their pale foreheads, the patients struggle to sit. They wear loose, light-blue, short-sleeved hospital gowns that hang limply from their weakened bodies. Their arms tremble noticeably, and a look of effort etched on their faces, their palms pressed deliberately against the mattress. As their torsos slowly rise, the wrinkled, thin, white hospital sheets, creased from restless sleep, slide down their chests and pool around their waists. The bed is a standard metal-framed model; in the slightly out-of-focus background, a metal IV rack stands sentry. The entire scene is illuminated by the cool, flat, unforgiving light of overhead fluorescent lamps, casting soft but sharp shadows that eliminate any sense of warmth and create a somber and deeply personal atmosphere filled with quiet struggle and vulnerability.", + "seed": 30898, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_1/index_2018_seed_30898_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12432523630559444, + "vlm_logprob": -0.12432523630559444, + "vlm_yes_no_margin": 2.474999988079071, + "vlm_token_credit": -0.12432523630559444, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04796530306339264 + ], + "opposite_logprob_tokens": [ + -3.5479652881622314 + ], + "logprob_reward": -0.04796530306339264, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04796530306339264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23462004959583282, + 0.14383088052272797, + 0.04642457142472267, + 0.11178826540708542, + 0.039023954421281815, + 0.0709976777434349, + 0.1059858575463295, + 0.07352054119110107, + 0.1386181265115738, + 0.03519007936120033 + ], + "credit_map": [ + 0.018786070868372917, + 0.029031170532107353, + 0.02886384166777134, + 0.03348550572991371, + 0.043344952166080475, + 0.0233494620770216, + 0.03174426406621933, + 0.026014773175120354, + 0.012077424675226212, + 0.019532864913344383, + 0.014402547851204872, + 0.027275823056697845, + 0.01904238387942314, + 0.017477771267294884, + 0.01561641599982977, + 0.018405651673674583, + 0.004136273637413979, + 0.005816340912133455, + 0.005455842241644859, + 0.009239749982953072, + 0.006178748328238726, + 0.005442243535071611, + 0.006567679345607758, + 0.0035876939073204994, + 0.009764635935425758, + 0.013356518000364304, + 0.012969240546226501, + 0.009268069639801979, + 0.023422805592417717, + 0.012608028016984463, + 0.013632997870445251, + 0.016765963286161423, + 0.0035853213630616665, + 0.0057040490210056305, + 0.004236051347106695, + 0.004912275820970535, + 0.007774452678859234, + 0.005634498316794634, + 0.004016889724880457, + 0.003160417079925537, + 0.008643554523587227, + 0.00959774386137724, + 0.006879117805510759, + 0.008249013684689999, + 0.011571625247597694, + 0.009650672785937786, + 0.009181196801364422, + 0.007224752567708492, + 0.00889491755515337, + 0.013460840098559856, + 0.01091061532497406, + 0.015801748260855675, + 0.011429413221776485, + 0.01402368862181902, + 0.015799429267644882, + 0.015665195882320404, + 0.006508716382086277, + 0.012422150000929832, + 0.006867693271487951, + 0.007329284213483334, + 0.012793834321200848, + 0.008398977108299732, + 0.01311121042817831, + 0.006088677328079939, + 0.009282619692385197, + 0.029542645439505577, + 0.010608669370412827, + 0.007387990597635508, + 0.016011277213692665, + 0.013444057665765285, + 0.04542260244488716, + 0.00691825058311224, + 0.0035488875582814217, + 0.003954935818910599, + 0.005366164725273848, + 0.004045978654175997, + 0.004513952881097794, + 0.0033732156734913588, + 0.006162113510072231, + 0.004224829841405153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.04796530306339264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004715502262115 + ], + "opposite_logprob_tokens": [ + -2.69504714012146 + ], + "logprob_reward": -0.07004715502262115, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004715502262115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15724056959152222, + 0.09796005487442017, + 0.04554080590605736, + 0.045351848006248474, + 0.06761261075735092, + 0.29632893204689026, + 0.12570405006408691, + 0.07089914381504059, + 0.04974262788891792, + 0.04361940920352936 + ], + "credit_map": [ + 0.010831824503839016, + 0.01779794506728649, + 0.020614907145500183, + 0.0299823060631752, + 0.02195744961500168, + 0.015082979574799538, + 0.023201512172818184, + 0.017771650105714798, + 0.009581753984093666, + 0.018150223419070244, + 0.010042307898402214, + 0.012800687924027443, + 0.013741063885390759, + 0.013374723494052887, + 0.012796608731150627, + 0.007472681347280741, + 0.004946527536958456, + 0.00819682702422142, + 0.0038622552528977394, + 0.006122528109699488, + 0.006265598349273205, + 0.005673517473042011, + 0.006324229296296835, + 0.004149320535361767, + 0.003958340268582106, + 0.00722169317305088, + 0.004565742798149586, + 0.0032060854136943817, + 0.006338491104543209, + 0.008486165665090084, + 0.005795255769044161, + 0.005780071951448917, + 0.006312315817922354, + 0.011022810824215412, + 0.0071730115450918674, + 0.008098735474050045, + 0.012653199024498463, + 0.008521956391632557, + 0.00862012431025505, + 0.005210460629314184, + 0.04013598710298538, + 0.041213855147361755, + 0.024045223370194435, + 0.02909320965409279, + 0.051191601902246475, + 0.0450034998357296, + 0.036776117980480194, + 0.028869453817605972, + 0.013136176392436028, + 0.012784389778971672, + 0.013835610821843147, + 0.01800082065165043, + 0.013743781484663486, + 0.02233417145907879, + 0.015308688394725323, + 0.016560401767492294, + 0.005856106989085674, + 0.00971818808466196, + 0.0068185823038220406, + 0.007676528766751289, + 0.01375682931393385, + 0.007653457578271627, + 0.014829698018729687, + 0.004589749034494162, + 0.003286972176283598, + 0.009980968199670315, + 0.003398254280909896, + 0.005113642662763596, + 0.005134090315550566, + 0.004318211693316698, + 0.01570071093738079, + 0.0028097741305828094, + 0.004075868055224419, + 0.008060801774263382, + 0.0059214127250015736, + 0.0056266142055392265, + 0.006041303277015686, + 0.005159722175449133, + 0.004132696893066168, + 0.00460099009796977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.07004715502262115 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603582501411438 + ], + "opposite_logprob_tokens": [ + -1.910358190536499 + ], + "logprob_reward": -0.1603582501411438, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603582501411438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1596306562423706, + 0.05550464987754822, + 0.032230209559202194, + 0.07579651474952698, + 0.09563390165567398, + 0.11421585828065872, + 0.2239668369293213, + 0.13055376708507538, + 0.055414892733097076, + 0.05705276504158974 + ], + "credit_map": [ + 0.013143474236130714, + 0.019108232110738754, + 0.01757337711751461, + 0.025648213922977448, + 0.024397818371653557, + 0.015207476913928986, + 0.02616562321782112, + 0.01838642545044422, + 0.004539770074188709, + 0.008329630829393864, + 0.006557640619575977, + 0.009405885823071003, + 0.007845722138881683, + 0.00679948041215539, + 0.007034490816295147, + 0.004992026370018721, + 0.004213996697217226, + 0.004882109817117453, + 0.003562271362170577, + 0.003894560504704714, + 0.003985769581049681, + 0.004383125342428684, + 0.004184768535196781, + 0.00312360399402678, + 0.007225756999105215, + 0.009732704609632492, + 0.009144873358309269, + 0.005601149518042803, + 0.014689627103507519, + 0.00961625948548317, + 0.008439159952104092, + 0.01134698186069727, + 0.009379181079566479, + 0.015425984747707844, + 0.011044418439269066, + 0.011989421211183071, + 0.017490530386567116, + 0.012139460071921349, + 0.011218513362109661, + 0.006946386303752661, + 0.014497478492558002, + 0.01591077260673046, + 0.009862465783953667, + 0.010680792853236198, + 0.019503487274050713, + 0.016518065705895424, + 0.015665121376514435, + 0.011577668599784374, + 0.023696834221482277, + 0.02378460392355919, + 0.024345947429537773, + 0.03128628805279732, + 0.024602370336651802, + 0.045476898550987244, + 0.023899376392364502, + 0.026874512434005737, + 0.008973830379545689, + 0.014428653754293919, + 0.013091547414660454, + 0.008753793314099312, + 0.030770668759942055, + 0.011877832002937794, + 0.034654710441827774, + 0.00800272636115551, + 0.005176952574402094, + 0.007091559935361147, + 0.006276196800172329, + 0.004744161386042833, + 0.006239207927137613, + 0.007259623147547245, + 0.014746122993528843, + 0.0038810677360743284, + 0.00556224538013339, + 0.005955478176474571, + 0.0073653399012982845, + 0.006046718917787075, + 0.0067816609516739845, + 0.004421419929713011, + 0.015153330750763416, + 0.005766573827713728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.1603582501411438 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31345587968826294 + ], + "opposite_logprob_tokens": [ + -1.3134558200836182 + ], + "logprob_reward": -0.31345587968826294, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31345587968826294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16628916561603546, + 0.14637604355812073, + 0.052338678389787674, + 0.1446230560541153, + 0.06723607331514359, + 0.157424658536911, + 0.11768069118261337, + 0.04615873843431473, + 0.06001381576061249, + 0.04185904189944267 + ], + "credit_map": [ + 0.013510243967175484, + 0.01908436231315136, + 0.024499235674738884, + 0.024548843502998352, + 0.030987795442342758, + 0.015638498589396477, + 0.02224622666835785, + 0.015773970633745193, + 0.013083403930068016, + 0.027215158566832542, + 0.013745811767876148, + 0.023732200264930725, + 0.020841514691710472, + 0.019985832273960114, + 0.014904147945344448, + 0.01286796759814024, + 0.006738077383488417, + 0.005581695586442947, + 0.008382136933505535, + 0.00949952658265829, + 0.005628072191029787, + 0.006052060984075069, + 0.005843726918101311, + 0.004613380879163742, + 0.009718775749206543, + 0.016105487942695618, + 0.016155267134308815, + 0.01034495234489441, + 0.031888969242572784, + 0.01851825788617134, + 0.017202796414494514, + 0.02468855492770672, + 0.006316137500107288, + 0.007950231432914734, + 0.007174062542617321, + 0.010442223399877548, + 0.011608486995100975, + 0.009889108128845692, + 0.01037843618541956, + 0.003477388294413686, + 0.023307953029870987, + 0.0249759703874588, + 0.014139825478196144, + 0.012817278504371643, + 0.027558818459510803, + 0.021722348406910896, + 0.01880924589931965, + 0.014093222096562386, + 0.010113206692039967, + 0.01448873057961464, + 0.011874351650476456, + 0.020035790279507637, + 0.015205384232103825, + 0.01542612910270691, + 0.01824161596596241, + 0.012295479886233807, + 0.005569717846810818, + 0.005656341090798378, + 0.008273988962173462, + 0.004062554333359003, + 0.007309603504836559, + 0.003808518871665001, + 0.007210879120975733, + 0.004267134238034487, + 0.00547467777505517, + 0.007153226528316736, + 0.007716428488492966, + 0.003782553132623434, + 0.006585963070392609, + 0.0077376277185976505, + 0.01730380952358246, + 0.004259532783180475, + 0.003040388925001025, + 0.00726022943854332, + 0.0037724915891885757, + 0.0051049599424004555, + 0.006083765998482704, + 0.005245002917945385, + 0.007102171424776316, + 0.004250030964612961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.31345587968826294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979959361255169 + ], + "opposite_logprob_tokens": [ + -3.529799699783325 + ], + "logprob_reward": -0.02979959361255169, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979959361255169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11620023846626282, + 0.05142701789736748, + 0.049965765327215195, + 0.1002659946680069, + 0.1239703819155693, + 0.19840148091316223, + 0.10291416943073273, + 0.08328379690647125, + 0.1407960057258606, + 0.032775189727544785 + ], + "credit_map": [ + 0.00982542335987091, + 0.011167098768055439, + 0.01881164312362671, + 0.023293355479836464, + 0.019159577786922455, + 0.01048838160932064, + 0.011690854094922543, + 0.01176389865577221, + 0.00578272482380271, + 0.007179656997323036, + 0.0053986236453056335, + 0.006450689863413572, + 0.008275344967842102, + 0.0062162671238183975, + 0.004535829182714224, + 0.007587878964841366, + 0.007111572660505772, + 0.00608123978599906, + 0.007069869432598352, + 0.0070635490119457245, + 0.007476259022951126, + 0.00577706890180707, + 0.004334839526563883, + 0.005051359999924898, + 0.010496857576072216, + 0.013532937504351139, + 0.009183415211737156, + 0.0075171934440732, + 0.023369010537862778, + 0.013642887584865093, + 0.008331180550158024, + 0.014192505739629269, + 0.011626104824244976, + 0.01857873611152172, + 0.013281727209687233, + 0.015568893402814865, + 0.019096016883850098, + 0.0177056435495615, + 0.017263488844037056, + 0.010849768295884132, + 0.02611042931675911, + 0.02636505663394928, + 0.014304938726127148, + 0.01875395141541958, + 0.03451652452349663, + 0.03470074012875557, + 0.025494905188679695, + 0.0181549321860075, + 0.009536097757518291, + 0.012139802798628807, + 0.00842183455824852, + 0.014445839449763298, + 0.014885631389915943, + 0.01437668688595295, + 0.017302682623267174, + 0.011805589310824871, + 0.005258188117295504, + 0.009480297565460205, + 0.0052450899966061115, + 0.005676182918250561, + 0.018809253349900246, + 0.008529474027454853, + 0.024055540561676025, + 0.006229768972843885, + 0.017503606155514717, + 0.022829104214906693, + 0.012154234573245049, + 0.012266339734196663, + 0.014818103983998299, + 0.01539069414138794, + 0.03471466898918152, + 0.011119253002107143, + 0.003105523996055126, + 0.004929924849420786, + 0.004832395352423191, + 0.003400741843506694, + 0.005233129486441612, + 0.0032947328872978687, + 0.004258797969669104, + 0.003719939151778817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.02979959361255169 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot, imbued with subtle film grain for a cinematic quality, captures exhausted patients in a sterile hospital ward. With damp, matted hair plastered to their pale foreheads, the patients struggle to sit. They wear loose, light-blue, short-sleeved hospital gowns that hang limply from their weakened bodies. Their arms tremble noticeably, and a look of effort etched on their faces, their palms pressed deliberately against the mattress. As their torsos slowly rise, the wrinkled, thin, white hospital sheets, creased from restless sleep, slide down their chests and pool around their waists. The bed is a standard metal-framed model; in the slightly out-of-focus background, a metal IV rack stands sentry. The entire scene is illuminated by the cool, flat, unforgiving light of overhead fluorescent lamps, casting soft but sharp shadows that eliminate any sense of warmth and create a somber and deeply personal atmosphere filled with quiet struggle and vulnerability.", + "seed": 30899, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_1/index_2018_seed_30899_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10599475502967834, + "vlm_logprob": -0.10599475502967834, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.10599475502967834, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056785959750413895 + ], + "opposite_logprob_tokens": [ + -3.431786060333252 + ], + "logprob_reward": -0.056785959750413895, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.056785959750413895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19343362748622894, + 0.07832308113574982, + 0.050317443907260895, + 0.18048100173473358, + 0.08963193744421005, + 0.07166700810194016, + 0.09154944121837616, + 0.047163624316453934, + 0.12004595249891281, + 0.07738687098026276 + ], + "credit_map": [ + 0.02731313183903694, + 0.025493571534752846, + 0.01816035620868206, + 0.024864355102181435, + 0.032552074640989304, + 0.02270679734647274, + 0.023902229964733124, + 0.01844111457467079, + 0.009478369727730751, + 0.010961629450321198, + 0.005196264013648033, + 0.007319241762161255, + 0.012888407334685326, + 0.014053946360945702, + 0.010107879526913166, + 0.008317350409924984, + 0.010561146773397923, + 0.006193453911691904, + 0.00563409086316824, + 0.004613521974533796, + 0.004642529413104057, + 0.006696450058370829, + 0.006725889630615711, + 0.005250357557088137, + 0.02663705311715603, + 0.032806601375341415, + 0.015747344121336937, + 0.01862798072397709, + 0.021742619574069977, + 0.027631022036075592, + 0.024673542007803917, + 0.012614840641617775, + 0.012719133868813515, + 0.01199229247868061, + 0.008993064053356647, + 0.009876369498670101, + 0.010895938612520695, + 0.009836624376475811, + 0.012308728881180286, + 0.013009784743189812, + 0.007305316627025604, + 0.011154474690556526, + 0.005215358454734087, + 0.012371392920613289, + 0.01521978061646223, + 0.005363586358726025, + 0.009291392751038074, + 0.00574570382013917, + 0.00641448050737381, + 0.025332797318696976, + 0.004385697655379772, + 0.026559418067336082, + 0.01122453436255455, + 0.004869681317359209, + 0.00687007512897253, + 0.005892749410122633, + 0.008618759922683239, + 0.005643877666443586, + 0.004039235878735781, + 0.0063765025697648525, + 0.005959067959338427, + 0.006100794300436974, + 0.004636175464838743, + 0.005789207294583321, + 0.01736554689705372, + 0.015560821630060673, + 0.016946515068411827, + 0.013270740397274494, + 0.014184368774294853, + 0.01605292782187462, + 0.014156801626086235, + 0.012508231215178967, + 0.014424558728933334, + 0.017128976061940193, + 0.007630789652466774, + 0.007097290363162756, + 0.010436383076012135, + 0.008537657558918, + 0.006339836400002241, + 0.005791387055069208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.056785959750413895 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005082815885544 + ], + "opposite_logprob_tokens": [ + -2.6950507164001465 + ], + "logprob_reward": -0.07005082815885544, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005082815885544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1553633064031601, + 0.07842082530260086, + 0.04691462591290474, + 0.07008209824562073, + 0.12398489564657211, + 0.10619501024484634, + 0.185686394572258, + 0.06162266433238983, + 0.10608452558517456, + 0.06564562767744064 + ], + "credit_map": [ + 0.024489402770996094, + 0.017636116594076157, + 0.02172652818262577, + 0.016885114833712578, + 0.028195537626743317, + 0.018879152834415436, + 0.014854731038212776, + 0.01269671693444252, + 0.009081616997718811, + 0.016386782750487328, + 0.004261125810444355, + 0.007742313668131828, + 0.01058870553970337, + 0.007929367944598198, + 0.014556905254721642, + 0.007874002680182457, + 0.008944152854382992, + 0.005620632320642471, + 0.007090059574693441, + 0.0064303213730454445, + 0.004737180192023516, + 0.003960542380809784, + 0.006390125025063753, + 0.003741608466953039, + 0.009919226169586182, + 0.01244870014488697, + 0.0067201838828623295, + 0.00823682639747858, + 0.008198871277272701, + 0.009061289951205254, + 0.010071314871311188, + 0.005425684154033661, + 0.014476899988949299, + 0.022337228059768677, + 0.009759059175848961, + 0.009047490544617176, + 0.02688979171216488, + 0.011672934517264366, + 0.013025733642280102, + 0.016775760799646378, + 0.010685904882848263, + 0.019676294177770615, + 0.007466649170964956, + 0.014345331117510796, + 0.02702384628355503, + 0.006821285467594862, + 0.011289539746940136, + 0.008886163122951984, + 0.009416130371391773, + 0.058422524482011795, + 0.005950398277491331, + 0.06552610546350479, + 0.020819641649723053, + 0.0055562411434948444, + 0.012110553681850433, + 0.007884812541306019, + 0.0062666162848472595, + 0.007763100787997246, + 0.006460323464125395, + 0.009574510157108307, + 0.009688878431916237, + 0.007820363156497478, + 0.006447094492614269, + 0.0076017784886062145, + 0.009838740341365337, + 0.014466466382145882, + 0.009765501134097576, + 0.01334321592003107, + 0.016033703461289406, + 0.013829766772687435, + 0.01748495362699032, + 0.01132218074053526, + 0.005677912384271622, + 0.01012770738452673, + 0.0047037904150784016, + 0.00645653810352087, + 0.009367546997964382, + 0.012051978148519993, + 0.009685833007097244, + 0.007574317045509815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.07005082815885544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129295751452446 + ], + "opposite_logprob_tokens": [ + -2.237929582595825 + ], + "logprob_reward": -0.1129295751452446, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129295751452446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17723532021045685, + 0.06637060642242432, + 0.08265151083469391, + 0.09650003165006638, + 0.10201706737279892, + 0.07589637488126755, + 0.048925865441560745, + 0.03723234683275223, + 0.13501441478729248, + 0.17815643548965454 + ], + "credit_map": [ + 0.02506955899298191, + 0.024417027831077576, + 0.02426866628229618, + 0.02119104377925396, + 0.029763611033558846, + 0.021562103182077408, + 0.017966559156775475, + 0.012996749021112919, + 0.0067212386056780815, + 0.011329807341098785, + 0.004446603823453188, + 0.009097205474972725, + 0.009558415971696377, + 0.007499877363443375, + 0.011358336545526981, + 0.00635911850258708, + 0.015258589759469032, + 0.011522716842591763, + 0.00848007295280695, + 0.01072815153747797, + 0.009736234322190285, + 0.008171153254806995, + 0.009658023715019226, + 0.009096572175621986, + 0.015121168456971645, + 0.01496172696352005, + 0.012319761328399181, + 0.013001857325434685, + 0.010344994254410267, + 0.008443456143140793, + 0.011690834537148476, + 0.010616229847073555, + 0.011691178195178509, + 0.01775178126990795, + 0.007273138966411352, + 0.009831773117184639, + 0.01866759918630123, + 0.010616523213684559, + 0.011146696284413338, + 0.01503837388008833, + 0.0075442735105752945, + 0.01254061795771122, + 0.005494997370988131, + 0.011129341088235378, + 0.015261674299836159, + 0.00844153668731451, + 0.008060296066105366, + 0.007423642557114363, + 0.004136044532060623, + 0.014208700507879257, + 0.0034660836681723595, + 0.0121011259034276, + 0.005662078969180584, + 0.0026334018912166357, + 0.0032834638841450214, + 0.003434965619817376, + 0.0062597221694886684, + 0.004925522953271866, + 0.0036874052602797747, + 0.004479037597775459, + 0.005668194964528084, + 0.004451098386198282, + 0.0036096421536058187, + 0.004151723813265562, + 0.012758860364556313, + 0.020630989223718643, + 0.014555605128407478, + 0.019254107028245926, + 0.01909603178501129, + 0.01319708488881588, + 0.022747566923499107, + 0.012774172239005566, + 0.01373344287276268, + 0.02338222786784172, + 0.007219801656901836, + 0.015400375239551067, + 0.029009008780121803, + 0.049792446196079254, + 0.022944152355194092, + 0.016674982383847237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.1129295751452446 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211137533187866 + ], + "opposite_logprob_tokens": [ + -1.5021114349365234 + ], + "logprob_reward": -0.25211137533187866, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211137533187866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29029661417007446, + 0.07190617173910141, + 0.11245793849229813, + 0.07683919370174408, + 0.05145132914185524, + 0.14568430185317993, + 0.10077168792486191, + 0.03989279270172119, + 0.06349605321884155, + 0.0472039096057415 + ], + "credit_map": [ + 0.039567235857248306, + 0.04164234176278114, + 0.023833001032471657, + 0.029655562713742256, + 0.04543495923280716, + 0.036155425012111664, + 0.04189227521419525, + 0.032115813344717026, + 0.007505102548748255, + 0.01148257963359356, + 0.009627314284443855, + 0.007226179353892803, + 0.010484812781214714, + 0.010854693129658699, + 0.009338168427348137, + 0.0053873201832175255, + 0.02454308234155178, + 0.011961604468524456, + 0.012301390059292316, + 0.010709079913794994, + 0.011966826394200325, + 0.012718320824205875, + 0.016644343733787537, + 0.011613287031650543, + 0.010714374482631683, + 0.013341978192329407, + 0.00863032229244709, + 0.007494806312024593, + 0.011667649261653423, + 0.00790125411003828, + 0.010207765735685825, + 0.006881044711917639, + 0.006914000492542982, + 0.010019121691584587, + 0.004703948274254799, + 0.004822990857064724, + 0.008161901496350765, + 0.004636250901967287, + 0.005880981683731079, + 0.006312134210020304, + 0.018976734951138496, + 0.029847199097275734, + 0.009237699210643768, + 0.023383157327771187, + 0.02863319031894207, + 0.010740622878074646, + 0.01471452321857214, + 0.010151165537536144, + 0.01193175371736288, + 0.01873890683054924, + 0.00957880262285471, + 0.018159113824367523, + 0.014295371249318123, + 0.008545775897800922, + 0.01056752447038889, + 0.00895443931221962, + 0.007379625458270311, + 0.0052031404338777065, + 0.0024345098063349724, + 0.004927446600049734, + 0.0055801295675337315, + 0.0051308865658938885, + 0.004475354682654142, + 0.0047616963274776936, + 0.006915195379406214, + 0.006763870362192392, + 0.007078668102622032, + 0.008345757611095905, + 0.007834761403501034, + 0.011373715475201607, + 0.007627357263118029, + 0.007556725759059191, + 0.006055349949747324, + 0.006750311236828566, + 0.0028284513391554356, + 0.005068106576800346, + 0.007007230538874865, + 0.009664957411587238, + 0.005407084710896015, + 0.004422419238835573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.25211137533187866 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809603676199913 + ], + "opposite_logprob_tokens": [ + -3.2880959510803223 + ], + "logprob_reward": -0.03809603676199913, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809603676199913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11994113773107529, + 0.08418168872594833, + 0.09459098428487778, + 0.09901754558086395, + 0.05263148993253708, + 0.10604734718799591, + 0.2380293756723404, + 0.1054517850279808, + 0.05555184930562973, + 0.044556789100170135 + ], + "credit_map": [ + 0.02070070430636406, + 0.015633976086974144, + 0.009956811554729939, + 0.015618045814335346, + 0.019409330561757088, + 0.014596468769013882, + 0.013295377604663372, + 0.010730434209108353, + 0.011116783134639263, + 0.012719150632619858, + 0.00670901034027338, + 0.01043288130313158, + 0.010492128320038319, + 0.007695185951888561, + 0.011660025455057621, + 0.01335652731359005, + 0.016015753149986267, + 0.011054357513785362, + 0.00961469765752554, + 0.012993833981454372, + 0.013796784915030003, + 0.009168985299766064, + 0.013773101381957531, + 0.00817346666008234, + 0.02424953319132328, + 0.01163148507475853, + 0.006446609273552895, + 0.012213646434247494, + 0.012512183748185635, + 0.009666098281741142, + 0.011535419151186943, + 0.010762574151158333, + 0.007873129099607468, + 0.011934816837310791, + 0.0029534855857491493, + 0.003968331962823868, + 0.010983575135469437, + 0.004601423628628254, + 0.004540512803941965, + 0.005776217672973871, + 0.011781807988882065, + 0.026480108499526978, + 0.004490469116717577, + 0.010559100657701492, + 0.03014911524951458, + 0.009185533970594406, + 0.007298758253455162, + 0.006102452520281076, + 0.015676872804760933, + 0.0748315155506134, + 0.008537773974239826, + 0.07644891738891602, + 0.0289115272462368, + 0.008852344006299973, + 0.013812321238219738, + 0.010958111844956875, + 0.017392510548233986, + 0.010939043015241623, + 0.010495754890143871, + 0.015926068648695946, + 0.01424138993024826, + 0.013292455114424229, + 0.012056011706590652, + 0.011108551174402237, + 0.0064496835693717, + 0.007213904056698084, + 0.004832625389099121, + 0.007531819399446249, + 0.007562709506601095, + 0.008142094127833843, + 0.008003778755664825, + 0.005815234500914812, + 0.004795360378921032, + 0.00743901077657938, + 0.003355126129463315, + 0.004694892559200525, + 0.006509730592370033, + 0.0057492926716804504, + 0.007159121334552765, + 0.004854259546846151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.03809603676199913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot, imbued with subtle film grain for a cinematic quality, captures exhausted patients in a sterile hospital ward. With damp, matted hair plastered to their pale foreheads, the patients struggle to sit. They wear loose, light-blue, short-sleeved hospital gowns that hang limply from their weakened bodies. Their arms tremble noticeably, and a look of effort etched on their faces, their palms pressed deliberately against the mattress. As their torsos slowly rise, the wrinkled, thin, white hospital sheets, creased from restless sleep, slide down their chests and pool around their waists. The bed is a standard metal-framed model; in the slightly out-of-focus background, a metal IV rack stands sentry. The entire scene is illuminated by the cool, flat, unforgiving light of overhead fluorescent lamps, casting soft but sharp shadows that eliminate any sense of warmth and create a somber and deeply personal atmosphere filled with quiet struggle and vulnerability.", + "seed": 30900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_1/index_2018_seed_30900_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15433891750872136, + "vlm_logprob": -0.15433891750872136, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.15433891750872136, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030283605679869652 + ], + "opposite_logprob_tokens": [ + -4.280283451080322 + ], + "logprob_reward": -0.030283605679869652, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.030283605679869652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12731541693210602, + 0.15422333776950836, + 0.06051066145300865, + 0.05444430187344551, + 0.02618299052119255, + 0.07969478517770767, + 0.13135729730129242, + 0.04395144805312157, + 0.04511527344584465, + 0.2772044837474823 + ], + "credit_map": [ + 0.01625269465148449, + 0.01965891383588314, + 0.012244747020304203, + 0.0183453056961298, + 0.00837346538901329, + 0.012075472623109818, + 0.016797566786408424, + 0.023567257449030876, + 0.01463305950164795, + 0.022787652909755707, + 0.01729029044508934, + 0.01937929540872574, + 0.016110876575112343, + 0.016258735209703445, + 0.024606138467788696, + 0.023157309740781784, + 0.0037442531902343035, + 0.00980739202350378, + 0.004778437316417694, + 0.011246773414313793, + 0.007084817625582218, + 0.005747789982706308, + 0.008971726521849632, + 0.009129480458796024, + 0.006163002923130989, + 0.006486011669039726, + 0.0038542093243449926, + 0.011040976271033287, + 0.006078389938920736, + 0.005461405031383038, + 0.009495774284005165, + 0.005864537786692381, + 0.002848230069503188, + 0.004027626011520624, + 0.0019396769348531961, + 0.004357040859758854, + 0.004255624953657389, + 0.0021895035170018673, + 0.004189746920019388, + 0.0023755463771522045, + 0.015540448017418385, + 0.010375766083598137, + 0.007861444726586342, + 0.012952650897204876, + 0.006705192849040031, + 0.008538785390555859, + 0.006878111977130175, + 0.010842392221093178, + 0.0066443877294659615, + 0.03926056623458862, + 0.005279644392430782, + 0.037572309374809265, + 0.013738507404923439, + 0.007586811203509569, + 0.014820895157754421, + 0.0064542037434875965, + 0.0051721143536269665, + 0.008847721852362156, + 0.004503875505179167, + 0.006364533677697182, + 0.005024243611842394, + 0.004195116925984621, + 0.004832446575164795, + 0.005011398810893297, + 0.004299422260373831, + 0.008047792129218578, + 0.0038391968701034784, + 0.007210581563413143, + 0.004876525606960058, + 0.0030488097108900547, + 0.00704916613176465, + 0.006743782199919224, + 0.011269835755228996, + 0.033186741173267365, + 0.010291374288499355, + 0.10812881588935852, + 0.013231021352112293, + 0.006442144978791475, + 0.08266457170248032, + 0.011989988386631012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.030283605679869652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812575712800026 + ], + "opposite_logprob_tokens": [ + -3.28812575340271 + ], + "logprob_reward": -0.03812575712800026, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812575712800026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22518935799598694, + 0.13068921864032745, + 0.06372646242380142, + 0.02895434945821762, + 0.06467454880475998, + 0.09051985293626785, + 0.06466451287269592, + 0.05015411600470543, + 0.09016767144203186, + 0.19125986099243164 + ], + "credit_map": [ + 0.019388120621442795, + 0.051855966448783875, + 0.009519296698272228, + 0.056950319558382034, + 0.018916022032499313, + 0.013341664336621761, + 0.03428396210074425, + 0.020934011787176132, + 0.01234125904738903, + 0.02561391517519951, + 0.012019386515021324, + 0.021802155300974846, + 0.0142038743942976, + 0.010954789817333221, + 0.021805981174111366, + 0.01194786187261343, + 0.004152069799602032, + 0.01598023995757103, + 0.00505708297714591, + 0.0076262084767222404, + 0.009336992166936398, + 0.006611405406147242, + 0.00594634935259819, + 0.009016113355755806, + 0.0034646508283913136, + 0.003830790286883712, + 0.0025589223951101303, + 0.004011737648397684, + 0.0032046029809862375, + 0.003383124014362693, + 0.005629592575132847, + 0.0028709268663078547, + 0.00975048914551735, + 0.008682872168719769, + 0.004643991123884916, + 0.00885059405118227, + 0.008554336614906788, + 0.008065899834036827, + 0.009616278111934662, + 0.006510086823254824, + 0.01318618655204773, + 0.011190150864422321, + 0.009397769346833229, + 0.017488809302449226, + 0.010502138175070286, + 0.00868851225823164, + 0.00936735887080431, + 0.010698927566409111, + 0.003629385493695736, + 0.014869021251797676, + 0.0028855709824711084, + 0.01816355064511299, + 0.006123879924416542, + 0.0036475134547799826, + 0.010645360685884953, + 0.00470022764056921, + 0.006389250513166189, + 0.00835100281983614, + 0.005699300207197666, + 0.008132928982377052, + 0.006146573927253485, + 0.005429493263363838, + 0.006404428277164698, + 0.0036011366173624992, + 0.009574471972882748, + 0.022845463827252388, + 0.006098526995629072, + 0.01232500746846199, + 0.010668283328413963, + 0.005196500103920698, + 0.011441389098763466, + 0.01201802771538496, + 0.012243908829987049, + 0.034990306943655014, + 0.01005389541387558, + 0.05997076258063316, + 0.01502485666424036, + 0.008322359062731266, + 0.03561192750930786, + 0.015041843988001347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.03812575712800026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902986347675323 + ], + "opposite_logprob_tokens": [ + -2.4640297889709473 + ], + "logprob_reward": -0.08902986347675323, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902986347675323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09380988776683807, + 0.08419469743967056, + 0.10296666622161865, + 0.08929766714572906, + 0.07318504899740219, + 0.047975216060876846, + 0.31383201479911804, + 0.04205610230565071, + 0.042333416640758514, + 0.11034927517175674 + ], + "credit_map": [ + 0.010925637558102608, + 0.013624634593725204, + 0.005809119436889887, + 0.023896189406514168, + 0.005241622216999531, + 0.008118059486150742, + 0.015086655505001545, + 0.011107977479696274, + 0.007712700869888067, + 0.010644693858921528, + 0.011945043690502644, + 0.015422605909407139, + 0.008387967944145203, + 0.007922281511127949, + 0.01216956041753292, + 0.009989834390580654, + 0.006477159447968006, + 0.015508211217820644, + 0.005476959049701691, + 0.02194984443485737, + 0.009532121010124683, + 0.008103957399725914, + 0.011529206298291683, + 0.024389207363128662, + 0.009048479609191418, + 0.012484312988817692, + 0.005285911727696657, + 0.011814978905022144, + 0.011814381927251816, + 0.009116411209106445, + 0.017308510839939117, + 0.012424681335687637, + 0.006393634248524904, + 0.011174974031746387, + 0.004139939788728952, + 0.016726555302739143, + 0.009756726212799549, + 0.005955939646810293, + 0.011652805842459202, + 0.007384480442851782, + 0.005373797845095396, + 0.008112436160445213, + 0.004753084387630224, + 0.006637556944042444, + 0.008738476783037186, + 0.003639754606410861, + 0.004913034848868847, + 0.005807071458548307, + 0.01570904441177845, + 0.09838999807834625, + 0.010734436102211475, + 0.09253959357738495, + 0.03296200931072235, + 0.015502492897212505, + 0.033787328749895096, + 0.014207101427018642, + 0.004328983835875988, + 0.007751436438411474, + 0.004100664518773556, + 0.00609158119186759, + 0.0048425765708088875, + 0.005326695274561644, + 0.00592707097530365, + 0.0036870944313704967, + 0.003469684859737754, + 0.005463503301143646, + 0.0029814252629876137, + 0.009259208105504513, + 0.004848102107644081, + 0.0030121318995952606, + 0.007467420306056738, + 0.005831941030919552, + 0.004900066647678614, + 0.01031818799674511, + 0.004877552855759859, + 0.038102515041828156, + 0.007058935705572367, + 0.0051481290720403194, + 0.02845505252480507, + 0.011488834396004677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.08902986347675323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761569142341614 + ], + "opposite_logprob_tokens": [ + -0.8261569142341614 + ], + "logprob_reward": -0.5761569142341614, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761569142341614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16265490651130676, + 0.06676532328128815, + 0.06635735183954239, + 0.02540486864745617, + 0.04979817569255829, + 0.12027458101511002, + 0.0961453765630722, + 0.029253773391246796, + 0.02304781787097454, + 0.3602977991104126 + ], + "credit_map": [ + 0.018231702968478203, + 0.018364746123552322, + 0.01454603485763073, + 0.04245994612574577, + 0.007871445268392563, + 0.013729983009397984, + 0.024492505937814713, + 0.022958537563681602, + 0.00958947278559208, + 0.0103585971519351, + 0.006395986769348383, + 0.009008617140352726, + 0.008494477719068527, + 0.0056460220366716385, + 0.008398914709687233, + 0.008873235434293747, + 0.003971781115978956, + 0.01222138199955225, + 0.004982247482985258, + 0.012503267265856266, + 0.007709184195846319, + 0.006139747332781553, + 0.011888769455254078, + 0.006940977182239294, + 0.0026093425694853067, + 0.002821338828653097, + 0.0017839678330346942, + 0.0035515136551111937, + 0.0030670478008687496, + 0.002460450166836381, + 0.005902766715735197, + 0.0032084393315017223, + 0.008609643206000328, + 0.00689343549311161, + 0.004391620866954327, + 0.005731412675231695, + 0.00760742137208581, + 0.004845118150115013, + 0.006944021675735712, + 0.0047754994593560696, + 0.014792203903198242, + 0.01562558487057686, + 0.014087822288274765, + 0.017826426774263382, + 0.01890639215707779, + 0.009302536956965923, + 0.01595139130949974, + 0.013782226480543613, + 0.005379056092351675, + 0.028274275362491608, + 0.003909468650817871, + 0.028203830122947693, + 0.008925755508244038, + 0.006282370071858168, + 0.010534625500440598, + 0.004635987803339958, + 0.003221145598217845, + 0.003898380557075143, + 0.0033377439249306917, + 0.003495266428217292, + 0.004862132016569376, + 0.002828815020620823, + 0.004768261220306158, + 0.0028420283924788237, + 0.002438376657664776, + 0.0034878544975072145, + 0.002460065996274352, + 0.003022831864655018, + 0.002970355562865734, + 0.0020665423944592476, + 0.003038451075553894, + 0.0035633384250104427, + 0.014076783321797848, + 0.04888852685689926, + 0.012976429425179958, + 0.130581796169281, + 0.022083790972828865, + 0.014967980794608593, + 0.08077701926231384, + 0.035945456475019455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.5761569142341614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the patients wearing loose, light-blue, short-sleeved hospital gowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the patients have damp, matted hair plastered to their pale foreheads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the patients shown struggling to sit up with trembling arms and wrinkled white sheets sliding down their chests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool, flat overhead fluorescent lamps casting sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot with a slightly out-of-focus background containing a metal IV rack?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038098447024822235 + ], + "opposite_logprob_tokens": [ + -3.2880983352661133 + ], + "logprob_reward": -0.038098447024822235, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038098447024822235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06071983277797699, + 0.03918115049600601, + 0.028136447072029114, + 0.04524689167737961, + 0.07500030100345612, + 0.37111160159111023, + 0.03207285329699516, + 0.04100044071674347, + 0.015613030642271042, + 0.29191747307777405 + ], + "credit_map": [ + 0.006552374921739101, + 0.006796861533075571, + 0.004369494505226612, + 0.018838420510292053, + 0.0032702158205211163, + 0.005488054361194372, + 0.010045734234154224, + 0.005358676891773939, + 0.003450657706707716, + 0.006642903666943312, + 0.003962111659348011, + 0.00812417920678854, + 0.004036606755107641, + 0.0030291706789284945, + 0.005460692569613457, + 0.00447482755407691, + 0.0018994244746863842, + 0.002969014924019575, + 0.001737773884087801, + 0.0073264967650175095, + 0.002103410428389907, + 0.0020680788438767195, + 0.0032176217064261436, + 0.006814626511186361, + 0.004823158495128155, + 0.005129607859998941, + 0.002359537174925208, + 0.013305632397532463, + 0.004239583387970924, + 0.002660703146830201, + 0.008409960195422173, + 0.00431870948523283, + 0.006742763798683882, + 0.009160223416984081, + 0.004643237683922052, + 0.016434235498309135, + 0.00982961431145668, + 0.008451638743281364, + 0.011854761280119419, + 0.007883823476731777, + 0.07050708681344986, + 0.04280481114983559, + 0.042450547218322754, + 0.04293575510382652, + 0.03362056612968445, + 0.05424383655190468, + 0.02969483472406864, + 0.05485416576266289, + 0.0017348742112517357, + 0.008378950878977776, + 0.0014425760600715876, + 0.010053100064396858, + 0.0027722518425434828, + 0.0015368927270174026, + 0.004610829055309296, + 0.001543377642519772, + 0.0038899776991456747, + 0.007581609766930342, + 0.003115567844361067, + 0.0063275923021137714, + 0.004553388338536024, + 0.0038228139746934175, + 0.00603587506338954, + 0.005673616658896208, + 0.0022804990876466036, + 0.0017671092646196485, + 0.0022373110987246037, + 0.0022456622682511806, + 0.0014981863787397742, + 0.0017513412749394774, + 0.0018779551610350609, + 0.0019549657590687275, + 0.009021682664752007, + 0.031748950481414795, + 0.007426254916936159, + 0.11467576771974564, + 0.014613780193030834, + 0.007434248458594084, + 0.08896897733211517, + 0.018027812242507935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.038098447024822235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_10/index_tensor([6785, 6785])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_10/index_tensor([6785, 6785])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..01875f212238fe90d6f6b6c0237ed21ee7ddc211 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_10/index_tensor([6785, 6785])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a display of documentary-style realism, a steady, eye-level medium shot captures a giant Japanese spider crab, Macrocheira kaempferi, illuminated by the powerful, focused beam of a submersible's light in the abyssal darkness of the deep seabed. The creature crouches low, its knobby, calcified carapace, patterned in pale orange and grayish-white speckles, gleaming in the cold, white light. Its incredibly spindly legs, each articulated like a complex mechanism and bent at sharp angles, keep its armored body balanced, motionless, on a bed of fine-powdered sediment. The surrounding water is a pitch-black void, the only movement the gentle, constant drift of marine snow, tiny grains of organic detritus catching the light like a slow-motion blizzard. The cinematic quality of the image emphasizes the creature's immense size and alien beauty, contrasting the sharp, detailed textures of its body with the soft, out-of-focus drift of particles in the vast, dark, silent water, creating an atmosphere of profound isolation and awe-inspiring scientific discovery.", + "seed": 26598, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_10/index_6785_seed_26598_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02322609964758158, + "vlm_logprob": -0.02322609964758158, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.02322609964758158, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026511870324611664 + ], + "opposite_logprob_tokens": [ + -3.7765119075775146 + ], + "logprob_reward": -0.026511870324611664, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026511870324611664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1496191918849945, + 0.06664524227380753, + 0.03288411349058151, + 0.22366869449615479, + 0.1916641891002655, + 0.06510908156633377, + 0.09366316348314285, + 0.07384509593248367, + 0.06133084371685982, + 0.04157035052776337 + ], + "credit_map": [ + 0.021962810307741165, + 0.017018521204590797, + 0.015357592143118382, + 0.012909211218357086, + 0.0244451891630888, + 0.018570128828287125, + 0.017788046970963478, + 0.021567685529589653, + 0.0074142348021268845, + 0.008793249726295471, + 0.0074069444090127945, + 0.007850462570786476, + 0.009211641736328602, + 0.008884159848093987, + 0.010015416890382767, + 0.007069132756441832, + 0.005045355763286352, + 0.0035979533568024635, + 0.002863385947421193, + 0.003620118834078312, + 0.003194078104570508, + 0.004618180450052023, + 0.0033280595671385527, + 0.006616982165724039, + 0.028442738577723503, + 0.021383874118328094, + 0.021113725379109383, + 0.03784424811601639, + 0.026808617636561394, + 0.02821853756904602, + 0.04024125635623932, + 0.019615722820162773, + 0.016858482733368874, + 0.018514441326260567, + 0.012644055299460888, + 0.012174639850854874, + 0.044890325516462326, + 0.02885582111775875, + 0.025032516568899155, + 0.03269392251968384, + 0.008000684902071953, + 0.0044722966849803925, + 0.005294451490044594, + 0.007139229215681553, + 0.005158041138201952, + 0.01008556503802538, + 0.009447033517062664, + 0.015511778183281422, + 0.012839059345424175, + 0.007859477773308754, + 0.00634772377088666, + 0.009625835344195366, + 0.01639343798160553, + 0.013048939406871796, + 0.008532116189599037, + 0.019016575068235397, + 0.009812521748244762, + 0.004574521444737911, + 0.005404917988926172, + 0.006816117092967033, + 0.005358783062547445, + 0.012531193904578686, + 0.006410065107047558, + 0.022936983034014702, + 0.007051389664411545, + 0.005398564040660858, + 0.004190370906144381, + 0.00717132305726409, + 0.0074477167800068855, + 0.009176567196846008, + 0.005646936595439911, + 0.015247983857989311, + 0.0095625389367342, + 0.003525035222992301, + 0.0030908179469406605, + 0.0032966488506644964, + 0.004735717084258795, + 0.006044146604835987, + 0.003919218201190233, + 0.007396229542791843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.026511870324611664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00766183715313673 + ], + "opposite_logprob_tokens": [ + -4.882661819458008 + ], + "logprob_reward": -0.00766183715313673, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00766183715313673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1257614940404892, + 0.21881960332393646, + 0.11742490530014038, + 0.07040655612945557, + 0.04086657613515854, + 0.021233266219496727, + 0.019239991903305054, + 0.08277773857116699, + 0.14959405362606049, + 0.1538759022951126 + ], + "credit_map": [ + 0.01593044586479664, + 0.010724596679210663, + 0.015401391312479973, + 0.014946953393518925, + 0.019300179556012154, + 0.018267543986439705, + 0.01508499588817358, + 0.016105366870760918, + 0.031992606818675995, + 0.015059743076562881, + 0.016172710806131363, + 0.019669566303491592, + 0.018537823110818863, + 0.03634839504957199, + 0.026727914810180664, + 0.05431082844734192, + 0.01623431034386158, + 0.009949115104973316, + 0.011539271101355553, + 0.011305410414934158, + 0.011721848510205746, + 0.017505887895822525, + 0.01244622003287077, + 0.026722831651568413, + 0.008245354518294334, + 0.006256371270865202, + 0.006813977379351854, + 0.011342816986143589, + 0.0042136176489293575, + 0.009619680233299732, + 0.014789492823183537, + 0.0091252401471138, + 0.005076435394585133, + 0.0035058401990681887, + 0.003325249534100294, + 0.003242997219786048, + 0.008521058596670628, + 0.006741205230355263, + 0.004922688007354736, + 0.0055310954339802265, + 0.002761157928034663, + 0.002071759896352887, + 0.0024869285989552736, + 0.0020236081909388304, + 0.0027569718658924103, + 0.003445571754127741, + 0.003018184332177043, + 0.002669081324711442, + 0.0023073258344084024, + 0.0017371888970956206, + 0.0018043277086690068, + 0.0024867290630936623, + 0.0025315554812550545, + 0.002293097786605358, + 0.0031707072630524635, + 0.0029090591706335545, + 0.011542365886271, + 0.008011638186872005, + 0.007118944078683853, + 0.007962852716445923, + 0.00815659575164318, + 0.012207754887640476, + 0.009365173056721687, + 0.018412401899695396, + 0.021012211218476295, + 0.00987918023020029, + 0.009411854669451714, + 0.01808212511241436, + 0.013812610879540443, + 0.022966155782341957, + 0.015181582421064377, + 0.03924831748008728, + 0.01653885655105114, + 0.01236393116414547, + 0.015587205067276955, + 0.012259072624146938, + 0.0155800711363554, + 0.02592548169195652, + 0.012335595674812794, + 0.04328565672039986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.00766183715313673 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292265832424164 + ], + "opposite_logprob_tokens": [ + -3.773292303085327 + ], + "logprob_reward": -0.023292265832424164, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292265832424164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15835720300674438, + 0.2580380141735077, + 0.09121724218130112, + 0.09942195564508438, + 0.08515947312116623, + 0.0569387823343277, + 0.04840104654431343, + 0.10478740185499191, + 0.03237947076559067, + 0.06529948115348816 + ], + "credit_map": [ + 0.02294558845460415, + 0.021696899086236954, + 0.01735164225101471, + 0.018655814230442047, + 0.020593546330928802, + 0.021463783457875252, + 0.01805110275745392, + 0.01759883016347885, + 0.0394442155957222, + 0.020253077149391174, + 0.017284050583839417, + 0.024612654000520706, + 0.031213274225592613, + 0.04083065688610077, + 0.025766080245375633, + 0.05863397568464279, + 0.008696887642145157, + 0.011078539304435253, + 0.009287510998547077, + 0.0176069475710392, + 0.008235340937972069, + 0.012034566141664982, + 0.010448474436998367, + 0.013828977942466736, + 0.011414985172450542, + 0.012015331536531448, + 0.008849721401929855, + 0.01692967303097248, + 0.008804157376289368, + 0.01247574295848608, + 0.018179358914494514, + 0.010752984322607517, + 0.008264190517365932, + 0.010328397154808044, + 0.009662122465670109, + 0.012223961763083935, + 0.006968803238123655, + 0.011072631925344467, + 0.01623883843421936, + 0.010400522500276566, + 0.008364713750779629, + 0.0053568496368825436, + 0.008442450314760208, + 0.007432673126459122, + 0.004025861155241728, + 0.007815725170075893, + 0.0076256440952420235, + 0.007874865084886551, + 0.010369407944381237, + 0.004723698832094669, + 0.00414696242660284, + 0.004609376657754183, + 0.004928236361593008, + 0.006612528581172228, + 0.0058387876488268375, + 0.007172046694904566, + 0.01900302618741989, + 0.010286570526659489, + 0.008924701251089573, + 0.007792979944497347, + 0.011009786278009415, + 0.015607504174113274, + 0.008238262496888638, + 0.0239245742559433, + 0.00484228041023016, + 0.003695949912071228, + 0.0029783034697175026, + 0.003511721733957529, + 0.004022300709038973, + 0.004922881256788969, + 0.002948520239442587, + 0.005457514431327581, + 0.011542193591594696, + 0.006165549159049988, + 0.006129289977252483, + 0.005259084515273571, + 0.008192704059183598, + 0.009427406825125217, + 0.004670583177357912, + 0.013912669382989407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.023292265832424164 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808513656258583 + ], + "opposite_logprob_tokens": [ + -3.2880852222442627 + ], + "logprob_reward": -0.03808513656258583, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808513656258583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10245631635189056, + 0.1885128617286682, + 0.12615090608596802, + 0.1007019653916359, + 0.09406053274869919, + 0.06057615205645561, + 0.06395120918750763, + 0.11698902398347855, + 0.061288364231586456, + 0.08531267195940018 + ], + "credit_map": [ + 0.011705034412443638, + 0.016442064195871353, + 0.011721327900886536, + 0.013060574419796467, + 0.013128206133842468, + 0.012997527606785297, + 0.013081797398626804, + 0.010319778695702553, + 0.02634396031498909, + 0.012364990077912807, + 0.014605891890823841, + 0.014810232445597649, + 0.015491686761379242, + 0.03240127116441727, + 0.022898737341165543, + 0.04959610477089882, + 0.0144843440502882, + 0.01177047099918127, + 0.013260547071695328, + 0.017662392929196358, + 0.010413658805191517, + 0.018429573625326157, + 0.01690998300909996, + 0.023219941183924675, + 0.012343408539891243, + 0.010320670902729034, + 0.007735344581305981, + 0.006641666404902935, + 0.021686747670173645, + 0.016444874927401543, + 0.010640830732882023, + 0.014888420701026917, + 0.0075038522481918335, + 0.008397758938372135, + 0.005018885713070631, + 0.008176456205546856, + 0.01936383545398712, + 0.014796404168009758, + 0.00889796856790781, + 0.021905377507209778, + 0.009085781872272491, + 0.005777440033853054, + 0.0041939313523471355, + 0.007161593064665794, + 0.007096681743860245, + 0.00856320746243, + 0.005746322218328714, + 0.012951196171343327, + 0.009034284390509129, + 0.0047922697849571705, + 0.005221424624323845, + 0.005066174082458019, + 0.010925446636974812, + 0.009841726161539555, + 0.006536286789923906, + 0.012533596716821194, + 0.017817657440900803, + 0.01041733380407095, + 0.009492640383541584, + 0.010640357621014118, + 0.016278544440865517, + 0.016829000785946846, + 0.011748737655580044, + 0.023764753714203835, + 0.012529246509075165, + 0.00511879613623023, + 0.005741470959037542, + 0.005798028316348791, + 0.007309374865144491, + 0.00860824715346098, + 0.006231893785297871, + 0.009951304644346237, + 0.02330024540424347, + 0.007882153615355492, + 0.007216149941086769, + 0.00778483971953392, + 0.006661777384579182, + 0.012554137036204338, + 0.007472741883248091, + 0.0124406348913908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.03808513656258583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579388365149498 + ], + "opposite_logprob_tokens": [ + -3.8955793380737305 + ], + "logprob_reward": -0.020579388365149498, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579388365149498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11163034290075302, + 0.2753393352031708, + 0.31393682956695557, + 0.06207415834069252, + 0.027158057317137718, + 0.022397687658667564, + 0.027468813583254814, + 0.041761692613363266, + 0.055664677172899246, + 0.06256838887929916 + ], + "credit_map": [ + 0.0149347810074687, + 0.009611809626221657, + 0.018762733787298203, + 0.01309122983366251, + 0.013768537901341915, + 0.013735407032072544, + 0.013844257220625877, + 0.013881583698093891, + 0.02944827638566494, + 0.02918262779712677, + 0.01828175224363804, + 0.028504492715001106, + 0.03206716850399971, + 0.03716752305626869, + 0.04658111557364464, + 0.05410637706518173, + 0.043293893337249756, + 0.027880363166332245, + 0.032087624073028564, + 0.03134646266698837, + 0.03781238570809364, + 0.041624270379543304, + 0.033736009150743484, + 0.0661557987332344, + 0.0070475987158715725, + 0.007568785920739174, + 0.0042269267141819, + 0.007293016649782658, + 0.012947204522788525, + 0.008811995387077332, + 0.0070690736174583435, + 0.007109551224857569, + 0.0034345118328928947, + 0.0024400041438639164, + 0.002980103250592947, + 0.002671367023140192, + 0.0022870670072734356, + 0.003877842565998435, + 0.0036105522885918617, + 0.005856606177985668, + 0.0023047944996505976, + 0.0020028159487992525, + 0.0019330700160935521, + 0.002193302381783724, + 0.002345592947676778, + 0.003769905073568225, + 0.0024483492597937584, + 0.005399854388087988, + 0.004808465950191021, + 0.002599111059680581, + 0.002653165953233838, + 0.0023411870934069157, + 0.002146483166143298, + 0.0040419697761535645, + 0.003008598228916526, + 0.005869831889867783, + 0.005354392807930708, + 0.003402267349883914, + 0.0034850663505494595, + 0.0039489176124334335, + 0.0038456094916909933, + 0.006462789606302977, + 0.00401658471673727, + 0.011246060952544212, + 0.006508192513138056, + 0.005888607818633318, + 0.006159279961138964, + 0.005886073689907789, + 0.006373973097652197, + 0.00832454115152359, + 0.00628018518909812, + 0.010243820957839489, + 0.008430512621998787, + 0.0067421384155750275, + 0.005608201026916504, + 0.006040367763489485, + 0.009800155647099018, + 0.008805674500763416, + 0.005862799473106861, + 0.011278534308075905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.020579388365149498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a display of documentary-style realism, a steady, eye-level medium shot captures a giant Japanese spider crab, Macrocheira kaempferi, illuminated by the powerful, focused beam of a submersible's light in the abyssal darkness of the deep seabed. The creature crouches low, its knobby, calcified carapace, patterned in pale orange and grayish-white speckles, gleaming in the cold, white light. Its incredibly spindly legs, each articulated like a complex mechanism and bent at sharp angles, keep its armored body balanced, motionless, on a bed of fine-powdered sediment. The surrounding water is a pitch-black void, the only movement the gentle, constant drift of marine snow, tiny grains of organic detritus catching the light like a slow-motion blizzard. The cinematic quality of the image emphasizes the creature's immense size and alien beauty, contrasting the sharp, detailed textures of its body with the soft, out-of-focus drift of particles in the vast, dark, silent water, creating an atmosphere of profound isolation and awe-inspiring scientific discovery.", + "seed": 26599, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_10/index_6785_seed_26599_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07369041256606579, + "vlm_logprob": -0.07369041256606579, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.07369041256606579, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043336376547813416 + ], + "opposite_logprob_tokens": [ + -3.2933363914489746 + ], + "logprob_reward": -0.043336376547813416, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.043336376547813416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.125821053981781, + 0.050394207239151, + 0.18600477278232574, + 0.10117709636688232, + 0.07035093009471893, + 0.09047440439462662, + 0.08330941200256348, + 0.13757748901844025, + 0.11534566432237625, + 0.039545051753520966 + ], + "credit_map": [ + 0.014564125798642635, + 0.014875288121402264, + 0.018133925274014473, + 0.015910636633634567, + 0.015719326213002205, + 0.013023857958614826, + 0.01428957935422659, + 0.019304320216178894, + 0.005172377452254295, + 0.005854467395693064, + 0.009193316102027893, + 0.007769153453409672, + 0.005640885792672634, + 0.005561846774071455, + 0.005849520210176706, + 0.005352638196200132, + 0.022942431271076202, + 0.020617270842194557, + 0.028548801317811012, + 0.029580408707261086, + 0.01998184062540531, + 0.021624403074383736, + 0.0198377538472414, + 0.022871853783726692, + 0.009848825633525848, + 0.01302179042249918, + 0.014934505335986614, + 0.015065626241266727, + 0.011775139719247818, + 0.011362777091562748, + 0.010980156250298023, + 0.014188271947205067, + 0.006608189549297094, + 0.012389265932142735, + 0.00800891313701868, + 0.011646413244307041, + 0.009513038210570812, + 0.007551697548478842, + 0.007829603739082813, + 0.0068038031458854675, + 0.010384433902800083, + 0.021039115265011787, + 0.009397879242897034, + 0.00810459814965725, + 0.014382261782884598, + 0.009852227754890919, + 0.009484914131462574, + 0.007828973233699799, + 0.009670326486229897, + 0.014131959527730942, + 0.012537133879959583, + 0.010351255536079407, + 0.010334634222090244, + 0.008718007244169712, + 0.008170010522007942, + 0.00939608458429575, + 0.015163535252213478, + 0.015853064134716988, + 0.021448899060487747, + 0.01376180537045002, + 0.01561582088470459, + 0.01576574333012104, + 0.021177353337407112, + 0.01879126764833927, + 0.013487077318131924, + 0.020712921395897865, + 0.016816431656479836, + 0.012953328900039196, + 0.008861782029271126, + 0.01150794792920351, + 0.01612464338541031, + 0.014881537295877934, + 0.005438830703496933, + 0.008940624073147774, + 0.005034899804741144, + 0.003970813937485218, + 0.004437914118170738, + 0.004197038244456053, + 0.003045502584427595, + 0.004479428753256798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.043336376547813416 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208503246307373 + ], + "opposite_logprob_tokens": [ + -2.8120851516723633 + ], + "logprob_reward": -0.06208503246307373, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208503246307373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1042223572731018, + 0.10468882322311401, + 0.0767795741558075, + 0.09983768314123154, + 0.1190280169248581, + 0.22381319105625153, + 0.18592269718647003, + 0.039292801171541214, + 0.019479336217045784, + 0.026935501024127007 + ], + "credit_map": [ + 0.012712734751403332, + 0.011441635899245739, + 0.018587840721011162, + 0.015064612030982971, + 0.013796026818454266, + 0.009646407328546047, + 0.01235360000282526, + 0.010619504377245903, + 0.014139440841972828, + 0.014356476254761219, + 0.017647067084908485, + 0.013684618286788464, + 0.011783012188971043, + 0.012550932355225086, + 0.00989551655948162, + 0.01063176617026329, + 0.006811683066189289, + 0.017718132585287094, + 0.00781311560422182, + 0.011202888563275337, + 0.009171821177005768, + 0.007715804036706686, + 0.009147468954324722, + 0.007198663428425789, + 0.015716593712568283, + 0.007913528941571712, + 0.015188533812761307, + 0.009150177240371704, + 0.01071364339441061, + 0.012630991637706757, + 0.010240929201245308, + 0.018283283337950706, + 0.018185537308454514, + 0.012071232311427593, + 0.02117493376135826, + 0.013302482664585114, + 0.012309608981013298, + 0.013589026406407356, + 0.010494782589375973, + 0.017900407314300537, + 0.025058535858988762, + 0.026357803493738174, + 0.03638801723718643, + 0.029313165694475174, + 0.03303350880742073, + 0.025252005085349083, + 0.024108055979013443, + 0.024302087724208832, + 0.025496968999505043, + 0.019462723284959793, + 0.030189454555511475, + 0.02077602967619896, + 0.025561880320310593, + 0.024069085717201233, + 0.01885128580033779, + 0.021515268832445145, + 0.0043335966765880585, + 0.005896692629903555, + 0.005408488679677248, + 0.003865453414618969, + 0.00435788556933403, + 0.004397967364639044, + 0.005992759484797716, + 0.005039958283305168, + 0.002364545362070203, + 0.0030016335658729076, + 0.002371720504015684, + 0.002229064702987671, + 0.002660505473613739, + 0.0020424353424459696, + 0.0022838327568024397, + 0.002525599440559745, + 0.0029832518193870783, + 0.006602731999009848, + 0.002240103902295232, + 0.0021000003907829523, + 0.0034408466890454292, + 0.003301553661003709, + 0.0029185630846768618, + 0.0033484494779258966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.06208503246307373 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294946178793907 + ], + "opposite_logprob_tokens": [ + -3.7732949256896973 + ], + "logprob_reward": -0.023294946178793907, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294946178793907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12179520726203918, + 0.08498447388410568, + 0.07250701636075974, + 0.10136920213699341, + 0.06679891794919968, + 0.11217299848794937, + 0.09591301530599594, + 0.2500349283218384, + 0.03997209668159485, + 0.054452210664749146 + ], + "credit_map": [ + 0.011678270995616913, + 0.016688114032149315, + 0.018290534615516663, + 0.014587651006877422, + 0.018463237211108208, + 0.013865834102034569, + 0.014477896504104137, + 0.013743662275373936, + 0.008698411285877228, + 0.0076722935773432255, + 0.017259597778320312, + 0.011827448382973671, + 0.009495772421360016, + 0.011930050328373909, + 0.009200453758239746, + 0.008900444023311138, + 0.00784994475543499, + 0.012790563516318798, + 0.009815589524805546, + 0.0067716133780777454, + 0.009524110704660416, + 0.009080782532691956, + 0.007404145784676075, + 0.009270268492400646, + 0.01246856153011322, + 0.012810141779482365, + 0.015338052995502949, + 0.01248119119554758, + 0.013430900871753693, + 0.01116665918380022, + 0.01033769641071558, + 0.013336004689335823, + 0.007406862918287516, + 0.007672056555747986, + 0.009908173233270645, + 0.010470753535628319, + 0.00877908430993557, + 0.007165329996496439, + 0.006449379492551088, + 0.008947272785007954, + 0.011726128868758678, + 0.016007443889975548, + 0.016255978494882584, + 0.013405769132077694, + 0.018282024189829826, + 0.011056809686124325, + 0.012431503273546696, + 0.0130073307082057, + 0.010831237770617008, + 0.01422936748713255, + 0.014824414625763893, + 0.011143915355205536, + 0.014431009069085121, + 0.009293611161410809, + 0.010010889731347561, + 0.011148571036756039, + 0.023609967902302742, + 0.026152802631258965, + 0.031902894377708435, + 0.04988010227680206, + 0.0203282181173563, + 0.022385306656360626, + 0.03709465637803078, + 0.03868095949292183, + 0.005228426307439804, + 0.0033415150828659534, + 0.005683005787432194, + 0.004779990296810865, + 0.004483692813664675, + 0.00451125530526042, + 0.005435839761048555, + 0.006508370861411095, + 0.005127859767526388, + 0.012189103290438652, + 0.005517063662409782, + 0.005447424948215485, + 0.006766646169126034, + 0.005380011163651943, + 0.00613362668082118, + 0.007890471257269382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.023294946178793907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255399376153946 + ], + "opposite_logprob_tokens": [ + -1.6005399227142334 + ], + "logprob_reward": -0.2255399376153946, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255399376153946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16593454778194427, + 0.2092108130455017, + 0.1122984066605568, + 0.05305250361561775, + 0.07043935358524323, + 0.0932932123541832, + 0.09971416741609573, + 0.09011490643024445, + 0.06004082411527634, + 0.04590126499533653 + ], + "credit_map": [ + 0.016305893659591675, + 0.0186222605407238, + 0.030449310317635536, + 0.019385091960430145, + 0.020309463143348694, + 0.020371047779917717, + 0.01721367798745632, + 0.02327781915664673, + 0.026327159255743027, + 0.022183071821928024, + 0.03690481185913086, + 0.034883614629507065, + 0.020651279017329216, + 0.022807907313108444, + 0.02324586547911167, + 0.02220708690583706, + 0.012813064269721508, + 0.015412257052958012, + 0.014905653893947601, + 0.01920742355287075, + 0.012806501239538193, + 0.01109169889241457, + 0.01494637131690979, + 0.011115431785583496, + 0.009167547337710857, + 0.005602716468274593, + 0.006906033959239721, + 0.004707789979875088, + 0.005723922047764063, + 0.006574870087206364, + 0.005516704637557268, + 0.008852919563651085, + 0.007935941219329834, + 0.017007261514663696, + 0.008107911795377731, + 0.006539315450936556, + 0.011876698583364487, + 0.008042746223509312, + 0.005186799913644791, + 0.005742673762142658, + 0.008874486200511456, + 0.013952359557151794, + 0.013302365317940712, + 0.013424312695860863, + 0.011039755307137966, + 0.009775137528777122, + 0.010614666156470776, + 0.012310128659009933, + 0.01069138664752245, + 0.009264197200536728, + 0.014276090078055859, + 0.013737264089286327, + 0.014314545318484306, + 0.011317584663629532, + 0.011039849370718002, + 0.015073256567120552, + 0.009676214307546616, + 0.01905115880072117, + 0.00843333825469017, + 0.010849795304238796, + 0.009027653373777866, + 0.009532678872346878, + 0.012688782997429371, + 0.010855281725525856, + 0.005920725874602795, + 0.008078223094344139, + 0.00629347562789917, + 0.006621646229177713, + 0.008492431603372097, + 0.006117305718362331, + 0.009296764619648457, + 0.00922025553882122, + 0.006469504442065954, + 0.004951813258230686, + 0.00621400261297822, + 0.007824606262147427, + 0.004973893985152245, + 0.00486851716414094, + 0.004961615893989801, + 0.005637311842292547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.2255399376153946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195770025253296 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195770025253296, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195770025253296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1076200008392334, + 0.11544928699731827, + 0.11043830960988998, + 0.15708647668361664, + 0.072708860039711, + 0.057875316590070724, + 0.0864548534154892, + 0.1736556440591812, + 0.08665098249912262, + 0.03206025809049606 + ], + "credit_map": [ + 0.010862687602639198, + 0.01144412811845541, + 0.021790774539113045, + 0.01722104661166668, + 0.008725411258637905, + 0.009861193597316742, + 0.013341717422008514, + 0.014373043552041054, + 0.011624668724834919, + 0.012024212628602982, + 0.020189542323350906, + 0.024876514449715614, + 0.010296104475855827, + 0.010877094231545925, + 0.013912312686443329, + 0.011648832820355892, + 0.017427779734134674, + 0.01297000888735056, + 0.01531397458165884, + 0.013570544309914112, + 0.01284553948789835, + 0.01205708459019661, + 0.014081536792218685, + 0.012171842157840729, + 0.018547849729657173, + 0.02946580946445465, + 0.020473960787057877, + 0.023844458162784576, + 0.015092053450644016, + 0.01779724843800068, + 0.01562210451811552, + 0.016242988407611847, + 0.00752448383718729, + 0.011420642025768757, + 0.009175790473818779, + 0.00972084328532219, + 0.0075622121803462505, + 0.009407349862158298, + 0.007361858617514372, + 0.010535676963627338, + 0.006481787655502558, + 0.008655750192701817, + 0.005884387996047735, + 0.006696678232401609, + 0.008892551995813847, + 0.0059922258369624615, + 0.007514779921621084, + 0.007757153362035751, + 0.009704818949103355, + 0.017791232094168663, + 0.012935762293636799, + 0.008181793615221977, + 0.011568408459424973, + 0.008028656244277954, + 0.008866813965141773, + 0.009377365000545979, + 0.018022075295448303, + 0.023930322378873825, + 0.02181001380085945, + 0.026271788403391838, + 0.016202609986066818, + 0.0173875130712986, + 0.024834243580698967, + 0.025197098031640053, + 0.009411085397005081, + 0.01919112168252468, + 0.009970228187739849, + 0.008682913146913052, + 0.010321816429495811, + 0.00842384435236454, + 0.011281570419669151, + 0.009368401020765305, + 0.004700254183262587, + 0.003894357942044735, + 0.005031890701502562, + 0.002871021395549178, + 0.0031021381728351116, + 0.005058723967522383, + 0.003605652367696166, + 0.003796219127252698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.014195770025253296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a display of documentary-style realism, a steady, eye-level medium shot captures a giant Japanese spider crab, Macrocheira kaempferi, illuminated by the powerful, focused beam of a submersible's light in the abyssal darkness of the deep seabed. The creature crouches low, its knobby, calcified carapace, patterned in pale orange and grayish-white speckles, gleaming in the cold, white light. Its incredibly spindly legs, each articulated like a complex mechanism and bent at sharp angles, keep its armored body balanced, motionless, on a bed of fine-powdered sediment. The surrounding water is a pitch-black void, the only movement the gentle, constant drift of marine snow, tiny grains of organic detritus catching the light like a slow-motion blizzard. The cinematic quality of the image emphasizes the creature's immense size and alien beauty, contrasting the sharp, detailed textures of its body with the soft, out-of-focus drift of particles in the vast, dark, silent water, creating an atmosphere of profound isolation and awe-inspiring scientific discovery.", + "seed": 26600, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_10/index_6785_seed_26600_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09437008984386921, + "vlm_logprob": -0.09437008984386921, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.09437008984386921, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0471578910946846 + ], + "opposite_logprob_tokens": [ + -3.1721580028533936 + ], + "logprob_reward": -0.0471578910946846, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0471578910946846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22545500099658966, + 0.07168059051036835, + 0.0555998720228672, + 0.19732140004634857, + 0.05610780790448189, + 0.03925739973783493, + 0.06501854956150055, + 0.13128332793712616, + 0.10937770456075668, + 0.048898402601480484 + ], + "credit_map": [ + 0.04910169914364815, + 0.014512617141008377, + 0.02239144779741764, + 0.020661186426877975, + 0.03581113740801811, + 0.03355829790234566, + 0.01939372718334198, + 0.03002488799393177, + 0.021684734150767326, + 0.004525137133896351, + 0.006122353486716747, + 0.005053992848843336, + 0.008446100167930126, + 0.01313020195811987, + 0.005037311930209398, + 0.007680761627852917, + 0.01007872261106968, + 0.0042607891373336315, + 0.005901625379920006, + 0.004824645817279816, + 0.0061503523029387, + 0.00865765567868948, + 0.005598605610430241, + 0.010127476416528225, + 0.04707576334476471, + 0.0139394486322999, + 0.01703079603612423, + 0.014462155289947987, + 0.016669986769557, + 0.033796753734350204, + 0.013896130956709385, + 0.04045036435127258, + 0.008294183760881424, + 0.0033766410779207945, + 0.005456112325191498, + 0.004185082390904427, + 0.0048996382392942905, + 0.00913085788488388, + 0.004307442810386419, + 0.016457850113511086, + 0.010980542749166489, + 0.003155733924359083, + 0.00286269374191761, + 0.003786413697525859, + 0.0038511601742357016, + 0.006669714115560055, + 0.0028760896530002356, + 0.0050750537775456905, + 0.01617196388542652, + 0.0027764032129198313, + 0.0051748910918831825, + 0.004708394408226013, + 0.007049406412988901, + 0.010101443156599998, + 0.004283618181943893, + 0.014752427116036415, + 0.02823180891573429, + 0.008246546611189842, + 0.010503825731575489, + 0.007515596691519022, + 0.012736582197248936, + 0.021318908780813217, + 0.008434277027845383, + 0.03429577872157097, + 0.021300259977579117, + 0.008382739499211311, + 0.009912088513374329, + 0.007393351290374994, + 0.010950527153909206, + 0.017568103969097137, + 0.009436686523258686, + 0.02443394996225834, + 0.011157477274537086, + 0.0023814397864043713, + 0.004554464481770992, + 0.0036176673602312803, + 0.0045106858015060425, + 0.00866780523210764, + 0.0035401349887251854, + 0.010468724183738232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0471578910946846 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012599628418684006 + ], + "opposite_logprob_tokens": [ + -4.387599468231201 + ], + "logprob_reward": -0.012599628418684006, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012599628418684006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2123321294784546, + 0.083387091755867, + 0.09037113934755325, + 0.1464671641588211, + 0.02148931659758091, + 0.01628904975950718, + 0.027913350611925125, + 0.16185905039310455, + 0.191258043050766, + 0.048633698374032974 + ], + "credit_map": [ + 0.05009636655449867, + 0.016890550032258034, + 0.018106535077095032, + 0.02057981863617897, + 0.02974345162510872, + 0.03553761541843414, + 0.018475787714123726, + 0.022902021184563637, + 0.021975606679916382, + 0.005860368721187115, + 0.007685701362788677, + 0.007988161407411098, + 0.009035706520080566, + 0.015598522499203682, + 0.007272936869412661, + 0.007970084436237812, + 0.02627895027399063, + 0.006411000154912472, + 0.008506225422024727, + 0.00664684409275651, + 0.010207897052168846, + 0.015037880279123783, + 0.0057497331872582436, + 0.011532609350979328, + 0.03732234239578247, + 0.010336523875594139, + 0.011409744620323181, + 0.010104690678417683, + 0.01184038259088993, + 0.02696182020008564, + 0.010124832391738892, + 0.028366830199956894, + 0.0025864499621093273, + 0.0017619789578020573, + 0.0020517383236438036, + 0.0026846283581107855, + 0.002917021280154586, + 0.0027403009589761496, + 0.002606593305245042, + 0.004140603821724653, + 0.003053144318982959, + 0.001660304144024849, + 0.001472124713473022, + 0.0015381327830255032, + 0.0016674719518050551, + 0.002232785802334547, + 0.001705282717011869, + 0.0029598018154501915, + 0.005722861271351576, + 0.001428069081157446, + 0.002819838933646679, + 0.002665699692443013, + 0.0035136686637997627, + 0.0042587751522660255, + 0.0029996605589985847, + 0.004504778888076544, + 0.03708887845277786, + 0.010541900992393494, + 0.012286785989999771, + 0.00885502714663744, + 0.016229834407567978, + 0.026800472289323807, + 0.010906603187322617, + 0.03914955258369446, + 0.04568251967430115, + 0.01331225037574768, + 0.016196757555007935, + 0.010801191441714764, + 0.016754722222685814, + 0.032442592084407806, + 0.011766741052269936, + 0.04430127888917923, + 0.011265941895544529, + 0.0032835279125720263, + 0.0047660935670137405, + 0.004366263281553984, + 0.006700801197439432, + 0.007802561856806278, + 0.004076299723237753, + 0.0063722082413733006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.012599628418684006 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801329597830772 + ], + "opposite_logprob_tokens": [ + -3.529801368713379 + ], + "logprob_reward": -0.029801329597830772, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801329597830772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16477258503437042, + 0.09838689863681793, + 0.19479110836982727, + 0.1548907607793808, + 0.10073074698448181, + 0.06298161298036575, + 0.028249582275748253, + 0.03590412437915802, + 0.13049615919589996, + 0.028796417638659477 + ], + "credit_map": [ + 0.042013414204120636, + 0.009716308675706387, + 0.01468067429959774, + 0.009439568035304546, + 0.020428480580449104, + 0.027442265301942825, + 0.013110584579408169, + 0.02794129215180874, + 0.019839785993099213, + 0.007689486723393202, + 0.008675881661474705, + 0.008678686805069447, + 0.00980455707758665, + 0.014984331093728542, + 0.00824010744690895, + 0.020474063232541084, + 0.046162642538547516, + 0.013787792064249516, + 0.01802900619804859, + 0.01564045064151287, + 0.021799316629767418, + 0.030371258035302162, + 0.012965263798832893, + 0.03603536635637283, + 0.04007336497306824, + 0.009777816943824291, + 0.013209770433604717, + 0.010326740331947803, + 0.011037468910217285, + 0.029452990740537643, + 0.009987934492528439, + 0.031024673953652382, + 0.014712723903357983, + 0.006634230725467205, + 0.011126706376671791, + 0.009340949356555939, + 0.008647867478430271, + 0.015229840762913227, + 0.007892371155321598, + 0.027146056294441223, + 0.01699705235660076, + 0.003832566784694791, + 0.005193002987653017, + 0.004580357111990452, + 0.004733980633318424, + 0.011370173655450344, + 0.004005007911473513, + 0.01226947084069252, + 0.006691460032016039, + 0.001956494292244315, + 0.0020155818201601505, + 0.0017333407886326313, + 0.0035635801032185555, + 0.004484177101403475, + 0.00212951865978539, + 0.005675430875271559, + 0.009392871521413326, + 0.00280567305162549, + 0.002594101708382368, + 0.0025916993618011475, + 0.0042361905798316, + 0.006325593683868647, + 0.0021774140186607838, + 0.0057805804535746574, + 0.028315257281064987, + 0.008110787719488144, + 0.011577374301850796, + 0.007130493875592947, + 0.010535391047596931, + 0.020970435813069344, + 0.00918215699493885, + 0.034674275666475296, + 0.00496244290843606, + 0.0021674633026123047, + 0.0029575263615697622, + 0.0029220443684607744, + 0.005160953383892775, + 0.004126702435314655, + 0.0025757011026144028, + 0.0039235856384038925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.029801329597830772 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34859776496887207 + ], + "opposite_logprob_tokens": [ + -1.223597764968872 + ], + "logprob_reward": -0.34859776496887207, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34859776496887207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06571896374225616, + 0.06545048207044601, + 0.02760111168026924, + 0.5252546072006226, + 0.09832226485013962, + 0.0733904018998146, + 0.0353381484746933, + 0.026253685355186462, + 0.05693504586815834, + 0.025735339149832726 + ], + "credit_map": [ + 0.011209536343812943, + 0.0047799767926335335, + 0.007518494967371225, + 0.008042052388191223, + 0.008641546592116356, + 0.008809572085738182, + 0.007305519189685583, + 0.009412265382707119, + 0.014351313933730125, + 0.0036650088150054216, + 0.004999184049665928, + 0.004341022111475468, + 0.006240230984985828, + 0.010694281198084354, + 0.004271115176379681, + 0.016888322308659554, + 0.008303945884108543, + 0.0018429725896567106, + 0.0021240320056676865, + 0.002335744444280863, + 0.0026443125680088997, + 0.0048228669911623, + 0.002226713579148054, + 0.0033005219884216785, + 0.13827215135097504, + 0.03853648528456688, + 0.04477082937955856, + 0.03380325064063072, + 0.04227215796709061, + 0.09609644114971161, + 0.0322590135037899, + 0.09924423694610596, + 0.0204976387321949, + 0.004969505593180656, + 0.012117798440158367, + 0.007893613539636135, + 0.008546457625925541, + 0.015600664541125298, + 0.007568138185888529, + 0.021128445863723755, + 0.01645529828965664, + 0.004570516292005777, + 0.007546409964561462, + 0.0059448061510920525, + 0.007298693526536226, + 0.010843394324183464, + 0.006344623398035765, + 0.014386650174856186, + 0.007920772768557072, + 0.002474664244800806, + 0.002768357750028372, + 0.0037788436748087406, + 0.004478016868233681, + 0.005790953058749437, + 0.002724113641306758, + 0.005402423907071352, + 0.005408324766904116, + 0.0015681703807786107, + 0.0026107095181941986, + 0.0016065646195784211, + 0.0025001473259180784, + 0.004087925422936678, + 0.0014894007472321391, + 0.006982442457228899, + 0.014039522968232632, + 0.004289929289370775, + 0.0046058944426476955, + 0.0035715457051992416, + 0.005525951273739338, + 0.01003525871783495, + 0.0036629873793572187, + 0.011203951202332973, + 0.0051019685342907906, + 0.0013009515823796391, + 0.0028438465669751167, + 0.0015445675235241652, + 0.0026185261085629463, + 0.00386625062674284, + 0.0017090640030801296, + 0.006750162225216627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.34859776496887207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336938351392746 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.0336938351392746, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336938351392746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14619432389736176, + 0.17332208156585693, + 0.18663908541202545, + 0.27961546182632446, + 0.018132532015442848, + 0.031252987682819366, + 0.10555265098810196, + 0.014599528163671494, + 0.021991975605487823, + 0.02269936166703701 + ], + "credit_map": [ + 0.027726391330361366, + 0.011691920459270477, + 0.01434720866382122, + 0.01623525097966194, + 0.023136630654335022, + 0.020230691879987717, + 0.014415394514799118, + 0.018410827964544296, + 0.04414226487278938, + 0.009904970414936543, + 0.013434105552732944, + 0.010054659098386765, + 0.014565667137503624, + 0.03117847815155983, + 0.01054017897695303, + 0.03950176015496254, + 0.04857507720589638, + 0.011508028022944927, + 0.016658926382660866, + 0.010159552097320557, + 0.019973840564489365, + 0.031807929277420044, + 0.012806549668312073, + 0.035149190574884415, + 0.07814475148916245, + 0.01624327339231968, + 0.021516408771276474, + 0.015019492246210575, + 0.020570753142237663, + 0.0544406957924366, + 0.016099926084280014, + 0.057580165565013885, + 0.0028919847682118416, + 0.0015720538794994354, + 0.0021750617306679487, + 0.0015666037797927856, + 0.002165812999010086, + 0.002771230647340417, + 0.0016798785654827952, + 0.0033099050633609295, + 0.007004921790212393, + 0.0016995397163555026, + 0.0033445742446929216, + 0.00231984187848866, + 0.003183758584782481, + 0.0046478742733597755, + 0.0023871667217463255, + 0.006665310822427273, + 0.02479800209403038, + 0.00631481409072876, + 0.0078091612085700035, + 0.006250102072954178, + 0.00857043918222189, + 0.017099035903811455, + 0.0068598841316998005, + 0.027851218357682228, + 0.002822695067152381, + 0.0010085696121677756, + 0.0016506167594343424, + 0.0011312148999422789, + 0.0022913585416972637, + 0.002301792846992612, + 0.0009072772227227688, + 0.002486004028469324, + 0.005493007134646177, + 0.0018926585325971246, + 0.001875175628811121, + 0.001782407402060926, + 0.002697369083762169, + 0.004046515561640263, + 0.001345124444924295, + 0.0028597174677997828, + 0.004742557182908058, + 0.0012254875618964434, + 0.002714230213314295, + 0.0019107538973912597, + 0.0027804281562566757, + 0.003539070952683687, + 0.0017397665651515126, + 0.00404706783592701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.0336938351392746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a display of documentary-style realism, a steady, eye-level medium shot captures a giant Japanese spider crab, Macrocheira kaempferi, illuminated by the powerful, focused beam of a submersible's light in the abyssal darkness of the deep seabed. The creature crouches low, its knobby, calcified carapace, patterned in pale orange and grayish-white speckles, gleaming in the cold, white light. Its incredibly spindly legs, each articulated like a complex mechanism and bent at sharp angles, keep its armored body balanced, motionless, on a bed of fine-powdered sediment. The surrounding water is a pitch-black void, the only movement the gentle, constant drift of marine snow, tiny grains of organic detritus catching the light like a slow-motion blizzard. The cinematic quality of the image emphasizes the creature's immense size and alien beauty, contrasting the sharp, detailed textures of its body with the soft, out-of-focus drift of particles in the vast, dark, silent water, creating an atmosphere of profound isolation and awe-inspiring scientific discovery.", + "seed": 26601, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_10/index_6785_seed_26601_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12060092948377132, + "vlm_logprob": -0.12060092948377132, + "vlm_yes_no_margin": 2.824999988079071, + "vlm_token_credit": -0.12060092948377132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022843893617391586 + ], + "opposite_logprob_tokens": [ + -4.022843837738037 + ], + "logprob_reward": -0.022843893617391586, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.022843893617391586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15142667293548584, + 0.24005696177482605, + 0.02029685117304325, + 0.05959143117070198, + 0.1331864595413208, + 0.20511043071746826, + 0.07886185497045517, + 0.03657706454396248, + 0.042536381632089615, + 0.0323559045791626 + ], + "credit_map": [ + 0.020122554153203964, + 0.015365495346486568, + 0.014869315549731255, + 0.023207154124975204, + 0.017401641234755516, + 0.023415420204401016, + 0.021549640223383904, + 0.015495432540774345, + 0.04441260173916817, + 0.02172345295548439, + 0.021574491634964943, + 0.021694108843803406, + 0.03747481852769852, + 0.03790133073925972, + 0.03199887275695801, + 0.02327726222574711, + 0.002415561582893133, + 0.002292237477377057, + 0.0025861666072160006, + 0.0028443229384720325, + 0.002149391220882535, + 0.002773544518277049, + 0.002835697727277875, + 0.002399926772341132, + 0.010443956591188908, + 0.006703472696244717, + 0.004787239246070385, + 0.0046617137268185616, + 0.012543123215436935, + 0.009756645187735558, + 0.0060478271916508675, + 0.004647446796298027, + 0.04121096804738045, + 0.020201105624437332, + 0.012882058508694172, + 0.008016987703740597, + 0.013975459150969982, + 0.023491835221648216, + 0.005989176221191883, + 0.007418858353048563, + 0.05949315428733826, + 0.026767082512378693, + 0.021711979061365128, + 0.012578028254210949, + 0.019665542989969254, + 0.03394851088523865, + 0.013260367326438427, + 0.017685741186141968, + 0.021461835131049156, + 0.0127016082406044, + 0.006965161766856909, + 0.004839685745537281, + 0.008109153248369694, + 0.013142969459295273, + 0.004624559078365564, + 0.00701687578111887, + 0.007627391256392002, + 0.006103767082095146, + 0.0029721481259912252, + 0.0030159587040543556, + 0.005516075063496828, + 0.005287750158458948, + 0.002969521563500166, + 0.003084450727328658, + 0.0049307215958833694, + 0.005770496558398008, + 0.0043367729522287846, + 0.0036881016567349434, + 0.006759734824299812, + 0.006782487966120243, + 0.004828483797609806, + 0.005439574830234051, + 0.004267229698598385, + 0.003648339072242379, + 0.0030814537312835455, + 0.0044433362782001495, + 0.004924507346004248, + 0.004703715909272432, + 0.0036568823270499706, + 0.0036304413806647062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.022843893617391586 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256016731262207 + ], + "opposite_logprob_tokens": [ + -1.6006016731262207 + ], + "logprob_reward": -0.2256016731262207, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256016731262207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20851372182369232, + 0.17759892344474792, + 0.057322051376104355, + 0.22596265375614166, + 0.15736894309520721, + 0.06047600507736206, + 0.016266223043203354, + 0.04763248190283775, + 0.022641334682703018, + 0.026217691600322723 + ], + "credit_map": [ + 0.03625934571027756, + 0.018970530480146408, + 0.021718857809901237, + 0.024159690365195274, + 0.029416874051094055, + 0.0309474878013134, + 0.03030131570994854, + 0.01673959754407406, + 0.031223999336361885, + 0.015918375924229622, + 0.014621629379689693, + 0.017802640795707703, + 0.029694776982069016, + 0.028176860883831978, + 0.024448055773973465, + 0.015712585300207138, + 0.006542637012898922, + 0.0064369430765509605, + 0.004721667151898146, + 0.00712327565997839, + 0.009471091441810131, + 0.008464663289487362, + 0.008851734921336174, + 0.005710036493837833, + 0.0608009435236454, + 0.038702335208654404, + 0.022034449502825737, + 0.017212534323334694, + 0.022402115166187286, + 0.037118516862392426, + 0.012828114442527294, + 0.014863653108477592, + 0.04099302366375923, + 0.022144559770822525, + 0.016002727672457695, + 0.011607593856751919, + 0.015972455963492393, + 0.025751430541276932, + 0.011325802654027939, + 0.013571356423199177, + 0.016335804015398026, + 0.009653455577790737, + 0.006020078435540199, + 0.00405670003965497, + 0.005672787316143513, + 0.009742162190377712, + 0.0037789777852594852, + 0.005216038785874844, + 0.0022347080521285534, + 0.0030108715873211622, + 0.0017217568820342422, + 0.0015108271036297083, + 0.0023533336352556944, + 0.0027298603672534227, + 0.0012750804889947176, + 0.0014297851594164968, + 0.009898080490529537, + 0.009351070038974285, + 0.003960188943892717, + 0.003953556530177593, + 0.006789007224142551, + 0.006984366569668055, + 0.003140073735266924, + 0.0035561376716941595, + 0.002182646654546261, + 0.0031088448595255613, + 0.0025217686779797077, + 0.002554928418248892, + 0.004097725730389357, + 0.0030946482438594103, + 0.0025339345447719097, + 0.0025468366220593452, + 0.003214098047465086, + 0.0033299890346825123, + 0.0022873899433761835, + 0.0027299788780510426, + 0.005716925952583551, + 0.003460734849795699, + 0.0025911384727805853, + 0.002887437120079994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.2256016731262207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583592355251312 + ], + "opposite_logprob_tokens": [ + -3.8955836296081543 + ], + "logprob_reward": -0.020583592355251312, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583592355251312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09562601894140244, + 0.2650257349014282, + 0.014284757897257805, + 0.21907934546470642, + 0.218242347240448, + 0.044506460428237915, + 0.04718136414885521, + 0.015923956409096718, + 0.013363768346607685, + 0.06676626950502396 + ], + "credit_map": [ + 0.013597538694739342, + 0.009402336552739143, + 0.006993521936237812, + 0.014498432166874409, + 0.015637565404176712, + 0.015392647124826908, + 0.012020368129014969, + 0.008083605207502842, + 0.04813874512910843, + 0.024244757369160652, + 0.022994376718997955, + 0.024951146915555, + 0.04073411226272583, + 0.04164716601371765, + 0.036166004836559296, + 0.026149427518248558, + 0.0010920579079538584, + 0.002268889918923378, + 0.0011336723109707236, + 0.0021674600429832935, + 0.002024053130298853, + 0.002222564537078142, + 0.0019167891005054116, + 0.0014592708321288228, + 0.057013992220163345, + 0.040111251175403595, + 0.02109796367585659, + 0.017086882144212723, + 0.02296692691743374, + 0.034978386014699936, + 0.010921100154519081, + 0.014902844093739986, + 0.06040521338582039, + 0.03187422826886177, + 0.021166758611798286, + 0.015104004181921482, + 0.02258303202688694, + 0.03681157901883125, + 0.014158968813717365, + 0.016138561069965363, + 0.008514069952070713, + 0.005907390266656876, + 0.003948265686631203, + 0.005657300818711519, + 0.004222969990223646, + 0.006747894920408726, + 0.004192806780338287, + 0.005315765272825956, + 0.007813842035830021, + 0.008081137202680111, + 0.004264465998858213, + 0.0037263331469148397, + 0.007295127958059311, + 0.007226310204714537, + 0.004645902663469315, + 0.004128244239836931, + 0.0025269724428653717, + 0.0025827987119555473, + 0.0013040630146861076, + 0.0016904280055314302, + 0.0023357633035629988, + 0.00237210001796484, + 0.0015453448286280036, + 0.0015664868988096714, + 0.0019579457584768534, + 0.00133787642698735, + 0.001304329140111804, + 0.0016476665623486042, + 0.0022073062136769295, + 0.0017483948031440377, + 0.0015411233762279153, + 0.00161912536714226, + 0.009939587675035, + 0.00891003105789423, + 0.006441459059715271, + 0.009065020829439163, + 0.009467716328799725, + 0.009477755054831505, + 0.006933818571269512, + 0.006530878134071827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.020583592355251312 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31339341402053833 + ], + "opposite_logprob_tokens": [ + -1.3133933544158936 + ], + "logprob_reward": -0.31339341402053833, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31339341402053833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07053782045841217, + 0.14212433993816376, + 0.040674369782209396, + 0.11019240319728851, + 0.2934250831604004, + 0.10178620368242264, + 0.07661934196949005, + 0.1017443910241127, + 0.013578815385699272, + 0.04931728169322014 + ], + "credit_map": [ + 0.010895914398133755, + 0.006455180235207081, + 0.006008630618453026, + 0.00963607244193554, + 0.011361091397702694, + 0.010259142145514488, + 0.009672223590314388, + 0.006249560508877039, + 0.022119132801890373, + 0.011219574138522148, + 0.011968336999416351, + 0.014011652208864689, + 0.025153346359729767, + 0.02243518829345703, + 0.020915009081363678, + 0.014302102848887444, + 0.0036647473461925983, + 0.004390500485897064, + 0.004729267675429583, + 0.005518778692930937, + 0.0063543058931827545, + 0.006185762584209442, + 0.005067958030849695, + 0.004763052798807621, + 0.02477160654962063, + 0.02189376950263977, + 0.008942182175815105, + 0.00859703216701746, + 0.014564632438123226, + 0.017122669145464897, + 0.006750285625457764, + 0.007550220005214214, + 0.07556076347827911, + 0.039487291127443314, + 0.031086832284927368, + 0.022463688626885414, + 0.028776787221431732, + 0.048572611063718796, + 0.022149620577692986, + 0.025327494367957115, + 0.027376271784305573, + 0.008753666654229164, + 0.010861102491617203, + 0.006650944240391254, + 0.012684510089457035, + 0.017027800902724266, + 0.011416353285312653, + 0.007015555631369352, + 0.012100040912628174, + 0.01028123963624239, + 0.006522041745483875, + 0.005047421436756849, + 0.016985423862934113, + 0.012528536841273308, + 0.006650920491665602, + 0.006503709591925144, + 0.0204587634652853, + 0.01988121308386326, + 0.00839718896895647, + 0.00934094563126564, + 0.013335701078176498, + 0.014830493368208408, + 0.006986353546380997, + 0.008513730950653553, + 0.0015520970337092876, + 0.001727546565234661, + 0.001084851217456162, + 0.0017976951785385609, + 0.002025818917900324, + 0.0021586616057902575, + 0.0018905374454334378, + 0.0013416067231446505, + 0.005571442656219006, + 0.0059920200146734715, + 0.005648854188621044, + 0.004706245381385088, + 0.008924909867346287, + 0.006802653893828392, + 0.006856306456029415, + 0.0048148478381335735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.31339341402053833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a giant Japanese spider crab (Macrocheira kaempferi)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab illuminated by a focused beam of light in an otherwise pitch-black environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab's carapace patterned with pale orange and grayish-white speckles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crab crouching motionless on a bed of fine-powdered sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot steady, eye-level, and framed as a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058207429945469 + ], + "opposite_logprob_tokens": [ + -3.8955819606781006 + ], + "logprob_reward": -0.02058207429945469, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058207429945469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13972455263137817, + 0.07543415576219559, + 0.03745167329907417, + 0.3114231526851654, + 0.149595707654953, + 0.10228363424539566, + 0.048330582678318024, + 0.032874953001737595, + 0.044839583337306976, + 0.058042027056217194 + ], + "credit_map": [ + 0.015134606510400772, + 0.013010995462536812, + 0.02037862502038479, + 0.016914445906877518, + 0.016375700011849403, + 0.02026093751192093, + 0.021589437499642372, + 0.01605979911983013, + 0.011578907258808613, + 0.009501240216195583, + 0.008129497058689594, + 0.010606680996716022, + 0.009933915920555592, + 0.010539853945374489, + 0.009144456125795841, + 0.005999599117785692, + 0.0038069591391831636, + 0.005394387990236282, + 0.003878618823364377, + 0.004153369925916195, + 0.005505161825567484, + 0.006000282242894173, + 0.004135476890951395, + 0.00457741878926754, + 0.08153024315834045, + 0.05693918839097023, + 0.029475552961230278, + 0.023743893951177597, + 0.03392991051077843, + 0.04944990575313568, + 0.015316917560994625, + 0.021037554368376732, + 0.0421038419008255, + 0.02017943374812603, + 0.014297466725111008, + 0.009940105490386486, + 0.01669827103614807, + 0.025422528386116028, + 0.009299232624471188, + 0.011654829606413841, + 0.024227021262049675, + 0.011615552008152008, + 0.010659396648406982, + 0.008638289757072926, + 0.010468808934092522, + 0.01640152931213379, + 0.007769291754812002, + 0.012503743171691895, + 0.00929027795791626, + 0.006409213878214359, + 0.004186095204204321, + 0.0033679150510579348, + 0.009762253612279892, + 0.007215287070721388, + 0.00373320234939456, + 0.004366337321698666, + 0.0062703522853553295, + 0.004191240295767784, + 0.003005705773830414, + 0.0034779785200953484, + 0.004178953357040882, + 0.004628994036465883, + 0.0033087434712797403, + 0.003812988055869937, + 0.004454423673450947, + 0.006466808263212442, + 0.004594304133206606, + 0.004817264620214701, + 0.009867578744888306, + 0.00661375280469656, + 0.004013439174741507, + 0.004012007266283035, + 0.007060634437948465, + 0.008979043923318386, + 0.003671664744615555, + 0.008058241568505764, + 0.008928583934903145, + 0.008428557775914669, + 0.00817476399242878, + 0.0047405376099050045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.02058207429945469 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_11/index_tensor([1332, 1332])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_11/index_tensor([1332, 1332])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c4dcd452eef9fc5681c02b79ae36cbc384edcc2f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_11/index_tensor([1332, 1332])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight tracking shot, shot from the side and slightly from an upward angle, closely follows the resolute face of a rickshaw puller. He wears a worn, dark cloth cap, his bronzed cheeks crisscrossed with deep wrinkles. Large beads of sweat trickle down his furrowed brow and clenched jawline, gleaming faintly in the warm sunset. He leans forward, his taut upper body clad only in a sweat-soaked, clinging gray sweatshirt. With every ounce of his strength, he pulls a yellow freight tricycle, its paint peeling in patches, revealing the rust beneath. The wheels creak heavily and laboriously as they bump against the uneven, time-worn bluestone slabs of an ancient stone arch bridge. In the background, a moss-covered stone railing and a hazy view of the bridge below loom. The entire picture is stained with a warm yellow hue of retro art, like a faded old photo. The soft golden light lengthens the shadows of people and cars, profoundly depicting an indomitable resilience of life and a profound sense of the times in the difficult life.", + "seed": 31989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_11/index_1332_seed_31989_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040626536775380376, + "vlm_logprob": -0.040626536775380376, + "vlm_yes_no_margin": 3.599999952316284, + "vlm_token_credit": -0.040626536775380376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054565392434597015 + ], + "opposite_logprob_tokens": [ + -3.0545654296875 + ], + "logprob_reward": -0.054565392434597015, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.054565392434597015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12264671921730042, + 0.050712354481220245, + 0.207100510597229, + 0.07059691846370697, + 0.07600310444831848, + 0.2233516126871109, + 0.07429657131433487, + 0.10152572393417358, + 0.019124558195471764, + 0.05464194715023041 + ], + "credit_map": [ + 0.010800029151141644, + 0.01864451728761196, + 0.014132441021502018, + 0.013772866688668728, + 0.019527889788150787, + 0.0200850460678339, + 0.01408599503338337, + 0.01159793883562088, + 0.004661387763917446, + 0.004556773230433464, + 0.007293714210391045, + 0.008034351281821728, + 0.00852999184280634, + 0.005773589946329594, + 0.006774063687771559, + 0.005088480189442635, + 0.01803567074239254, + 0.020829372107982635, + 0.03669465333223343, + 0.02549757994711399, + 0.03783136606216431, + 0.026615815237164497, + 0.023449702188372612, + 0.018146349117159843, + 0.004416351206600666, + 0.006954375654459, + 0.009687569923698902, + 0.015436820685863495, + 0.009465038776397705, + 0.01265462301671505, + 0.005542938597500324, + 0.00643920386210084, + 0.00787061545997858, + 0.011487677693367004, + 0.011953691020607948, + 0.006225713528692722, + 0.010332053527235985, + 0.006159368436783552, + 0.005958836060017347, + 0.016015153378248215, + 0.00843424815684557, + 0.01222438458353281, + 0.016217850148677826, + 0.07976892590522766, + 0.021943528205156326, + 0.01382121816277504, + 0.012395370751619339, + 0.05854608491063118, + 0.006680878344923258, + 0.004722895100712776, + 0.00871853344142437, + 0.009730630554258823, + 0.0109822116792202, + 0.009249003604054451, + 0.009731576777994633, + 0.014480846002697945, + 0.006358740385621786, + 0.009686910547316074, + 0.010740452446043491, + 0.03412347659468651, + 0.013742503710091114, + 0.004707647487521172, + 0.008861541748046875, + 0.013304446823894978, + 0.0017334098229184747, + 0.003210214199498296, + 0.0015606465749442577, + 0.0033232783898711205, + 0.003351970575749874, + 0.001838536816649139, + 0.0020226300694048405, + 0.002083870815113187, + 0.005611240863800049, + 0.0048067159950733185, + 0.005829586647450924, + 0.0074796415865421295, + 0.013033529743552208, + 0.007949462160468102, + 0.00482005812227726, + 0.005111712496727705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.054565392434597015 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308105766773224 + ], + "opposite_logprob_tokens": [ + -3.773308038711548 + ], + "logprob_reward": -0.023308105766773224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308105766773224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08483611047267914, + 0.0705966055393219, + 0.20738548040390015, + 0.15481412410736084, + 0.05705835297703743, + 0.049223557114601135, + 0.024766255170106888, + 0.10290780663490295, + 0.06895864009857178, + 0.17945311963558197 + ], + "credit_map": [ + 0.005367204546928406, + 0.010630722157657146, + 0.008510311134159565, + 0.011171400547027588, + 0.014304625801742077, + 0.015650276094675064, + 0.010274243541061878, + 0.008927319198846817, + 0.006489220540970564, + 0.004700899589806795, + 0.005919584538787603, + 0.006923094857484102, + 0.013532011769711971, + 0.009613650850951672, + 0.009488037787377834, + 0.013930105604231358, + 0.025487370789051056, + 0.01768695004284382, + 0.04818342998623848, + 0.026239456608891487, + 0.024352846667170525, + 0.021475937217473984, + 0.02085142955183983, + 0.02310805395245552, + 0.01378466933965683, + 0.014580111019313335, + 0.010511069558560848, + 0.022781431674957275, + 0.02709013968706131, + 0.023739099502563477, + 0.025499094277620316, + 0.016828512772917747, + 0.005569551605731249, + 0.008231491781771183, + 0.008070523850619793, + 0.006100568920373917, + 0.012470684014260769, + 0.004571923054754734, + 0.00413438118994236, + 0.007909227162599564, + 0.002484730677679181, + 0.003356187604367733, + 0.004095012787729502, + 0.01695798896253109, + 0.0053134215995669365, + 0.0038955938071012497, + 0.0028877677395939827, + 0.010232851840555668, + 0.002870312426239252, + 0.0035274254623800516, + 0.002118783537298441, + 0.003157873172312975, + 0.0023400033824145794, + 0.003521199570968747, + 0.003022874938324094, + 0.0042077843099832535, + 0.005329021718353033, + 0.011272428557276726, + 0.0067128646187484264, + 0.03704231604933739, + 0.012438793666660786, + 0.003661569207906723, + 0.008660346269607544, + 0.01779046282172203, + 0.0032062730751931667, + 0.005046014208346605, + 0.002612179843708873, + 0.024802356958389282, + 0.010474230162799358, + 0.0041665988974273205, + 0.004732629284262657, + 0.013918358832597733, + 0.017808450385928154, + 0.00893310271203518, + 0.0504717119038105, + 0.026736389845609665, + 0.01977989822626114, + 0.023000922054052353, + 0.015525786206126213, + 0.017196858301758766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.023308105766773224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738083902746439 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738083902746439, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738083902746439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14817650616168976, + 0.0688219740986824, + 0.0678177922964096, + 0.07480940967798233, + 0.0945059135556221, + 0.28291916847229004, + 0.03743279352784157, + 0.07718901336193085, + 0.04839635640382767, + 0.0999310314655304 + ], + "credit_map": [ + 0.011603022925555706, + 0.010286490432918072, + 0.013412026688456535, + 0.01773807965219021, + 0.027932919561862946, + 0.036124832928180695, + 0.014034537598490715, + 0.017044605687260628, + 0.005198497325181961, + 0.005734134465456009, + 0.0062097725458443165, + 0.0076071773655712605, + 0.013154488056898117, + 0.00959869660437107, + 0.008249092847108841, + 0.013070115819573402, + 0.005812800023704767, + 0.0035667200572788715, + 0.00967935286462307, + 0.012529811821877956, + 0.007669663522392511, + 0.009676755405962467, + 0.008988523855805397, + 0.00989417266100645, + 0.006777256261557341, + 0.008230635896325111, + 0.006720268167555332, + 0.007722506765276194, + 0.009024368599057198, + 0.019136905670166016, + 0.008064290508627892, + 0.009133180603384972, + 0.007912720553576946, + 0.012259275652468204, + 0.019172174856066704, + 0.010572458617389202, + 0.011159711517393589, + 0.008906512521207333, + 0.006528319325298071, + 0.017994750291109085, + 0.010166192427277565, + 0.012901965528726578, + 0.01871357299387455, + 0.10197191685438156, + 0.0273008793592453, + 0.01912716031074524, + 0.01618334650993347, + 0.07655416429042816, + 0.003386699827387929, + 0.003992154728621244, + 0.0030884118750691414, + 0.004936530254781246, + 0.007353797554969788, + 0.0030422548297792673, + 0.005383085459470749, + 0.0062498608604073524, + 0.003874578047543764, + 0.008238148875534534, + 0.005888660904020071, + 0.024624550715088844, + 0.007658906746655703, + 0.00425514904782176, + 0.007961036637425423, + 0.014687991701066494, + 0.0024823849089443684, + 0.003995081409811974, + 0.003209443297237158, + 0.014397908002138138, + 0.006985972169786692, + 0.004423009697347879, + 0.004678288009017706, + 0.00822427123785019, + 0.011921146884560585, + 0.007023973856121302, + 0.014503740705549717, + 0.013781892135739326, + 0.022456206381320953, + 0.01439749263226986, + 0.007442498579621315, + 0.008404087275266647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.006738083902746439 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033965110778809 + ], + "opposite_logprob_tokens": [ + -2.350339651107788 + ], + "logprob_reward": -0.10033965110778809, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033965110778809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19725391268730164, + 0.07286859303712845, + 0.05298756808042526, + 0.20231112837791443, + 0.08510497212409973, + 0.07680286467075348, + 0.0465700589120388, + 0.07201749086380005, + 0.03750302270054817, + 0.15658038854599 + ], + "credit_map": [ + 0.011813096702098846, + 0.017481787130236626, + 0.015459183603525162, + 0.02187551185488701, + 0.02346641756594181, + 0.060518186539411545, + 0.021040290594100952, + 0.025599444285035133, + 0.004562825430184603, + 0.005967015400528908, + 0.006717510521411896, + 0.009145457297563553, + 0.013173011131584644, + 0.015055151656270027, + 0.010118970647454262, + 0.008128651417791843, + 0.004780054558068514, + 0.004737816751003265, + 0.0074407728388905525, + 0.007274267263710499, + 0.006822061724960804, + 0.008312297984957695, + 0.006961105391383171, + 0.006659192964434624, + 0.018650325015187263, + 0.016770804300904274, + 0.01595318503677845, + 0.024115197360515594, + 0.02331177145242691, + 0.059877101331949234, + 0.01783672161400318, + 0.025796037167310715, + 0.007151767145842314, + 0.008425764739513397, + 0.009419657289981842, + 0.009223616681993008, + 0.009210784919559956, + 0.013817550614476204, + 0.006127497181296349, + 0.021728338673710823, + 0.0027458511758595705, + 0.005155955441296101, + 0.005052736960351467, + 0.02220967970788479, + 0.007361287251114845, + 0.011169017292559147, + 0.006880213972181082, + 0.016228128224611282, + 0.0032774589490145445, + 0.00417328393086791, + 0.0033586272038519382, + 0.004560651257634163, + 0.004421260207891464, + 0.013873955234885216, + 0.005366384517401457, + 0.007538438308984041, + 0.0035998658277094364, + 0.008171642199158669, + 0.006035007536411285, + 0.024414153769612312, + 0.007815384306013584, + 0.002793736057356, + 0.007724007125943899, + 0.011463695205748081, + 0.0027162397745996714, + 0.005515386816114187, + 0.0028764090966433287, + 0.008727744221687317, + 0.005682657472789288, + 0.003467387054115534, + 0.0037881259340792894, + 0.004729073960334063, + 0.020675862208008766, + 0.012471181340515614, + 0.021844930946826935, + 0.019254431128501892, + 0.03471978008747101, + 0.021156297996640205, + 0.012646337039768696, + 0.013811561278998852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.10033965110778809 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181814506649971 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.0181814506649971, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181814506649971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07636407762765884, + 0.03079916536808014, + 0.07594344019889832, + 0.03389084339141846, + 0.0551748089492321, + 0.5271198153495789, + 0.04407574608922005, + 0.10307562351226807, + 0.01752507872879505, + 0.036031439900398254 + ], + "credit_map": [ + 0.005351721774786711, + 0.010461598634719849, + 0.009391880594193935, + 0.01354400534182787, + 0.013961337506771088, + 0.006360916420817375, + 0.00873494241386652, + 0.008557680994272232, + 0.002918279031291604, + 0.003910085652023554, + 0.003178744111210108, + 0.0034594805911183357, + 0.004444860853254795, + 0.00408101174980402, + 0.0030121973250061274, + 0.005794505123049021, + 0.007012564223259687, + 0.005115458741784096, + 0.009735751897096634, + 0.01276854146271944, + 0.009719669818878174, + 0.0065550366416573524, + 0.010813030414283276, + 0.01422338280826807, + 0.002673896262422204, + 0.003342758398503065, + 0.0068662515841424465, + 0.005122676957398653, + 0.005656323861330748, + 0.0036261966452002525, + 0.003303465200588107, + 0.0032992742490023375, + 0.0053758942522108555, + 0.008376890793442726, + 0.010728464461863041, + 0.005422510206699371, + 0.009297305718064308, + 0.0041558705270290375, + 0.0034316875971853733, + 0.008386187255382538, + 0.01923195831477642, + 0.02554934285581112, + 0.03426137566566467, + 0.19244015216827393, + 0.05277979001402855, + 0.03406883776187897, + 0.030230112373828888, + 0.13855819404125214, + 0.0029698535799980164, + 0.003130638739094138, + 0.00290189147926867, + 0.006568863522261381, + 0.004386096261441708, + 0.013985571451485157, + 0.0038879101630300283, + 0.006244918331503868, + 0.00449287798255682, + 0.010284073650836945, + 0.007618285249918699, + 0.03557688742876053, + 0.01058188732713461, + 0.004444012884050608, + 0.0088832788169384, + 0.021194305270910263, + 0.0012027488555759192, + 0.0016473716823384166, + 0.001643806230276823, + 0.003303967649117112, + 0.0024408085737377405, + 0.0025793903041630983, + 0.0023569585755467415, + 0.002350026275962591, + 0.003379976376891136, + 0.0033070140052586794, + 0.004512316081672907, + 0.005386857315897942, + 0.006099049933254719, + 0.005646888632327318, + 0.0033456471282988787, + 0.004353689961135387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.0181814506649971 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight tracking shot, shot from the side and slightly from an upward angle, closely follows the resolute face of a rickshaw puller. He wears a worn, dark cloth cap, his bronzed cheeks crisscrossed with deep wrinkles. Large beads of sweat trickle down his furrowed brow and clenched jawline, gleaming faintly in the warm sunset. He leans forward, his taut upper body clad only in a sweat-soaked, clinging gray sweatshirt. With every ounce of his strength, he pulls a yellow freight tricycle, its paint peeling in patches, revealing the rust beneath. The wheels creak heavily and laboriously as they bump against the uneven, time-worn bluestone slabs of an ancient stone arch bridge. In the background, a moss-covered stone railing and a hazy view of the bridge below loom. The entire picture is stained with a warm yellow hue of retro art, like a faded old photo. The soft golden light lengthens the shadows of people and cars, profoundly depicting an indomitable resilience of life and a profound sense of the times in the difficult life.", + "seed": 31990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_11/index_1332_seed_31990_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11609661467373371, + "vlm_logprob": -0.11609661467373371, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.11609661467373371, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03832632303237915 + ], + "opposite_logprob_tokens": [ + -3.4133262634277344 + ], + "logprob_reward": -0.03832632303237915, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03832632303237915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0976610854268074, + 0.08259878307580948, + 0.04227520152926445, + 0.0637139081954956, + 0.18577080965042114, + 0.37449026107788086, + 0.0534711591899395, + 0.04197540134191513, + 0.03240980580449104, + 0.025633538141846657 + ], + "credit_map": [ + 0.013962216675281525, + 0.011038151569664478, + 0.011289594694972038, + 0.011630150489509106, + 0.01487946417182684, + 0.01312852744013071, + 0.013501542620360851, + 0.008231432177126408, + 0.012904255650937557, + 0.011934850364923477, + 0.011273941025137901, + 0.00731956260278821, + 0.009694038890302181, + 0.011875484138727188, + 0.010887126438319683, + 0.006709529086947441, + 0.004872466437518597, + 0.00493990583345294, + 0.004582039546221495, + 0.00503066461533308, + 0.006542158778756857, + 0.006620469503104687, + 0.004757711198180914, + 0.004929784685373306, + 0.008732800371944904, + 0.004230814520269632, + 0.005671681836247444, + 0.00804075039923191, + 0.005831720307469368, + 0.010796123184263706, + 0.01648220606148243, + 0.003927808720618486, + 0.008906260132789612, + 0.024130016565322876, + 0.014317678287625313, + 0.02875123731791973, + 0.059849418699741364, + 0.015174505300819874, + 0.024627070873975754, + 0.01001463457942009, + 0.031290896236896515, + 0.06075514107942581, + 0.057211004197597504, + 0.05271630734205246, + 0.03390966355800629, + 0.035662367939949036, + 0.04291650652885437, + 0.06002839282155037, + 0.005816992372274399, + 0.008751272223889828, + 0.007214410696178675, + 0.00893209408968687, + 0.00624864362180233, + 0.0037461528554558754, + 0.00810813345015049, + 0.0046534608118236065, + 0.0044467877596616745, + 0.00434867711737752, + 0.006773966830223799, + 0.005561373196542263, + 0.004338886588811874, + 0.005511439871042967, + 0.00522575294598937, + 0.005768518894910812, + 0.0034389791544526815, + 0.005041263531893492, + 0.003764210967347026, + 0.0051247538067400455, + 0.003598055336624384, + 0.0025452743284404278, + 0.005694222170859575, + 0.0032030451111495495, + 0.005322617013007402, + 0.0028915898874402046, + 0.0033289433922618628, + 0.002842071931809187, + 0.0025676290970295668, + 0.003108702600002289, + 0.0029814555309712887, + 0.002590527758002281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.03832632303237915 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360709220170975 + ], + "opposite_logprob_tokens": [ + -3.6513607501983643 + ], + "logprob_reward": -0.026360709220170975, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360709220170975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17696455121040344, + 0.2459517866373062, + 0.06365832686424255, + 0.06351160258054733, + 0.09779377281665802, + 0.07086148858070374, + 0.09087198227643967, + 0.0678766667842865, + 0.05956371873617172, + 0.06294607371091843 + ], + "credit_map": [ + 0.020076990127563477, + 0.014143326319754124, + 0.01757168024778366, + 0.01830996572971344, + 0.035747379064559937, + 0.027462858706712723, + 0.023386454209685326, + 0.020265907049179077, + 0.015430308878421783, + 0.015872333198785782, + 0.04263553395867348, + 0.038737084716558456, + 0.03686648979783058, + 0.016504568979144096, + 0.05525178462266922, + 0.02465367503464222, + 0.006760308984667063, + 0.00769772008061409, + 0.009724552743136883, + 0.007816803641617298, + 0.008084739558398724, + 0.008881964720785618, + 0.008859584107995033, + 0.005832658149302006, + 0.0036131292581558228, + 0.00575009360909462, + 0.008387361653149128, + 0.00808742642402649, + 0.009328695945441723, + 0.013939475640654564, + 0.010736274532973766, + 0.00366914551705122, + 0.009493592195212841, + 0.012973005883395672, + 0.009057195857167244, + 0.01100215408951044, + 0.025058593600988388, + 0.012819457799196243, + 0.012599914334714413, + 0.004789866507053375, + 0.010677526704967022, + 0.011708814650774002, + 0.012832381762564182, + 0.009397084824740887, + 0.006266273092478514, + 0.006600346881896257, + 0.007545409258455038, + 0.00583365885540843, + 0.011438462883234024, + 0.016392122954130173, + 0.013533404096961021, + 0.01568586565554142, + 0.008788895793259144, + 0.007721899542957544, + 0.008794997818768024, + 0.008516344241797924, + 0.010162864811718464, + 0.011401033028960228, + 0.013081233948469162, + 0.009913529269397259, + 0.005258406978100538, + 0.0062628514133393764, + 0.0066554853692650795, + 0.005141272209584713, + 0.008075227029621601, + 0.0075593614019453526, + 0.0075263711623847485, + 0.009975632652640343, + 0.005968149751424789, + 0.0059728482738137245, + 0.01005372405052185, + 0.0044324155896902084, + 0.011005723848938942, + 0.005945301614701748, + 0.008796906098723412, + 0.007064223755151033, + 0.00750710116699338, + 0.007986702024936676, + 0.008605502545833588, + 0.006034614518284798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.026360709220170975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818636804819107 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.01818636804819107, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818636804819107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1934637725353241, + 0.23724722862243652, + 0.059998027980327606, + 0.03620891273021698, + 0.1386677622795105, + 0.04114750772714615, + 0.06382633745670319, + 0.12671130895614624, + 0.053153228014707565, + 0.04957582801580429 + ], + "credit_map": [ + 0.043022219091653824, + 0.016200443729758263, + 0.03096577152609825, + 0.023039400577545166, + 0.022188078612089157, + 0.021370796486735344, + 0.025245655328035355, + 0.011431427672505379, + 0.030098263174295425, + 0.02207271382212639, + 0.04885798692703247, + 0.015631457790732384, + 0.0252637080848217, + 0.01839515194296837, + 0.06251347810029984, + 0.014414507895708084, + 0.00774418655782938, + 0.006976673845201731, + 0.007981901057064533, + 0.006699013523757458, + 0.007777261547744274, + 0.007763213478028774, + 0.009725645184516907, + 0.005330141633749008, + 0.00314100063405931, + 0.0027882743161171675, + 0.004158565774559975, + 0.005469819065183401, + 0.004501515533775091, + 0.006066830363124609, + 0.006522772368043661, + 0.003560140263289213, + 0.010578105226159096, + 0.015414606779813766, + 0.01586376503109932, + 0.01959208957850933, + 0.03421648219227791, + 0.01939629204571247, + 0.01753184199333191, + 0.006074601784348488, + 0.004292549565434456, + 0.006863482296466827, + 0.005598384886980057, + 0.006149105727672577, + 0.0029409900307655334, + 0.0035878170747309923, + 0.005189809016883373, + 0.006525375880300999, + 0.006806850899010897, + 0.01300586573779583, + 0.010900146327912807, + 0.010402685031294823, + 0.007022419944405556, + 0.006037497892975807, + 0.004928974434733391, + 0.004721904639154673, + 0.015137610025703907, + 0.015337585471570492, + 0.0217971820384264, + 0.019544456154108047, + 0.010359733365476131, + 0.014265197329223156, + 0.015606955625116825, + 0.014662597328424454, + 0.008307416923344135, + 0.007335437927395105, + 0.008058219216763973, + 0.006525353062897921, + 0.005852316040545702, + 0.0060464139096438885, + 0.007031060289591551, + 0.003997023217380047, + 0.006599529646337032, + 0.005801819264888763, + 0.005693317390978336, + 0.005454637575894594, + 0.00802835263311863, + 0.006638417020440102, + 0.006670810282230377, + 0.004688952583819628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.01818636804819107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743213653564453 + ], + "opposite_logprob_tokens": [ + -0.9743213653564453 + ], + "logprob_reward": -0.4743213653564453, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743213653564453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25686389207839966, + 0.1575317531824112, + 0.09626032412052155, + 0.03609231859445572, + 0.16930091381072998, + 0.05074639245867729, + 0.06530215591192245, + 0.03502991050481796, + 0.07748645544052124, + 0.055385835468769073 + ], + "credit_map": [ + 0.07627029716968536, + 0.024449819698929787, + 0.033029910176992416, + 0.026128489524126053, + 0.02107624150812626, + 0.032962094992399216, + 0.02724415250122547, + 0.015702886506915092, + 0.024143077433109283, + 0.018818970769643784, + 0.015421802178025246, + 0.009667002595961094, + 0.023007644340395927, + 0.017213748767971992, + 0.038646090775728226, + 0.010613423772156239, + 0.012415895238518715, + 0.01032012328505516, + 0.017528222873806953, + 0.008339868858456612, + 0.012891273945569992, + 0.012534373439848423, + 0.014928948134183884, + 0.007301620673388243, + 0.002427372382953763, + 0.0030219333712011576, + 0.003713317448273301, + 0.005886171478778124, + 0.00442472193390131, + 0.0065848249942064285, + 0.005992836784571409, + 0.004041143227368593, + 0.020458625629544258, + 0.01767563447356224, + 0.019925609230995178, + 0.02339348755776882, + 0.031371794641017914, + 0.028085216879844666, + 0.02134501188993454, + 0.007045568898320198, + 0.005385224241763353, + 0.005705769173800945, + 0.006071188952773809, + 0.007132690865546465, + 0.0035425610840320587, + 0.005912007763981819, + 0.006695258896797895, + 0.010301691479980946, + 0.007000630721449852, + 0.01351949106901884, + 0.010547983460128307, + 0.01049343403428793, + 0.00605111476033926, + 0.00499412976205349, + 0.006880305241793394, + 0.005815071519464254, + 0.0051329233683645725, + 0.004858978092670441, + 0.007591933012008667, + 0.005478405859321356, + 0.002388455206528306, + 0.0034832172095775604, + 0.0031034478452056646, + 0.0029925545677542686, + 0.013692429289221764, + 0.010157397948205471, + 0.00991921778768301, + 0.010185916908085346, + 0.008081766776740551, + 0.008548692800104618, + 0.010995377786457539, + 0.005905660334974527, + 0.007104041054844856, + 0.006294189020991325, + 0.007636691443622112, + 0.006618678569793701, + 0.008216064423322678, + 0.007455408573150635, + 0.007298602256923914, + 0.004762164782732725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.4743213653564453 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288307711482048 + ], + "opposite_logprob_tokens": [ + -3.7732882499694824 + ], + "logprob_reward": -0.023288307711482048, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288307711482048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21556547284126282, + 0.1556963473558426, + 0.08641441911458969, + 0.05949948728084564, + 0.11692608147859573, + 0.07328014820814133, + 0.09003904461860657, + 0.0435335636138916, + 0.08098453283309937, + 0.07806088030338287 + ], + "credit_map": [ + 0.055356334894895554, + 0.01524980366230011, + 0.025581305846571922, + 0.021397078409790993, + 0.02043963596224785, + 0.036340389400720596, + 0.02402966096997261, + 0.017171276733279228, + 0.025252077728509903, + 0.023401865735650063, + 0.016343753784894943, + 0.012042607180774212, + 0.021503355354070663, + 0.019223880022764206, + 0.025714239105582237, + 0.012214571237564087, + 0.007148980628699064, + 0.0061583137139678, + 0.016174931079149246, + 0.008752157911658287, + 0.010340931825339794, + 0.009627111256122589, + 0.01864554174244404, + 0.009566453285515308, + 0.00525255873799324, + 0.00560747692361474, + 0.008465767838060856, + 0.011059756390750408, + 0.008617945946753025, + 0.006569996010512114, + 0.008449711836874485, + 0.00547627592459321, + 0.01095972117036581, + 0.014423687011003494, + 0.01403663121163845, + 0.017604228109121323, + 0.023250557482242584, + 0.01572221703827381, + 0.014928732998669147, + 0.006000306457281113, + 0.007207245100289583, + 0.008263994008302689, + 0.0068605272099375725, + 0.01026563998311758, + 0.0055538564920425415, + 0.01000890415161848, + 0.009433825500309467, + 0.01568615436553955, + 0.010337701067328453, + 0.01813321001827717, + 0.013353750109672546, + 0.015273530967533588, + 0.009886507876217365, + 0.008212285116314888, + 0.007303682621568441, + 0.007538372650742531, + 0.005624414887279272, + 0.006193287670612335, + 0.009046162478625774, + 0.0066339410841465, + 0.003280924167484045, + 0.0045919292606413364, + 0.0038049849681556225, + 0.004357917699962854, + 0.008749259635806084, + 0.012574797496199608, + 0.009049012325704098, + 0.013911972753703594, + 0.008936766535043716, + 0.007010199595242739, + 0.013215997256338596, + 0.007536530494689941, + 0.008785474114120007, + 0.007912871427834034, + 0.01222112774848938, + 0.011097272858023643, + 0.011114034801721573, + 0.007663412019610405, + 0.010017883032560349, + 0.009248800575733185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.023288307711482048 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight tracking shot, shot from the side and slightly from an upward angle, closely follows the resolute face of a rickshaw puller. He wears a worn, dark cloth cap, his bronzed cheeks crisscrossed with deep wrinkles. Large beads of sweat trickle down his furrowed brow and clenched jawline, gleaming faintly in the warm sunset. He leans forward, his taut upper body clad only in a sweat-soaked, clinging gray sweatshirt. With every ounce of his strength, he pulls a yellow freight tricycle, its paint peeling in patches, revealing the rust beneath. The wheels creak heavily and laboriously as they bump against the uneven, time-worn bluestone slabs of an ancient stone arch bridge. In the background, a moss-covered stone railing and a hazy view of the bridge below loom. The entire picture is stained with a warm yellow hue of retro art, like a faded old photo. The soft golden light lengthens the shadows of people and cars, profoundly depicting an indomitable resilience of life and a profound sense of the times in the difficult life.", + "seed": 31991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_11/index_1332_seed_31991_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10458665806800127, + "vlm_logprob": -0.10458665806800127, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.10458665806800127, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038946688175201416 + ], + "opposite_logprob_tokens": [ + -3.4139466285705566 + ], + "logprob_reward": -0.038946688175201416, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.038946688175201416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23022449016571045, + 0.12354206293821335, + 0.09413132816553116, + 0.10468920320272446, + 0.12094692140817642, + 0.12417175620794296, + 0.09096856415271759, + 0.02651458978652954, + 0.05042903870344162, + 0.03438199311494827 + ], + "credit_map": [ + 0.030179794877767563, + 0.03618824481964111, + 0.02257624827325344, + 0.031521186232566833, + 0.02133607305586338, + 0.0467248409986496, + 0.024255728349089622, + 0.017442375421524048, + 0.00888498593121767, + 0.019534440711140633, + 0.026066195219755173, + 0.014319470152258873, + 0.017398623749613762, + 0.01045242603868246, + 0.014543911442160606, + 0.012342024594545364, + 0.006302000489085913, + 0.02576005458831787, + 0.008086667396128178, + 0.027693748474121094, + 0.008914241567254066, + 0.004337632097303867, + 0.0073486375622451305, + 0.005688354838639498, + 0.012488466687500477, + 0.009441080503165722, + 0.025794003158807755, + 0.010289832018315792, + 0.010507036000490189, + 0.021731669083237648, + 0.0066998484544456005, + 0.007737277075648308, + 0.007654487155377865, + 0.009212205186486244, + 0.02610410563647747, + 0.008233544416725636, + 0.008565345779061317, + 0.011258657090365887, + 0.03898195922374725, + 0.01093662716448307, + 0.0099182752892375, + 0.008270992897450924, + 0.01444177981466055, + 0.02916882000863552, + 0.0116419093683362, + 0.030689619481563568, + 0.01321902684867382, + 0.006821340881288052, + 0.007810106500983238, + 0.009657922200858593, + 0.013984281569719315, + 0.01271761767566204, + 0.011584710329771042, + 0.010164091363549232, + 0.011888806708157063, + 0.013161039911210537, + 0.0017012308817356825, + 0.003856643568724394, + 0.003142312169075012, + 0.003140215063467622, + 0.0037491347175091505, + 0.00313349487259984, + 0.004167754668742418, + 0.00362380500882864, + 0.0034643644466996193, + 0.004673339892178774, + 0.006077412981539965, + 0.006658095866441727, + 0.009874406270682812, + 0.0051916916854679585, + 0.008065490052103996, + 0.006424236577004194, + 0.002298546489328146, + 0.003388389479368925, + 0.0034327455796301365, + 0.004671383649110794, + 0.0062889885157346725, + 0.005335123743861914, + 0.005204806104302406, + 0.0037620109505951405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.038946688175201416 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636488899588585 + ], + "opposite_logprob_tokens": [ + -3.651364803314209 + ], + "logprob_reward": -0.02636488899588585, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636488899588585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09345532208681107, + 0.07203646004199982, + 0.052883733063936234, + 0.0812031701207161, + 0.06693582236766815, + 0.32725679874420166, + 0.05747176706790924, + 0.02791660465300083, + 0.11951208859682083, + 0.10132817178964615 + ], + "credit_map": [ + 0.013259450905025005, + 0.015477992594242096, + 0.012991166673600674, + 0.013935796916484833, + 0.008010456338524818, + 0.010637945495545864, + 0.012291629798710346, + 0.0068508838303387165, + 0.005410690791904926, + 0.01037097629159689, + 0.00770188495516777, + 0.010028792545199394, + 0.011632396839559078, + 0.010418388061225414, + 0.010606667958199978, + 0.005866665858775377, + 0.005742647219449282, + 0.007751356344670057, + 0.007273297756910324, + 0.009602515026926994, + 0.006825071293860674, + 0.004168651532381773, + 0.00690151983872056, + 0.004618672654032707, + 0.011229949072003365, + 0.010037452913820744, + 0.00914147961884737, + 0.00945199839770794, + 0.013937135227024555, + 0.007620096206665039, + 0.008242592215538025, + 0.011542475782334805, + 0.003935488406568766, + 0.007822400890290737, + 0.010185923427343369, + 0.0049245418049395084, + 0.00844302773475647, + 0.010168345645070076, + 0.015848079696297646, + 0.005608019419014454, + 0.04495968297123909, + 0.019375160336494446, + 0.023640280589461327, + 0.07840315252542496, + 0.028215734288096428, + 0.07852284610271454, + 0.028364630416035652, + 0.025775335729122162, + 0.00412165280431509, + 0.006286428775638342, + 0.009541715495288372, + 0.008476126939058304, + 0.007829674519598484, + 0.005208615679293871, + 0.00828054640442133, + 0.007727011572569609, + 0.0021127876825630665, + 0.005859097465872765, + 0.004072709009051323, + 0.004066601395606995, + 0.003337861504405737, + 0.0018478721613064408, + 0.0038836749736219645, + 0.002736002439633012, + 0.009366476908326149, + 0.0082362350076437, + 0.009914196096360683, + 0.01096457988023758, + 0.015909887850284576, + 0.012919940054416656, + 0.013878866098821163, + 0.038321904838085175, + 0.007321435492485762, + 0.007701459806412458, + 0.016091078519821167, + 0.013765254989266396, + 0.015302996151149273, + 0.008800623007118702, + 0.01996959187090397, + 0.012375729158520699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.02636488899588585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608443818986416 + ], + "opposite_logprob_tokens": [ + -4.141084671020508 + ], + "logprob_reward": -0.01608443818986416, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608443818986416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16931867599487305, + 0.054972317069768906, + 0.13655075430870056, + 0.09843210875988007, + 0.15154898166656494, + 0.1324763000011444, + 0.1000678613781929, + 0.03713949769735336, + 0.06876417994499207, + 0.05072926729917526 + ], + "credit_map": [ + 0.02239980176091194, + 0.025298859924077988, + 0.021929921582341194, + 0.02360519953072071, + 0.02088864892721176, + 0.027015672996640205, + 0.01618872582912445, + 0.011991861276328564, + 0.003195445518940687, + 0.008579954504966736, + 0.005688999313861132, + 0.007277588360011578, + 0.009092921391129494, + 0.0060370005667209625, + 0.010530021041631699, + 0.004570393357425928, + 0.007981126196682453, + 0.04666885733604431, + 0.006427375599741936, + 0.04598069190979004, + 0.009324455633759499, + 0.0055932109244167805, + 0.00839637778699398, + 0.00617868360131979, + 0.012220118194818497, + 0.010096507146954536, + 0.01519786287099123, + 0.00961532723158598, + 0.018360204994678497, + 0.014981559477746487, + 0.008040736429393291, + 0.009919795207679272, + 0.010995624586939812, + 0.023947760462760925, + 0.020834168419241905, + 0.013990242965519428, + 0.021728985011577606, + 0.023526083678007126, + 0.018761049956083298, + 0.017765076830983162, + 0.011730165220797062, + 0.007430770434439182, + 0.008062518201768398, + 0.03350954130291939, + 0.016964353621006012, + 0.03698800131678581, + 0.011670858599245548, + 0.006120095495134592, + 0.00821459200233221, + 0.013282164931297302, + 0.013140814378857613, + 0.01312955655157566, + 0.01642497256398201, + 0.012079963460564613, + 0.014737575314939022, + 0.009058221243321896, + 0.002106107072904706, + 0.006920517887920141, + 0.004104682244360447, + 0.004319461062550545, + 0.005991311743855476, + 0.004374073352664709, + 0.005381826777011156, + 0.003941521514207125, + 0.003944463096559048, + 0.005415975581854582, + 0.007218324113637209, + 0.007025422062724829, + 0.0110776387155056, + 0.00523070665076375, + 0.009663560427725315, + 0.01918809674680233, + 0.0038135098293423653, + 0.004404568113386631, + 0.0065268115140497684, + 0.0030705800745636225, + 0.008270157501101494, + 0.00564434053376317, + 0.010462569072842598, + 0.008536729961633682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.01608443818986416 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289990961551666 + ], + "opposite_logprob_tokens": [ + -1.0539990663528442 + ], + "logprob_reward": -0.4289990961551666, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289990961551666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1648850291967392, + 0.07514627277851105, + 0.06991183012723923, + 0.11050550639629364, + 0.0722527876496315, + 0.12363686412572861, + 0.05081305652856827, + 0.0255608968436718, + 0.26437142491340637, + 0.042916301637887955 + ], + "credit_map": [ + 0.01802896335721016, + 0.02969641238451004, + 0.018453728407621384, + 0.026253364980220795, + 0.020897574722766876, + 0.024777373299002647, + 0.01461818814277649, + 0.012159434147179127, + 0.005470163654536009, + 0.013593695126473904, + 0.0093582384288311, + 0.008727485314011574, + 0.009701017290353775, + 0.008790411986410618, + 0.012485862709581852, + 0.007019408978521824, + 0.005803941749036312, + 0.01567215286195278, + 0.004700102843344212, + 0.01418250985443592, + 0.006297826301306486, + 0.006776539608836174, + 0.00703353201970458, + 0.009445230476558208, + 0.012432333081960678, + 0.013449694961309433, + 0.015757806599140167, + 0.011419408954679966, + 0.02117556892335415, + 0.014045101590454578, + 0.010109415277838707, + 0.01211619284003973, + 0.004382571671158075, + 0.008575818501412868, + 0.011069698259234428, + 0.005750663112848997, + 0.00800353940576315, + 0.0136128980666399, + 0.01506650261580944, + 0.005791096948087215, + 0.007721332833170891, + 0.006544260308146477, + 0.00873643159866333, + 0.035146456211805344, + 0.0115779098123312, + 0.035916171967983246, + 0.010477293282747269, + 0.007517015561461449, + 0.004681359510868788, + 0.005982343107461929, + 0.004992043599486351, + 0.0058444286696612835, + 0.007702786475419998, + 0.007145773619413376, + 0.007984070107340813, + 0.0064802588894963264, + 0.0017405692487955093, + 0.003775663673877716, + 0.0028683620039373636, + 0.003771519288420677, + 0.0034721926786005497, + 0.003442415501922369, + 0.0031782961450517178, + 0.003311879700049758, + 0.01974884793162346, + 0.014558856375515461, + 0.01891712099313736, + 0.024409648030996323, + 0.03217016160488129, + 0.03204605355858803, + 0.02937120385468006, + 0.09314955770969391, + 0.003697220468893647, + 0.004130203742533922, + 0.0029080333188176155, + 0.0025584374088793993, + 0.00795838888734579, + 0.0038252423983067274, + 0.011176679283380508, + 0.006662100553512573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.4289990961551666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538178823888302 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538178823888302, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538178823888302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14641021192073822, + 0.1521580070257187, + 0.048448193818330765, + 0.11978674679994583, + 0.11463934928178787, + 0.15293951332569122, + 0.062092721462249756, + 0.0473485067486763, + 0.07726746052503586, + 0.0789092555642128 + ], + "credit_map": [ + 0.02153017185628414, + 0.01826777309179306, + 0.015047603286802769, + 0.0237601175904274, + 0.016331393271684647, + 0.02466101571917534, + 0.015839219093322754, + 0.010972926393151283, + 0.011412420310080051, + 0.026739219203591347, + 0.016285495832562447, + 0.019394898787140846, + 0.020437197759747505, + 0.017835017293691635, + 0.02303462289273739, + 0.017019137740135193, + 0.004738321527838707, + 0.006476403679698706, + 0.008139228448271751, + 0.007769090123474598, + 0.004765691701322794, + 0.005198820028454065, + 0.006235167849808931, + 0.005125471390783787, + 0.01257855910807848, + 0.011062465608119965, + 0.025663310661911964, + 0.009630030021071434, + 0.020168857648968697, + 0.021054962649941444, + 0.008537225425243378, + 0.011091341264545918, + 0.007394909393042326, + 0.020088495686650276, + 0.010500721633434296, + 0.011015601456165314, + 0.022760072723031044, + 0.020076986402273178, + 0.010633050464093685, + 0.012169509194791317, + 0.020887790247797966, + 0.010021290741860867, + 0.013017323799431324, + 0.035438284277915955, + 0.014267626218497753, + 0.03246811777353287, + 0.014902648515999317, + 0.011936438269913197, + 0.006092282943427563, + 0.007228055503219366, + 0.005496020428836346, + 0.007131650578230619, + 0.00972515344619751, + 0.008620900101959705, + 0.010099909268319607, + 0.007698747795075178, + 0.0030368955340236425, + 0.009986096993088722, + 0.0038663300219923258, + 0.006216473411768675, + 0.0076943314634263515, + 0.006185377016663551, + 0.005412600468844175, + 0.004950404167175293, + 0.005808440502732992, + 0.006500305607914925, + 0.011603530496358871, + 0.0056451126001775265, + 0.01655934378504753, + 0.006918576546013355, + 0.01075360644608736, + 0.013478545472025871, + 0.008127564564347267, + 0.008347144350409508, + 0.008991178125143051, + 0.005641057156026363, + 0.016003984957933426, + 0.0064516207203269005, + 0.016023587435483932, + 0.009323124773800373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.012538178823888302 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight tracking shot, shot from the side and slightly from an upward angle, closely follows the resolute face of a rickshaw puller. He wears a worn, dark cloth cap, his bronzed cheeks crisscrossed with deep wrinkles. Large beads of sweat trickle down his furrowed brow and clenched jawline, gleaming faintly in the warm sunset. He leans forward, his taut upper body clad only in a sweat-soaked, clinging gray sweatshirt. With every ounce of his strength, he pulls a yellow freight tricycle, its paint peeling in patches, revealing the rust beneath. The wheels creak heavily and laboriously as they bump against the uneven, time-worn bluestone slabs of an ancient stone arch bridge. In the background, a moss-covered stone railing and a hazy view of the bridge below loom. The entire picture is stained with a warm yellow hue of retro art, like a faded old photo. The soft golden light lengthens the shadows of people and cars, profoundly depicting an indomitable resilience of life and a profound sense of the times in the difficult life.", + "seed": 31992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_11/index_1332_seed_31992_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05855872873216868, + "vlm_logprob": -0.05855872873216868, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.05855872873216868, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047690290957689285 + ], + "opposite_logprob_tokens": [ + -3.1726903915405273 + ], + "logprob_reward": -0.047690290957689285, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.047690290957689285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33845293521881104, + 0.023271819576621056, + 0.0726175308227539, + 0.05632372573018074, + 0.1339167058467865, + 0.05480099469423294, + 0.03782837465405464, + 0.033506546169519424, + 0.17078594863414764, + 0.07849541306495667 + ], + "credit_map": [ + 0.030870646238327026, + 0.04702214524149895, + 0.02632572315633297, + 0.04914996027946472, + 0.0488571934401989, + 0.04868347942829132, + 0.06475289165973663, + 0.022790879011154175, + 0.003454147605225444, + 0.004369661677628756, + 0.0018990635871887207, + 0.0035962711554020643, + 0.0028742568101733923, + 0.0018271482549607754, + 0.002978038275614381, + 0.0022732317447662354, + 0.0066176182590425014, + 0.007234352640807629, + 0.009982757270336151, + 0.011133711785078049, + 0.017897242680191994, + 0.004987809807062149, + 0.008633391000330448, + 0.006130651105195284, + 0.004043556749820709, + 0.006920912768691778, + 0.005815598182380199, + 0.011776755563914776, + 0.007947318255901337, + 0.0064753321930766106, + 0.008145075291395187, + 0.005199179518967867, + 0.009894230403006077, + 0.008751808665692806, + 0.012055727653205395, + 0.025866122916340828, + 0.018442999571561813, + 0.04066774249076843, + 0.010997749865055084, + 0.007240329869091511, + 0.0034828493371605873, + 0.0037684787530452013, + 0.006100821308791637, + 0.00852730218321085, + 0.010515525937080383, + 0.007405281998217106, + 0.009757395833730698, + 0.005243339110165834, + 0.002598650986328721, + 0.003313241759315133, + 0.0031540202908217907, + 0.007398239336907864, + 0.010724351741373539, + 0.003078082576394081, + 0.004579165019094944, + 0.0029826227109879255, + 0.00244915415532887, + 0.0034758911933749914, + 0.0039912257343530655, + 0.004362898413091898, + 0.005841655656695366, + 0.003015895141288638, + 0.004761131480336189, + 0.005608694162219763, + 0.012114782817661762, + 0.02234608307480812, + 0.02325594425201416, + 0.021215368062257767, + 0.057875171303749084, + 0.012353288941085339, + 0.01607942022383213, + 0.0055458806455135345, + 0.00656169606372714, + 0.010230282321572304, + 0.011019949801266193, + 0.015503651462495327, + 0.020434774458408356, + 0.003828141838312149, + 0.007416850887238979, + 0.0035000641364604235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.047690290957689285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820802316069603 + ], + "opposite_logprob_tokens": [ + -4.018208026885986 + ], + "logprob_reward": -0.01820802316069603, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820802316069603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.126962348818779, + 0.11165565252304077, + 0.10724011808633804, + 0.02765924111008644, + 0.19551409780979156, + 0.027281545102596283, + 0.060315415263175964, + 0.08326739072799683, + 0.09735535085201263, + 0.1627487987279892 + ], + "credit_map": [ + 0.0161215178668499, + 0.015394347719848156, + 0.014637527987360954, + 0.014670446515083313, + 0.02828173153102398, + 0.014980081468820572, + 0.015591512434184551, + 0.007285183761268854, + 0.013999233953654766, + 0.018984248861670494, + 0.010966621339321136, + 0.015266607515513897, + 0.01672249846160412, + 0.011011885479092598, + 0.010672538541257381, + 0.014032024890184402, + 0.017025889828801155, + 0.01040321122854948, + 0.013057759962975979, + 0.014482573606073856, + 0.015833774581551552, + 0.010395132936537266, + 0.01440974697470665, + 0.011632036417722702, + 0.0016640728572383523, + 0.0026985283475369215, + 0.0027125030755996704, + 0.004192212596535683, + 0.003389843972399831, + 0.003963257186114788, + 0.005430157296359539, + 0.0036086675245314837, + 0.010251818224787712, + 0.009984022006392479, + 0.011190950870513916, + 0.044675882905721664, + 0.01680673286318779, + 0.07674941420555115, + 0.01748218759894371, + 0.008373111486434937, + 0.0027193494606763124, + 0.0029243456665426493, + 0.0029020097572356462, + 0.0036699422635138035, + 0.003786928951740265, + 0.003664940595626831, + 0.004902880638837814, + 0.0027111507952213287, + 0.004650422837585211, + 0.005888062994927168, + 0.006424557883292437, + 0.007937009446322918, + 0.007540426217019558, + 0.009542322717607021, + 0.011570727452635765, + 0.0067618852481245995, + 0.004107498563826084, + 0.007010017987340689, + 0.012257156893610954, + 0.01840914972126484, + 0.018282514065504074, + 0.004160498268902302, + 0.012493904680013657, + 0.006546658929437399, + 0.006763989105820656, + 0.015255738981068134, + 0.017577774822711945, + 0.012300283648073673, + 0.028186511248350143, + 0.007446856237947941, + 0.005233826115727425, + 0.0045903753489255905, + 0.010480684228241444, + 0.020221594721078873, + 0.03621033951640129, + 0.0260142982006073, + 0.03842291608452797, + 0.007030222099274397, + 0.018693948164582253, + 0.005674810614436865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.01820802316069603 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081107892096043 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081107892096043, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081107892096043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.367445707321167, + 0.036186277866363525, + 0.04595714434981346, + 0.05346654728055, + 0.12955494225025177, + 0.09344369173049927, + 0.08184514939785004, + 0.03227877616882324, + 0.04769333079457283, + 0.11212848126888275 + ], + "credit_map": [ + 0.02910284698009491, + 0.04775705188512802, + 0.07202626764774323, + 0.04783981665968895, + 0.047613415867090225, + 0.038025401532649994, + 0.06180064380168915, + 0.02328023500740528, + 0.005027241073548794, + 0.004773498512804508, + 0.002796849003061652, + 0.005240540020167828, + 0.004878796171396971, + 0.003686249954625964, + 0.003511544317007065, + 0.006271556951105595, + 0.005165210459381342, + 0.006200948264449835, + 0.006546960677951574, + 0.008099310100078583, + 0.007476156111806631, + 0.003099593333899975, + 0.004708334803581238, + 0.004660629201680422, + 0.0051773409359157085, + 0.006096934899687767, + 0.00496881315484643, + 0.009781580418348312, + 0.00683147506788373, + 0.006716789212077856, + 0.00867688748985529, + 0.0052167256362736225, + 0.010468468070030212, + 0.009855476208031178, + 0.014951611869037151, + 0.025082139298319817, + 0.014989642426371574, + 0.03431588038802147, + 0.013995594345033169, + 0.005896127317100763, + 0.006074571516364813, + 0.0070351059548556805, + 0.010181576944887638, + 0.014733687043190002, + 0.015364731661975384, + 0.013074825517833233, + 0.017489947378635406, + 0.009489252232015133, + 0.005493866745382547, + 0.007397704757750034, + 0.007104986812919378, + 0.014888863079249859, + 0.020648274570703506, + 0.00744987465441227, + 0.010654270648956299, + 0.008207317441701889, + 0.0026733321137726307, + 0.004403375089168549, + 0.003080661641433835, + 0.0042290412820875645, + 0.005812478717416525, + 0.002520062495023012, + 0.0059639643877744675, + 0.0035958625376224518, + 0.0059739090502262115, + 0.004550767131149769, + 0.007984142750501633, + 0.005307022947818041, + 0.007516381796449423, + 0.006360413506627083, + 0.004258699249476194, + 0.0057419962249696255, + 0.007458026055246592, + 0.006552909966558218, + 0.01263006217777729, + 0.027657324448227882, + 0.04002697765827179, + 0.0049139875918626785, + 0.007575590163469315, + 0.005313608329743147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.011081107892096043 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161974430084229 + ], + "opposite_logprob_tokens": [ + -1.7016197443008423 + ], + "logprob_reward": -0.20161974430084229, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161974430084229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19713708758354187, + 0.08571525663137436, + 0.057680342346429825, + 0.04481847211718559, + 0.165384441614151, + 0.057743314653635025, + 0.09081467241048813, + 0.05049557983875275, + 0.07172501087188721, + 0.17848582565784454 + ], + "credit_map": [ + 0.020579054951667786, + 0.036854129284620285, + 0.037820298224687576, + 0.02327416092157364, + 0.021865690127015114, + 0.024598538875579834, + 0.019516225904226303, + 0.012628992088139057, + 0.010391605086624622, + 0.014375816099345684, + 0.007935857400298119, + 0.01237359270453453, + 0.012490502558648586, + 0.008516289293766022, + 0.009401431307196617, + 0.01023016031831503, + 0.007893417030572891, + 0.005546603351831436, + 0.009321882389485836, + 0.008441075682640076, + 0.008154202252626419, + 0.0044878385961055756, + 0.006764248479157686, + 0.007071078289300203, + 0.004094674717634916, + 0.0040560876950621605, + 0.0038612124044448137, + 0.008366907015442848, + 0.006798936054110527, + 0.0048913597129285336, + 0.00926860049366951, + 0.003480694256722927, + 0.014642822556197643, + 0.012757129967212677, + 0.01787983812391758, + 0.03229813650250435, + 0.01753772236406803, + 0.047166354954242706, + 0.01571204885840416, + 0.00739039434120059, + 0.005794535856693983, + 0.006996176205575466, + 0.0044938200153410435, + 0.00607985258102417, + 0.011461419984698296, + 0.008283332921564579, + 0.007604480255395174, + 0.007029695436358452, + 0.006307001691311598, + 0.008414693176746368, + 0.009458367712795734, + 0.0128309465944767, + 0.01641787588596344, + 0.01207529567182064, + 0.014567735604941845, + 0.010742753744125366, + 0.004775672685354948, + 0.005326293874531984, + 0.006343471352010965, + 0.007557861972600222, + 0.006577019114047289, + 0.005165875889360905, + 0.007125146221369505, + 0.007624238729476929, + 0.00823705643415451, + 0.007221250794827938, + 0.010840461589396, + 0.008427668362855911, + 0.009159487672150135, + 0.009480840526521206, + 0.010102314874529839, + 0.008255934342741966, + 0.012323182076215744, + 0.01054292730987072, + 0.019146060571074486, + 0.04428252577781677, + 0.06552054733037949, + 0.007098163012415171, + 0.012236793525516987, + 0.007335628382861614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.20161974430084229 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a rickshaw puller with a resolute expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puller wearing a worn dark cloth cap and a sweat-soaked gray sweatshirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the puller leaning forward while pulling a yellow freight tricycle with peeling paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an ancient stone arch bridge with uneven bluestone slabs and a moss-covered railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from the side and slightly from an upward angle in a tight tracking shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419447734951973 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.01419447734951973, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419447734951973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1760862022638321, + 0.04415103793144226, + 0.05354394018650055, + 0.05300385504961014, + 0.16028518974781036, + 0.08837093412876129, + 0.09046671539545059, + 0.042631521821022034, + 0.08365421742200851, + 0.20780639350414276 + ], + "credit_map": [ + 0.01888887584209442, + 0.03199043124914169, + 0.012090652249753475, + 0.021897142753005028, + 0.028188686817884445, + 0.024711973965168, + 0.027373187243938446, + 0.010945260524749756, + 0.004486520309001207, + 0.006142959464341402, + 0.003964954987168312, + 0.006610470823943615, + 0.006278541870415211, + 0.005252884700894356, + 0.005564672872424126, + 0.005850035231560469, + 0.005411873571574688, + 0.007574392482638359, + 0.007039348129183054, + 0.007583840750157833, + 0.009894534945487976, + 0.003744747955352068, + 0.007300343364477158, + 0.004994863178580999, + 0.004524887539446354, + 0.004307535942643881, + 0.004037533421069384, + 0.007714755833148956, + 0.007602533791214228, + 0.007884902879595757, + 0.011806909926235676, + 0.005124793853610754, + 0.011315546929836273, + 0.0116762425750494, + 0.013658956624567509, + 0.03445868939161301, + 0.014811012893915176, + 0.05119696632027626, + 0.015743862837553024, + 0.007423911243677139, + 0.0074423858895897865, + 0.008114800788462162, + 0.009670431725680828, + 0.013319392688572407, + 0.014116302132606506, + 0.011485523544251919, + 0.01485129538923502, + 0.009370801970362663, + 0.006901429034769535, + 0.010582230053842068, + 0.008568533696234226, + 0.012953492812812328, + 0.015454781241714954, + 0.012014871463179588, + 0.013340027071535587, + 0.01065134909003973, + 0.004563568159937859, + 0.005911719519644976, + 0.003917634952813387, + 0.006861213129013777, + 0.007255474105477333, + 0.0038970063906162977, + 0.006822062656283379, + 0.003402841743081808, + 0.006480921525508165, + 0.012767858803272247, + 0.014702258631587029, + 0.009687911719083786, + 0.017358504235744476, + 0.008052184246480465, + 0.008233927190303802, + 0.006370648741722107, + 0.013102419674396515, + 0.01368901040405035, + 0.022990843281149864, + 0.04971877485513687, + 0.07274384051561356, + 0.009837070479989052, + 0.015839554369449615, + 0.009884873405098915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.01419447734951973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_12/index_tensor([5190, 5190])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_12/index_tensor([5190, 5190])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..83b9399fd4a70b9a84bbc4576df32d1406d4aa8d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_12/index_tensor([5190, 5190])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium shot captures the rear of a lone, classic black-and-white police cruiser as it slowly winds its way through the heart of a deserted city street at night. The vehicle's glossy paint, slick with rain, vividly reflects the silent, rhythmic pulse of red and blue emergency light strips, casting sweeping, ephemeral bursts of color across the wet, black asphalt. The camera then begins a smooth, steady crane motion, pulling upward and backward, gradually widening the frame. As the view rises, the street expands into a complex grid of empty streets, all shimmering in the cold, ambient glow and hazy atmosphere of distant streetlights. The shot culminates in a breathtaking high-angle, extreme wide shot, revealing a vast, desolate metropolis composed of towering, dark skyscrapers whose windows are barely lit, creating a powerful sense of scale and profound emptiness. The lone police cruiser is now a small, almost insignificant vehicle, its flashing lights the sole beacon of activity in the vast, silent cityscape.", + "seed": 11196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_12/index_5190_seed_11196_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0840082566253841, + "vlm_logprob": -0.0840082566253841, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.0840082566253841, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3244094252586365 + ], + "opposite_logprob_tokens": [ + -1.3244094848632812 + ], + "logprob_reward": -0.3244094252586365, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3244094252586365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.191039577126503, + 0.09867201000452042, + 0.12462154030799866, + 0.152930349111557, + 0.08359640836715698, + 0.039171647280454636, + 0.12925291061401367, + 0.04063451290130615, + 0.055294983088970184, + 0.0847860798239708 + ], + "credit_map": [ + 0.01803809590637684, + 0.019883189350366592, + 0.019663257524371147, + 0.0324353352189064, + 0.02417692355811596, + 0.022569147869944572, + 0.024167055264115334, + 0.030106570571660995, + 0.012284504249691963, + 0.0074660577811300755, + 0.009716193191707134, + 0.008380910381674767, + 0.01247600931674242, + 0.017274705693125725, + 0.017607487738132477, + 0.013466141186654568, + 0.017430854961276054, + 0.009617750532925129, + 0.013588509522378445, + 0.02733544073998928, + 0.010557885281741619, + 0.011328591033816338, + 0.017201125621795654, + 0.017561381682753563, + 0.023924002423882484, + 0.01380197610706091, + 0.03598652780056, + 0.014843424782156944, + 0.009181837551295757, + 0.020229725167155266, + 0.01006337720900774, + 0.024899471551179886, + 0.009183942340314388, + 0.008455803617835045, + 0.009663092903792858, + 0.015782928094267845, + 0.006904329173266888, + 0.009400932118296623, + 0.010162439197301865, + 0.014042936265468597, + 0.007375913206487894, + 0.003930927254259586, + 0.004791334271430969, + 0.0040773432701826096, + 0.00372096779756248, + 0.006949497386813164, + 0.0034722965210676193, + 0.004853361286222935, + 0.015836264938116074, + 0.013744083233177662, + 0.019851945340633392, + 0.012667794711887836, + 0.012954947538673878, + 0.018925676122307777, + 0.014161146245896816, + 0.02111104317009449, + 0.005179836414754391, + 0.005094275809824467, + 0.00704279774799943, + 0.004770540166646242, + 0.004103127401322126, + 0.003995638806372881, + 0.004402250982820988, + 0.00604604184627533, + 0.0094230892136693, + 0.006420166697353125, + 0.006554820109158754, + 0.005492334719747305, + 0.0061316778883337975, + 0.005895947106182575, + 0.004262958653271198, + 0.011113988235592842, + 0.01007667277008295, + 0.011072291061282158, + 0.011239467188715935, + 0.010135826654732227, + 0.007765904534608126, + 0.015159683302044868, + 0.007090527098625898, + 0.012245704419910908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.3244094252586365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014233849011361599 + ], + "opposite_logprob_tokens": [ + -4.2642340660095215 + ], + "logprob_reward": -0.014233849011361599, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014233849011361599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09749328345060349, + 0.07210778445005417, + 0.061449941247701645, + 0.05513228103518486, + 0.03734167665243149, + 0.10128384083509445, + 0.36329612135887146, + 0.07733525335788727, + 0.027062566950917244, + 0.10749729722738266 + ], + "credit_map": [ + 0.008933326229453087, + 0.007409735582768917, + 0.0067448397167027, + 0.010845665819942951, + 0.012290621176362038, + 0.015282221138477325, + 0.013758220709860325, + 0.022228652611374855, + 0.006081217434257269, + 0.006568605545908213, + 0.006146921776235104, + 0.006395546719431877, + 0.006106801331043243, + 0.013600530102849007, + 0.009993111714720726, + 0.017215045168995857, + 0.0084937559440732, + 0.005069764796644449, + 0.01099305972456932, + 0.00664104986935854, + 0.005911579355597496, + 0.01131310686469078, + 0.00569967832416296, + 0.0073279389180243015, + 0.0063546039164066315, + 0.0036709089763462543, + 0.008314687758684158, + 0.008623399771749973, + 0.005716255400329828, + 0.008227596990764141, + 0.006168841850012541, + 0.008055990561842918, + 0.005772823467850685, + 0.003391363425180316, + 0.005882124416530132, + 0.0032435725443065166, + 0.003118876600638032, + 0.006525950040668249, + 0.0035406253300607204, + 0.005866343155503273, + 0.016714585945010185, + 0.00917638186365366, + 0.012690137140452862, + 0.009854753501713276, + 0.011034588329494, + 0.013732586055994034, + 0.01179318968206644, + 0.01628761552274227, + 0.063503198325634, + 0.024899637326598167, + 0.0659257024526596, + 0.022230111062526703, + 0.021252745762467384, + 0.06721525639295578, + 0.015480230562388897, + 0.08278923481702805, + 0.018560077995061874, + 0.007347363978624344, + 0.013434561900794506, + 0.006047110538929701, + 0.005988996475934982, + 0.009838801808655262, + 0.005635200999677181, + 0.010483138263225555, + 0.003342234995216131, + 0.002023812383413315, + 0.0037439013831317425, + 0.0035297363065183163, + 0.0028923519421368837, + 0.004484118428081274, + 0.0027590924873948097, + 0.004287318326532841, + 0.010947193950414658, + 0.014084902592003345, + 0.025428801774978638, + 0.01530732586979866, + 0.008680857717990875, + 0.008704327046871185, + 0.007814254611730576, + 0.016529630869627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.014233849011361599 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640998784452677 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640998784452677, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640998784452677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12819276750087738, + 0.05860154703259468, + 0.07452165335416794, + 0.15347230434417725, + 0.09628234058618546, + 0.07563771307468414, + 0.07679915428161621, + 0.07854083925485611, + 0.14716856181621552, + 0.11078320443630219 + ], + "credit_map": [ + 0.015255517326295376, + 0.013460787013173103, + 0.01306977216154337, + 0.014511555433273315, + 0.01849990524351597, + 0.02045448124408722, + 0.013808929361402988, + 0.019131818786263466, + 0.006648661103099585, + 0.00433924188837409, + 0.007030670531094074, + 0.006765548139810562, + 0.00596639234572649, + 0.0101133668795228, + 0.007975896820425987, + 0.009761768393218517, + 0.01434400212019682, + 0.005549043882638216, + 0.012477867305278778, + 0.00750349136069417, + 0.006171096581965685, + 0.008442975580692291, + 0.007827101275324821, + 0.01220607478171587, + 0.01746290735900402, + 0.014478025957942009, + 0.03454587608575821, + 0.0189539585262537, + 0.00986512377858162, + 0.018346518278121948, + 0.01713736727833748, + 0.022682521492242813, + 0.015850648283958435, + 0.012644839473068714, + 0.013994098640978336, + 0.009680638089776039, + 0.008113379590213299, + 0.014608621597290039, + 0.007625320460647345, + 0.013764792121946812, + 0.012286413460969925, + 0.006994965020567179, + 0.010136784985661507, + 0.005257968790829182, + 0.008579373359680176, + 0.014888045378029346, + 0.006205827463418245, + 0.011288336478173733, + 0.010967034846544266, + 0.009477104060351849, + 0.014656280167400837, + 0.006508577615022659, + 0.005634133704006672, + 0.011113418266177177, + 0.006578410044312477, + 0.011864199303090572, + 0.009475694969296455, + 0.010729534551501274, + 0.01121184416115284, + 0.009738739579916, + 0.009775150567293167, + 0.01007500384002924, + 0.007843704894185066, + 0.009691163897514343, + 0.022888146340847015, + 0.014856851659715176, + 0.017766639590263367, + 0.013754624873399734, + 0.019787773489952087, + 0.021376926451921463, + 0.010680009610950947, + 0.026057595387101173, + 0.013085025362670422, + 0.011677277274429798, + 0.018278319388628006, + 0.012827729806303978, + 0.010818559676408768, + 0.022282909601926804, + 0.006504348944872618, + 0.01530903484672308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.004640998784452677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370628505945206 + ], + "opposite_logprob_tokens": [ + -3.4087061882019043 + ], + "logprob_reward": -0.03370628505945206, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370628505945206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07823217660188675, + 0.08185075223445892, + 0.07621549814939499, + 0.10952641814947128, + 0.10721970349550247, + 0.0690273568034172, + 0.1711292266845703, + 0.19415771961212158, + 0.09267628937959671, + 0.0199647955596447 + ], + "credit_map": [ + 0.007859227247536182, + 0.008038063533604145, + 0.006411922164261341, + 0.0069025722332298756, + 0.009425611235201359, + 0.014309252612292767, + 0.009085503406822681, + 0.016200032085180283, + 0.007191319018602371, + 0.009518721140921116, + 0.008438004180788994, + 0.010062175802886486, + 0.009224175475537777, + 0.012533320114016533, + 0.011637423187494278, + 0.013245617039501667, + 0.005990958772599697, + 0.00724410405382514, + 0.012210514396429062, + 0.014660526067018509, + 0.007059914059937, + 0.009698716923594475, + 0.009065951220691204, + 0.010284814983606339, + 0.008646506816148758, + 0.00770560884848237, + 0.01828029379248619, + 0.02393534407019615, + 0.008590612560510635, + 0.009781351312994957, + 0.017470356076955795, + 0.015116357244551182, + 0.013526907190680504, + 0.012299318797886372, + 0.015632372349500656, + 0.012660346925258636, + 0.00890535768121481, + 0.014898334629833698, + 0.013366106897592545, + 0.015930969268083572, + 0.011786701157689095, + 0.006667668465524912, + 0.008875686675310135, + 0.007190176751464605, + 0.0072606285102665424, + 0.01097640860825777, + 0.008603219874203205, + 0.007666868157684803, + 0.02799152210354805, + 0.015057604759931564, + 0.031526703387498856, + 0.009688559919595718, + 0.011124173179268837, + 0.023100852966308594, + 0.013360190205276012, + 0.03927963599562645, + 0.034842923283576965, + 0.01611795462667942, + 0.03619811683893204, + 0.019659778103232384, + 0.014486104249954224, + 0.025576701387763023, + 0.021173706278204918, + 0.026102447882294655, + 0.010534500703215599, + 0.006521729286760092, + 0.009348654188215733, + 0.010155047290027142, + 0.010978794656693935, + 0.008970593102276325, + 0.023288818076252937, + 0.012878154404461384, + 0.0021096598356962204, + 0.0019671667832881212, + 0.003982509486377239, + 0.001796685392037034, + 0.0019474466098472476, + 0.0027099521830677986, + 0.002784744603559375, + 0.0026666312478482723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.03370628505945206 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043050725013017654 + ], + "opposite_logprob_tokens": [ + -3.168050765991211 + ], + "logprob_reward": -0.043050725013017654, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043050725013017654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1478469967842102, + 0.07974118739366531, + 0.07405537366867065, + 0.07973762601613998, + 0.11198469996452332, + 0.15113534033298492, + 0.07481549680233002, + 0.07786592841148376, + 0.08663396537303925, + 0.1161833256483078 + ], + "credit_map": [ + 0.020773908123373985, + 0.015478038229048252, + 0.015481537207961082, + 0.019026704132556915, + 0.015973910689353943, + 0.018523365259170532, + 0.01688552275300026, + 0.0257040336728096, + 0.009719646535813808, + 0.007396094035357237, + 0.007883891463279724, + 0.007039617281407118, + 0.010411915369331837, + 0.017041688784956932, + 0.009876654483377934, + 0.010371685959398746, + 0.011147553101181984, + 0.005078185815364122, + 0.01030407939106226, + 0.009272575378417969, + 0.007493053562939167, + 0.011667381040751934, + 0.0076066586188972, + 0.011485893279314041, + 0.008191661909222603, + 0.005589015316218138, + 0.017509376630187035, + 0.013011670671403408, + 0.007181861903518438, + 0.006992368493229151, + 0.009904976934194565, + 0.011356698349118233, + 0.018153121694922447, + 0.01389367412775755, + 0.01754377968609333, + 0.010086923837661743, + 0.010118026286363602, + 0.015887459740042686, + 0.011990071274340153, + 0.014311647973954678, + 0.027372021228075027, + 0.013981364667415619, + 0.029404330998659134, + 0.010889029130339622, + 0.011995555832982063, + 0.0274627935141325, + 0.00884008314460516, + 0.021190166473388672, + 0.014889155514538288, + 0.006602659821510315, + 0.012735892087221146, + 0.005573262460529804, + 0.005716557148844004, + 0.00934530608355999, + 0.008560027927160263, + 0.011392639018595219, + 0.009299004450440407, + 0.009618965908885002, + 0.012196865864098072, + 0.008964858017861843, + 0.00835228431969881, + 0.008897846564650536, + 0.009724712930619717, + 0.010811393149197102, + 0.010313427075743675, + 0.0082858856767416, + 0.009624666534364223, + 0.009384389035403728, + 0.008646761998534203, + 0.018891235813498497, + 0.00880774762481451, + 0.01267986185848713, + 0.013571321964263916, + 0.013283506035804749, + 0.020739471539855003, + 0.015347071923315525, + 0.00842214748263359, + 0.0187701266258955, + 0.009584412910044193, + 0.01646527275443077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.043050725013017654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot captures the rear of a lone, classic black-and-white police cruiser as it slowly winds its way through the heart of a deserted city street at night. The vehicle's glossy paint, slick with rain, vividly reflects the silent, rhythmic pulse of red and blue emergency light strips, casting sweeping, ephemeral bursts of color across the wet, black asphalt. The camera then begins a smooth, steady crane motion, pulling upward and backward, gradually widening the frame. As the view rises, the street expands into a complex grid of empty streets, all shimmering in the cold, ambient glow and hazy atmosphere of distant streetlights. The shot culminates in a breathtaking high-angle, extreme wide shot, revealing a vast, desolate metropolis composed of towering, dark skyscrapers whose windows are barely lit, creating a powerful sense of scale and profound emptiness. The lone police cruiser is now a small, almost insignificant vehicle, its flashing lights the sole beacon of activity in the vast, silent cityscape.", + "seed": 11197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_12/index_5190_seed_11197_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03145325230434537, + "vlm_logprob": -0.03145325230434537, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03145325230434537, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1027781292796135 + ], + "opposite_logprob_tokens": [ + -2.477778196334839 + ], + "logprob_reward": -0.1027781292796135, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1027781292796135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39028093218803406, + 0.09513873606920242, + 0.10985855013132095, + 0.08433020114898682, + 0.054813940078020096, + 0.04560818895697594, + 0.06128079444169998, + 0.04929962754249573, + 0.05010047182440758, + 0.05928851664066315 + ], + "credit_map": [ + 0.0272339079529047, + 0.01991840824484825, + 0.026146933436393738, + 0.029386110603809357, + 0.025162894278764725, + 0.1309128999710083, + 0.029075833037495613, + 0.10244393348693848, + 0.01805994100868702, + 0.007994601503014565, + 0.009025610983371735, + 0.010799149982631207, + 0.00950315035879612, + 0.013803429901599884, + 0.012079523876309395, + 0.013873321935534477, + 0.0170171819627285, + 0.007954740896821022, + 0.02595433034002781, + 0.012288646772503853, + 0.006828878074884415, + 0.011855123564600945, + 0.012025363743305206, + 0.015934288501739502, + 0.008255619555711746, + 0.007772644050419331, + 0.010650835931301117, + 0.010182280093431473, + 0.01233791932463646, + 0.009974264539778233, + 0.012756743468344212, + 0.012399897910654545, + 0.006241436116397381, + 0.008007739670574665, + 0.007972638122737408, + 0.006902002263814211, + 0.00623429985716939, + 0.00726958317682147, + 0.006238230038434267, + 0.0059480140917003155, + 0.008552351966500282, + 0.005433752667158842, + 0.006823801435530186, + 0.006544848904013634, + 0.0037541596684604883, + 0.007459087762981653, + 0.003289161017164588, + 0.003751022508367896, + 0.007113127503544092, + 0.008687863126397133, + 0.0081873694434762, + 0.008033408783376217, + 0.008844377472996712, + 0.008245481178164482, + 0.005028772633522749, + 0.007140397094190121, + 0.004739564843475819, + 0.0070984698832035065, + 0.005500580184161663, + 0.005849514156579971, + 0.007266763132065535, + 0.007793285883963108, + 0.005202935542911291, + 0.005848515313118696, + 0.005689758807420731, + 0.009248830378055573, + 0.008040067739784718, + 0.0036870723124593496, + 0.008318720385432243, + 0.005034356843680143, + 0.005050007253885269, + 0.005031661596149206, + 0.004972034599632025, + 0.0067735714837908745, + 0.004700202960520983, + 0.007295508403331041, + 0.010541579686105251, + 0.005957718472927809, + 0.007244102656841278, + 0.011803800240159035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.1027781292796135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005976070184260607 + ], + "opposite_logprob_tokens": [ + -5.13097620010376 + ], + "logprob_reward": -0.005976070184260607, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005976070184260607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19255861639976501, + 0.06031624600291252, + 0.08446946740150452, + 0.1677705943584442, + 0.04700847715139389, + 0.058254364877939224, + 0.09976305067539215, + 0.05938895046710968, + 0.13380061089992523, + 0.09666962921619415 + ], + "credit_map": [ + 0.017877114936709404, + 0.010214606299996376, + 0.0191337950527668, + 0.02110520377755165, + 0.011666679754853249, + 0.04683739319443703, + 0.015921643003821373, + 0.049802180379629135, + 0.007938711903989315, + 0.006507753860205412, + 0.0059420084580779076, + 0.0057777101173996925, + 0.009777185507118702, + 0.008617534302175045, + 0.008999878540635109, + 0.006755460053682327, + 0.014578643254935741, + 0.007045543286949396, + 0.013926961459219456, + 0.010524451732635498, + 0.0070042842999100685, + 0.009531174786388874, + 0.009889107197523117, + 0.011969300918281078, + 0.02700508013367653, + 0.021164076402783394, + 0.02801368571817875, + 0.01773723214864731, + 0.008531051687896252, + 0.022042104974389076, + 0.009340344928205013, + 0.033937014639377594, + 0.006770112551748753, + 0.008134354837238789, + 0.006320904474705458, + 0.003976603504270315, + 0.004616658203303814, + 0.0067842579446733, + 0.004025781527161598, + 0.006379805039614439, + 0.007193720433861017, + 0.007312054745852947, + 0.008427359163761139, + 0.007152458652853966, + 0.005691729951649904, + 0.00966575462371111, + 0.005742410663515329, + 0.007068878039717674, + 0.016918661072850227, + 0.01151332352310419, + 0.013920296914875507, + 0.010953713208436966, + 0.011447347700595856, + 0.016521502286195755, + 0.007061580661684275, + 0.011426620185375214, + 0.006363990250974894, + 0.00821198895573616, + 0.008194996975362301, + 0.005849036388099194, + 0.009235158562660217, + 0.009020913392305374, + 0.006084871012717485, + 0.006427997723221779, + 0.0190865658223629, + 0.019953377544879913, + 0.018549522385001183, + 0.01767508126795292, + 0.01945110410451889, + 0.01849425956606865, + 0.009393136017024517, + 0.011197556741535664, + 0.008051451295614243, + 0.010985022410750389, + 0.010653330013155937, + 0.01169542782008648, + 0.02165648527443409, + 0.008483177050948143, + 0.014121773652732372, + 0.011022964492440224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.005976070184260607 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464111752808094 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.00464111752808094, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464111752808094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12742720544338226, + 0.05347829312086105, + 0.07329348474740982, + 0.08572764694690704, + 0.08566164970397949, + 0.10601956397294998, + 0.20779700577259064, + 0.04185975715517998, + 0.1402546912431717, + 0.07848074287176132 + ], + "credit_map": [ + 0.012168105691671371, + 0.006791179068386555, + 0.011765320785343647, + 0.01135086826980114, + 0.006513253785669804, + 0.03728588670492172, + 0.008746582083404064, + 0.03280601277947426, + 0.011073232628405094, + 0.004087063018232584, + 0.004440690390765667, + 0.005700198467820883, + 0.004793612286448479, + 0.008834377862513065, + 0.00661461753770709, + 0.00793449766933918, + 0.015715617686510086, + 0.008506270125508308, + 0.012824339792132378, + 0.006316679529845715, + 0.006218573544174433, + 0.007800500839948654, + 0.006618625018745661, + 0.009292876347899437, + 0.008088130503892899, + 0.009102839045226574, + 0.013173340819776058, + 0.01022737380117178, + 0.010360168293118477, + 0.009184914641082287, + 0.009885030798614025, + 0.01570584438741207, + 0.010137607343494892, + 0.013857894577085972, + 0.01197715662419796, + 0.007859435863792896, + 0.011198345571756363, + 0.012015082873404026, + 0.008935605175793171, + 0.009680521674454212, + 0.014841428026556969, + 0.015435362234711647, + 0.01614374853670597, + 0.01029010210186243, + 0.010246680118143559, + 0.016020627692341805, + 0.010351893492043018, + 0.012689710594713688, + 0.02828787826001644, + 0.031963471323251724, + 0.02925792708992958, + 0.02513214200735092, + 0.026170853525400162, + 0.03302557393908501, + 0.01243412122130394, + 0.02152504213154316, + 0.005508014466613531, + 0.007406436372548342, + 0.005858373828232288, + 0.0035329649690538645, + 0.006121522281318903, + 0.00567190907895565, + 0.003942163195461035, + 0.0038183745928108692, + 0.01735035888850689, + 0.029544688761234283, + 0.015927383676171303, + 0.013425934128463268, + 0.027324073016643524, + 0.015763726085424423, + 0.009593440219759941, + 0.011325093917548656, + 0.006714894436299801, + 0.009024934843182564, + 0.013761309906840324, + 0.012254292145371437, + 0.014497769996523857, + 0.007080956362187862, + 0.006778270937502384, + 0.008368308655917645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.00464111752808094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329576015472412 + ], + "opposite_logprob_tokens": [ + -3.7732958793640137 + ], + "logprob_reward": -0.02329576015472412, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329576015472412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10379109531641006, + 0.05581807345151901, + 0.07481097429990768, + 0.07767224311828613, + 0.0637703612446785, + 0.0809357613325119, + 0.18649981915950775, + 0.09504648298025131, + 0.16834788024425507, + 0.09330729395151138 + ], + "credit_map": [ + 0.007493754383176565, + 0.0067955441772937775, + 0.012144835665822029, + 0.011524558998644352, + 0.007376149296760559, + 0.023360570892691612, + 0.00919724628329277, + 0.025898437947034836, + 0.009767517447471619, + 0.006511213723570108, + 0.0046488880179822445, + 0.006488979794085026, + 0.005681078881025314, + 0.009073508903384209, + 0.0063985236920416355, + 0.007248366251587868, + 0.011364051140844822, + 0.00769173726439476, + 0.01847367361187935, + 0.008380739949643612, + 0.006034431979060173, + 0.006367706228047609, + 0.006575682200491428, + 0.009922953322529793, + 0.009321010671555996, + 0.006251334212720394, + 0.011143183335661888, + 0.012858699075877666, + 0.0083827069029212, + 0.008749953471124172, + 0.009955035522580147, + 0.011010324582457542, + 0.006173073314130306, + 0.006549519021064043, + 0.008841482922434807, + 0.006336372345685959, + 0.010261771269142628, + 0.009126417338848114, + 0.008795290254056454, + 0.007686437573283911, + 0.011278076097369194, + 0.007324704434722662, + 0.013142850250005722, + 0.008638612926006317, + 0.008915891870856285, + 0.013081923127174377, + 0.00854242593050003, + 0.010011278092861176, + 0.02909942902624607, + 0.024666214361786842, + 0.02643534727394581, + 0.020444214344024658, + 0.022699736058712006, + 0.030293850228190422, + 0.013208407908678055, + 0.019652612507343292, + 0.013992645777761936, + 0.016430210322141647, + 0.013868454843759537, + 0.010717935860157013, + 0.012546597048640251, + 0.013602693565189838, + 0.007658754475414753, + 0.006229190621525049, + 0.014145275577902794, + 0.04953606054186821, + 0.012656491249799728, + 0.008035668171942234, + 0.04496683180332184, + 0.01816292107105255, + 0.009866878390312195, + 0.010977747850120068, + 0.005921491421759129, + 0.010561841540038586, + 0.008584822528064251, + 0.0110696442425251, + 0.01183821726590395, + 0.008614443242549896, + 0.01220175065100193, + 0.02451508119702339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.02329576015472412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575184375047684 + ], + "opposite_logprob_tokens": [ + -3.8955752849578857 + ], + "logprob_reward": -0.020575184375047684, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575184375047684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08853647112846375, + 0.03504360467195511, + 0.08488393574953079, + 0.04607285559177399, + 0.06031275540590286, + 0.0616595596075058, + 0.3061128854751587, + 0.06780771166086197, + 0.1084582507610321, + 0.1411120444536209 + ], + "credit_map": [ + 0.006108996458351612, + 0.005530949216336012, + 0.013389063067734241, + 0.01122484914958477, + 0.007078980095684528, + 0.01843355968594551, + 0.007716678082942963, + 0.01905338279902935, + 0.006086179055273533, + 0.0030887448228895664, + 0.003076071385294199, + 0.003664920572191477, + 0.005563762970268726, + 0.005184173118323088, + 0.003969579469412565, + 0.004410167224705219, + 0.018913855776190758, + 0.007380730472505093, + 0.01572423428297043, + 0.007249234244227409, + 0.006758001167327166, + 0.008716046810150146, + 0.008113403804600239, + 0.01202841755002737, + 0.005288858897984028, + 0.0055280509404838085, + 0.005647026468068361, + 0.0070694610476493835, + 0.00537482462823391, + 0.005518125370144844, + 0.006039244122803211, + 0.005607254337519407, + 0.0067594884894788265, + 0.007791350595653057, + 0.009038316085934639, + 0.005654624197632074, + 0.007941266521811485, + 0.010711144655942917, + 0.006593648809939623, + 0.005822909064590931, + 0.010817302390933037, + 0.009908010251820087, + 0.009491871111094952, + 0.006769655272364616, + 0.005507927853614092, + 0.008876430802047253, + 0.005198347382247448, + 0.005090011283755302, + 0.04032037407159805, + 0.048125676810741425, + 0.04161753132939339, + 0.03756020590662956, + 0.0370895154774189, + 0.04730313643813133, + 0.021124497056007385, + 0.03297191113233566, + 0.008581027388572693, + 0.011948073282837868, + 0.011369067244231701, + 0.007952069863677025, + 0.0079838577657938, + 0.009069668129086494, + 0.005782826337963343, + 0.005121113732457161, + 0.00839532632380724, + 0.03200892359018326, + 0.008261625654995441, + 0.005268505774438381, + 0.029283151030540466, + 0.011544058099389076, + 0.006108672823756933, + 0.007587973959743977, + 0.010096544399857521, + 0.015336764976382256, + 0.02103954367339611, + 0.016379516571760178, + 0.01862047053873539, + 0.015055209398269653, + 0.017966218292713165, + 0.02661777287721634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.020575184375047684 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot captures the rear of a lone, classic black-and-white police cruiser as it slowly winds its way through the heart of a deserted city street at night. The vehicle's glossy paint, slick with rain, vividly reflects the silent, rhythmic pulse of red and blue emergency light strips, casting sweeping, ephemeral bursts of color across the wet, black asphalt. The camera then begins a smooth, steady crane motion, pulling upward and backward, gradually widening the frame. As the view rises, the street expands into a complex grid of empty streets, all shimmering in the cold, ambient glow and hazy atmosphere of distant streetlights. The shot culminates in a breathtaking high-angle, extreme wide shot, revealing a vast, desolate metropolis composed of towering, dark skyscrapers whose windows are barely lit, creating a powerful sense of scale and profound emptiness. The lone police cruiser is now a small, almost insignificant vehicle, its flashing lights the sole beacon of activity in the vast, silent cityscape.", + "seed": 11198, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_12/index_5190_seed_11198_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04641271438449621, + "vlm_logprob": -0.04641271438449621, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.04641271438449621, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18542475998401642 + ], + "opposite_logprob_tokens": [ + -1.8104248046875 + ], + "logprob_reward": -0.18542475998401642, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18542475998401642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5686703324317932, + 0.041518814861774445, + 0.08216319978237152, + 0.1049548014998436, + 0.06250372529029846, + 0.018913445994257927, + 0.04335338994860649, + 0.022054806351661682, + 0.032857056707143784, + 0.023010466247797012 + ], + "credit_map": [ + 0.0575723759829998, + 0.013301919214427471, + 0.19487673044204712, + 0.052072398364543915, + 0.01612820476293564, + 0.031551994383335114, + 0.03570391237735748, + 0.16746275126934052, + 0.005254330579191446, + 0.006071195006370544, + 0.0050105927512049675, + 0.0038229224737733603, + 0.0046828570775687695, + 0.006630501244217157, + 0.005123754497617483, + 0.004922656342387199, + 0.009262693114578724, + 0.011600183323025703, + 0.006464528385549784, + 0.009375396184623241, + 0.009465305134654045, + 0.01058713998645544, + 0.0155657809227705, + 0.009842167608439922, + 0.011050495319068432, + 0.0145866172388196, + 0.0077784089371562, + 0.013349361717700958, + 0.011027800850570202, + 0.014510786160826683, + 0.019057858735322952, + 0.013593479059636593, + 0.005360705312341452, + 0.0037923811469227076, + 0.01218369323760271, + 0.004628046881407499, + 0.008130805566906929, + 0.006195716559886932, + 0.005038812290877104, + 0.017173565924167633, + 0.002970022615045309, + 0.0022932286374270916, + 0.0026835629250854254, + 0.00273695751093328, + 0.002153804525732994, + 0.002341692568734288, + 0.0019275861559435725, + 0.0018065907061100006, + 0.00346574024297297, + 0.004636249504983425, + 0.01036026794463396, + 0.006960304453969002, + 0.0039832256734371185, + 0.0046016257256269455, + 0.004955087788403034, + 0.004390893038362265, + 0.0017465705750510097, + 0.0025336870457977057, + 0.002655767137184739, + 0.0032869158312678337, + 0.0024138169828802347, + 0.0031121927313506603, + 0.0027801564428955317, + 0.003525698324665427, + 0.0018810357432812452, + 0.0025546113029122353, + 0.003324982011690736, + 0.0020027540158480406, + 0.0016072908183559775, + 0.009760831482708454, + 0.002244646195322275, + 0.009480905719101429, + 0.002280331216752529, + 0.003127617295831442, + 0.0034389093052595854, + 0.004420924466103315, + 0.001196866505779326, + 0.003472014330327511, + 0.0019090382847934961, + 0.0031647668220102787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.18542475998401642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005973581690341234 + ], + "opposite_logprob_tokens": [ + -5.130973815917969 + ], + "logprob_reward": -0.005973581690341234, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005973581690341234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22463874518871307, + 0.056747786700725555, + 0.04556744918227196, + 0.1024533212184906, + 0.08728276938199997, + 0.02667800709605217, + 0.045581016689538956, + 0.3240177631378174, + 0.041546810418367386, + 0.04548622667789459 + ], + "credit_map": [ + 0.02242666855454445, + 0.005782817956060171, + 0.07706741243600845, + 0.02285841293632984, + 0.006371004041284323, + 0.012955356389284134, + 0.015069945715367794, + 0.0621071495115757, + 0.0072162700816988945, + 0.006228617858141661, + 0.005333583801984787, + 0.009318486787378788, + 0.0065881917253136635, + 0.006356146652251482, + 0.005143024958670139, + 0.010563469491899014, + 0.005284303333610296, + 0.005979230161756277, + 0.004438790492713451, + 0.006902207154780626, + 0.003920939285308123, + 0.004532679915428162, + 0.008461883291602135, + 0.006047412287443876, + 0.013656772673130035, + 0.017540311440825462, + 0.00846483651548624, + 0.009784596972167492, + 0.010461162775754929, + 0.010714351199567318, + 0.014924543909728527, + 0.016906747594475746, + 0.008570193313062191, + 0.004415958654135466, + 0.01760793849825859, + 0.005693470593541861, + 0.00798523984849453, + 0.009804835543036461, + 0.005865546874701977, + 0.027339590713381767, + 0.002680395031347871, + 0.0029763192869722843, + 0.0038906491827219725, + 0.003683678572997451, + 0.0033163477201014757, + 0.0028849344234913588, + 0.0042179482989013195, + 0.00302773411385715, + 0.003377860877662897, + 0.005467183887958527, + 0.007834237068891525, + 0.007961993105709553, + 0.004537115804851055, + 0.005777537357062101, + 0.006180807016789913, + 0.00444428576156497, + 0.012083452194929123, + 0.015364741906523705, + 0.014553862623870373, + 0.021210649982094765, + 0.008779872208833694, + 0.11208288371562958, + 0.013127891346812248, + 0.12681441009044647, + 0.0028817204292863607, + 0.004812867846339941, + 0.004158870782703161, + 0.005097089800983667, + 0.00245287106372416, + 0.008382567204535007, + 0.0020562256686389446, + 0.011704595759510994, + 0.0024784149136394262, + 0.004185460973531008, + 0.0037351695355027914, + 0.005394722800701857, + 0.002823008457198739, + 0.01073608361184597, + 0.0037782564759254456, + 0.012355105951428413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.005973581690341234 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003192449454218149 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.003192449454218149, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003192449454218149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19612908363342285, + 0.11525403708219528, + 0.07528447359800339, + 0.08629678934812546, + 0.07073716074228287, + 0.09631986916065216, + 0.04173706844449043, + 0.12569719552993774, + 0.0441751629114151, + 0.14836914837360382 + ], + "credit_map": [ + 0.02138749696314335, + 0.0065343487076461315, + 0.06515999138355255, + 0.025128787383437157, + 0.006109720095992088, + 0.00930606946349144, + 0.014575153589248657, + 0.04792751371860504, + 0.018341224640607834, + 0.014044215902686119, + 0.016818776726722717, + 0.02015463262796402, + 0.005555087234824896, + 0.011420364491641521, + 0.008220788091421127, + 0.020698953419923782, + 0.009650642983615398, + 0.0125162648037076, + 0.008134105242788792, + 0.00913812406361103, + 0.0066337790340185165, + 0.007526252884417772, + 0.011375104077160358, + 0.010310200974345207, + 0.009006982669234276, + 0.014692872762680054, + 0.006236319430172443, + 0.011302074417471886, + 0.011015472002327442, + 0.008563194423913956, + 0.014287682250142097, + 0.01119219046086073, + 0.006816300563514233, + 0.006929741706699133, + 0.011754992417991161, + 0.008026006631553173, + 0.0121056092903018, + 0.007928245700895786, + 0.007546377833932638, + 0.009629892185330391, + 0.008263117633759975, + 0.007192438002675772, + 0.021420367062091827, + 0.008778361603617668, + 0.01469658687710762, + 0.010733393020927906, + 0.011981145478785038, + 0.013254459016025066, + 0.0044204299338161945, + 0.005637848284095526, + 0.005878382362425327, + 0.006874127313494682, + 0.003879481926560402, + 0.006431044079363346, + 0.0038377013988792896, + 0.004778054542839527, + 0.004439860116690397, + 0.0056400420144200325, + 0.0065543511882424355, + 0.0069126528687775135, + 0.0035731648094952106, + 0.04226412624120712, + 0.00552867166697979, + 0.05078434199094772, + 0.004444292280822992, + 0.005421127658337355, + 0.0056506977416574955, + 0.0039120023138821125, + 0.0023847483098506927, + 0.01055078487843275, + 0.0024900001008063555, + 0.009321508929133415, + 0.006516783498227596, + 0.014081882312893867, + 0.01653236337006092, + 0.018994992598891258, + 0.007438874803483486, + 0.02915889210999012, + 0.009699853137135506, + 0.045945510268211365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.003192449454218149 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195887371897697 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195887371897697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195887371897697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20335863530635834, + 0.04320153594017029, + 0.04504087567329407, + 0.04851861298084259, + 0.04588140919804573, + 0.03149912878870964, + 0.0405820794403553, + 0.11616743355989456, + 0.3581213653087616, + 0.0676289051771164 + ], + "credit_map": [ + 0.019151795655488968, + 0.005556988995522261, + 0.06900449842214584, + 0.02014259435236454, + 0.005536315497010946, + 0.011221951805055141, + 0.015332018956542015, + 0.05741246044635773, + 0.0051038432866334915, + 0.004307753872126341, + 0.003981994464993477, + 0.008924262598156929, + 0.0033726422116160393, + 0.0047683753073215485, + 0.0033388962037861347, + 0.009403767995536327, + 0.005095849744975567, + 0.006757182069122791, + 0.0037760345730930567, + 0.0055358391255140305, + 0.005602599587291479, + 0.004929876420646906, + 0.007571945432573557, + 0.005771544761955738, + 0.007294161710888147, + 0.0075057572685182095, + 0.0034032363910228014, + 0.004192041233181953, + 0.005542743485420942, + 0.0060660345479846, + 0.005462363827973604, + 0.009052272886037827, + 0.005283040925860405, + 0.005492492113262415, + 0.0052268593572080135, + 0.006647519301623106, + 0.006049436051398516, + 0.006347005255520344, + 0.004889047704637051, + 0.005946008488535881, + 0.004034883342683315, + 0.0029436699114739895, + 0.0039352597668766975, + 0.00515015609562397, + 0.0047408477403223515, + 0.0031471310649067163, + 0.003679092740640044, + 0.0038680864963680506, + 0.0030152820982038975, + 0.00503881648182869, + 0.0064537134021520615, + 0.00810664240270853, + 0.003045578021556139, + 0.005509908311069012, + 0.004698921926319599, + 0.004713216330856085, + 0.005365938879549503, + 0.007212246302515268, + 0.008397908881306648, + 0.009812881238758564, + 0.006813928484916687, + 0.03435758873820305, + 0.008077926002442837, + 0.036129020154476166, + 0.01631833054125309, + 0.022704679518938065, + 0.026253007352352142, + 0.017986182123422623, + 0.011965764686465263, + 0.12285064905881882, + 0.014592140913009644, + 0.12545061111450195, + 0.004343067761510611, + 0.007787374779582024, + 0.011028501205146313, + 0.009925906546413898, + 0.0036038595717400312, + 0.010072717443108559, + 0.005485497415065765, + 0.015381983481347561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.014195887371897697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327689342200756 + ], + "opposite_logprob_tokens": [ + -3.7732768058776855 + ], + "logprob_reward": -0.02327689342200756, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327689342200756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22359807789325714, + 0.0637425109744072, + 0.023098837584257126, + 0.02666338160634041, + 0.08208231627941132, + 0.0346512570977211, + 0.026659393683075905, + 0.053298741579055786, + 0.15163391828536987, + 0.31457164883613586 + ], + "credit_map": [ + 0.021859843283891678, + 0.006210574880242348, + 0.07387325912714005, + 0.02777346968650818, + 0.009009548462927341, + 0.011623685248196125, + 0.01732255145907402, + 0.055925123393535614, + 0.009403578005731106, + 0.006775431800633669, + 0.006348772440105677, + 0.012600315734744072, + 0.004083758220076561, + 0.006708980072289705, + 0.004320381209254265, + 0.013501285575330257, + 0.002368553774431348, + 0.0023289357777684927, + 0.002400791272521019, + 0.003452599747106433, + 0.002928707515820861, + 0.003626076737418771, + 0.0021416936069726944, + 0.003851476591080427, + 0.004715215414762497, + 0.0033782226964831352, + 0.001968976343050599, + 0.0032123022247105837, + 0.004211625549942255, + 0.002798251574859023, + 0.003442053683102131, + 0.0029367299284785986, + 0.008680237457156181, + 0.008043385110795498, + 0.012300673872232437, + 0.01022500079125166, + 0.00581737793982029, + 0.009965376928448677, + 0.006370298098772764, + 0.020679958164691925, + 0.005087059456855059, + 0.003185648936778307, + 0.005603963043540716, + 0.00622810423374176, + 0.0032726216595619917, + 0.004532684572041035, + 0.0028032311238348484, + 0.003937939647585154, + 0.0027218537870794535, + 0.003886872436851263, + 0.0038935879711061716, + 0.0027056720573455095, + 0.003671369282528758, + 0.0038066201377660036, + 0.0033665436785668135, + 0.0026068715378642082, + 0.0034421628806740046, + 0.004944751039147377, + 0.0055654640309512615, + 0.00714534567669034, + 0.0047625391744077206, + 0.010402422398328781, + 0.0044939229264855385, + 0.012542128562927246, + 0.015505473129451275, + 0.020052311941981316, + 0.023210817947983742, + 0.016498833894729614, + 0.010068118572235107, + 0.03219122812151909, + 0.009015649557113647, + 0.025091471150517464, + 0.014715543948113918, + 0.030854228883981705, + 0.040901850908994675, + 0.04612276330590248, + 0.015181710943579674, + 0.05336406081914902, + 0.021681496873497963, + 0.09174994379281998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.02327689342200756 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot captures the rear of a lone, classic black-and-white police cruiser as it slowly winds its way through the heart of a deserted city street at night. The vehicle's glossy paint, slick with rain, vividly reflects the silent, rhythmic pulse of red and blue emergency light strips, casting sweeping, ephemeral bursts of color across the wet, black asphalt. The camera then begins a smooth, steady crane motion, pulling upward and backward, gradually widening the frame. As the view rises, the street expands into a complex grid of empty streets, all shimmering in the cold, ambient glow and hazy atmosphere of distant streetlights. The shot culminates in a breathtaking high-angle, extreme wide shot, revealing a vast, desolate metropolis composed of towering, dark skyscrapers whose windows are barely lit, creating a powerful sense of scale and profound emptiness. The lone police cruiser is now a small, almost insignificant vehicle, its flashing lights the sole beacon of activity in the vast, silent cityscape.", + "seed": 11199, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_12/index_5190_seed_11199_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03383256141096354, + "vlm_logprob": -0.03383256141096354, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03383256141096354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1186646893620491 + ], + "opposite_logprob_tokens": [ + -2.2436647415161133 + ], + "logprob_reward": -0.1186646893620491, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1186646893620491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19779455661773682, + 0.09827884286642075, + 0.07391972094774246, + 0.08891435712575912, + 0.06223519518971443, + 0.04361388459801674, + 0.0764138251543045, + 0.02711787447333336, + 0.22383315861225128, + 0.10787859559059143 + ], + "credit_map": [ + 0.022855976596474648, + 0.024843188002705574, + 0.014890629798173904, + 0.019912298768758774, + 0.03425920009613037, + 0.027346445247530937, + 0.02584654651582241, + 0.02784026600420475, + 0.008381305262446404, + 0.014040268957614899, + 0.010936071164906025, + 0.008168704807758331, + 0.013270875439047813, + 0.012594078667461872, + 0.018174054101109505, + 0.012713490054011345, + 0.0041412957943975925, + 0.0056372834369540215, + 0.006668176501989365, + 0.006016267463564873, + 0.010754059068858624, + 0.011435233987867832, + 0.01875133253633976, + 0.010516071692109108, + 0.004335342440754175, + 0.005900839809328318, + 0.005191478878259659, + 0.012808932922780514, + 0.011746047995984554, + 0.021476812660694122, + 0.012041687965393066, + 0.015413215383887291, + 0.003581321332603693, + 0.005317374132573605, + 0.003590587293729186, + 0.009586975909769535, + 0.008112883195281029, + 0.009102324023842812, + 0.011072740890085697, + 0.011870991438627243, + 0.004876555409282446, + 0.004267493262887001, + 0.0024337605573236942, + 0.005575407296419144, + 0.005182687658816576, + 0.007535859942436218, + 0.006944513414055109, + 0.006797604728490114, + 0.013909511268138885, + 0.005789437331259251, + 0.005913742817938328, + 0.008507339283823967, + 0.006587319076061249, + 0.008737022057175636, + 0.011648407205939293, + 0.015321043320000172, + 0.004521082621067762, + 0.00465480936691165, + 0.001914623542688787, + 0.0019373579416424036, + 0.002901067491620779, + 0.0035521446261554956, + 0.001775780925527215, + 0.005861007142812014, + 0.03248396888375282, + 0.02818525955080986, + 0.023860320448875427, + 0.02176288329064846, + 0.030415508896112442, + 0.03431205078959465, + 0.01916402019560337, + 0.033649150282144547, + 0.009152140468358994, + 0.016654813662171364, + 0.015720773488283157, + 0.0108999814838171, + 0.017051415517926216, + 0.016855336725711823, + 0.010831029154360294, + 0.010713109746575356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.1186646893620491 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007656395435333252 + ], + "opposite_logprob_tokens": [ + -4.882656574249268 + ], + "logprob_reward": -0.007656395435333252, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007656395435333252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1704515665769577, + 0.08829844743013382, + 0.040587976574897766, + 0.06084589287638664, + 0.06372854858636856, + 0.14100222289562225, + 0.13325119018554688, + 0.09144666790962219, + 0.11868084967136383, + 0.09170670062303543 + ], + "credit_map": [ + 0.021035628393292427, + 0.02186889387667179, + 0.016122734174132347, + 0.015915807336568832, + 0.027529766783118248, + 0.02456032671034336, + 0.021045848727226257, + 0.022372547537088394, + 0.008176679722964764, + 0.009592191316187382, + 0.014636515639722347, + 0.0076407333835959435, + 0.010723014362156391, + 0.013559448532760143, + 0.011459756642580032, + 0.012510107830166817, + 0.0023032445460557938, + 0.004570282530039549, + 0.004268345423042774, + 0.0044405641965568066, + 0.0031243839766830206, + 0.008842124603688717, + 0.004774415399879217, + 0.008264617063105106, + 0.003215207252651453, + 0.005006223451346159, + 0.005206770729273558, + 0.011066727340221405, + 0.00690044742077589, + 0.01131833903491497, + 0.009086398407816887, + 0.009045783430337906, + 0.0054345340467989445, + 0.006177160423249006, + 0.006578505504876375, + 0.007805987261235714, + 0.00913752056658268, + 0.011531263589859009, + 0.006181386299431324, + 0.01088219229131937, + 0.020179612562060356, + 0.017978264018893242, + 0.01509317010641098, + 0.018371539190411568, + 0.015432963147759438, + 0.016646668314933777, + 0.01775503344833851, + 0.019544968381524086, + 0.023342976346611977, + 0.013894634321331978, + 0.013670893386006355, + 0.01818186230957508, + 0.012232224456965923, + 0.013699464499950409, + 0.016233714297413826, + 0.021995441988110542, + 0.018711794167757034, + 0.008118309080600739, + 0.009191923774778843, + 0.009332963265478611, + 0.00889559555798769, + 0.008924867957830429, + 0.007860446348786354, + 0.020410768687725067, + 0.011358554475009441, + 0.01641252636909485, + 0.012552578933537006, + 0.013830033130943775, + 0.015680765733122826, + 0.01829567365348339, + 0.011339280754327774, + 0.019211433827877045, + 0.00894987303763628, + 0.010005725547671318, + 0.011553922668099403, + 0.009859565645456314, + 0.018283970654010773, + 0.013332781381905079, + 0.010097615420818329, + 0.009623243473470211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.007656395435333252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978105328977108 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.00978105328977108, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978105328977108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04756329581141472, + 0.0361078642308712, + 0.020198212936520576, + 0.033120959997177124, + 0.032436639070510864, + 0.03563963621854782, + 0.039627768099308014, + 0.10111724585294724, + 0.12957720458507538, + 0.5246111750602722 + ], + "credit_map": [ + 0.007183374837040901, + 0.005869466811418533, + 0.004809525795280933, + 0.0051708403043448925, + 0.005425584968179464, + 0.0068258740939199924, + 0.006667837966233492, + 0.005610796622931957, + 0.0026565492153167725, + 0.00430191308259964, + 0.00266178697347641, + 0.004129481967538595, + 0.004017193801701069, + 0.005275013856589794, + 0.007947603240609169, + 0.005118326749652624, + 0.0013521334622055292, + 0.0021836129017174244, + 0.0023497161455452442, + 0.002430805703625083, + 0.0017257995205000043, + 0.0031193336471915245, + 0.004299168474972248, + 0.002737644826993346, + 0.0012058706488460302, + 0.002696661278605461, + 0.0019492292776703835, + 0.005316106602549553, + 0.004448858555406332, + 0.006275129038840532, + 0.005775851663202047, + 0.005453259684145451, + 0.002648622030392289, + 0.003113583894446492, + 0.0026165645103901625, + 0.004526775795966387, + 0.0032615207601338625, + 0.006156011484563351, + 0.004236006177961826, + 0.0058775609359145164, + 0.0042246561497449875, + 0.004062282387167215, + 0.0029891948215663433, + 0.004390044137835503, + 0.004121069796383381, + 0.005158820655196905, + 0.005891729146242142, + 0.004801844712346792, + 0.007478188257664442, + 0.005651501938700676, + 0.0028939831536263227, + 0.004163081757724285, + 0.0035486212000250816, + 0.0056230747140944, + 0.003974991384893656, + 0.006294329650700092, + 0.021176127716898918, + 0.01106354407966137, + 0.0073441374115645885, + 0.009462852962315083, + 0.008523136377334595, + 0.011783492751419544, + 0.008873121812939644, + 0.022890837863087654, + 0.022756893187761307, + 0.015915563330054283, + 0.013508663512766361, + 0.0164423156529665, + 0.017287706956267357, + 0.019551467150449753, + 0.009258698672056198, + 0.014855911955237389, + 0.05088508501648903, + 0.06031617894768715, + 0.06209035962820053, + 0.05525902286171913, + 0.09771782904863358, + 0.07945869117975235, + 0.060087915509939194, + 0.05879610776901245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.00978105328977108 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978435855358839 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.00978435855358839, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978435855358839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08714082092046738, + 0.0452461801469326, + 0.030645320191979408, + 0.052507027983665466, + 0.07041250914335251, + 0.05054863169789314, + 0.14771497249603271, + 0.0942016988992691, + 0.23568354547023773, + 0.18589922785758972 + ], + "credit_map": [ + 0.014541280455887318, + 0.00794784165918827, + 0.009983960539102554, + 0.008732779882848263, + 0.009710111655294895, + 0.012721112929284573, + 0.012804076075553894, + 0.010699663311243057, + 0.004270720761269331, + 0.004321024287492037, + 0.007935083471238613, + 0.004495204892009497, + 0.00412776181474328, + 0.006109226029366255, + 0.007336540147662163, + 0.006650626193732023, + 0.0016432912088930607, + 0.0035732630640268326, + 0.003680790076032281, + 0.002735074842348695, + 0.003101109294220805, + 0.004637913312762976, + 0.007146649993956089, + 0.0041272323578596115, + 0.0021422209683805704, + 0.00405488396063447, + 0.002994315465912223, + 0.00723183061927557, + 0.00890191737562418, + 0.00810861773788929, + 0.010805921629071236, + 0.008267329074442387, + 0.005707154516130686, + 0.007501648273319006, + 0.005803852342069149, + 0.011418852023780346, + 0.008833060972392559, + 0.007499233353883028, + 0.01152366679161787, + 0.012125048786401749, + 0.006539674010127783, + 0.006262455601245165, + 0.003526239423081279, + 0.006216332316398621, + 0.006100423168390989, + 0.005502735730260611, + 0.008857031352818012, + 0.007543746381998062, + 0.021591298282146454, + 0.019881298765540123, + 0.012665030546486378, + 0.01381105650216341, + 0.016422729939222336, + 0.03049588017165661, + 0.013934674672782421, + 0.018913021311163902, + 0.016458116471767426, + 0.02541736140847206, + 0.004694437608122826, + 0.005293827038258314, + 0.0070557608269155025, + 0.016992531716823578, + 0.005720841232687235, + 0.012568837031722069, + 0.03722056746482849, + 0.031041966751217842, + 0.02432246319949627, + 0.022533932700753212, + 0.030566122382879257, + 0.039466120302677155, + 0.015790969133377075, + 0.034741438925266266, + 0.022564176470041275, + 0.027270905673503876, + 0.02203400246798992, + 0.013845437206327915, + 0.037537313997745514, + 0.027440570294857025, + 0.017223570495843887, + 0.017983270809054375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.00978435855358839 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a classic black-and-white police cruiser viewed from the rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle's paint glossy and wet with rain, reflecting red and blue emergency light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a deserted city street at night with wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle medium shot and execute a smooth crane motion pulling upward and backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot reveal a vast, desolate metropolis with towering dark skyscrapers and barely lit windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327631041407585 + ], + "opposite_logprob_tokens": [ + -3.7732763290405273 + ], + "logprob_reward": -0.02327631041407585, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327631041407585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08657650649547577, + 0.06495636701583862, + 0.027096526697278023, + 0.0535210482776165, + 0.049089204519987106, + 0.04392771050333977, + 0.07713261246681213, + 0.0832877829670906, + 0.1782398819923401, + 0.33617234230041504 + ], + "credit_map": [ + 0.014055383391678333, + 0.00882427766919136, + 0.011143447831273079, + 0.010555371642112732, + 0.008477933704853058, + 0.011793744750320911, + 0.010652701370418072, + 0.011073650792241096, + 0.0062435101717710495, + 0.006873805541545153, + 0.010370367206633091, + 0.006283062044531107, + 0.006221582181751728, + 0.009039252996444702, + 0.009625229984521866, + 0.010299557819962502, + 0.0013951826840639114, + 0.0031070478726178408, + 0.0027993409894406796, + 0.0034497773740440607, + 0.0024745860137045383, + 0.0039450726471841335, + 0.006038474850356579, + 0.0038870451971888542, + 0.0022902265191078186, + 0.004530706442892551, + 0.0028030797839164734, + 0.009880836121737957, + 0.007821443490684032, + 0.009379028342664242, + 0.008539380505681038, + 0.008276348002254963, + 0.003726729191839695, + 0.00511428015306592, + 0.0034500346519052982, + 0.006445535458624363, + 0.007750983815640211, + 0.006851649843156338, + 0.007573711685836315, + 0.008176280185580254, + 0.005572868976742029, + 0.004984037019312382, + 0.0031556531321257353, + 0.005872079636901617, + 0.004921240732073784, + 0.005381522700190544, + 0.007242841646075249, + 0.006797467358410358, + 0.011759726330637932, + 0.009202644228935242, + 0.004185607191175222, + 0.007786657195538282, + 0.008930540643632412, + 0.01599281281232834, + 0.007919124327600002, + 0.011355496942996979, + 0.01729901321232319, + 0.02826380915939808, + 0.004231571685522795, + 0.005835132207721472, + 0.007151432801038027, + 0.01080999430269003, + 0.0038286095950752497, + 0.005868229083716869, + 0.029467111453413963, + 0.025289755314588547, + 0.017603296786546707, + 0.023342827335000038, + 0.026652077212929726, + 0.022928010672330856, + 0.01564093865454197, + 0.01731586642563343, + 0.033023007214069366, + 0.03889275714755058, + 0.04002964124083519, + 0.027618667110800743, + 0.06685534119606018, + 0.04982774704694748, + 0.04028107970952988, + 0.039644137024879456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.02327631041407585 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_13/index_tensor([8216, 8216])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_13/index_tensor([8216, 8216])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8c9463392df91486c3ea726393adc954ff663afc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_13/index_tensor([8216, 8216])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up macro shot, taken from directly above, focuses on the inner world of a complex mechanical watch, its open, skeletonized dial revealing its intricate mechanics. A slender, mirror-polished golden seconds hand sweeps across the dial with impeccable, steady, and hypnotic smoothness, its tip tracing the brass and steel forest of mechanisms beneath with unerring precision. As the seconds hand rotates, the escapement beneath it vibrates at a visible rate, setting the interlocking layers of tiny gears in a silent and precise ballet. A soft, single-point studio light shines from the upper left, sculpting flowing, sharp highlights on the chamfered edges and brushed metal surfaces of each gear, while casting subtle, dynamic shadows within the deeper mechanisms. In the picture, the warm tones of gold and brass, the cool tones of fine steel, and the deep red artificial rubies used as bearings embellish it together form a rich and harmonious color painting. The whole scene is filled with a kind of concentration and tranquil atmosphere of awe-inspiring respect for the ultimate craftsmanship, timeless design and mechanical beauty.", + "seed": 1914, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_13/index_8216_seed_1914_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04268131270073354, + "vlm_logprob": -0.04268131270073354, + "vlm_yes_no_margin": 3.774999976158142, + "vlm_token_credit": -0.04268131270073354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03617390990257263 + ], + "opposite_logprob_tokens": [ + -3.6611738204956055 + ], + "logprob_reward": -0.03617390990257263, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03617390990257263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08767952024936676, + 0.05304925516247749, + 0.09256309270858765, + 0.21326184272766113, + 0.12907986342906952, + 0.14614787697792053, + 0.13912424445152283, + 0.04812377318739891, + 0.052996691316366196, + 0.037973854690790176 + ], + "credit_map": [ + 0.011806798167526722, + 0.0075225080363452435, + 0.009884263388812542, + 0.008713962510228157, + 0.012174861505627632, + 0.010422509163618088, + 0.007945945486426353, + 0.019208675250411034, + 0.006339230574667454, + 0.0043765828013420105, + 0.004717627074569464, + 0.008019418455660343, + 0.007710916921496391, + 0.005535054951906204, + 0.007535561919212341, + 0.008814861066639423, + 0.010829075239598751, + 0.009745744988322258, + 0.018345633521676064, + 0.01194697804749012, + 0.007572940085083246, + 0.010442117229104042, + 0.011436110362410545, + 0.012244493700563908, + 0.009057826362550259, + 0.007369065657258034, + 0.01328666415065527, + 0.006288249045610428, + 0.006088606081902981, + 0.087404265999794, + 0.011528837494552135, + 0.07223831862211227, + 0.009244809858500957, + 0.010246865451335907, + 0.009414543397724628, + 0.03932171314954758, + 0.008250569924712181, + 0.007077852264046669, + 0.013838891871273518, + 0.03168461471796036, + 0.012416239827871323, + 0.02351430431008339, + 0.01752089150249958, + 0.016807038336992264, + 0.017845267429947853, + 0.019047312438488007, + 0.023378316313028336, + 0.01561850868165493, + 0.031156914308667183, + 0.01696888916194439, + 0.010494460351765156, + 0.02358771488070488, + 0.01641719974577427, + 0.014067541807889938, + 0.012797209434211254, + 0.013634318485856056, + 0.005519659724086523, + 0.005578635726124048, + 0.006562523078173399, + 0.007167133968323469, + 0.0063089714385569096, + 0.004311080556362867, + 0.0065828044898808, + 0.006092965602874756, + 0.009988120757043362, + 0.007745195180177689, + 0.003972026519477367, + 0.0040655010379850864, + 0.005782724358141422, + 0.00752587616443634, + 0.005742561537772417, + 0.00817468948662281, + 0.005210100207477808, + 0.004123707767575979, + 0.008460629731416702, + 0.003562465077266097, + 0.0039032550994306803, + 0.002844093833118677, + 0.004895705264061689, + 0.004973899573087692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.03617390990257263 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003219066886231303 + ], + "opposite_logprob_tokens": [ + -5.753219127655029 + ], + "logprob_reward": -0.003219066886231303, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003219066886231303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1250007003545761, + 0.07355595380067825, + 0.03746955096721649, + 0.17219938337802887, + 0.15595348179340363, + 0.05148628354072571, + 0.2186867594718933, + 0.03377804160118103, + 0.0787457823753357, + 0.05312405899167061 + ], + "credit_map": [ + 0.013064328581094742, + 0.0103911766782403, + 0.015426787547767162, + 0.020684324204921722, + 0.01674165576696396, + 0.012181599624454975, + 0.019130609929561615, + 0.017380230128765106, + 0.016850221902132034, + 0.0036338374484330416, + 0.005618594586849213, + 0.008288145996630192, + 0.004640502855181694, + 0.006713635288178921, + 0.007435866631567478, + 0.020375153049826622, + 0.004078036639839411, + 0.0045625073835253716, + 0.007818695157766342, + 0.00328146037645638, + 0.004126374609768391, + 0.004288805183023214, + 0.004033328499644995, + 0.005280346143990755, + 0.007619758136570454, + 0.0062303305603563786, + 0.0037030905950814486, + 0.004161527846008539, + 0.005480092950165272, + 0.07174389064311981, + 0.009418854489922523, + 0.06384185701608658, + 0.00913168117403984, + 0.007804763037711382, + 0.006529107224196196, + 0.06114735081791878, + 0.005108883138746023, + 0.008846355602145195, + 0.008806484751403332, + 0.04857888072729111, + 0.004893444012850523, + 0.006221129093319178, + 0.006386312190443277, + 0.006806021090596914, + 0.007371066603809595, + 0.0040711890906095505, + 0.009178620763123035, + 0.006558505352586508, + 0.0342596098780632, + 0.03193299472332001, + 0.026214037090539932, + 0.029826965183019638, + 0.022931478917598724, + 0.019175926223397255, + 0.024811333045363426, + 0.029534438624978065, + 0.003267682855948806, + 0.004007620736956596, + 0.006105536594986916, + 0.004396864213049412, + 0.004397990647703409, + 0.003190746996551752, + 0.004505873657763004, + 0.0039057289250195026, + 0.012791120447218418, + 0.008683525957167149, + 0.006662143860012293, + 0.00715296994894743, + 0.007847798056900501, + 0.012214791029691696, + 0.008531666360795498, + 0.014861776493489742, + 0.008761920034885406, + 0.005045090336352587, + 0.010730050504207611, + 0.0050782556645572186, + 0.004835629370063543, + 0.004363463260233402, + 0.008220207877457142, + 0.006089443806558847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.003219066886231303 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020599473267793655 + ], + "opposite_logprob_tokens": [ + -3.895599365234375 + ], + "logprob_reward": -0.020599473267793655, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020599473267793655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10500788688659668, + 0.07718410342931747, + 0.04703327640891075, + 0.09183319658041, + 0.09218734502792358, + 0.16662460565567017, + 0.23990322649478912, + 0.05006822571158409, + 0.06422003358602524, + 0.0659380853176117 + ], + "credit_map": [ + 0.013920472003519535, + 0.012851694598793983, + 0.018560657277703285, + 0.011791929602622986, + 0.010517343878746033, + 0.011749908328056335, + 0.010889751836657524, + 0.014726139605045319, + 0.018853794783353806, + 0.003596297698095441, + 0.006369051989167929, + 0.007824025116860867, + 0.005452511832118034, + 0.005410039331763983, + 0.0066143302246928215, + 0.023064056411385536, + 0.005200491286814213, + 0.007142455782741308, + 0.005636226385831833, + 0.005157935433089733, + 0.004781064577400684, + 0.005290687084197998, + 0.004313764162361622, + 0.009510654024779797, + 0.0040209428407251835, + 0.0036144449841231108, + 0.009887436404824257, + 0.0035869081038981676, + 0.004605080001056194, + 0.033851128071546555, + 0.005826679989695549, + 0.026440590620040894, + 0.00552928214892745, + 0.005441112909466028, + 0.00355085963383317, + 0.03726575896143913, + 0.004141112323850393, + 0.0054357317276299, + 0.00615037651732564, + 0.024673113599419594, + 0.01535345520824194, + 0.0312309879809618, + 0.018137574195861816, + 0.016769809648394585, + 0.019354183226823807, + 0.025549184530973434, + 0.021961163729429245, + 0.01826826110482216, + 0.04875730723142624, + 0.02131643332540989, + 0.039389196783304214, + 0.01712786592543125, + 0.018764281645417213, + 0.029249045997858047, + 0.018456192687153816, + 0.04684291034936905, + 0.007532699033617973, + 0.005337232258170843, + 0.006661025807261467, + 0.006326568312942982, + 0.006897112354636192, + 0.005264807492494583, + 0.005793957505375147, + 0.00625482527539134, + 0.0067241257056593895, + 0.010729948990046978, + 0.004795832093805075, + 0.007863150909543037, + 0.0075746700167655945, + 0.009937942959368229, + 0.008031594567000866, + 0.00856277346611023, + 0.008665207773447037, + 0.004910753108561039, + 0.014995397999882698, + 0.005493008531630039, + 0.005316965747624636, + 0.0058386302553117275, + 0.013388831168413162, + 0.007329294923692942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.020599473267793655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706884741783142 + ], + "opposite_logprob_tokens": [ + -2.1270687580108643 + ], + "logprob_reward": -0.12706884741783142, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12706884741783142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07946613430976868, + 0.043476734310388565, + 0.14002834260463715, + 0.1367190033197403, + 0.12900254130363464, + 0.11177512258291245, + 0.11618871241807938, + 0.14231863617897034, + 0.035664673894643784, + 0.06536000221967697 + ], + "credit_map": [ + 0.008469711989164352, + 0.007333385292440653, + 0.0126024279743433, + 0.008448483422398567, + 0.00816163420677185, + 0.008443910628557205, + 0.01349586620926857, + 0.012510726228356361, + 0.005336785223335028, + 0.005132208578288555, + 0.004400629084557295, + 0.007579597644507885, + 0.005500079598277807, + 0.005747330840677023, + 0.004672341048717499, + 0.005107769276946783, + 0.019940774887800217, + 0.022334475070238113, + 0.022608129307627678, + 0.015714533627033234, + 0.01832498237490654, + 0.010087885893881321, + 0.01264822669327259, + 0.01836935058236122, + 0.0062615880742669106, + 0.005479795392602682, + 0.003957872744649649, + 0.005796412006020546, + 0.0065556298941373825, + 0.0533473864197731, + 0.008372520096600056, + 0.04694781452417374, + 0.009470204822719097, + 0.0075686038471758366, + 0.006684089545160532, + 0.0421774797141552, + 0.0073013887740671635, + 0.00939376000314951, + 0.00977671891450882, + 0.03663031756877899, + 0.010495229624211788, + 0.02246106043457985, + 0.011997747235000134, + 0.013459288515150547, + 0.010674542747437954, + 0.016396209597587585, + 0.014075902290642262, + 0.012215144000947475, + 0.019723134115338326, + 0.010164809413254261, + 0.015387563966214657, + 0.011399677023291588, + 0.008652816526591778, + 0.017038369551301003, + 0.011055379174649715, + 0.022766975685954094, + 0.013472946360707283, + 0.020447829738259315, + 0.018977204337716103, + 0.0242836344987154, + 0.019163282588124275, + 0.012043808586895466, + 0.01689857803285122, + 0.01703137345612049, + 0.0035512165632098913, + 0.00528855761513114, + 0.002941205631941557, + 0.003933178260922432, + 0.0042176940478384495, + 0.005126393865793943, + 0.0044786762446165085, + 0.006127751898020506, + 0.00737338000908494, + 0.005368183832615614, + 0.014958593063056469, + 0.005071945022791624, + 0.005404808092862368, + 0.005424294155091047, + 0.014705509878695011, + 0.007053291890770197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.12706884741783142 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263452660292387 + ], + "opposite_logprob_tokens": [ + -3.6513452529907227 + ], + "logprob_reward": -0.0263452660292387, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263452660292387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03290197253227234, + 0.04722331836819649, + 0.12332271784543991, + 0.15255168080329895, + 0.03911587595939636, + 0.15398922562599182, + 0.09379100799560547, + 0.04665761813521385, + 0.06009892746806145, + 0.25034764409065247 + ], + "credit_map": [ + 0.0038310957606881857, + 0.0033634272404015064, + 0.004212417174130678, + 0.0034556768368929625, + 0.0031082218047231436, + 0.004477465990930796, + 0.005449396558105946, + 0.005004270002245903, + 0.011384026147425175, + 0.0023100494872778654, + 0.003983325324952602, + 0.007320830132812262, + 0.0025715359952300787, + 0.003976435400545597, + 0.003976211417466402, + 0.011700901202857494, + 0.019977623596787453, + 0.017314419150352478, + 0.021594446152448654, + 0.01153161283582449, + 0.01617070659995079, + 0.009968100115656853, + 0.010491948574781418, + 0.01627386547625065, + 0.006660584360361099, + 0.005443584639579058, + 0.009723112918436527, + 0.0039189523085951805, + 0.005692788399755955, + 0.061828695237636566, + 0.007989369332790375, + 0.05129459127783775, + 0.002777602756395936, + 0.0041356259025633335, + 0.004132864065468311, + 0.008859779685735703, + 0.004427327774465084, + 0.0023974296636879444, + 0.006174017209559679, + 0.006211233325302601, + 0.014268610626459122, + 0.0283109862357378, + 0.01715870201587677, + 0.015519320964813232, + 0.01988171972334385, + 0.02092278003692627, + 0.021268395707011223, + 0.016658706590533257, + 0.014711138792335987, + 0.009591029956936836, + 0.009072750806808472, + 0.0133192865177989, + 0.018347110599279404, + 0.01028495654463768, + 0.009588941931724548, + 0.008875790983438492, + 0.0051401820965111256, + 0.005164125468581915, + 0.008837657049298286, + 0.00630599819123745, + 0.006226575933396816, + 0.00411881972104311, + 0.0060004000551998615, + 0.004863859154284, + 0.010112284682691097, + 0.00816394668072462, + 0.004994018003344536, + 0.005115015432238579, + 0.005093357991427183, + 0.010143148712813854, + 0.006100165657699108, + 0.010376994498074055, + 0.02441152185201645, + 0.024100884795188904, + 0.07071465998888016, + 0.02635306864976883, + 0.021443236619234085, + 0.021743644028902054, + 0.029808958992362022, + 0.03177168592810631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.0263452660292387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro shot, taken from directly above, focuses on the inner world of a complex mechanical watch, its open, skeletonized dial revealing its intricate mechanics. A slender, mirror-polished golden seconds hand sweeps across the dial with impeccable, steady, and hypnotic smoothness, its tip tracing the brass and steel forest of mechanisms beneath with unerring precision. As the seconds hand rotates, the escapement beneath it vibrates at a visible rate, setting the interlocking layers of tiny gears in a silent and precise ballet. A soft, single-point studio light shines from the upper left, sculpting flowing, sharp highlights on the chamfered edges and brushed metal surfaces of each gear, while casting subtle, dynamic shadows within the deeper mechanisms. In the picture, the warm tones of gold and brass, the cool tones of fine steel, and the deep red artificial rubies used as bearings embellish it together form a rich and harmonious color painting. The whole scene is filled with a kind of concentration and tranquil atmosphere of awe-inspiring respect for the ultimate craftsmanship, timeless design and mechanical beauty.", + "seed": 1915, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_13/index_8216_seed_1915_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04167626744601875, + "vlm_logprob": -0.04167626744601875, + "vlm_yes_no_margin": 4.249999976158142, + "vlm_token_credit": -0.04167626744601875, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02796577289700508 + ], + "opposite_logprob_tokens": [ + -4.027965545654297 + ], + "logprob_reward": -0.02796577289700508, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02796577289700508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06463933736085892, + 0.07804284989833832, + 0.06914003193378448, + 0.1643289029598236, + 0.2585468888282776, + 0.11161945760250092, + 0.015978414565324783, + 0.09069819003343582, + 0.10499714314937592, + 0.042008787393569946 + ], + "credit_map": [ + 0.009137282148003578, + 0.008510077372193336, + 0.00717906141653657, + 0.00698093231767416, + 0.009302016347646713, + 0.007386950775980949, + 0.00774101447314024, + 0.008402004837989807, + 0.008484764024615288, + 0.0080543402582407, + 0.009014413692057133, + 0.015465520322322845, + 0.011800511740148067, + 0.007527345325797796, + 0.007678851950913668, + 0.010017105378210545, + 0.009064543060958385, + 0.007165863644331694, + 0.0056295013055205345, + 0.007007874548435211, + 0.009457535110414028, + 0.017231008037924767, + 0.006664304994046688, + 0.0069194091483950615, + 0.024417903274297714, + 0.01848253421485424, + 0.022195931524038315, + 0.022687477990984917, + 0.02275281772017479, + 0.017573753371834755, + 0.02118622697889805, + 0.015032265335321426, + 0.031278546899557114, + 0.02126879058778286, + 0.01787448860704899, + 0.012379051186144352, + 0.09595484286546707, + 0.01856249012053013, + 0.014127262867987156, + 0.04710144177079201, + 0.01401879545301199, + 0.013118241913616657, + 0.006800631992518902, + 0.008021221496164799, + 0.026983585208654404, + 0.014688832685351372, + 0.008960869163274765, + 0.019027283415198326, + 0.0020051109604537487, + 0.0029625254683196545, + 0.0009786223527044058, + 0.0027439717669039965, + 0.00151729432400316, + 0.002058484125882387, + 0.002007249277085066, + 0.0017051578033715487, + 0.010960974730551243, + 0.011583401821553707, + 0.010087208822369576, + 0.011382349766790867, + 0.012728150933980942, + 0.008508190512657166, + 0.014060776680707932, + 0.011387153528630733, + 0.010481061413884163, + 0.018670303747057915, + 0.010621125809848309, + 0.025716673582792282, + 0.011726930737495422, + 0.009684992954134941, + 0.008854479528963566, + 0.009241574443876743, + 0.006173002067953348, + 0.009434972889721394, + 0.0034981875214725733, + 0.005186179652810097, + 0.004758097697049379, + 0.005140568595379591, + 0.0036975406110286713, + 0.004120241850614548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.02796577289700508 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010617101797834039 + ], + "opposite_logprob_tokens": [ + -6.876061916351318 + ], + "logprob_reward": -0.0010617101797834039, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010617101797834039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10256115347146988, + 0.20874202251434326, + 0.06807348877191544, + 0.0583425909280777, + 0.08408644050359726, + 0.05433338135480881, + 0.0969318374991417, + 0.21123871207237244, + 0.0717150866985321, + 0.04397531598806381 + ], + "credit_map": [ + 0.010668610222637653, + 0.016320904716849327, + 0.011731317266821861, + 0.012696164660155773, + 0.016536470502614975, + 0.01062797475606203, + 0.010590427555143833, + 0.013389280997216702, + 0.02933036908507347, + 0.02909024991095066, + 0.025766819715499878, + 0.025420477613806725, + 0.025514710694551468, + 0.02307046391069889, + 0.024380728602409363, + 0.0261682141572237, + 0.01137864962220192, + 0.00842309184372425, + 0.009745221585035324, + 0.008193581365048885, + 0.006920369807630777, + 0.009381684474647045, + 0.007789699360728264, + 0.006241193041205406, + 0.007124726660549641, + 0.006590558215975761, + 0.006827063392847776, + 0.006185413803905249, + 0.009987966157495975, + 0.0071584610268473625, + 0.008512898348271847, + 0.005955501925200224, + 0.016550986096262932, + 0.009074066765606403, + 0.013363718055188656, + 0.0076033324003219604, + 0.00891365110874176, + 0.010027827695012093, + 0.010014558210968971, + 0.008538302965462208, + 0.0054494100622832775, + 0.004429439082741737, + 0.0038752364926040173, + 0.006618208717554808, + 0.013079909607768059, + 0.006378888618201017, + 0.005165164824575186, + 0.009337124414741993, + 0.012286901473999023, + 0.01819084770977497, + 0.010647370480000973, + 0.011488577350974083, + 0.009962700307369232, + 0.010946356691420078, + 0.011538260616362095, + 0.011870821006596088, + 0.024820823222398758, + 0.027619292959570885, + 0.025858456268906593, + 0.029245469719171524, + 0.026032021269202232, + 0.023818643763661385, + 0.027781406417489052, + 0.026062577962875366, + 0.007461829576641321, + 0.01295362040400505, + 0.007731324061751366, + 0.016561053693294525, + 0.007109123282134533, + 0.006879652850329876, + 0.006182842422276735, + 0.006835642270743847, + 0.004626522772014141, + 0.005040761549025774, + 0.007537118159234524, + 0.008424847386777401, + 0.005379556678235531, + 0.004714702256023884, + 0.005033701192587614, + 0.0032181073911488056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.0010617101797834039 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810123145580292 + ], + "opposite_logprob_tokens": [ + -3.5298101902008057 + ], + "logprob_reward": -0.029810123145580292, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810123145580292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12389182299375534, + 0.08735349029302597, + 0.027422821149230003, + 0.10037583112716675, + 0.12019547820091248, + 0.03418101370334625, + 0.02347201481461525, + 0.15372209250926971, + 0.0437418632209301, + 0.2856435477733612 + ], + "credit_map": [ + 0.015068618580698967, + 0.016804177314043045, + 0.014466953463852406, + 0.014734567143023014, + 0.01845446228981018, + 0.01557985320687294, + 0.01509150955826044, + 0.013691680505871773, + 0.009463593363761902, + 0.007866842672228813, + 0.009483259171247482, + 0.01802769862115383, + 0.01137111522257328, + 0.011338569223880768, + 0.008356076665222645, + 0.011446332558989525, + 0.004995394963771105, + 0.002604776993393898, + 0.0025953524746000767, + 0.0032076379284262657, + 0.0033674633596092463, + 0.004986243322491646, + 0.0022892379201948643, + 0.003376713255420327, + 0.014567503705620766, + 0.009342484176158905, + 0.01898329146206379, + 0.015391181223094463, + 0.009987978264689445, + 0.009685838595032692, + 0.012406688183546066, + 0.010010865516960621, + 0.00850122980773449, + 0.012470121495425701, + 0.006306113675236702, + 0.004264838062226772, + 0.0511200986802578, + 0.008078479208052158, + 0.006059648934751749, + 0.023394953459501266, + 0.00379018671810627, + 0.003703919006511569, + 0.00357684469781816, + 0.0029499633237719536, + 0.005302121862769127, + 0.0038181934505701065, + 0.0037916447035968304, + 0.007248139474540949, + 0.0025270653422921896, + 0.00296447379514575, + 0.0024988215882331133, + 0.003482022788375616, + 0.0027870943304151297, + 0.0031853823456913233, + 0.0035851080901920795, + 0.002442045835778117, + 0.016517849639058113, + 0.021669453009963036, + 0.01750526763498783, + 0.01856926642358303, + 0.024939417839050293, + 0.016602838411927223, + 0.01963745430111885, + 0.018280526623129845, + 0.0047625647857785225, + 0.008816084824502468, + 0.003928471822291613, + 0.005255177617073059, + 0.004403592552989721, + 0.006770351901650429, + 0.0037360196001827717, + 0.006069600582122803, + 0.022254159674048424, + 0.05417509749531746, + 0.032103780657052994, + 0.06115957722067833, + 0.024345045909285545, + 0.020457616075873375, + 0.029538875445723534, + 0.04160941764712334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.029810123145580292 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280955493450165 + ], + "opposite_logprob_tokens": [ + -2.0178096294403076 + ], + "logprob_reward": -0.14280955493450165, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14280955493450165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1404968649148941, + 0.07700151950120926, + 0.13447123765945435, + 0.049501340836286545, + 0.05865325778722763, + 0.04347981512546539, + 0.06830181181430817, + 0.18827740848064423, + 0.05094192549586296, + 0.18887490034103394 + ], + "credit_map": [ + 0.01026378758251667, + 0.03452587127685547, + 0.009781884029507637, + 0.009642957709729671, + 0.02122349850833416, + 0.012563931755721569, + 0.010180959478020668, + 0.032313961535692215, + 0.009775938466191292, + 0.00919684674590826, + 0.0073078996501863, + 0.013088846579194069, + 0.009356042370200157, + 0.010380582883954048, + 0.0066046747379004955, + 0.011290681548416615, + 0.022955674678087234, + 0.01443382166326046, + 0.013243583030998707, + 0.011780724860727787, + 0.01306950394064188, + 0.03133116662502289, + 0.01024441048502922, + 0.017412342131137848, + 0.0061779203824698925, + 0.005295237991958857, + 0.007351200561970472, + 0.008008535020053387, + 0.006700410507619381, + 0.004977615550160408, + 0.006517124827951193, + 0.004473294131457806, + 0.005811530165374279, + 0.005842234473675489, + 0.004149176646023989, + 0.003490774892270565, + 0.02103441022336483, + 0.00423035491257906, + 0.004855163395404816, + 0.00923960842192173, + 0.0047247945331037045, + 0.002952075796201825, + 0.0025455919094383717, + 0.0034655826166272163, + 0.012298988178372383, + 0.0048571075312793255, + 0.003617771901190281, + 0.009017899632453918, + 0.006627338007092476, + 0.013026979751884937, + 0.007467666640877724, + 0.009093692526221275, + 0.00691197207197547, + 0.007901204749941826, + 0.009364425204694271, + 0.007908529601991177, + 0.0189353059977293, + 0.02614092454314232, + 0.02260507643222809, + 0.02394768036901951, + 0.031237231567502022, + 0.018395105376839638, + 0.026063473895192146, + 0.020952589809894562, + 0.005725721828639507, + 0.010930576361715794, + 0.005304635968059301, + 0.007262952160090208, + 0.006771608721464872, + 0.0055288635194301605, + 0.003985508810728788, + 0.005432058125734329, + 0.01681648939847946, + 0.032683227211236954, + 0.019234642386436462, + 0.04136960580945015, + 0.015220378525555134, + 0.015021445229649544, + 0.019356679171323776, + 0.02917243354022503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.14280955493450165 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734176073223352 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734176073223352, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734176073223352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19085688889026642, + 0.06687115132808685, + 0.09704123437404633, + 0.025511758401989937, + 0.11184236407279968, + 0.01711094193160534, + 0.03259245678782463, + 0.08102741092443466, + 0.13048778474330902, + 0.2466580867767334 + ], + "credit_map": [ + 0.02575579844415188, + 0.041034940630197525, + 0.012154914438724518, + 0.030345361679792404, + 0.022195464000105858, + 0.015565857291221619, + 0.012679675593972206, + 0.031124860048294067, + 0.007718835026025772, + 0.00873957946896553, + 0.007567159365862608, + 0.01240596454590559, + 0.007972931489348412, + 0.006979154888540506, + 0.006498638540506363, + 0.008988883346319199, + 0.013298708014190197, + 0.007131767924875021, + 0.007936336100101471, + 0.017942294478416443, + 0.012078176252543926, + 0.01815740205347538, + 0.010282953269779682, + 0.010213587433099747, + 0.0027506831102073193, + 0.002531783189624548, + 0.003211948089301586, + 0.0041253152303397655, + 0.004386175889521837, + 0.0023864966351538897, + 0.0034599120263010263, + 0.00265944330021739, + 0.006540751550346613, + 0.009586348198354244, + 0.005549004301428795, + 0.004027898423373699, + 0.04845708608627319, + 0.00709147984161973, + 0.004322886932641268, + 0.02626689523458481, + 0.0016285122837871313, + 0.0018442777218297124, + 0.0016873304266482592, + 0.003042575204744935, + 0.0018943207105621696, + 0.0025033948477357626, + 0.0022546506952494383, + 0.0022558788768947124, + 0.0027002005372196436, + 0.004561262670904398, + 0.004055723547935486, + 0.00711440434679389, + 0.0029109870083630085, + 0.003868313040584326, + 0.004959566053003073, + 0.00242199981585145, + 0.008100745268166065, + 0.01175951398909092, + 0.009369087405502796, + 0.009990372695028782, + 0.013699153438210487, + 0.009392647072672844, + 0.00983919482678175, + 0.008876685053110123, + 0.014778234995901585, + 0.02926190197467804, + 0.014683772809803486, + 0.017493052408099174, + 0.018636034801602364, + 0.014019851572811604, + 0.01075842883437872, + 0.010856494307518005, + 0.017960285767912865, + 0.041892826557159424, + 0.029875604435801506, + 0.054126545786857605, + 0.020002707839012146, + 0.01845826767385006, + 0.02583116851747036, + 0.03851066529750824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.006734176073223352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro shot, taken from directly above, focuses on the inner world of a complex mechanical watch, its open, skeletonized dial revealing its intricate mechanics. A slender, mirror-polished golden seconds hand sweeps across the dial with impeccable, steady, and hypnotic smoothness, its tip tracing the brass and steel forest of mechanisms beneath with unerring precision. As the seconds hand rotates, the escapement beneath it vibrates at a visible rate, setting the interlocking layers of tiny gears in a silent and precise ballet. A soft, single-point studio light shines from the upper left, sculpting flowing, sharp highlights on the chamfered edges and brushed metal surfaces of each gear, while casting subtle, dynamic shadows within the deeper mechanisms. In the picture, the warm tones of gold and brass, the cool tones of fine steel, and the deep red artificial rubies used as bearings embellish it together form a rich and harmonious color painting. The whole scene is filled with a kind of concentration and tranquil atmosphere of awe-inspiring respect for the ultimate craftsmanship, timeless design and mechanical beauty.", + "seed": 1916, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_13/index_8216_seed_1916_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10107256732881069, + "vlm_logprob": -0.10107256732881069, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.10107256732881069, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.053275030106306076 + ], + "opposite_logprob_tokens": [ + -3.0532751083374023 + ], + "logprob_reward": -0.053275030106306076, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.053275030106306076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06938305497169495, + 0.15670333802700043, + 0.06744727492332458, + 0.06486482173204422, + 0.04817511513829231, + 0.03187283128499985, + 0.03340701758861542, + 0.3345881998538971, + 0.058717381209135056, + 0.1348409354686737 + ], + "credit_map": [ + 0.006310742348432541, + 0.010600213892757893, + 0.0065451511181890965, + 0.007132872473448515, + 0.0074263582937419415, + 0.010947568342089653, + 0.010013904422521591, + 0.010406246408820152, + 0.011315403506159782, + 0.01009413879364729, + 0.009223572909832, + 0.012588322162628174, + 0.010989133268594742, + 0.019765028730034828, + 0.05080205574631691, + 0.03192568197846413, + 0.0064272088930010796, + 0.006192319560796022, + 0.007202617358416319, + 0.005916149355471134, + 0.008099991828203201, + 0.010396123863756657, + 0.01195139717310667, + 0.011261469684541225, + 0.005803447216749191, + 0.0069892872124910355, + 0.0072418199852108955, + 0.007226039655506611, + 0.005688272882252932, + 0.014897771179676056, + 0.00921458937227726, + 0.007803594693541527, + 0.004017876926809549, + 0.005222199484705925, + 0.004864228889346123, + 0.0052697970531880856, + 0.004559255205094814, + 0.011089883744716644, + 0.007159068249166012, + 0.0059928023256361485, + 0.00470972154289484, + 0.004031034652143717, + 0.0026073672343045473, + 0.002563673537224531, + 0.0030974294058978558, + 0.005064364057034254, + 0.00452785287052393, + 0.005271386355161667, + 0.0030392459593713284, + 0.00541516812518239, + 0.0038023574743419886, + 0.003104520495980978, + 0.003261581528931856, + 0.004520981572568417, + 0.00716126523911953, + 0.0031018934678286314, + 0.03670984506607056, + 0.04265156015753746, + 0.02735844999551773, + 0.029163509607315063, + 0.04111230745911598, + 0.07422640919685364, + 0.03992561995983124, + 0.04344046488404274, + 0.005862597841769457, + 0.009226006455719471, + 0.007047289051115513, + 0.006026755087077618, + 0.005577771924436092, + 0.010349422693252563, + 0.006194716785103083, + 0.008432813920080662, + 0.017857257276773453, + 0.016153713688254356, + 0.01862422190606594, + 0.012973358854651451, + 0.012301010079681873, + 0.01658755913376808, + 0.02341538481414318, + 0.016928425058722496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.053275030106306076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060963399708271 + ], + "opposite_logprob_tokens": [ + -3.8956096172332764 + ], + "logprob_reward": -0.02060963399708271, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060963399708271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22099338471889496, + 0.18162570893764496, + 0.06381665170192719, + 0.03857681155204773, + 0.03974835202097893, + 0.038549989461898804, + 0.05097224563360214, + 0.26720309257507324, + 0.04803141951560974, + 0.050482362508773804 + ], + "credit_map": [ + 0.026600439101457596, + 0.022745365276932716, + 0.01943923905491829, + 0.019366640597581863, + 0.02713846042752266, + 0.03265533596277237, + 0.045001424849033356, + 0.028046492487192154, + 0.013940992765128613, + 0.01044713705778122, + 0.01437382958829403, + 0.01326073706150055, + 0.011629962362349033, + 0.022887209430336952, + 0.047171689569950104, + 0.04791415482759476, + 0.004982377868145704, + 0.005676660221070051, + 0.0055175600573420525, + 0.004727528430521488, + 0.0063674477860331535, + 0.0069137392565608025, + 0.01881440356373787, + 0.010816930793225765, + 0.00428872462362051, + 0.00394344562664628, + 0.0047424728982150555, + 0.0033640158362686634, + 0.0043991440907120705, + 0.008500437252223492, + 0.005073804873973131, + 0.00426476402208209, + 0.005068275146186352, + 0.005720071494579315, + 0.004152753856033087, + 0.0048231990076601505, + 0.0037018228322267532, + 0.005986910313367844, + 0.005071082152426243, + 0.005224238149821758, + 0.003657338907942176, + 0.004813875537365675, + 0.004564232658594847, + 0.004004219546914101, + 0.003985227085649967, + 0.005769016221165657, + 0.005498423241078854, + 0.006257656496018171, + 0.004861262161284685, + 0.005045134574174881, + 0.007540521677583456, + 0.004631357733160257, + 0.003875202964991331, + 0.007867671549320221, + 0.008164132945239544, + 0.008986959233880043, + 0.0286009069532156, + 0.04391999915242195, + 0.025496363639831543, + 0.02575029619038105, + 0.045500852167606354, + 0.03189289942383766, + 0.03645177558064461, + 0.02958998642861843, + 0.004180592950433493, + 0.005796848330646753, + 0.005042048636823893, + 0.004321067128330469, + 0.006384475156664848, + 0.006751558277755976, + 0.008888752199709415, + 0.006666078232228756, + 0.0044528162106871605, + 0.008187644183635712, + 0.004726140759885311, + 0.005274515133351088, + 0.007759178522974253, + 0.006438806187361479, + 0.007770432159304619, + 0.005872828885912895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.02060963399708271 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900993525981903 + ], + "opposite_logprob_tokens": [ + -2.579010009765625 + ], + "logprob_reward": -0.07900993525981903, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900993525981903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19043585658073425, + 0.21876071393489838, + 0.10864149034023285, + 0.036588553339242935, + 0.07129447162151337, + 0.03879828006029129, + 0.06615854799747467, + 0.21181996166706085, + 0.027640115469694138, + 0.029862036928534508 + ], + "credit_map": [ + 0.014334147796034813, + 0.021437963470816612, + 0.014938959851861, + 0.017581116408109665, + 0.017508303746581078, + 0.02040744572877884, + 0.04394910857081413, + 0.04027881100773811, + 0.01682996191084385, + 0.013196732848882675, + 0.01740150712430477, + 0.014760483987629414, + 0.012924536131322384, + 0.020263124257326126, + 0.06475609540939331, + 0.05862826108932495, + 0.012450778856873512, + 0.00998232327401638, + 0.00822868850082159, + 0.007088012527674437, + 0.01301842276006937, + 0.01470233779400587, + 0.03124767541885376, + 0.011923255398869514, + 0.0037872677203267813, + 0.0034740588162094355, + 0.005176451988518238, + 0.0032822752837091684, + 0.003169371047988534, + 0.00795936118811369, + 0.0059564742259681225, + 0.003783295163884759, + 0.00800263974815607, + 0.008242619223892689, + 0.005266473162919283, + 0.006101164501160383, + 0.007791000884026289, + 0.013245237059891224, + 0.012568628415465355, + 0.010076706297695637, + 0.0035546545404940844, + 0.0055309138260781765, + 0.004136925563216209, + 0.00401036674156785, + 0.004116335418075323, + 0.006850859150290489, + 0.004712455905973911, + 0.005885769613087177, + 0.005459681618958712, + 0.006357818841934204, + 0.009477696381509304, + 0.005691325757652521, + 0.006213986314833164, + 0.008201254531741142, + 0.014094319194555283, + 0.010662468150258064, + 0.023535236716270447, + 0.02906164899468422, + 0.018670309334993362, + 0.01599840819835663, + 0.024132121354341507, + 0.034295566380023956, + 0.0438295342028141, + 0.02229713462293148, + 0.0027386953588575125, + 0.003078032284975052, + 0.0040538134053349495, + 0.0022019417956471443, + 0.0032862513326108456, + 0.0029590569902211428, + 0.004612530115991831, + 0.004709793720394373, + 0.003977365326136351, + 0.003891719039529562, + 0.0026976896915584803, + 0.002052606549113989, + 0.0025331471115350723, + 0.003966576419770718, + 0.003681587753817439, + 0.007061345968395472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.07900993525981903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033954679965973 + ], + "opposite_logprob_tokens": [ + -2.350339651107788 + ], + "logprob_reward": -0.10033954679965973, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033954679965973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06426316499710083, + 0.11994786560535431, + 0.18286985158920288, + 0.06637189537286758, + 0.1853816658258438, + 0.10533620417118073, + 0.06238590180873871, + 0.08808667957782745, + 0.05381302908062935, + 0.07154373824596405 + ], + "credit_map": [ + 0.00859924964606762, + 0.00915475096553564, + 0.006970565300434828, + 0.005979262292385101, + 0.007110733538866043, + 0.009145352058112621, + 0.00755878584459424, + 0.009744465351104736, + 0.012491817586123943, + 0.010000528767704964, + 0.007915286347270012, + 0.011232956312596798, + 0.012356029823422432, + 0.017414048314094543, + 0.029945362359285355, + 0.01859183982014656, + 0.013336963020265102, + 0.014652584679424763, + 0.013988812454044819, + 0.013787752948701382, + 0.017952091991901398, + 0.02294708415865898, + 0.052071865648031235, + 0.0341326966881752, + 0.00634436821565032, + 0.008504477329552174, + 0.007519387640058994, + 0.006436946336179972, + 0.007698343135416508, + 0.00836955476552248, + 0.013816475868225098, + 0.007682343479245901, + 0.018617838621139526, + 0.025605183094739914, + 0.01626736856997013, + 0.017709562554955482, + 0.025895535945892334, + 0.03193430230021477, + 0.02410099469125271, + 0.025250878185033798, + 0.009972633793950081, + 0.012117082253098488, + 0.01583007723093033, + 0.009439398534595966, + 0.010445800609886646, + 0.022455280646681786, + 0.012714878655970097, + 0.012361052446067333, + 0.005401745438575745, + 0.007246456574648619, + 0.008195752277970314, + 0.006912452634423971, + 0.005448319483548403, + 0.011531643569469452, + 0.010646114125847816, + 0.007003421895205975, + 0.010466577485203743, + 0.00962747074663639, + 0.0069999429397284985, + 0.007411569356918335, + 0.00891250278800726, + 0.015705350786447525, + 0.018069401383399963, + 0.010893862694501877, + 0.003945710137486458, + 0.006239698734134436, + 0.006350553594529629, + 0.0048926351591944695, + 0.004147846717387438, + 0.009932469576597214, + 0.012706300243735313, + 0.0055978125892579556, + 0.00865496788173914, + 0.00852690264582634, + 0.008067615330219269, + 0.007462638430297375, + 0.006654323544353247, + 0.008813975378870964, + 0.01430946122854948, + 0.009053854271769524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.10033954679965973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521286904811859 + ], + "opposite_logprob_tokens": [ + -1.5021287202835083 + ], + "logprob_reward": -0.2521286904811859, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521286904811859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19259341061115265, + 0.09252025932073593, + 0.10445355623960495, + 0.08946163952350616, + 0.17000411450862885, + 0.049324534833431244, + 0.0526624396443367, + 0.18299508094787598, + 0.02065957337617874, + 0.0453253909945488 + ], + "credit_map": [ + 0.019414786249399185, + 0.0193193219602108, + 0.01605379395186901, + 0.017955275252461433, + 0.016514355316758156, + 0.033560797572135925, + 0.03950619697570801, + 0.03026888333261013, + 0.008745762519538403, + 0.007930676452815533, + 0.008965043351054192, + 0.00869016908109188, + 0.009009587578475475, + 0.015823377296328545, + 0.015787161886692047, + 0.017568480223417282, + 0.0074594756588339806, + 0.011226622387766838, + 0.013723057694733143, + 0.009926974773406982, + 0.00993553176522255, + 0.018608473241329193, + 0.019680066034197807, + 0.013893348164856434, + 0.00867585651576519, + 0.007621303666383028, + 0.012725874781608582, + 0.007944121025502682, + 0.007242333609610796, + 0.015508249402046204, + 0.015239826403558254, + 0.014504080638289452, + 0.017782652750611305, + 0.02366904355585575, + 0.015966705977916718, + 0.014525158330798149, + 0.024119162932038307, + 0.02808152139186859, + 0.027311718091368675, + 0.018548157066106796, + 0.0057482533156871796, + 0.0064259907230734825, + 0.004894522950053215, + 0.0038993600755929947, + 0.00710903899744153, + 0.008089111186563969, + 0.006755846086889505, + 0.006402411963790655, + 0.004749204497784376, + 0.00506941182538867, + 0.007043641060590744, + 0.004302302375435829, + 0.005674911662936211, + 0.007286116946488619, + 0.008837290108203888, + 0.009699561633169651, + 0.01982268877327442, + 0.03294796869158745, + 0.018579065799713135, + 0.014918050728738308, + 0.028363512828946114, + 0.01846715249121189, + 0.03182429447770119, + 0.0180723387748003, + 0.0023039246443659067, + 0.0019313417142257094, + 0.00282422942109406, + 0.0017965325387194753, + 0.001775955781340599, + 0.003119284752756357, + 0.004104360938072205, + 0.002803943119943142, + 0.0060461764223873615, + 0.0052831582725048065, + 0.004988537169992924, + 0.0035225162282586098, + 0.004875281360000372, + 0.006872374564409256, + 0.007529859431087971, + 0.006207488477230072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.2521286904811859 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up macro shot, taken from directly above, focuses on the inner world of a complex mechanical watch, its open, skeletonized dial revealing its intricate mechanics. A slender, mirror-polished golden seconds hand sweeps across the dial with impeccable, steady, and hypnotic smoothness, its tip tracing the brass and steel forest of mechanisms beneath with unerring precision. As the seconds hand rotates, the escapement beneath it vibrates at a visible rate, setting the interlocking layers of tiny gears in a silent and precise ballet. A soft, single-point studio light shines from the upper left, sculpting flowing, sharp highlights on the chamfered edges and brushed metal surfaces of each gear, while casting subtle, dynamic shadows within the deeper mechanisms. In the picture, the warm tones of gold and brass, the cool tones of fine steel, and the deep red artificial rubies used as bearings embellish it together form a rich and harmonious color painting. The whole scene is filled with a kind of concentration and tranquil atmosphere of awe-inspiring respect for the ultimate craftsmanship, timeless design and mechanical beauty.", + "seed": 1917, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_13/index_8216_seed_1917_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06181558310054243, + "vlm_logprob": -0.06181558310054243, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.06181558310054243, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05280112102627754 + ], + "opposite_logprob_tokens": [ + -3.1778011322021484 + ], + "logprob_reward": -0.05280112102627754, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05280112102627754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2543593645095825, + 0.04025598615407944, + 0.048817284405231476, + 0.03538680076599121, + 0.3534260094165802, + 0.07096755504608154, + 0.09346392750740051, + 0.07581951469182968, + 0.017004232853651047, + 0.010499357245862484 + ], + "credit_map": [ + 0.018056195229291916, + 0.016221610829234123, + 0.012947971932590008, + 0.05690765380859375, + 0.0997009426355362, + 0.014768718741834164, + 0.020817359909415245, + 0.01493889931589365, + 0.005278623662889004, + 0.005239943973720074, + 0.006787099875509739, + 0.004495853558182716, + 0.0058614639565348625, + 0.0031420374289155006, + 0.005266756750643253, + 0.004184211138635874, + 0.008158739656209946, + 0.0068246484734117985, + 0.0066596828401088715, + 0.006369247101247311, + 0.005912390537559986, + 0.0040944903157651424, + 0.006361190229654312, + 0.004436897113919258, + 0.003670342965051532, + 0.005161740351468325, + 0.0037427162751555443, + 0.004217606503516436, + 0.008806146681308746, + 0.002677662530913949, + 0.004577556159347296, + 0.002533030929043889, + 0.020786141976714134, + 0.023059098049998283, + 0.005616925656795502, + 0.018544266000390053, + 0.1568785160779953, + 0.07606105506420135, + 0.035457294434309006, + 0.01702268421649933, + 0.0077752042561769485, + 0.01196309644728899, + 0.010733530856668949, + 0.007148962002247572, + 0.010246903635561466, + 0.008062160573899746, + 0.00917007215321064, + 0.005867628380656242, + 0.014849495142698288, + 0.003641766030341387, + 0.010360898450016975, + 0.003980222158133984, + 0.03187450021505356, + 0.020959898829460144, + 0.004290791694074869, + 0.003506358014419675, + 0.006614047568291426, + 0.0059470501728355885, + 0.005283263046294451, + 0.0062074256129562855, + 0.014425013214349747, + 0.012963409535586834, + 0.016770044341683388, + 0.0076092625968158245, + 0.0015115913702175021, + 0.0014694301644340158, + 0.0014818086056038737, + 0.0018440753920003772, + 0.004671537317335606, + 0.0023762749042361975, + 0.002240540459752083, + 0.001408973941579461, + 0.0010544483084231615, + 0.0014970236225053668, + 0.001135673956014216, + 0.0016362356254830956, + 0.0015576095320284367, + 0.0012242314405739307, + 0.0015286445850506425, + 0.0008654898265376687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.05280112102627754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025173898320645094 + ], + "opposite_logprob_tokens": [ + -6.002517223358154 + ], + "logprob_reward": -0.0025173898320645094, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025173898320645094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20568327605724335, + 0.05649903789162636, + 0.07527631521224976, + 0.044711265712976456, + 0.3791367709636688, + 0.022896381095051765, + 0.12911787629127502, + 0.03763243928551674, + 0.028853101655840874, + 0.020193448290228844 + ], + "credit_map": [ + 0.015224961563944817, + 0.013249466195702553, + 0.008898172527551651, + 0.043972063809633255, + 0.08150403946638107, + 0.011444213800132275, + 0.015097152441740036, + 0.016293218359351158, + 0.00799719151109457, + 0.0066103748977184296, + 0.009546667337417603, + 0.007288574241101742, + 0.01069752685725689, + 0.0043512918055057526, + 0.005885131191462278, + 0.004122286103665829, + 0.01160664763301611, + 0.009183336049318314, + 0.01158173382282257, + 0.010605084709823132, + 0.008177635259926319, + 0.0064126914367079735, + 0.012425505556166172, + 0.005283687729388475, + 0.003406325588002801, + 0.0037252360489219427, + 0.00235251197591424, + 0.007403338793665171, + 0.016074664890766144, + 0.004083224572241306, + 0.004678994417190552, + 0.0029869757127016783, + 0.022298987954854965, + 0.025288891047239304, + 0.008021759800612926, + 0.017822109162807465, + 0.16655920445919037, + 0.08220912516117096, + 0.03770673647522926, + 0.019230004400014877, + 0.002289803931489587, + 0.0032769928220659494, + 0.0028447932563722134, + 0.0022661592811346054, + 0.004185731988400221, + 0.002654082141816616, + 0.0037652053870260715, + 0.0016136140329763293, + 0.020873861387372017, + 0.0053570084273815155, + 0.011937729083001614, + 0.0050779907032847404, + 0.04694763571023941, + 0.028300078585743904, + 0.00599999725818634, + 0.004623591434210539, + 0.003332654945552349, + 0.0030220854096114635, + 0.0025164601393043995, + 0.003514413023367524, + 0.00706910016015172, + 0.004790926817804575, + 0.00867735780775547, + 0.004709446337074041, + 0.003996918443590403, + 0.002106452826410532, + 0.0024857267271727324, + 0.0027870098128914833, + 0.006586199160665274, + 0.005710172932595015, + 0.0033358531072735786, + 0.0018447715556249022, + 0.003348884405568242, + 0.002789280842989683, + 0.0032336220610886812, + 0.0027301760856062174, + 0.0020407531410455704, + 0.0017439029179513454, + 0.0022724010050296783, + 0.002034429693594575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.0025173898320645094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637290023267269 + ], + "opposite_logprob_tokens": [ + -3.6513729095458984 + ], + "logprob_reward": -0.02637290023267269, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637290023267269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09018087387084961, + 0.061928387731313705, + 0.08763982355594635, + 0.04134165868163109, + 0.4200941324234009, + 0.039411768317222595, + 0.18623095750808716, + 0.03340751677751541, + 0.028700832277536392, + 0.011064105667173862 + ], + "credit_map": [ + 0.010349792428314686, + 0.007499780505895615, + 0.01317379716783762, + 0.009241593070328236, + 0.025845956057310104, + 0.005173855926841497, + 0.008398711681365967, + 0.01049739122390747, + 0.005804084707051516, + 0.007710705511271954, + 0.008735449984669685, + 0.013778314925730228, + 0.00885392352938652, + 0.004705285653471947, + 0.007838609628379345, + 0.004502005875110626, + 0.009556245990097523, + 0.008879401721060276, + 0.013739402405917645, + 0.020695336163043976, + 0.010384849272668362, + 0.0059266346506774426, + 0.012073449790477753, + 0.006384498439729214, + 0.004063191823661327, + 0.003931528888642788, + 0.003786917310208082, + 0.007944352924823761, + 0.009379950352013111, + 0.0027873541694134474, + 0.005883951671421528, + 0.0035644040908664465, + 0.024580568075180054, + 0.02813534438610077, + 0.009481731802225113, + 0.019068049266934395, + 0.1833500862121582, + 0.09281700104475021, + 0.04110681265592575, + 0.021554486826062202, + 0.005489862523972988, + 0.0057653505355119705, + 0.004628672264516354, + 0.004087410867214203, + 0.006210935767740011, + 0.005069132428616285, + 0.004616997204720974, + 0.003543404396623373, + 0.015306689776480198, + 0.011087514460086823, + 0.005208549089729786, + 0.008330317214131355, + 0.08347151428461075, + 0.0451185368001461, + 0.010719715617597103, + 0.00698809651657939, + 0.0033122936729341745, + 0.0029782550409436226, + 0.0021262115333229303, + 0.003154521342366934, + 0.0075043137185275555, + 0.005000886972993612, + 0.006686984095722437, + 0.0026440497022122145, + 0.004007864277809858, + 0.0027969132643193007, + 0.003035191213712096, + 0.00289639038965106, + 0.006784599740058184, + 0.003670718753710389, + 0.0033933373633772135, + 0.0021158147137612104, + 0.0014970271149650216, + 0.0011237810831516981, + 0.001659429632127285, + 0.0013856381410732865, + 0.001743211643770337, + 0.001167554990388453, + 0.001695536426268518, + 0.0007919264025986195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.02637290023267269 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033781826496124 + ], + "opposite_logprob_tokens": [ + -2.3503377437591553 + ], + "logprob_reward": -0.10033781826496124, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033781826496124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1759721338748932, + 0.030750202015042305, + 0.026043931022286415, + 0.08850890398025513, + 0.28067126870155334, + 0.08375148475170135, + 0.22385123372077942, + 0.046723805367946625, + 0.03388479724526405, + 0.00984222162514925 + ], + "credit_map": [ + 0.012716061435639858, + 0.01137425284832716, + 0.007816839963197708, + 0.03742978721857071, + 0.07118875533342361, + 0.010279513895511627, + 0.012621084228157997, + 0.012545841746032238, + 0.00349612720310688, + 0.004354102537035942, + 0.00544294947758317, + 0.00458176014944911, + 0.003925816155970097, + 0.003115077270194888, + 0.003233850235119462, + 0.002600517589598894, + 0.002835002029314637, + 0.0034530959092080593, + 0.004085774999111891, + 0.0042615365236997604, + 0.0034922068007290363, + 0.0023093800991773605, + 0.0032270378433167934, + 0.002379897516220808, + 0.007696067448705435, + 0.010530954226851463, + 0.00744685810059309, + 0.01460967306047678, + 0.027962949126958847, + 0.007162508554756641, + 0.008054886013269424, + 0.005045007914304733, + 0.01606179215013981, + 0.018228227272629738, + 0.005309283267706633, + 0.014214182272553444, + 0.12334034591913223, + 0.06249554082751274, + 0.02687377855181694, + 0.014148135669529438, + 0.012064208276569843, + 0.01330607756972313, + 0.010765437968075275, + 0.008009477518498898, + 0.00967105757445097, + 0.011670218780636787, + 0.010091579519212246, + 0.008173429407179356, + 0.011453651823103428, + 0.012311826460063457, + 0.009480565786361694, + 0.01031221728771925, + 0.09481719136238098, + 0.06411300599575043, + 0.012170016765594482, + 0.009192739613354206, + 0.0039268299005925655, + 0.003478417871519923, + 0.0028499627951532602, + 0.004118107724934816, + 0.009032348170876503, + 0.0077024842612445354, + 0.010891425423324108, + 0.004724229220300913, + 0.002229084726423025, + 0.0021595892030745745, + 0.0024870517663657665, + 0.0033174108248203993, + 0.012607686221599579, + 0.00600635539740324, + 0.002911454299464822, + 0.002166166203096509, + 0.0011935390066355467, + 0.0009054961265064776, + 0.0009781603002920747, + 0.001291936612688005, + 0.001742538413964212, + 0.0009609758271835744, + 0.0017888149013742805, + 0.0009807602036744356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.10033781826496124 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view taken from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the watch dial open and skeletonized to reveal intricate internal mechanics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, mirror-polished golden seconds hand sweeping smoothly across the dial?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a soft, single-point studio light shining from the upper left creating sharp highlights and dynamic shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible deep red artificial rubies used as bearings within the gear mechanisms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704868614673615 + ], + "opposite_logprob_tokens": [ + -2.1270487308502197 + ], + "logprob_reward": -0.12704868614673615, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704868614673615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21498149633407593, + 0.03613993152976036, + 0.04328346997499466, + 0.0638403370976448, + 0.3879825472831726, + 0.08911316096782684, + 0.11632271856069565, + 0.0167083740234375, + 0.017608879134058952, + 0.014019101858139038 + ], + "credit_map": [ + 0.011944732628762722, + 0.013687764294445515, + 0.020670432597398758, + 0.052721016108989716, + 0.06438136100769043, + 0.010249472223222256, + 0.027728402987122536, + 0.013598302379250526, + 0.00403011217713356, + 0.0049454825930297375, + 0.00462397700175643, + 0.004810419399291277, + 0.005208312068134546, + 0.0033087246119976044, + 0.0060033854097127914, + 0.003209518501535058, + 0.005121223162859678, + 0.00396090978756547, + 0.006517962086945772, + 0.004806907381862402, + 0.00663419533520937, + 0.003245503408834338, + 0.00864705815911293, + 0.004349706694483757, + 0.00435563363134861, + 0.004728975705802441, + 0.004436157178133726, + 0.012933461926877499, + 0.021899838000535965, + 0.005075206980109215, + 0.0066786278039216995, + 0.0037324358709156513, + 0.02348613366484642, + 0.026362573727965355, + 0.0073168836534023285, + 0.023448143154382706, + 0.17207081615924835, + 0.07727605849504471, + 0.03936328738927841, + 0.01865866221487522, + 0.01082474086433649, + 0.008397353813052177, + 0.007550734560936689, + 0.007827582769095898, + 0.01753365807235241, + 0.008137122727930546, + 0.02070510946214199, + 0.008136856369674206, + 0.041375428438186646, + 0.005764979403465986, + 0.03013862483203411, + 0.005802906583994627, + 0.014349845238029957, + 0.007964550517499447, + 0.005203198175877333, + 0.005723182111978531, + 0.0017372245201840997, + 0.0013585954438894987, + 0.0013587322318926454, + 0.0015030952636152506, + 0.003288265550509095, + 0.0027383030392229557, + 0.0031545842066407204, + 0.0015695741167291999, + 0.001282897312194109, + 0.0015027892077341676, + 0.0016180407255887985, + 0.001945106196217239, + 0.004598540253937244, + 0.0030440306290984154, + 0.0023485571146011353, + 0.0012689182767644525, + 0.002118134405463934, + 0.0018212655559182167, + 0.0020270301029086113, + 0.0016638340894132853, + 0.0019655737560242414, + 0.0013242206769064069, + 0.0016953747253865004, + 0.0014036688953638077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.12704868614673615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_14/index_tensor([7602, 7602])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_14/index_tensor([7602, 7602])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2655d3c1228b59295a657b7bd188f3f1ad5d0861 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_14/index_tensor([7602, 7602])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro close-up, captured from a slightly elevated angle and slowly panning forward, focuses on an ant exploring a large, vibrant green leaf. The ant's glossy, dark brown exoskeleton reflects delicately in the soft morning light, its segmented body supported by slender, powerful legs. Its two slender, jointed antennae are the dynamic focal point of the image, quivering and swaying, tapping and sweeping across the leaf with a curious air, as if meticulously interpreting the scents and textures of its surroundings. The leaf itself resembles a miniature landscape, its intricate network of veins forming ridges and valleys, and a few glistening dewdrops clinging to its surface like tiny gems, catching and refracting the light. Soft, diffuse sunlight penetrates the leaf's translucent structure, bathing the scene in a tranquil, almost ethereal green aura, dominated by vibrant emerald greens and earthy browns. The ant moves cautiously along the \"highway\" formed by the leaf veins at a steady and leisurely pace. Its every movement is full of purpose, creating an atmosphere of focus, tranquility and the miracle of life.", + "seed": 6621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_14/index_7602_seed_6621_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0658965215086937, + "vlm_logprob": -0.0658965215086937, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.0658965215086937, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0064351484179496765 + ], + "opposite_logprob_tokens": [ + -6.756434917449951 + ], + "logprob_reward": -0.0064351484179496765, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0064351484179496765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1848689466714859, + 0.056983355432748795, + 0.036975935101509094, + 0.026684382930397987, + 0.022223979234695435, + 0.053322046995162964, + 0.0417574942111969, + 0.4099084436893463, + 0.050763681530952454, + 0.1165117397904396 + ], + "credit_map": [ + 0.01463689561933279, + 0.008824975229799747, + 0.014210861176252365, + 0.015656335279345512, + 0.010487640276551247, + 0.04167681559920311, + 0.01710912398993969, + 0.06226630508899689, + 0.006154984701424837, + 0.009739190340042114, + 0.005845215637236834, + 0.009551172144711018, + 0.0056612673215568066, + 0.005870738532394171, + 0.005428468808531761, + 0.008732316084206104, + 0.004042061511427164, + 0.006305955350399017, + 0.00400141067802906, + 0.0038230507634580135, + 0.004275278188288212, + 0.00595671683549881, + 0.0034072762355208397, + 0.0051641883328557014, + 0.0037015159614384174, + 0.0037791887298226357, + 0.0033943422604352236, + 0.004813218954950571, + 0.002176667330786586, + 0.0029348055832087994, + 0.0024946138728410006, + 0.0033900286071002483, + 0.0022640335373580456, + 0.002289028838276863, + 0.003302115947008133, + 0.002086053602397442, + 0.001628258964046836, + 0.003528063651174307, + 0.002754931803792715, + 0.004371493589133024, + 0.0036713434383273125, + 0.005499725230038166, + 0.013058006763458252, + 0.005196350626647472, + 0.003670042846351862, + 0.013839014805853367, + 0.004160439129918814, + 0.004227121360599995, + 0.005208882037550211, + 0.005711669567972422, + 0.007834464311599731, + 0.004678426776081324, + 0.003794937627390027, + 0.004749661311507225, + 0.004167263396084309, + 0.0056121861562132835, + 0.011563010513782501, + 0.010227627120912075, + 0.01950261928141117, + 0.007508501410484314, + 0.16980773210525513, + 0.15370319783687592, + 0.012598913162946701, + 0.02499685250222683, + 0.004111230373382568, + 0.00684925215318799, + 0.007708145771175623, + 0.007398975547403097, + 0.005530687980353832, + 0.004435822367668152, + 0.010061553679406643, + 0.004668015521019697, + 0.009448458440601826, + 0.00644725514575839, + 0.010107414796948433, + 0.006520500406622887, + 0.02412671595811844, + 0.02953108213841915, + 0.007336490321904421, + 0.022993825376033783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0064351484179496765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820661872625351 + ], + "opposite_logprob_tokens": [ + -4.018206596374512 + ], + "logprob_reward": -0.01820661872625351, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820661872625351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25211745500564575, + 0.15382520854473114, + 0.05329865217208862, + 0.024697043001651764, + 0.053688161075115204, + 0.06973214447498322, + 0.12680339813232422, + 0.09893058240413666, + 0.06458314508199692, + 0.1023242175579071 + ], + "credit_map": [ + 0.017685847356915474, + 0.03567245602607727, + 0.016856590285897255, + 0.02137908712029457, + 0.015602781437337399, + 0.06601430475711823, + 0.019764836877584457, + 0.059141553938388824, + 0.012050811201334, + 0.030022675171494484, + 0.011546322144567966, + 0.023628827184438705, + 0.01593484729528427, + 0.018571408465504646, + 0.01438616868108511, + 0.027684148401021957, + 0.0054176561534404755, + 0.007133014965802431, + 0.005908701103180647, + 0.005598349962383509, + 0.0052462294697761536, + 0.00847636628895998, + 0.005692601669579744, + 0.009825735352933407, + 0.003291610162705183, + 0.004738889168947935, + 0.0025079769548028708, + 0.0030278372578322887, + 0.002129859523847699, + 0.002389888046309352, + 0.0038010170683264732, + 0.0028099638875573874, + 0.005239263642579317, + 0.009503636509180069, + 0.005439048632979393, + 0.007611843291670084, + 0.004148527979850769, + 0.008124856278300285, + 0.007912798784673214, + 0.005708185490220785, + 0.005318707320839167, + 0.006688994355499744, + 0.014832834713160992, + 0.005756066180765629, + 0.005381402559578419, + 0.0172747690230608, + 0.004990106448531151, + 0.009489263407886028, + 0.013482213020324707, + 0.013716962188482285, + 0.032023802399635315, + 0.013181008398532867, + 0.01205296628177166, + 0.013834563083946705, + 0.010391218587756157, + 0.01812065951526165, + 0.007434355095028877, + 0.00996004231274128, + 0.01376576442271471, + 0.011121582239866257, + 0.02704363875091076, + 0.010189211927354336, + 0.009421887807548046, + 0.009994094260036945, + 0.006628140341490507, + 0.011615665629506111, + 0.007312566041946411, + 0.008376957848668098, + 0.007044616155326366, + 0.006463692989200354, + 0.010532419197261333, + 0.006609086412936449, + 0.008280646055936813, + 0.007528873160481453, + 0.009118825197219849, + 0.00725714024156332, + 0.025195861235260963, + 0.027725309133529663, + 0.006414979230612516, + 0.010802585631608963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.01820661872625351 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898668944835663 + ], + "opposite_logprob_tokens": [ + -2.578986644744873 + ], + "logprob_reward": -0.07898668944835663, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898668944835663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15596207976341248, + 0.08607354760169983, + 0.05305849015712738, + 0.0255882628262043, + 0.02973233535885811, + 0.17164744436740875, + 0.3367793560028076, + 0.03382081910967827, + 0.03004944883286953, + 0.0772881954908371 + ], + "credit_map": [ + 0.011078923009335995, + 0.007515545003116131, + 0.010631158016622066, + 0.010621960274875164, + 0.009308453649282455, + 0.04738623648881912, + 0.011715405620634556, + 0.047704409807920456, + 0.011312401853501797, + 0.009332728572189808, + 0.014443784020841122, + 0.012799885123968124, + 0.00651132594794035, + 0.008963963948190212, + 0.012921634130179882, + 0.009787820279598236, + 0.0058089555241167545, + 0.005168573930859566, + 0.006152126006782055, + 0.006410898175090551, + 0.004001064226031303, + 0.0053090485744178295, + 0.007276872638612986, + 0.012930947355926037, + 0.0036300215870141983, + 0.0030547643546015024, + 0.0036785139236599207, + 0.0038505741395056248, + 0.0016598394140601158, + 0.0025372514501214027, + 0.003619641065597534, + 0.0035576559603214264, + 0.0024876121897250414, + 0.0028509562835097313, + 0.003003406571224332, + 0.004420964512974024, + 0.0017750898841768503, + 0.006633039563894272, + 0.0038753512781113386, + 0.004685915540903807, + 0.012440361082553864, + 0.00629926472902298, + 0.05128674954175949, + 0.013188262470066547, + 0.007957342080771923, + 0.056892674416303635, + 0.007942887023091316, + 0.015639903023838997, + 0.034455519169569016, + 0.036828164011240005, + 0.08490674942731857, + 0.037232041358947754, + 0.030156796798110008, + 0.035965610295534134, + 0.02933778427541256, + 0.04789671301841736, + 0.0019996732007712126, + 0.0030560865998268127, + 0.0023794216103851795, + 0.003575849812477827, + 0.002391835441812873, + 0.008630442433059216, + 0.0025744549930095673, + 0.009213052690029144, + 0.0021550902165472507, + 0.0036318039055913687, + 0.004648979753255844, + 0.004292437341064215, + 0.0038805678486824036, + 0.0033276418689638376, + 0.004795747343450785, + 0.003317181719467044, + 0.006445284467190504, + 0.005105423741042614, + 0.008009865880012512, + 0.005122835747897625, + 0.0189277995377779, + 0.022296953946352005, + 0.00494128605350852, + 0.006438750308007002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.07898668944835663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129259541630745 + ], + "opposite_logprob_tokens": [ + -2.2379260063171387 + ], + "logprob_reward": -0.1129259541630745, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129259541630745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0716102123260498, + 0.04455415531992912, + 0.04373561590909958, + 0.03230676427483559, + 0.019502362236380577, + 0.09577812999486923, + 0.08435395359992981, + 0.494804322719574, + 0.024447619915008545, + 0.08890687674283981 + ], + "credit_map": [ + 0.006359456107020378, + 0.01051313616335392, + 0.007243185769766569, + 0.00834211241453886, + 0.005219197366386652, + 0.010101071558892727, + 0.006377023179084063, + 0.017455024644732475, + 0.004711646120995283, + 0.010953524149954319, + 0.006456906441599131, + 0.004193536937236786, + 0.0041763661429286, + 0.004404551349580288, + 0.004437734372913837, + 0.005219888407737017, + 0.004807697143405676, + 0.006814848631620407, + 0.005875939503312111, + 0.0051358588971197605, + 0.004113852512091398, + 0.00596208218485117, + 0.004987475462257862, + 0.006037855055183172, + 0.005153816659003496, + 0.006030594930052757, + 0.0037032184191048145, + 0.004037347622215748, + 0.0028198109939694405, + 0.0033963918685913086, + 0.003175790188834071, + 0.003989789634943008, + 0.0017277796287089586, + 0.004235540516674519, + 0.0016448855167254806, + 0.0018875952810049057, + 0.0015001001302152872, + 0.0034697235096246004, + 0.0018545060884207487, + 0.0031822286546230316, + 0.006221602205187082, + 0.00923647079616785, + 0.025965716689825058, + 0.008147208020091057, + 0.005055639427155256, + 0.02783139795064926, + 0.005908941384404898, + 0.007411136757582426, + 0.009493859484791756, + 0.009340919554233551, + 0.020040836185216904, + 0.009383783675730228, + 0.006995752919465303, + 0.009863554500043392, + 0.008847344666719437, + 0.010387897491455078, + 0.011991500854492188, + 0.014468876644968987, + 0.02221783623099327, + 0.013130929321050644, + 0.18112492561340332, + 0.1905575394630432, + 0.01359292771667242, + 0.047719746828079224, + 0.002318485639989376, + 0.0039155809208750725, + 0.0033850756008177996, + 0.0030590363312512636, + 0.00246278103441, + 0.002465017605572939, + 0.004116244148463011, + 0.0027253974694758654, + 0.006802656687796116, + 0.005986337084323168, + 0.008531376719474792, + 0.005671374034136534, + 0.023661019280552864, + 0.025775834918022156, + 0.005054335575550795, + 0.007423939183354378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.1129259541630745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292819678783417 + ], + "opposite_logprob_tokens": [ + -2.2379281520843506 + ], + "logprob_reward": -0.11292819678783417, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292819678783417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07645060122013092, + 0.020623238757252693, + 0.05571790039539337, + 0.027062347158789635, + 0.025363577529788017, + 0.03267933055758476, + 0.04584367573261261, + 0.6144334077835083, + 0.03267359733581543, + 0.0691523626446724 + ], + "credit_map": [ + 0.007066785357892513, + 0.008673868142068386, + 0.006129640154540539, + 0.008391406387090683, + 0.004654652439057827, + 0.015040025115013123, + 0.006113732233643532, + 0.02038048580288887, + 0.0018757064826786518, + 0.004191763699054718, + 0.0023264691699296236, + 0.0021803760901093483, + 0.002096143551170826, + 0.0020095473155379295, + 0.0026963497512042522, + 0.0032468819990754128, + 0.005042592063546181, + 0.004800952505320311, + 0.006522659212350845, + 0.007382087875157595, + 0.004297344014048576, + 0.006835952866822481, + 0.0070843929424881935, + 0.013751916587352753, + 0.0035156751982867718, + 0.003743034088984132, + 0.0032188191544264555, + 0.003572317771613598, + 0.0033556572161614895, + 0.0028308751061558723, + 0.0028895677533000708, + 0.003936401102691889, + 0.0028309947811067104, + 0.005524253007024527, + 0.0024513015523552895, + 0.003800251055508852, + 0.002415756229311228, + 0.0025678058154881, + 0.0033566951751708984, + 0.0024165206123143435, + 0.0032721671741455793, + 0.004835254047065973, + 0.006036151200532913, + 0.0036028060130774975, + 0.002356664277613163, + 0.006023323629051447, + 0.0035138698294758797, + 0.0030390964820981026, + 0.004974666517227888, + 0.004891278687864542, + 0.011186535470187664, + 0.004919050727039576, + 0.0037981518544256687, + 0.004607367794960737, + 0.0048923795111477375, + 0.006574247032403946, + 0.014310061000287533, + 0.016877470538020134, + 0.02646433562040329, + 0.013312109746038914, + 0.2250334918498993, + 0.23992450535297394, + 0.016054579988121986, + 0.06245686486363411, + 0.002506641671061516, + 0.004129764623939991, + 0.00397248100489378, + 0.0045785242691636086, + 0.004841678310185671, + 0.0038860624190419912, + 0.005134614650160074, + 0.0036238268949091434, + 0.005550050642341375, + 0.0044470904394984245, + 0.007315382361412048, + 0.004263792186975479, + 0.017320575192570686, + 0.019596602767705917, + 0.0041175829246640205, + 0.006541285198181868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.11292819678783417 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, captured from a slightly elevated angle and slowly panning forward, focuses on an ant exploring a large, vibrant green leaf. The ant's glossy, dark brown exoskeleton reflects delicately in the soft morning light, its segmented body supported by slender, powerful legs. Its two slender, jointed antennae are the dynamic focal point of the image, quivering and swaying, tapping and sweeping across the leaf with a curious air, as if meticulously interpreting the scents and textures of its surroundings. The leaf itself resembles a miniature landscape, its intricate network of veins forming ridges and valleys, and a few glistening dewdrops clinging to its surface like tiny gems, catching and refracting the light. Soft, diffuse sunlight penetrates the leaf's translucent structure, bathing the scene in a tranquil, almost ethereal green aura, dominated by vibrant emerald greens and earthy browns. The ant moves cautiously along the \"highway\" formed by the leaf veins at a steady and leisurely pace. Its every movement is full of purpose, creating an atmosphere of focus, tranquility and the miracle of life.", + "seed": 6622, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_14/index_7602_seed_6622_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05016691200435162, + "vlm_logprob": -0.05016691200435162, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.05016691200435162, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00568072684109211 + ], + "opposite_logprob_tokens": [ + -6.880680561065674 + ], + "logprob_reward": -0.00568072684109211, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.00568072684109211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13599741458892822, + 0.06745794415473938, + 0.03511567786335945, + 0.03417166694998741, + 0.037071093916893005, + 0.21068482100963593, + 0.2687935531139374, + 0.12204551696777344, + 0.04834992066025734, + 0.04031239449977875 + ], + "credit_map": [ + 0.02510906755924225, + 0.011444276198744774, + 0.010492539964616299, + 0.018169712275266647, + 0.011305874213576317, + 0.01112289261072874, + 0.01806809939444065, + 0.030284954234957695, + 0.013053487986326218, + 0.004489808343350887, + 0.006221455056220293, + 0.006721640471369028, + 0.004688838962465525, + 0.005391873884946108, + 0.021459195762872696, + 0.005431637167930603, + 0.004974952898919582, + 0.004281565546989441, + 0.004159990232437849, + 0.005149184726178646, + 0.002826792187988758, + 0.004950810223817825, + 0.005337132140994072, + 0.0034352487418800592, + 0.0038781692273914814, + 0.0052841645665466785, + 0.003400220302864909, + 0.005638273898512125, + 0.003292773850262165, + 0.004393809009343386, + 0.004364083521068096, + 0.003920173738151789, + 0.005566083360463381, + 0.004260524641722441, + 0.0031334729865193367, + 0.0049662161618471146, + 0.005072466563433409, + 0.005193037446588278, + 0.0035144882276654243, + 0.005364805459976196, + 0.023681744933128357, + 0.031813912093639374, + 0.023070763796567917, + 0.029663806781172752, + 0.026836905628442764, + 0.02969757840037346, + 0.020374150946736336, + 0.025545956566929817, + 0.03720201179385185, + 0.01066287700086832, + 0.07438868284225464, + 0.01500394195318222, + 0.016230467706918716, + 0.018480388447642326, + 0.01699632965028286, + 0.07982885092496872, + 0.01326213963329792, + 0.00980862881988287, + 0.029845891520380974, + 0.014024372212588787, + 0.01407972164452076, + 0.011334932409226894, + 0.013403907418251038, + 0.016285918653011322, + 0.003324005287140608, + 0.004765610210597515, + 0.0059613496996462345, + 0.005017516203224659, + 0.004404870327562094, + 0.004848152864724398, + 0.010808310471475124, + 0.009220104664564133, + 0.0031316597014665604, + 0.00477986503392458, + 0.006079863756895065, + 0.005530462134629488, + 0.0036515905521810055, + 0.006097529549151659, + 0.0077909016981720924, + 0.0032505204435437918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.00568072684109211 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608596369624138 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.01608596369624138, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608596369624138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10643287003040314, + 0.054197244346141815, + 0.02557833306491375, + 0.019597994163632393, + 0.01462905015796423, + 0.39129093289375305, + 0.18669652938842773, + 0.14543354511260986, + 0.03221549838781357, + 0.023927999660372734 + ], + "credit_map": [ + 0.03285762295126915, + 0.006919308099895716, + 0.005549649242311716, + 0.006442108191549778, + 0.004815069492906332, + 0.008317817933857441, + 0.00854914914816618, + 0.03298214450478554, + 0.0054974984377622604, + 0.008839991874992847, + 0.006310715805739164, + 0.009668505750596523, + 0.005089290905743837, + 0.005897117778658867, + 0.006984861101955175, + 0.005909258499741554, + 0.0031289425678551197, + 0.0019388687796890736, + 0.003951482940465212, + 0.002885014284402132, + 0.00229447428137064, + 0.003325952449813485, + 0.004004775080829859, + 0.004048821981996298, + 0.0019062852952629328, + 0.002820035209879279, + 0.002244719536975026, + 0.002786557422950864, + 0.0020035470370203257, + 0.0026211582589894533, + 0.0029713150579482317, + 0.0022443737834692, + 0.002003094181418419, + 0.0020468553993850946, + 0.0015635336749255657, + 0.002516380278393626, + 0.0014119230909273028, + 0.0019848328083753586, + 0.0011398019269108772, + 0.001962629845365882, + 0.04296411573886871, + 0.06205395609140396, + 0.0483643002808094, + 0.055605705827474594, + 0.04457452520728111, + 0.05153106153011322, + 0.04013746976852417, + 0.04605979844927788, + 0.023785719648003578, + 0.009944411925971508, + 0.0503871776163578, + 0.014401430264115334, + 0.01009401772171259, + 0.014872722327709198, + 0.013440090231597424, + 0.04977096989750862, + 0.01828080415725708, + 0.010856575332581997, + 0.03963805362582207, + 0.014267945662140846, + 0.012690523639321327, + 0.013520320877432823, + 0.020402533933520317, + 0.015776796266436577, + 0.0028131285216659307, + 0.00391019694507122, + 0.004284568130970001, + 0.0033335951156914234, + 0.003109572222456336, + 0.0042197187431156635, + 0.007035067770630121, + 0.003509650006890297, + 0.001680459245108068, + 0.0041046361438930035, + 0.0032772307749837637, + 0.0030586100183427334, + 0.0024110940285027027, + 0.003438913030549884, + 0.003035609843209386, + 0.0029214455280452967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.01608596369624138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003315538167953 + ], + "opposite_logprob_tokens": [ + -2.695033073425293 + ], + "logprob_reward": -0.07003315538167953, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003315538167953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20347458124160767, + 0.12694470584392548, + 0.050086189061403275, + 0.01504014153033495, + 0.020499741658568382, + 0.09169236570596695, + 0.05821529030799866, + 0.3560822904109955, + 0.05245288088917732, + 0.025511786341667175 + ], + "credit_map": [ + 0.07016301900148392, + 0.012593761086463928, + 0.01141720823943615, + 0.00879995059221983, + 0.006606608163565397, + 0.013656876049935818, + 0.01393057219684124, + 0.06630661338567734, + 0.015061835758388042, + 0.009490830823779106, + 0.017505427822470665, + 0.014400905929505825, + 0.011757727712392807, + 0.022325260564684868, + 0.016507677733898163, + 0.019895043224096298, + 0.0050326259806752205, + 0.004888780415058136, + 0.006826790515333414, + 0.005518331192433834, + 0.004664220847189426, + 0.008874747902154922, + 0.006948747672140598, + 0.007331944536417723, + 0.001647593337111175, + 0.0020967989694327116, + 0.0018528945511206985, + 0.0021837607491761446, + 0.0011811185395345092, + 0.001816709409467876, + 0.002483094809576869, + 0.0017781711649149656, + 0.0019441659096628428, + 0.0027333064936101437, + 0.002096936572343111, + 0.002723884303122759, + 0.0027924473397433758, + 0.0032181686256080866, + 0.0025044179055839777, + 0.002486415207386017, + 0.01131387148052454, + 0.014323842711746693, + 0.00971683394163847, + 0.011743742972612381, + 0.01232059858739376, + 0.012706393375992775, + 0.008655289188027382, + 0.010911795310676098, + 0.00572088360786438, + 0.003354712389409542, + 0.01338634267449379, + 0.0040384382009506226, + 0.005116499960422516, + 0.00689263641834259, + 0.005821811966598034, + 0.013883966021239758, + 0.04307635873556137, + 0.025708844885230064, + 0.09950019419193268, + 0.035838641226291656, + 0.032204464077949524, + 0.033268749713897705, + 0.049688637256622314, + 0.03679642453789711, + 0.002499438589438796, + 0.004011808428913355, + 0.005204655230045319, + 0.003332550171762705, + 0.0029354351572692394, + 0.0040279896929860115, + 0.017735078930854797, + 0.012705926783382893, + 0.0022473300341516733, + 0.002980513498187065, + 0.003624425735324621, + 0.003753840224817395, + 0.00267962459474802, + 0.004168541636317968, + 0.0031461152248084545, + 0.0029113953933119774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.07003315538167953 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000681012868881 + ], + "opposite_logprob_tokens": [ + -2.695006847381592 + ], + "logprob_reward": -0.07000681012868881, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000681012868881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019834695383906364, + 0.028968768194317818, + 0.029826907441020012, + 0.02229173108935356, + 0.013283876702189445, + 0.0386706218123436, + 0.306270956993103, + 0.49334362149238586, + 0.03324561566114426, + 0.014263233169913292 + ], + "credit_map": [ + 0.0023033353500068188, + 0.0026854579336941242, + 0.0018569058738648891, + 0.0022162452805787325, + 0.002236534608528018, + 0.0019953083246946335, + 0.0034342568833380938, + 0.003106652060523629, + 0.0036121965385973454, + 0.002567203715443611, + 0.0031503383070230484, + 0.004369980655610561, + 0.0027832509949803352, + 0.003096500178799033, + 0.005784360691905022, + 0.0036049371119588614, + 0.003732922952622175, + 0.003329073777422309, + 0.00348392641171813, + 0.003533119335770607, + 0.0028479264583438635, + 0.0028591593727469444, + 0.004716642200946808, + 0.005324136465787888, + 0.0024635924492031336, + 0.003300093114376068, + 0.0025617911014705896, + 0.0031032992992550135, + 0.0022003252524882555, + 0.0022730703931301832, + 0.0028244981076568365, + 0.0035650606732815504, + 0.0020983724389225245, + 0.0014006801648065448, + 0.001659879577346146, + 0.001692085643298924, + 0.0013500022469088435, + 0.0018237174954265356, + 0.0011304585495963693, + 0.0021286807022988796, + 0.005770985037088394, + 0.0040906574577093124, + 0.006879401858896017, + 0.00484221987426281, + 0.005020612385123968, + 0.004125283565372229, + 0.002820999827235937, + 0.005120462272316217, + 0.0399089939892292, + 0.014613972045481205, + 0.08628133684396744, + 0.023127330467104912, + 0.01488733384758234, + 0.021687280386686325, + 0.021285736933350563, + 0.0844789668917656, + 0.06086161732673645, + 0.03395349532365799, + 0.1395130455493927, + 0.04994739964604378, + 0.04387575387954712, + 0.0470120869576931, + 0.06810591369867325, + 0.050074320286512375, + 0.0019347474444657564, + 0.003390456549823284, + 0.00427204929292202, + 0.002675921656191349, + 0.0027918776031583548, + 0.003482975997030735, + 0.008649228140711784, + 0.006048360373824835, + 0.0011317443568259478, + 0.0020935952197760344, + 0.0016650470206514, + 0.00172074802685529, + 0.001300767995417118, + 0.0019721093121916056, + 0.0033512646332383156, + 0.0010279564885422587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.07000681012868881 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902790397405624 + ], + "opposite_logprob_tokens": [ + -2.4640278816223145 + ], + "logprob_reward": -0.08902790397405624, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902790397405624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0452723391354084, + 0.018935460597276688, + 0.01378535758703947, + 0.014577055349946022, + 0.014751589857041836, + 0.026289107277989388, + 0.2816970944404602, + 0.5544107556343079, + 0.022684378549456596, + 0.007596897892653942 + ], + "credit_map": [ + 0.011242777109146118, + 0.0027990464586764574, + 0.0028118996415287256, + 0.0029673788230866194, + 0.00300338352099061, + 0.004995041061192751, + 0.004353769589215517, + 0.01309904269874096, + 0.0021613778080791235, + 0.0027534510008990765, + 0.002117567230015993, + 0.002558480016887188, + 0.002323088003322482, + 0.0021502384915947914, + 0.0024201807100325823, + 0.0024510789662599564, + 0.001396389794535935, + 0.001802939805202186, + 0.001807554392144084, + 0.0020697866566479206, + 0.0012477299897000194, + 0.0017626857152208686, + 0.0016935236053541303, + 0.002004748908802867, + 0.0014414102770388126, + 0.002380951074883342, + 0.0015840051928535104, + 0.0022544690873473883, + 0.0016457297606393695, + 0.0015348130837082863, + 0.0017235585255548358, + 0.002012119395658374, + 0.00260825059376657, + 0.002074738033115864, + 0.001127820578403771, + 0.0014302562922239304, + 0.0015582502819597721, + 0.0013389186933636665, + 0.0013336841948330402, + 0.0032796708401292562, + 0.005257279146462679, + 0.002435245318338275, + 0.005999768152832985, + 0.0024648646358400583, + 0.002760368399322033, + 0.0022635417990386486, + 0.0020486866123974323, + 0.0030593527480959892, + 0.039241328835487366, + 0.01375984214246273, + 0.07581938058137894, + 0.02102007530629635, + 0.017412716522812843, + 0.022824671119451523, + 0.01847783848643303, + 0.07314125448465347, + 0.07165447622537613, + 0.039197735488414764, + 0.15871043503284454, + 0.05563565716147423, + 0.04970592260360718, + 0.04801512137055397, + 0.07571417838335037, + 0.05577724054455757, + 0.0014104688307270408, + 0.003687979420647025, + 0.002466503530740738, + 0.0030993036925792694, + 0.0028986018151044846, + 0.001799797872081399, + 0.0047375354915857315, + 0.0025841863825917244, + 0.0006330937030725181, + 0.00111099099740386, + 0.0009014317765831947, + 0.001247386448085308, + 0.0006882084417156875, + 0.0011851879535242915, + 0.001098291832022369, + 0.0007323069148696959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.08902790397405624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, captured from a slightly elevated angle and slowly panning forward, focuses on an ant exploring a large, vibrant green leaf. The ant's glossy, dark brown exoskeleton reflects delicately in the soft morning light, its segmented body supported by slender, powerful legs. Its two slender, jointed antennae are the dynamic focal point of the image, quivering and swaying, tapping and sweeping across the leaf with a curious air, as if meticulously interpreting the scents and textures of its surroundings. The leaf itself resembles a miniature landscape, its intricate network of veins forming ridges and valleys, and a few glistening dewdrops clinging to its surface like tiny gems, catching and refracting the light. Soft, diffuse sunlight penetrates the leaf's translucent structure, bathing the scene in a tranquil, almost ethereal green aura, dominated by vibrant emerald greens and earthy browns. The ant moves cautiously along the \"highway\" formed by the leaf veins at a steady and leisurely pace. Its every movement is full of purpose, creating an atmosphere of focus, tranquility and the miracle of life.", + "seed": 6623, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_14/index_7602_seed_6623_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0326247425749898, + "vlm_logprob": -0.0326247425749898, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.0326247425749898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005017546005547047 + ], + "opposite_logprob_tokens": [ + -7.0050177574157715 + ], + "logprob_reward": -0.005017546005547047, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.005017546005547047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1461648941040039, + 0.07254752516746521, + 0.11138885468244553, + 0.10154196619987488, + 0.12220677733421326, + 0.08995761722326279, + 0.0668230652809143, + 0.09959935396909714, + 0.05424640327692032, + 0.1355234831571579 + ], + "credit_map": [ + 0.012373501434922218, + 0.02093503810465336, + 0.01748347096145153, + 0.01697532646358013, + 0.025693759322166443, + 0.027264585718512535, + 0.012705154716968536, + 0.0127340629696846, + 0.0056707714684307575, + 0.00975210964679718, + 0.010895962826907635, + 0.00844770111143589, + 0.014437607489526272, + 0.00882663019001484, + 0.008003284223377705, + 0.006513457279652357, + 0.014326503500342369, + 0.012192204594612122, + 0.01282492559403181, + 0.01845305971801281, + 0.012059133499860764, + 0.014621168375015259, + 0.010730063542723656, + 0.01618180237710476, + 0.009342732839286327, + 0.01404231134802103, + 0.013021380640566349, + 0.014128976501524448, + 0.013594142161309719, + 0.015824468806385994, + 0.00797366350889206, + 0.0136142922565341, + 0.0073571461252868176, + 0.02328234165906906, + 0.03098694048821926, + 0.009317541494965553, + 0.014752603136003017, + 0.013967875391244888, + 0.007698121480643749, + 0.014844201505184174, + 0.0067276787012815475, + 0.013132014311850071, + 0.02387244626879692, + 0.0077894628047943115, + 0.012976138852536678, + 0.010105065070092678, + 0.008699079044163227, + 0.006655735895037651, + 0.005476444959640503, + 0.023060420528054237, + 0.014475286938250065, + 0.005196454469114542, + 0.005564945284277201, + 0.004030824173241854, + 0.004423798993229866, + 0.00459488807246089, + 0.011924916878342628, + 0.007577370852231979, + 0.013081187382340431, + 0.01842734031379223, + 0.014637548476457596, + 0.008443051017820835, + 0.013281128369271755, + 0.012226811610162258, + 0.005686392076313496, + 0.004912072792649269, + 0.006710379850119352, + 0.007835473865270615, + 0.008222654461860657, + 0.006406267639249563, + 0.008519986644387245, + 0.005953173618763685, + 0.015680933371186256, + 0.02433713898062706, + 0.01406308077275753, + 0.016839083284139633, + 0.028110790997743607, + 0.013720766641199589, + 0.01288542989641428, + 0.009886262938380241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.005017546005547047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255878061056137 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.01255878061056137, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255878061056137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17863623797893524, + 0.07077518105506897, + 0.0656013935804367, + 0.15423180162906647, + 0.045185912400484085, + 0.03847352787852287, + 0.0674554705619812, + 0.0784326046705246, + 0.07001963257789612, + 0.23118822276592255 + ], + "credit_map": [ + 0.015993719920516014, + 0.035936929285526276, + 0.04427381604909897, + 0.017473699524998665, + 0.0171852707862854, + 0.017466923221945763, + 0.0159212164580822, + 0.014384672977030277, + 0.005547976121306419, + 0.005819409620016813, + 0.008494766429066658, + 0.00910137128084898, + 0.012974926270544529, + 0.01019667275249958, + 0.00939632486552, + 0.009243743494153023, + 0.0077748713083565235, + 0.004631543066352606, + 0.009055120870471, + 0.010274779982864857, + 0.00930346455425024, + 0.010148594155907631, + 0.00778591213747859, + 0.006627114489674568, + 0.010897484607994556, + 0.03608779236674309, + 0.051820240914821625, + 0.013434953056275845, + 0.012913797050714493, + 0.01478102058172226, + 0.005440311972051859, + 0.008856209926307201, + 0.0032550208270549774, + 0.007182903587818146, + 0.01078843791037798, + 0.005453967954963446, + 0.0053800963796675205, + 0.004971602465957403, + 0.004319317638874054, + 0.0038345716893672943, + 0.002229784382507205, + 0.010531801730394363, + 0.012969788163900375, + 0.003014771966263652, + 0.002876952989026904, + 0.002884726505726576, + 0.002015381585806608, + 0.0019503259100019932, + 0.004047618247568607, + 0.020417701452970505, + 0.022569123655557632, + 0.004433929920196533, + 0.0045383223332464695, + 0.0036548979114741087, + 0.004634404089301825, + 0.003159474115818739, + 0.009102794341742992, + 0.006211447529494762, + 0.009628105908632278, + 0.014162663370370865, + 0.012176570482552052, + 0.00838242843747139, + 0.010671967640519142, + 0.008096632547676563, + 0.006720926146954298, + 0.006035220809280872, + 0.006528095342218876, + 0.007727138698101044, + 0.014586856588721275, + 0.010602006688714027, + 0.009987704455852509, + 0.007831687107682228, + 0.023797670379281044, + 0.044298477470874786, + 0.030661175027489662, + 0.03789624944329262, + 0.029274867847561836, + 0.017548400908708572, + 0.029655544087290764, + 0.018055859953165054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.01255878061056137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585693418979645 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585693418979645, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585693418979645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1334322690963745, + 0.07747629284858704, + 0.08546441048383713, + 0.1667766124010086, + 0.058898333460092545, + 0.057584088295698166, + 0.12645544111728668, + 0.13732929527759552, + 0.06990988552570343, + 0.08667335659265518 + ], + "credit_map": [ + 0.010083509609103203, + 0.03063107840716839, + 0.027699589729309082, + 0.018317626789212227, + 0.008861611597239971, + 0.011116726323962212, + 0.01563555933535099, + 0.01108656544238329, + 0.006281272042542696, + 0.007638240233063698, + 0.008571287617087364, + 0.012439937330782413, + 0.011282048188149929, + 0.009631911292672157, + 0.012692148797214031, + 0.00893944501876831, + 0.012648114003241062, + 0.01263795979321003, + 0.011500570923089981, + 0.017051056027412415, + 0.008467382751405239, + 0.008770476095378399, + 0.007007792126387358, + 0.007381059695035219, + 0.012793680652976036, + 0.0521409697830677, + 0.05269753187894821, + 0.012653353624045849, + 0.008785311132669449, + 0.011170153506100178, + 0.0073921228758990765, + 0.009143510833382607, + 0.004719531629234552, + 0.009293632581830025, + 0.010364537127315998, + 0.007887938991189003, + 0.006162300705909729, + 0.007525299210101366, + 0.004825972020626068, + 0.008119119331240654, + 0.005249458365142345, + 0.013768943026661873, + 0.018770305439829826, + 0.005953329149633646, + 0.0038615635130554438, + 0.003489677095785737, + 0.003159148618578911, + 0.0033316619228571653, + 0.012121429666876793, + 0.03320229426026344, + 0.04326426610350609, + 0.010399185121059418, + 0.00789431482553482, + 0.005912479478865862, + 0.006500564981251955, + 0.0071609048172831535, + 0.014860589988529682, + 0.01098315604031086, + 0.019599368795752525, + 0.020154740661382675, + 0.02106098085641861, + 0.011427652090787888, + 0.02236924134194851, + 0.016873562708497047, + 0.009019691497087479, + 0.0040911780670285225, + 0.007866993546485901, + 0.014615570195019245, + 0.0047593675553798676, + 0.0049173408187925816, + 0.01916608028113842, + 0.005473671481013298, + 0.011956469155848026, + 0.010298391804099083, + 0.013083445839583874, + 0.013802553527057171, + 0.009618673473596573, + 0.009542847983539104, + 0.009514234028756618, + 0.008856736123561859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.020585693418979645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054948654025793076 + ], + "opposite_logprob_tokens": [ + -2.929948568344116 + ], + "logprob_reward": -0.054948654025793076, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054948654025793076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20465360581874847, + 0.15224777162075043, + 0.17196696996688843, + 0.1223553866147995, + 0.03142211213707924, + 0.033251892775297165, + 0.07931888848543167, + 0.0432688407599926, + 0.06000540032982826, + 0.10150911659002304 + ], + "credit_map": [ + 0.01819998398423195, + 0.03410075232386589, + 0.05560648441314697, + 0.01184102799743414, + 0.025400320068001747, + 0.027609756216406822, + 0.015048624016344547, + 0.016846660524606705, + 0.011109260842204094, + 0.021521572023630142, + 0.0298638716340065, + 0.02201450988650322, + 0.020452946424484253, + 0.013878386467695236, + 0.020092930644750595, + 0.013314296491444111, + 0.025043686851859093, + 0.02143925242125988, + 0.02143239416182041, + 0.02342969924211502, + 0.018527017906308174, + 0.018822675570845604, + 0.023675328120589256, + 0.019596928730607033, + 0.0072466046549379826, + 0.031334687024354935, + 0.04796909913420677, + 0.013564559631049633, + 0.0053320941515266895, + 0.00849065650254488, + 0.0032161823473870754, + 0.005201504100114107, + 0.00286101084202528, + 0.006034215446561575, + 0.008140668272972107, + 0.003216055454686284, + 0.0035149622708559036, + 0.0031573534943163395, + 0.0022988372948020697, + 0.002199007896706462, + 0.002685349667444825, + 0.005650730803608894, + 0.010871011763811111, + 0.003922848496586084, + 0.0031455066055059433, + 0.0029850779101252556, + 0.0017335455631837249, + 0.002257823245599866, + 0.004481298383325338, + 0.026045972481369972, + 0.025877462700009346, + 0.004782951436936855, + 0.0050175865180790424, + 0.004389440640807152, + 0.004964098334312439, + 0.003760071936994791, + 0.0038363097701221704, + 0.003430647077038884, + 0.008794784545898438, + 0.005125393159687519, + 0.007746501825749874, + 0.00393954711034894, + 0.0033813866320997477, + 0.007014269009232521, + 0.004957215860486031, + 0.0041435821913182735, + 0.007110536098480225, + 0.008498760871589184, + 0.007921176962554455, + 0.0075109913013875484, + 0.013932449743151665, + 0.005930684041231871, + 0.013623781502246857, + 0.021486658602952957, + 0.012817470356822014, + 0.01151567604392767, + 0.01457479689270258, + 0.009609421715140343, + 0.010890700854361057, + 0.0069906143471598625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.054948654025793076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001303881406784 + ], + "opposite_logprob_tokens": [ + -2.6950130462646484 + ], + "logprob_reward": -0.07001303881406784, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001303881406784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1470288187265396, + 0.09619355201721191, + 0.07206208258867264, + 0.08466536551713943, + 0.11675850301980972, + 0.03920643404126167, + 0.17536300420761108, + 0.08146658539772034, + 0.1122920885682106, + 0.0749635398387909 + ], + "credit_map": [ + 0.011939090676605701, + 0.03691573441028595, + 0.02938227914273739, + 0.015477819368243217, + 0.018987584859132767, + 0.015518351458013058, + 0.009063327684998512, + 0.00974463764578104, + 0.007204296533018351, + 0.011063577607274055, + 0.009905007667839527, + 0.013366579078137875, + 0.0148349329829216, + 0.015165251679718494, + 0.015467491932213306, + 0.009186424314975739, + 0.008771956898272038, + 0.012337044812738895, + 0.014945639297366142, + 0.008840896189212799, + 0.008541533723473549, + 0.007369109895080328, + 0.005557724740356207, + 0.005698173772543669, + 0.009074189700186253, + 0.013096434064209461, + 0.01479073241353035, + 0.01202849019318819, + 0.010443786159157753, + 0.010664336383342743, + 0.0074056703597307205, + 0.007161736022680998, + 0.012424522079527378, + 0.021488957107067108, + 0.023793259635567665, + 0.0186734851449728, + 0.010263236239552498, + 0.008918103761970997, + 0.009394238702952862, + 0.01180270779877901, + 0.002909223549067974, + 0.004262351430952549, + 0.010442741215229034, + 0.005503103137016296, + 0.004955856595188379, + 0.0034696676302701235, + 0.004069076851010323, + 0.0035944145638495684, + 0.008986254222691059, + 0.060149896889925, + 0.055496346205472946, + 0.010991165414452553, + 0.012427598237991333, + 0.010255194269120693, + 0.010433683171868324, + 0.006622865796089172, + 0.007441862020641565, + 0.005868172738701105, + 0.008117916993796825, + 0.010885545052587986, + 0.013367436826229095, + 0.009603125974535942, + 0.009025832638144493, + 0.0171566940844059, + 0.014278598129749298, + 0.009620183147490025, + 0.011483347974717617, + 0.015610053203999996, + 0.018977472558617592, + 0.013221976347267628, + 0.016208169981837273, + 0.01289229467511177, + 0.009757374413311481, + 0.01564890705049038, + 0.0124479029327631, + 0.00752106960862875, + 0.00999836903065443, + 0.007170164957642555, + 0.00554037606343627, + 0.0068793813697993755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.07001303881406784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, captured from a slightly elevated angle and slowly panning forward, focuses on an ant exploring a large, vibrant green leaf. The ant's glossy, dark brown exoskeleton reflects delicately in the soft morning light, its segmented body supported by slender, powerful legs. Its two slender, jointed antennae are the dynamic focal point of the image, quivering and swaying, tapping and sweeping across the leaf with a curious air, as if meticulously interpreting the scents and textures of its surroundings. The leaf itself resembles a miniature landscape, its intricate network of veins forming ridges and valleys, and a few glistening dewdrops clinging to its surface like tiny gems, catching and refracting the light. Soft, diffuse sunlight penetrates the leaf's translucent structure, bathing the scene in a tranquil, almost ethereal green aura, dominated by vibrant emerald greens and earthy browns. The ant moves cautiously along the \"highway\" formed by the leaf veins at a steady and leisurely pace. Its every movement is full of purpose, creating an atmosphere of focus, tranquility and the miracle of life.", + "seed": 6624, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_14/index_7602_seed_6624_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03393457848578692, + "vlm_logprob": -0.03393457848578692, + "vlm_yes_no_margin": 4.0749999523162845, + "vlm_token_credit": -0.03393457848578692, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005678593181073666 + ], + "opposite_logprob_tokens": [ + -6.880678653717041 + ], + "logprob_reward": -0.005678593181073666, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.005678593181073666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02395574189722538, + 0.012608046643435955, + 0.0046181948855519295, + 0.00642260629683733, + 0.004075687378644943, + 0.00767502561211586, + 0.019255612045526505, + 0.8496633172035217, + 0.007111692801117897, + 0.06461407244205475 + ], + "credit_map": [ + 0.001620743190869689, + 0.0016203797422349453, + 0.00159204110968858, + 0.0015427969628944993, + 0.0016477728495374322, + 0.00830757711082697, + 0.0015971747925505042, + 0.006027257069945335, + 0.0013009363319724798, + 0.0025275342632085085, + 0.0016751968069002032, + 0.0013113495660945773, + 0.0014315489679574966, + 0.0011150306090712547, + 0.0011257198639214039, + 0.002120730932801962, + 0.00044521703966893256, + 0.0007567229913547635, + 0.0007252306095324457, + 0.0005089049809612334, + 0.0005344772944226861, + 0.0005183599423617125, + 0.0004853983409702778, + 0.0006438837735913694, + 0.0004753043467644602, + 0.0007014768780209124, + 0.00043229435686953366, + 0.0005584859754890203, + 0.000522323593031615, + 0.0018995192367583513, + 0.0004685764142777771, + 0.001364625641144812, + 0.0003467103815637529, + 0.0006054859841242433, + 0.0002972024376504123, + 0.0005030713509768248, + 0.00048771852743811905, + 0.0008044203277677298, + 0.00048243236960843205, + 0.0005486460868269205, + 0.0006325943977572024, + 0.0007272445946000516, + 0.0007295437389984727, + 0.000507367542013526, + 0.0006635349127463996, + 0.0022726256866008043, + 0.0004980884259566665, + 0.0016440264880657196, + 0.001774817705154419, + 0.0015588747337460518, + 0.004067325033247471, + 0.0012022771406918764, + 0.0017584204906597733, + 0.00624261936172843, + 0.001229601912200451, + 0.0014216770650818944, + 0.05152765288949013, + 0.02563486061990261, + 0.21463795006275177, + 0.027478616684675217, + 0.08271245658397675, + 0.3744703531265259, + 0.040765274316072464, + 0.03243613615632057, + 0.0007622448028996587, + 0.0008170903311111033, + 0.000671003945171833, + 0.0009572214330546558, + 0.0011033889604732394, + 0.0010009607067331672, + 0.0010424040956422687, + 0.0007573784678243101, + 0.007315110880881548, + 0.007426344323903322, + 0.012869387865066528, + 0.006915121804922819, + 0.0057979244738817215, + 0.006014770828187466, + 0.007510465569794178, + 0.01076494064182043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.005678593181073666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820252276957035 + ], + "opposite_logprob_tokens": [ + -4.018202304840088 + ], + "logprob_reward": -0.01820252276957035, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01820252276957035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07015926390886307, + 0.04488291218876839, + 0.015687676146626472, + 0.011429368518292904, + 0.015239179134368896, + 0.012000408954918385, + 0.03766946122050285, + 0.6243335604667664, + 0.019984045997262, + 0.14861419796943665 + ], + "credit_map": [ + 0.005510581657290459, + 0.00773710198700428, + 0.005011360161006451, + 0.006494821514934301, + 0.008182520978152752, + 0.014874442480504513, + 0.00697677955031395, + 0.015371664427220821, + 0.004987568594515324, + 0.005907025653868914, + 0.007702851202338934, + 0.0039550187066197395, + 0.004775679670274258, + 0.00483980868011713, + 0.003077286761254072, + 0.009637672454118729, + 0.0013481171336025, + 0.003575041890144348, + 0.0017986971652135253, + 0.0017400190699845552, + 0.0015192896826192737, + 0.0022718976251780987, + 0.001969385426491499, + 0.001465229201130569, + 0.0012546229409053922, + 0.0018252439331263304, + 0.0013417211594060063, + 0.0017036708304658532, + 0.001234422903507948, + 0.0017060829559341073, + 0.0011354380985721946, + 0.001228166976943612, + 0.0013468553079292178, + 0.0032677638810127974, + 0.0011935918591916561, + 0.0016867308877408504, + 0.0023319979663938284, + 0.0014875195920467377, + 0.0021639177575707436, + 0.0017608023481443524, + 0.001468375907279551, + 0.0015532245161011815, + 0.0016704251756891608, + 0.0011658453149721026, + 0.0014510219916701317, + 0.002179897390305996, + 0.0011625394690781832, + 0.0013490814017131925, + 0.003690508659929037, + 0.002890061354264617, + 0.008869525976479053, + 0.002362817758694291, + 0.004027654882520437, + 0.0031694613862782717, + 0.0030444760341197252, + 0.009614959359169006, + 0.03773698955774307, + 0.019758453592658043, + 0.15565168857574463, + 0.019204773008823395, + 0.06131887808442116, + 0.2745656371116638, + 0.030658502131700516, + 0.025438634678721428, + 0.002201916417106986, + 0.0033240558113902807, + 0.002364500192925334, + 0.0022683413699269295, + 0.0028860720340162516, + 0.002350830938667059, + 0.0024815716315060854, + 0.002106759464368224, + 0.014531705528497696, + 0.02289888635277748, + 0.02413840964436531, + 0.013754334300756454, + 0.026139365509152412, + 0.01573740877211094, + 0.019186360761523247, + 0.012227743864059448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.01820252276957035 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818554848432541 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.01818554848432541, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818554848432541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12799343466758728, + 0.15543632209300995, + 0.04520944505929947, + 0.014092997647821903, + 0.018127446994185448, + 0.014958675019443035, + 0.09311097860336304, + 0.4682886004447937, + 0.027330271899700165, + 0.03545190393924713 + ], + "credit_map": [ + 0.010155478492379189, + 0.007314557209610939, + 0.009038076736032963, + 0.008126690983772278, + 0.01183886919170618, + 0.04502855986356735, + 0.007893786765635014, + 0.028597423806786537, + 0.013306524604558945, + 0.011667526327073574, + 0.0243512112647295, + 0.011867446824908257, + 0.021690228953957558, + 0.028835652396082878, + 0.013534551486372948, + 0.03018317185342312, + 0.0034063151106238365, + 0.006383518688380718, + 0.006165994796901941, + 0.0038488709833472967, + 0.0052415309473872185, + 0.00847039557993412, + 0.0057578920386731625, + 0.005934923887252808, + 0.0011589322239160538, + 0.0015247748233377934, + 0.0011841098312288523, + 0.0014041743706911802, + 0.001312220236286521, + 0.003779076039791107, + 0.0009353413479402661, + 0.002794368891045451, + 0.0016105222748592496, + 0.0031898252200335264, + 0.0016634351341053843, + 0.0021411548368632793, + 0.00233005010522902, + 0.0016570748994126916, + 0.0029792480636388063, + 0.00255613774061203, + 0.002271465491503477, + 0.0013949829153716564, + 0.002012502634897828, + 0.0015145805664360523, + 0.0016704157460480928, + 0.003180209780111909, + 0.0016770579386502504, + 0.0012374600628390908, + 0.008594011887907982, + 0.004816398490220308, + 0.025664547458291054, + 0.0051151663064956665, + 0.008485445752739906, + 0.031161310151219368, + 0.0044115749187767506, + 0.004862526431679726, + 0.028530849143862724, + 0.014583772048354149, + 0.1162802129983902, + 0.014194863848388195, + 0.046132974326610565, + 0.20714759826660156, + 0.023029949516057968, + 0.01838834583759308, + 0.002577058970928192, + 0.003477621590718627, + 0.002653059782460332, + 0.0033267836552113295, + 0.004424265585839748, + 0.0041598910465836525, + 0.004010786768049002, + 0.00270080310292542, + 0.003043042030185461, + 0.004726721439510584, + 0.0031016296707093716, + 0.004917935002595186, + 0.004825368523597717, + 0.003781381994485855, + 0.0043142568320035934, + 0.006741567514836788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.01818554848432541 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864410310983658 + ], + "opposite_logprob_tokens": [ + -3.0486440658569336 + ], + "logprob_reward": -0.04864410310983658, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864410310983658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.006280736532062292, + 0.0039147622883319855, + 0.003991524688899517, + 0.002232012338936329, + 0.0031997866462916136, + 0.0066490801982581615, + 0.027103513479232788, + 0.9318223595619202, + 0.008668131195008755, + 0.006138079334050417 + ], + "credit_map": [ + 0.0004502752562984824, + 0.0004679459088947624, + 0.0005189160583540797, + 0.0004728594794869423, + 0.0005570457433350384, + 0.0020990429911762476, + 0.00028141902294009924, + 0.0014332319842651486, + 0.0003642166848294437, + 0.0006440955330617726, + 0.0004740890581160784, + 0.0003769146278500557, + 0.0005264737410470843, + 0.000705118989571929, + 0.00034157923073507845, + 0.00048227416118606925, + 0.00039962519076652825, + 0.001172581221908331, + 0.0004755459085572511, + 0.0004229379410389811, + 0.00026886750129051507, + 0.0003674931649584323, + 0.0004888880648650229, + 0.0003955858410336077, + 0.00025058360188268125, + 0.00041475362377241254, + 0.00018494311370886862, + 0.000273410725640133, + 0.00023785702069289982, + 0.00032113766064867377, + 0.00022806627384852618, + 0.000321260275086388, + 0.0003590092237573117, + 0.0006526685319840908, + 0.0002343414380447939, + 0.00038690396468155086, + 0.00037921874900348485, + 0.00042004286660812795, + 0.0005406949785538018, + 0.0002269067772431299, + 0.0007239495171234012, + 0.0015858231345191598, + 0.0005164077738299966, + 0.0006121278274804354, + 0.0007434199214912951, + 0.0012184572406113148, + 0.0004799797316081822, + 0.0007689150515943766, + 0.0027224060613662004, + 0.0016873992281034589, + 0.0056533534079790115, + 0.0012201122008264065, + 0.0033581042662262917, + 0.009453735314309597, + 0.0014637288404628634, + 0.0015446726465597749, + 0.05675135925412178, + 0.028124472126364708, + 0.23469290137290955, + 0.030332155525684357, + 0.09070396423339844, + 0.41118916869163513, + 0.04462311416864395, + 0.035405233502388, + 0.001392202451825142, + 0.0010562930256128311, + 0.0013648148160427809, + 0.0009086431819014251, + 0.0009527419460937381, + 0.0010956086916849017, + 0.0008846080745570362, + 0.0010132190072908998, + 0.0008387884590774775, + 0.0006388896727003157, + 0.0014199630822986364, + 0.0005565806059166789, + 0.0005009744199924171, + 0.000655609299428761, + 0.0006205251556821167, + 0.0009067484643310308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.04864410310983658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of an ant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ant's exoskeleton glossy and dark brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening dewdrops visible on the leaf surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan forward from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ant moving cautiously along the leaf veins at a steady pace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896212488412857 + ], + "opposite_logprob_tokens": [ + -2.5789620876312256 + ], + "logprob_reward": -0.07896212488412857, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896212488412857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025520719587802887, + 0.03209247812628746, + 0.012002550065517426, + 0.006162888370454311, + 0.01564982905983925, + 0.013380492106080055, + 0.015124879777431488, + 0.8441269993782043, + 0.015344119630753994, + 0.02059498429298401 + ], + "credit_map": [ + 0.0016809316584840417, + 0.003400034736841917, + 0.0015030096983537078, + 0.0021629296243190765, + 0.0019909339025616646, + 0.007160595618188381, + 0.0025094379670917988, + 0.0051128496415913105, + 0.002442206023260951, + 0.0036943338345736265, + 0.004808224271982908, + 0.0025606981944292784, + 0.004756024107336998, + 0.006331401877105236, + 0.0026886158157140017, + 0.004810978192836046, + 0.000993445049971342, + 0.0018131897086277604, + 0.0019036426674574614, + 0.0011648349463939667, + 0.0010063012596219778, + 0.002025140915066004, + 0.0012506595812737942, + 0.0018453368684276938, + 0.0006064121262170374, + 0.0010267692850902677, + 0.0006908458890393376, + 0.0007972785970196128, + 0.0006351537886075675, + 0.0011193802347406745, + 0.0004989727749489248, + 0.0007880761986598372, + 0.00151916837785393, + 0.004266397561877966, + 0.0009642367949709296, + 0.0016991323791444302, + 0.0017306437948718667, + 0.0026118336245417595, + 0.0016273201908916235, + 0.0012310984311625361, + 0.0014316305750980973, + 0.001227624830789864, + 0.0015803746646270156, + 0.0014213284011930227, + 0.0015905140899121761, + 0.0014288494130596519, + 0.003396156243979931, + 0.0013040160993114114, + 0.0013381008757278323, + 0.0013693749206140637, + 0.003141740569844842, + 0.0008661469328217208, + 0.0013600335223600268, + 0.005159671418368816, + 0.0008991018985398114, + 0.000990710686892271, + 0.05126499757170677, + 0.025528253987431526, + 0.2131381332874298, + 0.027555789798498154, + 0.08194998651742935, + 0.3721172511577606, + 0.04051905870437622, + 0.032053619623184204, + 0.002502521499991417, + 0.002431150060147047, + 0.0017764981603249907, + 0.0016552527667954564, + 0.0017921997932717204, + 0.002721178811043501, + 0.0013703497825190425, + 0.0010949698043987155, + 0.0035118788946419954, + 0.0023578605614602566, + 0.003559785196557641, + 0.0014247143408283591, + 0.00282078143209219, + 0.0021875272504985332, + 0.0022524904925376177, + 0.0024799471721053123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.07896212488412857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_15/index_tensor([4225, 4225])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_15/index_tensor([4225, 4225])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e560626333b8ff1ee598811bddfd2a6452bdf2d9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_15/index_tensor([4225, 4225])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, shot from a low angle, gently tilts upward and pans to the right, tracking a creature soaring through the sky. A large, bright full moon dominates the background, its cratered silver surface rendered in razor-sharp detail. A massive, ancient dragon circles before it with a deliberate grace and deliberate slowness. Its obsidian-smooth scales trace a graceful arc, its massive, leathery wings flapping with a powerful yet deliberate rhythm, each downward stroke silently and powerfully displacing the air. The full moon's cool, bright light serves as a powerful backlight, casting a sharp silver rimlight across the dragon's silhouette. The light sweeps across its body, causing its black scales to shimmer with an iridescent luster as it moves, like flowing metal. Wisps of wispy clouds drift lazily across the deep, velvety night sky, further emphasizing the dragon's immense size and height. The entire scene is filled with profound ancient majesty and quiet transcendental power, creating a breathtaking, mythical atmosphere.", + "seed": 7753, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_15/index_4225_seed_7753_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03149513229727745, + "vlm_logprob": -0.03149513229727745, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03149513229727745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016154352575540543 + ], + "opposite_logprob_tokens": [ + -4.3911542892456055 + ], + "logprob_reward": -0.016154352575540543, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016154352575540543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2847876250743866, + 0.054860670119524, + 0.06343740969896317, + 0.040355004370212555, + 0.06545072793960571, + 0.213407963514328, + 0.10781628638505936, + 0.06850180774927139, + 0.06206193566322327, + 0.03932053595781326 + ], + "credit_map": [ + 0.040495555847883224, + 0.021136503666639328, + 0.02862250804901123, + 0.02353782393038273, + 0.02031957171857357, + 0.015516330488026142, + 0.050435878336429596, + 0.08472345024347305, + 0.008314277045428753, + 0.004197039641439915, + 0.009673384949564934, + 0.005856481846421957, + 0.00586758041754365, + 0.005477596540004015, + 0.008917790837585926, + 0.00655651930719614, + 0.008917015977203846, + 0.0064802467823028564, + 0.00877404399216175, + 0.006039315368980169, + 0.007000607904046774, + 0.009560013189911842, + 0.0055555133149027824, + 0.011110656894743443, + 0.006970452610403299, + 0.0037327466998249292, + 0.007453053258359432, + 0.003129218239337206, + 0.003191499039530754, + 0.006291520316153765, + 0.002650441136211157, + 0.006936072371900082, + 0.010781283490359783, + 0.005758538376539946, + 0.008714832365512848, + 0.0052407835610210896, + 0.009785541333258152, + 0.00921027921140194, + 0.006717722397297621, + 0.009241744875907898, + 0.01982777751982212, + 0.0060040634125471115, + 0.0061397855170071125, + 0.0066240583546459675, + 0.008433563634753227, + 0.0739363506436348, + 0.008197915740311146, + 0.08424443006515503, + 0.00927269458770752, + 0.014534290879964828, + 0.018372882157564163, + 0.012380414642393589, + 0.01735875755548477, + 0.012973194941878319, + 0.010979118756949902, + 0.011944931000471115, + 0.014057589694857597, + 0.005287396255880594, + 0.0067274426110088825, + 0.004734647460281849, + 0.005520246922969818, + 0.007963903248310089, + 0.010529174469411373, + 0.013681414537131786, + 0.007114934269338846, + 0.009259319864213467, + 0.006713133305311203, + 0.008266451768577099, + 0.010054066777229309, + 0.00767152663320303, + 0.007934222929179668, + 0.005048274528235197, + 0.00693404208868742, + 0.004750857129693031, + 0.003257836913689971, + 0.0033176427241414785, + 0.0032602448482066393, + 0.009343380108475685, + 0.003927439916878939, + 0.004529092460870743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.016154352575540543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002881914377213 + ], + "opposite_logprob_tokens": [ + -2.695028781890869 + ], + "logprob_reward": -0.07002881914377213, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002881914377213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20109067857265472, + 0.018224991858005524, + 0.059971217066049576, + 0.08435894548892975, + 0.062218356877565384, + 0.37584352493286133, + 0.041870880872011185, + 0.033710233867168427, + 0.08930786699056625, + 0.03340335190296173 + ], + "credit_map": [ + 0.02440066635608673, + 0.011372664012014866, + 0.01744481362402439, + 0.01624339073896408, + 0.010728847235441208, + 0.009343583136796951, + 0.045791372656822205, + 0.06576532125473022, + 0.0022741977591067553, + 0.0019282253924757242, + 0.0023029877338558435, + 0.003304501064121723, + 0.002132245572283864, + 0.002207714132964611, + 0.002017878694459796, + 0.002057240344583988, + 0.008813063614070415, + 0.004476442933082581, + 0.008739493787288666, + 0.008838941343128681, + 0.003971341531723738, + 0.008888143114745617, + 0.005357097368687391, + 0.01088668592274189, + 0.012573165819048882, + 0.007160248700529337, + 0.013947970233857632, + 0.009540215134620667, + 0.0076536634005606174, + 0.01384245976805687, + 0.007108793593943119, + 0.012532426044344902, + 0.007232559844851494, + 0.004530671983957291, + 0.010833707638084888, + 0.00428764009848237, + 0.011284752748906612, + 0.011189979501068592, + 0.005885344929993153, + 0.006973695009946823, + 0.030731480568647385, + 0.01570926420390606, + 0.00812650565057993, + 0.013906256295740604, + 0.015921302139759064, + 0.13387897610664368, + 0.0129143837839365, + 0.14465531706809998, + 0.0036532857920974493, + 0.004937987774610519, + 0.008205205202102661, + 0.004679861478507519, + 0.0062776533886790276, + 0.005854601040482521, + 0.0040151034481823444, + 0.004247183911502361, + 0.004200385883450508, + 0.004134428221732378, + 0.004541509784758091, + 0.002998724579811096, + 0.0037048282101750374, + 0.0039032974746078253, + 0.00505115045234561, + 0.005175905767828226, + 0.009616442956030369, + 0.01002297829836607, + 0.00928465649485588, + 0.00787513330578804, + 0.01679343357682228, + 0.016742806881666183, + 0.009555920958518982, + 0.009416485205292702, + 0.0071866512298583984, + 0.002606945112347603, + 0.0015986409271135926, + 0.0021299934014678, + 0.0031656986102461815, + 0.009456143714487553, + 0.004311488941311836, + 0.002947787521407008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.07002881914377213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637230679392815 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.007637230679392815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637230679392815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46510425209999084, + 0.02246866375207901, + 0.050849009305238724, + 0.15071675181388855, + 0.15084271132946014, + 0.05493064969778061, + 0.01709318719804287, + 0.026332538574934006, + 0.054982807487249374, + 0.006679490674287081 + ], + "credit_map": [ + 0.07188508659601212, + 0.0234660841524601, + 0.044744547456502914, + 0.0352938249707222, + 0.029540205374360085, + 0.02414090745151043, + 0.08956212550401688, + 0.14647147059440613, + 0.002379771089181304, + 0.003136246232315898, + 0.0027032752986997366, + 0.003082419978454709, + 0.003592099528759718, + 0.003012892557308078, + 0.002254301216453314, + 0.0023076573852449656, + 0.006398113444447517, + 0.004863859619945288, + 0.01028600987046957, + 0.005414056126028299, + 0.005316628143191338, + 0.006383956875652075, + 0.006200733128935099, + 0.0059856511652469635, + 0.01678292639553547, + 0.017292017117142677, + 0.020379193127155304, + 0.016831956803798676, + 0.0160699263215065, + 0.022915836423635483, + 0.016946088522672653, + 0.02349880523979664, + 0.01886921003460884, + 0.01241519395262003, + 0.028531353920698166, + 0.013216842897236347, + 0.01676863245666027, + 0.02052430436015129, + 0.024591047316789627, + 0.01592612825334072, + 0.004455875139683485, + 0.0032909091096371412, + 0.0025032153353095055, + 0.0032871910370886326, + 0.003254610113799572, + 0.01681908592581749, + 0.002926314016804099, + 0.01839345134794712, + 0.001840405398979783, + 0.002430130960419774, + 0.003274981863796711, + 0.002009386895224452, + 0.0020594908855855465, + 0.0021388898603618145, + 0.001628437777981162, + 0.0017114650690928102, + 0.0037346710450947285, + 0.0022287871688604355, + 0.002537206280976534, + 0.0026821643114089966, + 0.002592140808701515, + 0.0033578085713088512, + 0.005141810979694128, + 0.0040579503402113914, + 0.0058637759648263454, + 0.005804297048598528, + 0.0070542823523283005, + 0.004469654522836208, + 0.00949926022440195, + 0.008961793035268784, + 0.007207588292658329, + 0.0061221555806696415, + 0.0009798728860914707, + 0.0008233063854277134, + 0.0005028154700994492, + 0.0008643539040349424, + 0.0007373699336312711, + 0.0009468203061260283, + 0.0009115721913985908, + 0.0009133793064393103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.007637230679392815 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307984188199043 + ], + "opposite_logprob_tokens": [ + -3.1680798530578613 + ], + "logprob_reward": -0.04307984188199043, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307984188199043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15134452283382416, + 0.048289623111486435, + 0.0671800971031189, + 0.10986432433128357, + 0.050628166645765305, + 0.09673740714788437, + 0.08027514070272446, + 0.14676342904567719, + 0.17390677332878113, + 0.0750105232000351 + ], + "credit_map": [ + 0.025421738624572754, + 0.016039397567510605, + 0.018193088471889496, + 0.01552543044090271, + 0.015396994538605213, + 0.010351421311497688, + 0.03462296351790428, + 0.015793487429618835, + 0.0057642352767288685, + 0.005668749567121267, + 0.005880364216864109, + 0.005977511405944824, + 0.006079858634620905, + 0.005920188035815954, + 0.007277441211044788, + 0.00572127103805542, + 0.008474776521325111, + 0.005918862298130989, + 0.015016929246485233, + 0.00875761080533266, + 0.005862427409738302, + 0.010160287842154503, + 0.0050897798500955105, + 0.007899426855146885, + 0.015033233910799026, + 0.00982161145657301, + 0.0184602290391922, + 0.010022404603660107, + 0.011196723207831383, + 0.018086135387420654, + 0.009859834797680378, + 0.017384150996804237, + 0.0059389397501945496, + 0.006117039825767279, + 0.009507903829216957, + 0.004420156124979258, + 0.007396132219582796, + 0.005419584922492504, + 0.00544336624443531, + 0.006385043263435364, + 0.009996960870921612, + 0.004594007041305304, + 0.00426896195858717, + 0.004852627404034138, + 0.005913943517953157, + 0.027986647561192513, + 0.005072948057204485, + 0.034051313996315, + 0.009746557101607323, + 0.006751595064997673, + 0.015651166439056396, + 0.007247278466820717, + 0.013860147446393967, + 0.010556281544268131, + 0.006974690593779087, + 0.009487420320510864, + 0.02718510664999485, + 0.01972769945859909, + 0.017276346683502197, + 0.012156713753938675, + 0.017105428501963615, + 0.01613551191985607, + 0.017745940014719963, + 0.019430693238973618, + 0.01864919438958168, + 0.016317447647452354, + 0.018786007538437843, + 0.02346254140138626, + 0.0297989621758461, + 0.027452252805233, + 0.02240341156721115, + 0.017036955803632736, + 0.01618214137852192, + 0.007265961263328791, + 0.0032589274924248457, + 0.005128559190779924, + 0.008233588188886642, + 0.022099876776337624, + 0.007443100679665804, + 0.005398368928581476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.04307984188199043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575417205691338 + ], + "opposite_logprob_tokens": [ + -3.895575523376465 + ], + "logprob_reward": -0.020575417205691338, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575417205691338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32012882828712463, + 0.018417296931147575, + 0.13657675683498383, + 0.062487050890922546, + 0.09794575721025467, + 0.1672748178243637, + 0.04272989183664322, + 0.030670538544654846, + 0.06691176444292068, + 0.05685734003782272 + ], + "credit_map": [ + 0.04076781123876572, + 0.013652700930833817, + 0.029732437804341316, + 0.01959175430238247, + 0.015617049299180508, + 0.015745727345347404, + 0.07841041684150696, + 0.10661091655492783, + 0.002883837092667818, + 0.002364457119256258, + 0.0018696775659918785, + 0.0029375171288847923, + 0.002064227592200041, + 0.0016290663043037057, + 0.002596346428617835, + 0.0020721671171486378, + 0.01806788146495819, + 0.011795736849308014, + 0.02407124824821949, + 0.015318459831178188, + 0.01402023620903492, + 0.018869759514927864, + 0.013073408976197243, + 0.021360019221901894, + 0.009800567291676998, + 0.0045679835602641106, + 0.0098565137013793, + 0.0052779181860387325, + 0.007050583604723215, + 0.0083466786891222, + 0.008881526067852974, + 0.008705281652510166, + 0.013960233889520168, + 0.005603406112641096, + 0.023347845301032066, + 0.00784347951412201, + 0.011695747263729572, + 0.015599982813000679, + 0.009281767532229424, + 0.01061328873038292, + 0.02023382857441902, + 0.02524811401963234, + 0.01899109221994877, + 0.018127083778381348, + 0.025777621194720268, + 0.030312931165099144, + 0.014985978603363037, + 0.013598162680864334, + 0.004690345376729965, + 0.0036775495391339064, + 0.010276573710143566, + 0.004483977798372507, + 0.007987884804606438, + 0.0049134837463498116, + 0.0029212720692157745, + 0.0037788073532283306, + 0.004903116729110479, + 0.0025460333563387394, + 0.003946857526898384, + 0.0027552845422178507, + 0.0032408370170742273, + 0.003637033049017191, + 0.0055703832767903805, + 0.004070993047207594, + 0.0065190489403903484, + 0.00924890860915184, + 0.006356276571750641, + 0.006000186782330275, + 0.012040959671139717, + 0.011628685519099236, + 0.008588462136685848, + 0.0065292418003082275, + 0.009190312586724758, + 0.006721593905240297, + 0.00339060858823359, + 0.0062349517829716206, + 0.007560453377664089, + 0.011531373485922813, + 0.006980370730161667, + 0.005247674882411957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.020575417205691338 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a low angle, gently tilts upward and pans to the right, tracking a creature soaring through the sky. A large, bright full moon dominates the background, its cratered silver surface rendered in razor-sharp detail. A massive, ancient dragon circles before it with a deliberate grace and deliberate slowness. Its obsidian-smooth scales trace a graceful arc, its massive, leathery wings flapping with a powerful yet deliberate rhythm, each downward stroke silently and powerfully displacing the air. The full moon's cool, bright light serves as a powerful backlight, casting a sharp silver rimlight across the dragon's silhouette. The light sweeps across its body, causing its black scales to shimmer with an iridescent luster as it moves, like flowing metal. Wisps of wispy clouds drift lazily across the deep, velvety night sky, further emphasizing the dragon's immense size and height. The entire scene is filled with profound ancient majesty and quiet transcendental power, creating a breathtaking, mythical atmosphere.", + "seed": 7754, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_15/index_4225_seed_7754_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11515146843157709, + "vlm_logprob": -0.11515146843157709, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.11515146843157709, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25738024711608887 + ], + "opposite_logprob_tokens": [ + -1.5073802471160889 + ], + "logprob_reward": -0.25738024711608887, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25738024711608887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07345907390117645, + 0.08686838299036026, + 0.0740387961268425, + 0.3819151222705841, + 0.19103869795799255, + 0.05331519618630409, + 0.0454988069832325, + 0.009946689009666443, + 0.024039002135396004, + 0.05988021567463875 + ], + "credit_map": [ + 0.00665229232981801, + 0.01639382913708687, + 0.008731341920793056, + 0.008501050062477589, + 0.005373148713260889, + 0.007796389516443014, + 0.00983508862555027, + 0.010175943374633789, + 0.0056487787514925, + 0.007600088138133287, + 0.008519068360328674, + 0.004943158943206072, + 0.005998078268021345, + 0.024657785892486572, + 0.009352007880806923, + 0.02014942653477192, + 0.005802280735224485, + 0.0023707496002316475, + 0.014059750363230705, + 0.0032766000367701054, + 0.001914903405122459, + 0.015267597511410713, + 0.005126735661178827, + 0.026220185682177544, + 0.03413918986916542, + 0.017093585804104805, + 0.050010886043310165, + 0.01149770524352789, + 0.015241546556353569, + 0.10070332139730453, + 0.019416557624936104, + 0.13381235301494598, + 0.01623770408332348, + 0.0038838237524032593, + 0.07927542924880981, + 0.004944267682731152, + 0.0049635604955255985, + 0.011618085205554962, + 0.0032490871381014585, + 0.06686674803495407, + 0.009534606710076332, + 0.006044217385351658, + 0.00927449855953455, + 0.005957713816314936, + 0.0043889256194233894, + 0.0073836189694702625, + 0.004115101881325245, + 0.006616518832743168, + 0.007197883911430836, + 0.004269229248166084, + 0.006852254271507263, + 0.003787109162658453, + 0.003311104141175747, + 0.008988458663225174, + 0.004025168716907501, + 0.007067601662129164, + 0.0018154438585042953, + 0.0011815059697255492, + 0.0011329347034916282, + 0.0009454721584916115, + 0.0009984972421079874, + 0.0012345552677288651, + 0.001124334055930376, + 0.001513947150669992, + 0.0034200043883174658, + 0.0011554667726159096, + 0.0050981552340090275, + 0.0020308969542384148, + 0.002892718417569995, + 0.0040610493160784245, + 0.00210268865339458, + 0.0032780237961560488, + 0.006416010204702616, + 0.008148889057338238, + 0.00823123101145029, + 0.0041473181918263435, + 0.010104884393513203, + 0.012744258157908916, + 0.003008094150573015, + 0.0070795356296002865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.25738024711608887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990797758102417 + ], + "opposite_logprob_tokens": [ + -1.804908037185669 + ], + "logprob_reward": -0.17990797758102417, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990797758102417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018539849668741226, + 0.021863926202058792, + 0.01875842735171318, + 0.4448646605014801, + 0.4132463037967682, + 0.022185906767845154, + 0.023534782230854034, + 0.009762095287442207, + 0.019780052825808525, + 0.007463973015546799 + ], + "credit_map": [ + 0.0028572978917509317, + 0.002168923383578658, + 0.0029137837700545788, + 0.0017188984202221036, + 0.0012947958894073963, + 0.0025620574597269297, + 0.002564269583672285, + 0.0024598250165581703, + 0.000928295950870961, + 0.0009407242177985609, + 0.0015080912271514535, + 0.001169021474197507, + 0.0010853480780497193, + 0.007274379953742027, + 0.001455345074646175, + 0.007502720225602388, + 0.0015868778573349118, + 0.0010101289954036474, + 0.002843243069946766, + 0.0007697074906900525, + 0.0007130688754841685, + 0.004048571456223726, + 0.0018626474775373936, + 0.005924181547015905, + 0.03897079825401306, + 0.018674977123737335, + 0.05145496875047684, + 0.011945204809308052, + 0.016904447227716446, + 0.1291804015636444, + 0.015892717987298965, + 0.16184115409851074, + 0.03422890603542328, + 0.005674728192389011, + 0.17734910547733307, + 0.009547468274831772, + 0.003555134404450655, + 0.025575783103704453, + 0.00528649752959609, + 0.15202867984771729, + 0.0016753956442698836, + 0.0016399776795879006, + 0.007229579612612724, + 0.0013617402873933315, + 0.0007803302141837776, + 0.0017174402019008994, + 0.0006915713893249631, + 0.0070898705162107944, + 0.002926100743934512, + 0.0014914919156581163, + 0.005450595170259476, + 0.0015222576912492514, + 0.0015075111296027899, + 0.006192633882164955, + 0.002190785249695182, + 0.002253406448289752, + 0.0011925228172913194, + 0.001124347560107708, + 0.0010818245355039835, + 0.0010711662471294403, + 0.0010208019521087408, + 0.0012660701759159565, + 0.0012038899585604668, + 0.001801471458747983, + 0.001083256327547133, + 0.0009128284291364253, + 0.006449554581195116, + 0.0014533267822116613, + 0.0015901169972494245, + 0.005994673818349838, + 0.0011815120233222842, + 0.0011147841578349471, + 0.0010580241214483976, + 0.0012500741286203265, + 0.000989584717899561, + 0.0006319439271464944, + 0.0010706802131608129, + 0.001232335576787591, + 0.00047718422138132155, + 0.0007541460799984634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.17990797758102417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022046088706701994 + ], + "opposite_logprob_tokens": [ + -6.127204418182373 + ], + "logprob_reward": -0.0022046088706701994, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022046088706701994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050379518419504166, + 0.10000024735927582, + 0.17012862861156464, + 0.3507493734359741, + 0.2278858870267868, + 0.028383145108819008, + 0.022709807381033897, + 0.009385118260979652, + 0.01481518242508173, + 0.025563083589076996 + ], + "credit_map": [ + 0.0062713962979614735, + 0.010430734604597092, + 0.006066750735044479, + 0.007129118777811527, + 0.003045290242880583, + 0.005266947206109762, + 0.006570721045136452, + 0.005598555784672499, + 0.005178520455956459, + 0.005109522491693497, + 0.007601512596011162, + 0.004790826700627804, + 0.005317683331668377, + 0.03367342799901962, + 0.008611557073891163, + 0.029717199504375458, + 0.012748991139233112, + 0.004505131859332323, + 0.0323135107755661, + 0.005391932092607021, + 0.004262455273419619, + 0.03753476217389107, + 0.009442497044801712, + 0.06392934173345566, + 0.030636489391326904, + 0.014218972064554691, + 0.0414765290915966, + 0.00933051947504282, + 0.012650497257709503, + 0.10136739909648895, + 0.012995380908250809, + 0.128073588013649, + 0.018851052969694138, + 0.0031293199863284826, + 0.09718496352434158, + 0.005180108826607466, + 0.002957603195682168, + 0.013450517319142818, + 0.0029575922526419163, + 0.0841747373342514, + 0.0051398249343037605, + 0.0038183375727385283, + 0.004046884831041098, + 0.0035230808425694704, + 0.002415537601336837, + 0.00391477532684803, + 0.0022849796805530787, + 0.0032397243194282055, + 0.003682647831737995, + 0.0020328271202743053, + 0.0035212445072829723, + 0.0015002215513959527, + 0.0017665865598246455, + 0.003336674300953746, + 0.002753067761659622, + 0.004116537049412727, + 0.001280918368138373, + 0.000787685567047447, + 0.0011579387355595827, + 0.0011309213005006313, + 0.0011109717888757586, + 0.001207202672958374, + 0.001085100113414228, + 0.0016243795398622751, + 0.0013242895947769284, + 0.0011116542154923081, + 0.003083746414631605, + 0.0016717980615794659, + 0.0019360161386430264, + 0.00260179047472775, + 0.0014373903395608068, + 0.0016484960215166211, + 0.0026467759162187576, + 0.003975552041083574, + 0.0031396395061165094, + 0.001879650168120861, + 0.00397566007450223, + 0.004849580582231283, + 0.0015842070570215583, + 0.0035120167303830385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.0022046088706701994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298132687807083 + ], + "opposite_logprob_tokens": [ + -2.2379813194274902 + ], + "logprob_reward": -0.11298132687807083, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298132687807083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020817479118704796, + 0.019083036109805107, + 0.008944500237703323, + 0.13278363645076752, + 0.7311514019966125, + 0.018683040514588356, + 0.04910755157470703, + 0.005092311650514603, + 0.0064966389909386635, + 0.007840442471206188 + ], + "credit_map": [ + 0.003125897143036127, + 0.003401780966669321, + 0.00267921295017004, + 0.0023028231225907803, + 0.0012117853621020913, + 0.0026727644726634026, + 0.0027421670965850353, + 0.0026810476556420326, + 0.0018400093540549278, + 0.001974798273295164, + 0.001776146120391786, + 0.0012469214852899313, + 0.0014938092790544033, + 0.0037680275272578, + 0.002729360479861498, + 0.004253963474184275, + 0.000795039813965559, + 0.000914926000405103, + 0.0012639528140425682, + 0.0007313736714422703, + 0.0007966877892613411, + 0.0018405279843136668, + 0.0012609019177034497, + 0.0013410903047770262, + 0.011871157214045525, + 0.006016024854034185, + 0.01642632856965065, + 0.003876605536788702, + 0.0054858592338860035, + 0.03704224154353142, + 0.005094300489872694, + 0.04697111248970032, + 0.06082717329263687, + 0.00964271742850542, + 0.31393495202064514, + 0.016732212156057358, + 0.007330409251153469, + 0.04401475191116333, + 0.00925592053681612, + 0.2694132328033447, + 0.0022910083644092083, + 0.002124210586771369, + 0.004007242154330015, + 0.0017486931756138802, + 0.0011267727240920067, + 0.0023446606937795877, + 0.0010264189913868904, + 0.0040140352211892605, + 0.004335193894803524, + 0.003840535180643201, + 0.014849408529698849, + 0.002703757956624031, + 0.0037249610759317875, + 0.01089541893452406, + 0.0028623808175325394, + 0.005895892158150673, + 0.0008884947746992111, + 0.000484865449834615, + 0.0006508856313303113, + 0.00042186531936749816, + 0.0005928017199039459, + 0.000640113721601665, + 0.0005639194860123098, + 0.0008493653149344027, + 0.0007072799489833415, + 0.0002871706092264503, + 0.0013596907956525683, + 0.0005302266799844801, + 0.0007388780941255391, + 0.001617206260561943, + 0.000511312682647258, + 0.0007448741816915572, + 0.0007776210550218821, + 0.0009669269784353673, + 0.0013094805181026459, + 0.0005393961910158396, + 0.001243597362190485, + 0.0015483751194551587, + 0.000373738439520821, + 0.0010813067201524973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.11298132687807083 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283181712031364 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283181712031364, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283181712031364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031985994428396225, + 0.01843000389635563, + 0.0488990880548954, + 0.44019022583961487, + 0.33746659755706787, + 0.03423114866018295, + 0.040865227580070496, + 0.00901090819388628, + 0.01831912435591221, + 0.020601661875844002 + ], + "credit_map": [ + 0.004036962054669857, + 0.006728949025273323, + 0.003984767943620682, + 0.004854833707213402, + 0.0016739871352910995, + 0.0035476998891681433, + 0.003784783883020282, + 0.003374011255800724, + 0.0012205413077026606, + 0.0011495642829686403, + 0.0020768919494003057, + 0.001308814506046474, + 0.0010255567030981183, + 0.005370980128645897, + 0.0018577207811176777, + 0.004419935401529074, + 0.0041272700764238834, + 0.0014609319623559713, + 0.00957600399851799, + 0.0017287038499489427, + 0.0010808598017320037, + 0.01042980793863535, + 0.002818258246406913, + 0.01767725497484207, + 0.03891920670866966, + 0.01835153065621853, + 0.05101161450147629, + 0.011716995388269424, + 0.016761578619480133, + 0.12843824923038483, + 0.015157009474933147, + 0.15983404219150543, + 0.0279551949352026, + 0.004387798719108105, + 0.14455482363700867, + 0.0076494417153298855, + 0.004097426310181618, + 0.020008321851491928, + 0.004387914203107357, + 0.12442568689584732, + 0.0023433298338204622, + 0.0023708920925855637, + 0.011917232535779476, + 0.0017986756283789873, + 0.0010997208300977945, + 0.0020660904701799154, + 0.0008929898031055927, + 0.011742215603590012, + 0.005673837848007679, + 0.005209013354033232, + 0.006283389870077372, + 0.0035964353010058403, + 0.0042594061233103275, + 0.005824308842420578, + 0.003919614944607019, + 0.006099223159253597, + 0.0014379817293956876, + 0.0008579473360441625, + 0.0008130663773044944, + 0.0008356605540029705, + 0.0013423115015029907, + 0.0012226776452735066, + 0.0010457064490765333, + 0.0014555572997778654, + 0.0018187450477853417, + 0.0006432574591599405, + 0.004250805359333754, + 0.0015211703721433878, + 0.0017065884312614799, + 0.004824470728635788, + 0.0012192775029689074, + 0.0023348084650933743, + 0.0026181088760495186, + 0.002515999833121896, + 0.0031966192182153463, + 0.0011900760000571609, + 0.0029188371263444424, + 0.004524250980466604, + 0.0010070777498185635, + 0.002630692208185792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.023283181712031364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a low angle, gently tilts upward and pans to the right, tracking a creature soaring through the sky. A large, bright full moon dominates the background, its cratered silver surface rendered in razor-sharp detail. A massive, ancient dragon circles before it with a deliberate grace and deliberate slowness. Its obsidian-smooth scales trace a graceful arc, its massive, leathery wings flapping with a powerful yet deliberate rhythm, each downward stroke silently and powerfully displacing the air. The full moon's cool, bright light serves as a powerful backlight, casting a sharp silver rimlight across the dragon's silhouette. The light sweeps across its body, causing its black scales to shimmer with an iridescent luster as it moves, like flowing metal. Wisps of wispy clouds drift lazily across the deep, velvety night sky, further emphasizing the dragon's immense size and height. The entire scene is filled with profound ancient majesty and quiet transcendental power, creating a breathtaking, mythical atmosphere.", + "seed": 7755, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_15/index_4225_seed_7755_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.053283601603470745, + "vlm_logprob": -0.053283601603470745, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.053283601603470745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14583586156368256 + ], + "opposite_logprob_tokens": [ + -2.0208358764648438 + ], + "logprob_reward": -0.14583586156368256, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14583586156368256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1802908033132553, + 0.10454938560724258, + 0.12490545213222504, + 0.02749445289373398, + 0.028552990406751633, + 0.03572535887360573, + 0.05168375000357628, + 0.06369124352931976, + 0.11789193749427795, + 0.2652145326137543 + ], + "credit_map": [ + 0.020877627655863762, + 0.012311112135648727, + 0.024497153237462044, + 0.02146897092461586, + 0.011979732662439346, + 0.02756269834935665, + 0.02454525977373123, + 0.03704826906323433, + 0.011539553292095661, + 0.010323146358132362, + 0.008010908029973507, + 0.011611733585596085, + 0.011741997674107552, + 0.01570742018520832, + 0.014923975802958012, + 0.020690664649009705, + 0.011648104526102543, + 0.007459226064383984, + 0.006027486640959978, + 0.007078514434397221, + 0.0060943663120269775, + 0.03448733314871788, + 0.006398583296686411, + 0.04571186751127243, + 0.003657063702121377, + 0.002971136011183262, + 0.0036776154302060604, + 0.0033572667744010687, + 0.0015347994631156325, + 0.004103927873075008, + 0.00262443907558918, + 0.005568208638578653, + 0.004406031686812639, + 0.004236336797475815, + 0.003477769438177347, + 0.0020638250280171633, + 0.002009194577112794, + 0.003636379027739167, + 0.0032311438117176294, + 0.005492315627634525, + 0.004231503698974848, + 0.004778528586030006, + 0.004443873185664415, + 0.0031580531504005194, + 0.00419368501752615, + 0.003673510393127799, + 0.005326994229108095, + 0.005919211078435183, + 0.006514404434710741, + 0.004459389019757509, + 0.005123192444443703, + 0.006476745940744877, + 0.0042162248864769936, + 0.008620820939540863, + 0.004286031238734722, + 0.011986950412392616, + 0.009856109507381916, + 0.00738159054890275, + 0.010500162839889526, + 0.007148937322199345, + 0.00618613138794899, + 0.009224886074662209, + 0.006109780631959438, + 0.007283658720552921, + 0.02573787234723568, + 0.011247615329921246, + 0.017647668719291687, + 0.01223108172416687, + 0.009837635792791843, + 0.019245190545916557, + 0.007407040800899267, + 0.014537843875586987, + 0.02233847975730896, + 0.024389995262026787, + 0.04344268515706062, + 0.03693240135908127, + 0.040379393845796585, + 0.042277731001377106, + 0.029894955456256866, + 0.02555893547832966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.14583586156368256 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207270920276642 + ], + "opposite_logprob_tokens": [ + -2.81207275390625 + ], + "logprob_reward": -0.06207270920276642, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207270920276642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11604362726211548, + 0.04278826341032982, + 0.08212093263864517, + 0.05861291661858559, + 0.05185011401772499, + 0.06145869195461273, + 0.0684407502412796, + 0.0823742002248764, + 0.16499531269073486, + 0.2713151276111603 + ], + "credit_map": [ + 0.013294260948896408, + 0.010791354812681675, + 0.019417377188801765, + 0.010858898051083088, + 0.010683107189834118, + 0.014739984646439552, + 0.016612283885478973, + 0.019646363332867622, + 0.004703052341938019, + 0.009677675552666187, + 0.003238026285544038, + 0.006350060924887657, + 0.004026117734611034, + 0.004509173799306154, + 0.004517396911978722, + 0.005766760092228651, + 0.00881994143128395, + 0.006045324262231588, + 0.0060739764012396336, + 0.007372604217380285, + 0.006584973074495792, + 0.016384799033403397, + 0.009325088001787663, + 0.021514220163226128, + 0.007279298733919859, + 0.0070650228299200535, + 0.006402018945664167, + 0.00497270654886961, + 0.0048243640922009945, + 0.008772527799010277, + 0.005163145717233419, + 0.014133831486105919, + 0.007432871963828802, + 0.006998204626142979, + 0.005436742212623358, + 0.005118956323713064, + 0.0035046839620918036, + 0.0064947642385959625, + 0.006224163342267275, + 0.010639729909598827, + 0.008750895969569683, + 0.0076624080538749695, + 0.008314643055200577, + 0.00646150903776288, + 0.0063992515206336975, + 0.0071206591092050076, + 0.006848035380244255, + 0.009901289828121662, + 0.005520987324416637, + 0.005206620786339045, + 0.005952056497335434, + 0.00894583947956562, + 0.005014523398131132, + 0.012689914554357529, + 0.005934278946369886, + 0.019176531583070755, + 0.012020325288176537, + 0.009427139535546303, + 0.010740836150944233, + 0.012726228684186935, + 0.008480614051222801, + 0.013125083409249783, + 0.008189283311367035, + 0.007664685603231192, + 0.030442887917160988, + 0.014265940524637699, + 0.02897808700799942, + 0.014004072174429893, + 0.02119949832558632, + 0.02697625569999218, + 0.012152429670095444, + 0.016976136714220047, + 0.024229789152741432, + 0.03596944734454155, + 0.05260482430458069, + 0.03672963008284569, + 0.026265911757946014, + 0.04792024940252304, + 0.024236619472503662, + 0.02335866540670395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.06207270920276642 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011857867939397693 + ], + "opposite_logprob_tokens": [ + -6.751185894012451 + ], + "logprob_reward": -0.0011857867939397693, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011857867939397693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08751893043518066, + 0.0812455266714096, + 0.5762557983398438, + 0.02247503772377968, + 0.049447111785411835, + 0.02651190757751465, + 0.03404257073998451, + 0.03449032083153725, + 0.03782710060477257, + 0.050185661762952805 + ], + "credit_map": [ + 0.008726957254111767, + 0.009858239442110062, + 0.009294919669628143, + 0.009674605913460255, + 0.007386196870356798, + 0.014090759679675102, + 0.010608576238155365, + 0.01787867583334446, + 0.007897727191448212, + 0.009832821786403656, + 0.00479563744738698, + 0.007805052679032087, + 0.006634288467466831, + 0.015938106924295425, + 0.007175438106060028, + 0.021166447550058365, + 0.05581478402018547, + 0.02533378079533577, + 0.023382162675261497, + 0.030241820961236954, + 0.026659199967980385, + 0.16465483605861664, + 0.029479270800948143, + 0.22068995237350464, + 0.002586295595392585, + 0.004351625684648752, + 0.0028863882180303335, + 0.002929783659055829, + 0.0022386545315384865, + 0.0023022594396024942, + 0.002870005089789629, + 0.002310023410245776, + 0.006893557030707598, + 0.006902717053890228, + 0.004493371583521366, + 0.0064518041908741, + 0.005130853038281202, + 0.00495641166344285, + 0.00778432609513402, + 0.006834065541625023, + 0.004083351232111454, + 0.0029859060887247324, + 0.003177566919475794, + 0.0021443883888423443, + 0.0024354709312319756, + 0.0036005363799631596, + 0.0030185880605131388, + 0.0050660972483456135, + 0.004951222334057093, + 0.0032659436110407114, + 0.0037269957829266787, + 0.005851496942341328, + 0.00322250509634614, + 0.004357925150543451, + 0.0032751637045294046, + 0.005391317419707775, + 0.0035437713377177715, + 0.0044242884032428265, + 0.005924933589994907, + 0.0057537793181836605, + 0.003929463680833578, + 0.0040718805976212025, + 0.00284013245254755, + 0.004002070520073175, + 0.004841150715947151, + 0.0021683021914213896, + 0.009394901804625988, + 0.0043520694598555565, + 0.0031000380404293537, + 0.007438456639647484, + 0.002734128152951598, + 0.003798052668571472, + 0.005431441590189934, + 0.004966392647475004, + 0.008147824555635452, + 0.007701203227043152, + 0.00751158082857728, + 0.006757055874913931, + 0.00496886158362031, + 0.004701298661530018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0011857867939397693 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048686910420656204 + ], + "opposite_logprob_tokens": [ + -3.048686981201172 + ], + "logprob_reward": -0.048686910420656204, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048686910420656204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1446770429611206, + 0.09313343465328217, + 0.17380918562412262, + 0.05011625215411186, + 0.032891642302274704, + 0.04008302092552185, + 0.06472630053758621, + 0.12323536723852158, + 0.14694800972938538, + 0.13037973642349243 + ], + "credit_map": [ + 0.014761514030396938, + 0.013427983969449997, + 0.018416084349155426, + 0.017604518681764603, + 0.01684260182082653, + 0.0179461557418108, + 0.02251008339226246, + 0.02316809818148613, + 0.009986123070120811, + 0.01354298647493124, + 0.006966003216803074, + 0.012462652288377285, + 0.011484634131193161, + 0.011559754610061646, + 0.01226526778191328, + 0.014866016805171967, + 0.01736682467162609, + 0.00762588856741786, + 0.008685880340635777, + 0.005576334428042173, + 0.007256551645696163, + 0.05088682472705841, + 0.008964122273027897, + 0.0674467608332634, + 0.0049414667300879955, + 0.00815102644264698, + 0.00487689021974802, + 0.00707929627969861, + 0.004472802858799696, + 0.00637234328314662, + 0.005654574371874332, + 0.008567850105464458, + 0.0059875305742025375, + 0.004062923602759838, + 0.003633461892604828, + 0.0035652779042720795, + 0.002503829775378108, + 0.0032311398535966873, + 0.004026115406304598, + 0.0058813681825995445, + 0.004884655587375164, + 0.003972902428358793, + 0.00606981385499239, + 0.0039926753379404545, + 0.0035965617280453444, + 0.005735923070460558, + 0.004925027955323458, + 0.006905458867549896, + 0.011814702302217484, + 0.008538540452718735, + 0.005808035843074322, + 0.007346320431679487, + 0.005411670543253422, + 0.008513434790074825, + 0.005817114375531673, + 0.011476477608084679, + 0.01348220743238926, + 0.01391622144728899, + 0.01682525873184204, + 0.01650932990014553, + 0.0193962249904871, + 0.016264772042632103, + 0.013806717470288277, + 0.013034647330641747, + 0.015850810334086418, + 0.010339162312448025, + 0.0393521785736084, + 0.017135178670287132, + 0.013517231680452824, + 0.030781246721744537, + 0.0079084737226367, + 0.012063713744282722, + 0.01335237268358469, + 0.012539439834654331, + 0.0212862528860569, + 0.021581171080470085, + 0.013991308398544788, + 0.016507800668478012, + 0.01666969619691372, + 0.01445169746875763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.048686910420656204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12443187087774277, + 0.08042977750301361, + 0.1796465367078781, + 0.023804740980267525, + 0.039316073060035706, + 0.0505097471177578, + 0.08370159566402435, + 0.12128335237503052, + 0.06304391473531723, + 0.2338324338197708 + ], + "credit_map": [ + 0.013116135261952877, + 0.0104008037596941, + 0.017840126529335976, + 0.013165457174181938, + 0.010754143819212914, + 0.016925817355513573, + 0.01871674880385399, + 0.023512620478868484, + 0.007820832543075085, + 0.007668098900467157, + 0.006924233864992857, + 0.006938891019672155, + 0.00574274268001318, + 0.015505645424127579, + 0.0068484023213386536, + 0.022980930283665657, + 0.01862207055091858, + 0.008375500328838825, + 0.00935815554112196, + 0.007250902242958546, + 0.00856035016477108, + 0.049993205815553665, + 0.010402016341686249, + 0.06708431243896484, + 0.002604403765872121, + 0.0032761034090071917, + 0.0022539393976330757, + 0.0013431648258119822, + 0.0017377672484144568, + 0.004253368824720383, + 0.002023198874667287, + 0.006312794052064419, + 0.005233762785792351, + 0.005294577218592167, + 0.003915728069841862, + 0.002596202539280057, + 0.0028386355843394995, + 0.0058969538658857346, + 0.003574070753529668, + 0.009966141544282436, + 0.006274878978729248, + 0.005180065054446459, + 0.004843426868319511, + 0.00512685626745224, + 0.005047076381742954, + 0.006280616857111454, + 0.008744562976062298, + 0.009012260474264622, + 0.01426753867417574, + 0.0063826944679021835, + 0.007517898455262184, + 0.006157116033136845, + 0.006653091870248318, + 0.014623934403061867, + 0.006684517953544855, + 0.021414803341031075, + 0.009402004070580006, + 0.010783825069665909, + 0.03447287902235985, + 0.013319349847733974, + 0.009075245819985867, + 0.020944494754076004, + 0.007518390659242868, + 0.015767155215144157, + 0.008830681443214417, + 0.004809195641428232, + 0.013919341377913952, + 0.008630367927253246, + 0.006837812718003988, + 0.011292188428342342, + 0.003543831640854478, + 0.005180492997169495, + 0.02193276956677437, + 0.02971392683684826, + 0.0449582077562809, + 0.024295741692185402, + 0.027808785438537598, + 0.04454934224486351, + 0.01836461015045643, + 0.022209035232663155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.008636740036308765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a low angle, gently tilts upward and pans to the right, tracking a creature soaring through the sky. A large, bright full moon dominates the background, its cratered silver surface rendered in razor-sharp detail. A massive, ancient dragon circles before it with a deliberate grace and deliberate slowness. Its obsidian-smooth scales trace a graceful arc, its massive, leathery wings flapping with a powerful yet deliberate rhythm, each downward stroke silently and powerfully displacing the air. The full moon's cool, bright light serves as a powerful backlight, casting a sharp silver rimlight across the dragon's silhouette. The light sweeps across its body, causing its black scales to shimmer with an iridescent luster as it moves, like flowing metal. Wisps of wispy clouds drift lazily across the deep, velvety night sky, further emphasizing the dragon's immense size and height. The entire scene is filled with profound ancient majesty and quiet transcendental power, creating a breathtaking, mythical atmosphere.", + "seed": 7756, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_15/index_4225_seed_7756_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029326659243088214, + "vlm_logprob": -0.029326659243088214, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.029326659243088214, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014414937235414982 + ], + "opposite_logprob_tokens": [ + -4.6394147872924805 + ], + "logprob_reward": -0.014414937235414982, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.014414937235414982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1210489422082901, + 0.09464935213327408, + 0.16754639148712158, + 0.0696416050195694, + 0.0988532155752182, + 0.08300571888685226, + 0.08037075400352478, + 0.1504475623369217, + 0.07079100608825684, + 0.06364546716213226 + ], + "credit_map": [ + 0.012779341079294682, + 0.007581715472042561, + 0.020512597635388374, + 0.013978887349367142, + 0.008381886407732964, + 0.014508898369967937, + 0.029384898021817207, + 0.01392071507871151, + 0.009156315587460995, + 0.006005346775054932, + 0.013740126974880695, + 0.009132545441389084, + 0.009244920685887337, + 0.011858542449772358, + 0.01349030714482069, + 0.02202124334871769, + 0.011708348989486694, + 0.03325324133038521, + 0.010896128602325916, + 0.013984033837914467, + 0.01681257039308548, + 0.02247658744454384, + 0.02321244589984417, + 0.03520301356911659, + 0.005446044262498617, + 0.00939524918794632, + 0.004720406141132116, + 0.005582649726420641, + 0.008561454713344574, + 0.010359477251768112, + 0.011839200742542744, + 0.013737117871642113, + 0.010389070957899094, + 0.011721542105078697, + 0.008063186891376972, + 0.00909004732966423, + 0.014044616371393204, + 0.013971773907542229, + 0.020128969103097916, + 0.011444006115198135, + 0.00616834731772542, + 0.006322218105196953, + 0.005361918359994888, + 0.008019509725272655, + 0.010226127691566944, + 0.022146722301840782, + 0.007317030802369118, + 0.017443843185901642, + 0.008995977230370045, + 0.006821468006819487, + 0.012601512484252453, + 0.018439508974552155, + 0.006745601072907448, + 0.010146440006792545, + 0.005081769544631243, + 0.01153846736997366, + 0.009470466524362564, + 0.013633687980473042, + 0.03218024596571922, + 0.012113484553992748, + 0.009082197211682796, + 0.05052032321691513, + 0.005221100989729166, + 0.018226059153676033, + 0.0051297638565301895, + 0.00723975570872426, + 0.01719457097351551, + 0.007852863520383835, + 0.006092258263379335, + 0.014784090220928192, + 0.0073890346102416515, + 0.00510866753757, + 0.007833520881831646, + 0.0040139066986739635, + 0.01659991219639778, + 0.004168277140706778, + 0.004932093434035778, + 0.01527240127325058, + 0.007249387912452221, + 0.0035759652964770794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.014414937235414982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899516820907593 + ], + "opposite_logprob_tokens": [ + -2.5789952278137207 + ], + "logprob_reward": -0.07899516820907593, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899516820907593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07115718722343445, + 0.05482947826385498, + 0.09846410155296326, + 0.06331706047058105, + 0.0702342838048935, + 0.07028545439243317, + 0.15894733369350433, + 0.14788690209388733, + 0.1196506917476654, + 0.1452275514602661 + ], + "credit_map": [ + 0.006451927125453949, + 0.004275430925190449, + 0.013568357564508915, + 0.009077729657292366, + 0.0038861650973558426, + 0.008284230716526508, + 0.01233269926160574, + 0.01328064501285553, + 0.005081018432974815, + 0.005121112335473299, + 0.0110270194709301, + 0.006815838161855936, + 0.005550113040953875, + 0.0064125810749828815, + 0.00860302709043026, + 0.006218764465302229, + 0.007350059691816568, + 0.008006316609680653, + 0.008403811603784561, + 0.005754532292485237, + 0.013964359648525715, + 0.011285357177257538, + 0.017212197184562683, + 0.02648746967315674, + 0.004927200265228748, + 0.008140246383845806, + 0.0046471403911709785, + 0.004845807794481516, + 0.010092925280332565, + 0.01107766292989254, + 0.011426864191889763, + 0.008159209974110126, + 0.005513615906238556, + 0.007796879857778549, + 0.0059545221738517284, + 0.0085062887519598, + 0.008242054842412472, + 0.0104074003174901, + 0.01185951754450798, + 0.011953999288380146, + 0.004122335929423571, + 0.008244865573942661, + 0.004825659561902285, + 0.006118331104516983, + 0.008912435732781887, + 0.01719244383275509, + 0.00626139622181654, + 0.014607991091907024, + 0.02089804969727993, + 0.014738455414772034, + 0.03214770555496216, + 0.03205133602023125, + 0.013439382426440716, + 0.020156487822532654, + 0.009248506277799606, + 0.016267403960227966, + 0.007070422172546387, + 0.01661289669573307, + 0.024479031562805176, + 0.013938767835497856, + 0.008451673202216625, + 0.049517758190631866, + 0.005483932793140411, + 0.022332435473799706, + 0.005646673496812582, + 0.004851686302572489, + 0.04169400408864021, + 0.007582484744489193, + 0.004499396309256554, + 0.044710274785757065, + 0.0051832981407642365, + 0.005482867360115051, + 0.016852740198373795, + 0.016641277819871902, + 0.015531271696090698, + 0.015783336013555527, + 0.01671605370938778, + 0.02393163926899433, + 0.028840789571404457, + 0.0109304403886199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.07899516820907593 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009255892946384847 + ], + "opposite_logprob_tokens": [ + -7.000925540924072 + ], + "logprob_reward": -0.0009255892946384847, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009255892946384847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07932018488645554, + 0.07231631129980087, + 0.1004948616027832, + 0.06923016160726547, + 0.03930221125483513, + 0.06034737452864647, + 0.05559073016047478, + 0.09432126581668854, + 0.3970597982406616, + 0.032017093151807785 + ], + "credit_map": [ + 0.008530659601092339, + 0.005515166092664003, + 0.01547963172197342, + 0.008377579972147942, + 0.0069125499576330185, + 0.009845857508480549, + 0.009208676405251026, + 0.015450065024197102, + 0.0065584611147642136, + 0.004725031089037657, + 0.009471111930906773, + 0.006104719825088978, + 0.004892786033451557, + 0.009613344445824623, + 0.009851428680121899, + 0.021099437028169632, + 0.0071629248559474945, + 0.014783550053834915, + 0.007010621950030327, + 0.007658042479306459, + 0.009118549525737762, + 0.013860671781003475, + 0.014216781593859196, + 0.02668372169137001, + 0.007756748236715794, + 0.009397958405315876, + 0.007559876423329115, + 0.005525453016161919, + 0.009484758600592613, + 0.009997713379561901, + 0.009788122028112411, + 0.009719531051814556, + 0.0034976850729435682, + 0.003667928511276841, + 0.004382808692753315, + 0.0031986068934202194, + 0.0053282794542610645, + 0.00650062644854188, + 0.0056951045989990234, + 0.007031173910945654, + 0.004473825450986624, + 0.005627535283565521, + 0.004615962039679289, + 0.005344657227396965, + 0.007189735770225525, + 0.014725223183631897, + 0.004610513802617788, + 0.01375992875546217, + 0.005611192435026169, + 0.007342031225562096, + 0.007551233749836683, + 0.009547879919409752, + 0.0057530878111720085, + 0.00806865282356739, + 0.004654754884541035, + 0.0070618996396660805, + 0.006126439198851585, + 0.01106914971023798, + 0.01895524188876152, + 0.007611910812556744, + 0.0052844989113509655, + 0.028963861986994743, + 0.004245581105351448, + 0.01206459291279316, + 0.016047492623329163, + 0.01048718299716711, + 0.14853747189044952, + 0.014351001009345055, + 0.011844167485833168, + 0.16760733723640442, + 0.008595215156674385, + 0.01958993822336197, + 0.002090722555294633, + 0.002678332617506385, + 0.006843049544841051, + 0.00268925866112113, + 0.0024342939723283052, + 0.008550959639251232, + 0.004362449981272221, + 0.0023680280428379774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.0009255892946384847 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810923546552658 + ], + "opposite_logprob_tokens": [ + -3.288109302520752 + ], + "logprob_reward": -0.03810923546552658, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810923546552658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1229611411690712, + 0.07439164072275162, + 0.10322257876396179, + 0.0767841562628746, + 0.12707501649856567, + 0.12336057424545288, + 0.06962593644857407, + 0.12389922142028809, + 0.08911391347646713, + 0.08956579864025116 + ], + "credit_map": [ + 0.012927137315273285, + 0.006992975249886513, + 0.03268135339021683, + 0.01842963695526123, + 0.006898332387208939, + 0.016302064061164856, + 0.01833871379494667, + 0.010390928946435452, + 0.00685463473200798, + 0.004639160819351673, + 0.016687050461769104, + 0.00834742747247219, + 0.005576382856816053, + 0.008708028122782707, + 0.012667088769376278, + 0.010911861434578896, + 0.00891135260462761, + 0.0193990059196949, + 0.006790064740926027, + 0.006834833417087793, + 0.013652645982801914, + 0.013325478881597519, + 0.01252541784197092, + 0.02178378403186798, + 0.007927533239126205, + 0.011374272406101227, + 0.009643207304179668, + 0.0047851549461483955, + 0.008333055302500725, + 0.009433256462216377, + 0.013271098956465721, + 0.012016580440104008, + 0.0092331413179636, + 0.013666605576872826, + 0.009262139908969402, + 0.012775758281350136, + 0.02386835403740406, + 0.02213280461728573, + 0.018519192934036255, + 0.017617028206586838, + 0.011344465427100658, + 0.011266372166574001, + 0.009793028235435486, + 0.011658374220132828, + 0.016544140875339508, + 0.027830446138978004, + 0.011274253018200397, + 0.023649495095014572, + 0.006734713912010193, + 0.008024630136787891, + 0.012850098311901093, + 0.010718729346990585, + 0.006980300415307283, + 0.008863898925483227, + 0.006649471819400787, + 0.008804090321063995, + 0.01297138910740614, + 0.026061318814754486, + 0.01794896088540554, + 0.015543794259428978, + 0.014477117918431759, + 0.018668651580810547, + 0.008574334904551506, + 0.00965365581214428, + 0.00724756857380271, + 0.009821489453315735, + 0.016758805140852928, + 0.012821984477341175, + 0.010381155647337437, + 0.012426212430000305, + 0.009853791445493698, + 0.009802906773984432, + 0.007865923456847668, + 0.010449004359543324, + 0.015982331708073616, + 0.008959191851317883, + 0.01059173047542572, + 0.012887721881270409, + 0.01622559130191803, + 0.006604298017919064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.03810923546552658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a massive dragon soaring in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle, gently tilting upward and panning to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, bright full moon with a cratered silver surface visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon have obsidian-smooth scales and massive leathery wings flapping with a deliberate rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp silver rimlight on the dragon's silhouette caused by the full moon's backlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188366010785103 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188366010785103, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188366010785103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08893661946058273, + 0.051155395805835724, + 0.05191918835043907, + 0.035205479711294174, + 0.05537399277091026, + 0.04347281530499458, + 0.06776654720306396, + 0.1324877291917801, + 0.3921276032924652, + 0.0815545991063118 + ], + "credit_map": [ + 0.008914458565413952, + 0.004825699608772993, + 0.019246375188231468, + 0.013207484036684036, + 0.006813903339207172, + 0.012430879287421703, + 0.007546864449977875, + 0.01595095731317997, + 0.005217797122895718, + 0.0027886591851711273, + 0.007019771728664637, + 0.005329366773366928, + 0.004711378365755081, + 0.0075654033571481705, + 0.007122972514480352, + 0.011400045827031136, + 0.0038336580619215965, + 0.0038934233598411083, + 0.003628681879490614, + 0.003510893788188696, + 0.0064254580065608025, + 0.007484906818717718, + 0.00825522281229496, + 0.014886942692101002, + 0.004104590974748135, + 0.005274353548884392, + 0.004202743526548147, + 0.0020889544393867254, + 0.004134169779717922, + 0.0036886853631585836, + 0.006082952953875065, + 0.00562902819365263, + 0.004429651889950037, + 0.004734894260764122, + 0.004007186274975538, + 0.006787100341171026, + 0.007168238516896963, + 0.009395494125783443, + 0.007834222167730331, + 0.011017206124961376, + 0.003766375593841076, + 0.00552851939573884, + 0.004081845283508301, + 0.004097435157746077, + 0.00498016644269228, + 0.007227796595543623, + 0.0037096538580954075, + 0.010081022046506405, + 0.006544033531099558, + 0.008941944688558578, + 0.0065811630338430405, + 0.01605290174484253, + 0.008417034521698952, + 0.008222382515668869, + 0.00605496671050787, + 0.006952118128538132, + 0.008016353473067284, + 0.018557701259851456, + 0.02241421677172184, + 0.012128392234444618, + 0.010231105610728264, + 0.03820384666323662, + 0.0066207824274897575, + 0.016315322369337082, + 0.01815141923725605, + 0.010256486013531685, + 0.15245318412780762, + 0.013559352606534958, + 0.010403321124613285, + 0.16447050869464874, + 0.005886910483241081, + 0.01694641262292862, + 0.00722090108320117, + 0.007773815654218197, + 0.021646475419402122, + 0.005366216879338026, + 0.006133702117949724, + 0.01957559585571289, + 0.008806327357888222, + 0.005031558219343424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.014188366010785103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_2/index_tensor([7917, 7917])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_2/index_tensor([7917, 7917])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cf485cf6b2a7022682ca5e46ff3677339c218673 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_2/index_tensor([7917, 7917])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up shot from a low angle, slowly and relentlessly zooming in, intensifies the focus on a grieving young woman. Her brown hair is pulled back in a simple ponytail, with loose strands catching the sunlight like thin golden threads. She wears thin wire-rimmed glasses and a crisp light blue cotton dress shirt. The entire frame is dominated by the raw physicality of a violent cough that convulses her body; her shoulders arch violently, and her head jerk forward with each spasmodic, uncontrollable thump. Her eyes, already watery with underlying sorrow, squeeze tightly in agony, and her face flushes. The scene is bathed in a warm, ethereal golden hour light, creating a soft halo around her body against the blurred backdrop of an outdoor night. The stark contrast between the serene, beautiful light and her inner anguish creates an intense, poignant, and intimate atmosphere of vulnerability and pain.", + "seed": 11920, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_2/index_7917_seed_11920_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.37890769839286803, + "vlm_logprob": -0.37890769839286803, + "vlm_yes_no_margin": 1.375, + "vlm_token_credit": -0.37890769839286803, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5507336854934692 + ], + "opposite_logprob_tokens": [ + -0.9257336854934692 + ], + "logprob_reward": -0.5507336854934692, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5507336854934692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36073023080825806, + 0.09010906517505646, + 0.03867809846997261, + 0.09749716520309448, + 0.07334250956773758, + 0.04086357355117798, + 0.023472441360354424, + 0.0887831598520279, + 0.08531849086284637, + 0.1012052670121193 + ], + "credit_map": [ + 0.08149804919958115, + 0.04485829547047615, + 0.072410449385643, + 0.053067076951265335, + 0.02667447365820408, + 0.02205825038254261, + 0.038246385753154755, + 0.021917277947068214, + 0.010946989990770817, + 0.012324719689786434, + 0.01186718512326479, + 0.0123943742364645, + 0.01177581399679184, + 0.007119990885257721, + 0.0164282638579607, + 0.007251729257404804, + 0.005555179435759783, + 0.0033626717049628496, + 0.0064135268330574036, + 0.00453728623688221, + 0.006741444114595652, + 0.0034003371838480234, + 0.005042009521275759, + 0.0036256469320505857, + 0.014263169839978218, + 0.01674741879105568, + 0.014028494246304035, + 0.007318410556763411, + 0.014967252500355244, + 0.009756962768733501, + 0.011183339171111584, + 0.009232121519744396, + 0.01736713945865631, + 0.00684552313759923, + 0.019545748829841614, + 0.0044763535261154175, + 0.008210329338908195, + 0.006551417056471109, + 0.004664160311222076, + 0.005681853275746107, + 0.0050405533984303474, + 0.00418497109785676, + 0.006525283679366112, + 0.004784861579537392, + 0.007202130276709795, + 0.0033751095179468393, + 0.0060054040513932705, + 0.0037452622782438993, + 0.003314934903755784, + 0.0037217512726783752, + 0.003638157621026039, + 0.0021612823475152254, + 0.003744814544916153, + 0.0016579278744757175, + 0.002896498888731003, + 0.0023370743729174137, + 0.013052498921751976, + 0.01285757776349783, + 0.008132532238960266, + 0.010798436589539051, + 0.008761988952755928, + 0.004119775723665953, + 0.02158539555966854, + 0.009474962949752808, + 0.011304061859846115, + 0.016888724640011787, + 0.011263636872172356, + 0.012933612801134586, + 0.009754485450685024, + 0.011071868240833282, + 0.006930329836905003, + 0.005171774886548519, + 0.014190558344125748, + 0.014146438799798489, + 0.014987154863774776, + 0.02396218664944172, + 0.012607182376086712, + 0.0057213488034904, + 0.01012398861348629, + 0.00546641880646348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.5507336854934692 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934441924095154 + ], + "opposite_logprob_tokens": [ + -0.6934441924095154 + ], + "logprob_reward": -0.6934441924095154, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934441924095154, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.31705185770988464, + 0.11014123260974884, + 0.027468537911772728, + 0.07285143435001373, + 0.0999135673046112, + 0.03626317158341408, + 0.0796368271112442, + 0.03979567065834999, + 0.07728459686040878, + 0.13959307968616486 + ], + "credit_map": [ + 0.06764205545186996, + 0.03814336657524109, + 0.05574074387550354, + 0.044459279626607895, + 0.032088134437799454, + 0.02198057994246483, + 0.03588742017745972, + 0.02111029252409935, + 0.013593695126473904, + 0.013437922112643719, + 0.014484476298093796, + 0.016295725479722023, + 0.015261285938322544, + 0.008573847822844982, + 0.017228055745363235, + 0.011266223154962063, + 0.0030624319333583117, + 0.0038239690475165844, + 0.0031442579347640276, + 0.003181004896759987, + 0.00622968003153801, + 0.0017015249468386173, + 0.0032707680948078632, + 0.0030549007933586836, + 0.011535423807799816, + 0.007461278699338436, + 0.014261789619922638, + 0.006870432291179895, + 0.013333279639482498, + 0.00810855533927679, + 0.006553414277732372, + 0.004727266263216734, + 0.030861716717481613, + 0.005482176318764687, + 0.033449918031692505, + 0.004565923009067774, + 0.00441662548109889, + 0.007696507964283228, + 0.008476659655570984, + 0.004964033607393503, + 0.004839085042476654, + 0.0037786851171404123, + 0.004948859568685293, + 0.003602329408749938, + 0.008008305914700031, + 0.002411034656688571, + 0.006165174301713705, + 0.002509698737412691, + 0.015727754682302475, + 0.00716077396646142, + 0.0212142001837492, + 0.007260704413056374, + 0.00927375815808773, + 0.005869945045560598, + 0.0068627409636974335, + 0.0062669506296515465, + 0.006749384105205536, + 0.007390422280877829, + 0.005143705289810896, + 0.003611142747104168, + 0.004477288573980331, + 0.0026086936704814434, + 0.006177862174808979, + 0.0036371699534356594, + 0.00993622001260519, + 0.010460876859724522, + 0.012037742882966995, + 0.007635562680661678, + 0.012631242163479328, + 0.016585610806941986, + 0.0038447685074061155, + 0.0041525703854858875, + 0.022858263924717903, + 0.016322415322065353, + 0.0241036769002676, + 0.03952468931674957, + 0.0127434516325593, + 0.006323229521512985, + 0.010687853209674358, + 0.007029503583908081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.6934441924095154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084526777267456 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084526777267456, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084526777267456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3408067524433136, + 0.16191227734088898, + 0.026259692385792732, + 0.16782251000404358, + 0.04826335608959198, + 0.018506471067667007, + 0.0831596776843071, + 0.031148508191108704, + 0.06065497547388077, + 0.06146576255559921 + ], + "credit_map": [ + 0.06475505977869034, + 0.04563082009553909, + 0.05136289447546005, + 0.049348827451467514, + 0.04045336693525314, + 0.01958627440035343, + 0.048088230192661285, + 0.021581267938017845, + 0.01912178471684456, + 0.013393163681030273, + 0.02032742276787758, + 0.01761651784181595, + 0.023370111361145973, + 0.014990512281656265, + 0.03854034096002579, + 0.014552432112395763, + 0.003038001013919711, + 0.0035443624947220087, + 0.0033100189175456762, + 0.0033015580847859383, + 0.005573211703449488, + 0.0017335133161395788, + 0.0035919572692364454, + 0.0021670707501471043, + 0.02653510682284832, + 0.021820953115820885, + 0.027732035145163536, + 0.011006315238773823, + 0.036085136234760284, + 0.015970805659890175, + 0.019061140716075897, + 0.009611031971871853, + 0.009623253718018532, + 0.004043992608785629, + 0.009491515345871449, + 0.002390296431258321, + 0.006368901580572128, + 0.006092874798923731, + 0.005396671127527952, + 0.004855851177126169, + 0.0020610049832612276, + 0.00243024667724967, + 0.002519316039979458, + 0.0023217948619276285, + 0.0033267659600824118, + 0.0016165346605703235, + 0.002957745222374797, + 0.0012730640592053533, + 0.0209925826638937, + 0.007830866612493992, + 0.02523694559931755, + 0.007306399755179882, + 0.007855314761400223, + 0.003645149525254965, + 0.005928660277277231, + 0.004363756626844406, + 0.005227859131991863, + 0.007813000120222569, + 0.00474836491048336, + 0.002145771635696292, + 0.004257949069142342, + 0.002261384157463908, + 0.0024925130419433117, + 0.0022016658913344145, + 0.007464534603059292, + 0.008857261389493942, + 0.007570597808808088, + 0.010950607247650623, + 0.008403727784752846, + 0.005017007701098919, + 0.008023370988667011, + 0.004367862828075886, + 0.010389812290668488, + 0.006901192478835583, + 0.009846826083958149, + 0.01707877218723297, + 0.006687283981591463, + 0.0025797460693866014, + 0.0050562908872962, + 0.002925839740782976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.011084526777267456 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871159851551056 + ], + "opposite_logprob_tokens": [ + -1.1371159553527832 + ], + "logprob_reward": -0.3871159851551056, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871159851551056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2698836326599121, + 0.1868724375963211, + 0.028415609151124954, + 0.054954010993242264, + 0.058251503854990005, + 0.043652474880218506, + 0.22925008833408356, + 0.02364572323858738, + 0.07172858715057373, + 0.03334598243236542 + ], + "credit_map": [ + 0.05228754132986069, + 0.041010063141584396, + 0.044028390198946, + 0.04102570563554764, + 0.02384275756776333, + 0.020063837990164757, + 0.03047540783882141, + 0.01714988425374031, + 0.021589362993836403, + 0.01915811374783516, + 0.026341067627072334, + 0.022020364180207253, + 0.031539976596832275, + 0.012969381175935268, + 0.03274710476398468, + 0.02050706185400486, + 0.003401167457923293, + 0.002679607830941677, + 0.0038216067478060722, + 0.004464094061404467, + 0.005563464015722275, + 0.002432099776342511, + 0.0034024964552372694, + 0.002651070710271597, + 0.011562257073819637, + 0.0051461998373270035, + 0.013016332872211933, + 0.003558149328455329, + 0.009210392832756042, + 0.004767597187310457, + 0.004526298493146896, + 0.003166778478771448, + 0.013861502520740032, + 0.004459195304661989, + 0.01439233124256134, + 0.0037398941349238157, + 0.004855279810726643, + 0.005185600835829973, + 0.0051080621778965, + 0.006649632006883621, + 0.005291462875902653, + 0.004615736193954945, + 0.004807521589100361, + 0.0045557776466012, + 0.009110222570598125, + 0.002867342671379447, + 0.008779603987932205, + 0.0036248061805963516, + 0.04256394878029823, + 0.026039281859993935, + 0.05248399078845978, + 0.022999050095677376, + 0.029981553554534912, + 0.014898608438670635, + 0.021914929151535034, + 0.018368704244494438, + 0.0028591116424649954, + 0.0027689540293067694, + 0.002308045979589224, + 0.002448513638228178, + 0.0028862471226602793, + 0.0017657701391726732, + 0.00423876428976655, + 0.004370314069092274, + 0.011178107932209969, + 0.010329722426831722, + 0.010341784916818142, + 0.0054578338749706745, + 0.00811034720391035, + 0.018210887908935547, + 0.00363387749530375, + 0.004466019570827484, + 0.0043384539894759655, + 0.0042145890183746815, + 0.004824600648134947, + 0.004410307854413986, + 0.004914646968245506, + 0.002862878842279315, + 0.004929184913635254, + 0.002851315541192889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.3871159851551056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25216010212898254 + ], + "opposite_logprob_tokens": [ + -1.5021600723266602 + ], + "logprob_reward": -0.25216010212898254, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25216010212898254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24840761721134186, + 0.1227845698595047, + 0.047842755913734436, + 0.19682855904102325, + 0.060761380940675735, + 0.11422748118638992, + 0.07642584294080734, + 0.015256819315254688, + 0.029242467135190964, + 0.08822257071733475 + ], + "credit_map": [ + 0.05391492694616318, + 0.034760307520627975, + 0.04214814305305481, + 0.03381314501166344, + 0.02333100140094757, + 0.015278028324246407, + 0.028920026496052742, + 0.016242019832134247, + 0.012404422275722027, + 0.018196646124124527, + 0.01696433313190937, + 0.02924918197095394, + 0.017552556470036507, + 0.0069320169277489185, + 0.013954630121588707, + 0.007530759554356337, + 0.006884693633764982, + 0.0065815686248242855, + 0.0076001607812941074, + 0.008005666546523571, + 0.007413438055664301, + 0.0028062344063073397, + 0.004973313305526972, + 0.0035776752047240734, + 0.04056510329246521, + 0.017862174659967422, + 0.04591478407382965, + 0.01516321487724781, + 0.028046471998095512, + 0.01419675350189209, + 0.0181666761636734, + 0.016913359984755516, + 0.019731516018509865, + 0.003987282980233431, + 0.020885659381747246, + 0.0023659723810851574, + 0.004075617995113134, + 0.0034285392612218857, + 0.003544931299984455, + 0.0027418523095548153, + 0.014310864731669426, + 0.0160480085760355, + 0.011582323350012302, + 0.013062315061688423, + 0.021925002336502075, + 0.010262064635753632, + 0.018894169479608536, + 0.008142726495862007, + 0.020029662176966667, + 0.0062357233837246895, + 0.023351117968559265, + 0.006076028570532799, + 0.007296394556760788, + 0.003842428093776107, + 0.004926526453346014, + 0.004667951259762049, + 0.0022719765547662973, + 0.001997558865696192, + 0.001806261483579874, + 0.0014527934836223722, + 0.001777254045009613, + 0.0011615168768912554, + 0.0027614461723715067, + 0.0020280100870877504, + 0.004026110749691725, + 0.0051726484671235085, + 0.0034710539039224386, + 0.0035100304521620274, + 0.003779452061280608, + 0.005214782431721687, + 0.002104050014168024, + 0.0019643381237983704, + 0.01272218581289053, + 0.012990409508347511, + 0.014232832007110119, + 0.021363848820328712, + 0.009928376413881779, + 0.0046902005560696125, + 0.008207334205508232, + 0.004087377805262804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.25216010212898254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a low angle, slowly and relentlessly zooming in, intensifies the focus on a grieving young woman. Her brown hair is pulled back in a simple ponytail, with loose strands catching the sunlight like thin golden threads. She wears thin wire-rimmed glasses and a crisp light blue cotton dress shirt. The entire frame is dominated by the raw physicality of a violent cough that convulses her body; her shoulders arch violently, and her head jerk forward with each spasmodic, uncontrollable thump. Her eyes, already watery with underlying sorrow, squeeze tightly in agony, and her face flushes. The scene is bathed in a warm, ethereal golden hour light, creating a soft halo around her body against the blurred backdrop of an outdoor night. The stark contrast between the serene, beautiful light and her inner anguish creates an intense, poignant, and intimate atmosphere of vulnerability and pain.", + "seed": 11921, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_2/index_7917_seed_11921_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3514769710600376, + "vlm_logprob": -0.3514769710600376, + "vlm_yes_no_margin": 1.25, + "vlm_token_credit": -0.3514769710600376, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8494107127189636 + ], + "opposite_logprob_tokens": [ + -0.5994107127189636 + ], + "logprob_reward": -0.8494107127189636, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8494107127189636, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11628163605928421, + 0.06895028799772263, + 0.09519889205694199, + 0.0651998296380043, + 0.15929757058620453, + 0.05824928358197212, + 0.13641682267189026, + 0.04300404712557793, + 0.046438414603471756, + 0.21096324920654297 + ], + "credit_map": [ + 0.01174546405673027, + 0.012514729984104633, + 0.008657842874526978, + 0.023932548239827156, + 0.012354179285466671, + 0.019695037975907326, + 0.017127320170402527, + 0.01025451347231865, + 0.010747087188065052, + 0.005016156006604433, + 0.010800596326589584, + 0.005608255509287119, + 0.010547088459134102, + 0.009243686683475971, + 0.009643512777984142, + 0.007343906909227371, + 0.00954427756369114, + 0.009016314521431923, + 0.012112248688936234, + 0.00549220060929656, + 0.022536106407642365, + 0.009997192770242691, + 0.016520541161298752, + 0.009980014525353909, + 0.007497433107346296, + 0.009591019712388515, + 0.009874961338937283, + 0.006799311842769384, + 0.009548898786306381, + 0.0068560633808374405, + 0.00797894038259983, + 0.007053199224174023, + 0.04661177098751068, + 0.009205590933561325, + 0.06063804775476456, + 0.0065175509080290794, + 0.014088091440498829, + 0.008666623383760452, + 0.007688118610531092, + 0.005881774704903364, + 0.009024320170283318, + 0.004183557350188494, + 0.015092004090547562, + 0.0029191470239311457, + 0.009509607218205929, + 0.007435831241309643, + 0.006494451314210892, + 0.0035903621464967728, + 0.014772706665098667, + 0.023101774975657463, + 0.013570334762334824, + 0.01822955347597599, + 0.022099308669567108, + 0.019059451296925545, + 0.016056450083851814, + 0.009527239017188549, + 0.004641694016754627, + 0.006360449828207493, + 0.0046711210161447525, + 0.005538803059607744, + 0.0066403578966856, + 0.0052533517591655254, + 0.004378196783363819, + 0.005520072299987078, + 0.003086245618760586, + 0.008105878718197346, + 0.007858570665121078, + 0.00935366190969944, + 0.005229215603321791, + 0.0039029482286423445, + 0.005413023754954338, + 0.0034888722002506256, + 0.01157981064170599, + 0.011257522739470005, + 0.04801508039236069, + 0.00964347179979086, + 0.01937793754041195, + 0.015431501902639866, + 0.0855090469121933, + 0.010148866102099419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.8494107127189636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25217798352241516 + ], + "opposite_logprob_tokens": [ + -1.5021779537200928 + ], + "logprob_reward": -0.25217798352241516, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25217798352241516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05608953908085823, + 0.05164141207933426, + 0.02730947732925415, + 0.056737665086984634, + 0.06445341557264328, + 0.020944593474268913, + 0.041105080395936966, + 0.06566207855939865, + 0.08200230449438095, + 0.5340545177459717 + ], + "credit_map": [ + 0.005996349733322859, + 0.006222187075763941, + 0.005111653823405504, + 0.012146798893809319, + 0.004387513734400272, + 0.009367711842060089, + 0.007881054654717445, + 0.004976268857717514, + 0.008699961937963963, + 0.004500655457377434, + 0.0093470374122262, + 0.004056880716234446, + 0.008650081232190132, + 0.005314032547175884, + 0.006844475399702787, + 0.004228290170431137, + 0.004399447236210108, + 0.0022962887305766344, + 0.0043045938946306705, + 0.0013000776525586843, + 0.005036068614572287, + 0.0025097611360251904, + 0.005247201770544052, + 0.0022160394582897425, + 0.006180901080369949, + 0.005915379151701927, + 0.008219131268560886, + 0.0034114064183086157, + 0.00866687297821045, + 0.009005113504827023, + 0.00898830033838749, + 0.006350564770400524, + 0.019056707620620728, + 0.003576067276299, + 0.024025557562708855, + 0.0025230247993022203, + 0.005721071735024452, + 0.0031999675557017326, + 0.003389027202501893, + 0.0029619901906698942, + 0.003097607521340251, + 0.002507475670427084, + 0.00360024138353765, + 0.001404708018526435, + 0.004670822527259588, + 0.0020457960199564695, + 0.002098037861287594, + 0.0015199047047644854, + 0.0059121884405612946, + 0.007116770837455988, + 0.004345656838268042, + 0.006093594711273909, + 0.004777943249791861, + 0.005527885165065527, + 0.003979880828410387, + 0.0033511617220938206, + 0.009065150283277035, + 0.010793226771056652, + 0.006462071090936661, + 0.005833026021718979, + 0.009713683277368546, + 0.008858339861035347, + 0.00693998159840703, + 0.007996591739356518, + 0.003915244247764349, + 0.01725311577320099, + 0.011007444933056831, + 0.014458159916102886, + 0.01438947208225727, + 0.007197280880063772, + 0.007855736650526524, + 0.005925850942730904, + 0.028217628598213196, + 0.02796325646340847, + 0.11533938348293304, + 0.022999072447419167, + 0.05078711360692978, + 0.042967069894075394, + 0.21943016350269318, + 0.026350807398557663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.25217798352241516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307105019688606 + ], + "opposite_logprob_tokens": [ + -3.1680710315704346 + ], + "logprob_reward": -0.04307105019688606, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307105019688606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3097062110900879, + 0.05708123371005058, + 0.2013048231601715, + 0.01968478225171566, + 0.02314833551645279, + 0.028469430282711983, + 0.03376302495598793, + 0.0989033430814743, + 0.04392228648066521, + 0.18401655554771423 + ], + "credit_map": [ + 0.03248216211795807, + 0.04371828958392143, + 0.050199177116155624, + 0.06228041648864746, + 0.025637730956077576, + 0.029397090896964073, + 0.043384820222854614, + 0.022606506943702698, + 0.007602587342262268, + 0.003784733824431896, + 0.008107265457510948, + 0.0038643185980618, + 0.012863749638199806, + 0.005512802395969629, + 0.010154570452868938, + 0.005191202275454998, + 0.01774272695183754, + 0.023609815165400505, + 0.024356257170438766, + 0.010425405576825142, + 0.04850649833679199, + 0.02494899369776249, + 0.03109736554324627, + 0.020617753267288208, + 0.0022774485405534506, + 0.0025588232092559338, + 0.00246042781509459, + 0.0017965020379051566, + 0.002583838300779462, + 0.0026935371570289135, + 0.0027818758971989155, + 0.0025323289446532726, + 0.002847804222255945, + 0.002176918089389801, + 0.0029328966047614813, + 0.0012995762517675757, + 0.006576621904969215, + 0.0029681448359042406, + 0.003041112096980214, + 0.0013052606955170631, + 0.004530610051006079, + 0.003292804816737771, + 0.004513172432780266, + 0.0018724495312198997, + 0.004813231527805328, + 0.0029597189277410507, + 0.003322450676932931, + 0.0031649910379201174, + 0.003380900714546442, + 0.005578148178756237, + 0.002923553576692939, + 0.004613854922354221, + 0.004430445376783609, + 0.0044388240203261375, + 0.005272583104670048, + 0.003124713432043791, + 0.006659398321062326, + 0.012664960697293282, + 0.006901117041707039, + 0.015782799571752548, + 0.020516101270914078, + 0.014778507873415947, + 0.012396086007356644, + 0.009204359725117683, + 0.0023592952638864517, + 0.009210131131112576, + 0.004586277063935995, + 0.008620599284768105, + 0.005928437691181898, + 0.003839390818029642, + 0.005052411463111639, + 0.004325738176703453, + 0.009824576787650585, + 0.01039285771548748, + 0.03778758645057678, + 0.008647225797176361, + 0.018855063244700432, + 0.013874591328203678, + 0.07487957924604416, + 0.009755074977874756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.04307105019688606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561690211296082 + ], + "opposite_logprob_tokens": [ + -1.6006169319152832 + ], + "logprob_reward": -0.22561690211296082, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561690211296082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04544169083237648, + 0.07351048290729523, + 0.03376815468072891, + 0.022198794409632683, + 0.1291278749704361, + 0.03129494562745094, + 0.028876284137368202, + 0.1520562767982483, + 0.05302530899643898, + 0.4307001233100891 + ], + "credit_map": [ + 0.003807496279478073, + 0.005645452998578548, + 0.006248991005122662, + 0.0070895180106163025, + 0.006029436830431223, + 0.005747292190790176, + 0.007859698496758938, + 0.003013808513060212, + 0.00836768839508295, + 0.005843000486493111, + 0.009785294532775879, + 0.005090261809527874, + 0.015432404354214668, + 0.00797614362090826, + 0.012938318774104118, + 0.008077369071543217, + 0.0034356238320469856, + 0.0034057721495628357, + 0.003964228089898825, + 0.001621980918571353, + 0.008616823703050613, + 0.003952027764171362, + 0.006138389930129051, + 0.002633308991789818, + 0.00190915132407099, + 0.0028559344355016947, + 0.0031189152505248785, + 0.002118145814165473, + 0.0031143457163125277, + 0.0026031378656625748, + 0.005005569662898779, + 0.0014735936420038342, + 0.04167339578270912, + 0.006267160177230835, + 0.05297250300645828, + 0.0055706435814499855, + 0.010090349242091179, + 0.00513122696429491, + 0.0037144862581044436, + 0.003708100877702236, + 0.005744889844208956, + 0.004822113085538149, + 0.006435041315853596, + 0.00238368334248662, + 0.0032453001476824284, + 0.0033414612989872694, + 0.003256135154515505, + 0.00206631887704134, + 0.0023979840334504843, + 0.005590968765318394, + 0.0025348742492496967, + 0.003111462341621518, + 0.003895280184224248, + 0.003554840572178364, + 0.004502322990447283, + 0.003288551000878215, + 0.009821205399930477, + 0.022014619782567024, + 0.011825484223663807, + 0.032773036509752274, + 0.02460644207894802, + 0.020746227353811264, + 0.01857871189713478, + 0.011690548621118069, + 0.0032946288120001554, + 0.010241442359983921, + 0.010917240753769875, + 0.005304290447384119, + 0.007122008595615625, + 0.005380192771553993, + 0.0056412965059280396, + 0.0051242089830338955, + 0.021126987412571907, + 0.020898090675473213, + 0.09137457609176636, + 0.016567537561058998, + 0.03879261389374733, + 0.03307465463876724, + 0.18503712117671967, + 0.023828547447919846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.22561690211296082 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871082067489624 + ], + "opposite_logprob_tokens": [ + -1.1371082067489624 + ], + "logprob_reward": -0.3871082067489624, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871082067489624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06598905473947525, + 0.058666735887527466, + 0.08273431658744812, + 0.018619155511260033, + 0.24631240963935852, + 0.23253263533115387, + 0.021689346060156822, + 0.042180079966783524, + 0.05664832890033722, + 0.17462801933288574 + ], + "credit_map": [ + 0.007224626839160919, + 0.007222265470772982, + 0.010470299050211906, + 0.01135309413075447, + 0.0060679311864078045, + 0.007289509754627943, + 0.010113777592778206, + 0.006247540004551411, + 0.00942471344023943, + 0.0057852971367537975, + 0.010142196901142597, + 0.004401774611324072, + 0.008857820183038712, + 0.005814073141664267, + 0.008094795048236847, + 0.006146058440208435, + 0.008063461631536484, + 0.007913004606962204, + 0.012181872501969337, + 0.004029008559882641, + 0.020659761503338814, + 0.009694586507976055, + 0.013112873770296574, + 0.007079734466969967, + 0.0018477138364687562, + 0.0026607660111039877, + 0.0022610549349337816, + 0.0016807406209409237, + 0.0032124125864356756, + 0.0022560451179742813, + 0.0022600144147872925, + 0.002440404612571001, + 0.08043234795331955, + 0.011234140954911709, + 0.1019386574625969, + 0.011320670135319233, + 0.01891946792602539, + 0.008940945379436016, + 0.00666934996843338, + 0.006856820546090603, + 0.04243416339159012, + 0.05224725604057312, + 0.040799640119075775, + 0.02011796459555626, + 0.026187699288129807, + 0.027929134666919708, + 0.015797024592757225, + 0.007019727025181055, + 0.002471985062584281, + 0.0034776083193719387, + 0.0017929596360772848, + 0.0027560389135032892, + 0.0035656762775033712, + 0.003063189098611474, + 0.0028733601793646812, + 0.0016885255463421345, + 0.002788242418318987, + 0.005208759568631649, + 0.0036875249352306128, + 0.008886907249689102, + 0.006549375131726265, + 0.006240881979465485, + 0.004990422632545233, + 0.0038279618602246046, + 0.0031202416867017746, + 0.010204393416643143, + 0.007837964221835136, + 0.010604207403957844, + 0.008006767369806767, + 0.005731335375458002, + 0.0065752663649618626, + 0.004568145144730806, + 0.00899877492338419, + 0.008173429407179356, + 0.0377400778234005, + 0.006747181992977858, + 0.015502803027629852, + 0.013776049017906189, + 0.07414206117391586, + 0.009547610767185688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.3871082067489624 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a low angle, slowly and relentlessly zooming in, intensifies the focus on a grieving young woman. Her brown hair is pulled back in a simple ponytail, with loose strands catching the sunlight like thin golden threads. She wears thin wire-rimmed glasses and a crisp light blue cotton dress shirt. The entire frame is dominated by the raw physicality of a violent cough that convulses her body; her shoulders arch violently, and her head jerk forward with each spasmodic, uncontrollable thump. Her eyes, already watery with underlying sorrow, squeeze tightly in agony, and her face flushes. The scene is bathed in a warm, ethereal golden hour light, creating a soft halo around her body against the blurred backdrop of an outdoor night. The stark contrast between the serene, beautiful light and her inner anguish creates an intense, poignant, and intimate atmosphere of vulnerability and pain.", + "seed": 11922, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_2/index_7917_seed_11922_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3144844524562359, + "vlm_logprob": -0.3144844524562359, + "vlm_yes_no_margin": 1.675, + "vlm_token_credit": -0.3144844524562359, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11161582171916962 + ], + "opposite_logprob_tokens": [ + -2.4866158962249756 + ], + "logprob_reward": -0.11161582171916962, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11161582171916962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18173043429851532, + 0.08279767632484436, + 0.12539829313755035, + 0.06843655556440353, + 0.02891061082482338, + 0.0955217033624649, + 0.23861320316791534, + 0.06740713119506836, + 0.0915289968252182, + 0.0196554996073246 + ], + "credit_map": [ + 0.036146100610494614, + 0.009861089289188385, + 0.018037008121609688, + 0.0181971974670887, + 0.023807331919670105, + 0.005726668518036604, + 0.06041131168603897, + 0.00954373273998499, + 0.014146440662443638, + 0.004708273336291313, + 0.002692796755582094, + 0.003715276950970292, + 0.02949516475200653, + 0.003522006794810295, + 0.020385535433888435, + 0.004132185131311417, + 0.01152652595192194, + 0.031758811324834824, + 0.0113676143810153, + 0.012817133218050003, + 0.020935259759426117, + 0.004989096894860268, + 0.013965541496872902, + 0.018038298934698105, + 0.010446046479046345, + 0.019720161333680153, + 0.004746465478092432, + 0.004151508677750826, + 0.007985093630850315, + 0.0038268251810222864, + 0.011503119952976704, + 0.006057330872863531, + 0.004610034637153149, + 0.002284369431436062, + 0.0020721296314150095, + 0.002926130313426256, + 0.005835620686411858, + 0.001796505181118846, + 0.006677634082734585, + 0.0027081849984824657, + 0.01835128664970398, + 0.009243815205991268, + 0.012882583774626255, + 0.012234645895659924, + 0.013392618857324123, + 0.005816069897264242, + 0.016445700079202652, + 0.00715498486533761, + 0.05963008478283882, + 0.022633235901594162, + 0.007700593676418066, + 0.006042635068297386, + 0.10034477710723877, + 0.010592041537165642, + 0.02549339272081852, + 0.006176426075398922, + 0.011572828516364098, + 0.009656837210059166, + 0.0025480613112449646, + 0.003991732839494944, + 0.023234695196151733, + 0.004302297718822956, + 0.008385994471609592, + 0.003714683698490262, + 0.012676948681473732, + 0.008846012875437737, + 0.00747248250991106, + 0.010365039110183716, + 0.012582188472151756, + 0.009157920256257057, + 0.019219737499952316, + 0.011208668351173401, + 0.0025819845031946898, + 0.0022218129597604275, + 0.0015419431729242206, + 0.0024988385848701, + 0.0030307462438941, + 0.0012825033627450466, + 0.005025799851864576, + 0.0014718687161803246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.11161582171916962 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578915357589722 + ], + "opposite_logprob_tokens": [ + -0.6328915357589722 + ], + "logprob_reward": -0.7578915357589722, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578915357589722, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.056027982383966446, + 0.2719666063785553, + 0.15036776661872864, + 0.044444337487220764, + 0.054130371659994125, + 0.03532927855849266, + 0.1345958560705185, + 0.038242027163505554, + 0.17040593922138214, + 0.04448981583118439 + ], + "credit_map": [ + 0.008042100816965103, + 0.005264222621917725, + 0.006416305433958769, + 0.006596002262085676, + 0.008355307392776012, + 0.0038586310110986233, + 0.011996941640973091, + 0.005498469341546297, + 0.03472217172384262, + 0.01835222914814949, + 0.016170518472790718, + 0.01613360084593296, + 0.0829373300075531, + 0.01600627228617668, + 0.07030629366636276, + 0.01733820140361786, + 0.015389736741781235, + 0.029148535802960396, + 0.014861955307424068, + 0.014577213674783707, + 0.03045305609703064, + 0.007921733893454075, + 0.022186513990163803, + 0.015829017385840416, + 0.007189645431935787, + 0.009577391669154167, + 0.0038777657318860292, + 0.0038658292032778263, + 0.004405646119266748, + 0.002958455588668585, + 0.006864740513265133, + 0.005704862996935844, + 0.009077704511582851, + 0.004320093896239996, + 0.003528428031131625, + 0.003874461865052581, + 0.010996137745678425, + 0.0029058046638965607, + 0.014423445798456669, + 0.0050042979419231415, + 0.005294645670801401, + 0.0027798113878816366, + 0.00455064931884408, + 0.005141484551131725, + 0.00494352774694562, + 0.00238447985611856, + 0.00791100412607193, + 0.002323675900697708, + 0.026315605267882347, + 0.013880511745810509, + 0.006734018679708242, + 0.008113812655210495, + 0.04646661505103111, + 0.009106975980103016, + 0.019154418259859085, + 0.0048239100724458694, + 0.005184041801840067, + 0.0053916503675282, + 0.0017383098602294922, + 0.0026096347719430923, + 0.009909243322908878, + 0.005218390841037035, + 0.006043166853487492, + 0.0021475893445312977, + 0.024006854742765427, + 0.011921195313334465, + 0.013276426121592522, + 0.01942429132759571, + 0.027127742767333984, + 0.017899900674819946, + 0.03352927044034004, + 0.023220257833600044, + 0.004547117743641138, + 0.005585798993706703, + 0.0045076655223965645, + 0.006404118146747351, + 0.0055439588613808155, + 0.004815834108740091, + 0.00921931304037571, + 0.0038660059217363596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.7578915357589722 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635536715388298 + ], + "opposite_logprob_tokens": [ + -3.651355266571045 + ], + "logprob_reward": -0.02635536715388298, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635536715388298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07873210310935974, + 0.03453143686056137, + 0.05592038854956627, + 0.28156161308288574, + 0.02179071120917797, + 0.03060527704656124, + 0.45307594537734985, + 0.011516564525663853, + 0.020376157015562057, + 0.011889771558344364 + ], + "credit_map": [ + 0.015768589451909065, + 0.005880305543541908, + 0.006652230862528086, + 0.0056741563603281975, + 0.011258106678724289, + 0.004187548998743296, + 0.025144590064883232, + 0.004166578873991966, + 0.007862342521548271, + 0.00196253415197134, + 0.0018269841093569994, + 0.0015396903036162257, + 0.012528577819466591, + 0.001387922908179462, + 0.005584488157182932, + 0.0018388972384855151, + 0.006096378434449434, + 0.006884317379444838, + 0.005292107816785574, + 0.0053868843242526054, + 0.015508366748690605, + 0.0029146471060812473, + 0.008230503648519516, + 0.005607187282294035, + 0.04503631219267845, + 0.03390050679445267, + 0.013287417590618134, + 0.01562603749334812, + 0.02214907854795456, + 0.007740967907011509, + 0.12079185992479324, + 0.02302946150302887, + 0.0026848004199564457, + 0.0018883132142946124, + 0.0032141397241503, + 0.002800741931423545, + 0.0032654879614710808, + 0.0013079134514555335, + 0.004083315376192331, + 0.0025460002943873405, + 0.004271844867616892, + 0.0023643928579986095, + 0.0029560804832726717, + 0.003757127560675144, + 0.004618904087692499, + 0.0020770186092704535, + 0.007684777956455946, + 0.0028751313220709562, + 0.11865414679050446, + 0.04069779813289642, + 0.014006457291543484, + 0.009402602910995483, + 0.19477874040603638, + 0.019238736480474472, + 0.04485311359167099, + 0.01144437212496996, + 0.0015769494930282235, + 0.0012851523933932185, + 0.0009896684205159545, + 0.001260775257833302, + 0.0033476140815764666, + 0.0008250108221545815, + 0.0016399658052250743, + 0.0005914286011829972, + 0.00303631485439837, + 0.001095116720534861, + 0.0016382343601435423, + 0.0019681532867252827, + 0.0026808937545865774, + 0.0016394935082644224, + 0.005279023200273514, + 0.0030389265157282352, + 0.0020810524001717567, + 0.000727845064830035, + 0.0007022732752375305, + 0.0008142438600771129, + 0.002195116598159075, + 0.0005465895519591868, + 0.004019538406282663, + 0.000803113158326596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.02635536715388298 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035011172294617 + ], + "opposite_logprob_tokens": [ + -2.3503501415252686 + ], + "logprob_reward": -0.10035011172294617, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035011172294617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10894668102264404, + 0.12394814193248749, + 0.04717133939266205, + 0.19041970372200012, + 0.03782056272029877, + 0.06505007296800613, + 0.2694445252418518, + 0.015091832727193832, + 0.041240859776735306, + 0.10086628794670105 + ], + "credit_map": [ + 0.014777909964323044, + 0.0049712457694113255, + 0.010879370383918285, + 0.012558956630527973, + 0.013844303786754608, + 0.007091714534908533, + 0.03346578776836395, + 0.011357386596500874, + 0.022059282287955284, + 0.005889971740543842, + 0.004599196836352348, + 0.005901268683373928, + 0.04512997716665268, + 0.00441147992387414, + 0.029306884855031967, + 0.006650082767009735, + 0.00520169036462903, + 0.006640739273279905, + 0.0038309504743665457, + 0.005401477217674255, + 0.010649054311215878, + 0.0032258343417197466, + 0.006945214234292507, + 0.005276377312839031, + 0.029427943751215935, + 0.03006455861032009, + 0.008768359199166298, + 0.0103181516751647, + 0.01500939391553402, + 0.006102959159761667, + 0.07721900194883347, + 0.013509346172213554, + 0.006882181856781244, + 0.0026281762402504683, + 0.0025271084159612656, + 0.003008225467056036, + 0.00764260720461607, + 0.0018644207157194614, + 0.010258919559419155, + 0.0030089239589869976, + 0.00924768764525652, + 0.004947038367390633, + 0.005881949793547392, + 0.006662530358880758, + 0.00947394035756588, + 0.004602395463734865, + 0.01958915963768959, + 0.004645366687327623, + 0.0689128115773201, + 0.022954512387514114, + 0.012420657090842724, + 0.012991110794246197, + 0.10302302986383438, + 0.013303413055837154, + 0.027004271745681763, + 0.008834712207317352, + 0.0020310610998421907, + 0.0018632200080901384, + 0.0010636781807988882, + 0.0011070248438045382, + 0.003705894574522972, + 0.0010633886558935046, + 0.0031475152354687452, + 0.0011100491974502802, + 0.005525014828890562, + 0.0027618107851594687, + 0.002854557242244482, + 0.004350034054368734, + 0.007480808068066835, + 0.0025714996736496687, + 0.012110256589949131, + 0.0035868787672370672, + 0.018019376322627068, + 0.006384796928614378, + 0.004553557373583317, + 0.006827754434198141, + 0.018210357055068016, + 0.003373195417225361, + 0.03926834464073181, + 0.004228904843330383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.10035011172294617 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762094259262085 + ], + "opposite_logprob_tokens": [ + -0.8262094259262085 + ], + "logprob_reward": -0.5762094259262085, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762094259262085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23725609481334686, + 0.039670541882514954, + 0.05670652538537979, + 0.11757565289735794, + 0.02529117278754711, + 0.02314032055437565, + 0.24128064513206482, + 0.06616827100515366, + 0.04803433641791344, + 0.14487652480602264 + ], + "credit_map": [ + 0.036288995295763016, + 0.017113354057073593, + 0.024728091433644295, + 0.022626250982284546, + 0.029759163036942482, + 0.02362789958715439, + 0.04183219000697136, + 0.04128013178706169, + 0.005240676924586296, + 0.0038652990479022264, + 0.002330022631213069, + 0.004366218112409115, + 0.008254609070718288, + 0.0030500071588903666, + 0.010347835719585419, + 0.0022158718202263117, + 0.005529175046831369, + 0.008329210802912712, + 0.004923075903207064, + 0.006525153759866953, + 0.009825888089835644, + 0.003722015768289566, + 0.0059292553924024105, + 0.01192274410277605, + 0.01967225782573223, + 0.020101025700569153, + 0.0062536089681088924, + 0.007974042557179928, + 0.009670598432421684, + 0.005261396523565054, + 0.04070831835269928, + 0.007934389635920525, + 0.004808735102415085, + 0.001385629060678184, + 0.0015915854601189494, + 0.001818573335185647, + 0.004513161722570658, + 0.0011040299432352185, + 0.008272654376924038, + 0.0017968012252822518, + 0.0027557832654565573, + 0.002247289754450321, + 0.003820384619757533, + 0.00243570888414979, + 0.003380734007805586, + 0.002104891464114189, + 0.0034357246477156878, + 0.0029598025139421225, + 0.05996892228722572, + 0.022420236840844154, + 0.008104033768177032, + 0.006139167118817568, + 0.10084155946969986, + 0.01204284280538559, + 0.025358140468597412, + 0.006405717693269253, + 0.010842808522284031, + 0.010264396667480469, + 0.0028011819813400507, + 0.00424947077408433, + 0.021848564967513084, + 0.00482740206643939, + 0.007399079389870167, + 0.0039353580214083195, + 0.006525683216750622, + 0.004180970136076212, + 0.003977048676460981, + 0.005042349454015493, + 0.007506199646741152, + 0.004783899988979101, + 0.010010107420384884, + 0.006008075550198555, + 0.02484869956970215, + 0.009051269851624966, + 0.006214327644556761, + 0.010128777474164963, + 0.02717547118663788, + 0.005657675210386515, + 0.056058112531900406, + 0.005742167122662067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.5762094259262085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a low angle, slowly and relentlessly zooming in, intensifies the focus on a grieving young woman. Her brown hair is pulled back in a simple ponytail, with loose strands catching the sunlight like thin golden threads. She wears thin wire-rimmed glasses and a crisp light blue cotton dress shirt. The entire frame is dominated by the raw physicality of a violent cough that convulses her body; her shoulders arch violently, and her head jerk forward with each spasmodic, uncontrollable thump. Her eyes, already watery with underlying sorrow, squeeze tightly in agony, and her face flushes. The scene is bathed in a warm, ethereal golden hour light, creating a soft halo around her body against the blurred backdrop of an outdoor night. The stark contrast between the serene, beautiful light and her inner anguish creates an intense, poignant, and intimate atmosphere of vulnerability and pain.", + "seed": 11923, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_2/index_7917_seed_11923_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2741312185302377, + "vlm_logprob": -0.2741312185302377, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.2741312185302377, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1884937435388565 + ], + "opposite_logprob_tokens": [ + -1.9384937286376953 + ], + "logprob_reward": -0.1884937435388565, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1884937435388565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2871910631656647, + 0.10412617027759552, + 0.14800097048282623, + 0.08989959210157394, + 0.05622942000627518, + 0.08178959786891937, + 0.07921474426984787, + 0.05956791341304779, + 0.052865125238895416, + 0.04111543297767639 + ], + "credit_map": [ + 0.04927599057555199, + 0.020780600607395172, + 0.024606848135590553, + 0.02359814941883087, + 0.03025859408080578, + 0.067937932908535, + 0.042810820043087006, + 0.02792210690677166, + 0.01041449699550867, + 0.0062626260332763195, + 0.013921065255999565, + 0.013118994422256947, + 0.011738850735127926, + 0.010135216638445854, + 0.026354633271694183, + 0.012180285528302193, + 0.025304751470685005, + 0.010570306330919266, + 0.03325515612959862, + 0.013062507845461369, + 0.021835938096046448, + 0.008585739880800247, + 0.025551987811923027, + 0.009834586642682552, + 0.012744414620101452, + 0.00998217985033989, + 0.010025709867477417, + 0.006181113421916962, + 0.016685500741004944, + 0.017026204615831375, + 0.011439084075391293, + 0.005815389100462198, + 0.008647950366139412, + 0.005913891829550266, + 0.01429995708167553, + 0.0026616742834448814, + 0.011762380599975586, + 0.006241658236831427, + 0.0037999176420271397, + 0.002901989035308361, + 0.012220359407365322, + 0.004649715963751078, + 0.015035046264529228, + 0.005321229342371225, + 0.016456371173262596, + 0.006612616591155529, + 0.013977332040667534, + 0.007516929879784584, + 0.011250787414610386, + 0.010933355428278446, + 0.008227895013988018, + 0.007750920485705137, + 0.011764938943088055, + 0.008750522509217262, + 0.012902381829917431, + 0.007633946370333433, + 0.012465126812458038, + 0.006754611153155565, + 0.002853232203051448, + 0.005731755867600441, + 0.006871753837913275, + 0.01650056056678295, + 0.005141886416822672, + 0.0032489842269569635, + 0.008060675114393234, + 0.00400859909132123, + 0.004977976903319359, + 0.003852671943604946, + 0.005529318004846573, + 0.011321580037474632, + 0.006678867153823376, + 0.008435438387095928, + 0.0073590646497905254, + 0.0053832754492759705, + 0.0037558176554739475, + 0.00706517742946744, + 0.0038620270788669586, + 0.005891144275665283, + 0.003954149316996336, + 0.0038447799161076546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.1884937435388565 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934507489204407 + ], + "opposite_logprob_tokens": [ + -0.6934507489204407 + ], + "logprob_reward": -0.6934507489204407, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934507489204407, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08433690667152405, + 0.11876411736011505, + 0.07357249408960342, + 0.24674436450004578, + 0.05302349478006363, + 0.12254318594932556, + 0.04042421653866768, + 0.058130186051130295, + 0.14058583974838257, + 0.06187526136636734 + ], + "credit_map": [ + 0.013390426523983479, + 0.007019719574600458, + 0.01139379944652319, + 0.0065958090126514435, + 0.012031628750264645, + 0.01839526742696762, + 0.008715479634702206, + 0.006794771179556847, + 0.01788420043885708, + 0.006528342608362436, + 0.02945982664823532, + 0.01222118642181158, + 0.00795177835971117, + 0.015142577700316906, + 0.02011789381504059, + 0.00945830438286066, + 0.011086782440543175, + 0.005529726855456829, + 0.015617736615240574, + 0.00749241141602397, + 0.010027273558080196, + 0.004064333159476519, + 0.01566239632666111, + 0.004091838840395212, + 0.03301136940717697, + 0.023876702412962914, + 0.02275955304503441, + 0.026726393029093742, + 0.03692791610956192, + 0.04005030170083046, + 0.0235451590269804, + 0.03984697163105011, + 0.0066514513455331326, + 0.006289706565439701, + 0.005776801146566868, + 0.0034660722594708204, + 0.015986090525984764, + 0.006346066016703844, + 0.005079358350485563, + 0.0034279439132660627, + 0.02127627469599247, + 0.006493504624813795, + 0.026575779542326927, + 0.005854242015630007, + 0.01592317782342434, + 0.00887441448867321, + 0.029004167765378952, + 0.008541623130440712, + 0.0055222343653440475, + 0.005489325616508722, + 0.004079359583556652, + 0.003261773381382227, + 0.007531397510319948, + 0.0049255420453846455, + 0.0067113246768713, + 0.0029032586608082056, + 0.012434082105755806, + 0.007031119894236326, + 0.004324363078922033, + 0.003498693462461233, + 0.006930844858288765, + 0.014590222388505936, + 0.006242853589355946, + 0.003078002482652664, + 0.027116408571600914, + 0.008634146302938461, + 0.020020360127091408, + 0.009184014052152634, + 0.011259857565164566, + 0.035057615488767624, + 0.010260030627250671, + 0.019053397700190544, + 0.007699271664023399, + 0.010032456368207932, + 0.005278232507407665, + 0.010616641491651535, + 0.008710617199540138, + 0.008805937133729458, + 0.005612952634692192, + 0.0051191565580666065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.6934507489204407 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108712051063776 + ], + "opposite_logprob_tokens": [ + -4.511086940765381 + ], + "logprob_reward": -0.01108712051063776, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108712051063776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1468842774629593, + 0.07659652084112167, + 0.20500192046165466, + 0.07335788011550903, + 0.21518100798130035, + 0.08911684900522232, + 0.0737948790192604, + 0.036057647317647934, + 0.06521017849445343, + 0.018798857927322388 + ], + "credit_map": [ + 0.019303107634186745, + 0.009622747078537941, + 0.016111154109239578, + 0.014593093656003475, + 0.023756016045808792, + 0.014459066092967987, + 0.03541514277458191, + 0.01362395565956831, + 0.007195706479251385, + 0.007789634168148041, + 0.011527047492563725, + 0.007111174054443836, + 0.009307361207902431, + 0.010116612538695335, + 0.017239484935998917, + 0.0063094948418438435, + 0.037384793162345886, + 0.011644650250673294, + 0.048314910382032394, + 0.015615918673574924, + 0.028324520215392113, + 0.010254592634737492, + 0.04270143061876297, + 0.010761089622974396, + 0.00951447430998087, + 0.006811756175011396, + 0.00924559123814106, + 0.009363715536892414, + 0.008875331841409206, + 0.011333635076880455, + 0.0066109923645854, + 0.011602385900914669, + 0.027089767158031464, + 0.020250190049409866, + 0.022663144394755363, + 0.011816930957138538, + 0.07444525510072708, + 0.021566815674304962, + 0.023060208186507225, + 0.01428869366645813, + 0.015908921137452126, + 0.0035475061740726233, + 0.021327853202819824, + 0.004885964095592499, + 0.01337449997663498, + 0.005572933703660965, + 0.020121980458498, + 0.004377192817628384, + 0.01156485266983509, + 0.009187806397676468, + 0.008304393850266933, + 0.005415861960500479, + 0.011362725868821144, + 0.013079153373837471, + 0.007792037446051836, + 0.007088047917932272, + 0.010369905270636082, + 0.0025277226231992245, + 0.0021498803980648518, + 0.0013204957358539104, + 0.0028598466888070107, + 0.012880700640380383, + 0.0018663284135982394, + 0.002082766965031624, + 0.015788711607456207, + 0.0031291344203054905, + 0.0068709091283380985, + 0.002343629952520132, + 0.00300401053391397, + 0.021341638639569283, + 0.0036148906219750643, + 0.009117256850004196, + 0.001966682029888034, + 0.0032491714227944613, + 0.0018829244654625654, + 0.0030808711890131235, + 0.0026270709931850433, + 0.0024778530932962894, + 0.0015932732494547963, + 0.0019210110185667872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.01108712051063776 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867646470665932 + ], + "opposite_logprob_tokens": [ + -3.0486764907836914 + ], + "logprob_reward": -0.04867646470665932, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867646470665932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08430800586938858, + 0.07483167946338654, + 0.12257051467895508, + 0.1568317860364914, + 0.02849147655069828, + 0.3064579963684082, + 0.031203633174300194, + 0.03569941967725754, + 0.13663753867149353, + 0.02296789549291134 + ], + "credit_map": [ + 0.01473702397197485, + 0.006230645813047886, + 0.011023645289242268, + 0.009301196783781052, + 0.008295098319649696, + 0.01518898643553257, + 0.010577564127743244, + 0.008953841403126717, + 0.007532036397606134, + 0.010642735287547112, + 0.010831700637936592, + 0.010989591479301453, + 0.008360488340258598, + 0.008257875218987465, + 0.012155808508396149, + 0.00606144405901432, + 0.015611086040735245, + 0.014406793750822544, + 0.020333297550678253, + 0.010471230372786522, + 0.018779294565320015, + 0.008351154625415802, + 0.02715142071247101, + 0.007466228678822517, + 0.020843705162405968, + 0.014912883751094341, + 0.014752782881259918, + 0.01759709231555462, + 0.02318369224667549, + 0.024453289806842804, + 0.014774241484701633, + 0.026314103975892067, + 0.003957616165280342, + 0.0024445124436169863, + 0.00726673798635602, + 0.0018430000636726618, + 0.004466236103326082, + 0.003725520335137844, + 0.002734218258410692, + 0.0020536354277282953, + 0.03560666739940643, + 0.024706367403268814, + 0.04823219031095505, + 0.01863750070333481, + 0.05105968937277794, + 0.02137756161391735, + 0.08186584711074829, + 0.02497217059135437, + 0.0040925778448581696, + 0.003062022849917412, + 0.003144999733194709, + 0.0035376171581447124, + 0.005633357912302017, + 0.003112060483545065, + 0.004543254151940346, + 0.004077744670212269, + 0.010237745009362698, + 0.0028159148059785366, + 0.0020982027053833008, + 0.0013664744328707457, + 0.00300035672262311, + 0.012210715562105179, + 0.001983824884518981, + 0.001986186718568206, + 0.033034417778253555, + 0.008002683520317078, + 0.013001291081309319, + 0.006850776262581348, + 0.007656724192202091, + 0.04376531019806862, + 0.010242797434329987, + 0.01408353727310896, + 0.0029214771930128336, + 0.003885129000991583, + 0.0021093119867146015, + 0.004489607643336058, + 0.0026514935307204723, + 0.002911644522100687, + 0.002099771751090884, + 0.0018994588172063231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.04867646470665932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and relentlessly zoom in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject a young woman with brown hair in a ponytail wearing wire-rimmed glasses and a light blue cotton dress shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman exhibit violent coughing with convulsing shoulders, jerking head, and tightly squeezed eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a blurred outdoor night backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289480149745941 + ], + "opposite_logprob_tokens": [ + -1.0539480447769165 + ], + "logprob_reward": -0.4289480149745941, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289480149745941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20711661875247955, + 0.1042814701795578, + 0.23180654644966125, + 0.1691918522119522, + 0.045484937727451324, + 0.04853544011712074, + 0.035506438463926315, + 0.06415925920009613, + 0.06264184415340424, + 0.031275615096092224 + ], + "credit_map": [ + 0.02687094919383526, + 0.017879929393529892, + 0.027413886040449142, + 0.01485870499163866, + 0.03157529607415199, + 0.02917235717177391, + 0.045299820601940155, + 0.014045667834579945, + 0.016298437491059303, + 0.006980626378208399, + 0.02226262353360653, + 0.009644068777561188, + 0.013211023062467575, + 0.013822019100189209, + 0.016059551388025284, + 0.006003122311085463, + 0.05167769268155098, + 0.016577737405896187, + 0.07322347164154053, + 0.014997600577771664, + 0.020734108984470367, + 0.024010907858610153, + 0.017048293724656105, + 0.013536724261939526, + 0.029843131080269814, + 0.019465675577521324, + 0.01877828687429428, + 0.014896446838974953, + 0.015839478000998497, + 0.03945767879486084, + 0.012131058610975742, + 0.018780099228024483, + 0.007662328891456127, + 0.004199518822133541, + 0.010075037367641926, + 0.0034045695792883635, + 0.005491435062140226, + 0.00765229482203722, + 0.002859636675566435, + 0.0041401139460504055, + 0.006189634092152119, + 0.004448411054909229, + 0.00777675723657012, + 0.0030079479329288006, + 0.007904620841145515, + 0.0041656214743852615, + 0.012222275137901306, + 0.002820169785991311, + 0.004356143996119499, + 0.004564378876239061, + 0.002442275872454047, + 0.004351289477199316, + 0.006212315056473017, + 0.00385358021594584, + 0.006170810200273991, + 0.003555647563189268, + 0.01849023438990116, + 0.0038097319193184376, + 0.0036394516937434673, + 0.004524309653788805, + 0.004494332708418369, + 0.022404756397008896, + 0.00357508915476501, + 0.0032213572412729263, + 0.011125901713967323, + 0.003963265102356672, + 0.008162743411958218, + 0.004171367269009352, + 0.005622957367449999, + 0.014548401348292828, + 0.006018861196935177, + 0.009028344415128231, + 0.003991618752479553, + 0.004575475584715605, + 0.002775952685624361, + 0.005095637869089842, + 0.004110765177756548, + 0.0040771071799099445, + 0.0038648187182843685, + 0.0027842374984174967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.4289480149745941 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_3/index_tensor([3892, 3892])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_3/index_tensor([3892, 3892])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..870afd65aff02b542cb451b4d78162b388ad1f25 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_3/index_tensor([3892, 3892])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "still, extreme close-up focuses squarely on an eye brimming with tears from intense emotion. The amber iris appears remarkably clear around the slightly dilated pupil, the white of the eye bloodshot with the fine lines of emotion. A soft, sideways windowlight casts a precise spotlight on the eye, revealing the reddened skin around it and the long, slightly curled black lashes. As the camera gazes, a pool of clear tears quickly and silently flows from the depths of the eye, instantly filling the edge of the lower lid, forming a plump, trembling surface that drenches the entire eye socket with a shimmering sheen. A few crystalline droplets of water have already formed, delicately dangling from the tips of the eyelashes, reflecting a brilliant diamond-like light. At this moment, the tears have reached the critical point of overflowing, as if any slight blink or tremor of breath will make this hot tear break free and slide down. The whole picture is full of suppressed sadness and extreme fragility, creating a heartbreaking tranquil atmosphere.", + "seed": 25941, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_3/index_3892_seed_25941_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07576241604983806, + "vlm_logprob": -0.07576241604983806, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.07576241604983806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004547255113720894 + ], + "opposite_logprob_tokens": [ + -6.379547119140625 + ], + "logprob_reward": -0.004547255113720894, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.004547255113720894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08485185354948044, + 0.07025317847728729, + 0.34440305829048157, + 0.046925127506256104, + 0.08250262588262558, + 0.059955183416604996, + 0.07536538690328598, + 0.04789126291871071, + 0.10583236068487167, + 0.08201996237039566 + ], + "credit_map": [ + 0.008322471752762794, + 0.01660488359630108, + 0.010249362327158451, + 0.010176868177950382, + 0.01046990416944027, + 0.008435056544840336, + 0.009368167258799076, + 0.01122513972222805, + 0.006617238279432058, + 0.00774725154042244, + 0.009855203330516815, + 0.006080005783587694, + 0.007261082995682955, + 0.01194394938647747, + 0.012803586199879646, + 0.007944858632981777, + 0.042296092957258224, + 0.05984257906675339, + 0.02422928810119629, + 0.03196737542748451, + 0.04445357993245125, + 0.052813753485679626, + 0.044375594705343246, + 0.04442478343844414, + 0.00499033322557807, + 0.006855262443423271, + 0.004695745650678873, + 0.006289465818554163, + 0.0051742796786129475, + 0.005828604102134705, + 0.007117039989680052, + 0.005974396597594023, + 0.008010306395590305, + 0.007863275706768036, + 0.009140446782112122, + 0.015567940659821033, + 0.01024120394140482, + 0.009307239204645157, + 0.012499450705945492, + 0.009872754104435444, + 0.007373169530183077, + 0.008475696668028831, + 0.006662305910140276, + 0.005256651435047388, + 0.010556782595813274, + 0.007236337289214134, + 0.007465546019375324, + 0.006928686983883381, + 0.007416058797389269, + 0.009970284067094326, + 0.007210718002170324, + 0.009717549197375774, + 0.017178121954202652, + 0.008733672089874744, + 0.008241048082709312, + 0.006897933315485716, + 0.005877678748220205, + 0.00646964879706502, + 0.00488246139138937, + 0.005467727314680815, + 0.007325274869799614, + 0.00399344926699996, + 0.007356264162808657, + 0.006518755108118057, + 0.013115579262375832, + 0.011449596844613552, + 0.014954164624214172, + 0.018675468862056732, + 0.010185877792537212, + 0.009859570302069187, + 0.01119561493396759, + 0.01639648713171482, + 0.009445221163332462, + 0.013001962564885616, + 0.00809748750180006, + 0.01010341476649046, + 0.014792000874876976, + 0.0074014319106936455, + 0.01048369612544775, + 0.008694748394191265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.004547255113720894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870939254760742 + ], + "opposite_logprob_tokens": [ + -3.0487093925476074 + ], + "logprob_reward": -0.04870939254760742, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870939254760742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17874467372894287, + 0.07240220159292221, + 0.12512408196926117, + 0.03006763756275177, + 0.04530490189790726, + 0.021436311304569244, + 0.11221592873334885, + 0.030557164922356606, + 0.1147979199886322, + 0.26934918761253357 + ], + "credit_map": [ + 0.016748232766985893, + 0.037147972732782364, + 0.021822327747941017, + 0.019689057022333145, + 0.020661473274230957, + 0.02303636632859707, + 0.020784789696335793, + 0.018854444846510887, + 0.005371037404984236, + 0.010114726610481739, + 0.00818560179322958, + 0.01075836457312107, + 0.008894854225218296, + 0.00965296570211649, + 0.011355881579220295, + 0.008068759925663471, + 0.01260184682905674, + 0.01979031227529049, + 0.012930450029671192, + 0.012608366087079048, + 0.015419993549585342, + 0.018944941461086273, + 0.017476772889494896, + 0.015351388603448868, + 0.003947587683796883, + 0.0043621789664030075, + 0.0027828195597976446, + 0.0037970689591020346, + 0.0036546194460242987, + 0.003071168204769492, + 0.003959708847105503, + 0.004492484033107758, + 0.007560653146356344, + 0.0053067272529006, + 0.0059391590766608715, + 0.00450415862724185, + 0.0059547726996243, + 0.004725804086774588, + 0.004850973375141621, + 0.006462653633207083, + 0.0023058836814016104, + 0.0030709553975611925, + 0.0027952357195317745, + 0.0024966797791421413, + 0.0021718700882047415, + 0.002285405993461609, + 0.0031780190765857697, + 0.003132259240373969, + 0.010874401777982712, + 0.015990549698472023, + 0.011511409655213356, + 0.013832337222993374, + 0.02165309526026249, + 0.01178658101707697, + 0.013247030787169933, + 0.013320514000952244, + 0.0038306971546262503, + 0.004447511862963438, + 0.0036876981612294912, + 0.0035262417513877153, + 0.004787163808941841, + 0.003916020039469004, + 0.0031249774619936943, + 0.0032368546817451715, + 0.010148366913199425, + 0.015618765726685524, + 0.011296088807284832, + 0.021413011476397514, + 0.014490414410829544, + 0.011218871921300888, + 0.016386406496167183, + 0.014225996099412441, + 0.024392200633883476, + 0.048106856644153595, + 0.0263300109654665, + 0.03632354363799095, + 0.04990329220890999, + 0.025584623217582703, + 0.02719726413488388, + 0.03151138499379158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.04870939254760742 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810739889740944 + ], + "opposite_logprob_tokens": [ + -3.288107395172119 + ], + "logprob_reward": -0.03810739889740944, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810739889740944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15483900904655457, + 0.10783802717924118, + 0.13117827475070953, + 0.041045207530260086, + 0.16562174260616302, + 0.05460292473435402, + 0.07466010749340057, + 0.030725974589586258, + 0.13777495920658112, + 0.10171379894018173 + ], + "credit_map": [ + 0.01950533501803875, + 0.018613891676068306, + 0.013870648108422756, + 0.020265081897377968, + 0.01796208694577217, + 0.01774129644036293, + 0.02137499675154686, + 0.02550567500293255, + 0.013660123571753502, + 0.010688086040318012, + 0.017355823889374733, + 0.010979916900396347, + 0.01208187360316515, + 0.020004553720355034, + 0.012240508571267128, + 0.010827137157320976, + 0.01674947328865528, + 0.017086248844861984, + 0.01856270805001259, + 0.012682541273534298, + 0.018284164369106293, + 0.01965368166565895, + 0.01615334302186966, + 0.012006115168333054, + 0.004936814308166504, + 0.005508360918611288, + 0.0037891650572419167, + 0.006756787188351154, + 0.005575752351433039, + 0.0039901891723275185, + 0.005955692380666733, + 0.004532448947429657, + 0.01686360314488411, + 0.027439264580607414, + 0.01383343618363142, + 0.018837451934814453, + 0.022397926077246666, + 0.015579343773424625, + 0.029762614518404007, + 0.020908091217279434, + 0.006761979311704636, + 0.006041713058948517, + 0.005419889930635691, + 0.0066026425920426846, + 0.004739893600344658, + 0.005298364441841841, + 0.012021875940263271, + 0.0077165658585727215, + 0.010085663758218288, + 0.006914922036230564, + 0.01052567083388567, + 0.011179862543940544, + 0.006679894868284464, + 0.008983313106000423, + 0.010359075851738453, + 0.009931701235473156, + 0.003186374669894576, + 0.0044684940949082375, + 0.0028258732054382563, + 0.004950117785483599, + 0.002853043843060732, + 0.004011427517980337, + 0.00444862712174654, + 0.003982014954090118, + 0.018062710762023926, + 0.015739681199193, + 0.01949664019048214, + 0.025290075689554214, + 0.010874712839722633, + 0.013048047199845314, + 0.01679242216050625, + 0.018470652401447296, + 0.009001228958368301, + 0.010653305798768997, + 0.011942761018872261, + 0.024245629087090492, + 0.008240138180553913, + 0.009333732537925243, + 0.014340933412313461, + 0.013956073671579361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.03810739889740944 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706947326660156 + ], + "opposite_logprob_tokens": [ + -2.1270694732666016 + ], + "logprob_reward": -0.12706947326660156, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706947326660156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1322290301322937, + 0.09836810827255249, + 0.28851839900016785, + 0.13630717992782593, + 0.06893031299114227, + 0.05159255117177963, + 0.043808046728372574, + 0.032816071063280106, + 0.03156638145446777, + 0.11586388945579529 + ], + "credit_map": [ + 0.014122256077826023, + 0.021586567163467407, + 0.01413514744490385, + 0.01608404330909252, + 0.01615711860358715, + 0.018445096909999847, + 0.013469425030052662, + 0.018229376524686813, + 0.01261800155043602, + 0.012379187159240246, + 0.011803421191871166, + 0.011454843915998936, + 0.011532754637300968, + 0.013770127668976784, + 0.00884166732430458, + 0.015968112275004387, + 0.03596678376197815, + 0.0480513721704483, + 0.03205770254135132, + 0.026289168745279312, + 0.038273099809885025, + 0.04601867496967316, + 0.027130262926220894, + 0.03473132848739624, + 0.01724931225180626, + 0.01791069470345974, + 0.01708589680492878, + 0.020644372329115868, + 0.01428448036313057, + 0.014377433806657791, + 0.017359379678964615, + 0.017395608127117157, + 0.007310447283089161, + 0.009396592155098915, + 0.006937905680388212, + 0.008729543536901474, + 0.009264891035854816, + 0.008509313687682152, + 0.00861249677836895, + 0.010169128887355328, + 0.006479657720774412, + 0.007842146791517735, + 0.00678359204903245, + 0.005286010913550854, + 0.007199899293482304, + 0.005241832695901394, + 0.006823393050581217, + 0.005936020985245705, + 0.006698061246424913, + 0.003567233681678772, + 0.006030100863426924, + 0.006558089982718229, + 0.0036854734644293785, + 0.004652718082070351, + 0.006587240379303694, + 0.006029128096997738, + 0.0035195383243262768, + 0.003547807689756155, + 0.003237281460314989, + 0.0035271879751235247, + 0.005776230711489916, + 0.004183054435998201, + 0.003899326780810952, + 0.005125644616782665, + 0.0027725668624043465, + 0.0046766819432377815, + 0.0032919927034527063, + 0.004719975404441357, + 0.006325858179479837, + 0.0028384774923324585, + 0.0035239960998296738, + 0.0034168337006121874, + 0.012641405686736107, + 0.01155675295740366, + 0.009913265705108643, + 0.017784230411052704, + 0.008738434873521328, + 0.009885956533253193, + 0.021446455270051956, + 0.023897381499409676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.12706947326660156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.160378560423851 + ], + "opposite_logprob_tokens": [ + -1.9103785753250122 + ], + "logprob_reward": -0.160378560423851, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.160378560423851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12118268758058548, + 0.14130060374736786, + 0.19756220281124115, + 0.07798337191343307, + 0.16702347993850708, + 0.051010534167289734, + 0.03909847140312195, + 0.057626668363809586, + 0.10817424207925797, + 0.03903770446777344 + ], + "credit_map": [ + 0.012891784310340881, + 0.018982550129294395, + 0.011588498018682003, + 0.015165936201810837, + 0.0163568127900362, + 0.016125855967402458, + 0.014110038056969643, + 0.01596120372414589, + 0.013163912110030651, + 0.01986122690141201, + 0.013955219648778439, + 0.014584489166736603, + 0.018608637154102325, + 0.02388705126941204, + 0.019031543284654617, + 0.018208520486950874, + 0.02482500486075878, + 0.027460597455501556, + 0.024473339319229126, + 0.018305305391550064, + 0.024262450635433197, + 0.03002000041306019, + 0.028313685208559036, + 0.0199018195271492, + 0.009230989962816238, + 0.009286690503358841, + 0.008101432584226131, + 0.015252647921442986, + 0.00870864000171423, + 0.006851697340607643, + 0.011468656361103058, + 0.009082614444196224, + 0.016699787229299545, + 0.01603054814040661, + 0.017431773245334625, + 0.029665928333997726, + 0.017077285796403885, + 0.014776506461203098, + 0.033921804279088974, + 0.021419845521450043, + 0.005180165637284517, + 0.004851094447076321, + 0.005058811511844397, + 0.00812661461532116, + 0.004119338002055883, + 0.00502969091758132, + 0.012967689894139767, + 0.005677124951034784, + 0.004595674108713865, + 0.003670048201456666, + 0.005066705401986837, + 0.0072675831615924835, + 0.003953233361244202, + 0.004292325582355261, + 0.005365684162825346, + 0.0048872195184230804, + 0.004664393607527018, + 0.007741250097751617, + 0.006983596365898848, + 0.0069399685598909855, + 0.006948328111320734, + 0.00886234175413847, + 0.007732456550002098, + 0.007754328195005655, + 0.010793029330670834, + 0.011590435169637203, + 0.014532159082591534, + 0.02398725412786007, + 0.00894236285239458, + 0.010705781169235706, + 0.01242869719862938, + 0.015194525010883808, + 0.003853751812130213, + 0.003416238119825721, + 0.0057050189934670925, + 0.008427942171692848, + 0.002705727471038699, + 0.0045873564667999744, + 0.005203623324632645, + 0.005138046108186245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.160378560423851 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, extreme close-up focuses squarely on an eye brimming with tears from intense emotion. The amber iris appears remarkably clear around the slightly dilated pupil, the white of the eye bloodshot with the fine lines of emotion. A soft, sideways windowlight casts a precise spotlight on the eye, revealing the reddened skin around it and the long, slightly curled black lashes. As the camera gazes, a pool of clear tears quickly and silently flows from the depths of the eye, instantly filling the edge of the lower lid, forming a plump, trembling surface that drenches the entire eye socket with a shimmering sheen. A few crystalline droplets of water have already formed, delicately dangling from the tips of the eyelashes, reflecting a brilliant diamond-like light. At this moment, the tears have reached the critical point of overflowing, as if any slight blink or tremor of breath will make this hot tear break free and slide down. The whole picture is full of suppressed sadness and extreme fragility, creating a heartbreaking tranquil atmosphere.", + "seed": 25942, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_3/index_3892_seed_25942_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.039896738110110165, + "vlm_logprob": -0.039896738110110165, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.039896738110110165, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019535997416824102 + ], + "opposite_logprob_tokens": [ + -8.376953125 + ], + "logprob_reward": -0.0019535997416824102, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.0019535997416824102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16415712237358093, + 0.17801512777805328, + 0.08700928837060928, + 0.04936499148607254, + 0.10602699220180511, + 0.16709114611148834, + 0.1009531021118164, + 0.04163527488708496, + 0.06464999914169312, + 0.04109698906540871 + ], + "credit_map": [ + 0.020295636728405952, + 0.01386379636824131, + 0.021582024171948433, + 0.01845308020710945, + 0.020246252417564392, + 0.021597182378172874, + 0.018227089196443558, + 0.029892047867178917, + 0.013746102340519428, + 0.010068959556519985, + 0.0295022651553154, + 0.02093089185655117, + 0.014043854549527168, + 0.015327486209571362, + 0.0271417535841465, + 0.0472538135945797, + 0.011270851828157902, + 0.005405247211456299, + 0.010980003513395786, + 0.01108836941421032, + 0.007578506600111723, + 0.008247975260019302, + 0.010575627908110619, + 0.021862709894776344, + 0.005619262810796499, + 0.004207329824566841, + 0.005645087454468012, + 0.008162988349795341, + 0.00417185015976429, + 0.005141930188983679, + 0.007043990306556225, + 0.009372550994157791, + 0.008802598342299461, + 0.006429618690162897, + 0.027223220095038414, + 0.009905966930091381, + 0.009417528286576271, + 0.00730682909488678, + 0.008821627125144005, + 0.028119606897234917, + 0.017293861135840416, + 0.008929437957704067, + 0.01859986037015915, + 0.015852246433496475, + 0.02074413001537323, + 0.020203249529004097, + 0.01644829660654068, + 0.049020055681467056, + 0.009541758336126804, + 0.006978947203606367, + 0.009621293283998966, + 0.011548377573490143, + 0.011262227781116962, + 0.013297866098582745, + 0.010686798952519894, + 0.028015833348035812, + 0.006883757654577494, + 0.0037469277158379555, + 0.004535045009106398, + 0.006620448548346758, + 0.005148543044924736, + 0.004372940864413977, + 0.004848218522965908, + 0.005479391198605299, + 0.008516475558280945, + 0.004247234668582678, + 0.00918483268469572, + 0.009317979216575623, + 0.007354862056672573, + 0.004894002806395292, + 0.007355398032814264, + 0.013779216445982456, + 0.005209567025303841, + 0.0033015687949955463, + 0.004303935915231705, + 0.007600771263241768, + 0.005573804024606943, + 0.005369825754314661, + 0.005003191530704498, + 0.004734322894364595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.0019535997416824102 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498645454645157 + ], + "opposite_logprob_tokens": [ + -2.9299864768981934 + ], + "logprob_reward": -0.05498645454645157, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498645454645157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11555468291044235, + 0.19520270824432373, + 0.15374591946601868, + 0.05023037642240524, + 0.22997450828552246, + 0.03515854477882385, + 0.10634931921958923, + 0.03954290971159935, + 0.05756806209683418, + 0.016672920435667038 + ], + "credit_map": [ + 0.013256244361400604, + 0.013008737936615944, + 0.014083676040172577, + 0.01131584495306015, + 0.015400880016386509, + 0.020442258566617966, + 0.012505348771810532, + 0.015541684813797474, + 0.015072689391672611, + 0.01171858049929142, + 0.03411644697189331, + 0.022191818803548813, + 0.016290927305817604, + 0.016435343772172928, + 0.01892753690481186, + 0.06044936180114746, + 0.01800556667149067, + 0.015834888443350792, + 0.018412034958600998, + 0.013022716157138348, + 0.028211301192641258, + 0.021331684663891792, + 0.01980040967464447, + 0.019127318635582924, + 0.005968750920146704, + 0.0044646309688687325, + 0.0055529032833874226, + 0.008248109370470047, + 0.007033422123640776, + 0.006371965166181326, + 0.00731133297085762, + 0.005279263947159052, + 0.015819285064935684, + 0.010551678948104382, + 0.05995938554406166, + 0.02163834683597088, + 0.015445264056324959, + 0.015620144084095955, + 0.01955258660018444, + 0.07138780504465103, + 0.0030089050997048616, + 0.002796824788674712, + 0.0035814654547721148, + 0.004727232735604048, + 0.005993434228003025, + 0.0038063880056142807, + 0.00426167668774724, + 0.006982619408518076, + 0.010822035372257233, + 0.006875523366034031, + 0.014676333405077457, + 0.012447923421859741, + 0.01287867221981287, + 0.01094843540340662, + 0.013552953489124775, + 0.024147439748048782, + 0.004160316661000252, + 0.004656516015529633, + 0.0036572895478457212, + 0.0051974584348499775, + 0.005433726590126753, + 0.004890272859483957, + 0.005815268028527498, + 0.005732061807066202, + 0.0052086603827774525, + 0.004361859057098627, + 0.005975700449198484, + 0.011066913604736328, + 0.005993746221065521, + 0.005811380688101053, + 0.010718123987317085, + 0.008431676775217056, + 0.0015592187410220504, + 0.001441635307855904, + 0.0017257037106901407, + 0.0031402225140482187, + 0.002196911023929715, + 0.0021821835543960333, + 0.002090156776830554, + 0.0023368909023702145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.05498645454645157 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898624986410141 + ], + "opposite_logprob_tokens": [ + -2.578986167907715 + ], + "logprob_reward": -0.07898624986410141, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898624986410141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3697434961795807, + 0.11140497028827667, + 0.10392758250236511, + 0.09605865925550461, + 0.13428397476673126, + 0.07019862532615662, + 0.043128851801157, + 0.025706399232149124, + 0.02298441343009472, + 0.022562989965081215 + ], + "credit_map": [ + 0.031034383922815323, + 0.03391563147306442, + 0.03725167736411095, + 0.06793596595525742, + 0.036398082971572876, + 0.045555610209703445, + 0.06543450057506561, + 0.05221763998270035, + 0.007021251134574413, + 0.005930410232394934, + 0.021316982805728912, + 0.01305676344782114, + 0.008147847838699818, + 0.01026227604597807, + 0.012327099218964577, + 0.033342331647872925, + 0.011243278160691261, + 0.00904876459389925, + 0.01554115954786539, + 0.011877353303134441, + 0.012975183315575123, + 0.011401604861021042, + 0.013641297817230225, + 0.01819894090294838, + 0.012579761445522308, + 0.008551331236958504, + 0.009947696700692177, + 0.012567372061312199, + 0.01338926050812006, + 0.012369262054562569, + 0.012361674569547176, + 0.014292300678789616, + 0.007881129160523415, + 0.006257978267967701, + 0.026872284710407257, + 0.020247355103492737, + 0.008929554373025894, + 0.009156177751719952, + 0.0155021408572793, + 0.03943735733628273, + 0.004942936357110739, + 0.005442487075924873, + 0.008528055623173714, + 0.015346015803515911, + 0.007107648998498917, + 0.005852861795574427, + 0.01082693412899971, + 0.0121516864746809, + 0.004381053615361452, + 0.0026980782859027386, + 0.0054449234157800674, + 0.006814960390329361, + 0.004372905008494854, + 0.004425838124006987, + 0.0049622380174696445, + 0.010028854943811893, + 0.0023335381411015987, + 0.0025979483034461737, + 0.002365951193496585, + 0.004541274160146713, + 0.0030222306959331036, + 0.0025085618253797293, + 0.004287791438400745, + 0.004049102310091257, + 0.0021569940727204084, + 0.0020158032421022654, + 0.002435559406876564, + 0.004280965309590101, + 0.0023264905903488398, + 0.0023469659499824047, + 0.003962527960538864, + 0.003459108294919133, + 0.002089583547785878, + 0.0019864197820425034, + 0.002361913211643696, + 0.004399995319545269, + 0.002758343704044819, + 0.0031533006113022566, + 0.0029976684600114822, + 0.0028157660271972418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.07898624986410141 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204290688037872 + ], + "opposite_logprob_tokens": [ + -2.8120429515838623 + ], + "logprob_reward": -0.06204290688037872, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204290688037872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10484188050031662, + 0.2591315507888794, + 0.04199458658695221, + 0.056177955120801926, + 0.1883126050233841, + 0.05420089513063431, + 0.041623376309871674, + 0.14484110474586487, + 0.06585734337568283, + 0.0430186502635479 + ], + "credit_map": [ + 0.011285393498837948, + 0.011096978560090065, + 0.011723419651389122, + 0.014724637381732464, + 0.012690169736742973, + 0.015478902496397495, + 0.017219500616192818, + 0.010622888803482056, + 0.018387574702501297, + 0.012031947262585163, + 0.05317721143364906, + 0.027881832793354988, + 0.015177853405475616, + 0.019351040944457054, + 0.024170950055122375, + 0.08895315229892731, + 0.005503762047737837, + 0.004181254655122757, + 0.004165911115705967, + 0.007298627868294716, + 0.004772950895130634, + 0.004696148447692394, + 0.005902755074203014, + 0.005473176017403603, + 0.004744238220155239, + 0.004604436922818422, + 0.0054480936378240585, + 0.013270466588437557, + 0.005427635740488768, + 0.00479543162509799, + 0.009187553077936172, + 0.00870010256767273, + 0.012187345884740353, + 0.009067167527973652, + 0.03932413458824158, + 0.026723161339759827, + 0.014222752302885056, + 0.013620146550238132, + 0.02170693129301071, + 0.05146099254488945, + 0.0036963876336812973, + 0.0036120631266385317, + 0.006440527271479368, + 0.011530770920217037, + 0.004980689380317926, + 0.005707933567464352, + 0.0070001487620174885, + 0.011232376098632812, + 0.004917314276099205, + 0.0034945751540362835, + 0.004490314982831478, + 0.008995840325951576, + 0.0049834237433969975, + 0.004298064857721329, + 0.005022536963224411, + 0.005421311128884554, + 0.019990254193544388, + 0.016089823096990585, + 0.01407334953546524, + 0.023367688059806824, + 0.017914552241563797, + 0.014491959474980831, + 0.02054733783006668, + 0.018366146832704544, + 0.00826416164636612, + 0.004194538574665785, + 0.007075142115354538, + 0.012495431117713451, + 0.005688832141458988, + 0.007653791457414627, + 0.011127835139632225, + 0.009357616305351257, + 0.0035184232983738184, + 0.003500042948871851, + 0.004538519307971001, + 0.008516250178217888, + 0.005335730034857988, + 0.005598708987236023, + 0.005467910785228014, + 0.0065430691465735435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.06204290688037872 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015144795179367065 + ], + "opposite_logprob_tokens": [ + -6.501514434814453 + ], + "logprob_reward": -0.0015144795179367065, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015144795179367065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1093786358833313, + 0.32235679030418396, + 0.06914187967777252, + 0.05127299204468727, + 0.0808267742395401, + 0.060373514890670776, + 0.09031344205141068, + 0.08016595989465714, + 0.05580022931098938, + 0.08036979287862778 + ], + "credit_map": [ + 0.00905625056475401, + 0.010509137995541096, + 0.011883953586220741, + 0.016614332795143127, + 0.012850392609834671, + 0.017296202480793, + 0.016990555450320244, + 0.014177823439240456, + 0.021919650956988335, + 0.01508866436779499, + 0.0619945228099823, + 0.04965920373797417, + 0.01912648230791092, + 0.024547772482037544, + 0.03342676907777786, + 0.09659374505281448, + 0.006130342837423086, + 0.004404156003147364, + 0.00894500594586134, + 0.011444211937487125, + 0.006235694978386164, + 0.006211096420884132, + 0.009603755548596382, + 0.016167618334293365, + 0.005391490645706654, + 0.0034773158840835094, + 0.00618216535076499, + 0.006956950295716524, + 0.004924801643937826, + 0.005382672883570194, + 0.006291504018008709, + 0.012666095048189163, + 0.005311201326549053, + 0.004048827104270458, + 0.02013293094933033, + 0.006253537721931934, + 0.006021502427756786, + 0.005536692216992378, + 0.006765189114958048, + 0.026756903156638145, + 0.0072277323342859745, + 0.004496120382100344, + 0.006128822453320026, + 0.012257939204573631, + 0.006923690438270569, + 0.005485024768859148, + 0.008789082057774067, + 0.009065104648470879, + 0.008932274766266346, + 0.0057045575231313705, + 0.011929364874958992, + 0.009020883589982986, + 0.01137795764952898, + 0.009696406312286854, + 0.011935190297663212, + 0.02171681821346283, + 0.009442043490707874, + 0.008657027035951614, + 0.007801107596606016, + 0.014483219012618065, + 0.01013571210205555, + 0.007581495679914951, + 0.01167258433997631, + 0.01039277482777834, + 0.003848814871162176, + 0.0041658999398350716, + 0.006705224979668856, + 0.0109587786719203, + 0.0052812425419688225, + 0.005885512102395296, + 0.009202149696648121, + 0.009752614423632622, + 0.008592530153691769, + 0.008187218569219112, + 0.007755151484161615, + 0.01275371853262186, + 0.008025556802749634, + 0.0098927803337574, + 0.015148969367146492, + 0.010013868100941181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0015144795179367065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, extreme close-up focuses squarely on an eye brimming with tears from intense emotion. The amber iris appears remarkably clear around the slightly dilated pupil, the white of the eye bloodshot with the fine lines of emotion. A soft, sideways windowlight casts a precise spotlight on the eye, revealing the reddened skin around it and the long, slightly curled black lashes. As the camera gazes, a pool of clear tears quickly and silently flows from the depths of the eye, instantly filling the edge of the lower lid, forming a plump, trembling surface that drenches the entire eye socket with a shimmering sheen. A few crystalline droplets of water have already formed, delicately dangling from the tips of the eyelashes, reflecting a brilliant diamond-like light. At this moment, the tears have reached the critical point of overflowing, as if any slight blink or tremor of breath will make this hot tear break free and slide down. The whole picture is full of suppressed sadness and extreme fragility, creating a heartbreaking tranquil atmosphere.", + "seed": 25943, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_3/index_3892_seed_25943_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09172793021425604, + "vlm_logprob": -0.09172793021425604, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.09172793021425604, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0060349623672664165 + ], + "opposite_logprob_tokens": [ + -5.881034851074219 + ], + "logprob_reward": -0.0060349623672664165, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0060349623672664165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26552337408065796, + 0.06670452654361725, + 0.06417066603899002, + 0.041462242603302, + 0.06936916708946228, + 0.07439190149307251, + 0.06955453008413315, + 0.027551153674721718, + 0.2227649986743927, + 0.0985073521733284 + ], + "credit_map": [ + 0.040943827480077744, + 0.036736272275447845, + 0.04013201966881752, + 0.034272633492946625, + 0.03191321715712547, + 0.028711631894111633, + 0.02654736302793026, + 0.026266414672136307, + 0.01041856873780489, + 0.007131888065487146, + 0.010144921019673347, + 0.005604095291346312, + 0.006997461896389723, + 0.009539680555462837, + 0.005816237535327673, + 0.011051678098738194, + 0.010172544978559017, + 0.007013714872300625, + 0.007725437171757221, + 0.007341605611145496, + 0.008077024482190609, + 0.0067948708310723305, + 0.007643077056854963, + 0.009402391500771046, + 0.005366710480302572, + 0.005118193570524454, + 0.004470754414796829, + 0.00402375403791666, + 0.008290556259453297, + 0.005920477211475372, + 0.0044536348432302475, + 0.0038181652780622244, + 0.011967362836003304, + 0.008602630347013474, + 0.019065309315919876, + 0.005433291662484407, + 0.00632633687928319, + 0.0074297357350587845, + 0.006212717387825251, + 0.004331791307777166, + 0.008492866531014442, + 0.011192835867404938, + 0.012028285302221775, + 0.01035392563790083, + 0.009168241173028946, + 0.007982025854289532, + 0.008595607243478298, + 0.006578115280717611, + 0.007031756918877363, + 0.009237443096935749, + 0.013256720267236233, + 0.006489756051450968, + 0.007765132002532482, + 0.01069069467484951, + 0.008208665996789932, + 0.006874373648315668, + 0.0030762418173253536, + 0.0032440489158034325, + 0.004455433692783117, + 0.003614465007558465, + 0.002861252287402749, + 0.0041813175193965435, + 0.0035244885366410017, + 0.002593907294794917, + 0.03262358158826828, + 0.030023440718650818, + 0.03086852841079235, + 0.02215898223221302, + 0.026471111923456192, + 0.03412511199712753, + 0.017025262117385864, + 0.02946898154914379, + 0.013634126633405685, + 0.014536630362272263, + 0.01379048079252243, + 0.009557018056511879, + 0.00783024076372385, + 0.014199179597198963, + 0.008314951322972775, + 0.016644742339849472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.0060349623672664165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008650153875351 + ], + "opposite_logprob_tokens": [ + -2.6950864791870117 + ], + "logprob_reward": -0.07008650153875351, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008650153875351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13560321927070618, + 0.0743732675909996, + 0.12694881856441498, + 0.4673081338405609, + 0.02122999168932438, + 0.03199586644768715, + 0.018174251541495323, + 0.015121298842132092, + 0.09189888834953308, + 0.01734629087150097 + ], + "credit_map": [ + 0.019563546404242516, + 0.01897077076137066, + 0.019113844260573387, + 0.01723271794617176, + 0.01756334863603115, + 0.015621603466570377, + 0.01480065193027258, + 0.01273674052208662, + 0.008070763200521469, + 0.007976495660841465, + 0.011374550871551037, + 0.007252872455865145, + 0.012137006968259811, + 0.012063146568834782, + 0.007914869114756584, + 0.007583563216030598, + 0.01657267101109028, + 0.012397444806993008, + 0.0223527941852808, + 0.012272090651094913, + 0.017880376428365707, + 0.01457772497087717, + 0.018363598734140396, + 0.012532105669379234, + 0.04365386813879013, + 0.0528632253408432, + 0.04580068588256836, + 0.051699597388505936, + 0.11299725621938705, + 0.06871950626373291, + 0.0519896075129509, + 0.03958440199494362, + 0.002614825963973999, + 0.0028423815965652466, + 0.003690899582579732, + 0.0016340524889528751, + 0.003607552032917738, + 0.0023356580641120672, + 0.0022524676751345396, + 0.0022521549835801125, + 0.0029976647347211838, + 0.004439427051693201, + 0.004836495034396648, + 0.003381516085937619, + 0.006407658103853464, + 0.003224892308935523, + 0.004280270077288151, + 0.002427940256893635, + 0.0017572487704455853, + 0.002064056694507599, + 0.002444912912324071, + 0.002232483820989728, + 0.0017903585685417056, + 0.0022994638420641422, + 0.003533502807840705, + 0.0020522240083664656, + 0.0014556266833096743, + 0.002161777578294277, + 0.0023479328956454992, + 0.002491684863343835, + 0.0017848454881459475, + 0.0020056043285876513, + 0.0016548839630559087, + 0.0012189423432573676, + 0.013321556150913239, + 0.012489259243011475, + 0.012850774452090263, + 0.008473852649331093, + 0.0100101875141263, + 0.014537937007844448, + 0.007000146433711052, + 0.013215175829827785, + 0.001651776721701026, + 0.003244488500058651, + 0.002257001120597124, + 0.0018311190651729703, + 0.0015077076386660337, + 0.0028087524697184563, + 0.0018187464447692037, + 0.002226698212325573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.07008650153875351 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710274755954742 + ], + "opposite_logprob_tokens": [ + -2.127102851867676 + ], + "logprob_reward": -0.12710274755954742, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12710274755954742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11465410888195038, + 0.065364308655262, + 0.0880441665649414, + 0.037374891340732574, + 0.03742087259888649, + 0.12930786609649658, + 0.053387824445962906, + 0.06073974072933197, + 0.3078615665435791, + 0.10584463179111481 + ], + "credit_map": [ + 0.01846471056342125, + 0.014824535697698593, + 0.020166544243693352, + 0.010624882765114307, + 0.018570803105831146, + 0.012209231033921242, + 0.009001927450299263, + 0.010791474021971226, + 0.010066174902021885, + 0.007583823520690203, + 0.010164421051740646, + 0.006040670443326235, + 0.00661950558423996, + 0.009480034932494164, + 0.006437934935092926, + 0.008971745148301125, + 0.01146440114825964, + 0.00825041439384222, + 0.012372612953186035, + 0.010320273227989674, + 0.012915284372866154, + 0.010701206512749195, + 0.0113531444221735, + 0.010666833259165287, + 0.00500479293987155, + 0.005548290908336639, + 0.0056154001504182816, + 0.005259291734546423, + 0.003821661928668618, + 0.003732480574399233, + 0.004419133998453617, + 0.003973838873207569, + 0.004906302317976952, + 0.005113185849040747, + 0.004965465981513262, + 0.004170769359916449, + 0.005205026362091303, + 0.004507666919380426, + 0.003908634651452303, + 0.004643821623176336, + 0.01270701177418232, + 0.013989160768687725, + 0.02165241912007332, + 0.023766128346323967, + 0.013773380778729916, + 0.010478480719029903, + 0.019037429243326187, + 0.013903860002756119, + 0.00479848263785243, + 0.0066027152352035046, + 0.00921536609530449, + 0.006139988079667091, + 0.005453959107398987, + 0.006280888337641954, + 0.009078449569642544, + 0.005817974917590618, + 0.0070077404379844666, + 0.006489349063485861, + 0.010446867905557156, + 0.009149066172540188, + 0.005276037845760584, + 0.007435985375195742, + 0.009383714757859707, + 0.005550979636609554, + 0.04876081272959709, + 0.04178733006119728, + 0.04642908647656441, + 0.024374376982450485, + 0.031733304262161255, + 0.04647969827055931, + 0.020670438185334206, + 0.04762652888894081, + 0.014622076414525509, + 0.018619097769260406, + 0.013717494904994965, + 0.01129328552633524, + 0.00861231330782175, + 0.011963234283030033, + 0.01000071782618761, + 0.017016403377056122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.12710274755954742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029811859130859375 + ], + "opposite_logprob_tokens": [ + -3.5298118591308594 + ], + "logprob_reward": -0.029811859130859375, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029811859130859375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15002281963825226, + 0.18818241357803345, + 0.042299531400203705, + 0.05604678764939308, + 0.03609805554151535, + 0.2179870456457138, + 0.029495609924197197, + 0.139939084649086, + 0.06666233390569687, + 0.07326629012823105 + ], + "credit_map": [ + 0.020967641845345497, + 0.01665330119431019, + 0.024507157504558563, + 0.02214083820581436, + 0.01933400146663189, + 0.01811544969677925, + 0.015343105420470238, + 0.012961328029632568, + 0.04348268359899521, + 0.010428551584482193, + 0.06981121003627777, + 0.00822473969310522, + 0.009762784466147423, + 0.01945687271654606, + 0.009244957938790321, + 0.017770646139979362, + 0.00516121182590723, + 0.003586078993976116, + 0.009202818386256695, + 0.004737635143101215, + 0.00633086496964097, + 0.005345222540199757, + 0.004190608859062195, + 0.00374509091489017, + 0.006623938679695129, + 0.006667864043265581, + 0.005948497448116541, + 0.005561929661780596, + 0.013524009846150875, + 0.007109190337359905, + 0.005979960784316063, + 0.004631404764950275, + 0.006330497097223997, + 0.005553852766752243, + 0.0040778061375021935, + 0.003831063862890005, + 0.004912511445581913, + 0.004374865908175707, + 0.003974654246121645, + 0.003042803844437003, + 0.019833529368042946, + 0.025992438197135925, + 0.03298775479197502, + 0.04902471974492073, + 0.025853659957647324, + 0.020052898675203323, + 0.024225201457738876, + 0.020016847178339958, + 0.0025994270108640194, + 0.0037302852142602205, + 0.0041205622255802155, + 0.0031582338269799948, + 0.0032012127339839935, + 0.0048112645745277405, + 0.0046884347684681416, + 0.003186189802363515, + 0.011677361093461514, + 0.016345296055078506, + 0.020944636315107346, + 0.03103676438331604, + 0.012986041605472565, + 0.014472638256847858, + 0.01945362612605095, + 0.013022730126976967, + 0.007032244931906462, + 0.005996850784868002, + 0.013698236085474491, + 0.006585472263395786, + 0.008997599594295025, + 0.007930724881589413, + 0.007036953698843718, + 0.009384257718920708, + 0.008865131065249443, + 0.010022555477917194, + 0.010161944665014744, + 0.007906951010227203, + 0.005789571907371283, + 0.008793103508651257, + 0.007884359918534756, + 0.013842684216797352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.029811859130859375 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560358047485352 + ], + "opposite_logprob_tokens": [ + -1.6006035804748535 + ], + "logprob_reward": -0.22560358047485352, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560358047485352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11299257725477219, + 0.04904177039861679, + 0.030957698822021484, + 0.019330840557813644, + 0.023511914536356926, + 0.08137086778879166, + 0.045823756605386734, + 0.026520565152168274, + 0.3585265278816223, + 0.25192350149154663 + ], + "credit_map": [ + 0.013747350312769413, + 0.01604301668703556, + 0.016384262591600418, + 0.013035930693149567, + 0.017070120200514793, + 0.012106840498745441, + 0.012669539079070091, + 0.011935519985854626, + 0.006588159129023552, + 0.005969796795397997, + 0.009316506795585155, + 0.004675493109971285, + 0.00452706404030323, + 0.007287061773240566, + 0.004708752501755953, + 0.005968938115984201, + 0.003960557747632265, + 0.0029890004079788923, + 0.003934480715543032, + 0.00428283866494894, + 0.00525136012583971, + 0.003619555616751313, + 0.004183965735137463, + 0.0027359400410205126, + 0.0035165874287486076, + 0.0017478886293247342, + 0.002159459749236703, + 0.0020932601764798164, + 0.002142756013199687, + 0.0018528110813349485, + 0.002950278576463461, + 0.0028678004164248705, + 0.0029959881212562323, + 0.002555769169703126, + 0.0037998869083821774, + 0.002617381513118744, + 0.003978555556386709, + 0.0021806620061397552, + 0.002935120603069663, + 0.0024485501926392317, + 0.007640580181032419, + 0.012552355416119099, + 0.011420316994190216, + 0.01404515840113163, + 0.00813195388764143, + 0.006795189343392849, + 0.011451623402535915, + 0.009333694353699684, + 0.004681598395109177, + 0.005109363701194525, + 0.006677689030766487, + 0.00620722770690918, + 0.005041901487857103, + 0.006732969079166651, + 0.006422549020498991, + 0.004950458183884621, + 0.004367236979305744, + 0.0038750749081373215, + 0.004107914865016937, + 0.0022758508566766977, + 0.002482438925653696, + 0.003524053143337369, + 0.0032255484256893396, + 0.0026624477468430996, + 0.051025453954935074, + 0.05376015603542328, + 0.05304735153913498, + 0.027025405317544937, + 0.03900963068008423, + 0.06124770641326904, + 0.02381804771721363, + 0.0495927669107914, + 0.03435283899307251, + 0.03935794159770012, + 0.033978693187236786, + 0.02044704556465149, + 0.018878595903515816, + 0.029454439878463745, + 0.022356156259775162, + 0.05309776961803436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.22560358047485352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, extreme close-up focuses squarely on an eye brimming with tears from intense emotion. The amber iris appears remarkably clear around the slightly dilated pupil, the white of the eye bloodshot with the fine lines of emotion. A soft, sideways windowlight casts a precise spotlight on the eye, revealing the reddened skin around it and the long, slightly curled black lashes. As the camera gazes, a pool of clear tears quickly and silently flows from the depths of the eye, instantly filling the edge of the lower lid, forming a plump, trembling surface that drenches the entire eye socket with a shimmering sheen. A few crystalline droplets of water have already formed, delicately dangling from the tips of the eyelashes, reflecting a brilliant diamond-like light. At this moment, the tears have reached the critical point of overflowing, as if any slight blink or tremor of breath will make this hot tear break free and slide down. The whole picture is full of suppressed sadness and extreme fragility, creating a heartbreaking tranquil atmosphere.", + "seed": 25944, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_3/index_3892_seed_25944_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10043508429080247, + "vlm_logprob": -0.10043508429080247, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.10043508429080247, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004402943886816502 + ], + "opposite_logprob_tokens": [ + -8.129403114318848 + ], + "logprob_reward": -0.004402943886816502, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.004402943886816502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25380927324295044, + 0.06263326108455658, + 0.192640021443367, + 0.11908232420682907, + 0.037793781608343124, + 0.06489204615354538, + 0.10925225913524628, + 0.07507007569074631, + 0.041114434599876404, + 0.04371250793337822 + ], + "credit_map": [ + 0.032039571553468704, + 0.01851898990571499, + 0.04852566123008728, + 0.07475414127111435, + 0.008337972685694695, + 0.02611052244901657, + 0.01966278813779354, + 0.025859639048576355, + 0.008203989826142788, + 0.011205000802874565, + 0.008957911282777786, + 0.008548783138394356, + 0.005288115702569485, + 0.005562538746744394, + 0.006171190645545721, + 0.008695732802152634, + 0.024194922298192978, + 0.012928023934364319, + 0.04160476103425026, + 0.05741497129201889, + 0.00799446739256382, + 0.015307697467505932, + 0.011931333690881729, + 0.02126384526491165, + 0.019156888127326965, + 0.021988965570926666, + 0.016843553632497787, + 0.016949405893683434, + 0.008783992379903793, + 0.010971161536872387, + 0.00991733092814684, + 0.01447103638201952, + 0.0045989155769348145, + 0.003369217738509178, + 0.0050113690085709095, + 0.006837083958089352, + 0.002622036961838603, + 0.004620612133294344, + 0.004745177924633026, + 0.005989368073642254, + 0.01574545167386532, + 0.009163393639028072, + 0.008884236216545105, + 0.006839401088654995, + 0.004972327966243029, + 0.006989110726863146, + 0.005101023707538843, + 0.007197098340839148, + 0.013167613185942173, + 0.003174087032675743, + 0.024340378120541573, + 0.01353449746966362, + 0.003308763261884451, + 0.036413393914699554, + 0.007034992333501577, + 0.008278538472950459, + 0.009806521236896515, + 0.004614868201315403, + 0.010464178398251534, + 0.00738940667361021, + 0.005098310764878988, + 0.009091530926525593, + 0.009673243388533592, + 0.01893201656639576, + 0.005360233597457409, + 0.0028298720717430115, + 0.00558549864217639, + 0.004711260553449392, + 0.0034312682691961527, + 0.008488565683364868, + 0.004651261027902365, + 0.006056474056094885, + 0.006807612720876932, + 0.0037973131984472275, + 0.005678597837686539, + 0.007864435203373432, + 0.002531949896365404, + 0.007993225939571857, + 0.004134536255151033, + 0.004904836881905794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.004402943886816502 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208447366952896 + ], + "opposite_logprob_tokens": [ + -2.812084436416626 + ], + "logprob_reward": -0.06208447366952896, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208447366952896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06582093238830566, + 0.05138712376356125, + 0.4552546739578247, + 0.014227502048015594, + 0.025051580742001534, + 0.01301487535238266, + 0.2275134027004242, + 0.10350377857685089, + 0.028494952246546745, + 0.01573115400969982 + ], + "credit_map": [ + 0.007982385344803333, + 0.006561386398971081, + 0.008550149388611317, + 0.013982257805764675, + 0.005877691321074963, + 0.006892254110425711, + 0.007589526474475861, + 0.008385284803807735, + 0.006717363838106394, + 0.003709818935021758, + 0.0050620753318071365, + 0.00888762529939413, + 0.004452728200703859, + 0.0066458857618272305, + 0.006564693059772253, + 0.009346933104097843, + 0.052426718175411224, + 0.021621091291308403, + 0.10423209518194199, + 0.1516764611005783, + 0.016840830445289612, + 0.03729685768485069, + 0.028158586472272873, + 0.043002042919397354, + 0.002817991655319929, + 0.0010978656355291605, + 0.0027713775634765625, + 0.0017246347852051258, + 0.0010479380143806338, + 0.0013278056867420673, + 0.0019198209047317505, + 0.0015200687339529395, + 0.002679060911759734, + 0.0023627711925655603, + 0.0036286446265876293, + 0.003456864971667528, + 0.001955032115802169, + 0.0028247570153325796, + 0.0039062779396772385, + 0.004238171502947807, + 0.001478973194025457, + 0.001700238324701786, + 0.001326466561295092, + 0.0014849244616925716, + 0.0017260137246921659, + 0.0012995228171348572, + 0.002317855367437005, + 0.0016808821819722652, + 0.02827414497733116, + 0.004317102022469044, + 0.05339871346950531, + 0.027864474803209305, + 0.005715194623917341, + 0.07720188051462173, + 0.015618706122040749, + 0.01512319315224886, + 0.015754107385873795, + 0.00800285954028368, + 0.011100977659225464, + 0.010473930276930332, + 0.011564815416932106, + 0.00917082093656063, + 0.018432095646858215, + 0.01900416985154152, + 0.004066424909979105, + 0.0010524297831580043, + 0.00634269742295146, + 0.0031537918839603662, + 0.0008393117459490895, + 0.00870624091476202, + 0.0024561998434364796, + 0.0018778556259348989, + 0.002143277321010828, + 0.001349296304397285, + 0.0025138938799500465, + 0.0021518694702535868, + 0.0010818901937454939, + 0.003174444427713752, + 0.00175033591222018, + 0.001566146151162684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.06208447366952896 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295917630195618 + ], + "opposite_logprob_tokens": [ + -2.237959146499634 + ], + "logprob_reward": -0.11295917630195618, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295917630195618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12062641978263855, + 0.0423460491001606, + 0.3412264585494995, + 0.020617622882127762, + 0.23230086266994476, + 0.00997650995850563, + 0.1265355944633484, + 0.04402981698513031, + 0.02090701274573803, + 0.04143364354968071 + ], + "credit_map": [ + 0.014136258512735367, + 0.006877630017697811, + 0.01737689971923828, + 0.03511609882116318, + 0.0052223531529307365, + 0.01365082710981369, + 0.012494360096752644, + 0.015751997008919716, + 0.006609468720853329, + 0.00792787130922079, + 0.004463064018636942, + 0.0052056037820875645, + 0.0034005220513790846, + 0.00463677104562521, + 0.003837919794023037, + 0.006264824420213699, + 0.040190890431404114, + 0.015886085107922554, + 0.07632553577423096, + 0.1171777993440628, + 0.008607596158981323, + 0.028856297954916954, + 0.0200343020260334, + 0.0341479554772377, + 0.0030682957731187344, + 0.0022726363968104124, + 0.003122984664514661, + 0.0018282464006915689, + 0.002379428129643202, + 0.0012141079641878605, + 0.0041115242056548595, + 0.002620398998260498, + 0.022361140698194504, + 0.018813828006386757, + 0.024797886610031128, + 0.03050023503601551, + 0.022289909422397614, + 0.023311926051974297, + 0.058823373168706894, + 0.03140256553888321, + 0.0015150465769693255, + 0.000908640562556684, + 0.0012850461062043905, + 0.0009784572757780552, + 0.0014496151125058532, + 0.0007768927607685328, + 0.002270016586408019, + 0.0007927955011837184, + 0.015723563730716705, + 0.002390749054029584, + 0.029501527547836304, + 0.015315783210098743, + 0.0033488452900201082, + 0.042738232761621475, + 0.008502072654664516, + 0.009014823473989964, + 0.004579281434416771, + 0.002930274698883295, + 0.004420652985572815, + 0.005571604706346989, + 0.0027014785446226597, + 0.0059141903184354305, + 0.004402716178447008, + 0.013509619981050491, + 0.0024334874469786882, + 0.001028519356623292, + 0.003536604344844818, + 0.0025240120012313128, + 0.0009981156326830387, + 0.005977689288556576, + 0.0017424537800252438, + 0.0026661292649805546, + 0.0059145535342395306, + 0.0017616847762838006, + 0.007654539309442043, + 0.005716851446777582, + 0.0016366811469197273, + 0.012305988930165768, + 0.002476196736097336, + 0.00396714685484767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.11295917630195618 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281347393989563 + ], + "opposite_logprob_tokens": [ + -2.0178134441375732 + ], + "logprob_reward": -0.14281347393989563, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281347393989563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07519961893558502, + 0.06892713904380798, + 0.4681670665740967, + 0.14239384233951569, + 0.08055155724287033, + 0.019117580726742744, + 0.012718821875751019, + 0.011887354776263237, + 0.037472158670425415, + 0.08356490731239319 + ], + "credit_map": [ + 0.011016258969902992, + 0.006770748645067215, + 0.014167898334562778, + 0.01862979494035244, + 0.003789264941588044, + 0.0068711475469172, + 0.005974290426820517, + 0.00798021163791418, + 0.007496845442801714, + 0.007283990271389484, + 0.012373487465083599, + 0.01970263198018074, + 0.0031494363211095333, + 0.009489321149885654, + 0.005008898675441742, + 0.0044225240126252174, + 0.05661187693476677, + 0.024515187367796898, + 0.10810493677854538, + 0.15770775079727173, + 0.01184264849871397, + 0.03829237446188927, + 0.0261247381567955, + 0.04496751353144646, + 0.024735283106565475, + 0.022814670577645302, + 0.01958983577787876, + 0.01945001631975174, + 0.010897085070610046, + 0.013799680396914482, + 0.015166242606937885, + 0.015941020101308823, + 0.015070978552103043, + 0.00776259507983923, + 0.014418860897421837, + 0.009955341927707195, + 0.00562837952747941, + 0.01040596328675747, + 0.008704555220901966, + 0.008604882284998894, + 0.005306699313223362, + 0.001676430692896247, + 0.004065115936100483, + 0.0019987428095191717, + 0.0012173278955742717, + 0.0018383287824690342, + 0.0015589221147820354, + 0.0014560100389644504, + 0.001981006935238838, + 0.0007362873293459415, + 0.0026145719457417727, + 0.0014861541567370296, + 0.0008472778717987239, + 0.0028467567171901464, + 0.0013081735232844949, + 0.0008985921740531921, + 0.0013189851306378841, + 0.0009052936220541596, + 0.001194300246424973, + 0.001324659795500338, + 0.001286240410991013, + 0.0014007778372615576, + 0.0018750893650576472, + 0.0025820068549364805, + 0.00558071956038475, + 0.002074746647849679, + 0.007828157395124435, + 0.004216941073536873, + 0.002223958494141698, + 0.008018165826797485, + 0.0034965798258781433, + 0.004032887984067202, + 0.012711072340607643, + 0.0033035692758858204, + 0.015431896783411503, + 0.01070639118552208, + 0.003766839625313878, + 0.023693539202213287, + 0.0058153062127530575, + 0.008136292919516563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.14281347393989563 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing squarely on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris amber and the pupil slightly dilated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the white of the eye bloodshot with visible fine lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a soft sideways windowlight casting a spotlight on the eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are clear tears flowing to fill the lower lid and forming droplets on the eyelashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991535365581512 + ], + "opposite_logprob_tokens": [ + -1.8049153089523315 + ], + "logprob_reward": -0.17991535365581512, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991535365581512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09346374869346619, + 0.07894779741764069, + 0.07475245743989944, + 0.061190612614154816, + 0.4714619815349579, + 0.009603451937437057, + 0.01652296632528305, + 0.11459921300411224, + 0.037466973066329956, + 0.04199080541729927 + ], + "credit_map": [ + 0.009838184341788292, + 0.007622454781085253, + 0.012777001596987247, + 0.032291021198034286, + 0.0033785738050937653, + 0.010104484856128693, + 0.006836405955255032, + 0.010615621693432331, + 0.011223625391721725, + 0.014019764959812164, + 0.015364950522780418, + 0.011684381403028965, + 0.007082388736307621, + 0.0070764306001365185, + 0.005511983297765255, + 0.006984279025346041, + 0.008292862214148045, + 0.003668705001473427, + 0.013142607174813747, + 0.022049162536859512, + 0.0033349143341183662, + 0.007274187169969082, + 0.008653509430587292, + 0.008336513303220272, + 0.008475991897284985, + 0.010382497683167458, + 0.008916952647268772, + 0.005241650156676769, + 0.0062461672350764275, + 0.004138713236898184, + 0.011468508280813694, + 0.00632012914866209, + 0.04490784928202629, + 0.03887425735592842, + 0.04651448503136635, + 0.058277685195207596, + 0.04824989289045334, + 0.04776202514767647, + 0.12074188888072968, + 0.06613389402627945, + 0.002466646023094654, + 0.0009398737456649542, + 0.001528381952084601, + 0.001012916094623506, + 0.0008550737984478474, + 0.0008397071505896747, + 0.0012197461910545826, + 0.000741107389330864, + 0.002010934054851532, + 0.0007591943140141666, + 0.0032159043475985527, + 0.001911766710691154, + 0.0006954440614208579, + 0.005060467403382063, + 0.0013617152580991387, + 0.0015075403498485684, + 0.010740404948592186, + 0.007426816504448652, + 0.011770926415920258, + 0.015052709728479385, + 0.006718400400131941, + 0.016458658501505852, + 0.009643170051276684, + 0.03678813576698303, + 0.004165779799222946, + 0.001584016252309084, + 0.0065552471205592155, + 0.004881660453975201, + 0.0013971565058454871, + 0.012489177286624908, + 0.0033518210984766483, + 0.003042116528376937, + 0.0048967511393129826, + 0.002827373566105962, + 0.008112427778542042, + 0.005595148541033268, + 0.0017280856845900416, + 0.011829220689833164, + 0.00336161139421165, + 0.003640185110270977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.17991535365581512 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_4/index_tensor([3910, 3910])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_4/index_tensor([3910, 3910])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ca963c22230efe6956f20f6478b8200cf79a4aef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_4/index_tensor([3910, 3910])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "vibrant, low-angle, macro time-lapse shot follows a vibrant, multicolored ground ivy runner as it relentlessly stretches across a patch of rich, dark earth. The camera maintains a tight, parallel tracking motion, creating a sense of urgent forward momentum. At each node along the slender, reddish-green stem, life bursts forth: tiny white rootlets burst forth, quickly anchoring themselves in the fertile soil, while a pair of new leaves unfurl in a swift, crisp pattern. Initially wrinkled and small, the leaves quickly expand to reveal a distinctive, scalloped, kidney-shaped form with a vibrant green center surrounded by a sharp, creamy white border. The entire scene is bathed in cheerful sunlight, with golden sun patterns and soft shadows flickering across the ground, creating a dance of light that highlights the vibrant growth. The rapid camera movement, causing a slight motion blur in the soft-focus background, locks the viewer's eye on the ivy's determined spread and captures the powerful, energizing atmosphere of nature's vitality and joyful reproduction.", + "seed": 30247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_4/index_3910_seed_30247_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.014476615050807596, + "vlm_logprob": -0.014476615050807596, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.014476615050807596, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043543316423892975 + ], + "opposite_logprob_tokens": [ + -3.1685433387756348 + ], + "logprob_reward": -0.043543316423892975, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043543316423892975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1529114544391632, + 0.2922011613845825, + 0.07405505329370499, + 0.06035156920552254, + 0.05522027984261513, + 0.02703113481402397, + 0.0673210620880127, + 0.052169930189847946, + 0.18681234121322632, + 0.03192606195807457 + ], + "credit_map": [ + 0.012600572779774666, + 0.03339521214365959, + 0.01166834868490696, + 0.01520660612732172, + 0.01645498536527157, + 0.024382704868912697, + 0.02280336245894432, + 0.016399655491113663, + 0.026416730135679245, + 0.05889289826154709, + 0.02160761132836342, + 0.0245924424380064, + 0.06036198139190674, + 0.05123738572001457, + 0.023531032726168633, + 0.02556106261909008, + 0.005898337811231613, + 0.00961881224066019, + 0.005684772040694952, + 0.009539933875203133, + 0.009736761450767517, + 0.008452722802758217, + 0.015109891071915627, + 0.010013826191425323, + 0.004909921437501907, + 0.0071900044567883015, + 0.00649650115519762, + 0.009227689355611801, + 0.006391823757439852, + 0.007277379743754864, + 0.00954566989094019, + 0.009312580339610577, + 0.005159703548997641, + 0.010847573168575764, + 0.004726305603981018, + 0.005829301197081804, + 0.00992821529507637, + 0.006700934376567602, + 0.006739670876413584, + 0.005288579035550356, + 0.0027387693990021944, + 0.0051806773990392685, + 0.002274109050631523, + 0.0026436159387230873, + 0.0034598102793097496, + 0.0029198364354670048, + 0.004914493300020695, + 0.0028998227789998055, + 0.005975807551294565, + 0.008784553967416286, + 0.0056281560100615025, + 0.008563023991882801, + 0.008120069280266762, + 0.010985887609422207, + 0.012232694774866104, + 0.007030872628092766, + 0.005021292250603437, + 0.006760436110198498, + 0.005953858606517315, + 0.00874057412147522, + 0.005062959156930447, + 0.008920728228986263, + 0.005842167418450117, + 0.005867912899702787, + 0.015385349281132221, + 0.022380132228136063, + 0.026166578754782677, + 0.0240736473351717, + 0.018396856263279915, + 0.020233364775776863, + 0.031278084963560104, + 0.02889833226799965, + 0.003244356019422412, + 0.006803545635193586, + 0.004657076206058264, + 0.002579048275947571, + 0.004288234747946262, + 0.003452000906690955, + 0.0030323676764965057, + 0.003869429463520646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.043543316423892975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070596873760223 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070596873760223, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070596873760223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1671183556318283, + 0.32581132650375366, + 0.04047967121005058, + 0.03814856708049774, + 0.06864656507968903, + 0.026453634724020958, + 0.03577880188822746, + 0.0693979486823082, + 0.15663719177246094, + 0.07152803987264633 + ], + "credit_map": [ + 0.013750426471233368, + 0.06134800612926483, + 0.012919451110064983, + 0.008060485124588013, + 0.037760455161333084, + 0.02039860002696514, + 0.006432895082980394, + 0.006448004394769669, + 0.03053743951022625, + 0.0686352550983429, + 0.024215349927544594, + 0.029376404359936714, + 0.06473836302757263, + 0.05615675821900368, + 0.02164486236870289, + 0.030506838113069534, + 0.004582654219120741, + 0.004585304297506809, + 0.004681709222495556, + 0.00434299698099494, + 0.004883596673607826, + 0.004601708613336086, + 0.007676359266042709, + 0.005125336814671755, + 0.004274322185665369, + 0.005818868055939674, + 0.003543074941262603, + 0.006098518613725901, + 0.004486529156565666, + 0.005453415215015411, + 0.004521140828728676, + 0.003952689468860626, + 0.007227453403174877, + 0.01006246916949749, + 0.007211839314550161, + 0.008276065811514854, + 0.01090629305690527, + 0.007916317321360111, + 0.01081703882664442, + 0.006229075137525797, + 0.002923680702224374, + 0.006174121517688036, + 0.001821741578169167, + 0.002292014891281724, + 0.003886542981490493, + 0.0030783223919570446, + 0.0035607856698334217, + 0.0027164211496710777, + 0.005669236183166504, + 0.004546502139419317, + 0.0031376564875245094, + 0.0033277233596891165, + 0.005061674863100052, + 0.00413843197748065, + 0.006102033890783787, + 0.0037955332081764936, + 0.007024317514151335, + 0.008552285842597485, + 0.006725287530571222, + 0.009982692077755928, + 0.006216003093868494, + 0.012384437024593353, + 0.010814007371664047, + 0.007698908448219299, + 0.01741606369614601, + 0.018170615658164024, + 0.020555058494210243, + 0.02170719765126705, + 0.016926633194088936, + 0.01636052317917347, + 0.02369602397084236, + 0.021805040538311005, + 0.006354538258165121, + 0.011851029470562935, + 0.0094774030148983, + 0.01137069147080183, + 0.005923143122345209, + 0.008433803915977478, + 0.01032331958413124, + 0.007794105913490057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.016070596873760223 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021998509764671326 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021998509764671326, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021998509764671326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2788463532924652, + 0.11964373290538788, + 0.04620619863271713, + 0.06695323437452316, + 0.07210393249988556, + 0.09035085141658783, + 0.06440474838018417, + 0.07741919159889221, + 0.08926741778850555, + 0.09480433166027069 + ], + "credit_map": [ + 0.018737556412816048, + 0.09992323815822601, + 0.01105431653559208, + 0.011837073601782322, + 0.07922136038541794, + 0.029024584218859673, + 0.016484268009662628, + 0.012563970871269703, + 0.012484888546168804, + 0.025317862629890442, + 0.010195162147283554, + 0.012576020322740078, + 0.016971725970506668, + 0.02076239511370659, + 0.011073160916566849, + 0.010262527503073215, + 0.0037902386393398046, + 0.007984088733792305, + 0.004741613753139973, + 0.006434694863855839, + 0.005233691073954105, + 0.006060249637812376, + 0.007625296711921692, + 0.004336327780038118, + 0.007339488249272108, + 0.011187493801116943, + 0.007565733511000872, + 0.008452503941953182, + 0.00777056347578764, + 0.01048202347010374, + 0.007570116780698299, + 0.006585317198187113, + 0.00913812592625618, + 0.010308908298611641, + 0.008128670044243336, + 0.008842364884912968, + 0.01041924674063921, + 0.00894598476588726, + 0.009142609313130379, + 0.007178031373769045, + 0.009359537623822689, + 0.012151503935456276, + 0.008078441955149174, + 0.007626332808285952, + 0.019066661596298218, + 0.010489843785762787, + 0.009625769220292568, + 0.013952762819826603, + 0.00757997902110219, + 0.007970219478011131, + 0.006644624751061201, + 0.006356080994009972, + 0.007772771641612053, + 0.010007728822529316, + 0.011153641156852245, + 0.006919702980667353, + 0.007247446570545435, + 0.00910358689725399, + 0.008162843063473701, + 0.01229145284742117, + 0.0071917143650352955, + 0.013116061687469482, + 0.011433125473558903, + 0.00887296348810196, + 0.008675978519022465, + 0.013307581655681133, + 0.009938462637364864, + 0.01237554382532835, + 0.01220853254199028, + 0.011942983604967594, + 0.012212762609124184, + 0.008605577982962132, + 0.007649378851056099, + 0.014651508070528507, + 0.013210337609052658, + 0.010512717999517918, + 0.010442501865327358, + 0.00812873151153326, + 0.012332397513091564, + 0.01787676475942135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0021998509764671326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019370617810636759 + ], + "opposite_logprob_tokens": [ + -6.251936912536621 + ], + "logprob_reward": -0.0019370617810636759, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019370617810636759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2928735017776489, + 0.377960741519928, + 0.06972435861825943, + 0.04048415645956993, + 0.06685138493776321, + 0.019588282331824303, + 0.028835302218794823, + 0.021778490394353867, + 0.04896317049860954, + 0.032940614968538284 + ], + "credit_map": [ + 0.017524095252156258, + 0.11392393708229065, + 0.011266578920185566, + 0.00897405855357647, + 0.09130703657865524, + 0.03055180050432682, + 0.007324149366468191, + 0.012001829221844673, + 0.03420259431004524, + 0.07554797828197479, + 0.02742297202348709, + 0.03482483699917793, + 0.07955602556467056, + 0.06607998162508011, + 0.02964954450726509, + 0.03067680448293686, + 0.006763939745724201, + 0.017465533688664436, + 0.005384198855608702, + 0.005886052269488573, + 0.011125712655484676, + 0.01116795465350151, + 0.00676045473664999, + 0.0051705120131373405, + 0.004445331171154976, + 0.008135060779750347, + 0.005044774617999792, + 0.004265971947461367, + 0.0046204738318920135, + 0.005194198340177536, + 0.003665988566353917, + 0.005112356971949339, + 0.00849224254488945, + 0.008420092053711414, + 0.007134526502341032, + 0.006820145063102245, + 0.01475172583013773, + 0.008603530935943127, + 0.004911705385893583, + 0.007717412430793047, + 0.0027545103803277016, + 0.0034336689859628677, + 0.0021179711911827326, + 0.002144320635125041, + 0.0025945636443793774, + 0.0022291038185358047, + 0.0025739804841578007, + 0.0017401635413989425, + 0.0031143943779170513, + 0.0036005235742777586, + 0.0025988726411014795, + 0.003083647694438696, + 0.003453205805271864, + 0.004815948195755482, + 0.004956531338393688, + 0.003212179522961378, + 0.0029369457624852657, + 0.0028435452841222286, + 0.0024489853531122208, + 0.0033988934010267258, + 0.0019399287411943078, + 0.0030179740861058235, + 0.0026536337099969387, + 0.0025385855697095394, + 0.003996881190687418, + 0.0051514627411961555, + 0.005814595613628626, + 0.006109647452831268, + 0.009795923717319965, + 0.0055293249897658825, + 0.005173502955585718, + 0.0073918309062719345, + 0.0029663110617548227, + 0.006010427605360746, + 0.004284295719116926, + 0.005077929701656103, + 0.003263882827013731, + 0.0037326563615351915, + 0.003865080187097192, + 0.003740029875189066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0019370617810636759 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863224919885397 + ], + "opposite_logprob_tokens": [ + -4.758632183074951 + ], + "logprob_reward": -0.00863224919885397, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863224919885397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15601299703121185, + 0.3620251417160034, + 0.09649374336004257, + 0.020541343837976456, + 0.08566360175609589, + 0.10598820447921753, + 0.10587925463914871, + 0.015525436028838158, + 0.036467984318733215, + 0.015402317978441715 + ], + "credit_map": [ + 0.010607370175421238, + 0.056431498378515244, + 0.009232097305357456, + 0.009570646099746227, + 0.05062482878565788, + 0.008122889325022697, + 0.007366294506937265, + 0.004057357553392649, + 0.034245479851961136, + 0.07700259238481522, + 0.02711680717766285, + 0.03508078679442406, + 0.06740431487560272, + 0.06264851987361908, + 0.026242464780807495, + 0.03228415548801422, + 0.00958260241895914, + 0.020101012662053108, + 0.0075323269702494144, + 0.008530556224286556, + 0.014146397821605206, + 0.016435107216238976, + 0.012129055336117744, + 0.008036674931645393, + 0.002181112766265869, + 0.003180626081302762, + 0.0017781801288947463, + 0.0027028003241866827, + 0.001998528139665723, + 0.0026470969896763563, + 0.0028948732651770115, + 0.0031581250950694084, + 0.01006020326167345, + 0.012631433084607124, + 0.00899245124310255, + 0.009017123840749264, + 0.017747407779097557, + 0.011740782298147678, + 0.006985556334257126, + 0.008488636463880539, + 0.011097138747572899, + 0.014971088618040085, + 0.009770638309419155, + 0.008383306674659252, + 0.02198200300335884, + 0.011306644417345524, + 0.010851550847291946, + 0.017625823616981506, + 0.011408624239265919, + 0.0128358518704772, + 0.00974072702229023, + 0.010824409313499928, + 0.012688894756138325, + 0.016920126974582672, + 0.020655039697885513, + 0.010805567726492882, + 0.0012451265938580036, + 0.0012965162750333548, + 0.0016464053187519312, + 0.00447248388081789, + 0.0014378497144207358, + 0.0015533709665760398, + 0.0017700281459838152, + 0.002103652572259307, + 0.0037643073592334986, + 0.00575672322884202, + 0.002993292175233364, + 0.002974509494379163, + 0.008417720906436443, + 0.004899728577584028, + 0.0043371813371777534, + 0.0033245214726775885, + 0.0012928227661177516, + 0.002062741667032242, + 0.002175625180825591, + 0.0015654257731512189, + 0.001406244351528585, + 0.001587285427376628, + 0.0021216601599007845, + 0.0031905106734484434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.00863224919885397 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "vibrant, low-angle, macro time-lapse shot follows a vibrant, multicolored ground ivy runner as it relentlessly stretches across a patch of rich, dark earth. The camera maintains a tight, parallel tracking motion, creating a sense of urgent forward momentum. At each node along the slender, reddish-green stem, life bursts forth: tiny white rootlets burst forth, quickly anchoring themselves in the fertile soil, while a pair of new leaves unfurl in a swift, crisp pattern. Initially wrinkled and small, the leaves quickly expand to reveal a distinctive, scalloped, kidney-shaped form with a vibrant green center surrounded by a sharp, creamy white border. The entire scene is bathed in cheerful sunlight, with golden sun patterns and soft shadows flickering across the ground, creating a dance of light that highlights the vibrant growth. The rapid camera movement, causing a slight motion blur in the soft-focus background, locks the viewer's eye on the ivy's determined spread and captures the powerful, energizing atmosphere of nature's vitality and joyful reproduction.", + "seed": 30248, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_4/index_3910_seed_30248_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012132173217833043, + "vlm_logprob": -0.012132173217833043, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.012132173217833043, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024009548127651215 + ], + "opposite_logprob_tokens": [ + -3.7740094661712646 + ], + "logprob_reward": -0.024009548127651215, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.024009548127651215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17078973352909088, + 0.11510321497917175, + 0.08284620940685272, + 0.13593988120555878, + 0.056098438799381256, + 0.06586014479398727, + 0.0851343423128128, + 0.1939215213060379, + 0.028759770095348358, + 0.06554669886827469 + ], + "credit_map": [ + 0.018463686108589172, + 0.020586177706718445, + 0.01849050261080265, + 0.02134367637336254, + 0.023647552356123924, + 0.015844952315092087, + 0.029895097017288208, + 0.022518087178468704, + 0.01188117079436779, + 0.013011273927986622, + 0.01351830456405878, + 0.011480119079351425, + 0.014543678611516953, + 0.014932757243514061, + 0.015036355704069138, + 0.02069956250488758, + 0.008275956846773624, + 0.013213016092777252, + 0.007761566434055567, + 0.007354011759161949, + 0.007481797598302364, + 0.006713039707392454, + 0.02056906931102276, + 0.01147774700075388, + 0.01149478554725647, + 0.018938647583127022, + 0.012199405580759048, + 0.012153634801506996, + 0.01908355951309204, + 0.01655055209994316, + 0.03127816691994667, + 0.014241127297282219, + 0.006249436177313328, + 0.00621228851377964, + 0.005332164000719786, + 0.005708625540137291, + 0.006717413663864136, + 0.006976025179028511, + 0.011292833834886551, + 0.007609651889652014, + 0.008957520127296448, + 0.006885218434035778, + 0.005974766798317432, + 0.013104881159961224, + 0.00775922741740942, + 0.007535852957516909, + 0.01196963805705309, + 0.0036730405408889055, + 0.009099539369344711, + 0.010367368347942829, + 0.012202935293316841, + 0.008539507165551186, + 0.014371488243341446, + 0.011828373186290264, + 0.009980278089642525, + 0.00874484796077013, + 0.01274140179157257, + 0.053091131150722504, + 0.015093867667019367, + 0.009926778264343739, + 0.014230630360543728, + 0.06536741554737091, + 0.01267675869166851, + 0.010793528519570827, + 0.00311133055947721, + 0.002605041489005089, + 0.003544384613633156, + 0.003447008552029729, + 0.00441405177116394, + 0.004283254966139793, + 0.005326556507498026, + 0.0020281414035707712, + 0.007573044393211603, + 0.011234003119170666, + 0.006655251141637564, + 0.006477830000221729, + 0.009919186122715473, + 0.009405197575688362, + 0.008492193184792995, + 0.005789996590465307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.024009548127651215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193509429693222 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193509429693222, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193509429693222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15471956133842468, + 0.13080796599388123, + 0.15290462970733643, + 0.03506151959300041, + 0.1283305585384369, + 0.033810075372457504, + 0.09192778915166855, + 0.09404751658439636, + 0.06769392639398575, + 0.11069640517234802 + ], + "credit_map": [ + 0.020416337996721268, + 0.019614873453974724, + 0.012501435354351997, + 0.016890406608581543, + 0.02625548653304577, + 0.015645233914256096, + 0.025995463132858276, + 0.0174003466963768, + 0.013005640357732773, + 0.015049409121274948, + 0.02408847212791443, + 0.012684903107583523, + 0.019243665039539337, + 0.01637270115315914, + 0.017363574355840683, + 0.012999625876545906, + 0.012804973870515823, + 0.026767529547214508, + 0.013865170069038868, + 0.020957907661795616, + 0.017041746526956558, + 0.017040206119418144, + 0.03008214384317398, + 0.014344966039061546, + 0.003054805565625429, + 0.0051432447507977486, + 0.004092266317456961, + 0.0036609438247978687, + 0.005729116499423981, + 0.005316514987498522, + 0.005193108692765236, + 0.002871520584449172, + 0.012474753893911839, + 0.016713490709662437, + 0.009341925382614136, + 0.014847864396870136, + 0.01585984043776989, + 0.01466720923781395, + 0.03332138434052467, + 0.011104118078947067, + 0.0050023579970002174, + 0.00512642553076148, + 0.0037012596148997545, + 0.002943192608654499, + 0.0053391424007713795, + 0.004889929201453924, + 0.004301904235035181, + 0.0025058675091713667, + 0.010678470134735107, + 0.012206478975713253, + 0.012733188457787037, + 0.007762117777019739, + 0.01724541187286377, + 0.013751250691711903, + 0.008708992972970009, + 0.008841892704367638, + 0.007415482774376869, + 0.01252839993685484, + 0.007853829301893711, + 0.009329931810498238, + 0.017476724460721016, + 0.015061174519360065, + 0.016790280118584633, + 0.00759170250967145, + 0.007665474433451891, + 0.01205955445766449, + 0.006787432357668877, + 0.006720237899571657, + 0.01216399110853672, + 0.0063293734565377235, + 0.01114503014832735, + 0.004822843242436647, + 0.011958971619606018, + 0.02443942241370678, + 0.013382360339164734, + 0.010929033160209656, + 0.016061225906014442, + 0.01478278636932373, + 0.010763594880700111, + 0.008379018865525723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.018193509429693222 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040984689258039 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.0040984689258039, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040984689258039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13067223131656647, + 0.08619073033332825, + 0.13407719135284424, + 0.07026199996471405, + 0.2234053611755371, + 0.03589954599738121, + 0.022542646154761314, + 0.07468269020318985, + 0.1032627522945404, + 0.11900480836629868 + ], + "credit_map": [ + 0.011995089240372181, + 0.01781880110502243, + 0.012868894264101982, + 0.013539673760533333, + 0.017603814601898193, + 0.016423411667346954, + 0.022635284811258316, + 0.017787283286452293, + 0.0072896662168204784, + 0.007688490208238363, + 0.009754822589457035, + 0.013028261251747608, + 0.00901587400585413, + 0.00818838831037283, + 0.019294435158371925, + 0.011930793523788452, + 0.012074691243469715, + 0.022316843271255493, + 0.01531359925866127, + 0.016564227640628815, + 0.012766818515956402, + 0.018281888216733932, + 0.023544590920209885, + 0.013214550912380219, + 0.0062780496664345264, + 0.009076944552361965, + 0.007333659566938877, + 0.007418772205710411, + 0.012461966834962368, + 0.010830450803041458, + 0.011661899276077747, + 0.005200264509767294, + 0.02311178296804428, + 0.0278021227568388, + 0.019661372527480125, + 0.02564201131463051, + 0.02428939938545227, + 0.024081753566861153, + 0.0446394719183445, + 0.034177448600530624, + 0.005428782198578119, + 0.005240816622972488, + 0.004264517221599817, + 0.0028619638178497553, + 0.005166024900972843, + 0.004907981958240271, + 0.00550991902127862, + 0.002519544679671526, + 0.0022392291575670242, + 0.0036460859701037407, + 0.0024604233913123608, + 0.003024229547008872, + 0.003150585340335965, + 0.0030880088452249765, + 0.0029349804390221834, + 0.001999106490984559, + 0.008601886220276356, + 0.012653881683945656, + 0.009337937459349632, + 0.0056407637894153595, + 0.009280389174818993, + 0.016084730625152588, + 0.004872830118983984, + 0.008210275322198868, + 0.007718642242252827, + 0.028313472867012024, + 0.00852249376475811, + 0.003756413934752345, + 0.006888847332447767, + 0.035898059606552124, + 0.005785696674138308, + 0.006379131227731705, + 0.013710426166653633, + 0.02093685418367386, + 0.012461382895708084, + 0.011812346056103706, + 0.019917519763112068, + 0.018567940220236778, + 0.01243643183261156, + 0.00916192028671503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0040984689258039 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625637110322714 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625637110322714, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625637110322714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21097314357757568, + 0.07079043239355087, + 0.046939387917518616, + 0.1546473503112793, + 0.032491426914930344, + 0.04271505028009415, + 0.03219861164689064, + 0.33148473501205444, + 0.02450130507349968, + 0.053258560597896576 + ], + "credit_map": [ + 0.026337329298257828, + 0.047309018671512604, + 0.015225323848426342, + 0.017660455778241158, + 0.032899994403123856, + 0.035533882677555084, + 0.018121585249900818, + 0.01788555644452572, + 0.007640038151293993, + 0.009636439383029938, + 0.006118677090853453, + 0.008270918391644955, + 0.011516740545630455, + 0.00903253722935915, + 0.009233391843736172, + 0.009341686964035034, + 0.005069094244390726, + 0.005951636005192995, + 0.004637156147509813, + 0.004286005161702633, + 0.00937834195792675, + 0.006042837165296078, + 0.005079649388790131, + 0.006494669709354639, + 0.015431318432092667, + 0.02269582636654377, + 0.012731008231639862, + 0.014264632016420364, + 0.03562585636973381, + 0.024728702381253242, + 0.011362641118466854, + 0.0178073700517416, + 0.004166950471699238, + 0.0034078173339366913, + 0.003280112287029624, + 0.003130257362499833, + 0.00395397562533617, + 0.004132265690714121, + 0.005392525810748339, + 0.005027524195611477, + 0.00803090538829565, + 0.00580856017768383, + 0.0034039237070828676, + 0.0031882156617939472, + 0.007192725781351328, + 0.007652613800019026, + 0.004105614963918924, + 0.003332492196932435, + 0.003432008670642972, + 0.004292693920433521, + 0.004048327915370464, + 0.0026460157241672277, + 0.006689619738608599, + 0.005588325206190348, + 0.0024122532922774553, + 0.0030893643852323294, + 0.022212455049157143, + 0.10724714398384094, + 0.024853641167283058, + 0.008983319625258446, + 0.016559187322854996, + 0.13093972206115723, + 0.007561421487480402, + 0.01312782522290945, + 0.00210137153044343, + 0.003390838857740164, + 0.002908344380557537, + 0.0024766295682638884, + 0.0022931734565645456, + 0.006723614409565926, + 0.0029038647189736366, + 0.0017034695483744144, + 0.005100327078253031, + 0.006376141682267189, + 0.007387503050267696, + 0.0062829540111124516, + 0.00795774906873703, + 0.008910291828215122, + 0.006974734831601381, + 0.004268861375749111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.007625637110322714 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067337024956941605 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.0067337024956941605, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067337024956941605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08973918855190277, + 0.04198096692562103, + 0.03504738211631775, + 0.06630723178386688, + 0.02931738644838333, + 0.031663741916418076, + 0.11136920750141144, + 0.4659962058067322, + 0.09638717770576477, + 0.032191481441259384 + ], + "credit_map": [ + 0.014453986659646034, + 0.01064736396074295, + 0.010410836897790432, + 0.01101497933268547, + 0.011686161160469055, + 0.008327736519277096, + 0.01257473323494196, + 0.010623394511640072, + 0.005685513839125633, + 0.004228285048156977, + 0.005627533886581659, + 0.004976088181138039, + 0.005713700316846371, + 0.004236872773617506, + 0.006004841532558203, + 0.005508133675903082, + 0.004496919922530651, + 0.004067877773195505, + 0.003709463169798255, + 0.0028901654295623302, + 0.005695851985365152, + 0.005612675566226244, + 0.004308352712541819, + 0.004266079049557447, + 0.007000096142292023, + 0.008607558906078339, + 0.005792172625660896, + 0.006478716153651476, + 0.011869640089571476, + 0.01038411259651184, + 0.00677065784111619, + 0.00940428301692009, + 0.003714691149070859, + 0.003929744474589825, + 0.003200241830199957, + 0.0031347002368420362, + 0.003020056989043951, + 0.004544947296380997, + 0.004352600779384375, + 0.0034204055555164814, + 0.006306319497525692, + 0.004569897893816233, + 0.003418324049562216, + 0.002736346097663045, + 0.0045988960191607475, + 0.004827483557164669, + 0.0030241874046623707, + 0.002182289259508252, + 0.011847350746393204, + 0.01582840271294117, + 0.013209673576056957, + 0.014099215157330036, + 0.015093354508280754, + 0.014853897504508495, + 0.015506519004702568, + 0.010930804535746574, + 0.04203678295016289, + 0.11366735398769379, + 0.035648543387651443, + 0.026871126145124435, + 0.04840452969074249, + 0.13221478462219238, + 0.042430128902196884, + 0.02472298964858055, + 0.008207738399505615, + 0.021952617913484573, + 0.007546583656221628, + 0.009272179566323757, + 0.010496485978364944, + 0.021393142640590668, + 0.012372986413538456, + 0.005145449656993151, + 0.0037418559659272432, + 0.005139156710356474, + 0.005004350561648607, + 0.003873597364872694, + 0.0031475131399929523, + 0.002901762956753373, + 0.0036232119891792536, + 0.004760029260069132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.0067337024956941605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "vibrant, low-angle, macro time-lapse shot follows a vibrant, multicolored ground ivy runner as it relentlessly stretches across a patch of rich, dark earth. The camera maintains a tight, parallel tracking motion, creating a sense of urgent forward momentum. At each node along the slender, reddish-green stem, life bursts forth: tiny white rootlets burst forth, quickly anchoring themselves in the fertile soil, while a pair of new leaves unfurl in a swift, crisp pattern. Initially wrinkled and small, the leaves quickly expand to reveal a distinctive, scalloped, kidney-shaped form with a vibrant green center surrounded by a sharp, creamy white border. The entire scene is bathed in cheerful sunlight, with golden sun patterns and soft shadows flickering across the ground, creating a dance of light that highlights the vibrant growth. The rapid camera movement, causing a slight motion blur in the soft-focus background, locks the viewer's eye on the ivy's determined spread and captures the powerful, energizing atmosphere of nature's vitality and joyful reproduction.", + "seed": 30249, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_4/index_3910_seed_30249_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011115738376975059, + "vlm_logprob": -0.011115738376975059, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.011115738376975059, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0305193942040205 + ], + "opposite_logprob_tokens": [ + -3.530519485473633 + ], + "logprob_reward": -0.0305193942040205, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0305193942040205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17316552996635437, + 0.07645183801651001, + 0.07364558428525925, + 0.05424864962697029, + 0.10606668889522552, + 0.05393538624048233, + 0.022254740819334984, + 0.06302114576101303, + 0.290057897567749, + 0.08715257793664932 + ], + "credit_map": [ + 0.018553631380200386, + 0.022020643576979637, + 0.019692888483405113, + 0.025018392130732536, + 0.0174923837184906, + 0.018598532304167747, + 0.02899087592959404, + 0.022798188030719757, + 0.007758450694382191, + 0.012670565396547318, + 0.009035846218466759, + 0.009420414455235004, + 0.008850730955600739, + 0.008568303659558296, + 0.012899217195808887, + 0.0072483038529753685, + 0.009671464562416077, + 0.007807337213307619, + 0.007788151036947966, + 0.008380082435905933, + 0.00670141214504838, + 0.006389344576746225, + 0.01696518436074257, + 0.009942608885467052, + 0.008070241659879684, + 0.00486108660697937, + 0.0073755350895226, + 0.006605937611311674, + 0.004392577335238457, + 0.006122852675616741, + 0.005048741120845079, + 0.011771676130592823, + 0.026539873331785202, + 0.014011727645993233, + 0.010511023923754692, + 0.0071317958645522594, + 0.007017059717327356, + 0.00924835354089737, + 0.00999536830931902, + 0.021611491218209267, + 0.015290601179003716, + 0.00441382359713316, + 0.003220411716029048, + 0.002756984904408455, + 0.014547550119459629, + 0.005392750725150108, + 0.003186090849339962, + 0.00512717617675662, + 0.0027122017927467823, + 0.0029104314744472504, + 0.0028869814705103636, + 0.0027634939178824425, + 0.002504367148503661, + 0.002722887322306633, + 0.003136568935588002, + 0.0026178103871643543, + 0.0045426953583955765, + 0.00638102600350976, + 0.006582702975720167, + 0.005684361793100834, + 0.013414288870990276, + 0.012272694148123264, + 0.007328554056584835, + 0.006814820226281881, + 0.014818087220191956, + 0.04821952059864998, + 0.04191473871469498, + 0.01675986684858799, + 0.09594704210758209, + 0.0458875335752964, + 0.012205277569591999, + 0.014305825345218182, + 0.006303803995251656, + 0.012327242642641068, + 0.014124298468232155, + 0.01584426686167717, + 0.009147029370069504, + 0.008082601241767406, + 0.011868232861161232, + 0.009455102495849133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0305193942040205 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080164462327957 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080164462327957, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080164462327957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3027225434780121, + 0.09924677759408951, + 0.0717882588505745, + 0.045945923775434494, + 0.10710982978343964, + 0.06606153398752213, + 0.040557682514190674, + 0.1650649607181549, + 0.0564906932413578, + 0.04501179978251457 + ], + "credit_map": [ + 0.03740902990102768, + 0.03410318121314049, + 0.04728133976459503, + 0.04197922348976135, + 0.02999434992671013, + 0.027761083096265793, + 0.04490324482321739, + 0.03929109871387482, + 0.012003108859062195, + 0.016754280775785446, + 0.01135335210710764, + 0.010498974472284317, + 0.01139293983578682, + 0.014056462794542313, + 0.0134946433827281, + 0.009693016298115253, + 0.008080271072685719, + 0.007616284769028425, + 0.010860399343073368, + 0.008477439172565937, + 0.00787773635238409, + 0.007833350449800491, + 0.01162673532962799, + 0.009416037239134312, + 0.006894527934491634, + 0.005696867126971483, + 0.005090620834380388, + 0.003971025813370943, + 0.005842359736561775, + 0.004462838638573885, + 0.007157177664339542, + 0.006830504629760981, + 0.024522041901946068, + 0.013371233828365803, + 0.011316518299281597, + 0.006607582792639732, + 0.008471149951219559, + 0.009834377095103264, + 0.009904498234391212, + 0.023082425817847252, + 0.020019229501485825, + 0.005632925312966108, + 0.004494227934628725, + 0.005489986389875412, + 0.014418688602745533, + 0.0071654669009149075, + 0.0041446201503276825, + 0.004696385003626347, + 0.003612738335505128, + 0.004285614006221294, + 0.0033107115887105465, + 0.0066002835519611835, + 0.004894259851425886, + 0.006262661423534155, + 0.0041810027323663235, + 0.007410409860312939, + 0.009657572023570538, + 0.009317326359450817, + 0.007745304610580206, + 0.0046889749355614185, + 0.052480340003967285, + 0.06647954136133194, + 0.005571149289608002, + 0.009124755859375, + 0.0044432030990719795, + 0.01230336632579565, + 0.005036224611103535, + 0.0047941142693161964, + 0.0050435746088624, + 0.013002291321754456, + 0.00521501712501049, + 0.0066529023461043835, + 0.003473998047411442, + 0.003383943112567067, + 0.006317103281617165, + 0.008287818171083927, + 0.004148649983108044, + 0.004913583397865295, + 0.006590092089027166, + 0.007896608673036098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.011080164462327957 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024915861431509256 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024915861431509256, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024915861431509256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2948518991470337, + 0.08724217116832733, + 0.18471764028072357, + 0.05456290766596794, + 0.06800547987222672, + 0.02970314770936966, + 0.060013480484485626, + 0.08608859032392502, + 0.05610039830207825, + 0.07871434092521667 + ], + "credit_map": [ + 0.03896886855363846, + 0.04189629852771759, + 0.03827502951025963, + 0.03131547570228577, + 0.04541714861989021, + 0.03197207674384117, + 0.03160564601421356, + 0.035401351749897, + 0.012153593823313713, + 0.017851365730166435, + 0.010613011196255684, + 0.008278297260403633, + 0.01230091042816639, + 0.010570655576884747, + 0.006924029905349016, + 0.008550305850803852, + 0.0422559417784214, + 0.023019937798380852, + 0.015300768427550793, + 0.01531741302460432, + 0.015289690345525742, + 0.02403089962899685, + 0.017540879547595978, + 0.03196209669113159, + 0.008011477068066597, + 0.007716648746281862, + 0.006582342553883791, + 0.004375615157186985, + 0.0082930913195014, + 0.006819761358201504, + 0.006097453646361828, + 0.006666519213467836, + 0.014182076789438725, + 0.008696481585502625, + 0.007993677631020546, + 0.005338443908840418, + 0.0055472650565207005, + 0.0061924755573272705, + 0.006789575796574354, + 0.013265478424727917, + 0.0034855578560382128, + 0.002525745891034603, + 0.003797972807660699, + 0.004551628138870001, + 0.0026048896834254265, + 0.0038040357176214457, + 0.004514839965850115, + 0.004418475553393364, + 0.006894492544233799, + 0.005168471951037645, + 0.006620452739298344, + 0.01050371490418911, + 0.009225699119269848, + 0.0076393140479922295, + 0.005544448737055063, + 0.008416884578764439, + 0.005183666944503784, + 0.0054316152818500996, + 0.005146353971213102, + 0.0029331257101148367, + 0.025842228904366493, + 0.031122231855988503, + 0.004248202778398991, + 0.006181156728416681, + 0.003917159978300333, + 0.014993229880928993, + 0.005156293511390686, + 0.004595376085489988, + 0.012180089950561523, + 0.007294077891856432, + 0.003531060414388776, + 0.004433105234056711, + 0.006911964155733585, + 0.006925538182258606, + 0.010858549736440182, + 0.018303681164979935, + 0.006963695865124464, + 0.008982306346297264, + 0.009536133147776127, + 0.010232469066977501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0024915861431509256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017113338690251112 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.0017113338690251112, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017113338690251112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17313219606876373, + 0.05927196890115738, + 0.07935991138219833, + 0.0840071439743042, + 0.09371203929185867, + 0.11698617041110992, + 0.04183616116642952, + 0.2301269769668579, + 0.04905284568667412, + 0.07251458615064621 + ], + "credit_map": [ + 0.024232562631368637, + 0.022097213193774223, + 0.02145315706729889, + 0.01889009401202202, + 0.0317443385720253, + 0.022116785869002342, + 0.015752941370010376, + 0.0168450940400362, + 0.007340339943766594, + 0.007990174926817417, + 0.008065358735620975, + 0.005381606053560972, + 0.005523061845451593, + 0.008681959472596645, + 0.00710940919816494, + 0.009180057793855667, + 0.009972833096981049, + 0.017447257414460182, + 0.007448320277035236, + 0.009157299064099789, + 0.008533638902008533, + 0.008984762243926525, + 0.009636741131544113, + 0.00817906204611063, + 0.010473378002643585, + 0.02164529822766781, + 0.00872879009693861, + 0.006598163396120071, + 0.01076177041977644, + 0.011575503274798393, + 0.006161908619105816, + 0.008062328211963177, + 0.01299956627190113, + 0.024862511083483696, + 0.008705598302185535, + 0.008674408309161663, + 0.010118190199136734, + 0.010397857055068016, + 0.010271551087498665, + 0.007682363502681255, + 0.028996653854846954, + 0.010823017917573452, + 0.006812723353505135, + 0.010900584980845451, + 0.033926382660865784, + 0.01215091161429882, + 0.007626642938703299, + 0.005749247502535582, + 0.005861459765583277, + 0.004082330968230963, + 0.004734519403427839, + 0.00467331800609827, + 0.0063417935743927956, + 0.004483196418732405, + 0.004928359761834145, + 0.006731186993420124, + 0.013815761543810368, + 0.014274450950324535, + 0.011034907773137093, + 0.006245896220207214, + 0.07333039492368698, + 0.09387248754501343, + 0.0080635417252779, + 0.009489529766142368, + 0.004262261092662811, + 0.007344652898609638, + 0.0028591463342309, + 0.003921946976333857, + 0.009557545185089111, + 0.014223837293684483, + 0.0038662836886942387, + 0.0030171761754900217, + 0.006026033312082291, + 0.008473565801978111, + 0.01406540535390377, + 0.01288253627717495, + 0.008434764109551907, + 0.007488436531275511, + 0.008311908692121506, + 0.006831927224993706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.0017113338690251112 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776213206350803 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009776213206350803, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776213206350803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12568779289722443, + 0.08141425997018814, + 0.23914490640163422, + 0.11325262486934662, + 0.10398919880390167, + 0.0753074437379837, + 0.029574047774076462, + 0.030917104333639145, + 0.13883335888385773, + 0.06187920272350311 + ], + "credit_map": [ + 0.011231099255383015, + 0.01844821870326996, + 0.011506755836308002, + 0.014902106486260891, + 0.024981141090393066, + 0.01674339361488819, + 0.017824163660407066, + 0.01005090493708849, + 0.009512470103800297, + 0.010495727881789207, + 0.00945836491882801, + 0.012628316879272461, + 0.010743018239736557, + 0.01117048878222704, + 0.008712495677173138, + 0.00869338121265173, + 0.07298008352518082, + 0.03318047896027565, + 0.018915506079792976, + 0.014529799111187458, + 0.015880022197961807, + 0.02256118133664131, + 0.01709958165884018, + 0.043998248875141144, + 0.01462170947343111, + 0.014242664910852909, + 0.013542496599256992, + 0.009274064563214779, + 0.011260537430644035, + 0.013003061525523663, + 0.015628380700945854, + 0.021679719910025597, + 0.021499022841453552, + 0.011977850459516048, + 0.010624997317790985, + 0.007108661811798811, + 0.007334935013204813, + 0.011170987971127033, + 0.010660284198820591, + 0.023612450808286667, + 0.010803834535181522, + 0.006486887112259865, + 0.007017620373517275, + 0.0117006441578269, + 0.016319183632731438, + 0.009412098675966263, + 0.007381799165159464, + 0.006185383070260286, + 0.00285150739364326, + 0.005028434097766876, + 0.0035889495629817247, + 0.003945762757211924, + 0.005334225948899984, + 0.0032791297417134047, + 0.0020971058402210474, + 0.003448933595791459, + 0.0024740835651755333, + 0.001973790815100074, + 0.0023395242169499397, + 0.002006706316024065, + 0.006558526773005724, + 0.009436048567295074, + 0.0022418221924453974, + 0.0038866016548126936, + 0.00869867391884327, + 0.021566936746239662, + 0.022902097553014755, + 0.009747019968926907, + 0.03993576392531395, + 0.02279837429523468, + 0.006856048014014959, + 0.006328444927930832, + 0.005574864335358143, + 0.009823487140238285, + 0.007882537320256233, + 0.008997401222586632, + 0.0064174216240644455, + 0.006326718255877495, + 0.010239949449896812, + 0.006616821978241205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.009776213206350803 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "vibrant, low-angle, macro time-lapse shot follows a vibrant, multicolored ground ivy runner as it relentlessly stretches across a patch of rich, dark earth. The camera maintains a tight, parallel tracking motion, creating a sense of urgent forward momentum. At each node along the slender, reddish-green stem, life bursts forth: tiny white rootlets burst forth, quickly anchoring themselves in the fertile soil, while a pair of new leaves unfurl in a swift, crisp pattern. Initially wrinkled and small, the leaves quickly expand to reveal a distinctive, scalloped, kidney-shaped form with a vibrant green center surrounded by a sharp, creamy white border. The entire scene is bathed in cheerful sunlight, with golden sun patterns and soft shadows flickering across the ground, creating a dance of light that highlights the vibrant growth. The rapid camera movement, causing a slight motion blur in the soft-focus background, locks the viewer's eye on the ivy's determined spread and captures the powerful, energizing atmosphere of nature's vitality and joyful reproduction.", + "seed": 30250, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_4/index_3910_seed_30250_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02641098517924547, + "vlm_logprob": -0.02641098517924547, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02641098517924547, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06321626156568527 + ], + "opposite_logprob_tokens": [ + -2.813216209411621 + ], + "logprob_reward": -0.06321626156568527, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06321626156568527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10207220166921616, + 0.08083371073007584, + 0.030084004625678062, + 0.05187397077679634, + 0.09500787407159805, + 0.32534360885620117, + 0.03052660822868347, + 0.09822990745306015, + 0.09854677319526672, + 0.08748122304677963 + ], + "credit_map": [ + 0.013191373087465763, + 0.010117142461240292, + 0.010338446125388145, + 0.014641966670751572, + 0.016060324385762215, + 0.013851369731128216, + 0.012032751925289631, + 0.011838839389383793, + 0.01070259977132082, + 0.007199116051197052, + 0.023153124377131462, + 0.008426958695054054, + 0.006795508787035942, + 0.007969438098371029, + 0.008278720080852509, + 0.008308256976306438, + 0.004873776808381081, + 0.0030828106682747602, + 0.0033365434501320124, + 0.0037474322598427534, + 0.003666487755253911, + 0.0034986354876309633, + 0.004235445521771908, + 0.003642877098172903, + 0.00670006824657321, + 0.005694694351404905, + 0.004643079359084368, + 0.005860081873834133, + 0.006328161805868149, + 0.007569453213363886, + 0.009671534411609173, + 0.005406902637332678, + 0.02480410225689411, + 0.007422908209264278, + 0.0046736570075154305, + 0.005803105421364307, + 0.010581322945654392, + 0.026284758001565933, + 0.009109330363571644, + 0.006328696850687265, + 0.08071213960647583, + 0.023690572008490562, + 0.04945506155490875, + 0.023795193061232567, + 0.02660052292048931, + 0.07437419146299362, + 0.022615676745772362, + 0.024100286886096, + 0.004116259980946779, + 0.0033481575082987547, + 0.0034073542337864637, + 0.0036606951616704464, + 0.004892838653177023, + 0.0047432538121938705, + 0.0026660473085939884, + 0.003692005295306444, + 0.028738748282194138, + 0.007062752265483141, + 0.005866075865924358, + 0.005983649753034115, + 0.013235995545983315, + 0.026109738275408745, + 0.005987995304167271, + 0.005244965199381113, + 0.012282509356737137, + 0.014078275300562382, + 0.013401963748037815, + 0.011519129388034344, + 0.018165860325098038, + 0.012098473496735096, + 0.007506736554205418, + 0.00949383620172739, + 0.012257426977157593, + 0.017220349982380867, + 0.008946592919528484, + 0.007266178727149963, + 0.01449537742882967, + 0.0126168979331851, + 0.00810837373137474, + 0.006570028141140938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.06321626156568527 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233067087829113 + ], + "opposite_logprob_tokens": [ + -3.7733066082000732 + ], + "logprob_reward": -0.0233067087829113, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233067087829113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19403335452079773, + 0.09895926713943481, + 0.03766641020774841, + 0.053486477583646774, + 0.11033932119607925, + 0.3335982859134674, + 0.03218989446759224, + 0.02361121214926243, + 0.0617685467004776, + 0.05434723198413849 + ], + "credit_map": [ + 0.03186428174376488, + 0.019755465909838676, + 0.019535811617970467, + 0.017137326300144196, + 0.04335232824087143, + 0.02865227870643139, + 0.018331561237573624, + 0.01540428027510643, + 0.022005757316946983, + 0.006518424954265356, + 0.01644693873822689, + 0.005123614799231291, + 0.008695753291249275, + 0.025320325046777725, + 0.00945979543030262, + 0.005388644989579916, + 0.004814591258764267, + 0.004930576309561729, + 0.005990335252135992, + 0.004697680473327637, + 0.0038619665428996086, + 0.003839854383841157, + 0.0060552312061190605, + 0.0034761696588248014, + 0.006544713396579027, + 0.007047224324196577, + 0.009815934114158154, + 0.005862673278898001, + 0.0050432696007192135, + 0.007547129411250353, + 0.004480541218072176, + 0.0071449875831604, + 0.030252370983362198, + 0.007121965289115906, + 0.011445037089288235, + 0.005280339624732733, + 0.01113200094550848, + 0.032166723161935806, + 0.007694566622376442, + 0.005246309097856283, + 0.08461139351129532, + 0.022491440176963806, + 0.053542301058769226, + 0.024604598060250282, + 0.02375265769660473, + 0.07781782001256943, + 0.021359413862228394, + 0.02541867084801197, + 0.003932909108698368, + 0.004083206877112389, + 0.005278583616018295, + 0.0028810652438551188, + 0.0065142386592924595, + 0.0037411123048514128, + 0.0025527290999889374, + 0.0032060472294688225, + 0.0029409967828541994, + 0.0023745347280055285, + 0.0023086387664079666, + 0.0040338896214962006, + 0.003694614628329873, + 0.0036057811230421066, + 0.00252049183472991, + 0.002132263733074069, + 0.008116679266095161, + 0.009115368127822876, + 0.011461240239441395, + 0.0054641785100102425, + 0.006582923699170351, + 0.006916369777172804, + 0.006016327068209648, + 0.00809545163065195, + 0.007248448207974434, + 0.007958579808473587, + 0.006558131892234087, + 0.005220379680395126, + 0.008124485611915588, + 0.007790066301822662, + 0.0052788835018873215, + 0.006168249063193798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0233067087829113 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031989850103855133 + ], + "opposite_logprob_tokens": [ + -5.753199100494385 + ], + "logprob_reward": -0.0031989850103855133, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031989850103855133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10674668848514557, + 0.05838466435670853, + 0.04002262279391289, + 0.04671609029173851, + 0.14862117171287537, + 0.2442554384469986, + 0.047799963504076004, + 0.0704539343714714, + 0.09306085854768753, + 0.14393852651119232 + ], + "credit_map": [ + 0.015171675942838192, + 0.014913284219801426, + 0.010795697569847107, + 0.011923081241548061, + 0.01914139650762081, + 0.013656546361744404, + 0.013042258098721504, + 0.008102749474346638, + 0.008791094645857811, + 0.004948625806719065, + 0.01139472983777523, + 0.007958680391311646, + 0.0046396092511713505, + 0.006468154955655336, + 0.00926312617957592, + 0.004920651204884052, + 0.006112979259341955, + 0.0034915676806122065, + 0.006777547299861908, + 0.005236288532614708, + 0.0033637541346251965, + 0.00399235961958766, + 0.005206595174968243, + 0.0058415355160832405, + 0.006861219648271799, + 0.005642026197165251, + 0.005775332450866699, + 0.005911349784582853, + 0.00430290075019002, + 0.0052361078560352325, + 0.005702379625290632, + 0.00728477630764246, + 0.03775077685713768, + 0.01004137471318245, + 0.02164890617132187, + 0.00813665334135294, + 0.013824285939335823, + 0.04020886868238449, + 0.009707520715892315, + 0.007302796933799982, + 0.05633047968149185, + 0.01729007437825203, + 0.042174581438302994, + 0.01978646032512188, + 0.019087722525000572, + 0.052320562303066254, + 0.017819100990891457, + 0.01944645866751671, + 0.004665489308536053, + 0.0046026380732655525, + 0.008778976276516914, + 0.006988544017076492, + 0.007404833100736141, + 0.0053834873251616955, + 0.004498329944908619, + 0.005477669183164835, + 0.014388597570359707, + 0.007392485626041889, + 0.004418898373842239, + 0.0066486806608736515, + 0.014242253266274929, + 0.01270716916769743, + 0.006084905005991459, + 0.004570947028696537, + 0.011348908767104149, + 0.015189649537205696, + 0.01034670788794756, + 0.010461894795298576, + 0.013248455710709095, + 0.014631615951657295, + 0.009113774634897709, + 0.00871986337006092, + 0.018372148275375366, + 0.02471662312746048, + 0.01968012936413288, + 0.012019378133118153, + 0.019067339599132538, + 0.01811482198536396, + 0.014456725679337978, + 0.017511360347270966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.0031989850103855133 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798205941915512 + ], + "opposite_logprob_tokens": [ + -3.5297982692718506 + ], + "logprob_reward": -0.029798205941915512, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798205941915512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2856947183609009, + 0.09585172683000565, + 0.040520358830690384, + 0.04678942263126373, + 0.09219818562269211, + 0.17205698788166046, + 0.011846266686916351, + 0.0281375702470541, + 0.06166756898164749, + 0.16523727774620056 + ], + "credit_map": [ + 0.027401285246014595, + 0.04145050793886185, + 0.030738962814211845, + 0.05764821916818619, + 0.04660206288099289, + 0.026577940210700035, + 0.028096487745642662, + 0.027179203927516937, + 0.011873573996126652, + 0.008836962282657623, + 0.029283365234732628, + 0.009241562336683273, + 0.007880999706685543, + 0.011714410968124866, + 0.007526267319917679, + 0.009494578465819359, + 0.005047977436333895, + 0.005632518790662289, + 0.004693727008998394, + 0.007405013311654329, + 0.0042121228761971, + 0.004452965687960386, + 0.004760740790516138, + 0.004315290600061417, + 0.005873861722648144, + 0.005331788212060928, + 0.00985509343445301, + 0.006355649325996637, + 0.0030309874564409256, + 0.005235126707702875, + 0.0038627777248620987, + 0.007244132459163666, + 0.02666916511952877, + 0.006177201867103577, + 0.008064025081694126, + 0.003827499458566308, + 0.0101062236353755, + 0.02957383170723915, + 0.003378709312528372, + 0.004401514306664467, + 0.0323711521923542, + 0.012494378723204136, + 0.04469694569706917, + 0.014220074750483036, + 0.012507865205407143, + 0.026307255029678345, + 0.011372217908501625, + 0.01808708906173706, + 0.0014980575069785118, + 0.0010557490168139338, + 0.0023128895554691553, + 0.0008569007623009384, + 0.002166909398511052, + 0.001370391808450222, + 0.0009650138090364635, + 0.001620353665202856, + 0.0067665427923202515, + 0.0022843494080007076, + 0.002862956840544939, + 0.0020015714690089226, + 0.0033419227693229914, + 0.006742344703525305, + 0.0021204769145697355, + 0.0020174034871160984, + 0.0072356597520411015, + 0.011669653467833996, + 0.007562706712633371, + 0.009034444577991962, + 0.00880987849086523, + 0.0078270910307765, + 0.0047668875195086, + 0.004761243239045143, + 0.018717514351010323, + 0.030428491532802582, + 0.022401275113224983, + 0.01269533857703209, + 0.024578317999839783, + 0.02189822494983673, + 0.015559405088424683, + 0.018958693370223045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.029798205941915512 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant, multicolored ground ivy runner with a reddish-green stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do tiny white rootlets burst forth from the nodes to anchor in the rich, dark earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do new leaves unfurl rapidly to reveal a scalloped, kidney-shaped form with a vibrant green center and creamy white border?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in cheerful sunlight with golden sun patterns and soft shadows flickering across the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle macro time-lapse with a tight, parallel tracking motion causing slight motion blur in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534764595329762 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534764595329762, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534764595329762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23584406077861786, + 0.05977723002433777, + 0.05371341481804848, + 0.049963127821683884, + 0.07869619876146317, + 0.1912912279367447, + 0.032452963292598724, + 0.034356988966464996, + 0.09942904859781265, + 0.1644757091999054 + ], + "credit_map": [ + 0.02819047123193741, + 0.028943585231900215, + 0.022732747718691826, + 0.04256536811590195, + 0.0434710867702961, + 0.024291042238473892, + 0.021204093471169472, + 0.024445677176117897, + 0.011629479937255383, + 0.005004578270018101, + 0.01054447889328003, + 0.004135601222515106, + 0.004774992819875479, + 0.01345387939363718, + 0.005673621315509081, + 0.004560601897537708, + 0.0053117177449166775, + 0.005801782943308353, + 0.006061183288693428, + 0.006857444532215595, + 0.00717537896707654, + 0.006318195722997189, + 0.009165230207145214, + 0.007022488862276077, + 0.005162956193089485, + 0.0055252788588404655, + 0.004789395723491907, + 0.004520684015005827, + 0.0064467983320355415, + 0.0096313850954175, + 0.008784709498286247, + 0.0051019261591136456, + 0.009704669006168842, + 0.006786001846194267, + 0.019312141463160515, + 0.007298915181308985, + 0.009705167263746262, + 0.01026942953467369, + 0.009629481472074986, + 0.00599040137603879, + 0.029323481023311615, + 0.015662919729948044, + 0.05335791036486626, + 0.018917569890618324, + 0.015383253805339336, + 0.02011653035879135, + 0.012512832880020142, + 0.02601674571633339, + 0.004307688679546118, + 0.0035699696745723486, + 0.005848641507327557, + 0.0033836583606898785, + 0.005150815937668085, + 0.004632988013327122, + 0.002227850491181016, + 0.0033313550520688295, + 0.004205865319818258, + 0.0064119515009224415, + 0.0029158054385334253, + 0.004753713496029377, + 0.00558296637609601, + 0.004614659119397402, + 0.003236662596464157, + 0.0026353667490184307, + 0.012423563748598099, + 0.02076711691915989, + 0.013585486449301243, + 0.013715337961912155, + 0.010228926315903664, + 0.009632369503378868, + 0.006948526483029127, + 0.012127736583352089, + 0.0189723651856184, + 0.02501472644507885, + 0.026744849979877472, + 0.01271011307835579, + 0.021304698660969734, + 0.021373625844717026, + 0.01673727296292782, + 0.021618083119392395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.012534764595329762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_5/index_tensor([2572, 2572])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_5/index_tensor([2572, 2572])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7a7ff833fb446076022d1ab1a78aa4dc165ef6be --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_5/index_tensor([2572, 2572])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "handheld camera, in extreme close-up, cautiously advances, focusing on a pink peony blossoming in the soft morning sunlight. Perched at the center of the flower is a vibrant butterfly. Its velvety wings are covered in an intricate pattern of iridescent blue and fiery orange, edged with delicate black lines. A few glistening dewdrops dapple the petals and wings, shimmering with light. As the camera steadily and slowly approaches, the slight shakiness of the frame conveys a breathless tension. As if sensing an intruder, the butterfly's slender antennae tense. Then, without warning, it explodes, its wings powerfully flapping downward, launching it vertically upward at astonishing speed. The sudden gust of air lifts the fragile peony petals from their pistil, creating a shower of pink petals that swirls and falls gracefully in the air, shimmering with a golden halo. The camera instinctively jerks upward, trying to follow the butterfly, but ultimately captures it as a blurry, colorful dot of light disappearing into the bokeh-like, verdant garden background. The entire scene switches from an instant of extreme tranquility to a brief and brilliant commotion, filled with the vitality of nature and its fleeting beauty.", + "seed": 13745, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_5/index_2572_seed_13745_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024536303989589215, + "vlm_logprob": -0.024536303989589215, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.024536303989589215, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006013988982886076 + ], + "opposite_logprob_tokens": [ + -5.881013870239258 + ], + "logprob_reward": -0.006013988982886076, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006013988982886076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052081406116485596, + 0.017653778195381165, + 0.029766004532575607, + 0.6883629560470581, + 0.016169274225831032, + 0.04846834018826485, + 0.02989819087088108, + 0.02223532274365425, + 0.04768901318311691, + 0.04767569899559021 + ], + "credit_map": [ + 0.013880615122616291, + 0.0033424217253923416, + 0.011569460853934288, + 0.0030086992774158716, + 0.004603106994181871, + 0.0063778311014175415, + 0.003463883651420474, + 0.00583539018407464, + 0.002841293578967452, + 0.0014471168396994472, + 0.002298847772181034, + 0.0019116391194984317, + 0.0025763416197150946, + 0.0022006055805832148, + 0.0022536066826432943, + 0.002124327700585127, + 0.007997309789061546, + 0.0018491040682420135, + 0.007476888597011566, + 0.002925213659182191, + 0.0015963055193424225, + 0.002069025067612529, + 0.002329709008336067, + 0.0035224477760493755, + 0.2770802080631256, + 0.013243344612419605, + 0.2924272418022156, + 0.02055981010198593, + 0.010893810540437698, + 0.03972828760743141, + 0.018030842766165733, + 0.016399413347244263, + 0.0026989176403731108, + 0.002526156837120652, + 0.0024193432182073593, + 0.0014833002351224422, + 0.0020848670974373817, + 0.001375463674776256, + 0.00199179002083838, + 0.0015894350362941623, + 0.009716235101222992, + 0.009758859872817993, + 0.008533203974366188, + 0.0026749498210847378, + 0.004607096314430237, + 0.006263202056288719, + 0.003171849064528942, + 0.003742943750694394, + 0.00308864563703537, + 0.002595576224848628, + 0.003951336722820997, + 0.003342739073559642, + 0.003758038394153118, + 0.0036802038084715605, + 0.005987750831991434, + 0.0034938983153551817, + 0.0019616882782429457, + 0.002583991037681699, + 0.003459044499322772, + 0.0041973483748734, + 0.0020439240615814924, + 0.0030612184200435877, + 0.003134089056402445, + 0.001794019597582519, + 0.004312924575060606, + 0.008358356542885303, + 0.006337853614240885, + 0.005079516675323248, + 0.0060135009698569775, + 0.005561398342251778, + 0.008336411789059639, + 0.00368904834613204, + 0.005829420872032642, + 0.005067023914307356, + 0.007427515462040901, + 0.00810028612613678, + 0.00471525127068162, + 0.006656598765403032, + 0.005554674658924341, + 0.004324933048337698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.006013988982886076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797343984246254 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797343984246254, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797343984246254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049097929149866104, + 0.02200750820338726, + 0.05331011489033699, + 0.4757423996925354, + 0.0487036369740963, + 0.04572875797748566, + 0.12088944017887115, + 0.07723124325275421, + 0.04179081320762634, + 0.06549815088510513 + ], + "credit_map": [ + 0.010320701636373997, + 0.0031374809332191944, + 0.009116939269006252, + 0.004189024213701487, + 0.003961112350225449, + 0.006599401589483023, + 0.004166723694652319, + 0.007606544531881809, + 0.005289122462272644, + 0.001174513017758727, + 0.00502381194382906, + 0.0011942132841795683, + 0.002704457612708211, + 0.00249210256151855, + 0.0019724734593182802, + 0.0021568138618022203, + 0.02014746703207493, + 0.0018123857444152236, + 0.018426036462187767, + 0.0022773398086428642, + 0.0016095406608656049, + 0.0031429778318852186, + 0.002473460277542472, + 0.0034209091681987047, + 0.20767128467559814, + 0.010243791155517101, + 0.19896408915519714, + 0.014536499977111816, + 0.008062023669481277, + 0.013161108829081059, + 0.011633113026618958, + 0.011470449157059193, + 0.008342912420630455, + 0.005651998333632946, + 0.007136930711567402, + 0.004349880386143923, + 0.0057861595414578915, + 0.004908942151814699, + 0.005844222381711006, + 0.006682595703750849, + 0.007875079289078712, + 0.010029131546616554, + 0.008728361688554287, + 0.0028932890854775906, + 0.0037541284691542387, + 0.006270868703722954, + 0.0030395518988370895, + 0.0031383477617055178, + 0.01980695314705372, + 0.02537795715034008, + 0.01386590301990509, + 0.011121833696961403, + 0.016403494402766228, + 0.015932684764266014, + 0.009622953832149506, + 0.008757658302783966, + 0.006083356682211161, + 0.01089363545179367, + 0.010639875195920467, + 0.014560382813215256, + 0.010440154932439327, + 0.009525235742330551, + 0.010268178768455982, + 0.004820418078452349, + 0.003641389310359955, + 0.007561627775430679, + 0.008749887347221375, + 0.0037377476692199707, + 0.004775514360517263, + 0.00800950638949871, + 0.0027223837096244097, + 0.0025927559472620487, + 0.004686249885708094, + 0.01490743923932314, + 0.011803841218352318, + 0.007622068282216787, + 0.006745531223714352, + 0.008085393346846104, + 0.006960394326597452, + 0.004687231499701738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.009797343984246254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062046490609645844 + ], + "opposite_logprob_tokens": [ + -2.812046527862549 + ], + "logprob_reward": -0.062046490609645844, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062046490609645844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06159406527876854, + 0.02831956557929516, + 0.06670939177274704, + 0.3722648620605469, + 0.03714299947023392, + 0.04511024430394173, + 0.04079625383019447, + 0.05585043877363205, + 0.16512030363082886, + 0.12709185481071472 + ], + "credit_map": [ + 0.013801615685224533, + 0.004059347324073315, + 0.01283866073936224, + 0.004448076244443655, + 0.006486235652118921, + 0.008541010320186615, + 0.005037785042077303, + 0.0063813356682658195, + 0.004138509277254343, + 0.0018966374918818474, + 0.00440070778131485, + 0.0026109665632247925, + 0.004241690505295992, + 0.0025429800152778625, + 0.0036696824245154858, + 0.004818391520529985, + 0.027345405891537666, + 0.0017048290465027094, + 0.025737179443240166, + 0.0019487313693389297, + 0.0017932254122570157, + 0.0031315258238464594, + 0.0021998409647494555, + 0.0028486589435487986, + 0.15858900547027588, + 0.007763063535094261, + 0.15434123575687408, + 0.011517707258462906, + 0.007823983207345009, + 0.012078744359314442, + 0.009862666949629784, + 0.010288459248840809, + 0.006727568339556456, + 0.006411236245185137, + 0.005806998815387487, + 0.0036056709941476583, + 0.004686415195465088, + 0.0037299280520528555, + 0.003467944683507085, + 0.0027072390075773, + 0.008323856629431248, + 0.01090775616466999, + 0.009111864492297173, + 0.0024670457933098078, + 0.0036718512419611216, + 0.005039883311837912, + 0.002623107749968767, + 0.0029648817144334316, + 0.006108257919549942, + 0.007616082206368446, + 0.00511730695143342, + 0.0032403310760855675, + 0.005049325991421938, + 0.0048540011048316956, + 0.0052177621982991695, + 0.0035931896418333054, + 0.005524401552975178, + 0.007807194720953703, + 0.007088486570864916, + 0.005905120633542538, + 0.006585015915334225, + 0.010376403108239174, + 0.007069970015436411, + 0.005493844859302044, + 0.01178977731615305, + 0.017279157415032387, + 0.04828328266739845, + 0.012443662621080875, + 0.015635604038834572, + 0.04293355345726013, + 0.007634550333023071, + 0.009120718576014042, + 0.009792990051209927, + 0.029120953753590584, + 0.023609085008502007, + 0.013153725303709507, + 0.015388661995530128, + 0.013833018951117992, + 0.013471283949911594, + 0.008722133003175259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.062046490609645844 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038088809698820114 + ], + "opposite_logprob_tokens": [ + -3.288088798522949 + ], + "logprob_reward": -0.038088809698820114, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038088809698820114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0549997016787529, + 0.01875808648765087, + 0.04718950018286705, + 0.5425710678100586, + 0.04940081387758255, + 0.02878623828291893, + 0.10292378813028336, + 0.02722274325788021, + 0.05148134380578995, + 0.0766666978597641 + ], + "credit_map": [ + 0.012037593871355057, + 0.0032526955474168062, + 0.009228617884218693, + 0.0035128858871757984, + 0.004173559136688709, + 0.007155835162848234, + 0.003958309069275856, + 0.01168020348995924, + 0.003147448878735304, + 0.0014453008770942688, + 0.0030684766825288534, + 0.0013847312657162547, + 0.0022643539123237133, + 0.0026273878756910563, + 0.0018651719437912107, + 0.002955216681584716, + 0.01190309226512909, + 0.002756709000095725, + 0.009813852608203888, + 0.0030833701603114605, + 0.003752587828785181, + 0.006406551226973534, + 0.0034015497658401728, + 0.0060717882588505745, + 0.2230433076620102, + 0.01059104036539793, + 0.2288992553949356, + 0.015959620475769043, + 0.008473082445561886, + 0.02695455774664879, + 0.015105195343494415, + 0.01354499626904726, + 0.00631500780582428, + 0.007984179072082043, + 0.007380445953458548, + 0.003692593425512314, + 0.006072854157537222, + 0.007059411611407995, + 0.004429761786013842, + 0.006466560997068882, + 0.004285487812012434, + 0.007500757463276386, + 0.004940124694257975, + 0.002038943348452449, + 0.0029101986438035965, + 0.003201685380190611, + 0.0018961593741551042, + 0.002012883545830846, + 0.012755578383803368, + 0.018469257280230522, + 0.012681874446570873, + 0.008770249783992767, + 0.012292398139834404, + 0.016567401587963104, + 0.01386183500289917, + 0.007525195367634296, + 0.003811596892774105, + 0.003220995655283332, + 0.0038097549695521593, + 0.0029228334315121174, + 0.002491807099431753, + 0.004555081017315388, + 0.0038034310564398766, + 0.0026072433684021235, + 0.004475988447666168, + 0.009600264951586723, + 0.006571850273758173, + 0.005906519014388323, + 0.00533735565841198, + 0.009611696936190128, + 0.00602144468575716, + 0.003956223838031292, + 0.006282889284193516, + 0.012577054090797901, + 0.010693098418414593, + 0.011462453752756119, + 0.0099539365619421, + 0.011389309540390968, + 0.007230052258819342, + 0.007077902089804411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.038088809698820114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734886672347784 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734886672347784, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734886672347784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030725806951522827, + 0.013182347640395164, + 0.06069938838481903, + 0.30062633752822876, + 0.03266633674502373, + 0.032844822853803635, + 0.06817294657230377, + 0.04563646391034126, + 0.1405215561389923, + 0.27492398023605347 + ], + "credit_map": [ + 0.004327360540628433, + 0.002374502597376704, + 0.00460917130112648, + 0.003973935730755329, + 0.0033995199482887983, + 0.0032486715354025364, + 0.00435089785605669, + 0.004441746510565281, + 0.002758025424554944, + 0.000860063883010298, + 0.0022498772013932467, + 0.0010362803004682064, + 0.001933440798893571, + 0.0017081928672268987, + 0.0013389470987021923, + 0.0012975193094462156, + 0.023669444024562836, + 0.0017840771470218897, + 0.022474801167845726, + 0.002323942957445979, + 0.0018521102610975504, + 0.003670876147225499, + 0.0024069850333034992, + 0.0025171474553644657, + 0.12915962934494019, + 0.007186638191342354, + 0.11898624151945114, + 0.009332641959190369, + 0.007518936414271593, + 0.010670131072402, + 0.00896796490997076, + 0.008804171346127987, + 0.004916282836347818, + 0.005429744720458984, + 0.0051326327957212925, + 0.0028517544269561768, + 0.003988230135291815, + 0.0036934928502887487, + 0.0033352335449308157, + 0.0033189670648425817, + 0.004506720695644617, + 0.006682678125798702, + 0.005635214503854513, + 0.0024300934746861458, + 0.002899744315072894, + 0.006043008528649807, + 0.001998688792809844, + 0.0026486727874726057, + 0.007456362247467041, + 0.010777214542031288, + 0.008234238252043724, + 0.006562885828316212, + 0.006723305676132441, + 0.013205358758568764, + 0.008591756224632263, + 0.006621825974434614, + 0.004414557013660669, + 0.004527740180492401, + 0.006332050077617168, + 0.004561807960271835, + 0.00586438924074173, + 0.007165699731558561, + 0.0058251009322702885, + 0.006945121102035046, + 0.00927656702697277, + 0.0078534334897995, + 0.04802776500582695, + 0.01317183393985033, + 0.006835130508989096, + 0.03847133740782738, + 0.008887306787073612, + 0.00799819640815258, + 0.027387529611587524, + 0.03887017071247101, + 0.0516735315322876, + 0.03966187313199043, + 0.026544637978076935, + 0.03437885642051697, + 0.031559377908706665, + 0.024848001077771187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.006734886672347784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld camera, in extreme close-up, cautiously advances, focusing on a pink peony blossoming in the soft morning sunlight. Perched at the center of the flower is a vibrant butterfly. Its velvety wings are covered in an intricate pattern of iridescent blue and fiery orange, edged with delicate black lines. A few glistening dewdrops dapple the petals and wings, shimmering with light. As the camera steadily and slowly approaches, the slight shakiness of the frame conveys a breathless tension. As if sensing an intruder, the butterfly's slender antennae tense. Then, without warning, it explodes, its wings powerfully flapping downward, launching it vertically upward at astonishing speed. The sudden gust of air lifts the fragile peony petals from their pistil, creating a shower of pink petals that swirls and falls gracefully in the air, shimmering with a golden halo. The camera instinctively jerks upward, trying to follow the butterfly, but ultimately captures it as a blurry, colorful dot of light disappearing into the bokeh-like, verdant garden background. The entire scene switches from an instant of extreme tranquility to a brief and brilliant commotion, filled with the vitality of nature and its fleeting beauty.", + "seed": 13746, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_5/index_2572_seed_13746_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03012958951294422, + "vlm_logprob": -0.03012958951294422, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.03012958951294422, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005305614788085222 + ], + "opposite_logprob_tokens": [ + -5.880305767059326 + ], + "logprob_reward": -0.005305614788085222, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.005305614788085222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07641403377056122, + 0.07071215659379959, + 0.05054059624671936, + 0.046389542520046234, + 0.06689022481441498, + 0.0852404311299324, + 0.1925228387117386, + 0.08902135491371155, + 0.14304693043231964, + 0.17922183871269226 + ], + "credit_map": [ + 0.008269373327493668, + 0.015556052327156067, + 0.01545732282102108, + 0.006808750331401825, + 0.007670471910387278, + 0.006274438928812742, + 0.008714188821613789, + 0.007663443684577942, + 0.007952403277158737, + 0.007096827030181885, + 0.011171654798090458, + 0.008024493232369423, + 0.015128560364246368, + 0.006530534941703081, + 0.008643385022878647, + 0.006164303980767727, + 0.007212325930595398, + 0.0039533390663564205, + 0.010202636010944843, + 0.004069171845912933, + 0.008168411441147327, + 0.00502278795465827, + 0.007336003240197897, + 0.0045759291388094425, + 0.008123110048472881, + 0.004420254845172167, + 0.008878827095031738, + 0.004494939465075731, + 0.0035868242848664522, + 0.004732379224151373, + 0.005198227241635323, + 0.006954987999051809, + 0.010440998710691929, + 0.007011339534074068, + 0.011003791354596615, + 0.007284163031727076, + 0.010129764676094055, + 0.007032403722405434, + 0.008691440336406231, + 0.0052963304333388805, + 0.00890666339546442, + 0.010546991601586342, + 0.011585407890379429, + 0.009629380889236927, + 0.01762535236775875, + 0.008872629143297672, + 0.010188995860517025, + 0.00788501650094986, + 0.04767293110489845, + 0.007516920566558838, + 0.06633539497852325, + 0.010031254030764103, + 0.01100044697523117, + 0.02893548086285591, + 0.00947690661996603, + 0.011553524062037468, + 0.014346906915307045, + 0.006425280589610338, + 0.017120394855737686, + 0.008563496172428131, + 0.007358154281973839, + 0.01385030709207058, + 0.009052875451743603, + 0.012303952127695084, + 0.016665326431393623, + 0.01457527931779623, + 0.025829879567027092, + 0.02704966627061367, + 0.014958066865801811, + 0.014557593502104282, + 0.016315847635269165, + 0.013095296919345856, + 0.01766216941177845, + 0.019930696114897728, + 0.02884301170706749, + 0.01676676608622074, + 0.01632760651409626, + 0.02654932625591755, + 0.04029851779341698, + 0.012843756936490536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.005305614788085222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008652813732624054 + ], + "opposite_logprob_tokens": [ + -4.758652687072754 + ], + "logprob_reward": -0.008652813732624054, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652813732624054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07728540897369385, + 0.02519679069519043, + 0.06898637861013412, + 0.021285075694322586, + 0.03542144596576691, + 0.08961945027112961, + 0.10075962543487549, + 0.1805488020181656, + 0.14473678171634674, + 0.2561602294445038 + ], + "credit_map": [ + 0.01436420064419508, + 0.013065366074442863, + 0.008886244148015976, + 0.00888024177402258, + 0.007516409270465374, + 0.007136527448892593, + 0.00823861826211214, + 0.009197806008160114, + 0.004856451880186796, + 0.003592622932046652, + 0.0035676630213856697, + 0.0022908446844667196, + 0.003290003165602684, + 0.002942824037745595, + 0.002649162197485566, + 0.002007221570238471, + 0.009061138145625591, + 0.006221837364137173, + 0.01083345152437687, + 0.005371611565351486, + 0.011965575627982616, + 0.007282919716089964, + 0.01058860495686531, + 0.007661244831979275, + 0.0035712458193302155, + 0.0018241438083350658, + 0.0033002588897943497, + 0.0021780438255518675, + 0.00196624081581831, + 0.0021104218903928995, + 0.0026106350123882294, + 0.0037240865640342236, + 0.005405813455581665, + 0.00385196041315794, + 0.004692807327955961, + 0.003586018458008766, + 0.005532182287424803, + 0.0035718316212296486, + 0.005127531010657549, + 0.0036533025559037924, + 0.011114460416138172, + 0.011852196417748928, + 0.012592540122568607, + 0.011022035032510757, + 0.016941465437412262, + 0.006163852754980326, + 0.011485445313155651, + 0.00844745896756649, + 0.022100701928138733, + 0.00630598422139883, + 0.029040899127721786, + 0.006850309669971466, + 0.007091661915183067, + 0.01418232824653387, + 0.006606820039451122, + 0.008580927737057209, + 0.03139529004693031, + 0.013159786351025105, + 0.039656080305576324, + 0.0178548414260149, + 0.012712454423308372, + 0.026387814432382584, + 0.014300079084932804, + 0.025082480162382126, + 0.017927205190062523, + 0.016777021810412407, + 0.028971606865525246, + 0.018651526421308517, + 0.016475575044751167, + 0.015125659294426441, + 0.014860706403851509, + 0.01594749093055725, + 0.0315634086728096, + 0.014303223229944706, + 0.040165890008211136, + 0.020237931981682777, + 0.02622159942984581, + 0.020307956263422966, + 0.09114976227283478, + 0.012210475280880928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.008652813732624054 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001859694719315 + ], + "opposite_logprob_tokens": [ + -2.6950185298919678 + ], + "logprob_reward": -0.07001859694719315, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001859694719315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08074434846639633, + 0.03070003353059292, + 0.08907350897789001, + 0.02364674210548401, + 0.013657154515385628, + 0.061911460012197495, + 0.0596739761531353, + 0.06453628838062286, + 0.12035354226827621, + 0.4557029604911804 + ], + "credit_map": [ + 0.013655616901814938, + 0.015200837515294552, + 0.011987948790192604, + 0.00651514483615756, + 0.007265404798090458, + 0.008174064569175243, + 0.006600681692361832, + 0.01134464517235756, + 0.00535489059984684, + 0.003362667514011264, + 0.0057134185917675495, + 0.002206195378676057, + 0.0036425467114895582, + 0.004644895438104868, + 0.002717962022870779, + 0.003057455411180854, + 0.012173344381153584, + 0.007846537046134472, + 0.013986383564770222, + 0.008395413868129253, + 0.014309189282357693, + 0.0100984713062644, + 0.013647819869220257, + 0.008616349659860134, + 0.0037402305752038956, + 0.001849241554737091, + 0.0036468824837356806, + 0.002881463849917054, + 0.002114154165610671, + 0.002323170891031623, + 0.002813443075865507, + 0.004278156440705061, + 0.002318157348781824, + 0.0015607322566211224, + 0.0028244222048670053, + 0.0013507758267223835, + 0.0013864723732694983, + 0.0011698201997205615, + 0.0014056403888389468, + 0.001641133800148964, + 0.009059648029506207, + 0.007362237200140953, + 0.010476112365722656, + 0.006743222009390593, + 0.007467008661478758, + 0.006095428951084614, + 0.007101371884346008, + 0.00760642858222127, + 0.00737907225266099, + 0.005207480397075415, + 0.012794502079486847, + 0.005752428900450468, + 0.008547625504434109, + 0.006988706067204475, + 0.005867088679224253, + 0.007137066684663296, + 0.00922126043587923, + 0.007580591831356287, + 0.012112363241612911, + 0.006629041396081448, + 0.007236042525619268, + 0.006985038053244352, + 0.006843710318207741, + 0.007928240112960339, + 0.01593569479882717, + 0.0112778190523386, + 0.02084428444504738, + 0.018223002552986145, + 0.011422227136790752, + 0.01419042982161045, + 0.012593412771821022, + 0.01586667262017727, + 0.04838688671588898, + 0.03569282218813896, + 0.06105288490653038, + 0.03274032101035118, + 0.062344375997781754, + 0.052667856216430664, + 0.1367093324661255, + 0.026108473539352417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.07001859694719315 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062033720314502716 + ], + "opposite_logprob_tokens": [ + -2.8120336532592773 + ], + "logprob_reward": -0.062033720314502716, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062033720314502716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08015548437833786, + 0.01926537975668907, + 0.03646932914853096, + 0.08745510876178741, + 0.023109273985028267, + 0.09745915979146957, + 0.2756142020225525, + 0.04207319766283035, + 0.10535795986652374, + 0.2330409586429596 + ], + "credit_map": [ + 0.012577430345118046, + 0.013136755675077438, + 0.01133979856967926, + 0.00954082328826189, + 0.007239265367388725, + 0.007307624444365501, + 0.0070406850427389145, + 0.011973102577030659, + 0.0022526662796735764, + 0.002103211125358939, + 0.002598651684820652, + 0.0020587770268321037, + 0.003206956200301647, + 0.0025989545974880457, + 0.00236659892834723, + 0.0020795639138668776, + 0.004892788361757994, + 0.00324978563003242, + 0.0053011514246463776, + 0.004932281095534563, + 0.005165913607925177, + 0.0036902069114148617, + 0.006076164077967405, + 0.0031610378064215183, + 0.014694171957671642, + 0.00739567494019866, + 0.014631991274654865, + 0.006909300573170185, + 0.009880789555609226, + 0.00702829472720623, + 0.00926491804420948, + 0.017649976536631584, + 0.003439089749008417, + 0.0027887921314686537, + 0.003117506392300129, + 0.0029568627942353487, + 0.0036157025024294853, + 0.0023208260536193848, + 0.0025281726848334074, + 0.0023423214443027973, + 0.013928846456110477, + 0.010397548787295818, + 0.01759515330195427, + 0.011349762789905071, + 0.01510508544743061, + 0.011040137149393559, + 0.010666683316230774, + 0.007375940680503845, + 0.07276825606822968, + 0.007855496369302273, + 0.10195894539356232, + 0.010580696165561676, + 0.010864462703466415, + 0.04765348508954048, + 0.009928449988365173, + 0.014004411175847054, + 0.006692289374768734, + 0.003268333850428462, + 0.009532504715025425, + 0.004122253507375717, + 0.0028572569135576487, + 0.005730315111577511, + 0.0038401794154196978, + 0.006030069198459387, + 0.0100801857188344, + 0.014252539724111557, + 0.0133122643455863, + 0.02229752391576767, + 0.012371070683002472, + 0.010309046134352684, + 0.011021899990737438, + 0.011713423766195774, + 0.024846307933330536, + 0.020617535337805748, + 0.043663933873176575, + 0.019109424203634262, + 0.020786762237548828, + 0.027953455224633217, + 0.06341441720724106, + 0.012649118900299072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.062033720314502716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046372017823159695 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.0046372017823159695, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046372017823159695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02604222297668457, + 0.005587322171777487, + 0.008189693093299866, + 0.008364814333617687, + 0.006589310709387064, + 0.02602539025247097, + 0.15204507112503052, + 0.03798651322722435, + 0.05587425082921982, + 0.6732953786849976 + ], + "credit_map": [ + 0.003385319607332349, + 0.00418061763048172, + 0.0049474905245006084, + 0.0029175621457397938, + 0.0029283391777426004, + 0.0023672322276979685, + 0.0022827654611319304, + 0.003032898996025324, + 0.0011785521637648344, + 0.0007242248393595219, + 0.000999542768113315, + 0.0005113098886795342, + 0.0005189646035432816, + 0.00043524164357222617, + 0.0006799392285756767, + 0.0005395478801801801, + 0.0012370436452329159, + 0.0008888401789590716, + 0.0010524984681978822, + 0.0008044658461585641, + 0.001033439883030951, + 0.0009639605414122343, + 0.0011936682276427746, + 0.001015777001157403, + 0.0010319136781618, + 0.0008374243625439703, + 0.0011574022937566042, + 0.000939829449634999, + 0.0012150987749919295, + 0.0007830805843695998, + 0.0014032108010724187, + 0.0009968546219170094, + 0.0009971336694434285, + 0.000805961899459362, + 0.0009189057745970786, + 0.0007888568216003478, + 0.0008134294766932726, + 0.0008978015393950045, + 0.0007004349608905613, + 0.0006667869747616351, + 0.004006656352430582, + 0.0037297895178198814, + 0.004758595023304224, + 0.0019642827101051807, + 0.003323653480038047, + 0.0037720995023846626, + 0.002740761497989297, + 0.0017295543802902102, + 0.04141954332590103, + 0.003644806332886219, + 0.05759343132376671, + 0.005575148388743401, + 0.0066722785122692585, + 0.024931812658905983, + 0.005412837490439415, + 0.006795222405344248, + 0.006257069297134876, + 0.003088179975748062, + 0.006741981953382492, + 0.0032453329768031836, + 0.0038270982913672924, + 0.004792675841599703, + 0.0028913933783769608, + 0.007142781745642424, + 0.0055952477268874645, + 0.005815752781927586, + 0.0073830848559737206, + 0.016715798527002335, + 0.0039264969527721405, + 0.004100681282579899, + 0.00684338016435504, + 0.005493813659995794, + 0.07963074743747711, + 0.034543778747320175, + 0.12917622923851013, + 0.08079122006893158, + 0.05997517332434654, + 0.04555373266339302, + 0.21398594975471497, + 0.029638558626174927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.0046372017823159695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld camera, in extreme close-up, cautiously advances, focusing on a pink peony blossoming in the soft morning sunlight. Perched at the center of the flower is a vibrant butterfly. Its velvety wings are covered in an intricate pattern of iridescent blue and fiery orange, edged with delicate black lines. A few glistening dewdrops dapple the petals and wings, shimmering with light. As the camera steadily and slowly approaches, the slight shakiness of the frame conveys a breathless tension. As if sensing an intruder, the butterfly's slender antennae tense. Then, without warning, it explodes, its wings powerfully flapping downward, launching it vertically upward at astonishing speed. The sudden gust of air lifts the fragile peony petals from their pistil, creating a shower of pink petals that swirls and falls gracefully in the air, shimmering with a golden halo. The camera instinctively jerks upward, trying to follow the butterfly, but ultimately captures it as a blurry, colorful dot of light disappearing into the bokeh-like, verdant garden background. The entire scene switches from an instant of extreme tranquility to a brief and brilliant commotion, filled with the vitality of nature and its fleeting beauty.", + "seed": 13747, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_5/index_2572_seed_13747_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02187940515577793, + "vlm_logprob": -0.02187940515577793, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.02187940515577793, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007456093095242977 + ], + "opposite_logprob_tokens": [ + -6.132456302642822 + ], + "logprob_reward": -0.007456093095242977, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.007456093095242977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08610626310110092, + 0.04673149436712265, + 0.10091488063335419, + 0.042700886726379395, + 0.041316207498311996, + 0.18698100745677948, + 0.06626584380865097, + 0.06475717574357986, + 0.19408036768436432, + 0.1701458990573883 + ], + "credit_map": [ + 0.017661962658166885, + 0.016163572669029236, + 0.01276961900293827, + 0.007013880647718906, + 0.0071876938454806805, + 0.008566712960600853, + 0.007117471192032099, + 0.009625349193811417, + 0.00656192097812891, + 0.010654055513441563, + 0.006378124468028545, + 0.0036360006779432297, + 0.00514241773635149, + 0.005350883584469557, + 0.004514622036367655, + 0.004493468441069126, + 0.013740704394876957, + 0.013874334283173084, + 0.01075186487287283, + 0.00749181816354394, + 0.022863347083330154, + 0.012935818172991276, + 0.008837149478495121, + 0.010419847443699837, + 0.0044504692777991295, + 0.004480689764022827, + 0.00535491993650794, + 0.004815869498997927, + 0.008227812126278877, + 0.004561465699225664, + 0.005821690894663334, + 0.004987968131899834, + 0.004858141764998436, + 0.009440197609364986, + 0.0058924476616084576, + 0.0033788089640438557, + 0.0059830257669091225, + 0.003649441758170724, + 0.003101782640442252, + 0.005012363661080599, + 0.030521109700202942, + 0.01825513318181038, + 0.025251714512705803, + 0.019238566979765892, + 0.01863102987408638, + 0.030081911012530327, + 0.02044086717069149, + 0.02456069178879261, + 0.006820411421358585, + 0.010627523064613342, + 0.006884506903588772, + 0.007715507410466671, + 0.008158093318343163, + 0.006411339156329632, + 0.011123916134238243, + 0.008524550125002861, + 0.004798934794962406, + 0.0071434988640248775, + 0.007889647968113422, + 0.015335693955421448, + 0.008352147415280342, + 0.007913313806056976, + 0.00792018324136734, + 0.0054037561640143394, + 0.05769430473446846, + 0.027137532830238342, + 0.03268188238143921, + 0.011305373162031174, + 0.011386667378246784, + 0.01712995395064354, + 0.009530619718134403, + 0.027214037254452705, + 0.01767054945230484, + 0.017889155074954033, + 0.020097823813557625, + 0.024276886135339737, + 0.017255695536732674, + 0.019104687497019768, + 0.018681013956665993, + 0.03517008200287819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.007456093095242977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009808086790144444 + ], + "opposite_logprob_tokens": [ + -4.63480806350708 + ], + "logprob_reward": -0.009808086790144444, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009808086790144444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10759216547012329, + 0.07143484801054001, + 0.06816360354423523, + 0.1390911340713501, + 0.03802665323019028, + 0.07551184296607971, + 0.06418997049331665, + 0.15813685953617096, + 0.1711689829826355, + 0.10668396204710007 + ], + "credit_map": [ + 0.016902612522244453, + 0.021579205989837646, + 0.013410435058176517, + 0.010194798931479454, + 0.013254291377961636, + 0.012039166875183582, + 0.01104815024882555, + 0.00916350819170475, + 0.008334015496075153, + 0.013742326758801937, + 0.009695685468614101, + 0.007470381911844015, + 0.007627967745065689, + 0.009995430707931519, + 0.006282345857471228, + 0.008286693133413792, + 0.009555787779390812, + 0.009520416148006916, + 0.006990885362029076, + 0.004863375797867775, + 0.01558924000710249, + 0.00985688529908657, + 0.0060300687327980995, + 0.0057569402270019054, + 0.017194941639900208, + 0.031050628051161766, + 0.01564137265086174, + 0.010524667799472809, + 0.03381291404366493, + 0.010505462996661663, + 0.008280503563582897, + 0.012080635875463486, + 0.005080713890492916, + 0.008397177793085575, + 0.004722094628959894, + 0.002253691665828228, + 0.004926742520183325, + 0.002859575441107154, + 0.0022628328297287226, + 0.00752382492646575, + 0.008769143372774124, + 0.012631072662770748, + 0.016204969957470894, + 0.005802629981189966, + 0.008705522865056992, + 0.004593233112245798, + 0.00501291872933507, + 0.01379235740751028, + 0.007424851879477501, + 0.015242394059896469, + 0.004381193313747644, + 0.005961867514997721, + 0.008403465151786804, + 0.006545097567141056, + 0.00752653181552887, + 0.008704569190740585, + 0.014890260994434357, + 0.014905933290719986, + 0.019435351714491844, + 0.024601444602012634, + 0.017696186900138855, + 0.019109655171632767, + 0.020470324903726578, + 0.027027714997529984, + 0.011422253213822842, + 0.02303371950984001, + 0.01364136766642332, + 0.014329214580357075, + 0.029654167592525482, + 0.038596853613853455, + 0.015288012102246284, + 0.02520340122282505, + 0.009233631193637848, + 0.017318224534392357, + 0.01173836924135685, + 0.011606456711888313, + 0.022387992590665817, + 0.011882365681231022, + 0.010008150711655617, + 0.012508773244917393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.009808086790144444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033703286200761795 + ], + "opposite_logprob_tokens": [ + -3.408703327178955 + ], + "logprob_reward": -0.033703286200761795, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033703286200761795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08935601264238358, + 0.03749935328960419, + 0.045996226370334625, + 0.04189303517341614, + 0.03187887370586395, + 0.15997444093227386, + 0.05878738313913345, + 0.11118461936712265, + 0.06457510590553284, + 0.3588549792766571 + ], + "credit_map": [ + 0.010872592218220234, + 0.01412943098694086, + 0.00997063610702753, + 0.011862466111779213, + 0.013207587413489819, + 0.008435542695224285, + 0.010975368320941925, + 0.009902391582727432, + 0.0056172022596001625, + 0.0053448122926056385, + 0.004396910313516855, + 0.003783443011343479, + 0.004939342383295298, + 0.0052082836627960205, + 0.003647640347480774, + 0.004561718553304672, + 0.0051538110710680485, + 0.006610482465475798, + 0.005017163697630167, + 0.0034562291111797094, + 0.010886351577937603, + 0.004229308106005192, + 0.004730801563709974, + 0.00591207854449749, + 0.0045531741343438625, + 0.008581419475376606, + 0.004839922767132521, + 0.002808420918881893, + 0.009627810679376125, + 0.003098259912803769, + 0.0022871834225952625, + 0.006096845027059317, + 0.003933086059987545, + 0.00565132824704051, + 0.005574237089604139, + 0.0020645884796977043, + 0.00524792680516839, + 0.0027595506981015205, + 0.0017942842096090317, + 0.004853869788348675, + 0.02333812415599823, + 0.021295279264450073, + 0.016963927075266838, + 0.01569775678217411, + 0.016543637961149216, + 0.02215443179011345, + 0.018232906237244606, + 0.02574838139116764, + 0.007771017961204052, + 0.011714004911482334, + 0.005435947794467211, + 0.00419141910970211, + 0.009032953530550003, + 0.00613352470099926, + 0.006726667284965515, + 0.007781847845762968, + 0.009802762418985367, + 0.014968150295317173, + 0.011044648475944996, + 0.02365068532526493, + 0.020078426226973534, + 0.013955308124423027, + 0.010666617192327976, + 0.007018019445240498, + 0.0062225330621004105, + 0.011398282833397388, + 0.005456279031932354, + 0.004319409374147654, + 0.009744192473590374, + 0.009953814558684826, + 0.007076773326843977, + 0.010403824038803577, + 0.03406813368201256, + 0.05312686413526535, + 0.029643574729561806, + 0.0327615961432457, + 0.06188710033893585, + 0.04398293048143387, + 0.037273671478033066, + 0.06611111015081406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.033703286200761795 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864921420812607 + ], + "opposite_logprob_tokens": [ + -3.048649311065674 + ], + "logprob_reward": -0.04864921420812607, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864921420812607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1328713744878769, + 0.0858660414814949, + 0.08120529353618622, + 0.03459099680185318, + 0.022359514608979225, + 0.1900480091571808, + 0.05821334570646286, + 0.10117938369512558, + 0.13683836162090302, + 0.15682768821716309 + ], + "credit_map": [ + 0.03351597487926483, + 0.030490024015307426, + 0.019501863047480583, + 0.010096806101500988, + 0.011661814525723457, + 0.01232212595641613, + 0.007666278630495071, + 0.007616478484123945, + 0.011934296227991581, + 0.019846808165311813, + 0.011715888977050781, + 0.007830025628209114, + 0.010167768225073814, + 0.011405298486351967, + 0.007313322275876999, + 0.005652624182403088, + 0.010479046031832695, + 0.009348820894956589, + 0.00968980137258768, + 0.006631250027567148, + 0.018134860321879387, + 0.011837869882583618, + 0.006145510822534561, + 0.008938128128647804, + 0.004415958188474178, + 0.006304045673459768, + 0.004276831168681383, + 0.0034187904093414545, + 0.0068685756996273994, + 0.003351933555677533, + 0.002733509289100766, + 0.003221348859369755, + 0.003293850226327777, + 0.0037426031194627285, + 0.002811461454257369, + 0.0021941764280200005, + 0.0031736071687191725, + 0.0019689390901476145, + 0.0022247571032494307, + 0.002950116526335478, + 0.029156388714909554, + 0.02130666933953762, + 0.020173411816358566, + 0.01949954777956009, + 0.02005033567547798, + 0.0289889145642519, + 0.022043492645025253, + 0.02882923185825348, + 0.007289016153663397, + 0.009810354560613632, + 0.0032276224810630083, + 0.005038244649767876, + 0.009243418462574482, + 0.005888472311198711, + 0.008824887685477734, + 0.008891325443983078, + 0.0073098475113511086, + 0.008391927927732468, + 0.011032023467123508, + 0.027953140437602997, + 0.009035859256982803, + 0.013178488239645958, + 0.011746341362595558, + 0.01253175176680088, + 0.039623986929655075, + 0.017332056537270546, + 0.022762909531593323, + 0.010789399035274982, + 0.007848139852285385, + 0.010448706336319447, + 0.010008228942751884, + 0.018024925142526627, + 0.020095475018024445, + 0.014208456501364708, + 0.017717014998197556, + 0.018646320328116417, + 0.013963792473077774, + 0.011119230650365353, + 0.02004040963947773, + 0.041036974638700485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.04864921420812607 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780345484614372 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780345484614372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780345484614372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07205274701118469, + 0.028808938339352608, + 0.01850101910531521, + 0.045747607946395874, + 0.029335714876651764, + 0.44622117280960083, + 0.18548840284347534, + 0.04222685843706131, + 0.053626660257577896, + 0.07799087464809418 + ], + "credit_map": [ + 0.010158072225749493, + 0.009389965794980526, + 0.00852045789361, + 0.00610958319157362, + 0.01573810912668705, + 0.009164005517959595, + 0.008024560287594795, + 0.0049479869194328785, + 0.004639352206140757, + 0.005809559486806393, + 0.0035438777413219213, + 0.0029532432090491056, + 0.003076747991144657, + 0.0041325753554701805, + 0.002355380915105343, + 0.0022982023656368256, + 0.0023746828082948923, + 0.0029185181483626366, + 0.002361118094995618, + 0.0016585729317739606, + 0.003606866579502821, + 0.002006723079830408, + 0.001855408656410873, + 0.0017191283404827118, + 0.005965901538729668, + 0.007620842196047306, + 0.005094728898257017, + 0.0039790417067706585, + 0.010385592468082905, + 0.004064600449055433, + 0.0034558181650936604, + 0.005181082524359226, + 0.004255855455994606, + 0.004653454292565584, + 0.003966070711612701, + 0.0027733929455280304, + 0.004615433514118195, + 0.0026116184890270233, + 0.0026710378006100655, + 0.003788853995501995, + 0.07032117247581482, + 0.052345823496580124, + 0.05066090077161789, + 0.044771384447813034, + 0.04540615901350975, + 0.06945109367370605, + 0.05173783004283905, + 0.0615268237888813, + 0.019317660480737686, + 0.03240516036748886, + 0.015806833282113075, + 0.021305354312062263, + 0.020661817863583565, + 0.016226816922426224, + 0.034735243767499924, + 0.025029530748724937, + 0.0038023176603019238, + 0.00468286219984293, + 0.004871366545557976, + 0.007124891970306635, + 0.008617927320301533, + 0.0048588477075099945, + 0.004487711936235428, + 0.003780931932851672, + 0.005701972171664238, + 0.006167599465698004, + 0.005994787905365229, + 0.004733956418931484, + 0.007851607166230679, + 0.008790555410087109, + 0.006915961857885122, + 0.007470214739441872, + 0.011972786858677864, + 0.012024861760437489, + 0.007717084139585495, + 0.007423614617437124, + 0.01068150345236063, + 0.00991605594754219, + 0.008272712118923664, + 0.009982259012758732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.009780345484614372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld camera, in extreme close-up, cautiously advances, focusing on a pink peony blossoming in the soft morning sunlight. Perched at the center of the flower is a vibrant butterfly. Its velvety wings are covered in an intricate pattern of iridescent blue and fiery orange, edged with delicate black lines. A few glistening dewdrops dapple the petals and wings, shimmering with light. As the camera steadily and slowly approaches, the slight shakiness of the frame conveys a breathless tension. As if sensing an intruder, the butterfly's slender antennae tense. Then, without warning, it explodes, its wings powerfully flapping downward, launching it vertically upward at astonishing speed. The sudden gust of air lifts the fragile peony petals from their pistil, creating a shower of pink petals that swirls and falls gracefully in the air, shimmering with a golden halo. The camera instinctively jerks upward, trying to follow the butterfly, but ultimately captures it as a blurry, colorful dot of light disappearing into the bokeh-like, verdant garden background. The entire scene switches from an instant of extreme tranquility to a brief and brilliant commotion, filled with the vitality of nature and its fleeting beauty.", + "seed": 13748, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_5/index_2572_seed_13748_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02205933229997754, + "vlm_logprob": -0.02205933229997754, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.02205933229997754, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004330899100750685 + ], + "opposite_logprob_tokens": [ + -7.254331111907959 + ], + "logprob_reward": -0.004330899100750685, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.004330899100750685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03528415039181709, + 0.050013650208711624, + 0.11659010499715805, + 0.05177587643265724, + 0.042315177619457245, + 0.07378439605236053, + 0.14617428183555603, + 0.13053534924983978, + 0.142781600356102, + 0.2107454389333725 + ], + "credit_map": [ + 0.005861522629857063, + 0.007787598762661219, + 0.0039251563139259815, + 0.00277796876616776, + 0.006118935998529196, + 0.0029378854669630527, + 0.0035528175067156553, + 0.002322262851521373, + 0.007593437097966671, + 0.010309590958058834, + 0.010235658846795559, + 0.00565442955121398, + 0.005591689143329859, + 0.003143124282360077, + 0.003990142606198788, + 0.0034955726005136967, + 0.03987440466880798, + 0.0057927994057536125, + 0.046715181320905685, + 0.004999956116080284, + 0.004746375139802694, + 0.0036317482590675354, + 0.003219224978238344, + 0.00761040486395359, + 0.00869894027709961, + 0.008898207917809486, + 0.004221586976200342, + 0.0063903299160301685, + 0.007112825755029917, + 0.003486719448119402, + 0.0053529818542301655, + 0.007614278234541416, + 0.004945472814142704, + 0.0074777184054255486, + 0.0048374696634709835, + 0.00607647281140089, + 0.004284339491277933, + 0.0028275172226130962, + 0.008575007319450378, + 0.003291178261861205, + 0.006311655510216951, + 0.004122595302760601, + 0.008153139613568783, + 0.01536504365503788, + 0.007784162648022175, + 0.006776571273803711, + 0.013510404154658318, + 0.01176081970334053, + 0.021308457478880882, + 0.026380371302366257, + 0.02434288151562214, + 0.011428589932620525, + 0.01207294687628746, + 0.017120543867349625, + 0.011766448616981506, + 0.02175402082502842, + 0.010047638788819313, + 0.015371511690318584, + 0.011763056740164757, + 0.020093154162168503, + 0.018022777512669563, + 0.014093647710978985, + 0.02003321424126625, + 0.021110333502292633, + 0.014087042771279812, + 0.026962466537952423, + 0.015350714325904846, + 0.011134035885334015, + 0.01628219522535801, + 0.019433995708823204, + 0.02233561873435974, + 0.017195524647831917, + 0.016914011910557747, + 0.008609754033386707, + 0.02082720212638378, + 0.027325542643666267, + 0.01011930126696825, + 0.014028338715434074, + 0.07793852686882019, + 0.034982744604349136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.004330899100750685 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551599182188511 + ], + "opposite_logprob_tokens": [ + -4.387551784515381 + ], + "logprob_reward": -0.012551599182188511, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551599182188511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040902383625507355, + 0.014574402011930943, + 0.08526767045259476, + 0.016059335321187973, + 0.05105457082390785, + 0.02119290642440319, + 0.02692238986492157, + 0.03707116097211838, + 0.11192312836647034, + 0.5950321555137634 + ], + "credit_map": [ + 0.006113453768193722, + 0.011490238830447197, + 0.004004951566457748, + 0.0032532785553485155, + 0.006865163333714008, + 0.003980735316872597, + 0.0028534724842756987, + 0.0023410902358591557, + 0.0021607086528092623, + 0.0023809499107301235, + 0.0031429112423211336, + 0.0013273069635033607, + 0.00198753853328526, + 0.001151715638116002, + 0.0011505549773573875, + 0.001272716443054378, + 0.03118174895644188, + 0.0038281080778688192, + 0.03396200016140938, + 0.004251301288604736, + 0.003336586756631732, + 0.002725231694057584, + 0.0025235156062990427, + 0.0034591841977089643, + 0.002660562517121434, + 0.0029240690637379885, + 0.0011737574823200703, + 0.0019398201256990433, + 0.002035098848864436, + 0.0010131222661584616, + 0.001985853537917137, + 0.0023270505480468273, + 0.006739840842783451, + 0.007344523910433054, + 0.004739158321171999, + 0.006819284055382013, + 0.004271306563168764, + 0.003398110391572118, + 0.013223311863839626, + 0.004519034177064896, + 0.00171130639500916, + 0.0018435809761285782, + 0.002797103952616453, + 0.004122950602322817, + 0.0019628575537353754, + 0.0017592437798157334, + 0.004220150876790285, + 0.0027757107745856047, + 0.0027721335645765066, + 0.0055227577686309814, + 0.003084976691752672, + 0.002728149062022567, + 0.002865690039470792, + 0.0033441418781876564, + 0.002940058009698987, + 0.0036644816864281893, + 0.002564376685768366, + 0.004472192842513323, + 0.003290284425020218, + 0.0059113106690347195, + 0.005442272871732712, + 0.005596150644123554, + 0.005337302107363939, + 0.004457271657884121, + 0.010384171269834042, + 0.012488250620663166, + 0.009455474093556404, + 0.012570243328809738, + 0.014222946017980576, + 0.019132353365421295, + 0.023183248937129974, + 0.010486440733075142, + 0.05259637534618378, + 0.015975555405020714, + 0.06405055522918701, + 0.045054543763399124, + 0.023810068145394325, + 0.02230355143547058, + 0.26071488857269287, + 0.11052657663822174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.012551599182188511 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048654552549123764 + ], + "opposite_logprob_tokens": [ + -3.048654556274414 + ], + "logprob_reward": -0.048654552549123764, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048654552549123764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01545182429254055, + 0.012021256610751152, + 0.033974532037973404, + 0.008385146036744118, + 0.012617774307727814, + 0.015906011685729027, + 0.0307528804987669, + 0.060858044773340225, + 0.29237255454063416, + 0.5176599621772766 + ], + "credit_map": [ + 0.0020916468929499388, + 0.0029944179113954306, + 0.001831419882364571, + 0.0014057185035198927, + 0.0024050134234130383, + 0.00139338755980134, + 0.002299589803442359, + 0.0010306299664080143, + 0.0017625304171815515, + 0.002663754392415285, + 0.002214303705841303, + 0.0009189804550260305, + 0.0017729689134284854, + 0.0008434748160652816, + 0.0010450574336573482, + 0.000800186418928206, + 0.011355961672961712, + 0.0019410733366385102, + 0.011330200359225273, + 0.002374592237174511, + 0.0014378183986991644, + 0.0013464203802868724, + 0.0012037832057103515, + 0.0029846816323697567, + 0.0014540482079610229, + 0.0015997019363567233, + 0.0006346757872961462, + 0.0009078313014470041, + 0.0011340891942381859, + 0.0005327686085365713, + 0.0009176777093671262, + 0.0012043527094647288, + 0.0017737901071086526, + 0.0015601218910887837, + 0.001292724977247417, + 0.0019168382277712226, + 0.0011650831438601017, + 0.000865916081238538, + 0.0028713452629745007, + 0.0011719554895535111, + 0.0013089405838400126, + 0.0014007335994392633, + 0.00185209431219846, + 0.0027937230188399553, + 0.0014663657639175653, + 0.002087066415697336, + 0.0029699893202632666, + 0.0020270971581339836, + 0.0031428812071681023, + 0.0063686370849609375, + 0.0030055835377424955, + 0.003353258827701211, + 0.0035312373656779528, + 0.003672731341794133, + 0.002765954937785864, + 0.004912596195936203, + 0.003847187850624323, + 0.008978634141385555, + 0.004850972909480333, + 0.01006751973181963, + 0.010487491264939308, + 0.009925163350999355, + 0.007861805148422718, + 0.0048392717726528645, + 0.030362864956259727, + 0.031145021319389343, + 0.023771023377776146, + 0.03211519494652748, + 0.0384443998336792, + 0.04916611686348915, + 0.05923031270503998, + 0.02813761495053768, + 0.044030047953128815, + 0.014458777382969856, + 0.0561218224465847, + 0.04003990441560745, + 0.019930772483348846, + 0.01927102729678154, + 0.22678516805171967, + 0.09702251106500626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.048654552549123764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033688995987176895 + ], + "opposite_logprob_tokens": [ + -3.408689022064209 + ], + "logprob_reward": -0.033688995987176895, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033688995987176895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023755986243486404, + 0.024080364033579826, + 0.08628612756729126, + 0.014984885230660439, + 0.02512996271252632, + 0.02988123707473278, + 0.08527136594057083, + 0.207561194896698, + 0.10936108231544495, + 0.39368775486946106 + ], + "credit_map": [ + 0.0035052259918302298, + 0.004398982040584087, + 0.002581515582278371, + 0.0020676301792263985, + 0.004328275565057993, + 0.0020089629106223583, + 0.002923051593825221, + 0.001942346920259297, + 0.00406276062130928, + 0.005048106424510479, + 0.003997495397925377, + 0.0024809965398162603, + 0.002855168655514717, + 0.001809662557207048, + 0.001916677225381136, + 0.0019095013849437237, + 0.028729557991027832, + 0.0037606412079185247, + 0.027330594137310982, + 0.006826586090028286, + 0.0031547737307846546, + 0.002766065765172243, + 0.003072923282161355, + 0.010644998401403427, + 0.0022384384647011757, + 0.0029532855842262506, + 0.001102279988117516, + 0.0017519211396574974, + 0.002593075390905142, + 0.0009627257823012769, + 0.001551397959701717, + 0.001831762958317995, + 0.003576246788725257, + 0.0042429156601428986, + 0.0033582269679754972, + 0.002374190604314208, + 0.002416055416688323, + 0.0021729080472141504, + 0.004887607879936695, + 0.002101816004142165, + 0.0025604816619306803, + 0.002151713240891695, + 0.0029015722684562206, + 0.005438433028757572, + 0.003844278398901224, + 0.003240281017497182, + 0.005598738789558411, + 0.004145744256675243, + 0.009354034438729286, + 0.013308060355484486, + 0.011479312554001808, + 0.009377784095704556, + 0.008164653554558754, + 0.01325211301445961, + 0.008987581357359886, + 0.01134783960878849, + 0.012448360212147236, + 0.0281356293708086, + 0.017063861712813377, + 0.028010323643684387, + 0.03346843272447586, + 0.04649960249662399, + 0.02643975429236889, + 0.01549525000154972, + 0.009585920721292496, + 0.023807983845472336, + 0.012904486618936062, + 0.012036800384521484, + 0.01100804191082716, + 0.01316427905112505, + 0.017845723778009415, + 0.009007858112454414, + 0.034842249006032944, + 0.011495121754705906, + 0.04443151503801346, + 0.03076726570725441, + 0.014263292774558067, + 0.01506394986063242, + 0.1714145541191101, + 0.07140986621379852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.033688995987176895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme close-up of a pink peony with a vibrant butterfly perched on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the butterfly have velvety wings with an intricate pattern of iridescent blue and fiery orange edged with black lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glistening dewdrops on the petals and wings that shimmer with light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a handheld, shaky movement advancing slowly, then jerk upward to follow the butterfly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the butterfly launch vertically upward at high speed, causing a shower of pink petals to swirl and fall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107061468064785 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.01107061468064785, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107061468064785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027465282008051872, + 0.01004627626389265, + 0.04718146473169327, + 0.015089600346982479, + 0.0242262352257967, + 0.027514750137925148, + 0.09533905982971191, + 0.24276283383369446, + 0.0675700232386589, + 0.4428044259548187 + ], + "credit_map": [ + 0.003940707538276911, + 0.007018566597253084, + 0.003181345760822296, + 0.0020806752145290375, + 0.0043641189113259315, + 0.002686935244128108, + 0.002433110261335969, + 0.0017598230624571443, + 0.0016262931749224663, + 0.0010384401539340615, + 0.0014963241992518306, + 0.0012189532862976193, + 0.0013865889050066471, + 0.0011031648609787226, + 0.0011896600481122732, + 0.0009868517518043518, + 0.01477051991969347, + 0.0021269156131893396, + 0.013388494029641151, + 0.004570897202938795, + 0.001568091451190412, + 0.001709356321953237, + 0.001821254612877965, + 0.007225939072668552, + 0.0028059266041964293, + 0.0009898057905957103, + 0.0017041207756847143, + 0.0020442570094019175, + 0.001978730782866478, + 0.000991425709798932, + 0.0018423470901325345, + 0.002732988214120269, + 0.0033127593342214823, + 0.00465370574966073, + 0.002368748188018799, + 0.002762327902019024, + 0.0024792752228677273, + 0.0019039006438106298, + 0.004786864388734102, + 0.0019586551934480667, + 0.002253437414765358, + 0.0025559531059116125, + 0.003047267207875848, + 0.00477410014718771, + 0.0031034823041409254, + 0.0036921242717653513, + 0.00467129610478878, + 0.0034170900471508503, + 0.012641475535929203, + 0.012610496953129768, + 0.0112829664722085, + 0.011535797268152237, + 0.007883189246058464, + 0.015814855694770813, + 0.011027343571186066, + 0.012542943470180035, + 0.013715515844523907, + 0.033020466566085815, + 0.019826170057058334, + 0.0358491912484169, + 0.03773151710629463, + 0.04069502279162407, + 0.03817868232727051, + 0.02374628372490406, + 0.007782435044646263, + 0.009647602215409279, + 0.007101768162101507, + 0.009907720610499382, + 0.0062341829761862755, + 0.0062409318052232265, + 0.011524458415806293, + 0.009130925871431828, + 0.03810424357652664, + 0.012175224721431732, + 0.04747028648853302, + 0.032944921404123306, + 0.01691533252596855, + 0.016674159094691277, + 0.19512298703193665, + 0.0833972916007042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.01107061468064785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_6/index_tensor([5643, 5643])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_6/index_tensor([5643, 5643])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..22b707473d0c4c7e85d056c8d9ea3ff9ae214696 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_6/index_tensor([5643, 5643])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot shot from a horizontal angle centers on a playful crayon drawing on slightly textured white paper. The painting, rendered in highly saturated, rough-brushed crayons, depicts a simple story: a large, smiling yellow sun hangs high in the sky, beneath a small blue house with a red triangular roof and square windows, and beside it, a waving stick figure. The entire scene exudes an atmosphere of childlike joy. Suddenly, a sharp, sharply defined black shadow pierces the frame from beyond, poking precisely at the center. Immediately, the crayon-textured drawing, like a popped balloon, begins to rapidly sag inward from the point of impact. The once-flat surface of the paper instantly loses its tension, and the sun, house, and stick figure in the painting wrinkle and distort painfully as the image shrinks. The vibrant colors fade and blur, and ultimately, the entire painting collapses like a deflated balloon, shrunken and contracting into a ball, completely losing its original shape and story.", + "seed": 19980, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_6/index_5643_seed_19980_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08635750823887064, + "vlm_logprob": -0.08635750823887064, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.08635750823887064, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02685648202896118 + ], + "opposite_logprob_tokens": [ + -4.026856422424316 + ], + "logprob_reward": -0.02685648202896118, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02685648202896118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2253199815750122, + 0.4431297183036804, + 0.019188636913895607, + 0.02903723530471325, + 0.03456836938858032, + 0.056906163692474365, + 0.05644403398036957, + 0.030280418694019318, + 0.036537591367959976, + 0.0685877799987793 + ], + "credit_map": [ + 0.05699695646762848, + 0.014572830870747566, + 0.00933833047747612, + 0.015018668957054615, + 0.02531280927360058, + 0.08226235210895538, + 0.008134933188557625, + 0.013683122582733631, + 0.12487757951021194, + 0.03418055921792984, + 0.010711361654102802, + 0.023783504962921143, + 0.04617033526301384, + 0.17289799451828003, + 0.011223278939723969, + 0.019285142421722412, + 0.0024477364495396614, + 0.0017709082458168268, + 0.001440474996343255, + 0.0022327774204313755, + 0.004940800368785858, + 0.003206973196938634, + 0.0017388807609677315, + 0.001410085940733552, + 0.0035672869998961687, + 0.0030135675333440304, + 0.003657556837424636, + 0.003722911002114415, + 0.004968095570802689, + 0.003944720141589642, + 0.002582445042207837, + 0.0035806535743176937, + 0.005158458836376667, + 0.004030452575534582, + 0.0025374027900397778, + 0.005224144551903009, + 0.004179570823907852, + 0.0076282392255961895, + 0.0028381850570440292, + 0.0029719171579927206, + 0.0064542256295681, + 0.00766310840845108, + 0.003585363971069455, + 0.006337223574519157, + 0.011775055900216103, + 0.010040969587862492, + 0.005847733002156019, + 0.005202482920140028, + 0.007462632842361927, + 0.006266348995268345, + 0.005028472747653723, + 0.005476059857755899, + 0.005371155217289925, + 0.008096449077129364, + 0.005653891246765852, + 0.013089023530483246, + 0.00288521870970726, + 0.004888951312750578, + 0.00289875827729702, + 0.0034223170951008797, + 0.006881664041429758, + 0.004261919762939215, + 0.002668025204911828, + 0.002373564289882779, + 0.004395770374685526, + 0.0033190841786563396, + 0.0030086974147707224, + 0.005799220874905586, + 0.006734259892255068, + 0.006503637880086899, + 0.0035105817951261997, + 0.0032663403544574976, + 0.006348396651446819, + 0.009419593960046768, + 0.009307879023253918, + 0.015233561396598816, + 0.007066551595926285, + 0.00887441448867321, + 0.006431126967072487, + 0.005906250793486834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.02685648202896118 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870964050292969 + ], + "opposite_logprob_tokens": [ + -1.1370964050292969 + ], + "logprob_reward": -0.3870964050292969, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870964050292969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19693416357040405, + 0.3943520784378052, + 0.03202958405017853, + 0.10513661801815033, + 0.03409751132130623, + 0.020945584401488304, + 0.07849320024251938, + 0.05372614786028862, + 0.032059162855148315, + 0.052225999534130096 + ], + "credit_map": [ + 0.05438557267189026, + 0.01563616842031479, + 0.004725797567516565, + 0.010448827408254147, + 0.02026868797838688, + 0.07622676342725754, + 0.005832177121192217, + 0.009410161525011063, + 0.11164601892232895, + 0.030546898022294044, + 0.009551867842674255, + 0.021344607695937157, + 0.04058601334691048, + 0.153917595744133, + 0.010041618719696999, + 0.01671746000647545, + 0.007137522101402283, + 0.00205237977206707, + 0.0015778210945427418, + 0.002056132536381483, + 0.007046212907880545, + 0.008661190047860146, + 0.0016440190374851227, + 0.0018543072510510683, + 0.020495310425758362, + 0.013496223837137222, + 0.004522792529314756, + 0.007526638451963663, + 0.01921766996383667, + 0.025988411158323288, + 0.006500592455267906, + 0.0073889815248548985, + 0.005272996611893177, + 0.00535525381565094, + 0.0019009560346603394, + 0.003882547141984105, + 0.005460011772811413, + 0.006917908787727356, + 0.002762100426480174, + 0.0025457364972680807, + 0.0024327475111931562, + 0.0021942511666566133, + 0.0017815936589613557, + 0.0035613486543297768, + 0.002225629985332489, + 0.0039770216681063175, + 0.001946285250596702, + 0.0028267076704651117, + 0.008140824735164642, + 0.009251737967133522, + 0.004936666693538427, + 0.00869390182197094, + 0.009396946988999844, + 0.019796140491962433, + 0.00793447159230709, + 0.010342507623136044, + 0.004564323928207159, + 0.009250652976334095, + 0.0047769490629434586, + 0.006370146293193102, + 0.011290986090898514, + 0.007482875604182482, + 0.005351597908884287, + 0.004638616926968098, + 0.0033819330856204033, + 0.0056530022993683815, + 0.004289855714887381, + 0.004150014836341143, + 0.00417399313300848, + 0.004028975497931242, + 0.002926524030044675, + 0.0034548635594546795, + 0.00522286631166935, + 0.00830891914665699, + 0.005939141847193241, + 0.01109646912664175, + 0.005620528478175402, + 0.006021907087415457, + 0.00530779967084527, + 0.004708366002887487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.3870964050292969 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000726197671610862 + ], + "opposite_logprob_tokens": [ + -7.250726222991943 + ], + "logprob_reward": -0.000726197671610862, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.000726197671610862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12537921965122223, + 0.5440806746482849, + 0.05838220193982124, + 0.04242716729640961, + 0.024378621950745583, + 0.03533536195755005, + 0.021095581352710724, + 0.08259988576173782, + 0.02092178910970688, + 0.045399438589811325 + ], + "credit_map": [ + 0.011546522378921509, + 0.010327015072107315, + 0.009014398790895939, + 0.01418126467615366, + 0.03242749720811844, + 0.021995697170495987, + 0.011480565182864666, + 0.014406275935471058, + 0.15059232711791992, + 0.04282855987548828, + 0.013472056947648525, + 0.028061969205737114, + 0.059341851621866226, + 0.21324272453784943, + 0.01417198870331049, + 0.0223692636936903, + 0.008140625432133675, + 0.004734696354717016, + 0.006801735609769821, + 0.006175643764436245, + 0.010633590631186962, + 0.010871827602386475, + 0.0055826278403401375, + 0.005441460758447647, + 0.009100383147597313, + 0.004169203341007233, + 0.0032457413617521524, + 0.0036962328013032675, + 0.005053466651588678, + 0.011588651686906815, + 0.002005476038902998, + 0.0035680150613188744, + 0.0030336622148752213, + 0.002898693084716797, + 0.00199215323664248, + 0.0024929530918598175, + 0.003899897914379835, + 0.005274024326354265, + 0.002166725927963853, + 0.0026205137837678194, + 0.003936741501092911, + 0.005952833686023951, + 0.0030879206024110317, + 0.0030737929046154022, + 0.007571767549961805, + 0.005520393140614033, + 0.0031928184907883406, + 0.0029990968760102987, + 0.002425685292109847, + 0.002653016010299325, + 0.0023829005658626556, + 0.003057120367884636, + 0.002004035981371999, + 0.0028450842946767807, + 0.0026135211810469627, + 0.0031142185907810926, + 0.007645750883966684, + 0.016732824966311455, + 0.00493974145501852, + 0.0064100357703864574, + 0.021202977746725082, + 0.01292097195982933, + 0.007388127502053976, + 0.0053594643250107765, + 0.0025441546458750963, + 0.0023815908934921026, + 0.0022105444222688675, + 0.002965135034173727, + 0.0026573087088763714, + 0.003148038173094392, + 0.0026498474180698395, + 0.0023651716765016317, + 0.003440343774855137, + 0.004601750988513231, + 0.005397314205765724, + 0.009134975261986256, + 0.006251349579542875, + 0.005513512995094061, + 0.005450769327580929, + 0.005609429441392422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.000726197671610862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010434903670102358 + ], + "opposite_logprob_tokens": [ + -6.876043319702148 + ], + "logprob_reward": -0.0010434903670102358, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010434903670102358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15193292498588562, + 0.07485251128673553, + 0.0305490680038929, + 0.04380981996655464, + 0.05519479140639305, + 0.060649849474430084, + 0.40994179248809814, + 0.06041794270277023, + 0.06556052714586258, + 0.047090791165828705 + ], + "credit_map": [ + 0.03745013847947121, + 0.012543275021016598, + 0.005015712697058916, + 0.01020137406885624, + 0.016235023736953735, + 0.05454177036881447, + 0.005869210697710514, + 0.010076407343149185, + 0.018040349707007408, + 0.00642334995791316, + 0.0021494741085916758, + 0.004526390228420496, + 0.009513690136373043, + 0.027503129094839096, + 0.0026416934560984373, + 0.004054433666169643, + 0.005521143786609173, + 0.0029041997622698545, + 0.0016798864817246795, + 0.00248497212305665, + 0.005804621148854494, + 0.007935392670333385, + 0.0020150060299783945, + 0.0022038440220057964, + 0.007990475744009018, + 0.0044353436678647995, + 0.002413567854091525, + 0.003780257422477007, + 0.008004632778465748, + 0.010669915936887264, + 0.0033581897150725126, + 0.0031574370805174112, + 0.007902565412223339, + 0.009668275713920593, + 0.003333647735416889, + 0.005066693294793367, + 0.008901511318981647, + 0.010214625857770443, + 0.004989993758499622, + 0.005117481574416161, + 0.00831822119653225, + 0.007252806331962347, + 0.0040679736994206905, + 0.005210218019783497, + 0.007548081688582897, + 0.016138315200805664, + 0.004382158629596233, + 0.007732075173407793, + 0.045778077095746994, + 0.04454997181892395, + 0.036750178784132004, + 0.036745697259902954, + 0.028446299955248833, + 0.09015562385320663, + 0.04003562778234482, + 0.087480328977108, + 0.006102593149989843, + 0.011099803261458874, + 0.003597102127969265, + 0.0052568609826266766, + 0.014212138950824738, + 0.009881126694381237, + 0.0061242906376719475, + 0.004144024569541216, + 0.007551929913461208, + 0.007133675739169121, + 0.00795629434287548, + 0.009454388171434402, + 0.006934378296136856, + 0.010966571047902107, + 0.009248246438801289, + 0.00631504412740469, + 0.00490199401974678, + 0.00574340857565403, + 0.006034410558640957, + 0.006163994316011667, + 0.006108526606112719, + 0.00539722153916955, + 0.007456256076693535, + 0.005284980870783329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0010434903670102358 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064966097474098 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016064966097474098, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064966097474098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056562215089797974, + 0.20443667471408844, + 0.026508904993534088, + 0.029161343351006508, + 0.051571913063526154, + 0.050166670233011246, + 0.07308872789144516, + 0.08396351337432861, + 0.11311516910791397, + 0.31142476201057434 + ], + "credit_map": [ + 0.00880020298063755, + 0.007373353466391563, + 0.00494010653346777, + 0.008510301820933819, + 0.006739605218172073, + 0.010889546945691109, + 0.004558248445391655, + 0.004750861786305904, + 0.05588610842823982, + 0.01604796200990677, + 0.005399055778980255, + 0.010892407968640327, + 0.022695254534482956, + 0.07948014885187149, + 0.005484199151396751, + 0.008551575243473053, + 0.0037812504451721907, + 0.003227287670597434, + 0.002305764239281416, + 0.0030528083443641663, + 0.004295774735510349, + 0.0046326881274580956, + 0.002671621972694993, + 0.0025417134165763855, + 0.0039012590423226357, + 0.004220741800963879, + 0.0025571000296622515, + 0.004196952562779188, + 0.003658271860331297, + 0.005247971974313259, + 0.0024135562125593424, + 0.002965491497889161, + 0.009108614176511765, + 0.004890334792435169, + 0.003319117473438382, + 0.005248044617474079, + 0.00798086915165186, + 0.012170492671430111, + 0.004513985943049192, + 0.004340460058301687, + 0.007924306206405163, + 0.004466983489692211, + 0.00419692974537611, + 0.004094831179827452, + 0.005598495248705149, + 0.01406931970268488, + 0.0038292037788778543, + 0.005986606702208519, + 0.007905801758170128, + 0.007331353146582842, + 0.007519155275076628, + 0.0076902806758880615, + 0.009340374730527401, + 0.012811971828341484, + 0.009288579225540161, + 0.011201220564544201, + 0.008373481221497059, + 0.012086720205843449, + 0.007141981739550829, + 0.009999389760196209, + 0.016087274998426437, + 0.012024170719087124, + 0.010918937623500824, + 0.007331566885113716, + 0.013193277642130852, + 0.011574281379580498, + 0.013677805662155151, + 0.011465819552540779, + 0.016644850373268127, + 0.021101634949445724, + 0.014382532797753811, + 0.011074982583522797, + 0.029699821025133133, + 0.036890577524900436, + 0.037116389721632004, + 0.03806750103831291, + 0.0397101566195488, + 0.04075810685753822, + 0.05711027607321739, + 0.03207199275493622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.016064966097474098 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot shot from a horizontal angle centers on a playful crayon drawing on slightly textured white paper. The painting, rendered in highly saturated, rough-brushed crayons, depicts a simple story: a large, smiling yellow sun hangs high in the sky, beneath a small blue house with a red triangular roof and square windows, and beside it, a waving stick figure. The entire scene exudes an atmosphere of childlike joy. Suddenly, a sharp, sharply defined black shadow pierces the frame from beyond, poking precisely at the center. Immediately, the crayon-textured drawing, like a popped balloon, begins to rapidly sag inward from the point of impact. The once-flat surface of the paper instantly loses its tension, and the sun, house, and stick figure in the painting wrinkle and distort painfully as the image shrinks. The vibrant colors fade and blur, and ultimately, the entire painting collapses like a deflated balloon, shrunken and contracting into a ball, completely losing its original shape and story.", + "seed": 19981, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_6/index_5643_seed_19981_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024361598509131, + "vlm_logprob": -0.024361598509131, + "vlm_yes_no_margin": 5.4, + "vlm_token_credit": -0.024361598509131, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013416602276265621 + ], + "opposite_logprob_tokens": [ + -5.638416767120361 + ], + "logprob_reward": -0.013416602276265621, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.013416602276265621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19148372113704681, + 0.10332400351762772, + 0.05029032751917839, + 0.07648369669914246, + 0.03288623318076134, + 0.07460581511259079, + 0.16627207398414612, + 0.08356168866157532, + 0.040386803448200226, + 0.18070565164089203 + ], + "credit_map": [ + 0.02386079542338848, + 0.021966680884361267, + 0.037879422307014465, + 0.0204131118953228, + 0.017198016867041588, + 0.046920910477638245, + 0.014575782231986523, + 0.008668998256325722, + 0.01523752324283123, + 0.01053133700042963, + 0.020746350288391113, + 0.015526894479990005, + 0.0100752804428339, + 0.018736181780695915, + 0.007332966197282076, + 0.005137468222528696, + 0.005135674960911274, + 0.005770418327301741, + 0.0074664270505309105, + 0.007886677980422974, + 0.005389881320297718, + 0.010267265141010284, + 0.00522609194740653, + 0.0031478966120630503, + 0.007267762906849384, + 0.011791909113526344, + 0.007000952027738094, + 0.010300932452082634, + 0.011131702922284603, + 0.01596374623477459, + 0.00792575255036354, + 0.005100936628878117, + 0.0050921873189508915, + 0.003727375529706478, + 0.004255829378962517, + 0.003178807906806469, + 0.005799958482384682, + 0.004695011302828789, + 0.0037873596884310246, + 0.002349702874198556, + 0.006046804133802652, + 0.008000490255653858, + 0.008002158254384995, + 0.006526926998049021, + 0.021227749064564705, + 0.008813469670712948, + 0.009021648205816746, + 0.006966566666960716, + 0.013616519048810005, + 0.019594380632042885, + 0.011119968257844448, + 0.01068609207868576, + 0.02413320355117321, + 0.049476537853479385, + 0.008857048116624355, + 0.028788326308131218, + 0.013127341866493225, + 0.006857829634100199, + 0.011766579002141953, + 0.011398331262171268, + 0.012957376427948475, + 0.0071472530253231525, + 0.010768567211925983, + 0.009538409300148487, + 0.007065155543386936, + 0.0031844498589634895, + 0.006268173456192017, + 0.0037799151614308357, + 0.005306767765432596, + 0.0030575068667531013, + 0.006368459668010473, + 0.005356375593692064, + 0.04068080335855484, + 0.019375672563910484, + 0.029613811522722244, + 0.015623507089912891, + 0.012736789882183075, + 0.024093417450785637, + 0.01718173734843731, + 0.021399900317192078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.013416602276265621 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904742449522018 + ], + "opposite_logprob_tokens": [ + -2.464047431945801 + ], + "logprob_reward": -0.08904742449522018, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904742449522018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06374360620975494, + 0.05201127007603645, + 0.030763329938054085, + 0.035379476845264435, + 0.06724537163972855, + 0.08567526191473007, + 0.299167662858963, + 0.1011601909995079, + 0.052569977939128876, + 0.21228381991386414 + ], + "credit_map": [ + 0.007364643272012472, + 0.008126762695610523, + 0.0071013509295880795, + 0.00881427526473999, + 0.009752186946570873, + 0.009800994768738747, + 0.009191799908876419, + 0.00359159498475492, + 0.006364238914102316, + 0.005018956959247589, + 0.007384235505014658, + 0.010211285203695297, + 0.0069281854666769505, + 0.005964110605418682, + 0.0067822979763150215, + 0.003357955487444997, + 0.004410682246088982, + 0.003191667376086116, + 0.0039052546489983797, + 0.004236229229718447, + 0.005388340447098017, + 0.004124651197344065, + 0.0034503599163144827, + 0.00205614254809916, + 0.0029679625295102596, + 0.0035105720162391663, + 0.005000317469239235, + 0.004766111262142658, + 0.004241366405040026, + 0.00864128116518259, + 0.004134511109441519, + 0.0021173562854528427, + 0.007010574918240309, + 0.007297818083316088, + 0.009570750407874584, + 0.010314052924513817, + 0.010040419176220894, + 0.01144528016448021, + 0.007852749899029732, + 0.0037137314211577177, + 0.0070067173801362514, + 0.009750047698616982, + 0.010840930975973606, + 0.00847307313233614, + 0.019581343978643417, + 0.010742342099547386, + 0.010854263789951801, + 0.008426543325185776, + 0.018043309450149536, + 0.03628585860133171, + 0.041240692138671875, + 0.011908750049769878, + 0.025644943118095398, + 0.09101230651140213, + 0.015010182745754719, + 0.06002165377140045, + 0.01888287626206875, + 0.009641164913773537, + 0.015530097298324108, + 0.010959058068692684, + 0.014976397156715393, + 0.008730205707252026, + 0.010690102353692055, + 0.011750287376344204, + 0.007492280099540949, + 0.003979201894253492, + 0.008567256852984428, + 0.004669304471462965, + 0.008669707924127579, + 0.0050590792670845985, + 0.006946119945496321, + 0.007187025621533394, + 0.04181251302361488, + 0.031079184263944626, + 0.031272515654563904, + 0.021556884050369263, + 0.016278104856610298, + 0.025341611355543137, + 0.017300620675086975, + 0.027642391622066498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.08904742449522018 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000442407006630674 + ], + "opposite_logprob_tokens": [ + -7.7504425048828125 + ], + "logprob_reward": -0.000442407006630674, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.000442407006630674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10856557637453079, + 0.1402587890625, + 0.12741298973560333, + 0.09159568697214127, + 0.08376073092222214, + 0.10484807938337326, + 0.16824780404567719, + 0.05401962995529175, + 0.039198026061058044, + 0.08209267258644104 + ], + "credit_map": [ + 0.015927977859973907, + 0.014299063012003899, + 0.00940412562340498, + 0.011555536650121212, + 0.011002624407410622, + 0.022575821727514267, + 0.016182269901037216, + 0.00761816231533885, + 0.015356969088315964, + 0.01785215362906456, + 0.023043904453516006, + 0.017480365931987762, + 0.013112720102071762, + 0.03742919862270355, + 0.009055931121110916, + 0.006927546113729477, + 0.010963755659759045, + 0.016793953254818916, + 0.01716291718184948, + 0.02009330317378044, + 0.012424901127815247, + 0.03222091868519783, + 0.010995493270456791, + 0.006757751107215881, + 0.008341915905475616, + 0.014100617729127407, + 0.009875140152871609, + 0.01588401012122631, + 0.015405197627842426, + 0.013109722174704075, + 0.009170875884592533, + 0.0057082148268818855, + 0.007719562854617834, + 0.00982076395303011, + 0.013389850966632366, + 0.013834520243108273, + 0.007332894019782543, + 0.020692724734544754, + 0.006752863060683012, + 0.004217551089823246, + 0.008234681561589241, + 0.012059514410793781, + 0.0102412523701787, + 0.012889997102320194, + 0.023286398500204086, + 0.01272281352430582, + 0.015937380492687225, + 0.00947604514658451, + 0.012260370887815952, + 0.01912497729063034, + 0.017587751150131226, + 0.013685742393136024, + 0.01777886413037777, + 0.04506121575832367, + 0.015946201980113983, + 0.026802677661180496, + 0.006910066120326519, + 0.004525482654571533, + 0.008410235866904259, + 0.00752127543091774, + 0.008218486793339252, + 0.005512270610779524, + 0.00702940858900547, + 0.00589241087436676, + 0.005950800608843565, + 0.002786759752780199, + 0.006679957266896963, + 0.0034733719658106565, + 0.007847541943192482, + 0.002491909312084317, + 0.005056959111243486, + 0.004910729825496674, + 0.010888161137700081, + 0.01207701861858368, + 0.010518291965126991, + 0.006758386734873056, + 0.010535363107919693, + 0.013246666640043259, + 0.008362921886146069, + 0.009705860167741776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.000442407006630674 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007192885386757553 + ], + "opposite_logprob_tokens": [ + -7.25071907043457 + ], + "logprob_reward": -0.0007192885386757553, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007192885386757553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03262621909379959, + 0.08113608509302139, + 0.044132549315690994, + 0.009303119033575058, + 0.023479389026761055, + 0.1183488741517067, + 0.3381181061267853, + 0.14841780066490173, + 0.061835408210754395, + 0.14260250329971313 + ], + "credit_map": [ + 0.004244162701070309, + 0.0037046128418296576, + 0.004735371097922325, + 0.003144790418446064, + 0.004233019892126322, + 0.005257992073893547, + 0.005078704096376896, + 0.0022275650408118963, + 0.008355212397873402, + 0.008699839003384113, + 0.01554439403116703, + 0.010086064226925373, + 0.008167480118572712, + 0.020683957263827324, + 0.005946826655417681, + 0.003652313956990838, + 0.00522072846069932, + 0.00502906646579504, + 0.007487303577363491, + 0.005353888496756554, + 0.004896592814475298, + 0.01036158762872219, + 0.003701840527355671, + 0.0020815415773540735, + 0.001623552176170051, + 0.0016374402912333608, + 0.0010735800024122, + 0.0008056350634433329, + 0.001040792092680931, + 0.0016573206521570683, + 0.0008966434979811311, + 0.0005681548500433564, + 0.0029876139014959335, + 0.0037207775749266148, + 0.003193001262843609, + 0.0030433847568929195, + 0.0031122821383178234, + 0.003728550858795643, + 0.002478074748069048, + 0.0012157021556049585, + 0.010824996046721935, + 0.012208175845444202, + 0.008045745082199574, + 0.006184740923345089, + 0.040807317942380905, + 0.013969199731945992, + 0.018638314679265022, + 0.007670382037758827, + 0.031233349815011024, + 0.039785586297512054, + 0.020945463329553604, + 0.010945742949843407, + 0.046804074198007584, + 0.10730120539665222, + 0.0180963147431612, + 0.06300635635852814, + 0.0317845344543457, + 0.011194279417395592, + 0.01684747077524662, + 0.018360288813710213, + 0.02082056738436222, + 0.01708424836397171, + 0.01203172653913498, + 0.0202946700155735, + 0.012182948179543018, + 0.004418616183102131, + 0.011175919324159622, + 0.005169578827917576, + 0.00923896674066782, + 0.005159266758710146, + 0.005832016468048096, + 0.008658094331622124, + 0.021157193928956985, + 0.02116408571600914, + 0.017481697723269463, + 0.012218456715345383, + 0.01587301678955555, + 0.02181035652756691, + 0.01753988303244114, + 0.015357812866568565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0007192885386757553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182270228862762 + ], + "opposite_logprob_tokens": [ + -4.018182277679443 + ], + "logprob_reward": -0.018182270228862762, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018182270228862762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02644326351583004, + 0.03238861262798309, + 0.03242756053805351, + 0.013305415399372578, + 0.040429145097732544, + 0.2145070880651474, + 0.23285970091819763, + 0.04853835329413414, + 0.10581853240728378, + 0.253282368183136 + ], + "credit_map": [ + 0.004459848161786795, + 0.003524464787915349, + 0.002718204166740179, + 0.0028057789895683527, + 0.004301679786294699, + 0.0032728395890444517, + 0.0038610913325101137, + 0.0014993553049862385, + 0.004077525343745947, + 0.0032445346005260944, + 0.0068321372382342815, + 0.0045271990820765495, + 0.0031816023401916027, + 0.006759052164852619, + 0.0023046580608934164, + 0.0014619015855714679, + 0.004082245286554098, + 0.003391623729839921, + 0.0058638849295675755, + 0.004616404417902231, + 0.0035781969781965017, + 0.007147742435336113, + 0.0023211415391415358, + 0.0014263245975598693, + 0.0020388606935739517, + 0.002206707140430808, + 0.0018728377763181925, + 0.0013193492777645588, + 0.0013195901410654187, + 0.0027347621507942677, + 0.0011607802007347345, + 0.000652527844067663, + 0.003859591903164983, + 0.003988356329500675, + 0.007497035898268223, + 0.006889622192829847, + 0.00465269573032856, + 0.007084283512085676, + 0.004097513854503632, + 0.0023600466083735228, + 0.017924422398209572, + 0.024850400164723396, + 0.026132913306355476, + 0.017816422507166862, + 0.04767419770359993, + 0.03523615375161171, + 0.025994230061769485, + 0.01887834072113037, + 0.0133991539478302, + 0.024751340970396996, + 0.023016581311821938, + 0.01678811013698578, + 0.02498563937842846, + 0.06819996237754822, + 0.013086634688079357, + 0.04863228276371956, + 0.007596122100949287, + 0.0033457945100963116, + 0.007354817818850279, + 0.005889316089451313, + 0.008781437762081623, + 0.0052357264794409275, + 0.004638418089598417, + 0.005696720909327269, + 0.02526002563536167, + 0.005183451343327761, + 0.020184941589832306, + 0.01085397507995367, + 0.00817347876727581, + 0.007609239779412746, + 0.010007676668465137, + 0.018545737490057945, + 0.03146585822105408, + 0.03389299288392067, + 0.03402063623070717, + 0.023136137053370476, + 0.036681488156318665, + 0.03910814970731735, + 0.03347319737076759, + 0.021503901109099388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.018182270228862762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot shot from a horizontal angle centers on a playful crayon drawing on slightly textured white paper. The painting, rendered in highly saturated, rough-brushed crayons, depicts a simple story: a large, smiling yellow sun hangs high in the sky, beneath a small blue house with a red triangular roof and square windows, and beside it, a waving stick figure. The entire scene exudes an atmosphere of childlike joy. Suddenly, a sharp, sharply defined black shadow pierces the frame from beyond, poking precisely at the center. Immediately, the crayon-textured drawing, like a popped balloon, begins to rapidly sag inward from the point of impact. The once-flat surface of the paper instantly loses its tension, and the sun, house, and stick figure in the painting wrinkle and distort painfully as the image shrinks. The vibrant colors fade and blur, and ultimately, the entire painting collapses like a deflated balloon, shrunken and contracting into a ball, completely losing its original shape and story.", + "seed": 19982, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_6/index_5643_seed_19982_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025311199948191643, + "vlm_logprob": -0.025311199948191643, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.025311199948191643, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011381343007087708 + ], + "opposite_logprob_tokens": [ + -5.386381149291992 + ], + "logprob_reward": -0.011381343007087708, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.011381343007087708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09308518469333649, + 0.058862678706645966, + 0.04489261656999588, + 0.0657835379242897, + 0.06326600909233093, + 0.024414410814642906, + 0.11251386255025864, + 0.31031855940818787, + 0.08472011983394623, + 0.14214304089546204 + ], + "credit_map": [ + 0.016979936510324478, + 0.008363611064851284, + 0.018368538469076157, + 0.006715035531669855, + 0.0073129404336214066, + 0.0186449084430933, + 0.009050603955984116, + 0.007649606093764305, + 0.01678418181836605, + 0.004681464284658432, + 0.006393996998667717, + 0.0034415379632264376, + 0.0032608117908239365, + 0.013831904157996178, + 0.005658561363816261, + 0.004810219630599022, + 0.008246324956417084, + 0.0037365222815424204, + 0.008761199191212654, + 0.002741120057180524, + 0.003204927546903491, + 0.011894138529896736, + 0.0024357473012059927, + 0.0038726371712982655, + 0.019589802250266075, + 0.0061330171301960945, + 0.010949884541332722, + 0.00497072609141469, + 0.004654260817915201, + 0.00615164078772068, + 0.005267898552119732, + 0.008066311478614807, + 0.008465868420898914, + 0.006771945394575596, + 0.011721758171916008, + 0.008415991440415382, + 0.00589785585179925, + 0.00823510717600584, + 0.007631130516529083, + 0.006126353982836008, + 0.003961429465562105, + 0.003042543074116111, + 0.004539694171398878, + 0.0026513421908020973, + 0.0020312743727117777, + 0.002727764192968607, + 0.0029683609027415514, + 0.002491999650374055, + 0.025646984577178955, + 0.009575084783136845, + 0.02000323310494423, + 0.01613771542906761, + 0.005771217867732048, + 0.009728077799081802, + 0.015358667820692062, + 0.010292871855199337, + 0.06656047701835632, + 0.023594418540596962, + 0.06175042316317558, + 0.03754524141550064, + 0.01722106523811817, + 0.030280685052275658, + 0.045774444937705994, + 0.02759181708097458, + 0.022164948284626007, + 0.008341755717992783, + 0.01853131875395775, + 0.006993861403316259, + 0.007153165526688099, + 0.009598558768630028, + 0.00490537891164422, + 0.00703113479539752, + 0.04401730000972748, + 0.011383925564587116, + 0.009978009387850761, + 0.006657114718109369, + 0.006507705431431532, + 0.05089486762881279, + 0.005021127872169018, + 0.007682994473725557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.011381343007087708 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002148032188416 + ], + "opposite_logprob_tokens": [ + -2.695021390914917 + ], + "logprob_reward": -0.07002148032188416, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002148032188416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16066400706768036, + 0.06935396790504456, + 0.03965354710817337, + 0.0629737600684166, + 0.03350341320037842, + 0.06803177297115326, + 0.05384824424982071, + 0.055351559072732925, + 0.17150628566741943, + 0.2851135730743408 + ], + "credit_map": [ + 0.05755988508462906, + 0.010331647470593452, + 0.014540368691086769, + 0.006744335871189833, + 0.005846675485372543, + 0.04967253655195236, + 0.00788157619535923, + 0.00808697659522295, + 0.024846214801073074, + 0.004649611655622721, + 0.008987605571746826, + 0.0032151914201676846, + 0.003681889735162258, + 0.01723008044064045, + 0.0037886102218180895, + 0.0029547663871198893, + 0.009365874342620373, + 0.0037987481337040663, + 0.008938292041420937, + 0.003222133032977581, + 0.003300467738881707, + 0.005632753483951092, + 0.0022733428049832582, + 0.0031219327356666327, + 0.017516911029815674, + 0.006635877303779125, + 0.008195856586098671, + 0.0046216570772230625, + 0.004276070278137922, + 0.006653175689280033, + 0.0054181101731956005, + 0.009656100533902645, + 0.0070444131270051, + 0.0038829995319247246, + 0.004225573036819696, + 0.003495238022878766, + 0.0031082017812877893, + 0.0038838088512420654, + 0.0038522963877767324, + 0.004010880831629038, + 0.01497801672667265, + 0.007017907686531544, + 0.011679606512188911, + 0.008360682986676693, + 0.004106036387383938, + 0.006056636571884155, + 0.008137664757668972, + 0.007695222273468971, + 0.006463357713073492, + 0.005460325162857771, + 0.00901888869702816, + 0.006571836769580841, + 0.005048386286944151, + 0.006983114406466484, + 0.0074057625606656075, + 0.00689657311886549, + 0.010976923629641533, + 0.0047705210745334625, + 0.011151885613799095, + 0.005417420994490385, + 0.00355763197876513, + 0.0056799957528710365, + 0.007364389952272177, + 0.006432786583900452, + 0.04857347160577774, + 0.013875116594135761, + 0.030733438208699226, + 0.016694169491529465, + 0.010735996998846531, + 0.02667611464858055, + 0.013177500106394291, + 0.011040469631552696, + 0.023639701306819916, + 0.022607125341892242, + 0.09361688792705536, + 0.015915365889668465, + 0.013893990777432919, + 0.09014397114515305, + 0.0104330163449049, + 0.014863533899188042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.07002148032188416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028228219598531723 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.0028228219598531723, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028228219598531723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23635737597942352, + 0.10097010433673859, + 0.040585026144981384, + 0.061971455812454224, + 0.03317389637231827, + 0.03846072033047676, + 0.03919147700071335, + 0.05048062279820442, + 0.1272415667772293, + 0.27156776189804077 + ], + "credit_map": [ + 0.0752813071012497, + 0.016019029542803764, + 0.03299541398882866, + 0.012009352445602417, + 0.011416391469538212, + 0.0553748793900013, + 0.01768585480749607, + 0.015575158409774303, + 0.01921558938920498, + 0.009205687791109085, + 0.01333153061568737, + 0.008915293030440807, + 0.006026690360158682, + 0.013453390449285507, + 0.020084168761968613, + 0.010737755335867405, + 0.005056172143667936, + 0.003970968071371317, + 0.0070695094764232635, + 0.003760248888283968, + 0.0034137777984142303, + 0.007794354576617479, + 0.005982624366879463, + 0.0035373675636947155, + 0.019179826602339745, + 0.005498833954334259, + 0.011198896914720535, + 0.003969454672187567, + 0.0035439126659184694, + 0.00682411715388298, + 0.0038254570681601763, + 0.007930953986942768, + 0.0045193154364824295, + 0.0036164673510938883, + 0.005293056834489107, + 0.00544193759560585, + 0.0028756267856806517, + 0.003377809887751937, + 0.004906975664198399, + 0.003142708446830511, + 0.005352160427719355, + 0.004822339862585068, + 0.008040640503168106, + 0.003490941133350134, + 0.002926974091678858, + 0.004242222756147385, + 0.002946442225947976, + 0.006639002822339535, + 0.006766581907868385, + 0.004885593429207802, + 0.005713256541639566, + 0.003554883413016796, + 0.0029983266722410917, + 0.003904258366674185, + 0.003948930650949478, + 0.007419645320624113, + 0.008773448877036572, + 0.004352639429271221, + 0.009083815850317478, + 0.006711846217513084, + 0.0036210254766047, + 0.006758158095180988, + 0.00581678980961442, + 0.005362899973988533, + 0.040880195796489716, + 0.01282582152634859, + 0.03317468613386154, + 0.00798546802252531, + 0.0070716324262320995, + 0.010977457277476788, + 0.0061645787209272385, + 0.008161739446222782, + 0.0715455710887909, + 0.015529630705714226, + 0.10044524073600769, + 0.013795347884297371, + 0.011493254452943802, + 0.035920463502407074, + 0.009718868881464005, + 0.013119390234351158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0028228219598531723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537354603409767 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537354603409767, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537354603409767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06927667558193207, + 0.031203361228108406, + 0.03623669221997261, + 0.02567223459482193, + 0.02220040187239647, + 0.11490330845117569, + 0.04901716485619545, + 0.059005629271268845, + 0.19293367862701416, + 0.39955082535743713 + ], + "credit_map": [ + 0.012627394869923592, + 0.007463351357728243, + 0.010054457001388073, + 0.00463017076253891, + 0.005248401779681444, + 0.017185265198349953, + 0.007003230042755604, + 0.005064402241259813, + 0.0055592902936041355, + 0.002523857168853283, + 0.004074194934219122, + 0.002308201976120472, + 0.002193029038608074, + 0.006442434620112181, + 0.005457044113427401, + 0.0026453109458088875, + 0.005999154411256313, + 0.0031301339622586966, + 0.007673714775592089, + 0.0023363539949059486, + 0.0029111201874911785, + 0.008060338906943798, + 0.0031972608994692564, + 0.002928613917902112, + 0.003722471185028553, + 0.004008189309388399, + 0.0035410320851951838, + 0.002677494427189231, + 0.002730113919824362, + 0.002995579270645976, + 0.0032924702391028404, + 0.002704883925616741, + 0.004125779028981924, + 0.0019537026528269053, + 0.003650355851277709, + 0.00241983775049448, + 0.0014414233155548573, + 0.003418359672650695, + 0.002352768322452903, + 0.0028381748124957085, + 0.03129279240965843, + 0.011076410301029682, + 0.014283403754234314, + 0.012820282019674778, + 0.006174826063215733, + 0.011583631858229637, + 0.009626363404095173, + 0.01804560422897339, + 0.01073602493852377, + 0.004918355494737625, + 0.006815748754888773, + 0.005907509475946426, + 0.0030958866700530052, + 0.004996456205844879, + 0.006145469844341278, + 0.0064017134718596935, + 0.008004093542695045, + 0.006195227615535259, + 0.009977303445339203, + 0.008423964492976665, + 0.004830693360418081, + 0.0052882591262459755, + 0.007527990732342005, + 0.00875809881836176, + 0.04321848601102829, + 0.020550312474370003, + 0.03589987754821777, + 0.019989311695098877, + 0.01738668605685234, + 0.014759009703993797, + 0.01922626793384552, + 0.021903706714510918, + 0.06455279141664505, + 0.02561960369348526, + 0.13641519844532013, + 0.02531040832400322, + 0.02310289815068245, + 0.08474191278219223, + 0.018376067280769348, + 0.021431945264339447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.012537354603409767 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979299984872341 + ], + "opposite_logprob_tokens": [ + -3.5297930240631104 + ], + "logprob_reward": -0.02979299984872341, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979299984872341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047780461609363556, + 0.017030343413352966, + 0.023495879024267197, + 0.028934920206665993, + 0.021504152566194534, + 0.08207903802394867, + 0.034969210624694824, + 0.15615834295749664, + 0.0938921719789505, + 0.49415549635887146 + ], + "credit_map": [ + 0.0100172134116292, + 0.005087264347821474, + 0.0068180738016963005, + 0.002977986354380846, + 0.003600972704589367, + 0.011322437785565853, + 0.004323589149862528, + 0.003632929176092148, + 0.0020477951038628817, + 0.0020636857952922583, + 0.0023213226813822985, + 0.0014769124099984765, + 0.0013799798907712102, + 0.0029936302453279495, + 0.0030846912413835526, + 0.0016623286064714193, + 0.004038001876324415, + 0.0019134975736960769, + 0.0051283142529428005, + 0.0015487737255170941, + 0.001981265377253294, + 0.00523410364985466, + 0.0016308285994455218, + 0.0020210961811244488, + 0.004513357300311327, + 0.0028678979724645615, + 0.004394226241856813, + 0.003857035655528307, + 0.0025335047394037247, + 0.0037988126277923584, + 0.0032318453304469585, + 0.0037382461596280336, + 0.0032667468767613173, + 0.002272683894261718, + 0.005549360532313585, + 0.0023817785549908876, + 0.001738406834192574, + 0.0023064392153173685, + 0.001993192359805107, + 0.0019955451134592295, + 0.023421838879585266, + 0.008405184373259544, + 0.008530685678124428, + 0.007565719541162252, + 0.0046063936315476894, + 0.008304469287395477, + 0.006701247300952673, + 0.014543508179485798, + 0.005758336279541254, + 0.0042307376861572266, + 0.00472244480624795, + 0.0043071722611784935, + 0.002468820195645094, + 0.0035402902867645025, + 0.004715041257441044, + 0.005226371809840202, + 0.03249851241707802, + 0.012445361353456974, + 0.030350767076015472, + 0.02051505632698536, + 0.009207752533257008, + 0.015199578367173672, + 0.02080872282385826, + 0.015132601372897625, + 0.011822887696325779, + 0.011300927959382534, + 0.011433031409978867, + 0.01044828724116087, + 0.010729249566793442, + 0.011210084892809391, + 0.011597596108913422, + 0.015350119210779667, + 0.18670371174812317, + 0.031153246760368347, + 0.06391961872577667, + 0.019870024174451828, + 0.019182369112968445, + 0.13512030243873596, + 0.017404599115252495, + 0.020801648497581482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.02979299984872341 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot shot from a horizontal angle centers on a playful crayon drawing on slightly textured white paper. The painting, rendered in highly saturated, rough-brushed crayons, depicts a simple story: a large, smiling yellow sun hangs high in the sky, beneath a small blue house with a red triangular roof and square windows, and beside it, a waving stick figure. The entire scene exudes an atmosphere of childlike joy. Suddenly, a sharp, sharply defined black shadow pierces the frame from beyond, poking precisely at the center. Immediately, the crayon-textured drawing, like a popped balloon, begins to rapidly sag inward from the point of impact. The once-flat surface of the paper instantly loses its tension, and the sun, house, and stick figure in the painting wrinkle and distort painfully as the image shrinks. The vibrant colors fade and blur, and ultimately, the entire painting collapses like a deflated balloon, shrunken and contracting into a ball, completely losing its original shape and story.", + "seed": 19983, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_6/index_5643_seed_19983_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0318828874733299, + "vlm_logprob": -0.0318828874733299, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.0318828874733299, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03738730028271675 + ], + "opposite_logprob_tokens": [ + -3.6623873710632324 + ], + "logprob_reward": -0.03738730028271675, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03738730028271675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12161094695329666, + 0.01687481440603733, + 0.021064186468720436, + 0.016216030344367027, + 0.04245547577738762, + 0.570173442363739, + 0.06605923920869827, + 0.03140992298722267, + 0.05509096011519432, + 0.059045035392045975 + ], + "credit_map": [ + 0.01452645380049944, + 0.030877001583576202, + 0.00984891876578331, + 0.008281489834189415, + 0.009677101857960224, + 0.022918716073036194, + 0.008079593069851398, + 0.017401672899723053, + 0.0026659639552235603, + 0.003528719302266836, + 0.001262826845049858, + 0.001160274725407362, + 0.00183659337926656, + 0.0030916212126612663, + 0.0015781676629558206, + 0.001750646741129458, + 0.0024655249435454607, + 0.003728702897205949, + 0.0020607560873031616, + 0.0014234990812838078, + 0.003637232119217515, + 0.003443601541221142, + 0.002043111715465784, + 0.002261756919324398, + 0.003152600722387433, + 0.0019386002095416188, + 0.0022925371304154396, + 0.0011446551652625203, + 0.002271227305755019, + 0.0022789249196648598, + 0.0018669357523322105, + 0.0012705489061772823, + 0.0050172931514680386, + 0.0073361024260520935, + 0.0039807334542274475, + 0.002748076105490327, + 0.004652152769267559, + 0.00694769062101841, + 0.0043246387504041195, + 0.007448787335306406, + 0.06291540712118149, + 0.062146611511707306, + 0.13529440760612488, + 0.02648472599685192, + 0.04432147368788719, + 0.19353125989437103, + 0.017465610057115555, + 0.028013914823532104, + 0.0038069779984652996, + 0.0036673510912805796, + 0.0049797650426626205, + 0.008765675127506256, + 0.013602680526673794, + 0.010796996764838696, + 0.007149549666792154, + 0.01329024787992239, + 0.004471269436180592, + 0.0028652697801589966, + 0.003907276317477226, + 0.0028163015376776457, + 0.005910462234169245, + 0.005541714373975992, + 0.00342894671484828, + 0.0024686846882104874, + 0.009053264744579792, + 0.00894839596003294, + 0.004731796681880951, + 0.005016207695007324, + 0.005477011203765869, + 0.009501396678388119, + 0.006209034472703934, + 0.006153852213174105, + 0.011072162538766861, + 0.00657673692330718, + 0.007025295868515968, + 0.006289241369813681, + 0.007119606249034405, + 0.009642759338021278, + 0.006530329585075378, + 0.004788908176124096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.03738730028271675 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295555531978607 + ], + "opposite_logprob_tokens": [ + -2.2379555702209473 + ], + "logprob_reward": -0.11295555531978607, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295555531978607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060131609439849854, + 0.018940260633826256, + 0.030472608283162117, + 0.10500022768974304, + 0.027370069175958633, + 0.45992401242256165, + 0.13846652209758759, + 0.05656593665480614, + 0.07055617868900299, + 0.03257260471582413 + ], + "credit_map": [ + 0.007569014094769955, + 0.016972871497273445, + 0.0034033777192234993, + 0.002705896971747279, + 0.007634790614247322, + 0.011987127363681793, + 0.005683046765625477, + 0.004175483249127865, + 0.0023358396720141172, + 0.0038451014552265406, + 0.0020825753454118967, + 0.0018703859532251954, + 0.0016273555811494589, + 0.0030441870912909508, + 0.0018073029350489378, + 0.002327514113858342, + 0.004441613797098398, + 0.007197519764304161, + 0.0019278406398370862, + 0.0013257480459287763, + 0.0038946433924138546, + 0.005373089108616114, + 0.004268931690603495, + 0.0020432209130376577, + 0.019408447667956352, + 0.011276088654994965, + 0.011996527202427387, + 0.007280728314071894, + 0.019740436226129532, + 0.01525493711233139, + 0.013012917712330818, + 0.007030146196484566, + 0.004595932085067034, + 0.003067486919462681, + 0.003996009007096291, + 0.0025092046707868576, + 0.003662385744974017, + 0.003287644125521183, + 0.0027206058148294687, + 0.0035307996440678835, + 0.04835832864046097, + 0.0333339162170887, + 0.1441280096769333, + 0.022305376827716827, + 0.03603601083159447, + 0.14137670397758484, + 0.013184732757508755, + 0.02120092511177063, + 0.00854992214590311, + 0.005891292355954647, + 0.008186059072613716, + 0.010717479512095451, + 0.036971382796764374, + 0.024716513231396675, + 0.022000638768076897, + 0.021433237940073013, + 0.007351750507950783, + 0.005139043089002371, + 0.004651119466871023, + 0.0032725476194173098, + 0.01536551769822836, + 0.011158903129398823, + 0.004091260489076376, + 0.0055357892997562885, + 0.01086142472922802, + 0.0068444665521383286, + 0.006258248817175627, + 0.009015670977532864, + 0.00971164833754301, + 0.011176575906574726, + 0.011116997338831425, + 0.005571146495640278, + 0.003020223695784807, + 0.0032161674462258816, + 0.004133252426981926, + 0.00441773934289813, + 0.005414739716798067, + 0.004757710732519627, + 0.0036750866565853357, + 0.003937684465199709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.11295555531978607 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008245882345363498 + ], + "opposite_logprob_tokens": [ + -7.125824451446533 + ], + "logprob_reward": -0.0008245882345363498, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008245882345363498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09467491507530212, + 0.051410507410764694, + 0.11056890338659286, + 0.0557231605052948, + 0.038184527307748795, + 0.3195616900920868, + 0.050316184759140015, + 0.06445763260126114, + 0.13117721676826477, + 0.08392523974180222 + ], + "credit_map": [ + 0.011423888616263866, + 0.01988038793206215, + 0.006840001326054335, + 0.00683538056910038, + 0.010120709426701069, + 0.015862755477428436, + 0.009365187026560307, + 0.014346607960760593, + 0.006301541812717915, + 0.0067927660420536995, + 0.005301819648593664, + 0.005082572810351849, + 0.004628065042197704, + 0.008302777074277401, + 0.004666108172386885, + 0.010334858670830727, + 0.017829399555921555, + 0.015804901719093323, + 0.017144454643130302, + 0.010697271674871445, + 0.013820472173392773, + 0.01667536236345768, + 0.008071874268352985, + 0.010525167919695377, + 0.014908530749380589, + 0.007080581970512867, + 0.009219557978212833, + 0.004479260183870792, + 0.006551425438374281, + 0.0065027023665606976, + 0.0035765748471021652, + 0.0034045306965708733, + 0.004819945432245731, + 0.005114806815981865, + 0.004654113203287125, + 0.0038245227187871933, + 0.005553788971155882, + 0.005685438867658377, + 0.004449095111340284, + 0.004082813858985901, + 0.037648823112249374, + 0.1004510372877121, + 0.016653602942824364, + 0.010363358072936535, + 0.018928291276097298, + 0.10916168987751007, + 0.01085793599486351, + 0.015496970154345036, + 0.0052126566879451275, + 0.003060983493924141, + 0.007493938785046339, + 0.008648599497973919, + 0.006061881314963102, + 0.0071262177079916, + 0.0033152932301163673, + 0.009396613575518131, + 0.006123605649918318, + 0.005436772480607033, + 0.009389999322593212, + 0.00599751016125083, + 0.01487190742045641, + 0.01202764455229044, + 0.006608254741877317, + 0.004001936875283718, + 0.011210627853870392, + 0.010084669105708599, + 0.008563431911170483, + 0.016272390261292458, + 0.02201959677040577, + 0.019140038639307022, + 0.02705114707350731, + 0.01683531142771244, + 0.007233964744955301, + 0.005600479431450367, + 0.007952174171805382, + 0.01207686122506857, + 0.015575170516967773, + 0.010756593197584152, + 0.011275672353804111, + 0.013454322703182697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.0008245882345363498 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015136463334783912 + ], + "opposite_logprob_tokens": [ + -6.501513481140137 + ], + "logprob_reward": -0.0015136463334783912, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015136463334783912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05563235282897949, + 0.03355967998504639, + 0.0553651861846447, + 0.020382864400744438, + 0.05927174538373947, + 0.1975487619638443, + 0.05207296460866928, + 0.048583608120679855, + 0.09341786801815033, + 0.384164959192276 + ], + "credit_map": [ + 0.010718563571572304, + 0.012857099995017052, + 0.004807100165635347, + 0.0047537945210933685, + 0.004612917546182871, + 0.009700385853648186, + 0.003964112140238285, + 0.004218379966914654, + 0.006070393603295088, + 0.004788801074028015, + 0.0038796882145106792, + 0.003496646648272872, + 0.0031926438678056, + 0.005148118827491999, + 0.0038378702010959387, + 0.0031455177813768387, + 0.006202835123986006, + 0.007970941253006458, + 0.0037780003622174263, + 0.0026949187740683556, + 0.013439503498375416, + 0.009953424334526062, + 0.006957478355616331, + 0.0043680849485099316, + 0.002870044205337763, + 0.0032273672986775637, + 0.002960219979286194, + 0.001431736396625638, + 0.002307754009962082, + 0.003956981003284454, + 0.0020098898094147444, + 0.00161886983551085, + 0.009391079656779766, + 0.007138024549931288, + 0.0049714804627001286, + 0.004190688952803612, + 0.009400895796716213, + 0.013565503992140293, + 0.006502514239400625, + 0.004111557267606258, + 0.02212706208229065, + 0.025074070319533348, + 0.044816721230745316, + 0.008675141260027885, + 0.01484148669987917, + 0.06643060594797134, + 0.005826137028634548, + 0.009757555089890957, + 0.005393210332840681, + 0.0033892604988068342, + 0.002653419738635421, + 0.004493510816246271, + 0.0143131660297513, + 0.010106140747666359, + 0.00802062451839447, + 0.003703634487465024, + 0.005500401370227337, + 0.004091425333172083, + 0.00308424374088645, + 0.0044435905292630196, + 0.013185206800699234, + 0.008989791385829449, + 0.0057051582261919975, + 0.003583790734410286, + 0.01041371002793312, + 0.009363716468214989, + 0.012867076322436333, + 0.008389350958168507, + 0.016063092276453972, + 0.014964674599468708, + 0.01148182898759842, + 0.009874420240521431, + 0.044512372463941574, + 0.0410589799284935, + 0.04261736944317818, + 0.04318522661924362, + 0.0640474185347557, + 0.05095170810818672, + 0.04640703648328781, + 0.05138483643531799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.0015136463334783912 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crayon drawing on textured white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the drawing depict a yellow sun, a blue house with a red triangular roof, and a waving stick figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a sharp black shadow pierce the frame and cause the drawing to rapidly sag inward and wrinkle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the drawing eventually collapse into a shrunken ball, losing its original shape and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733347196131945 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.006733347196131945, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733347196131945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06559862196445465, + 0.03503939136862755, + 0.05288459360599518, + 0.01888805627822876, + 0.05014330893754959, + 0.16350194811820984, + 0.13420647382736206, + 0.06266604363918304, + 0.12519560754299164, + 0.2918759286403656 + ], + "credit_map": [ + 0.012546525336802006, + 0.012865828350186348, + 0.00552859203889966, + 0.004245434887707233, + 0.006102298386394978, + 0.01175525039434433, + 0.0038608252070844173, + 0.00869386363774538, + 0.005894774571061134, + 0.0044731684029102325, + 0.0038033106829971075, + 0.003821250516921282, + 0.003468708135187626, + 0.0059411004185676575, + 0.002760028000921011, + 0.004877049475908279, + 0.00609263451769948, + 0.00494579691439867, + 0.006420861929655075, + 0.004383867606520653, + 0.010364272631704807, + 0.007664940785616636, + 0.007874771021306515, + 0.005137447267770767, + 0.002094207564368844, + 0.0024409645702689886, + 0.002492023166269064, + 0.0016011081170290709, + 0.003152960678562522, + 0.002737591974437237, + 0.0020877323113381863, + 0.00228146743029356, + 0.00896291434764862, + 0.0056264279410243034, + 0.006149667780846357, + 0.0035750637762248516, + 0.009219502098858356, + 0.008221720345318317, + 0.004101982805877924, + 0.004286030773073435, + 0.015519358217716217, + 0.01916317827999592, + 0.05521515756845474, + 0.00831462163478136, + 0.011627880856394768, + 0.041311413049697876, + 0.005271316971629858, + 0.0070790075697004795, + 0.008557543158531189, + 0.005522083956748247, + 0.022917568683624268, + 0.02176244556903839, + 0.02546563372015953, + 0.020391767844557762, + 0.012211629189550877, + 0.01737780123949051, + 0.006019013002514839, + 0.005066646263003349, + 0.005201669409871101, + 0.005048306193202734, + 0.018881771713495255, + 0.012880120426416397, + 0.0045708296820521355, + 0.004997683688998222, + 0.0194657351821661, + 0.019542595371603966, + 0.011248758994042873, + 0.014457267709076405, + 0.01482066884636879, + 0.021567001938819885, + 0.013915934599936008, + 0.010177653282880783, + 0.02756679244339466, + 0.030298247933387756, + 0.036854833364486694, + 0.03838448226451874, + 0.06510251760482788, + 0.04384005442261696, + 0.025111259892582893, + 0.02471776492893696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.006733347196131945 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_tensor([4977, 4977])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_tensor([4977, 4977])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c0bb22d787d1702bf7aebc328f0d3810b6a3ec66 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_tensor([4977, 4977])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level, full-body shot captures the stunning courtship display of a male golden pheasant, its movements rendered in stunning 8K surreal and clean detail against the silent forest floor. The male, an explosion of saturated color, features a brilliant golden-yellow crown flowing into a metallic green upper back and deep, fiery red chest. With one swift and deliberate movement, he completely surrounds his head with his magnificent mantle—an intricate ruff of light orange feathers, each tipped with a precise black border—forming a perfect, vibrant semicircle, obscuring one eye. He dances in short, intense, rhythmic steps across the damp earth strewn with fallen leaves and moss, his body tilted toward a nearby hen. The hen, a study in her mottled brown camouflage feathers, watches him with a still, observing gesture. The background is a soft-focus tapestry of deep green ferns and dark tree trunks. Dappled sunlight filters through the treetops, capturing the iridescent sheen of the male's plumage and creating an atmosphere of dramatic, focused natural wonder.", + "seed": 16398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_4977_seed_16398_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.049357885890640316, + "vlm_logprob": -0.049357885890640316, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.049357885890640316, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024953769519925117 + ], + "opposite_logprob_tokens": [ + -4.024953842163086 + ], + "logprob_reward": -0.024953769519925117, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024953769519925117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20276901125907898, + 0.041720304638147354, + 0.04408557340502739, + 0.11874183267354965, + 0.07116713374853134, + 0.06140604242682457, + 0.19187650084495544, + 0.0520784817636013, + 0.07884810864925385, + 0.13730700314044952 + ], + "credit_map": [ + 0.02988230437040329, + 0.046614281833171844, + 0.013162975199520588, + 0.019547421485185623, + 0.028391774743795395, + 0.025416016578674316, + 0.019159775227308273, + 0.020594453439116478, + 0.0056411586701869965, + 0.005746643990278244, + 0.004367075394839048, + 0.006102442741394043, + 0.004576005972921848, + 0.0054176487028598785, + 0.002814928302541375, + 0.007054397836327553, + 0.003066521603614092, + 0.005593514535576105, + 0.004390694200992584, + 0.006441785953938961, + 0.00658965390175581, + 0.0069372691214084625, + 0.0038626205641776323, + 0.007203514222055674, + 0.01139905583113432, + 0.012304280884563923, + 0.014246096834540367, + 0.015231974422931671, + 0.016224393621087074, + 0.01703793555498123, + 0.018664201721549034, + 0.013633890077471733, + 0.006320194806903601, + 0.008652325719594955, + 0.008015459403395653, + 0.01318235695362091, + 0.009050756692886353, + 0.008895942009985447, + 0.01033939141780138, + 0.0067107100039720535, + 0.007858970202505589, + 0.009761670604348183, + 0.005245840642601252, + 0.006892471108585596, + 0.010225499048829079, + 0.0075182197615504265, + 0.007219505030661821, + 0.006683870684355497, + 0.01777888461947441, + 0.04233350604772568, + 0.0456228032708168, + 0.015020555816590786, + 0.024627933278679848, + 0.01459426712244749, + 0.017724381759762764, + 0.014174163341522217, + 0.005937629844993353, + 0.00747911911457777, + 0.005072020925581455, + 0.005539726931601763, + 0.006680084392428398, + 0.006457650568336248, + 0.010573071427643299, + 0.004339179024100304, + 0.010392429307103157, + 0.009910812601447105, + 0.006814856547862291, + 0.008771805092692375, + 0.012270273640751839, + 0.009484606795012951, + 0.008706723339855671, + 0.012496606446802616, + 0.015628179535269737, + 0.020280316472053528, + 0.010656160302460194, + 0.01518773939460516, + 0.02699166163802147, + 0.016702424734830856, + 0.01869341731071472, + 0.013167097233235836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.024953769519925117 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025081150233745575 + ], + "opposite_logprob_tokens": [ + -6.002508163452148 + ], + "logprob_reward": -0.0025081150233745575, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025081150233745575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03967510163784027, + 0.13721632957458496, + 0.031834784895181656, + 0.12873370945453644, + 0.11805298924446106, + 0.07726198434829712, + 0.3660677969455719, + 0.05399465188384056, + 0.01382166612893343, + 0.03334091231226921 + ], + "credit_map": [ + 0.0060280729085206985, + 0.00637546693906188, + 0.0030927457846701145, + 0.006622083950787783, + 0.004828088451176882, + 0.0046677738428115845, + 0.0031840279698371887, + 0.004876846447587013, + 0.01312574464827776, + 0.025709981098771095, + 0.01253539603203535, + 0.025275886058807373, + 0.018830420449376106, + 0.01510893739759922, + 0.011294645257294178, + 0.015335354022681713, + 0.0039819832891225815, + 0.005250218324363232, + 0.002884466666728258, + 0.0044219354167580605, + 0.004952992778271437, + 0.004994676914066076, + 0.002888829680159688, + 0.002459688810631633, + 0.01043574046343565, + 0.009911775588989258, + 0.015045899897813797, + 0.01586281880736351, + 0.02080005407333374, + 0.01962190866470337, + 0.02488483116030693, + 0.012170705012977123, + 0.011220544576644897, + 0.013571142219007015, + 0.010964689776301384, + 0.013992180116474628, + 0.02475506067276001, + 0.01902659237384796, + 0.012697843834757805, + 0.011824946850538254, + 0.010660544037818909, + 0.009124964475631714, + 0.009649842977523804, + 0.007965462282299995, + 0.00848847534507513, + 0.010017195716500282, + 0.010185938328504562, + 0.011169576086103916, + 0.03331728279590607, + 0.05816437304019928, + 0.06766307353973389, + 0.02778441272675991, + 0.06228761374950409, + 0.03737484663724899, + 0.045034438371658325, + 0.034441787749528885, + 0.007985989563167095, + 0.008618617430329323, + 0.005224943161010742, + 0.00501386309042573, + 0.006117175333201885, + 0.007038110867142677, + 0.00824661087244749, + 0.005749350879341364, + 0.0021834622602909803, + 0.0014446093700826168, + 0.0015594488941133022, + 0.0016572836320847273, + 0.0018733624601736665, + 0.0017064877320080996, + 0.0019559653010219336, + 0.0014410479925572872, + 0.004260131623595953, + 0.005355948116630316, + 0.004875242244452238, + 0.0047828080132603645, + 0.0032589014153927565, + 0.003258798271417618, + 0.003794205142185092, + 0.0037548826076090336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.0025081150233745575 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072122380137444 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072122380137444, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072122380137444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12330171465873718, + 0.06180601194500923, + 0.024897966533899307, + 0.06912090629339218, + 0.10361486673355103, + 0.07903346419334412, + 0.3890286684036255, + 0.06271767616271973, + 0.06443975120782852, + 0.022039026021957397 + ], + "credit_map": [ + 0.016995705664157867, + 0.02672184444963932, + 0.010318683460354805, + 0.012027162127196789, + 0.015865111723542213, + 0.01459601428359747, + 0.009729844518005848, + 0.017047345638275146, + 0.005789007991552353, + 0.007050690241158009, + 0.006612169556319714, + 0.016168922185897827, + 0.005429279990494251, + 0.008156174793839455, + 0.004842735826969147, + 0.007757030893117189, + 0.002146695973351598, + 0.003155279206112027, + 0.0034927683882415295, + 0.002484553726390004, + 0.003510075155645609, + 0.0037678827065974474, + 0.0026296540163457394, + 0.0037110564298927784, + 0.005966263823211193, + 0.00477549247443676, + 0.008204895071685314, + 0.00786528643220663, + 0.011926152743399143, + 0.011087424121797085, + 0.01205164659768343, + 0.007243742700666189, + 0.00862440001219511, + 0.013163321651518345, + 0.011463082395493984, + 0.010301548056304455, + 0.022896084934473038, + 0.017329853028059006, + 0.00926577765494585, + 0.010570802725851536, + 0.011846677400171757, + 0.010305901058018208, + 0.009009622037410736, + 0.007625483442097902, + 0.009399404749274254, + 0.010156103409826756, + 0.010127263143658638, + 0.010563005693256855, + 0.03512038290500641, + 0.06361430138349533, + 0.07405561953783035, + 0.029182501137256622, + 0.06557467579841614, + 0.03927628695964813, + 0.04682153835892677, + 0.035383328795433044, + 0.008165858685970306, + 0.009961917996406555, + 0.005621123593300581, + 0.005214003846049309, + 0.0072856382466852665, + 0.008480287156999111, + 0.012024451978504658, + 0.005964390002191067, + 0.008714869618415833, + 0.009350037202239037, + 0.0056719486601650715, + 0.00734059140086174, + 0.009490913711488247, + 0.008444567210972309, + 0.007023721933364868, + 0.008403103798627853, + 0.0032787341624498367, + 0.003966044168919325, + 0.0016787821659818292, + 0.0020648897625505924, + 0.0035011402796953917, + 0.002827397547662258, + 0.002887840149924159, + 0.00183419871609658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.016072122380137444 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015426605939865 + ], + "opposite_logprob_tokens": [ + -1.701542615890503 + ], + "logprob_reward": -0.2015426605939865, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015426605939865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07485885918140411, + 0.14888055622577667, + 0.04979129508137703, + 0.059627603739500046, + 0.18867436051368713, + 0.0810776948928833, + 0.2616412937641144, + 0.04426194727420807, + 0.016795886680483818, + 0.07439055293798447 + ], + "credit_map": [ + 0.013301771134138107, + 0.010503175668418407, + 0.006429997272789478, + 0.008465479128062725, + 0.009259220212697983, + 0.010063279420137405, + 0.007428287994116545, + 0.009407641366124153, + 0.014321881346404552, + 0.02886117622256279, + 0.01354279462248087, + 0.01920822449028492, + 0.021758923307061195, + 0.016864104196429253, + 0.01273761410266161, + 0.021585816517472267, + 0.006048037204891443, + 0.004832078702747822, + 0.006754292640835047, + 0.006817356683313847, + 0.005996709689497948, + 0.007038818206638098, + 0.005798863247036934, + 0.006505134049803019, + 0.0059617264196276665, + 0.0073497723788022995, + 0.005179745610803366, + 0.006595092825591564, + 0.008417641744017601, + 0.009359681978821754, + 0.009246282279491425, + 0.007517652586102486, + 0.016296474263072014, + 0.018952524289488792, + 0.01543410588055849, + 0.0366377979516983, + 0.02208837680518627, + 0.025563368573784828, + 0.0322878323495388, + 0.02141387015581131, + 0.010983247309923172, + 0.012965592555701733, + 0.00840052030980587, + 0.009014211595058441, + 0.011176872067153454, + 0.010534160770475864, + 0.01069922000169754, + 0.007303861901164055, + 0.025097036734223366, + 0.04203144460916519, + 0.04885760694742203, + 0.019605474546551704, + 0.043574776500463486, + 0.02612634375691414, + 0.032112639397382736, + 0.02423594333231449, + 0.006498205475509167, + 0.006495657376945019, + 0.004836481995880604, + 0.006085029803216457, + 0.004950971342623234, + 0.004670139402151108, + 0.005771723110228777, + 0.004953733645379543, + 0.0020943230483680964, + 0.001649585785344243, + 0.0017037633806467056, + 0.002725137397646904, + 0.0018542760517448187, + 0.0020347805693745613, + 0.002647075103595853, + 0.0020869439467787743, + 0.008623642846941948, + 0.011244981549680233, + 0.0062896087765693665, + 0.00919215939939022, + 0.012030590325593948, + 0.009196537546813488, + 0.010529732331633568, + 0.0072832959704101086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.2015426605939865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017127619357779622 + ], + "opposite_logprob_tokens": [ + -6.376712799072266 + ], + "logprob_reward": -0.0017127619357779622, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017127619357779622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09047289192676544, + 0.16930022835731506, + 0.021954810246825218, + 0.0584256611764431, + 0.09666227549314499, + 0.1675330102443695, + 0.22990529239177704, + 0.055049486458301544, + 0.033988866955041885, + 0.07670748233795166 + ], + "credit_map": [ + 0.01203006412833929, + 0.017935633659362793, + 0.00933112483471632, + 0.008858679793775082, + 0.012710307724773884, + 0.011862149462103844, + 0.00806332379579544, + 0.009681600145995617, + 0.016140947118401527, + 0.032097283750772476, + 0.015752451494336128, + 0.028399420902132988, + 0.023634759709239006, + 0.018971573561429977, + 0.014059829525649548, + 0.020243946462869644, + 0.0031715857330709696, + 0.0019049245165660977, + 0.0023180206771939993, + 0.0025348688941448927, + 0.003441914916038513, + 0.0038803122006356716, + 0.002019967418164015, + 0.002683213446289301, + 0.003351954510435462, + 0.003518973710015416, + 0.006833315826952457, + 0.004466440994292498, + 0.01380555983632803, + 0.010683971457183361, + 0.005255467724055052, + 0.010509973391890526, + 0.01063684280961752, + 0.008833044208586216, + 0.009300698526203632, + 0.008149855770170689, + 0.023492351174354553, + 0.01669752225279808, + 0.00962484534829855, + 0.009927106089890003, + 0.019585032016038895, + 0.02064860425889492, + 0.021224789321422577, + 0.018696123734116554, + 0.02026975527405739, + 0.02139442414045334, + 0.0212181955575943, + 0.02449607290327549, + 0.02254294790327549, + 0.04740428552031517, + 0.05297812074422836, + 0.016259346157312393, + 0.028886236250400543, + 0.018482573330402374, + 0.022916186600923538, + 0.02043559029698372, + 0.009065775200724602, + 0.008781933225691319, + 0.007981552742421627, + 0.006032868288457394, + 0.006120665930211544, + 0.005051895976066589, + 0.007614391390234232, + 0.00440039811655879, + 0.00476393336430192, + 0.0033428415190428495, + 0.003608875907957554, + 0.004088698886334896, + 0.0036727944388985634, + 0.005331940948963165, + 0.005097650922834873, + 0.00408212561160326, + 0.009195693768560886, + 0.011581402271986008, + 0.006650671362876892, + 0.008619923144578934, + 0.014561062678694725, + 0.00896924827247858, + 0.010330324061214924, + 0.006799144204705954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.0017127619357779622 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, full-body shot captures the stunning courtship display of a male golden pheasant, its movements rendered in stunning 8K surreal and clean detail against the silent forest floor. The male, an explosion of saturated color, features a brilliant golden-yellow crown flowing into a metallic green upper back and deep, fiery red chest. With one swift and deliberate movement, he completely surrounds his head with his magnificent mantle—an intricate ruff of light orange feathers, each tipped with a precise black border—forming a perfect, vibrant semicircle, obscuring one eye. He dances in short, intense, rhythmic steps across the damp earth strewn with fallen leaves and moss, his body tilted toward a nearby hen. The hen, a study in her mottled brown camouflage feathers, watches him with a still, observing gesture. The background is a soft-focus tapestry of deep green ferns and dark tree trunks. Dappled sunlight filters through the treetops, capturing the iridescent sheen of the male's plumage and creating an atmosphere of dramatic, focused natural wonder.", + "seed": 16399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_4977_seed_16399_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026106486585922538, + "vlm_logprob": -0.026106486585922538, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.026106486585922538, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0427999421954155 + ], + "opposite_logprob_tokens": [ + -3.292799949645996 + ], + "logprob_reward": -0.0427999421954155, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0427999421954155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15195390582084656, + 0.16670212149620056, + 0.04345713555812836, + 0.0652092695236206, + 0.4146866798400879, + 0.045057833194732666, + 0.015337024815380573, + 0.011757705360651016, + 0.06049909442663193, + 0.025339217856526375 + ], + "credit_map": [ + 0.028339281678199768, + 0.015708616003394127, + 0.014777200296521187, + 0.017517227679491043, + 0.013209528289735317, + 0.012296022847294807, + 0.01889514923095703, + 0.031210890039801598, + 0.022273747250437737, + 0.013748852536082268, + 0.025190718472003937, + 0.025673527270555496, + 0.016296325251460075, + 0.019234376028180122, + 0.020048247650265694, + 0.02423633635044098, + 0.00778695335611701, + 0.002954432275146246, + 0.005685500800609589, + 0.004894813988357782, + 0.004219972528517246, + 0.003965798765420914, + 0.0058406624011695385, + 0.008109000511467457, + 0.010295365005731583, + 0.004617683589458466, + 0.016624266281723976, + 0.00813498068600893, + 0.005191507749259472, + 0.005804278422147036, + 0.0067543950863182545, + 0.007786793168634176, + 0.06386911123991013, + 0.031193898990750313, + 0.06742534786462784, + 0.04485952481627464, + 0.03935321792960167, + 0.046792566776275635, + 0.05315805971622467, + 0.06803493946790695, + 0.008390326984226704, + 0.00377854285761714, + 0.004925545305013657, + 0.003931416664272547, + 0.006251809652894735, + 0.006717944052070379, + 0.0051614814437925816, + 0.00590076670050621, + 0.001644192379899323, + 0.001422181143425405, + 0.0012773725902661681, + 0.002519693924114108, + 0.0019556290935724974, + 0.0023116949014365673, + 0.002368259709328413, + 0.001838000607676804, + 0.0013927549589425325, + 0.0010432740673422813, + 0.0013008100213482976, + 0.0014838689239695668, + 0.0016216887161135674, + 0.0016286219470202923, + 0.001730382558889687, + 0.0015563045162707567, + 0.006574433762580156, + 0.0065348586067557335, + 0.004742363467812538, + 0.007577692624181509, + 0.010345524176955223, + 0.0068239266984164715, + 0.010941915214061737, + 0.0069583808071911335, + 0.003679681569337845, + 0.0039841486141085625, + 0.002177067566663027, + 0.003301246091723442, + 0.002836302388459444, + 0.003359112422913313, + 0.0032942513935267925, + 0.002707407111302018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0427999421954155 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003633802989497781 + ], + "opposite_logprob_tokens": [ + -5.628633975982666 + ], + "logprob_reward": -0.003633802989497781, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003633802989497781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2614333927631378, + 0.08029451966285706, + 0.07040707021951675, + 0.044427115470170975, + 0.13036003708839417, + 0.10976362973451614, + 0.034181661903858185, + 0.07727456092834473, + 0.06413488835096359, + 0.12772312760353088 + ], + "credit_map": [ + 0.05256277322769165, + 0.034925226122140884, + 0.021467819809913635, + 0.016371095553040504, + 0.02174469083547592, + 0.021017704159021378, + 0.02919117547571659, + 0.06415288150310516, + 0.011599339544773102, + 0.00633320352062583, + 0.010656716302037239, + 0.011629346758127213, + 0.008494400419294834, + 0.008243385702371597, + 0.009910053573548794, + 0.01342807151377201, + 0.011404205113649368, + 0.00471419095993042, + 0.01293234433978796, + 0.010850847698748112, + 0.005253088194876909, + 0.0048139323480427265, + 0.0069936648942530155, + 0.013444802723824978, + 0.006575223058462143, + 0.0028341226279735565, + 0.006107069551944733, + 0.0045540826395154, + 0.00634524691849947, + 0.004288890864700079, + 0.007872556336224079, + 0.00584992254152894, + 0.01941310428082943, + 0.010231303982436657, + 0.01802784763276577, + 0.01702922023832798, + 0.013168714940547943, + 0.013157430104911327, + 0.01702732779085636, + 0.022305091843008995, + 0.02087625116109848, + 0.008550469763576984, + 0.015102694742381573, + 0.008616429753601551, + 0.013267957605421543, + 0.01451038010418415, + 0.011808620765805244, + 0.017030833289027214, + 0.0052665825933218, + 0.002527910750359297, + 0.0036995222326368093, + 0.00401395233348012, + 0.00490317028015852, + 0.005304285325109959, + 0.003850681707262993, + 0.004615557845681906, + 0.010162660852074623, + 0.004382666200399399, + 0.011673245579004288, + 0.008157077245414257, + 0.006362729240208864, + 0.009859638288617134, + 0.018247149884700775, + 0.008429394103586674, + 0.008103656582534313, + 0.006412620656192303, + 0.003549155779182911, + 0.0077975234016776085, + 0.013311169110238552, + 0.006762966513633728, + 0.011532813310623169, + 0.006664982996881008, + 0.022915318608283997, + 0.014636622741818428, + 0.010184160433709621, + 0.013948002830147743, + 0.02072605863213539, + 0.012238924391567707, + 0.016296857967972755, + 0.016777189448475838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.003633802989497781 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362038123421371 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.00362038123421371, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362038123421371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.108071468770504, + 0.11597520112991333, + 0.09641335904598236, + 0.10991661995649338, + 0.39763781428337097, + 0.0551818311214447, + 0.02402474358677864, + 0.016462385654449463, + 0.04177543893456459, + 0.03454115241765976 + ], + "credit_map": [ + 0.02637079544365406, + 0.010312210768461227, + 0.010570166632533073, + 0.0073080286383628845, + 0.00893775187432766, + 0.008534395135939121, + 0.01001804694533348, + 0.026020072400569916, + 0.015921762213110924, + 0.0113456379622221, + 0.015816492959856987, + 0.01737767830491066, + 0.01241452619433403, + 0.013819131068885326, + 0.015090753324329853, + 0.014189226552844048, + 0.018515991047024727, + 0.006636507343500853, + 0.014646802097558975, + 0.009228884242475033, + 0.008612539619207382, + 0.00902089849114418, + 0.01332278735935688, + 0.016428949311375618, + 0.018053308129310608, + 0.007217032369226217, + 0.026159467175602913, + 0.013134470209479332, + 0.010874955914914608, + 0.008791460655629635, + 0.011349490843713284, + 0.014336436986923218, + 0.06044789403676987, + 0.030118832364678383, + 0.06301988661289215, + 0.0443197526037693, + 0.03901701793074608, + 0.04296271875500679, + 0.052581291645765305, + 0.06517043709754944, + 0.010701595805585384, + 0.0039924378506839275, + 0.0072881607338786125, + 0.005809750873595476, + 0.007162792608141899, + 0.005670879501849413, + 0.006449427455663681, + 0.008106786757707596, + 0.00313022849150002, + 0.002489827573299408, + 0.002321008127182722, + 0.002544958610087633, + 0.003546156222000718, + 0.004313010256737471, + 0.0029714524280279875, + 0.0027081004809588194, + 0.00287357484921813, + 0.001329690683633089, + 0.0017241284949705005, + 0.0019399194279685616, + 0.002202937612310052, + 0.0019947285763919353, + 0.0021567316725850105, + 0.002240674337372184, + 0.005247924942523241, + 0.004121654666960239, + 0.00304450374096632, + 0.004868417512625456, + 0.00803861953318119, + 0.00465488713234663, + 0.0075627276673913, + 0.004236701410263777, + 0.00502845598384738, + 0.004816648084670305, + 0.0031676378566771746, + 0.004533116705715656, + 0.00430213613435626, + 0.004224360920488834, + 0.003530903020873666, + 0.004937893245369196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.00362038123421371 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896608859300613 + ], + "opposite_logprob_tokens": [ + -2.5789661407470703 + ], + "logprob_reward": -0.07896608859300613, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896608859300613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3181462585926056, + 0.05007651448249817, + 0.024490496143698692, + 0.018358908593654633, + 0.3816346228122711, + 0.04750359058380127, + 0.025335054844617844, + 0.022884875535964966, + 0.05820852890610695, + 0.053361114114522934 + ], + "credit_map": [ + 0.050811704248189926, + 0.04928460344672203, + 0.042563002556562424, + 0.0283330250531435, + 0.037735465914011, + 0.027094978839159012, + 0.02870049513876438, + 0.053622983396053314, + 0.007909785956144333, + 0.0035342054907232523, + 0.007595856674015522, + 0.007520019076764584, + 0.004096572287380695, + 0.004579139873385429, + 0.005756991915404797, + 0.0090839434415102, + 0.003757581114768982, + 0.0014460236998274922, + 0.004365008324384689, + 0.0032784189097583294, + 0.0021388991735875607, + 0.0020770258270204067, + 0.003379057627171278, + 0.004048479720950127, + 0.0026881766971200705, + 0.0014630533987656236, + 0.0024485900066792965, + 0.002502101007848978, + 0.0015762833645567298, + 0.002470283070579171, + 0.0025614826008677483, + 0.00264893751591444, + 0.059075262397527695, + 0.027839669957756996, + 0.0652485117316246, + 0.03949609398841858, + 0.03554479405283928, + 0.042519085109233856, + 0.04720912501215935, + 0.0647021010518074, + 0.008779517374932766, + 0.004171716049313545, + 0.006774221546947956, + 0.0038349207025021315, + 0.005335257388651371, + 0.007099946029484272, + 0.004618003033101559, + 0.0068900068290531635, + 0.003201195038855076, + 0.0017604214372113347, + 0.0032622527796775103, + 0.0024771655444055796, + 0.003810050431638956, + 0.00419879425317049, + 0.0036927368491888046, + 0.0029324383940547705, + 0.004097132943570614, + 0.001675012637861073, + 0.002786531113088131, + 0.002556122373789549, + 0.0021781001705676317, + 0.003070350969210267, + 0.003812460694462061, + 0.0027091633528470993, + 0.006663577631115913, + 0.006202850956469774, + 0.004477676469832659, + 0.009150759316980839, + 0.009744266048073769, + 0.005828899331390858, + 0.0105916578322649, + 0.005548838526010513, + 0.008302529342472553, + 0.0051745581440627575, + 0.003499254584312439, + 0.005671117454767227, + 0.007814811542630196, + 0.007506994530558586, + 0.00882373470813036, + 0.006568113807588816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.07896608859300613 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015122179174795747 + ], + "opposite_logprob_tokens": [ + -6.501512050628662 + ], + "logprob_reward": -0.0015122179174795747, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015122179174795747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14134176075458527, + 0.09469103068113327, + 0.028606195002794266, + 0.09271295368671417, + 0.3905636668205261, + 0.060553427785634995, + 0.043261557817459106, + 0.029292097315192223, + 0.0645185261964798, + 0.054458752274513245 + ], + "credit_map": [ + 0.023546898737549782, + 0.01752760075032711, + 0.016611279919743538, + 0.020668519660830498, + 0.012570317834615707, + 0.009467912837862968, + 0.01671946793794632, + 0.02422977052628994, + 0.013532216660678387, + 0.0070211514830589294, + 0.016735713928937912, + 0.01637336052954197, + 0.006287207826972008, + 0.007236702833324671, + 0.01173012051731348, + 0.015774564817547798, + 0.00474576186388731, + 0.0016347848577424884, + 0.005502640735358, + 0.00391645822674036, + 0.0026581380516290665, + 0.0023155510425567627, + 0.0030595522839576006, + 0.004773310851305723, + 0.014256571419537067, + 0.006514042615890503, + 0.019494419917464256, + 0.013330462388694286, + 0.0067366561852395535, + 0.0085175521671772, + 0.012547707185149193, + 0.011315549723803997, + 0.06420814245939255, + 0.02742002345621586, + 0.07266419380903244, + 0.039216045290231705, + 0.030989959836006165, + 0.04237039387226105, + 0.04070202261209488, + 0.07299289852380753, + 0.010001786053180695, + 0.004405219107866287, + 0.010851564817130566, + 0.006979029625654221, + 0.005608425475656986, + 0.007790777832269669, + 0.0056641255505383015, + 0.009252500720322132, + 0.007891584187746048, + 0.002783339936286211, + 0.006399065256118774, + 0.0047293733805418015, + 0.004393087700009346, + 0.005987679120153189, + 0.005574781447649002, + 0.005502649582922459, + 0.005921860691159964, + 0.0015365456929430366, + 0.003970231395214796, + 0.003095603082329035, + 0.002202223753556609, + 0.003669174388051033, + 0.004260463640093803, + 0.004635995719581842, + 0.009398655034601688, + 0.004744927864521742, + 0.007225569803267717, + 0.007743991911411285, + 0.0058625079691410065, + 0.008841697126626968, + 0.008111411705613136, + 0.012589766643941402, + 0.00914474856108427, + 0.006010766141116619, + 0.0047293296083807945, + 0.006603028625249863, + 0.006741190794855356, + 0.00549738435074687, + 0.007263152860105038, + 0.008469156920909882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0015122179174795747 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, full-body shot captures the stunning courtship display of a male golden pheasant, its movements rendered in stunning 8K surreal and clean detail against the silent forest floor. The male, an explosion of saturated color, features a brilliant golden-yellow crown flowing into a metallic green upper back and deep, fiery red chest. With one swift and deliberate movement, he completely surrounds his head with his magnificent mantle—an intricate ruff of light orange feathers, each tipped with a precise black border—forming a perfect, vibrant semicircle, obscuring one eye. He dances in short, intense, rhythmic steps across the damp earth strewn with fallen leaves and moss, his body tilted toward a nearby hen. The hen, a study in her mottled brown camouflage feathers, watches him with a still, observing gesture. The background is a soft-focus tapestry of deep green ferns and dark tree trunks. Dappled sunlight filters through the treetops, capturing the iridescent sheen of the male's plumage and creating an atmosphere of dramatic, focused natural wonder.", + "seed": 16400, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_4977_seed_16400_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012369369622319937, + "vlm_logprob": -0.012369369622319937, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.012369369622319937, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021360348910093307 + ], + "opposite_logprob_tokens": [ + -4.146360397338867 + ], + "logprob_reward": -0.021360348910093307, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.021360348910093307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15733620524406433, + 0.14455115795135498, + 0.20185917615890503, + 0.11588574200868607, + 0.04517671465873718, + 0.07464328408241272, + 0.08517295122146606, + 0.09928734600543976, + 0.052616313099861145, + 0.02347109466791153 + ], + "credit_map": [ + 0.0262072142213583, + 0.025478951632976532, + 0.011813142336905003, + 0.019215328618884087, + 0.014710944145917892, + 0.019795330241322517, + 0.018102064728736877, + 0.022013241425156593, + 0.030326684936881065, + 0.013243201188743114, + 0.0360708013176918, + 0.021295521408319473, + 0.007456332910805941, + 0.006153801456093788, + 0.0150291183963418, + 0.014975694939494133, + 0.03784630075097084, + 0.017555665224790573, + 0.04006943851709366, + 0.04314746335148811, + 0.01430258248001337, + 0.013765950687229633, + 0.01484629511833191, + 0.020325463265180588, + 0.018785599619150162, + 0.010607706382870674, + 0.011889475397765636, + 0.01481260359287262, + 0.01752866804599762, + 0.016152273863554, + 0.012240631505846977, + 0.013868779875338078, + 0.007375578861683607, + 0.0031236291397362947, + 0.010849764570593834, + 0.005048121325671673, + 0.0029388924594968557, + 0.0046167937107384205, + 0.005326895043253899, + 0.005897043738514185, + 0.01209582481533289, + 0.004856339655816555, + 0.012442539446055889, + 0.008848530240356922, + 0.00871641468256712, + 0.008518923074007034, + 0.008804656565189362, + 0.010360056534409523, + 0.01770257018506527, + 0.007178490050137043, + 0.017330700531601906, + 0.0083225779235363, + 0.006626412272453308, + 0.009778601117432117, + 0.007668318226933479, + 0.01056528277695179, + 0.014527478255331516, + 0.007376410532742739, + 0.009655963629484177, + 0.010615995153784752, + 0.008520850911736488, + 0.018135368824005127, + 0.0116775743663311, + 0.01877770945429802, + 0.008454861119389534, + 0.005877861753106117, + 0.00689731165766716, + 0.005215951707214117, + 0.005804789252579212, + 0.006181890144944191, + 0.0055952128022909164, + 0.008588435128331184, + 0.003231193870306015, + 0.0028140749782323837, + 0.00430626654997468, + 0.002749189967289567, + 0.0031651498284190893, + 0.0020512596238404512, + 0.0025453802663832903, + 0.0026085807476192713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.021360348910093307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028348281048238277 + ], + "opposite_logprob_tokens": [ + -5.877834796905518 + ], + "logprob_reward": -0.0028348281048238277, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028348281048238277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19939373433589935, + 0.12856604158878326, + 0.046208929270505905, + 0.08298340439796448, + 0.0865587666630745, + 0.1380927562713623, + 0.05904749780893326, + 0.16397234797477722, + 0.05611243098974228, + 0.03906407207250595 + ], + "credit_map": [ + 0.03498228266835213, + 0.02914542146027088, + 0.020319905132055283, + 0.017623165622353554, + 0.0371210053563118, + 0.01983255334198475, + 0.021059131249785423, + 0.01931026764214039, + 0.022392932325601578, + 0.02185816317796707, + 0.02180290035903454, + 0.019305402413010597, + 0.007888872176408768, + 0.009389623999595642, + 0.009747160598635674, + 0.016181007027626038, + 0.008619243279099464, + 0.0029245486948639154, + 0.007811343763023615, + 0.0064761037938296795, + 0.0035586548037827015, + 0.003704819129779935, + 0.00863761268556118, + 0.004476605914533138, + 0.014307071454823017, + 0.005925668869167566, + 0.01118422020226717, + 0.013579383492469788, + 0.011482306756079197, + 0.006642268039286137, + 0.012635830789804459, + 0.0072266655042767525, + 0.012403291650116444, + 0.011282318271696568, + 0.01340273953974247, + 0.01220628246665001, + 0.013849708251655102, + 0.008295894600450993, + 0.008702212944626808, + 0.006416326854377985, + 0.0249239020049572, + 0.00980420596897602, + 0.016678154468536377, + 0.01651737093925476, + 0.012781460769474506, + 0.015996253117918968, + 0.016776086762547493, + 0.0246153324842453, + 0.00821107067167759, + 0.004925617948174477, + 0.0067202430218458176, + 0.006172408815473318, + 0.005303710233420134, + 0.012012992054224014, + 0.003765240078791976, + 0.011936217546463013, + 0.01409273874014616, + 0.005358375608921051, + 0.007725704926997423, + 0.00647208746522665, + 0.008011127822101116, + 0.054251302033662796, + 0.012379569001495838, + 0.05568144842982292, + 0.009485340677201748, + 0.0066836681216955185, + 0.005281026940792799, + 0.006594247184693813, + 0.006607052404433489, + 0.005495633464306593, + 0.0063215382397174835, + 0.00964392814785242, + 0.005208980292081833, + 0.0037892393302172422, + 0.005413294769823551, + 0.00679956516250968, + 0.0048097362741827965, + 0.003865759586915374, + 0.003999555017799139, + 0.005177944432944059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0028348281048238277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763498293235898 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.00763498293235898, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763498293235898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.145809605717659, + 0.05766928568482399, + 0.370462030172348, + 0.06614776700735092, + 0.05753915384411812, + 0.11749924719333649, + 0.0585971362888813, + 0.0740537941455841, + 0.031627919524908066, + 0.02059408649802208 + ], + "credit_map": [ + 0.020668242126703262, + 0.022155243903398514, + 0.014681973494589329, + 0.012133642099797726, + 0.02664339356124401, + 0.013953731395304203, + 0.01807098463177681, + 0.01750241033732891, + 0.011175614781677723, + 0.005451649893075228, + 0.012160088866949081, + 0.008598323911428452, + 0.004825192503631115, + 0.0032822636421769857, + 0.005453851073980331, + 0.006722303573042154, + 0.07085622847080231, + 0.02801075577735901, + 0.08050105720758438, + 0.07893827557563782, + 0.022967064753174782, + 0.024334384128451347, + 0.025900740176439285, + 0.03895353898406029, + 0.011084014549851418, + 0.006112587172538042, + 0.008377640508115292, + 0.01038938295096159, + 0.009204423055052757, + 0.006296180188655853, + 0.008267183788120747, + 0.006416355725377798, + 0.01026041992008686, + 0.006734068971127272, + 0.00977207999676466, + 0.009967884980142117, + 0.004544222727417946, + 0.00436742790043354, + 0.00564922858029604, + 0.006243817042559385, + 0.019975770264863968, + 0.007689318619668484, + 0.017075160518288612, + 0.014446859247982502, + 0.015009358525276184, + 0.012808101251721382, + 0.011861852370202541, + 0.01863282546401024, + 0.007442005444318056, + 0.004985631909221411, + 0.005273753311485052, + 0.005229190457612276, + 0.003680962836369872, + 0.013797999359667301, + 0.0041841487400233746, + 0.01400343794375658, + 0.008946110494434834, + 0.00641236174851656, + 0.00539321219548583, + 0.007061687298119068, + 0.0066407001577317715, + 0.01318341214209795, + 0.010597423650324345, + 0.015818890184164047, + 0.004233590327203274, + 0.003899338422343135, + 0.0018993315752595663, + 0.005090054590255022, + 0.0038752290420234203, + 0.004104021470993757, + 0.0036789318546652794, + 0.004847418516874313, + 0.0031782756559550762, + 0.0024639698676764965, + 0.002210704144090414, + 0.0028996458277106285, + 0.0034143771044909954, + 0.0020138367544859648, + 0.001842152327299118, + 0.0025711229536682367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.00763498293235898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280736058950424 + ], + "opposite_logprob_tokens": [ + -3.773280620574951 + ], + "logprob_reward": -0.023280736058950424, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280736058950424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18539588153362274, + 0.24773018062114716, + 0.18000765144824982, + 0.04583484306931496, + 0.1980939358472824, + 0.05775662139058113, + 0.015140385366976261, + 0.028494393453001976, + 0.02347302809357643, + 0.018073102459311485 + ], + "credit_map": [ + 0.045762017369270325, + 0.02862158976495266, + 0.02472599968314171, + 0.017556393519043922, + 0.013952935114502907, + 0.013619855977594852, + 0.016167819499969482, + 0.024989286437630653, + 0.04622384533286095, + 0.022173341363668442, + 0.05973701924085617, + 0.04750053957104683, + 0.01072737481445074, + 0.011096780188381672, + 0.019148405641317368, + 0.03112286701798439, + 0.03401463106274605, + 0.014328022487461567, + 0.036705613136291504, + 0.03924928233027458, + 0.01223782543092966, + 0.012599704787135124, + 0.013065801002085209, + 0.017806770280003548, + 0.007368965540081263, + 0.0033275624737143517, + 0.0058862571604549885, + 0.00818268209695816, + 0.005854267161339521, + 0.005643974058330059, + 0.005745502654463053, + 0.0038256326224654913, + 0.036315660923719406, + 0.012772753834724426, + 0.043873343616724014, + 0.031055768951773643, + 0.0170237198472023, + 0.01828300766646862, + 0.016810167580842972, + 0.021959524601697922, + 0.009867442771792412, + 0.0037075330037623644, + 0.0074865082278847694, + 0.00680014630779624, + 0.007160250563174486, + 0.007103214506059885, + 0.006024816073477268, + 0.009606705978512764, + 0.0022507905960083008, + 0.0014435610501095653, + 0.001848458661697805, + 0.001358801033347845, + 0.0016340988222509623, + 0.0022731611970812082, + 0.0018307121936231852, + 0.0025008018128573895, + 0.003228242974728346, + 0.0012957510771229863, + 0.0021991010289639235, + 0.0019177194917574525, + 0.0012293538311496377, + 0.008139554411172867, + 0.002171164844185114, + 0.008313504979014397, + 0.004484242759644985, + 0.0024891141802072525, + 0.0037559133488684893, + 0.002558748470619321, + 0.0024106353521347046, + 0.002752537839114666, + 0.0023365840315818787, + 0.0026852518785744905, + 0.003352368250489235, + 0.0019071924034506083, + 0.003073301864787936, + 0.0020544498693197966, + 0.0018362398259341717, + 0.0019100471399724483, + 0.001534168841317296, + 0.0024053347297012806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.023280736058950424 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735952105373144 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735952105373144, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735952105373144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25355589389801025, + 0.08825484663248062, + 0.08376599103212357, + 0.05041397735476494, + 0.033620309084653854, + 0.14597655832767487, + 0.1549832969903946, + 0.1516675353050232, + 0.025451937690377235, + 0.012309658341109753 + ], + "credit_map": [ + 0.054998818784952164, + 0.03958972543478012, + 0.03476615622639656, + 0.030479837208986282, + 0.016137422993779182, + 0.02025957778096199, + 0.020359525457024574, + 0.036964841187000275, + 0.01695646531879902, + 0.008702938444912434, + 0.019378935918211937, + 0.0167403407394886, + 0.00517367385327816, + 0.004613952245563269, + 0.006895715370774269, + 0.00979282520711422, + 0.015539921820163727, + 0.005836168769747019, + 0.016027923673391342, + 0.018555469810962677, + 0.005099914036691189, + 0.005271703936159611, + 0.006868590600788593, + 0.010566296987235546, + 0.008424364030361176, + 0.0037706377916038036, + 0.005948321428149939, + 0.008950995281338692, + 0.007378008216619492, + 0.006318582687526941, + 0.00536753423511982, + 0.004255530424416065, + 0.005377736873924732, + 0.0038051086012274027, + 0.004696174059063196, + 0.0045888363383710384, + 0.0037921450566500425, + 0.004455755930393934, + 0.003294416004791856, + 0.0036101366858929396, + 0.02533666603267193, + 0.011475260369479656, + 0.013214215636253357, + 0.014740364626049995, + 0.014910200610756874, + 0.01784464344382286, + 0.01515846885740757, + 0.033296726644039154, + 0.030325036495923996, + 0.01468395721167326, + 0.027983060106635094, + 0.015349527820944786, + 0.013505280017852783, + 0.02091003768146038, + 0.010615488514304161, + 0.02161089889705181, + 0.01765863597393036, + 0.0046896012499928474, + 0.016205228865146637, + 0.00624369690194726, + 0.005349809769541025, + 0.0484810471534729, + 0.005876939743757248, + 0.04716259241104126, + 0.004822896793484688, + 0.0028256529476493597, + 0.002600757172331214, + 0.0030896582175046206, + 0.002455639885738492, + 0.0031611432787030935, + 0.0035829138942062855, + 0.0029132745694369078, + 0.0020643663592636585, + 0.0012998618185520172, + 0.002018539933487773, + 0.00181891571264714, + 0.001290823332965374, + 0.001288328436203301, + 0.001097500673495233, + 0.0014313217252492905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.006735952105373144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, full-body shot captures the stunning courtship display of a male golden pheasant, its movements rendered in stunning 8K surreal and clean detail against the silent forest floor. The male, an explosion of saturated color, features a brilliant golden-yellow crown flowing into a metallic green upper back and deep, fiery red chest. With one swift and deliberate movement, he completely surrounds his head with his magnificent mantle—an intricate ruff of light orange feathers, each tipped with a precise black border—forming a perfect, vibrant semicircle, obscuring one eye. He dances in short, intense, rhythmic steps across the damp earth strewn with fallen leaves and moss, his body tilted toward a nearby hen. The hen, a study in her mottled brown camouflage feathers, watches him with a still, observing gesture. The background is a soft-focus tapestry of deep green ferns and dark tree trunks. Dappled sunlight filters through the treetops, capturing the iridescent sheen of the male's plumage and creating an atmosphere of dramatic, focused natural wonder.", + "seed": 16401, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_7/index_4977_seed_16401_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01494982298463583, + "vlm_logprob": -0.01494982298463583, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.01494982298463583, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02347591519355774 + ], + "opposite_logprob_tokens": [ + -4.0234761238098145 + ], + "logprob_reward": -0.02347591519355774, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02347591519355774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07509452104568481, + 0.039456430822610855, + 0.06731607764959335, + 0.3591565191745758, + 0.1406571865081787, + 0.14043393731117249, + 0.107408806681633, + 0.03249317407608032, + 0.020254148170351982, + 0.01772923767566681 + ], + "credit_map": [ + 0.0078001003712415695, + 0.00948159210383892, + 0.012331214733421803, + 0.00985669158399105, + 0.007671241648495197, + 0.0105036161839962, + 0.00911493320018053, + 0.008335129357874393, + 0.005340375006198883, + 0.005574760492891073, + 0.004461898468434811, + 0.006466549821197987, + 0.003394849831238389, + 0.005181802436709404, + 0.006286505609750748, + 0.0027496872935444117, + 0.00906588975340128, + 0.00545120844617486, + 0.00872464943677187, + 0.010760285891592503, + 0.006069208960980177, + 0.008768449537456036, + 0.010441528633236885, + 0.008034859783947468, + 0.05530572682619095, + 0.043508756905794144, + 0.055162981152534485, + 0.05570045858621597, + 0.03403378650546074, + 0.04730488359928131, + 0.03797663375735283, + 0.03016330674290657, + 0.02233796752989292, + 0.013593042269349098, + 0.029216952621936798, + 0.016311952844262123, + 0.017974236980080605, + 0.008336428552865982, + 0.019395101815462112, + 0.013491502963006496, + 0.02189488522708416, + 0.0120038827881217, + 0.02372472919523716, + 0.021767448633909225, + 0.01006784662604332, + 0.012003479525446892, + 0.015820950269699097, + 0.02315070852637291, + 0.026474734768271446, + 0.009382103569805622, + 0.009772676043212414, + 0.011117405258119106, + 0.009451828896999359, + 0.022754687815904617, + 0.010720123536884785, + 0.007735247258096933, + 0.00515313446521759, + 0.0025181835517287254, + 0.00561180105432868, + 0.004433987662196159, + 0.0027409899048507214, + 0.005367201287299395, + 0.004073453601449728, + 0.002594424644485116, + 0.002863661851733923, + 0.0019033633871003985, + 0.003266754560172558, + 0.00251956470310688, + 0.002303656190633774, + 0.0029948807787150145, + 0.0021790226455777884, + 0.002223244635388255, + 0.0032705937046557665, + 0.0010075413156300783, + 0.003614200744777918, + 0.003021060721948743, + 0.0013061246136203408, + 0.0028907605446875095, + 0.001336428220383823, + 0.001282527344301343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.02347591519355774 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002214600332081318 + ], + "opposite_logprob_tokens": [ + -6.127214431762695 + ], + "logprob_reward": -0.002214600332081318, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002214600332081318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09409243613481522, + 0.027217725291848183, + 0.16187743842601776, + 0.12967762351036072, + 0.05075899139046669, + 0.26554062962532043, + 0.15874311327934265, + 0.044826943427324295, + 0.02797423116862774, + 0.03929082304239273 + ], + "credit_map": [ + 0.0075141508132219315, + 0.008002229034900665, + 0.020387141034007072, + 0.015316498465836048, + 0.008671821095049381, + 0.013393465429544449, + 0.010970903560519218, + 0.009836219251155853, + 0.004544547293335199, + 0.003266831859946251, + 0.003530260641127825, + 0.0038906161207705736, + 0.0018243166850879788, + 0.004238009452819824, + 0.0030750285368412733, + 0.002848113654181361, + 0.02278008498251438, + 0.01914590410888195, + 0.024267328903079033, + 0.019551197066903114, + 0.012101870961487293, + 0.020432403311133385, + 0.018596211448311806, + 0.02500244416296482, + 0.02005859464406967, + 0.0145674217492342, + 0.02077123522758484, + 0.016027694568037987, + 0.012439940124750137, + 0.016218669712543488, + 0.01598663255572319, + 0.013607436791062355, + 0.007998083718121052, + 0.005101279355585575, + 0.009460526518523693, + 0.007410209625959396, + 0.0047339703887701035, + 0.003643762320280075, + 0.007069074083119631, + 0.00534208444878459, + 0.04012906551361084, + 0.021240025758743286, + 0.04478247091174126, + 0.040566280484199524, + 0.0182782169431448, + 0.020259199663996696, + 0.033489156514406204, + 0.04679623618721962, + 0.025553176179528236, + 0.012920557521283627, + 0.022554945200681686, + 0.022818000987172127, + 0.014356597326695919, + 0.02092711627483368, + 0.015815015882253647, + 0.02379770763218403, + 0.009048999287188053, + 0.0030588151421397924, + 0.007490117102861404, + 0.005534665193408728, + 0.002762269228696823, + 0.00932752713561058, + 0.004263716284185648, + 0.003340834053233266, + 0.004843851551413536, + 0.0023109421599656343, + 0.004267605021595955, + 0.0023920179810374975, + 0.004318798426538706, + 0.004111957270652056, + 0.0030799827072769403, + 0.0026490758173167706, + 0.005649521481245756, + 0.002468680264428258, + 0.008706286549568176, + 0.004693773575127125, + 0.004334712401032448, + 0.0067064035683870316, + 0.0039051263593137264, + 0.0028263211715966463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.002214600332081318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195776604115963 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.003195776604115963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195776604115963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04723623767495155, + 0.04145397990942001, + 0.09325911104679108, + 0.08297814428806305, + 0.14701613783836365, + 0.07777813822031021, + 0.4276699125766754, + 0.03565352037549019, + 0.027904020622372627, + 0.01905086077749729 + ], + "credit_map": [ + 0.0056731849908828735, + 0.005658819805830717, + 0.006634677294641733, + 0.006369377486407757, + 0.0054186368361115456, + 0.006877855863422155, + 0.004746510647237301, + 0.005857170093804598, + 0.005484430585056543, + 0.004913314711302519, + 0.005950637627393007, + 0.006891911383718252, + 0.003860844997689128, + 0.006392559502273798, + 0.00531374616548419, + 0.0026465309783816338, + 0.013250072486698627, + 0.00993969477713108, + 0.011180131696164608, + 0.013147294521331787, + 0.00880261231213808, + 0.01237163320183754, + 0.010724460706114769, + 0.013843200169503689, + 0.013949735090136528, + 0.008604669943451881, + 0.009131516329944134, + 0.011936169117689133, + 0.007826460525393486, + 0.011678025126457214, + 0.009363822638988495, + 0.010487743653357029, + 0.019555101171135902, + 0.015720004215836525, + 0.02526203729212284, + 0.016753679141402245, + 0.021079767495393753, + 0.011470968835055828, + 0.019250959157943726, + 0.017923615872859955, + 0.011630269698798656, + 0.00810704194009304, + 0.014589217491447926, + 0.010065500624477863, + 0.007378237787634134, + 0.00878117885440588, + 0.00985364243388176, + 0.0073730433359742165, + 0.06624720245599747, + 0.04289405792951584, + 0.056824665516614914, + 0.0637771338224411, + 0.04703104868531227, + 0.05307651683688164, + 0.040317781269550323, + 0.057501472532749176, + 0.00577304745092988, + 0.0030885578598827124, + 0.006857074797153473, + 0.004244015086442232, + 0.0028755501843988895, + 0.005373306106775999, + 0.004447351209819317, + 0.0029946172144263983, + 0.0036294248420745134, + 0.0027213001158088446, + 0.0051007578149437904, + 0.0031718681566417217, + 0.003009404055774212, + 0.0043251290917396545, + 0.0037052647676318884, + 0.002240868518128991, + 0.0031011125538498163, + 0.0015830739866942167, + 0.0038511536549776793, + 0.0023024235852062702, + 0.0017035247292369604, + 0.0026934773195534945, + 0.0024848224129527807, + 0.0013312717201188207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.003195776604115963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043045587837696075 + ], + "opposite_logprob_tokens": [ + -3.1680455207824707 + ], + "logprob_reward": -0.043045587837696075, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043045587837696075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08022740483283997, + 0.03672000393271446, + 0.08471348136663437, + 0.3724495470523834, + 0.11722570657730103, + 0.10774952918291092, + 0.12511642277240753, + 0.024027058854699135, + 0.016170259565114975, + 0.03560054302215576 + ], + "credit_map": [ + 0.009870283305644989, + 0.007741944398730993, + 0.011300715617835522, + 0.01130293495953083, + 0.007793364115059376, + 0.011378306895494461, + 0.00881521962583065, + 0.012024644762277603, + 0.005056827329099178, + 0.0033692147117108107, + 0.00495163956657052, + 0.004659644328057766, + 0.00338687258772552, + 0.005819287151098251, + 0.004940730519592762, + 0.0045357923954725266, + 0.013490978628396988, + 0.010422530584037304, + 0.009505830705165863, + 0.008216924965381622, + 0.005993455182760954, + 0.010749532841145992, + 0.008180317468941212, + 0.01815391704440117, + 0.06290948390960693, + 0.039500921964645386, + 0.052579496055841446, + 0.048374027013778687, + 0.03834755718708038, + 0.04940998926758766, + 0.03929867967963219, + 0.04202941060066223, + 0.015580837614834309, + 0.005977231077849865, + 0.016077086329460144, + 0.01784336194396019, + 0.010687240399420261, + 0.006838010624051094, + 0.023127425462007523, + 0.02109452709555626, + 0.01629634201526642, + 0.009028557687997818, + 0.017434529960155487, + 0.016073143109679222, + 0.008956297300755978, + 0.008669059723615646, + 0.013384787365794182, + 0.01790682040154934, + 0.03216273710131645, + 0.006682298611849546, + 0.010714576579630375, + 0.010203763842582703, + 0.007201099302619696, + 0.02783278003334999, + 0.01690671592950821, + 0.013412458822131157, + 0.003512782044708729, + 0.002811718499287963, + 0.003167065791785717, + 0.002360959304496646, + 0.002987261861562729, + 0.002929042326286435, + 0.0029298397712409496, + 0.0033283920492976904, + 0.0024871700443327427, + 0.0010968251153826714, + 0.0028996977489441633, + 0.002263045869767666, + 0.0023345169611275196, + 0.0026478220243006945, + 0.0010321368463337421, + 0.0014090457698330283, + 0.004878930281847715, + 0.002965566935017705, + 0.005511811934411526, + 0.005674944259226322, + 0.0040128109976649284, + 0.005217956844717264, + 0.003948773257434368, + 0.0033897513058036566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.043045587837696075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male golden pheasant displaying courtship behavior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the male pheasant have a brilliant golden-yellow crown, metallic green upper back, and deep red chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the male pheasant raise an intricate light orange mantle with black borders to form a semicircle around his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mottled brown hen visible in the scene watching the male pheasant?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of deep green ferns, dark tree trunks, and fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817234955728054 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817234955728054, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817234955728054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1360328197479248, + 0.05953025817871094, + 0.04982025548815727, + 0.08059895783662796, + 0.06925039738416672, + 0.35997089743614197, + 0.038946107029914856, + 0.12043130397796631, + 0.026032721623778343, + 0.0593862347304821 + ], + "credit_map": [ + 0.016137927770614624, + 0.013090087100863457, + 0.024723093956708908, + 0.017907345667481422, + 0.010392400436103344, + 0.02156059257686138, + 0.017077172175049782, + 0.01514421496540308, + 0.00660571176558733, + 0.006098724436014891, + 0.0053572584874928, + 0.008738168515264988, + 0.004641485400497913, + 0.008616985753178596, + 0.008735781535506248, + 0.010736143216490746, + 0.006367581896483898, + 0.006927003152668476, + 0.009443391114473343, + 0.0066141062416136265, + 0.003881406970322132, + 0.006585200782865286, + 0.0054548438638448715, + 0.004546718206256628, + 0.012014392763376236, + 0.00818732287734747, + 0.012882058508694172, + 0.012043052352964878, + 0.00605533504858613, + 0.009561467915773392, + 0.007532516028732061, + 0.012322807684540749, + 0.01075612474232912, + 0.005850245710462332, + 0.011635832488536835, + 0.00994198676198721, + 0.009470456279814243, + 0.004415812436491251, + 0.010158197954297066, + 0.007021742407232523, + 0.05242893099784851, + 0.029657430946826935, + 0.062103450298309326, + 0.05545743554830551, + 0.027392780408263206, + 0.029479941353201866, + 0.04523119702935219, + 0.05821973830461502, + 0.01075062993913889, + 0.002624586457386613, + 0.004388278350234032, + 0.004299126099795103, + 0.0027201503980904818, + 0.008212310262024403, + 0.0033533982932567596, + 0.0025976288598030806, + 0.020431693643331528, + 0.009680407121777534, + 0.021716613322496414, + 0.0173520278185606, + 0.007858031429350376, + 0.019894815981388092, + 0.013488784432411194, + 0.010008920915424824, + 0.004476976580917835, + 0.002562391571700573, + 0.0037392221856862307, + 0.0030644452199339867, + 0.0033155891578644514, + 0.0037937695160508156, + 0.002872808137908578, + 0.002207520417869091, + 0.009698711335659027, + 0.003059518989175558, + 0.012623686343431473, + 0.008643600158393383, + 0.005770839285105467, + 0.00901520811021328, + 0.005344443954527378, + 0.005230223760008812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.002817234955728054 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_8/index_tensor([1906, 1906])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_8/index_tensor([1906, 1906])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5b8868e0ac28a02394f2a0b7b768e0fae94ee701 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_8/index_tensor([1906, 1906])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly elevated close-up meticulously captures a quiet moment of transformation. A lone cube of melting ice rests within a heavy, clear glass partition, resting on a polished, dark countertop. Soft, directed light from an unseen source glints along the ice's damp, rounded edges, illuminating its inner crystalline structure and the delicate condensation on the glass's exterior. At the bottom of the glass, a small amount of perfectly transparent meltwater pools, partially submerging the cube's base. The action unfolds with a deliberate rhythm: as the base erodes, the ice cube begins to slowly tilt, its center of gravity shifting precariously. This slow tilt culminates in a sudden, fluid whirl in the water, causing it to sway gently for a moment before coming to complete stillness in its new orientation, the tiny ripples it created slowly dissipating into stillness. The entire scene is rendered in a cool, minimalist palette, creating an atmosphere of focused tranquility that emphasizes the simple, inevitable physics of melting.", + "seed": 6121, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_8/index_1906_seed_6121_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0205430697882548, + "vlm_logprob": -0.0205430697882548, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.0205430697882548, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017849527299404144 + ], + "opposite_logprob_tokens": [ + -5.017849445343018 + ], + "logprob_reward": -0.017849527299404144, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.017849527299404144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2311081439256668, + 0.07795386016368866, + 0.04617835953831673, + 0.02055310271680355, + 0.03038923256099224, + 0.29870936274528503, + 0.0568632036447525, + 0.0826646015048027, + 0.04924881458282471, + 0.10633133351802826 + ], + "credit_map": [ + 0.03339985013008118, + 0.02845478616654873, + 0.05346265435218811, + 0.02241409756243229, + 0.02140885405242443, + 0.024205129593610764, + 0.022157587110996246, + 0.025605181232094765, + 0.011936679482460022, + 0.007959232665598392, + 0.016522793099284172, + 0.007749656215310097, + 0.008035704493522644, + 0.008428598754107952, + 0.008431420661509037, + 0.008889773860573769, + 0.007301716599613428, + 0.006000079680234194, + 0.008973423391580582, + 0.006113937124609947, + 0.003400432877242565, + 0.004853919148445129, + 0.00407113041728735, + 0.005463721696287394, + 0.0031991791911423206, + 0.0023164867889136076, + 0.00404723035171628, + 0.002230030484497547, + 0.0024192542769014835, + 0.0019484033109620214, + 0.00231249094940722, + 0.0020800279453396797, + 0.004186567384749651, + 0.002964333165436983, + 0.005762090440839529, + 0.0023467345163226128, + 0.0032366206869482994, + 0.00406674575060606, + 0.003295899135991931, + 0.0045302389189600945, + 0.04136556386947632, + 0.03192083537578583, + 0.08077710121870041, + 0.020555684342980385, + 0.02780476026237011, + 0.03530081361532211, + 0.027677886188030243, + 0.033306729048490524, + 0.006632497068494558, + 0.008058157749474049, + 0.010920746251940727, + 0.005458058789372444, + 0.005359884351491928, + 0.006674631033092737, + 0.006721768528223038, + 0.0070374589413404465, + 0.009654239751398563, + 0.014404061250388622, + 0.009961449541151524, + 0.00719668623059988, + 0.00978703424334526, + 0.011736346408724785, + 0.009133599698543549, + 0.010791180655360222, + 0.004834401421248913, + 0.009876837022602558, + 0.0064170691184699535, + 0.004399782512336969, + 0.005143691319972277, + 0.005933045409619808, + 0.0067414576187729836, + 0.005902530625462532, + 0.009503614157438278, + 0.011956467293202877, + 0.007770578842610121, + 0.008328355848789215, + 0.01459866389632225, + 0.014106705784797668, + 0.02375616692006588, + 0.016310783103108406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.017849527299404144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089714244008064 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089714244008064, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089714244008064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09148555994033813, + 0.046572353690862656, + 0.0741468146443367, + 0.03441634029150009, + 0.06609701365232468, + 0.0999830812215805, + 0.11608822643756866, + 0.08791015297174454, + 0.08064116537570953, + 0.3026592433452606 + ], + "credit_map": [ + 0.014694361947476864, + 0.012205231934785843, + 0.024755826219916344, + 0.008074196986854076, + 0.007972802966833115, + 0.011214894242584705, + 0.005602579563856125, + 0.006965665612369776, + 0.007677948568016291, + 0.006774735637009144, + 0.010235232301056385, + 0.0038128518499433994, + 0.0036524138413369656, + 0.007009843830019236, + 0.003256869735196233, + 0.00415245583280921, + 0.010588588193058968, + 0.008125659078359604, + 0.012076343409717083, + 0.007890356704592705, + 0.006481430493295193, + 0.010323088616132736, + 0.00691038416698575, + 0.011750966310501099, + 0.004125102423131466, + 0.0037015704438090324, + 0.006798817776143551, + 0.0031362620647996664, + 0.0031112264841794968, + 0.0043690819293260574, + 0.0036044870503246784, + 0.005569793749600649, + 0.011565109714865685, + 0.006685630884021521, + 0.01658354513347149, + 0.004860397893935442, + 0.005976410582661629, + 0.009785025380551815, + 0.0046478379517793655, + 0.005993060301989317, + 0.013378029689192772, + 0.012048131786286831, + 0.02627529390156269, + 0.007138972170650959, + 0.009045942686498165, + 0.011448333039879799, + 0.009917707182466984, + 0.010730678215622902, + 0.013149728998541832, + 0.01430582907050848, + 0.0260919202119112, + 0.006600202992558479, + 0.009370465762913227, + 0.013222168199717999, + 0.014947529882192612, + 0.018400387838482857, + 0.009723334573209286, + 0.007927867583930492, + 0.01358376257121563, + 0.005090251099318266, + 0.007418311201035976, + 0.01363070122897625, + 0.01523691788315773, + 0.015299004502594471, + 0.0059381756000220776, + 0.008613894693553448, + 0.011332863941788673, + 0.0057902890257537365, + 0.007742127403616905, + 0.00933178048580885, + 0.015141811221837997, + 0.01675022579729557, + 0.02435532957315445, + 0.024954481050372124, + 0.022960206493735313, + 0.019996797665953636, + 0.03677833452820778, + 0.04091485217213631, + 0.062470436096191406, + 0.07022877782583237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.011089714244008064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787900373339653 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.009787900373339653, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787900373339653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04748789966106415, + 0.02022840827703476, + 0.015228820964694023, + 0.012076900340616703, + 0.08086831867694855, + 0.31267622113227844, + 0.373855322599411, + 0.04048154875636101, + 0.02873663417994976, + 0.06835988909006119 + ], + "credit_map": [ + 0.005876067094504833, + 0.006142258178442717, + 0.005746562033891678, + 0.004509580321609974, + 0.0062999427318573, + 0.006468884646892548, + 0.006046057678759098, + 0.0063985479064285755, + 0.0028936201706528664, + 0.0036601340398192406, + 0.003288879292085767, + 0.001978413201868534, + 0.0020561395213007927, + 0.002156416652724147, + 0.0019680242985486984, + 0.0022267811000347137, + 0.00198102043941617, + 0.0018941494636237621, + 0.0024760481901466846, + 0.0021112924441695213, + 0.0015169082907959819, + 0.0017852452583611012, + 0.0016251321649178863, + 0.0018390247132629156, + 0.0013755055842921138, + 0.001888349885120988, + 0.0016358067514374852, + 0.0015130589017644525, + 0.001430082367733121, + 0.001287625404074788, + 0.001371245482005179, + 0.0015752263134345412, + 0.012546387501060963, + 0.009233121760189533, + 0.019331185147166252, + 0.006213881075382233, + 0.007506129797548056, + 0.011207330040633678, + 0.006268239580094814, + 0.008562043309211731, + 0.04457486793398857, + 0.03492621332406998, + 0.08283267915248871, + 0.020040031522512436, + 0.025771725922822952, + 0.03785211965441704, + 0.0297840666025877, + 0.03689451888203621, + 0.041503362357616425, + 0.04996803030371666, + 0.08454614132642746, + 0.02091880328953266, + 0.032036323100328445, + 0.04493509978055954, + 0.04359721019864082, + 0.056350354105234146, + 0.005169976502656937, + 0.005598731338977814, + 0.006614042911678553, + 0.0032006825786083937, + 0.0028169897850602865, + 0.005909269209951162, + 0.005501948297023773, + 0.0056699118576943874, + 0.002762860618531704, + 0.005147107411175966, + 0.003831298090517521, + 0.0030896442476660013, + 0.002658932702615857, + 0.0032327331136912107, + 0.004106997046619654, + 0.003907059319317341, + 0.005379365291446447, + 0.005592226516455412, + 0.005984585266560316, + 0.004688260145485401, + 0.008738714270293713, + 0.00953406747430563, + 0.014581413939595222, + 0.013861254788935184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.009787900373339653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204436346888542 + ], + "opposite_logprob_tokens": [ + -2.812044382095337 + ], + "logprob_reward": -0.06204436346888542, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204436346888542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14479760825634003, + 0.04186594486236572, + 0.043045610189437866, + 0.020045718178153038, + 0.45134034752845764, + 0.02378394827246666, + 0.17676092684268951, + 0.04345566779375076, + 0.028544042259454727, + 0.026360243558883667 + ], + "credit_map": [ + 0.02244528755545616, + 0.024480147287249565, + 0.031171830371022224, + 0.013869163580238819, + 0.015188072808086872, + 0.013722243718802929, + 0.011096166446805, + 0.012824681587517262, + 0.0058441138826310635, + 0.004949226044118404, + 0.006470091640949249, + 0.004376833327114582, + 0.004543274641036987, + 0.005683224182575941, + 0.004767957609146833, + 0.005231220740824938, + 0.007348677609115839, + 0.006093842908740044, + 0.008084475994110107, + 0.004575217608362436, + 0.003560285782441497, + 0.0063123456202447414, + 0.0037928803358227015, + 0.003277879673987627, + 0.003231537062674761, + 0.001864545396529138, + 0.0035932431928813457, + 0.002609304618090391, + 0.0019940179772675037, + 0.002020334592089057, + 0.0018543015467002988, + 0.0028784300666302443, + 0.07299898564815521, + 0.0459786020219326, + 0.10348881781101227, + 0.036777008324861526, + 0.04235958680510521, + 0.06700189411640167, + 0.03666363283991814, + 0.04607177525758743, + 0.002782347844913602, + 0.002564453287050128, + 0.004580183885991573, + 0.002800969872623682, + 0.001838075346313417, + 0.003181028412654996, + 0.002494852291420102, + 0.0035420339554548264, + 0.019669732078909874, + 0.022585222497582436, + 0.03936098888516426, + 0.010952121578156948, + 0.014071433804929256, + 0.020708998665213585, + 0.020792247727513313, + 0.028620166704058647, + 0.004495434463024139, + 0.00632412638515234, + 0.006784212309867144, + 0.00428648991510272, + 0.002964907558634877, + 0.006476764567196369, + 0.006483318749815226, + 0.005640408489853144, + 0.003159495070576668, + 0.0034644650295376778, + 0.003892984939739108, + 0.0030794881749898195, + 0.003145488677546382, + 0.0034201662056148052, + 0.004462359938770533, + 0.003919590264558792, + 0.0017187574412673712, + 0.0021387736778706312, + 0.002075605094432831, + 0.0022887825034558773, + 0.0031181792728602886, + 0.00443190848454833, + 0.003320103045552969, + 0.007268130779266357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.06204436346888542 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001943843555636704 + ], + "opposite_logprob_tokens": [ + -6.251944065093994 + ], + "logprob_reward": -0.001943843555636704, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001943843555636704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09720096737146378, + 0.06805063039064407, + 0.03581332787871361, + 0.01307704672217369, + 0.12449691444635391, + 0.2953091263771057, + 0.10958178341388702, + 0.07818031311035156, + 0.11873093247413635, + 0.05955902487039566 + ], + "credit_map": [ + 0.01454150304198265, + 0.014283765107393265, + 0.020988352596759796, + 0.009675747714936733, + 0.009501920081675053, + 0.011339467018842697, + 0.008503212593495846, + 0.00836698617786169, + 0.010159600526094437, + 0.010759792290627956, + 0.009163713082671165, + 0.007105725351721048, + 0.008797272108495235, + 0.008681025356054306, + 0.006776189897209406, + 0.006607303861528635, + 0.006919730454683304, + 0.004182226490229368, + 0.006696941796690226, + 0.004039708524942398, + 0.002492142841219902, + 0.00433401158079505, + 0.002651673275977373, + 0.004496891051530838, + 0.0023551760241389275, + 0.001152853132225573, + 0.0029116603545844555, + 0.0012760215904563665, + 0.0011830473085865378, + 0.0014722913037985563, + 0.0008677092846482992, + 0.001858286326751113, + 0.017942290753126144, + 0.01363119576126337, + 0.024909675121307373, + 0.011789180338382721, + 0.013421432115137577, + 0.015873044729232788, + 0.011655369773507118, + 0.015274706296622753, + 0.03987676277756691, + 0.03166652098298073, + 0.08033405989408493, + 0.018452558666467667, + 0.026015639305114746, + 0.035057246685028076, + 0.02793101966381073, + 0.03597528859972954, + 0.012160885147750378, + 0.01396727841347456, + 0.026349598541855812, + 0.0072149960324168205, + 0.009423874318599701, + 0.012389017269015312, + 0.014347424730658531, + 0.01372869685292244, + 0.009723976254463196, + 0.00912250392138958, + 0.013939752243459225, + 0.00746601028367877, + 0.005515615921467543, + 0.01042067352682352, + 0.009406883269548416, + 0.01258489117026329, + 0.012153283692896366, + 0.027908241376280785, + 0.013082277961075306, + 0.011711913160979748, + 0.012782849371433258, + 0.013904151506721973, + 0.014807870611548424, + 0.012380332686007023, + 0.006560611072927713, + 0.008034789003431797, + 0.006032670848071575, + 0.004767719656229019, + 0.006583753507584333, + 0.007992388680577278, + 0.01072247326374054, + 0.008864615112543106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.001943843555636704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly elevated close-up meticulously captures a quiet moment of transformation. A lone cube of melting ice rests within a heavy, clear glass partition, resting on a polished, dark countertop. Soft, directed light from an unseen source glints along the ice's damp, rounded edges, illuminating its inner crystalline structure and the delicate condensation on the glass's exterior. At the bottom of the glass, a small amount of perfectly transparent meltwater pools, partially submerging the cube's base. The action unfolds with a deliberate rhythm: as the base erodes, the ice cube begins to slowly tilt, its center of gravity shifting precariously. This slow tilt culminates in a sudden, fluid whirl in the water, causing it to sway gently for a moment before coming to complete stillness in its new orientation, the tiny ripples it created slowly dissipating into stillness. The entire scene is rendered in a cool, minimalist palette, creating an atmosphere of focused tranquility that emphasizes the simple, inevitable physics of melting.", + "seed": 6122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_8/index_1906_seed_6122_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022080093249678613, + "vlm_logprob": -0.022080093249678613, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.022080093249678613, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009854713454842567 + ], + "opposite_logprob_tokens": [ + -6.134854793548584 + ], + "logprob_reward": -0.009854713454842567, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.009854713454842567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09342874586582184, + 0.09147916734218597, + 0.16883480548858643, + 0.07675953209400177, + 0.09730825573205948, + 0.05656415969133377, + 0.0620652511715889, + 0.07263847440481186, + 0.1275816261768341, + 0.15333989262580872 + ], + "credit_map": [ + 0.01123130228370428, + 0.009033968672156334, + 0.012688891030848026, + 0.010808471590280533, + 0.007742008194327354, + 0.011785266920924187, + 0.01244653481990099, + 0.017692308872938156, + 0.009583263657987118, + 0.009049986489117146, + 0.011779945343732834, + 0.009887482970952988, + 0.013143215328454971, + 0.010963305830955505, + 0.011547164991497993, + 0.015524812042713165, + 0.02037246897816658, + 0.02118898183107376, + 0.011957862414419651, + 0.02138843573629856, + 0.023184554651379585, + 0.01806463487446308, + 0.026053769513964653, + 0.026624122634530067, + 0.010188102722167969, + 0.00821705162525177, + 0.012218169867992401, + 0.006841658148914576, + 0.008883754722774029, + 0.0089788269251585, + 0.008051100187003613, + 0.013380873948335648, + 0.009193566627800465, + 0.012714159674942493, + 0.010045565664768219, + 0.011408112943172455, + 0.014600598253309727, + 0.0127103915438056, + 0.012195675633847713, + 0.014440204948186874, + 0.006207853555679321, + 0.003368111327290535, + 0.004091921262443066, + 0.003955420106649399, + 0.0031554833985865116, + 0.015634577721357346, + 0.002989425789564848, + 0.01716136932373047, + 0.010488857515156269, + 0.005899870302528143, + 0.006569786928594112, + 0.00608304375782609, + 0.005608237348496914, + 0.008258402347564697, + 0.007655809633433819, + 0.011501251719892025, + 0.006990700960159302, + 0.008919843472540379, + 0.008101033978164196, + 0.009578380733728409, + 0.007562519516795874, + 0.008321693167090416, + 0.011317749507725239, + 0.011846563778817654, + 0.013939308933913708, + 0.013853158801794052, + 0.011407042853534222, + 0.01318995002657175, + 0.01088603027164936, + 0.02217692881822586, + 0.015439748764038086, + 0.02668948285281658, + 0.012647313065826893, + 0.01816924661397934, + 0.015118401497602463, + 0.014033230021595955, + 0.020266814157366753, + 0.017887592315673828, + 0.023317357525229454, + 0.03189995139837265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.009854713454842567 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038125645369291306 + ], + "opposite_logprob_tokens": [ + -3.28812575340271 + ], + "logprob_reward": -0.038125645369291306, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038125645369291306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18884432315826416, + 0.0779627338051796, + 0.06517047435045242, + 0.028417006134986877, + 0.030813485383987427, + 0.08473299443721771, + 0.030345192179083824, + 0.04985811188817024, + 0.12370318174362183, + 0.3201524615287781 + ], + "credit_map": [ + 0.019467299804091454, + 0.019795654341578484, + 0.03634437173604965, + 0.021182384341955185, + 0.019507480785250664, + 0.021003631874918938, + 0.02455541491508484, + 0.02698809839785099, + 0.009607836604118347, + 0.008947627618908882, + 0.010038807988166809, + 0.009215131402015686, + 0.011260817758738995, + 0.008096088655292988, + 0.010090716183185577, + 0.010705708526074886, + 0.017202241346240044, + 0.004525603726506233, + 0.007979641668498516, + 0.004282729234546423, + 0.005042567383497953, + 0.010161780752241611, + 0.00514265475794673, + 0.010833258740603924, + 0.006108174100518227, + 0.002470889128744602, + 0.0046098544262349606, + 0.00216533406637609, + 0.0029126389417797327, + 0.004049346782267094, + 0.002122477861121297, + 0.003978292923420668, + 0.0037535270676016808, + 0.0023485629353672266, + 0.0032047885470092297, + 0.0032676528207957745, + 0.003198243910446763, + 0.005545453168451786, + 0.003270357381552458, + 0.006224902346730232, + 0.010392903350293636, + 0.0034513494465500116, + 0.0040737176313996315, + 0.004517838824540377, + 0.003161452244967222, + 0.026040755212306976, + 0.004031839314848185, + 0.029063133522868156, + 0.00358005752786994, + 0.003474657190963626, + 0.002978219883516431, + 0.0029276199638843536, + 0.002696913667023182, + 0.004314737394452095, + 0.00419586431235075, + 0.006177123170346022, + 0.008137360215187073, + 0.0041379076428711414, + 0.00829311553388834, + 0.005539268255233765, + 0.004198702983558178, + 0.005841152276843786, + 0.006065737456083298, + 0.007644871715456247, + 0.016489902511239052, + 0.011997533030807972, + 0.01774396374821663, + 0.01481059193611145, + 0.012866898439824581, + 0.01471810694783926, + 0.01592068187892437, + 0.019155504181981087, + 0.026075590401887894, + 0.039402756839990616, + 0.026057260110974312, + 0.028467701748013496, + 0.04228249937295914, + 0.03660311549901962, + 0.052644357085227966, + 0.06861919164657593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.038125645369291306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046457452699542046 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.0046457452699542046, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046457452699542046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06775008141994476, + 0.07583382725715637, + 0.21819306910037994, + 0.10103451460599899, + 0.10731814056634903, + 0.05141148716211319, + 0.0688772052526474, + 0.045324619859457016, + 0.1421510875225067, + 0.1221059113740921 + ], + "credit_map": [ + 0.009143014438450336, + 0.005823273677378893, + 0.01088695041835308, + 0.008342020213603973, + 0.004731182008981705, + 0.009329129010438919, + 0.006540039554238319, + 0.012954473495483398, + 0.012434610165655613, + 0.009500429034233093, + 0.005970299709588289, + 0.007556738331913948, + 0.009841997176408768, + 0.008251379244029522, + 0.011709732934832573, + 0.01056863646954298, + 0.04270882159471512, + 0.01885305903851986, + 0.02079310081899166, + 0.018848642706871033, + 0.023164812475442886, + 0.030080391094088554, + 0.018628424033522606, + 0.04511582851409912, + 0.009367763064801693, + 0.017112545669078827, + 0.006712910253554583, + 0.014190256595611572, + 0.01664825901389122, + 0.010303865186870098, + 0.013388761319220066, + 0.01331015583127737, + 0.011173241771757603, + 0.014166139997541904, + 0.01289188303053379, + 0.015146812424063683, + 0.01226422656327486, + 0.014050012454390526, + 0.012240174226462841, + 0.015385654754936695, + 0.004008367192000151, + 0.006169550120830536, + 0.004826038144528866, + 0.007657590322196484, + 0.005682636052370071, + 0.007657004985958338, + 0.006329570896923542, + 0.009080731309950352, + 0.00650379853323102, + 0.007369248196482658, + 0.00734113110229373, + 0.007611316628754139, + 0.00674712099134922, + 0.009981434792280197, + 0.009723763912916183, + 0.013599391095340252, + 0.004178254399448633, + 0.0050115445628762245, + 0.004801092203706503, + 0.005196040496230125, + 0.004777777940034866, + 0.006462015211582184, + 0.006754081230610609, + 0.008143812417984009, + 0.01568763144314289, + 0.019538700580596924, + 0.013676496222615242, + 0.012789851985871792, + 0.015933014452457428, + 0.020116273313760757, + 0.019697194918990135, + 0.02471192739903927, + 0.009498964995145798, + 0.015240148641169071, + 0.010555175133049488, + 0.011251102201640606, + 0.016700945794582367, + 0.013869745656847954, + 0.019238855689764023, + 0.02575097791850567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.0046457452699542046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495249107480049 + ], + "opposite_logprob_tokens": [ + -2.929952383041382 + ], + "logprob_reward": -0.05495249107480049, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495249107480049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16632860898971558, + 0.03900394216179848, + 0.15339675545692444, + 0.10623856633901596, + 0.07537668198347092, + 0.03603331372141838, + 0.05254412069916725, + 0.032156653702259064, + 0.21057023108005524, + 0.1283511519432068 + ], + "credit_map": [ + 0.017368055880069733, + 0.018290117383003235, + 0.03593716025352478, + 0.019142096862196922, + 0.017033269628882408, + 0.018939999863505363, + 0.021505063399672508, + 0.01811283268034458, + 0.00586662394925952, + 0.00392262265086174, + 0.0047571309842169285, + 0.004988545551896095, + 0.004802812822163105, + 0.004408982116729021, + 0.005038387142121792, + 0.005218835547566414, + 0.03200973570346832, + 0.011922060512006283, + 0.014653176069259644, + 0.012170727364718914, + 0.014793317764997482, + 0.022462008520960808, + 0.010373239405453205, + 0.035012491047382355, + 0.012584734708070755, + 0.01153853815048933, + 0.015770770609378815, + 0.013154844753444195, + 0.009762097150087357, + 0.012596490792930126, + 0.014647246338427067, + 0.01618383824825287, + 0.008427942171692848, + 0.007637627888470888, + 0.008180123753845692, + 0.008466158993542194, + 0.006705036386847496, + 0.014606904238462448, + 0.008188656531274319, + 0.013164231553673744, + 0.0028584492392838, + 0.005078043323010206, + 0.003037274582311511, + 0.005171246360987425, + 0.0052678207866847515, + 0.00394272617995739, + 0.004344793036580086, + 0.006332959048449993, + 0.004635128192603588, + 0.00500750495120883, + 0.006271100603044033, + 0.006545376963913441, + 0.004624456167221069, + 0.007147838827222586, + 0.008487601764500141, + 0.009825111366808414, + 0.0030616316944360733, + 0.003818337805569172, + 0.005309537518769503, + 0.004463242366909981, + 0.003853334579616785, + 0.003227156586945057, + 0.004044634755700827, + 0.004378773272037506, + 0.02115192636847496, + 0.03213458135724068, + 0.021460898220539093, + 0.027969608083367348, + 0.028464168310165405, + 0.019974621012806892, + 0.03402232378721237, + 0.025392109528183937, + 0.012085864320397377, + 0.015246578492224216, + 0.008815385401248932, + 0.011043229140341282, + 0.017650125548243523, + 0.01530213188380003, + 0.021236808970570564, + 0.026971030980348587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.05495249107480049 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00282187107950449 + ], + "opposite_logprob_tokens": [ + -5.877821922302246 + ], + "logprob_reward": -0.00282187107950449, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00282187107950449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18789976835250854, + 0.0555921271443367, + 0.25626328587532043, + 0.16871215403079987, + 0.05419411137700081, + 0.021947652101516724, + 0.051988985389471054, + 0.018069889396429062, + 0.09912938624620438, + 0.08620268851518631 + ], + "credit_map": [ + 0.020427033305168152, + 0.020826753228902817, + 0.032956548035144806, + 0.02022343873977661, + 0.022544121369719505, + 0.01991487853229046, + 0.02472640387713909, + 0.02628057450056076, + 0.005155911203473806, + 0.008019068278372288, + 0.005748398136347532, + 0.006323587149381638, + 0.0077208722941577435, + 0.005822876933962107, + 0.007114946376532316, + 0.009686467237770557, + 0.05717914551496506, + 0.018144860863685608, + 0.026265988126397133, + 0.017677713185548782, + 0.02164509706199169, + 0.03871459141373634, + 0.01693909801542759, + 0.05969677120447159, + 0.017026159912347794, + 0.020367009565234184, + 0.021302878856658936, + 0.023827973753213882, + 0.018397372215986252, + 0.016927722841501236, + 0.02521788515150547, + 0.025645148009061813, + 0.006412643939256668, + 0.0052187805995345116, + 0.005071316845715046, + 0.005808153189718723, + 0.004566430579870939, + 0.011376061476767063, + 0.005254988558590412, + 0.010485733859241009, + 0.002179962582886219, + 0.0023280081804841757, + 0.00274628191255033, + 0.0026049260050058365, + 0.002115699928253889, + 0.0032710779923945665, + 0.002587971044704318, + 0.004113724455237389, + 0.006598440930247307, + 0.005480221938341856, + 0.005318996962159872, + 0.005433953832834959, + 0.004838751628994942, + 0.008041697554290295, + 0.005889795254915953, + 0.010387125425040722, + 0.002011640463024378, + 0.0026914961636066437, + 0.0020951279439032078, + 0.0020643386524170637, + 0.0016847652150318027, + 0.0018067668424919248, + 0.0032659159041941166, + 0.0024498384445905685, + 0.008517093025147915, + 0.014448374509811401, + 0.010246855206787586, + 0.013641675002872944, + 0.011656354181468487, + 0.0097076166421175, + 0.016694631427526474, + 0.014216783456504345, + 0.007837366312742233, + 0.010165320709347725, + 0.007827025838196278, + 0.00810710247606039, + 0.010600513778626919, + 0.009429200552403927, + 0.014773963950574398, + 0.017462190240621567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.00282187107950449 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly elevated close-up meticulously captures a quiet moment of transformation. A lone cube of melting ice rests within a heavy, clear glass partition, resting on a polished, dark countertop. Soft, directed light from an unseen source glints along the ice's damp, rounded edges, illuminating its inner crystalline structure and the delicate condensation on the glass's exterior. At the bottom of the glass, a small amount of perfectly transparent meltwater pools, partially submerging the cube's base. The action unfolds with a deliberate rhythm: as the base erodes, the ice cube begins to slowly tilt, its center of gravity shifting precariously. This slow tilt culminates in a sudden, fluid whirl in the water, causing it to sway gently for a moment before coming to complete stillness in its new orientation, the tiny ripples it created slowly dissipating into stillness. The entire scene is rendered in a cool, minimalist palette, creating an atmosphere of focused tranquility that emphasizes the simple, inevitable physics of melting.", + "seed": 6123, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_8/index_1906_seed_6123_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05382908792234957, + "vlm_logprob": -0.05382908792234957, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.05382908792234957, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015758272260427475 + ], + "opposite_logprob_tokens": [ + -5.390758037567139 + ], + "logprob_reward": -0.015758272260427475, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.015758272260427475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19352786242961884, + 0.11539434641599655, + 0.05140206217765808, + 0.12045207619667053, + 0.05169904604554176, + 0.1278284788131714, + 0.09376707673072815, + 0.08959927409887314, + 0.06412658095359802, + 0.09220321476459503 + ], + "credit_map": [ + 0.020266825333237648, + 0.03188237547874451, + 0.03288343921303749, + 0.03477971628308296, + 0.019897937774658203, + 0.017016544938087463, + 0.01574598252773285, + 0.02105502039194107, + 0.012406068854033947, + 0.012952692806720734, + 0.01984153874218464, + 0.011172458529472351, + 0.010022449307143688, + 0.011569210328161716, + 0.01129148155450821, + 0.026138439774513245, + 0.0059533799067139626, + 0.004732450470328331, + 0.008986522443592548, + 0.004590674303472042, + 0.005047832150012255, + 0.007451594341546297, + 0.0053450362756848335, + 0.00929456576704979, + 0.018478278070688248, + 0.014226495288312435, + 0.014897387474775314, + 0.01242872979491949, + 0.014657587744295597, + 0.01922672986984253, + 0.013323924504220486, + 0.013212942518293858, + 0.010395549237728119, + 0.004603350069373846, + 0.010224255733191967, + 0.004136697854846716, + 0.004687103908509016, + 0.006626888178288937, + 0.005279873963445425, + 0.00574532151222229, + 0.015955228358507156, + 0.009523585438728333, + 0.02816114015877247, + 0.010988517664372921, + 0.012262559495866299, + 0.019390832632780075, + 0.012145903892815113, + 0.01940070278942585, + 0.01201139111071825, + 0.006975033786147833, + 0.01694706454873085, + 0.009427753277122974, + 0.008965331129729748, + 0.009391254745423794, + 0.014886382967233658, + 0.015162862837314606, + 0.010428172536194324, + 0.007519454695284367, + 0.017158396542072296, + 0.011323457583785057, + 0.007329398766160011, + 0.009656115435063839, + 0.014047264121472836, + 0.012137006036937237, + 0.007898621261119843, + 0.006265427451580763, + 0.01189531572163105, + 0.008928307332098484, + 0.0060936519876122475, + 0.006091338582336903, + 0.008148014545440674, + 0.008805898949503899, + 0.010784492827951908, + 0.01318343635648489, + 0.010924508795142174, + 0.011135312728583813, + 0.008216002956032753, + 0.01276704017072916, + 0.01497291587293148, + 0.010219495743513107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.015758272260427475 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048699744045734406 + ], + "opposite_logprob_tokens": [ + -3.0486998558044434 + ], + "logprob_reward": -0.048699744045734406, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048699744045734406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40217047929763794, + 0.11081994324922562, + 0.0767633467912674, + 0.07625378668308258, + 0.03569947928190231, + 0.032884467393159866, + 0.04486138001084328, + 0.0318986177444458, + 0.08093491196632385, + 0.10771355777978897 + ], + "credit_map": [ + 0.04670199751853943, + 0.06027020514011383, + 0.06494606286287308, + 0.06595215201377869, + 0.04167693480849266, + 0.03711593151092529, + 0.04768392816185951, + 0.03782331198453903, + 0.013808517716825008, + 0.01120410393923521, + 0.01914760284125805, + 0.01370236650109291, + 0.008357040584087372, + 0.011906009167432785, + 0.013536711223423481, + 0.019157588481903076, + 0.008363754488527775, + 0.007919184863567352, + 0.011703749187290668, + 0.008265258744359016, + 0.006678539793938398, + 0.010983973741531372, + 0.007741487585008144, + 0.015107402577996254, + 0.01393115520477295, + 0.008645255118608475, + 0.010077014565467834, + 0.007573565002530813, + 0.00793441105633974, + 0.012992330826818943, + 0.0077657452784478664, + 0.007334314752370119, + 0.004497411195188761, + 0.0045211161486804485, + 0.0043914709240198135, + 0.003529716981574893, + 0.0035016187466681004, + 0.005002692807465792, + 0.005276846699416637, + 0.0049786074087023735, + 0.006126926746219397, + 0.002769511193037033, + 0.004416667856276035, + 0.00352860102429986, + 0.004178598523139954, + 0.0030348319560289383, + 0.005001758690923452, + 0.00382757605984807, + 0.008084340021014214, + 0.003128225915133953, + 0.009253975003957748, + 0.004635374061763287, + 0.003308698069304228, + 0.003704636124894023, + 0.007249778136610985, + 0.0054963575676083565, + 0.006723771803081036, + 0.0029578059911727905, + 0.005148086231201887, + 0.0034082343336194754, + 0.0027112700045108795, + 0.003024506848305464, + 0.003420251188799739, + 0.004504697863012552, + 0.0071531618013978004, + 0.009510382078588009, + 0.012151876464486122, + 0.010531517677009106, + 0.008213050663471222, + 0.011221268214285374, + 0.012997633777558804, + 0.00915602594614029, + 0.014815540984272957, + 0.015929149463772774, + 0.012842349708080292, + 0.013353224843740463, + 0.010471601039171219, + 0.011770572513341904, + 0.018700871616601944, + 0.009830260649323463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.048699744045734406 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292966187000275 + ], + "opposite_logprob_tokens": [ + -3.7732930183410645 + ], + "logprob_reward": -0.023292966187000275, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292966187000275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19508551061153412, + 0.08172182738780975, + 0.08952385932207108, + 0.06790704280138016, + 0.07726240903139114, + 0.05066683515906334, + 0.16670463979244232, + 0.06835342943668365, + 0.15480312705039978, + 0.04797128215432167 + ], + "credit_map": [ + 0.022510509938001633, + 0.026283061131834984, + 0.0303067397326231, + 0.023067019879817963, + 0.022724485024809837, + 0.025153126567602158, + 0.02465054579079151, + 0.02039002813398838, + 0.012677651830017567, + 0.011943687684834003, + 0.011666837148368359, + 0.008861273527145386, + 0.006968068890273571, + 0.011758454144001007, + 0.009656419977545738, + 0.008189432322978973, + 0.011228160932660103, + 0.012636951170861721, + 0.013247101567685604, + 0.009559530764818192, + 0.009476091712713242, + 0.01080748438835144, + 0.013496878556907177, + 0.009071665816009045, + 0.005166618153452873, + 0.011193905025720596, + 0.008903522044420242, + 0.008120349608361721, + 0.00826909951865673, + 0.007097126916050911, + 0.011939192190766335, + 0.007217233534902334, + 0.013565762899816036, + 0.006176042836159468, + 0.012377172708511353, + 0.0067728678695857525, + 0.007252751849591732, + 0.013749218545854092, + 0.01023343950510025, + 0.007135148625820875, + 0.005367056000977755, + 0.005175734870135784, + 0.006572391837835312, + 0.006362142972648144, + 0.006994024384766817, + 0.006428739987313747, + 0.008475684560835361, + 0.00529106380417943, + 0.016786299645900726, + 0.01532246544957161, + 0.02209353633224964, + 0.020361125469207764, + 0.01839226856827736, + 0.015652010217308998, + 0.038689181208610535, + 0.01940775476396084, + 0.00657498138025403, + 0.008888469077646732, + 0.01032839622348547, + 0.009099764749407768, + 0.008561304770410061, + 0.007719275541603565, + 0.010823794640600681, + 0.0063574425876140594, + 0.015057162381708622, + 0.016972750425338745, + 0.020983483642339706, + 0.023078273981809616, + 0.025068284943699837, + 0.01678609848022461, + 0.021388905122876167, + 0.015468166209757328, + 0.0038229450583457947, + 0.0060989223420619965, + 0.005749003496021032, + 0.0047861384227871895, + 0.005581954028457403, + 0.007194819860160351, + 0.008617266081273556, + 0.006120232865214348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.023292966187000275 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987950146198273 + ], + "opposite_logprob_tokens": [ + -1.8048795461654663 + ], + "logprob_reward": -0.17987950146198273, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987950146198273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18165788054466248, + 0.059663280844688416, + 0.0661395788192749, + 0.048848655074834824, + 0.045943986624479294, + 0.07184354960918427, + 0.10880020260810852, + 0.08386960625648499, + 0.20870092511177063, + 0.12453232705593109 + ], + "credit_map": [ + 0.019172566011548042, + 0.02481057122349739, + 0.026275672018527985, + 0.02280384674668312, + 0.022554725408554077, + 0.0248061902821064, + 0.022542506456375122, + 0.018691806122660637, + 0.007376769091933966, + 0.006113341543823481, + 0.010185512714087963, + 0.006355087272822857, + 0.004544754046946764, + 0.005949764512479305, + 0.007421094924211502, + 0.011716955341398716, + 0.006619976367801428, + 0.009784305468201637, + 0.010891549289226532, + 0.007120279595255852, + 0.007079497445374727, + 0.007312683388590813, + 0.0071355304680764675, + 0.010195759125053883, + 0.005385439842939377, + 0.008105509914457798, + 0.006631162948906422, + 0.005560509394854307, + 0.004516421351581812, + 0.0039046502206474543, + 0.008099593222141266, + 0.006645361427217722, + 0.003819676348939538, + 0.006512704771012068, + 0.005840130150318146, + 0.003938755486160517, + 0.0063361250795423985, + 0.006816498003900051, + 0.006162677425891161, + 0.006517421454191208, + 0.009212437085807323, + 0.006633928045630455, + 0.009617957286536694, + 0.008270875550806522, + 0.010676891542971134, + 0.010771093890070915, + 0.01006270945072174, + 0.006597661878913641, + 0.012058909982442856, + 0.00858974363654852, + 0.019409339874982834, + 0.012036378495395184, + 0.012654488906264305, + 0.012132343836128712, + 0.018311914056539536, + 0.013607081025838852, + 0.008967298083007336, + 0.008131911978125572, + 0.01285935752093792, + 0.010366094298660755, + 0.010055684484541416, + 0.010710122063755989, + 0.012068529613316059, + 0.01071060262620449, + 0.019732557237148285, + 0.024536872282624245, + 0.026628613471984863, + 0.022169651463627815, + 0.032048773020505905, + 0.020123157650232315, + 0.04068078473210335, + 0.02278052270412445, + 0.013399350456893444, + 0.01486179605126381, + 0.01778390072286129, + 0.01190847996622324, + 0.014496047981083393, + 0.01936333440244198, + 0.015857169404625893, + 0.01686224341392517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.17987950146198273 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015149556566029787 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015149556566029787, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015149556566029787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23390811681747437, + 0.11801240593194962, + 0.05205857381224632, + 0.0449090413749218, + 0.06299683451652527, + 0.10982219874858856, + 0.07143855094909668, + 0.05085218325257301, + 0.16127333045005798, + 0.0947287529706955 + ], + "credit_map": [ + 0.03009629435837269, + 0.0321168452501297, + 0.034664008766412735, + 0.025428542867302895, + 0.023932626470923424, + 0.04103061184287071, + 0.02156611531972885, + 0.02507307380437851, + 0.0182273481041193, + 0.015493090264499187, + 0.0171543937176466, + 0.012395479716360569, + 0.009876170195639133, + 0.017868272960186005, + 0.012237864546477795, + 0.014759794808924198, + 0.008020042441785336, + 0.006195425055921078, + 0.008115272037684917, + 0.00495537742972374, + 0.004937739111483097, + 0.005572533700615168, + 0.004766392987221479, + 0.009495791979134083, + 0.005018596071749926, + 0.005477453116327524, + 0.006036711391061544, + 0.005547912325710058, + 0.005992860998958349, + 0.004397629294544458, + 0.00849093683063984, + 0.003946941811591387, + 0.008751258254051208, + 0.006979009136557579, + 0.009218182414770126, + 0.006347276736050844, + 0.006785617209970951, + 0.006985445041209459, + 0.011077135801315308, + 0.006852904800325632, + 0.0168598685413599, + 0.008901670575141907, + 0.025496717542409897, + 0.008034924045205116, + 0.010776716284453869, + 0.014391040429472923, + 0.010771619156002998, + 0.014589635655283928, + 0.012206809595227242, + 0.0054852026514709, + 0.016226179897785187, + 0.007434670813381672, + 0.005225128028541803, + 0.005779519211500883, + 0.010565942153334618, + 0.008515099063515663, + 0.00526968389749527, + 0.005813684314489365, + 0.0083160400390625, + 0.006642547901719809, + 0.006327119655907154, + 0.006046966649591923, + 0.00710676284506917, + 0.005329377017915249, + 0.014413179829716682, + 0.013273455202579498, + 0.027714014053344727, + 0.026589930057525635, + 0.0228558462113142, + 0.015648912638425827, + 0.02136830985546112, + 0.019409693777561188, + 0.013414768502116203, + 0.012758765369653702, + 0.014921731315553188, + 0.00967276282608509, + 0.011473811231553555, + 0.012240875512361526, + 0.01030806079506874, + 0.009937983006238937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0015149556566029787 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly elevated close-up meticulously captures a quiet moment of transformation. A lone cube of melting ice rests within a heavy, clear glass partition, resting on a polished, dark countertop. Soft, directed light from an unseen source glints along the ice's damp, rounded edges, illuminating its inner crystalline structure and the delicate condensation on the glass's exterior. At the bottom of the glass, a small amount of perfectly transparent meltwater pools, partially submerging the cube's base. The action unfolds with a deliberate rhythm: as the base erodes, the ice cube begins to slowly tilt, its center of gravity shifting precariously. This slow tilt culminates in a sudden, fluid whirl in the water, causing it to sway gently for a moment before coming to complete stillness in its new orientation, the tiny ripples it created slowly dissipating into stillness. The entire scene is rendered in a cool, minimalist palette, creating an atmosphere of focused tranquility that emphasizes the simple, inevitable physics of melting.", + "seed": 6124, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_8/index_1906_seed_6124_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027799622155725956, + "vlm_logprob": -0.027799622155725956, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.027799622155725956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011759249493479729 + ], + "opposite_logprob_tokens": [ + -5.511759281158447 + ], + "logprob_reward": -0.011759249493479729, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.011759249493479729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0792202353477478, + 0.05120306462049484, + 0.04601496458053589, + 0.03204802796244621, + 0.0950312688946724, + 0.09772395342588425, + 0.1863194853067398, + 0.2144589126110077, + 0.11932128667831421, + 0.07865877449512482 + ], + "credit_map": [ + 0.015987345948815346, + 0.006854129955172539, + 0.014569965191185474, + 0.008197810500860214, + 0.007548464462161064, + 0.009358344599604607, + 0.007701536640524864, + 0.009002652950584888, + 0.007769681513309479, + 0.006343376357108355, + 0.00948418490588665, + 0.00516865449026227, + 0.0037295627407729626, + 0.005706914234906435, + 0.005880094598978758, + 0.007120602764189243, + 0.006193927023559809, + 0.006617926526814699, + 0.006794482935220003, + 0.004052134696394205, + 0.005852704402059317, + 0.0057056499645113945, + 0.005068476311862469, + 0.0057296655140817165, + 0.004234991502016783, + 0.0041245040483772755, + 0.004254043102264404, + 0.00230417144484818, + 0.003574419068172574, + 0.005705707706511021, + 0.003086518496274948, + 0.004763673525303602, + 0.017774181440472603, + 0.009610051289200783, + 0.00914368499070406, + 0.008197084069252014, + 0.009628777392208576, + 0.025898827239871025, + 0.0079593975096941, + 0.0068192752078175545, + 0.01090823020786047, + 0.014241673052310944, + 0.014668481424450874, + 0.010831162333488464, + 0.012789994478225708, + 0.011613292619585991, + 0.012740427628159523, + 0.009930693544447422, + 0.030315212905406952, + 0.019240571185946465, + 0.019305549561977386, + 0.01702679507434368, + 0.027428671717643738, + 0.045537881553173065, + 0.016633840277791023, + 0.010830983519554138, + 0.037126366049051285, + 0.016321443021297455, + 0.01915663108229637, + 0.018724627792835236, + 0.034222740679979324, + 0.05877337604761124, + 0.016729848459362984, + 0.013403882272541523, + 0.016716793179512024, + 0.010220950469374657, + 0.009001370519399643, + 0.013680406846106052, + 0.034399036318063736, + 0.015171588398516178, + 0.01195053942501545, + 0.00818061362951994, + 0.0176065843552351, + 0.006758940406143665, + 0.007766835857182741, + 0.005910937208682299, + 0.01844380423426628, + 0.008048363961279392, + 0.006702489219605923, + 0.007420823443681002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.011759249493479729 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810475766658783 + ], + "opposite_logprob_tokens": [ + -3.288104772567749 + ], + "logprob_reward": -0.03810475766658783, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810475766658783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16820162534713745, + 0.07689034193754196, + 0.08987779915332794, + 0.03116408735513687, + 0.13513760268688202, + 0.15283043682575226, + 0.07601366937160492, + 0.1906546801328659, + 0.03814132511615753, + 0.04108842462301254 + ], + "credit_map": [ + 0.023415591567754745, + 0.016393210738897324, + 0.033495474606752396, + 0.01594727300107479, + 0.016759123653173447, + 0.019302034750580788, + 0.021021965891122818, + 0.02186696231365204, + 0.013665162958204746, + 0.006872738245874643, + 0.013359407894313335, + 0.006720785051584244, + 0.005620240233838558, + 0.009164455346763134, + 0.009858806617558002, + 0.011628750711679459, + 0.011987991631031036, + 0.010234447196125984, + 0.018801093101501465, + 0.008036855608224869, + 0.007088337559252977, + 0.012214386835694313, + 0.007625246420502663, + 0.013889440335333347, + 0.004291093908250332, + 0.005673368461430073, + 0.004604885820299387, + 0.0025537230540066957, + 0.003445914713665843, + 0.003324890276417136, + 0.0030235762242227793, + 0.004246635362505913, + 0.0189648550003767, + 0.009805707260966301, + 0.028816349804401398, + 0.006630569230765104, + 0.008536280132830143, + 0.02255156636238098, + 0.01199878565967083, + 0.0278334878385067, + 0.0318041667342186, + 0.009140660986304283, + 0.01691592112183571, + 0.008381924591958523, + 0.011921745724976063, + 0.050588518381118774, + 0.008951926603913307, + 0.015125561505556107, + 0.014314444735646248, + 0.006000300403684378, + 0.008104904554784298, + 0.005706921219825745, + 0.007968435995280743, + 0.016807248815894127, + 0.006535205524414778, + 0.010576206259429455, + 0.04186335206031799, + 0.01378193311393261, + 0.02198365516960621, + 0.01140348706394434, + 0.013233405537903309, + 0.05241868272423744, + 0.012566651217639446, + 0.02340351790189743, + 0.006447760853916407, + 0.0036051974166184664, + 0.005769442766904831, + 0.0043104467913508415, + 0.0061106267385184765, + 0.0036554262042045593, + 0.0034718511160463095, + 0.004770571831613779, + 0.0070133572444319725, + 0.002799402456730604, + 0.004698425531387329, + 0.0019397275755181909, + 0.00533299008384347, + 0.006688928697258234, + 0.00260070338845253, + 0.01001488883048296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.03810475766658783 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643358945846558 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643358945846558, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643358945846558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0661211833357811, + 0.03299316018819809, + 0.04782316833734512, + 0.025861358270049095, + 0.05295578017830849, + 0.06948524713516235, + 0.18659575283527374, + 0.14161236584186554, + 0.18310560286045074, + 0.19344639778137207 + ], + "credit_map": [ + 0.009620077908039093, + 0.00856205727905035, + 0.012559748254716396, + 0.005428547039628029, + 0.008445080369710922, + 0.009142246097326279, + 0.005937953479588032, + 0.006425473839044571, + 0.005917772650718689, + 0.0044593592174351215, + 0.005861269775778055, + 0.003936599940061569, + 0.0028119285125285387, + 0.0032186764292418957, + 0.0034892645198851824, + 0.0032982900738716125, + 0.007645257283002138, + 0.0055096326395869255, + 0.008154149167239666, + 0.004966736771166325, + 0.005865699611604214, + 0.006707610562443733, + 0.003982670605182648, + 0.0049914103001356125, + 0.004370959009975195, + 0.0037180474027991295, + 0.0038539220113307238, + 0.0029706391505897045, + 0.002696720650419593, + 0.002787692239508033, + 0.002955584554001689, + 0.002507793949916959, + 0.007552149705588818, + 0.007641713134944439, + 0.006475093774497509, + 0.005561284255236387, + 0.005978334695100784, + 0.005752382334321737, + 0.0068297963589429855, + 0.007165024988353252, + 0.011555973440408707, + 0.008473190478980541, + 0.008095839992165565, + 0.006494712550193071, + 0.006994165480136871, + 0.016863243654370308, + 0.005635295994579792, + 0.005372826475650072, + 0.02444872260093689, + 0.024895844981074333, + 0.022276565432548523, + 0.021345488727092743, + 0.0302725862711668, + 0.029403170570731163, + 0.020825808867812157, + 0.013127570040524006, + 0.02666560746729374, + 0.013478448614478111, + 0.011056413874030113, + 0.011713983491063118, + 0.017044207081198692, + 0.03822974115610123, + 0.01267118752002716, + 0.010752764530479908, + 0.02311869151890278, + 0.02064775675535202, + 0.02476911060512066, + 0.019228830933570862, + 0.04303648695349693, + 0.025676602497696877, + 0.016075290739536285, + 0.010552828200161457, + 0.03674944490194321, + 0.01918899454176426, + 0.019712941721081734, + 0.014922326430678368, + 0.05759397894144058, + 0.012459032237529755, + 0.02112189680337906, + 0.01169779896736145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.008643358945846558 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897567003965378 + ], + "opposite_logprob_tokens": [ + -2.5789756774902344 + ], + "logprob_reward": -0.07897567003965378, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897567003965378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09615480899810791, + 0.07354278862476349, + 0.0491018109023571, + 0.043967947363853455, + 0.10477237403392792, + 0.09600023180246353, + 0.13517771661281586, + 0.1165776252746582, + 0.1972046047449112, + 0.08750003576278687 + ], + "credit_map": [ + 0.013953079469501972, + 0.012861744500696659, + 0.015102575533092022, + 0.008218992501497269, + 0.01144278235733509, + 0.01341293752193451, + 0.01165931485593319, + 0.00950339064002037, + 0.010371998883783817, + 0.01127940509468317, + 0.011819767765700817, + 0.008175256662070751, + 0.007571194786578417, + 0.008892275393009186, + 0.009169794619083405, + 0.006263106130063534, + 0.0072376178577542305, + 0.005396086256951094, + 0.007355718407779932, + 0.00780476164072752, + 0.004660386126488447, + 0.004897637292742729, + 0.006269556935876608, + 0.005480046849697828, + 0.006255956832319498, + 0.006039183121174574, + 0.005985327996313572, + 0.004568231757730246, + 0.0052879927679896355, + 0.005086934193968773, + 0.0056786430068314075, + 0.005065679084509611, + 0.01670742779970169, + 0.014746776781976223, + 0.012868535704910755, + 0.0123505974188447, + 0.012896522879600525, + 0.01678120344877243, + 0.009718035347759724, + 0.008703287690877914, + 0.01616460457444191, + 0.009673049673438072, + 0.014173371717333794, + 0.010294744744896889, + 0.010641735047101974, + 0.01760886050760746, + 0.010220452211797237, + 0.007223418448120356, + 0.024588579311966896, + 0.015554476529359818, + 0.013590233400464058, + 0.011853729374706745, + 0.02165798656642437, + 0.028778929263353348, + 0.012059992179274559, + 0.007093807216733694, + 0.01445009745657444, + 0.014088405296206474, + 0.01278156228363514, + 0.012597739696502686, + 0.016611594706773758, + 0.02307206019759178, + 0.012875877320766449, + 0.010100295767188072, + 0.022308548912405968, + 0.019501250237226486, + 0.02104959823191166, + 0.021071312949061394, + 0.0394970066845417, + 0.028423292562365532, + 0.020609108731150627, + 0.024744519963860512, + 0.012455830350518227, + 0.011067565530538559, + 0.013444612734019756, + 0.008320032618939877, + 0.012956635095179081, + 0.010312649421393871, + 0.007244723383337259, + 0.011697995476424694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.07897567003965378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a lone cube of melting ice visible inside a heavy, clear glass partition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass partition resting on a polished, dark countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ice cube slowly tilt as its base erodes, causing its center of gravity to shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tilting ice cube suddenly whirl in the meltwater at the bottom of the glass before coming to stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a cool, minimalist palette with soft, directed light glinting off the ice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015150746330618858 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015150746330618858, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015150746330618858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15951752662658691, + 0.07055997848510742, + 0.05775856599211693, + 0.034402064979076385, + 0.057419151067733765, + 0.06842852383852005, + 0.05548004433512688, + 0.16463609039783478, + 0.1931947022676468, + 0.1386033594608307 + ], + "credit_map": [ + 0.026680639013648033, + 0.016553131863474846, + 0.024090316146612167, + 0.01237387116998434, + 0.015302678570151329, + 0.03008407913148403, + 0.014160191640257835, + 0.020272616297006607, + 0.01403264794498682, + 0.00641695037484169, + 0.010063765570521355, + 0.007297196425497532, + 0.005642253439873457, + 0.009712135419249535, + 0.008395826444029808, + 0.008999203331768513, + 0.008767198771238327, + 0.00568067841231823, + 0.011233799159526825, + 0.004583233501762152, + 0.005240870174020529, + 0.00942291971296072, + 0.004206386394798756, + 0.008623477071523666, + 0.005364877637475729, + 0.003402777248993516, + 0.00525338901206851, + 0.0040739779360592365, + 0.0032565451692789793, + 0.005352975334972143, + 0.0031903344206511974, + 0.004507187753915787, + 0.008346013724803925, + 0.008937451988458633, + 0.005909476429224014, + 0.007205433677881956, + 0.007332874462008476, + 0.007651586085557938, + 0.006901875138282776, + 0.005134436767548323, + 0.008320624940097332, + 0.00788833200931549, + 0.010531647130846977, + 0.006380068138241768, + 0.008742716163396835, + 0.007401677314192057, + 0.00817413255572319, + 0.010989324189722538, + 0.008845324628055096, + 0.005900432355701923, + 0.006351951509714127, + 0.004633815959095955, + 0.0075758942402899265, + 0.011158098466694355, + 0.0045815021730959415, + 0.0064330268651247025, + 0.024626314640045166, + 0.021684376522898674, + 0.013339051976799965, + 0.015353997237980366, + 0.023614294826984406, + 0.04186602309346199, + 0.014392678625881672, + 0.009759348817169666, + 0.02247019112110138, + 0.014979294501245022, + 0.024594835937023163, + 0.015208917669951916, + 0.039049651473760605, + 0.031953588128089905, + 0.017664296552538872, + 0.02727392315864563, + 0.02372060902416706, + 0.01105011347681284, + 0.013624570332467556, + 0.01025792770087719, + 0.04027225822210312, + 0.013081980869174004, + 0.01352233998477459, + 0.013073551468551159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0015150746330618858 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_9/index_tensor([2931, 2931])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_9/index_tensor([2931, 2931])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..41a779cd73cd70b224346c4c6fdee76b494d4198 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_9/index_tensor([2931, 2931])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly shaky close-up captures a whimsical scene unfolding from eye level against the sun-drenched, textured black plastic dashboard of a moving car. Two small, fantastical furry creatures, one with fluffy, warm brown fur and the other with creamy white fur speckled with soft gray, engage in a playful tussle. They tumble energetically and roll over each other, their large, black, expressive eyes wide with excitement as they harmlessly slap each other with their soft little paws and gently bite in a joyful mock battle. Bright sunlight streams directly through the front windshield, giving their fur a golden halo and casting sharp, dancing shadows beneath them. Through the clear glass, the vibrant green of the roadside trees and the blue of the sky blur into soft, continuous motion, reinforcing the feeling of a pleasant journey. The warm, saturated color palette and the creatures' joyful antics combine to create an atmosphere that is relaxing, tranquil, and utterly captivating.", + "seed": 19073, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_9/index_2931_seed_19073_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.007700573746114969, + "vlm_logprob": -0.007700573746114969, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.007700573746114969, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007684077136218548 + ], + "opposite_logprob_tokens": [ + -5.132684230804443 + ], + "logprob_reward": -0.007684077136218548, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007684077136218548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11159023642539978, + 0.0498172752559185, + 0.15421314537525177, + 0.050639886409044266, + 0.0710102990269661, + 0.04692453891038895, + 0.14456132054328918, + 0.1840750128030777, + 0.11154688894748688, + 0.07562142610549927 + ], + "credit_map": [ + 0.007937770336866379, + 0.008952830918133259, + 0.007771533913910389, + 0.011091607622802258, + 0.010229319334030151, + 0.030152088031172752, + 0.0087966313585639, + 0.02665845677256584, + 0.005200013052672148, + 0.005931926891207695, + 0.004838236607611179, + 0.005348989740014076, + 0.007100964430719614, + 0.008391211740672588, + 0.0063789002597332, + 0.006627032533288002, + 0.014323540031909943, + 0.016225267201662064, + 0.01269569993019104, + 0.015672368928790092, + 0.020706644281744957, + 0.020941661670804024, + 0.031248975545167923, + 0.02239898405969143, + 0.005175143480300903, + 0.010722747072577477, + 0.0037712801713496447, + 0.005461638793349266, + 0.008502480573952198, + 0.006145196035504341, + 0.006032989826053381, + 0.004828410688787699, + 0.0053427377715706825, + 0.008694600313901901, + 0.00645955977961421, + 0.01160473469644785, + 0.011308578774333, + 0.00831760186702013, + 0.009685085155069828, + 0.009597403928637505, + 0.003937277942895889, + 0.004690473433583975, + 0.002958157332614064, + 0.005394563544541597, + 0.008245955221354961, + 0.005870291963219643, + 0.005391970276832581, + 0.01043584942817688, + 0.014311441220343113, + 0.006715129129588604, + 0.005802917294204235, + 0.02393932268023491, + 0.007717132568359375, + 0.018820419907569885, + 0.010705051012337208, + 0.05654991418123245, + 0.013927417807281017, + 0.016461629420518875, + 0.018731648102402687, + 0.03356804698705673, + 0.01886070892214775, + 0.019991056993603706, + 0.024470986798405647, + 0.03806351125240326, + 0.009512078016996384, + 0.015671508386731148, + 0.010595863685011864, + 0.018464041873812675, + 0.020210137590765953, + 0.011450182646512985, + 0.013583162799477577, + 0.012059906497597694, + 0.004849429242312908, + 0.0090974485501647, + 0.003915910143405199, + 0.00856094527989626, + 0.018410176038742065, + 0.01713593676686287, + 0.0074097709730267525, + 0.006241815630346537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.007684077136218548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036281016655266285 + ], + "opposite_logprob_tokens": [ + -5.628628253936768 + ], + "logprob_reward": -0.0036281016655266285, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036281016655266285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12757791578769684, + 0.05103231221437454, + 0.08132671564817429, + 0.05838124826550484, + 0.04943610727787018, + 0.01998903974890709, + 0.13688258826732635, + 0.2905276119709015, + 0.08447817713022232, + 0.10036830604076385 + ], + "credit_map": [ + 0.008287490345537663, + 0.01891425810754299, + 0.016774721443653107, + 0.01920562982559204, + 0.0237425584346056, + 0.013779704459011555, + 0.012728258967399597, + 0.014145292341709137, + 0.004405430518090725, + 0.006753230933099985, + 0.004029573407024145, + 0.00940700527280569, + 0.007239624857902527, + 0.00601074518635869, + 0.0075492169708013535, + 0.005637484602630138, + 0.008470944128930569, + 0.010589373297989368, + 0.0051247538067400455, + 0.012597111985087395, + 0.013897581025958061, + 0.012308903969824314, + 0.010802535340189934, + 0.0075355092994868755, + 0.005506799090653658, + 0.009981351904571056, + 0.004159164149314165, + 0.007795688230544329, + 0.012537193484604359, + 0.004702355246990919, + 0.006555845960974693, + 0.007142852991819382, + 0.00294834328815341, + 0.00495028356090188, + 0.003433335805311799, + 0.006868240889161825, + 0.006139884702861309, + 0.010170475579798222, + 0.005750707350671291, + 0.009174837730824947, + 0.001662449911236763, + 0.0020475878845900297, + 0.001954628387466073, + 0.0022343930322676897, + 0.0038309304509311914, + 0.003184210043400526, + 0.0023338990285992622, + 0.0027409407775849104, + 0.014699474908411503, + 0.006458599120378494, + 0.006428161635994911, + 0.021021295338869095, + 0.00676122959703207, + 0.018653646111488342, + 0.008811277337372303, + 0.05404890701174736, + 0.023897627368569374, + 0.025229468941688538, + 0.03323378413915634, + 0.04934653639793396, + 0.029401419684290886, + 0.030358105897903442, + 0.04171996936202049, + 0.05734071508049965, + 0.008075546473264694, + 0.015799907967448235, + 0.005617223214358091, + 0.014454747550189495, + 0.012180759571492672, + 0.008723671548068523, + 0.010152316652238369, + 0.009474002756178379, + 0.006415671668946743, + 0.008318599313497543, + 0.002391544869169593, + 0.0030701695941388607, + 0.03665851429104805, + 0.031091829761862755, + 0.006116244476288557, + 0.006305734161287546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0036281016655266285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062267124652863 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.016062267124652863, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062267124652863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07778649777173996, + 0.0488886833190918, + 0.2125035524368286, + 0.14260609447956085, + 0.07750610262155533, + 0.05786995217204094, + 0.06885617971420288, + 0.07472886890172958, + 0.07490426301956177, + 0.16434982419013977 + ], + "credit_map": [ + 0.006503429263830185, + 0.011602446436882019, + 0.008585909381508827, + 0.00987321138381958, + 0.011571547016501427, + 0.008130736649036407, + 0.0084002036601305, + 0.013119013048708439, + 0.005001893732696772, + 0.005805832799524069, + 0.0035383938811719418, + 0.004795412067323923, + 0.007471567951142788, + 0.009562528692185879, + 0.00653827004134655, + 0.006174786482006311, + 0.03058280609548092, + 0.023779695853590965, + 0.014471433125436306, + 0.019680943340063095, + 0.027192072942852974, + 0.03584006428718567, + 0.025933783501386642, + 0.03502274304628372, + 0.014143490232527256, + 0.021288858726620674, + 0.015964234247803688, + 0.014506157487630844, + 0.015100466087460518, + 0.01431487686932087, + 0.020027486607432365, + 0.027260517701506615, + 0.005872005131095648, + 0.005781759507954121, + 0.006248957943171263, + 0.006542207207530737, + 0.006004530470818281, + 0.01995091512799263, + 0.007137174718081951, + 0.01996855065226555, + 0.006210014224052429, + 0.005248407367616892, + 0.005316907539963722, + 0.007334427442401648, + 0.009295902214944363, + 0.008893203921616077, + 0.007862620055675507, + 0.007708468474447727, + 0.006740038748830557, + 0.004187760874629021, + 0.004830891732126474, + 0.008973659947514534, + 0.004890979267656803, + 0.010401681996881962, + 0.006170216482132673, + 0.02266094461083412, + 0.007365858647972345, + 0.006459514610469341, + 0.009917772375047207, + 0.008461925201117992, + 0.007250088267028332, + 0.008815660141408443, + 0.010213283821940422, + 0.01624477095901966, + 0.008202040567994118, + 0.011055806651711464, + 0.009903918020427227, + 0.010183531790971756, + 0.00929956417530775, + 0.007684305775910616, + 0.009242779575288296, + 0.009332315996289253, + 0.010126669891178608, + 0.013317402452230453, + 0.004883765708655119, + 0.006582757458090782, + 0.05927537754178047, + 0.04796230047941208, + 0.010945546440780163, + 0.011256012134253979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.016062267124652863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640403859317303 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640403859317303, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640403859317303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17401041090488434, + 0.1879943460226059, + 0.05409039556980133, + 0.08870145678520203, + 0.1177339032292366, + 0.05538410693407059, + 0.1196666732430458, + 0.07267512381076813, + 0.09353780001401901, + 0.03620577231049538 + ], + "credit_map": [ + 0.011752087622880936, + 0.012200347147881985, + 0.01386391930282116, + 0.013474544510245323, + 0.012918859720230103, + 0.05317730829119682, + 0.013247781433165073, + 0.04337558150291443, + 0.018110010772943497, + 0.02350166253745556, + 0.020217129960656166, + 0.025742366909980774, + 0.02427706867456436, + 0.03531713783740997, + 0.022593101486563683, + 0.01823587343096733, + 0.004820979665964842, + 0.007380816154181957, + 0.00421623419970274, + 0.006767522543668747, + 0.008829937316477299, + 0.006117368116974831, + 0.008406031876802444, + 0.007551508489996195, + 0.007113242521882057, + 0.015911834314465523, + 0.007515120320022106, + 0.010079688392579556, + 0.019144300371408463, + 0.009646833874285221, + 0.009253866970539093, + 0.010036574676632881, + 0.007959594950079918, + 0.011003225110471249, + 0.007104367949068546, + 0.012988289818167686, + 0.013367265462875366, + 0.02987026609480381, + 0.011467469856142998, + 0.02397342398762703, + 0.004033968783915043, + 0.006139114964753389, + 0.004581087734550238, + 0.007947021164000034, + 0.007419191300868988, + 0.006588562857359648, + 0.011314603500068188, + 0.007360562216490507, + 0.011497126892209053, + 0.005868521053344011, + 0.007721289061009884, + 0.01758105680346489, + 0.006880361586809158, + 0.015753863379359245, + 0.009417898952960968, + 0.04494656249880791, + 0.0061121718026697636, + 0.008649943396449089, + 0.008853192441165447, + 0.008877086453139782, + 0.009199623018503189, + 0.007645502686500549, + 0.010709288530051708, + 0.012628327123820782, + 0.011677533388137817, + 0.013995502144098282, + 0.012460393831133842, + 0.01357751153409481, + 0.012957610189914703, + 0.01153438352048397, + 0.009034357964992523, + 0.00830051675438881, + 0.0030493568629026413, + 0.003850000211969018, + 0.002699315082281828, + 0.003644336946308613, + 0.0077366335317492485, + 0.006524764001369476, + 0.004987451713532209, + 0.0037139146588742733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.008640403859317303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024880189448595047 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.0024880189448595047, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024880189448595047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13611169159412384, + 0.09575509279966354, + 0.1805853694677353, + 0.062328264117240906, + 0.13098512589931488, + 0.03597487509250641, + 0.05516459047794342, + 0.1318032443523407, + 0.07446520775556564, + 0.09682657569646835 + ], + "credit_map": [ + 0.00982667226344347, + 0.018233900889754295, + 0.015452343039214611, + 0.021590204909443855, + 0.01635422743856907, + 0.01557116024196148, + 0.014731823466718197, + 0.02435135841369629, + 0.008877512067556381, + 0.013370378874242306, + 0.011565478518605232, + 0.013095098547637463, + 0.011796899139881134, + 0.016835875809192657, + 0.009769477881491184, + 0.010444375686347485, + 0.017678283154964447, + 0.018487734720110893, + 0.011659238487482071, + 0.020235974341630936, + 0.025152994319796562, + 0.027688270434737206, + 0.03194418549537659, + 0.027738673612475395, + 0.005699428729712963, + 0.010806843638420105, + 0.006155907642096281, + 0.00708869518712163, + 0.012668536975979805, + 0.00591621920466423, + 0.007229682989418507, + 0.0067629506811499596, + 0.009683902375400066, + 0.00884748250246048, + 0.008086656220257282, + 0.009708359837532043, + 0.007257985416799784, + 0.03814754635095596, + 0.010862233117222786, + 0.038390956819057465, + 0.0032344183418899775, + 0.003943121526390314, + 0.0041444068774580956, + 0.004904051311314106, + 0.00492760818451643, + 0.005820292979478836, + 0.004643901251256466, + 0.004357075784355402, + 0.004963700193911791, + 0.0032684297766536474, + 0.0038912652526050806, + 0.008329077623784542, + 0.0038232177030295134, + 0.007857800461351871, + 0.005701033398509026, + 0.01733006350696087, + 0.01070067472755909, + 0.008473319001495838, + 0.012380131520330906, + 0.021375132724642754, + 0.01254242192953825, + 0.016271675005555153, + 0.020190386101603508, + 0.029869507998228073, + 0.007698912173509598, + 0.014193044044077396, + 0.01027164701372385, + 0.006381911225616932, + 0.010507519356906414, + 0.0083992974832654, + 0.007053514011204243, + 0.00995936430990696, + 0.006148156709969044, + 0.009113501757383347, + 0.00440420676022768, + 0.00697346031665802, + 0.030250947922468185, + 0.023407336324453354, + 0.008864556439220905, + 0.007664409000426531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.0024880189448595047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly shaky close-up captures a whimsical scene unfolding from eye level against the sun-drenched, textured black plastic dashboard of a moving car. Two small, fantastical furry creatures, one with fluffy, warm brown fur and the other with creamy white fur speckled with soft gray, engage in a playful tussle. They tumble energetically and roll over each other, their large, black, expressive eyes wide with excitement as they harmlessly slap each other with their soft little paws and gently bite in a joyful mock battle. Bright sunlight streams directly through the front windshield, giving their fur a golden halo and casting sharp, dancing shadows beneath them. Through the clear glass, the vibrant green of the roadside trees and the blue of the sky blur into soft, continuous motion, reinforcing the feeling of a pleasant journey. The warm, saturated color palette and the creatures' joyful antics combine to create an atmosphere that is relaxing, tranquil, and utterly captivating.", + "seed": 19074, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_9/index_2931_seed_19074_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.00738478759303689, + "vlm_logprob": -0.00738478759303689, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.00738478759303689, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008988158777356148 + ], + "opposite_logprob_tokens": [ + -4.883988380432129 + ], + "logprob_reward": -0.008988158777356148, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008988158777356148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04290930554270744, + 0.0387130044400692, + 0.06754954159259796, + 0.03984396532177925, + 0.13342203199863434, + 0.21402084827423096, + 0.1793937385082245, + 0.07787708193063736, + 0.1908314824104309, + 0.015438972041010857 + ], + "credit_map": [ + 0.004776845686137676, + 0.005959605798125267, + 0.005158421117812395, + 0.006850606296211481, + 0.0059502688236534595, + 0.005121767986565828, + 0.005262827966362238, + 0.0038289676886051893, + 0.004236420150846243, + 0.006406860891729593, + 0.004394614603370428, + 0.007308777887374163, + 0.0044385408982634544, + 0.005004073027521372, + 0.004337268881499767, + 0.002586451591923833, + 0.015293634496629238, + 0.004734401125460863, + 0.020114537328481674, + 0.007704411167651415, + 0.005579366814345121, + 0.005800121929496527, + 0.005144636612385511, + 0.0031784423626959324, + 0.006258151959627867, + 0.006456159055233002, + 0.004668815061450005, + 0.008146503008902073, + 0.0036612646654248238, + 0.0031899292953312397, + 0.004662090912461281, + 0.0028010543901473284, + 0.01175004243850708, + 0.017748964950442314, + 0.013884871266782284, + 0.017456524074077606, + 0.02535211481153965, + 0.013469552621245384, + 0.022058697417378426, + 0.011701278388500214, + 0.01327765267342329, + 0.008556319400668144, + 0.07967020571231842, + 0.0196699108928442, + 0.011140097863972187, + 0.06349267810583115, + 0.009943983517587185, + 0.00826999917626381, + 0.056930385529994965, + 0.01034134067595005, + 0.0724247619509697, + 0.0135830482468009, + 0.006630236748605967, + 0.008857501670718193, + 0.005412241909652948, + 0.005214222706854343, + 0.008829536847770214, + 0.006669689901173115, + 0.009974398650228977, + 0.008698013611137867, + 0.01242122519761324, + 0.011773524805903435, + 0.009430118836462498, + 0.010080575942993164, + 0.06772542744874954, + 0.010656828992068768, + 0.07421602308750153, + 0.011540002189576626, + 0.0061404709704220295, + 0.009891664609313011, + 0.00596600491553545, + 0.004695083945989609, + 0.002455356763675809, + 0.001395634957589209, + 0.0021618844475597143, + 0.0022321792785078287, + 0.0012744141276925802, + 0.003297146875411272, + 0.0015082857571542263, + 0.0011140709975734353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.008988158777356148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362691399641335 + ], + "opposite_logprob_tokens": [ + -5.628626823425293 + ], + "logprob_reward": -0.00362691399641335, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362691399641335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11745913326740265, + 0.0820997804403305, + 0.05055315047502518, + 0.08052992820739746, + 0.11123135685920715, + 0.2121882289648056, + 0.09513849020004272, + 0.05156019702553749, + 0.13427695631980896, + 0.06496278941631317 + ], + "credit_map": [ + 0.01573009416460991, + 0.014517980627715588, + 0.009944820776581764, + 0.016814759001135826, + 0.01229802519083023, + 0.02036072500050068, + 0.01920539326965809, + 0.008587338030338287, + 0.008397909812629223, + 0.014880035072565079, + 0.009651181288063526, + 0.014394799247384071, + 0.012513212859630585, + 0.008162801153957844, + 0.00957501120865345, + 0.004524832125753164, + 0.008278323337435722, + 0.005147865973412991, + 0.006054998841136694, + 0.007264380343258381, + 0.005708644166588783, + 0.006857187952846289, + 0.005479720886796713, + 0.005762028973549604, + 0.01071688812226057, + 0.011355436407029629, + 0.011906217783689499, + 0.011855070479214191, + 0.008145946078002453, + 0.007907475344836712, + 0.012694177217781544, + 0.005948707461357117, + 0.010666691698133945, + 0.015387829393148422, + 0.011892431415617466, + 0.01302378997206688, + 0.021352576091885567, + 0.012203877791762352, + 0.017572734504938126, + 0.009131429716944695, + 0.013874197378754616, + 0.006901697255671024, + 0.08006488531827927, + 0.018710292875766754, + 0.010748382657766342, + 0.06567271053791046, + 0.0075566451996564865, + 0.008659401908516884, + 0.028317267075181007, + 0.00609979685395956, + 0.03809883072972298, + 0.007940858602523804, + 0.004074451047927141, + 0.004552191123366356, + 0.003327669808641076, + 0.0027274286840111017, + 0.007972772233188152, + 0.0037890837993472815, + 0.01032248791307211, + 0.0066006602719426155, + 0.004958237521350384, + 0.004445069935172796, + 0.00775640644133091, + 0.00571547681465745, + 0.04708864167332649, + 0.007235439959913492, + 0.049130503088235855, + 0.00865265354514122, + 0.004684083629399538, + 0.008517034351825714, + 0.0035819527693092823, + 0.005386634264141321, + 0.010521475225687027, + 0.007434013299643993, + 0.008812922984361649, + 0.009793752804398537, + 0.005404873285442591, + 0.012054519727826118, + 0.006340244319289923, + 0.004600986372679472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.00362691399641335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191186055541039 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191186055541039, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191186055541039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07900170981884003, + 0.09152503311634064, + 0.03134226053953171, + 0.1268336921930313, + 0.06685113906860352, + 0.0795680582523346, + 0.12547659873962402, + 0.03173031657934189, + 0.309516578912735, + 0.05815455690026283 + ], + "credit_map": [ + 0.010569119825959206, + 0.007928206585347652, + 0.012447159737348557, + 0.011769975535571575, + 0.009579675272107124, + 0.00899916235357523, + 0.010735667310655117, + 0.006972744129598141, + 0.009908747859299183, + 0.00930827111005783, + 0.010655416175723076, + 0.0130962198600173, + 0.01159603614360094, + 0.012300955131649971, + 0.01631573960185051, + 0.008343648165464401, + 0.003838869510218501, + 0.003245092462748289, + 0.0038711789529770613, + 0.004951755981892347, + 0.0033660284243524075, + 0.0039122565649449825, + 0.0051999143324792385, + 0.002957163145765662, + 0.025651229545474052, + 0.009183324873447418, + 0.008137895725667477, + 0.012313286773860455, + 0.011143957264721394, + 0.009624486789107323, + 0.041890520602464676, + 0.008888992480933666, + 0.00753286387771368, + 0.005554461386054754, + 0.006481361109763384, + 0.00684354966506362, + 0.011382842436432838, + 0.009991869330406189, + 0.012067841365933418, + 0.006996350362896919, + 0.0057878270745277405, + 0.0037820434663444757, + 0.026942484080791473, + 0.008002733811736107, + 0.006274320650845766, + 0.018841180950403214, + 0.006550832651555538, + 0.0033866397570818663, + 0.03671746328473091, + 0.007624883204698563, + 0.054021965712308884, + 0.009800394997000694, + 0.004996049217879772, + 0.00498719559982419, + 0.00394492456689477, + 0.0033837188966572285, + 0.0041730040684342384, + 0.0037923569325357676, + 0.004018839448690414, + 0.00458437530323863, + 0.004186553880572319, + 0.004088211338967085, + 0.003961027134209871, + 0.0029259524308145046, + 0.11378393322229385, + 0.014806116931140423, + 0.12544815242290497, + 0.016259057447314262, + 0.009497265331447124, + 0.014926303178071976, + 0.007561790756881237, + 0.0072339666076004505, + 0.008181420154869556, + 0.00830051302909851, + 0.007607783190906048, + 0.007094740867614746, + 0.004655188415199518, + 0.01320718601346016, + 0.005573614034801722, + 0.0035341139882802963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.014191186055541039 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076294224709272385 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.0076294224709272385, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076294224709272385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11727901548147202, + 0.07881402969360352, + 0.05790073424577713, + 0.3750689923763275, + 0.075351782143116, + 0.08390418440103531, + 0.0359504371881485, + 0.0749509334564209, + 0.05146816000342369, + 0.04931169003248215 + ], + "credit_map": [ + 0.014555923640727997, + 0.010491213761270046, + 0.015025713481009007, + 0.01905304379761219, + 0.012835616245865822, + 0.014601369388401508, + 0.020864436402916908, + 0.009851709008216858, + 0.01175256259739399, + 0.008166911080479622, + 0.015382024459540844, + 0.009297004900872707, + 0.007795686833560467, + 0.01258287113159895, + 0.008738795295357704, + 0.005098182708024979, + 0.008475136011838913, + 0.006631078198552132, + 0.006864757277071476, + 0.013293723575770855, + 0.005912987049669027, + 0.007437385618686676, + 0.005255198571830988, + 0.004030477721244097, + 0.0427716039121151, + 0.021500345319509506, + 0.045930515974760056, + 0.040009673684835434, + 0.037240926176309586, + 0.016712673008441925, + 0.14596691727638245, + 0.024936361238360405, + 0.012907055206596851, + 0.00869099423289299, + 0.008902588859200478, + 0.012830878607928753, + 0.007491524796932936, + 0.010149775072932243, + 0.009123108349740505, + 0.005255866330116987, + 0.00757632264867425, + 0.009329975582659245, + 0.010213851928710938, + 0.008808426558971405, + 0.011660669930279255, + 0.02032386139035225, + 0.00852320808917284, + 0.007467880379408598, + 0.004266389645636082, + 0.005860103294253349, + 0.0036330576986074448, + 0.005010217893868685, + 0.005531455855816603, + 0.004012084566056728, + 0.0038902878295630217, + 0.003746845992282033, + 0.009658142924308777, + 0.005898549687117338, + 0.008830429054796696, + 0.004495236091315746, + 0.01036172267049551, + 0.01992153748869896, + 0.005425815470516682, + 0.010359502397477627, + 0.007526068948209286, + 0.00535476254299283, + 0.006611592136323452, + 0.009559697471559048, + 0.005140242166817188, + 0.006154133006930351, + 0.005088561214506626, + 0.006033106707036495, + 0.010061768814921379, + 0.005344545003026724, + 0.006653379648923874, + 0.007193723227828741, + 0.0032190470956265926, + 0.008765908889472485, + 0.0038604617584496737, + 0.004212861880660057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0076294224709272385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024882566649466753 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.0024882566649466753, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024882566649466753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0718846470117569, + 0.041443243622779846, + 0.045582402497529984, + 0.1503952294588089, + 0.0913943350315094, + 0.06807250529527664, + 0.1386183649301529, + 0.07922161370515823, + 0.2495601922273636, + 0.06382754445075989 + ], + "credit_map": [ + 0.010227832943201065, + 0.011106389574706554, + 0.009246468544006348, + 0.011848111636936665, + 0.008764596655964851, + 0.007299585267901421, + 0.007953010499477386, + 0.00543864956125617, + 0.005747455172240734, + 0.0040834201499819756, + 0.006071014329791069, + 0.005651055369526148, + 0.005071153398603201, + 0.005545405671000481, + 0.005506088491529226, + 0.00376765220426023, + 0.010615022853016853, + 0.0033308390993624926, + 0.012474123388528824, + 0.004514205735176802, + 0.002630081260576844, + 0.005378133617341518, + 0.0031511581037193537, + 0.003488837042823434, + 0.014037870801985264, + 0.011024516075849533, + 0.018138457089662552, + 0.02044750191271305, + 0.016005804762244225, + 0.009259381331503391, + 0.05139905586838722, + 0.010082644410431385, + 0.012127249501645565, + 0.0076477923430502415, + 0.009144251234829426, + 0.012076783925294876, + 0.013660403899848461, + 0.01309924479573965, + 0.014753065072000027, + 0.008885539136826992, + 0.007155290339142084, + 0.010212831199169159, + 0.005462354049086571, + 0.009757897816598415, + 0.009884491562843323, + 0.010785133577883244, + 0.009356777183711529, + 0.005457730498164892, + 0.04072245955467224, + 0.009363343007862568, + 0.057559575885534286, + 0.011233632452785969, + 0.0055611152201890945, + 0.005840870086103678, + 0.004301808774471283, + 0.004035572987049818, + 0.009747247211635113, + 0.004880181513726711, + 0.009507299400866032, + 0.005860485602170229, + 0.011438988149166107, + 0.019364265725016594, + 0.007407376542687416, + 0.01101576816290617, + 0.08260735869407654, + 0.014939147979021072, + 0.08974284678697586, + 0.017677167430520058, + 0.009718981571495533, + 0.017473692074418068, + 0.007481217849999666, + 0.009919770993292332, + 0.0067695872858166695, + 0.010773890651762486, + 0.007973008789122105, + 0.015112567692995071, + 0.0038241632282733917, + 0.007399108726531267, + 0.006180889438837767, + 0.005794328637421131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.0024882566649466753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly shaky close-up captures a whimsical scene unfolding from eye level against the sun-drenched, textured black plastic dashboard of a moving car. Two small, fantastical furry creatures, one with fluffy, warm brown fur and the other with creamy white fur speckled with soft gray, engage in a playful tussle. They tumble energetically and roll over each other, their large, black, expressive eyes wide with excitement as they harmlessly slap each other with their soft little paws and gently bite in a joyful mock battle. Bright sunlight streams directly through the front windshield, giving their fur a golden halo and casting sharp, dancing shadows beneath them. Through the clear glass, the vibrant green of the roadside trees and the blue of the sky blur into soft, continuous motion, reinforcing the feeling of a pleasant journey. The warm, saturated color palette and the creatures' joyful antics combine to create an atmosphere that is relaxing, tranquil, and utterly captivating.", + "seed": 19075, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_9/index_2931_seed_19075_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.007210622075945139, + "vlm_logprob": -0.007210622075945139, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.007210622075945139, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005447306204587221 + ], + "opposite_logprob_tokens": [ + -5.5054473876953125 + ], + "logprob_reward": -0.005447306204587221, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005447306204587221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11259277909994125, + 0.107661172747612, + 0.11148760467767715, + 0.05244791507720947, + 0.21886250376701355, + 0.08763668686151505, + 0.06260021030902863, + 0.054457955062389374, + 0.12007437646389008, + 0.07217879593372345 + ], + "credit_map": [ + 0.014627702534198761, + 0.02231406420469284, + 0.007198774255812168, + 0.013622595928609371, + 0.013060332275927067, + 0.017429053783416748, + 0.014128469862043858, + 0.010211789980530739, + 0.011385845020413399, + 0.014229386113584042, + 0.008741245605051517, + 0.015400521457195282, + 0.014447489753365517, + 0.014164629392325878, + 0.016910018399357796, + 0.01238203514367342, + 0.005450327415019274, + 0.008357790298759937, + 0.004036858212202787, + 0.007583336904644966, + 0.007921197451651096, + 0.033494733273983, + 0.010162059217691422, + 0.03448128327727318, + 0.00433984212577343, + 0.009669712744653225, + 0.005760025233030319, + 0.012096654623746872, + 0.005037303548306227, + 0.004669147543609142, + 0.005145442206412554, + 0.005729785654693842, + 0.013870076276361942, + 0.017292650416493416, + 0.011633316054940224, + 0.031485602259635925, + 0.028188440948724747, + 0.05470693111419678, + 0.025050997734069824, + 0.03663448989391327, + 0.008906873874366283, + 0.011825066059827805, + 0.0068936655297875404, + 0.014928469434380531, + 0.008262129500508308, + 0.00932166539132595, + 0.010829020291566849, + 0.016669804230332375, + 0.005537560675293207, + 0.007303876336663961, + 0.012330794706940651, + 0.011120597831904888, + 0.0062481616623699665, + 0.006573592312633991, + 0.008685664273798466, + 0.004799965303391218, + 0.008581146597862244, + 0.00592973455786705, + 0.003536379896104336, + 0.007743522524833679, + 0.0065947165712714195, + 0.008542640134692192, + 0.0071577695198357105, + 0.006372045259922743, + 0.011752716265618801, + 0.011850918643176556, + 0.008659501560032368, + 0.010630824603140354, + 0.033830586820840836, + 0.012060186825692654, + 0.018437543883919716, + 0.012852100655436516, + 0.008495845831930637, + 0.010422319173812866, + 0.003420029766857624, + 0.01352796982973814, + 0.008530895225703716, + 0.010356580838561058, + 0.0093882717192173, + 0.008036881685256958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.005447306204587221 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028313808143138885 + ], + "opposite_logprob_tokens": [ + -5.87783145904541 + ], + "logprob_reward": -0.0028313808143138885, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028313808143138885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06962795555591583, + 0.17021878063678741, + 0.21424564719200134, + 0.053851254284381866, + 0.05688461661338806, + 0.06464902311563492, + 0.05969593673944473, + 0.06800036877393723, + 0.14203804731369019, + 0.10078837722539902 + ], + "credit_map": [ + 0.007949702441692352, + 0.016164004802703857, + 0.0061848582699894905, + 0.01081897784024477, + 0.0053842272609472275, + 0.007373225409537554, + 0.007383324671536684, + 0.008369628340005875, + 0.017595594748854637, + 0.011412038467824459, + 0.009447352029383183, + 0.0176725871860981, + 0.04104894772171974, + 0.02522008679807186, + 0.029115257784724236, + 0.018706921488046646, + 0.008938893675804138, + 0.007867532782256603, + 0.006229951046407223, + 0.01760380156338215, + 0.015138368122279644, + 0.0621880479156971, + 0.029410405084490776, + 0.0668686255812645, + 0.005140210036188364, + 0.007769427727907896, + 0.004756466951221228, + 0.011847136542201042, + 0.005805430468171835, + 0.004251901060342789, + 0.008597561158239841, + 0.005683119408786297, + 0.004463120363652706, + 0.006953117903321981, + 0.0034068566747009754, + 0.010447083041071892, + 0.008218890056014061, + 0.005922297947108746, + 0.010993382893502712, + 0.006479867734014988, + 0.0035408406984061003, + 0.007462528068572283, + 0.0036833069752901793, + 0.008012698963284492, + 0.01847112365067005, + 0.006924241781234741, + 0.009298088029026985, + 0.007256199140101671, + 0.005261584185063839, + 0.0064537706784904, + 0.00884359423071146, + 0.009631345979869366, + 0.008885111659765244, + 0.006025663111358881, + 0.009671611711382866, + 0.004923255182802677, + 0.009274489246308804, + 0.006142578553408384, + 0.004045629408210516, + 0.008178338408470154, + 0.010129062458872795, + 0.00873151421546936, + 0.011876238510012627, + 0.009622522629797459, + 0.014566294848918915, + 0.016308462247252464, + 0.010081926360726357, + 0.013227278366684914, + 0.03654451668262482, + 0.014124871231615543, + 0.021792195737361908, + 0.015392511151731014, + 0.01120736077427864, + 0.011329141445457935, + 0.006564444396644831, + 0.011096831411123276, + 0.012992210686206818, + 0.021684424951672554, + 0.012026128359138966, + 0.013887831941246986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.0028313808143138885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577985793352127 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577985793352127, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577985793352127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12295866757631302, + 0.046983424574136734, + 0.25004884600639343, + 0.03084581531584263, + 0.19244384765625, + 0.04560346528887749, + 0.05745955556631088, + 0.03801925480365753, + 0.04188057407736778, + 0.17375656962394714 + ], + "credit_map": [ + 0.015215405263006687, + 0.016915366053581238, + 0.009100336581468582, + 0.016384540125727654, + 0.016153382137417793, + 0.02097969688475132, + 0.01533181220293045, + 0.01287812739610672, + 0.0052904183976352215, + 0.005237040109932423, + 0.003951340913772583, + 0.0066810548305511475, + 0.005132575985044241, + 0.006150932516902685, + 0.005969856400042772, + 0.008570200763642788, + 0.010338310152292252, + 0.008004953153431416, + 0.007062210235744715, + 0.01929071731865406, + 0.021056296303868294, + 0.07428988069295883, + 0.02926868386566639, + 0.08073782175779343, + 0.0022929406259208918, + 0.004482444375753403, + 0.0021511027589440346, + 0.006356487050652504, + 0.00455322302877903, + 0.003133833874017, + 0.004870881326496601, + 0.003004902508109808, + 0.013389717787504196, + 0.012465107254683971, + 0.009317927993834019, + 0.026516210287809372, + 0.027826793491840363, + 0.040163855999708176, + 0.030134523287415504, + 0.03262970969080925, + 0.0036868304014205933, + 0.004581702873110771, + 0.003133536549285054, + 0.008414949290454388, + 0.00625024875625968, + 0.004046717192977667, + 0.005566924810409546, + 0.009922553785145283, + 0.0044885072857141495, + 0.006195942405611277, + 0.006935485173016787, + 0.007067000027745962, + 0.012906004674732685, + 0.005141071043908596, + 0.008635904639959335, + 0.0060896375216543674, + 0.004277947824448347, + 0.003697376698255539, + 0.00284044467844069, + 0.005710039287805557, + 0.0037000568117946386, + 0.004639591556042433, + 0.00554837379604578, + 0.007605427876114845, + 0.004398057237267494, + 0.004696353804320097, + 0.0029239594005048275, + 0.005414059851318598, + 0.005772174336016178, + 0.00400451710447669, + 0.008172444067895412, + 0.00649900920689106, + 0.01918540894985199, + 0.015561297535896301, + 0.009930920787155628, + 0.017369410023093224, + 0.023613085970282555, + 0.04551348462700844, + 0.021404186263680458, + 0.021178776398301125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.020577985793352127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257115233689547 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005257115233689547, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257115233689547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11277897655963898, + 0.13296066224575043, + 0.13289155066013336, + 0.0968928262591362, + 0.0607168935239315, + 0.14658541977405548, + 0.05342703312635422, + 0.08298448473215103, + 0.059934984892606735, + 0.12082713842391968 + ], + "credit_map": [ + 0.01259396318346262, + 0.018562812358140945, + 0.007881652563810349, + 0.01180942915380001, + 0.010729334317147732, + 0.01722106523811817, + 0.021286357194185257, + 0.012694359757006168, + 0.012401945888996124, + 0.010080945678055286, + 0.012517232447862625, + 0.019691718742251396, + 0.01648607663810253, + 0.016399281099438667, + 0.022491799667477608, + 0.022891657426953316, + 0.005344073288142681, + 0.0074954344891011715, + 0.003402347443625331, + 0.009844491258263588, + 0.008132747374475002, + 0.039046723395586014, + 0.012348359450697899, + 0.04727737605571747, + 0.009927026927471161, + 0.0150540079921484, + 0.009556856006383896, + 0.024342138320207596, + 0.00732164504006505, + 0.007649345323443413, + 0.010626805946230888, + 0.012415008619427681, + 0.004304689355194569, + 0.007211503107100725, + 0.004325905814766884, + 0.010731738992035389, + 0.010171708650887012, + 0.009715547785162926, + 0.008107260800898075, + 0.006148533429950476, + 0.011604475788772106, + 0.02021646685898304, + 0.012481726706027985, + 0.025155702605843544, + 0.017301548272371292, + 0.017407983541488647, + 0.019114777445793152, + 0.023302745074033737, + 0.004580625798553228, + 0.006415588781237602, + 0.009544839151203632, + 0.008043083362281322, + 0.0077610923908650875, + 0.006123026367276907, + 0.006014065816998482, + 0.0049447109922766685, + 0.00770202511921525, + 0.008963397704064846, + 0.005835825577378273, + 0.008989079855382442, + 0.010171919129788876, + 0.0112548703327775, + 0.014508125372231007, + 0.01555924117565155, + 0.006521322764456272, + 0.005960376467555761, + 0.0038106513675302267, + 0.00618892814964056, + 0.01500792521983385, + 0.007466863840818405, + 0.007729832548648119, + 0.00724908709526062, + 0.014171677641570568, + 0.011952074244618416, + 0.007150612771511078, + 0.012467436492443085, + 0.016288641840219498, + 0.028562260791659355, + 0.013986865989863873, + 0.016247564926743507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.005257115233689547 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019393223337829113 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019393223337829113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019393223337829113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1401468962430954, + 0.07754895091056824, + 0.3141147196292877, + 0.04527156427502632, + 0.07599091529846191, + 0.03641720861196518, + 0.08550499379634857, + 0.038604818284511566, + 0.06988293677568436, + 0.11651699244976044 + ], + "credit_map": [ + 0.015982341021299362, + 0.025270242244005203, + 0.013140015304088593, + 0.01768164336681366, + 0.014428200200200081, + 0.020994018763303757, + 0.020768651738762856, + 0.01188179012387991, + 0.006927426438778639, + 0.008129134774208069, + 0.007931004278361797, + 0.009415560401976109, + 0.012196016497910023, + 0.01239379495382309, + 0.011219194158911705, + 0.009336818940937519, + 0.011805039830505848, + 0.009439480490982533, + 0.007308689411729574, + 0.021767614409327507, + 0.02196545898914337, + 0.09906688332557678, + 0.03412681072950363, + 0.10863473266363144, + 0.0044174944050610065, + 0.005622695200145245, + 0.004914463963359594, + 0.009069203399121761, + 0.006201290991157293, + 0.0035987612791359425, + 0.007150890771299601, + 0.004296762868762016, + 0.006184931844472885, + 0.007043927442282438, + 0.005751166958361864, + 0.01377773005515337, + 0.009097862988710403, + 0.01268708985298872, + 0.01080863643437624, + 0.01063957717269659, + 0.00325951399281621, + 0.004128327593207359, + 0.003436789847910404, + 0.009676088578999043, + 0.003916760440915823, + 0.003453930839896202, + 0.0038489496801048517, + 0.004696847405284643, + 0.00661000981926918, + 0.007978060282766819, + 0.007358449976891279, + 0.008447710424661636, + 0.02312515117228031, + 0.008124182000756264, + 0.013924448750913143, + 0.009936988353729248, + 0.005687428172677755, + 0.004021582659333944, + 0.0028879758901894093, + 0.0052852327935397625, + 0.00456741638481617, + 0.005112032406032085, + 0.005199921317398548, + 0.005843227729201317, + 0.00631071487441659, + 0.006849842146039009, + 0.004219610709697008, + 0.008303738199174404, + 0.017296915873885155, + 0.007038315292447805, + 0.01099067460745573, + 0.008873122744262218, + 0.012580014765262604, + 0.011598050594329834, + 0.007511927280575037, + 0.013723599724471569, + 0.013508441857993603, + 0.0228755883872509, + 0.01643947698175907, + 0.018279893323779106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.0019393223337829113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly shaky close-up captures a whimsical scene unfolding from eye level against the sun-drenched, textured black plastic dashboard of a moving car. Two small, fantastical furry creatures, one with fluffy, warm brown fur and the other with creamy white fur speckled with soft gray, engage in a playful tussle. They tumble energetically and roll over each other, their large, black, expressive eyes wide with excitement as they harmlessly slap each other with their soft little paws and gently bite in a joyful mock battle. Bright sunlight streams directly through the front windshield, giving their fur a golden halo and casting sharp, dancing shadows beneath them. Through the clear glass, the vibrant green of the roadside trees and the blue of the sky blur into soft, continuous motion, reinforcing the feeling of a pleasant journey. The warm, saturated color palette and the creatures' joyful antics combine to create an atmosphere that is relaxing, tranquil, and utterly captivating.", + "seed": 19076, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000083/group_9/index_2931_seed_19076_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.008001167792826891, + "vlm_logprob": -0.008001167792826891, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.008001167792826891, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00677171116694808 + ], + "opposite_logprob_tokens": [ + -5.131771564483643 + ], + "logprob_reward": -0.00677171116694808, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00677171116694808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09308299422264099, + 0.11369320750236511, + 0.12618423998355865, + 0.08836257457733154, + 0.04763887822628021, + 0.16335175931453705, + 0.08742762356996536, + 0.0846257284283638, + 0.14776693284511566, + 0.047866061329841614 + ], + "credit_map": [ + 0.0063024056144058704, + 0.02543366700410843, + 0.006231071427464485, + 0.00832015834748745, + 0.012824509292840958, + 0.01150599680840969, + 0.011074444279074669, + 0.01139074843376875, + 0.005971666891127825, + 0.01499311812222004, + 0.011372625827789307, + 0.021241290494799614, + 0.011336157098412514, + 0.019302090629935265, + 0.01290424820035696, + 0.016572019085288048, + 0.007047154475003481, + 0.020897822454571724, + 0.0072778817266225815, + 0.01630563475191593, + 0.012941348366439342, + 0.02822422981262207, + 0.01218670979142189, + 0.021303460001945496, + 0.005626311991363764, + 0.010872144252061844, + 0.005806624889373779, + 0.01672707311809063, + 0.014645582064986229, + 0.008449549786746502, + 0.011654390953481197, + 0.014580895192921162, + 0.0021343526896089315, + 0.006194444838911295, + 0.0030961534939706326, + 0.007289949804544449, + 0.008500986732542515, + 0.00648734113201499, + 0.007370638195425272, + 0.00656501529738307, + 0.006606392562389374, + 0.015413003042340279, + 0.03228253871202469, + 0.013240518979728222, + 0.029845865443348885, + 0.019539566710591316, + 0.032057907432317734, + 0.014365964569151402, + 0.005384540650993586, + 0.009999050758779049, + 0.010457843542098999, + 0.01829935424029827, + 0.011919645592570305, + 0.008684969507157803, + 0.013437091372907162, + 0.009245127439498901, + 0.004740428179502487, + 0.019673390313982964, + 0.004154215566813946, + 0.012068106792867184, + 0.01458109263330698, + 0.011183490045368671, + 0.01067210454493761, + 0.00755290174856782, + 0.008523163385689259, + 0.017625823616981506, + 0.007823970168828964, + 0.02381085231900215, + 0.027523508295416832, + 0.018801946192979813, + 0.029516136273741722, + 0.01414152979850769, + 0.0029728044755756855, + 0.006738979369401932, + 0.0038235026877373457, + 0.0063725421205163, + 0.007647725287824869, + 0.0075223310850560665, + 0.009186722338199615, + 0.003601453499868512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.00677171116694808 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002502050483599305 + ], + "opposite_logprob_tokens": [ + -6.002501964569092 + ], + "logprob_reward": -0.002502050483599305, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002502050483599305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09116334468126297, + 0.0622529610991478, + 0.2848721146583557, + 0.03339288756251335, + 0.029827475547790527, + 0.22547012567520142, + 0.051982298493385315, + 0.05961863696575165, + 0.12009014934301376, + 0.041330017149448395 + ], + "credit_map": [ + 0.00629264023154974, + 0.02195674553513527, + 0.004799503367394209, + 0.0075351763516664505, + 0.015873609110713005, + 0.01350241806358099, + 0.0071512493304908276, + 0.01405201107263565, + 0.0033309583086520433, + 0.014174392446875572, + 0.004603684414178133, + 0.010774059221148491, + 0.0060516768135130405, + 0.00952217634767294, + 0.008021521382033825, + 0.005774494260549545, + 0.017069611698389053, + 0.05333035811781883, + 0.012951016426086426, + 0.022445855662226677, + 0.03325017914175987, + 0.08015985786914825, + 0.0279886145144701, + 0.0376766175031662, + 0.0023874305188655853, + 0.004675874020904303, + 0.002740431809797883, + 0.005206710193306208, + 0.005310987588018179, + 0.003371928120031953, + 0.0059637087397277355, + 0.0037358184345066547, + 0.0013346214545890689, + 0.0046245865523815155, + 0.0013413962442427874, + 0.004396659787744284, + 0.0038013411685824394, + 0.005052053835242987, + 0.0069305612705647945, + 0.002346254652366042, + 0.010180838406085968, + 0.0211213119328022, + 0.038550663739442825, + 0.017937401309609413, + 0.04025421664118767, + 0.06289017200469971, + 0.02310965396463871, + 0.011425850912928581, + 0.002949761226773262, + 0.006975134368985891, + 0.005434085614979267, + 0.008914721198379993, + 0.007135332562029362, + 0.006190478336066008, + 0.007905789650976658, + 0.006476996932178736, + 0.003202607622370124, + 0.007243677508085966, + 0.0033012470230460167, + 0.0080728055909276, + 0.008823410607874393, + 0.007482271641492844, + 0.01267891563475132, + 0.00881369598209858, + 0.007225721143186092, + 0.020959343761205673, + 0.0073616779409348965, + 0.020422438159585, + 0.01621764712035656, + 0.015711462125182152, + 0.02061339281499386, + 0.01157846488058567, + 0.0023142064455896616, + 0.007134954445064068, + 0.0028385489713400602, + 0.005771643482148647, + 0.006290864665061235, + 0.00628129206597805, + 0.006762173026800156, + 0.003936332650482655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.002502050483599305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327502891421318 + ], + "opposite_logprob_tokens": [ + -3.773275136947632 + ], + "logprob_reward": -0.02327502891421318, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327502891421318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11611686646938324, + 0.07332104444503784, + 0.09965766966342926, + 0.11788725107908249, + 0.10295379906892776, + 0.16197244822978973, + 0.14025020599365234, + 0.10862188786268234, + 0.05471891909837723, + 0.024499941617250443 + ], + "credit_map": [ + 0.009504634886980057, + 0.023158451542258263, + 0.009562894701957703, + 0.009067430160939693, + 0.018366927281022072, + 0.014212166890501976, + 0.01304298173636198, + 0.019201377406716347, + 0.00850924663245678, + 0.007551182061433792, + 0.009593568742275238, + 0.011545616202056408, + 0.006893671117722988, + 0.014643510803580284, + 0.007175950333476067, + 0.007408300880342722, + 0.0042732348665595055, + 0.030446767807006836, + 0.00428518932312727, + 0.010861418209969997, + 0.008379774168133736, + 0.022079411894083023, + 0.010051158256828785, + 0.009280716069042683, + 0.011014132760465145, + 0.02092983014881611, + 0.010359266772866249, + 0.021758466958999634, + 0.021251019090414047, + 0.007335229776799679, + 0.014761846512556076, + 0.010477454401552677, + 0.006042063236236572, + 0.011463330127298832, + 0.007153406273573637, + 0.017298400402069092, + 0.014148086309432983, + 0.01718868315219879, + 0.016598980873823166, + 0.013060850091278553, + 0.007243126630783081, + 0.01471408549696207, + 0.01953132078051567, + 0.013573182746767998, + 0.03139924630522728, + 0.041890405118465424, + 0.01840980350971222, + 0.015211284160614014, + 0.007286234758794308, + 0.018097465857863426, + 0.01847967877984047, + 0.022011343389749527, + 0.017103178426623344, + 0.013154013082385063, + 0.018363414332270622, + 0.02575487643480301, + 0.008981605060398579, + 0.02948608435690403, + 0.005167135037481785, + 0.01116117462515831, + 0.016147490590810776, + 0.013358230702579021, + 0.016036782413721085, + 0.008283383212983608, + 0.0040664030238986015, + 0.00941807497292757, + 0.0036329859867691994, + 0.009169328026473522, + 0.006507528480142355, + 0.007435866165906191, + 0.009223271161317825, + 0.005265458021312952, + 0.0018826387822628021, + 0.005103765521198511, + 0.0017168433405458927, + 0.002756816102191806, + 0.002987207379192114, + 0.003400883637368679, + 0.0037426657509058714, + 0.002909122034907341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.02327502891421318 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639100283384323 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639100283384323, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639100283384323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1823272556066513, + 0.1106477901339531, + 0.1958526074886322, + 0.060941729694604874, + 0.04282297194004059, + 0.17658230662345886, + 0.051806557923555374, + 0.051325373351573944, + 0.08846395462751389, + 0.03922945261001587 + ], + "credit_map": [ + 0.013801570050418377, + 0.04016204550862312, + 0.013839383609592915, + 0.019431667402386665, + 0.025855956599116325, + 0.02704186551272869, + 0.020512225106358528, + 0.02168254740536213, + 0.009648812934756279, + 0.014818401075899601, + 0.014524689875543118, + 0.012957561761140823, + 0.010532265529036522, + 0.019485484808683395, + 0.011405589990317822, + 0.01727498136460781, + 0.012632188387215137, + 0.035685811191797256, + 0.009608438238501549, + 0.021924663335084915, + 0.021971160545945168, + 0.04648051783442497, + 0.023175014182925224, + 0.024374820291996002, + 0.00469469279050827, + 0.010691499337553978, + 0.004727471154183149, + 0.008373819291591644, + 0.008914041332900524, + 0.007089543621987104, + 0.006878279615193605, + 0.00957238208502531, + 0.0025729285553097725, + 0.005628886166960001, + 0.00339135411195457, + 0.0060406834818422794, + 0.00547728082165122, + 0.0074803465977311134, + 0.004967211280018091, + 0.007264280691742897, + 0.008513670414686203, + 0.010380560532212257, + 0.0532841719686985, + 0.014585352502763271, + 0.01448275987058878, + 0.0458863228559494, + 0.02304314635694027, + 0.006406325846910477, + 0.002638766774907708, + 0.008907424286007881, + 0.004081008490175009, + 0.009624729864299297, + 0.008180966600775719, + 0.0057579935528337955, + 0.0070732347667217255, + 0.005542431958019733, + 0.00358303333632648, + 0.011348973028361797, + 0.003360220929607749, + 0.008549804799258709, + 0.0061982255429029465, + 0.007297451142221689, + 0.005537522956728935, + 0.005450141616165638, + 0.004926296416670084, + 0.01104382798075676, + 0.005338519345968962, + 0.012201439589262009, + 0.011549507267773151, + 0.013410583138465881, + 0.021474650129675865, + 0.008519130758941174, + 0.0031399508006870747, + 0.004643578082323074, + 0.003057837951928377, + 0.004706356208771467, + 0.004901929758489132, + 0.005973872262984514, + 0.008416581898927689, + 0.0043893479742109776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.004639100283384323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two small, furry creatures visible engaging in a playful tussle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one creature covered in fluffy warm brown fur and the other in creamy white fur speckled with soft gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set against the sun-drenched, textured black plastic dashboard of a moving car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a static, slightly shaky close-up shot from eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is bright sunlight streaming through the windshield, creating a golden halo on the fur and casting sharp, dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817948115989566 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.002817948115989566, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817948115989566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15531085431575775, + 0.06171413138508797, + 0.1120222881436348, + 0.050910912454128265, + 0.043454691767692566, + 0.27701061964035034, + 0.07916709780693054, + 0.11466658860445023, + 0.06602209061384201, + 0.039720725268125534 + ], + "credit_map": [ + 0.010264521464705467, + 0.036105476319789886, + 0.009598645381629467, + 0.011994545347988605, + 0.029051892459392548, + 0.0202467180788517, + 0.010616568848490715, + 0.027432484552264214, + 0.0033498448319733143, + 0.011219479143619537, + 0.005107060540467501, + 0.01313844695687294, + 0.005438160616904497, + 0.007839336059987545, + 0.008229582570493221, + 0.007392214145511389, + 0.006200513802468777, + 0.01973208226263523, + 0.004634741228073835, + 0.012162646278738976, + 0.011177293956279755, + 0.02994714491069317, + 0.011328562162816525, + 0.016839299350976944, + 0.0039021556731313467, + 0.009411119855940342, + 0.004323997534811497, + 0.009332293644547462, + 0.009067357517778873, + 0.0035231730435043573, + 0.006443322170525789, + 0.004907488822937012, + 0.002420208416879177, + 0.004738734103739262, + 0.0030645469669252634, + 0.005754954181611538, + 0.0054588112980127335, + 0.008220382034778595, + 0.00557782081887126, + 0.008219229057431221, + 0.011889902874827385, + 0.019136721268296242, + 0.05731241777539253, + 0.023476846516132355, + 0.040441010147333145, + 0.07595892250537872, + 0.028419658541679382, + 0.020375143736600876, + 0.0044938246719539165, + 0.01135479286313057, + 0.010119929909706116, + 0.008841603994369507, + 0.009806573390960693, + 0.008884662762284279, + 0.008844451978802681, + 0.016821255907416344, + 0.008469438180327415, + 0.027539731934666634, + 0.0046601188369095325, + 0.012761170975863934, + 0.017131786793470383, + 0.013979119248688221, + 0.019377976655960083, + 0.01074723619967699, + 0.004908260889351368, + 0.011254648678004742, + 0.00473751500248909, + 0.009959269315004349, + 0.008267970755696297, + 0.009594508446753025, + 0.01123394537717104, + 0.006065967492759228, + 0.0026687164790928364, + 0.01001504622399807, + 0.0024152675177901983, + 0.0035786358639597893, + 0.005571207031607628, + 0.006086528301239014, + 0.005574795417487621, + 0.0038105263374745846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.002817948115989566 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_0/index_tensor([3239, 3239])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_0/index_tensor([3239, 3239])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e1c9b236d7cdf675bfb0fe37ed608c5087abffda --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_0/index_tensor([3239, 3239])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level, medium shot clinically surveys the interior of a vast, biomechanical, Giger-esque chamber, its walls a nightmarish amalgam of vast petrified ribcages and weeping black metal pipes. At the center of the frame, a grotesque rat king lies on a damp, grated industrial floor. This horrific amalgam consists of the carcasses of numerous rats, their gray fur and pale, decaying flesh intertwined with and pierced by a network of smooth, polished black metal sinews that gleam in the cold light. The entire body begins to experience a sickening, hyper-attenuated vibration; a low-frequency, visceral shudder that causes metallic ligaments to creak and individual carcasses to twitch in grotesquely synchronized spasms. The tremors intensify into violent convulsions, the entire body contracting tightly before movement ceases with an abrupt, jarring finality. The creature lapses into an unnatural, frozen silence, a monstrous sculpture of flesh and steel. The scene is lit by harsh, sterile top-down lighting, creating sharp, damp highlights on the creatures’ metallic parts and the polished floor, while casting the depths of the room in deep, inky shadows, reinforcing the cold, detached, and deeply unsettling atmosphere.", + "seed": 18003, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_0/index_3239_seed_18003_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1778275619726628, + "vlm_logprob": -0.1778275619726628, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.1778275619726628, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018849648535251617 + ], + "opposite_logprob_tokens": [ + -4.268849849700928 + ], + "logprob_reward": -0.018849648535251617, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018849648535251617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14802783727645874, + 0.10554113984107971, + 0.09061805158853531, + 0.16231399774551392, + 0.11616578698158264, + 0.11480951309204102, + 0.06865860521793365, + 0.06018684804439545, + 0.10529576987028122, + 0.02838243916630745 + ], + "credit_map": [ + 0.02757246233522892, + 0.018369639292359352, + 0.014846639707684517, + 0.017885752022266388, + 0.01854332908987999, + 0.0169941745698452, + 0.01962343044579029, + 0.014192414470016956, + 0.0157953929156065, + 0.011805294081568718, + 0.012852651067078114, + 0.02320890501141548, + 0.010936045087873936, + 0.009551305323839188, + 0.012918373569846153, + 0.008473170921206474, + 0.010173490270972252, + 0.011238130740821362, + 0.0088594239205122, + 0.013892789371311665, + 0.011609588749706745, + 0.011943361721932888, + 0.013981960713863373, + 0.008919310756027699, + 0.02491680718958378, + 0.01171189546585083, + 0.008654098957777023, + 0.027446430176496506, + 0.011999600566923618, + 0.01698504015803337, + 0.017952628433704376, + 0.04264749959111214, + 0.012907318770885468, + 0.006960304919630289, + 0.007751076947897673, + 0.015606005676090717, + 0.009096227586269379, + 0.008547449484467506, + 0.0318862609565258, + 0.023411136120557785, + 0.027015047147870064, + 0.009319853037595749, + 0.006316984072327614, + 0.023334894329309464, + 0.012588785961270332, + 0.010724754072725773, + 0.016716936603188515, + 0.00879225879907608, + 0.013664142228662968, + 0.006148769520223141, + 0.0028267349116504192, + 0.006218076217919588, + 0.005306299310177565, + 0.006350236013531685, + 0.02152313105762005, + 0.006621222477406263, + 0.0072875842452049255, + 0.005729693453758955, + 0.00529959611594677, + 0.009565168060362339, + 0.006825427524745464, + 0.00845012441277504, + 0.01189375389367342, + 0.005135500803589821, + 0.011256342753767967, + 0.008762015961110592, + 0.00900222361087799, + 0.02045859768986702, + 0.012296175584197044, + 0.01359099242836237, + 0.020979559049010277, + 0.008949865587055683, + 0.006419918965548277, + 0.002240607049316168, + 0.0027714965399354696, + 0.003281898330897093, + 0.003091441700235009, + 0.0037352964282035828, + 0.00345326098613441, + 0.003388519398868084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.018849648535251617 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017270424868911505 + ], + "opposite_logprob_tokens": [ + -6.376727104187012 + ], + "logprob_reward": -0.0017270424868911505, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017270424868911505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12169421464204788, + 0.1282653957605362, + 0.07799634337425232, + 0.1660017967224121, + 0.08887351304292679, + 0.15612246096134186, + 0.05555249750614166, + 0.04053094610571861, + 0.12200764566659927, + 0.04295520856976509 + ], + "credit_map": [ + 0.011161092668771744, + 0.013698408380150795, + 0.012865087948739529, + 0.016961239278316498, + 0.014576246030628681, + 0.013834563083946705, + 0.021641580387949944, + 0.01695599965751171, + 0.01642264425754547, + 0.008371870033442974, + 0.007962181232869625, + 0.03774195909500122, + 0.012258700095117092, + 0.009448221884667873, + 0.02797558903694153, + 0.008084232918918133, + 0.008056852035224438, + 0.009288091212511063, + 0.0045303101651370525, + 0.012706561014056206, + 0.009704772382974625, + 0.008423484861850739, + 0.018418092280626297, + 0.006868179887533188, + 0.02757343277335167, + 0.013081821613013744, + 0.010604099370539188, + 0.021868323907256126, + 0.010149692185223103, + 0.01824473962187767, + 0.015934986993670464, + 0.04854469746351242, + 0.026521144434809685, + 0.007296326570212841, + 0.006785486824810505, + 0.009414969943463802, + 0.00524985371157527, + 0.007208887953311205, + 0.012563822790980339, + 0.013833021745085716, + 0.03771446272730827, + 0.016646821051836014, + 0.008159724064171314, + 0.02910778485238552, + 0.01884884014725685, + 0.016191821545362473, + 0.016768837347626686, + 0.012684167362749577, + 0.012398737482726574, + 0.004518471658229828, + 0.002302245469763875, + 0.005760719068348408, + 0.004154396243393421, + 0.003154342994093895, + 0.01975463703274727, + 0.0035089473240077496, + 0.006860936060547829, + 0.002857377054169774, + 0.0031117070466279984, + 0.0035019535571336746, + 0.0030786250717937946, + 0.0038859834894537926, + 0.007021680474281311, + 0.010212681256234646, + 0.012306418269872665, + 0.009941515512764454, + 0.011808691546320915, + 0.022003430873155594, + 0.01700122281908989, + 0.015535758808255196, + 0.02199666202068329, + 0.011413943022489548, + 0.007438900880515575, + 0.005118811968713999, + 0.004093848168849945, + 0.004062136169523001, + 0.003967332653701305, + 0.008296252228319645, + 0.004645074252039194, + 0.005332851782441139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.0017270424868911505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294502019882202 + ], + "opposite_logprob_tokens": [ + -2.237945079803467 + ], + "logprob_reward": -0.11294502019882202, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294502019882202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13402068614959717, + 0.06993759423494339, + 0.09631416946649551, + 0.12667328119277954, + 0.12238219380378723, + 0.0855831578373909, + 0.05881814658641815, + 0.08732718229293823, + 0.18480610847473145, + 0.034137483686208725 + ], + "credit_map": [ + 0.02933122031390667, + 0.013065231963992119, + 0.008369829505681992, + 0.03678597882390022, + 0.014168119058012962, + 0.01200582180172205, + 0.012865095399320126, + 0.007429386023432016, + 0.010716279037296772, + 0.0061145201325416565, + 0.005238555837422609, + 0.014749472960829735, + 0.011383896693587303, + 0.006173837464302778, + 0.010958084836602211, + 0.004602947272360325, + 0.010198294185101986, + 0.005391022190451622, + 0.006590783130377531, + 0.014366884715855122, + 0.0081452876329422, + 0.010457661934196949, + 0.03177085518836975, + 0.009393374435603619, + 0.021145714446902275, + 0.009448212571442127, + 0.0078986631706357, + 0.021274253726005554, + 0.008177844807505608, + 0.013388833031058311, + 0.010309685952961445, + 0.035030074417591095, + 0.01695815660059452, + 0.005531605798751116, + 0.005493031349033117, + 0.01546125952154398, + 0.010781866498291492, + 0.008163418620824814, + 0.04281682148575783, + 0.01717604137957096, + 0.014544893987476826, + 0.008126496337354183, + 0.0050012338906526566, + 0.012610663659870625, + 0.013039649464190006, + 0.00987402256578207, + 0.015304135158658028, + 0.007082057651132345, + 0.013459007255733013, + 0.005661229137331247, + 0.0040484764613211155, + 0.007803349755704403, + 0.0064260936342179775, + 0.00486674252897501, + 0.011751911602914333, + 0.004801338072866201, + 0.013019917532801628, + 0.005483187735080719, + 0.008028588257730007, + 0.007413459941744804, + 0.008202438242733479, + 0.007293811999261379, + 0.020535193383693695, + 0.017350589856505394, + 0.02093237265944481, + 0.01369100995361805, + 0.018726816400885582, + 0.03883172571659088, + 0.01878894865512848, + 0.022691618651151657, + 0.032729875296354294, + 0.018413733690977097, + 0.0072420863434672356, + 0.0029161369893699884, + 0.003226366825401783, + 0.003972294274717569, + 0.0025732312351465225, + 0.007628357969224453, + 0.002730895532295108, + 0.003848114050924778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.11294502019882202 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28132304549217224 + ], + "opposite_logprob_tokens": [ + -1.4063230752944946 + ], + "logprob_reward": -0.28132304549217224, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28132304549217224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12842237949371338, + 0.14947567880153656, + 0.04357650876045227, + 0.04712921008467674, + 0.0763281062245369, + 0.12923938035964966, + 0.09376856684684753, + 0.10144020617008209, + 0.1864859014749527, + 0.044133976101875305 + ], + "credit_map": [ + 0.017344659194350243, + 0.013776437379419804, + 0.011104089207947254, + 0.033197108656167984, + 0.015318406745791435, + 0.013253414072096348, + 0.017628304660320282, + 0.006799974013119936, + 0.018680820241570473, + 0.018871761858463287, + 0.020957427099347115, + 0.023606522008776665, + 0.014001837000250816, + 0.01481515821069479, + 0.025856947526335716, + 0.012685221619904041, + 0.005361686460673809, + 0.00410502590239048, + 0.003532891860231757, + 0.005740907043218613, + 0.0050276340916752815, + 0.004679539706557989, + 0.010216179303824902, + 0.004912647884339094, + 0.005444395821541548, + 0.0033868919126689434, + 0.003499068785458803, + 0.013455190695822239, + 0.004824620671570301, + 0.004817820619791746, + 0.007232924457639456, + 0.0044683050364255905, + 0.010156276635825634, + 0.005738035310059786, + 0.006980251520872116, + 0.01224751677364111, + 0.006970311515033245, + 0.006223554722964764, + 0.018002664670348167, + 0.01000950112938881, + 0.024178728461265564, + 0.010667216964066029, + 0.00589363370090723, + 0.021144965663552284, + 0.016437625512480736, + 0.010630295611917973, + 0.028626544401049614, + 0.011660383082926273, + 0.016279857605695724, + 0.0077217998914420605, + 0.00689726835116744, + 0.01273440569639206, + 0.010376139543950558, + 0.008971026167273521, + 0.015513339079916477, + 0.015274737030267715, + 0.024269169196486473, + 0.007715744432061911, + 0.006220768205821514, + 0.012306397780776024, + 0.007104379590600729, + 0.008555421605706215, + 0.015112385153770447, + 0.020155956968665123, + 0.020190656185150146, + 0.014101131819188595, + 0.016642332077026367, + 0.03756875544786453, + 0.02134447731077671, + 0.02230616845190525, + 0.03434830531477928, + 0.019984088838100433, + 0.010087186470627785, + 0.0030540679581463337, + 0.004721376113593578, + 0.007725396193563938, + 0.0036590248346328735, + 0.0056320675648748875, + 0.0043334937654435635, + 0.004921368323266506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.28132304549217224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.474293053150177 + ], + "opposite_logprob_tokens": [ + -0.974293053150177 + ], + "logprob_reward": -0.474293053150177, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.474293053150177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16073280572891235, + 0.18018358945846558, + 0.035486720502376556, + 0.1753881871700287, + 0.07157687842845917, + 0.13891524076461792, + 0.05438753589987755, + 0.045175302773714066, + 0.08565617352724075, + 0.052497487515211105 + ], + "credit_map": [ + 0.0335957407951355, + 0.019290996715426445, + 0.015250533819198608, + 0.03388834372162819, + 0.015917224809527397, + 0.018352363258600235, + 0.012305563315749168, + 0.012132036499679089, + 0.021156316623091698, + 0.022810693830251694, + 0.023299962282180786, + 0.0322522409260273, + 0.016068296507000923, + 0.017864204943180084, + 0.030535398051142693, + 0.016196465119719505, + 0.004233781713992357, + 0.0039356485940515995, + 0.0023491911124438047, + 0.006845342926681042, + 0.0037493505515158176, + 0.004964273888617754, + 0.0065106963738799095, + 0.002898436738178134, + 0.027768928557634354, + 0.012448853813111782, + 0.009512089192867279, + 0.03266836702823639, + 0.011623211205005646, + 0.019000092521309853, + 0.01445572916418314, + 0.04791092127561569, + 0.008105444721877575, + 0.00555019685998559, + 0.007467574905604124, + 0.00873596128076315, + 0.007852901704609394, + 0.005728448275476694, + 0.016372451558709145, + 0.011763900518417358, + 0.020887956023216248, + 0.016263505443930626, + 0.009151719510555267, + 0.02083769626915455, + 0.02120407298207283, + 0.015096668154001236, + 0.025096753612160683, + 0.010376876220107079, + 0.012351641431450844, + 0.004432466812431812, + 0.0025387764908373356, + 0.004993570037186146, + 0.003933182917535305, + 0.00367074110545218, + 0.018510470166802406, + 0.003956690430641174, + 0.007646209094673395, + 0.003883851459249854, + 0.0030799133237451315, + 0.0069830771535634995, + 0.0048978584818542, + 0.004595080856233835, + 0.009759703651070595, + 0.004329606890678406, + 0.012910516932606697, + 0.006427091080695391, + 0.007194007746875286, + 0.01255710981786251, + 0.011873779818415642, + 0.011050920933485031, + 0.016346529126167297, + 0.00729621946811676, + 0.013457306660711765, + 0.00543651357293129, + 0.004741065204143524, + 0.005317866336554289, + 0.004393568728119135, + 0.007649017032235861, + 0.00474144471809268, + 0.0067607080563902855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.474293053150177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot clinically surveys the interior of a vast, biomechanical, Giger-esque chamber, its walls a nightmarish amalgam of vast petrified ribcages and weeping black metal pipes. At the center of the frame, a grotesque rat king lies on a damp, grated industrial floor. This horrific amalgam consists of the carcasses of numerous rats, their gray fur and pale, decaying flesh intertwined with and pierced by a network of smooth, polished black metal sinews that gleam in the cold light. The entire body begins to experience a sickening, hyper-attenuated vibration; a low-frequency, visceral shudder that causes metallic ligaments to creak and individual carcasses to twitch in grotesquely synchronized spasms. The tremors intensify into violent convulsions, the entire body contracting tightly before movement ceases with an abrupt, jarring finality. The creature lapses into an unnatural, frozen silence, a monstrous sculpture of flesh and steel. The scene is lit by harsh, sterile top-down lighting, creating sharp, damp highlights on the creatures’ metallic parts and the polished floor, while casting the depths of the room in deep, inky shadows, reinforcing the cold, detached, and deeply unsettling atmosphere.", + "seed": 18004, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_0/index_3239_seed_18004_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08985470135230571, + "vlm_logprob": -0.08985470135230571, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.08985470135230571, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011842305772006512 + ], + "opposite_logprob_tokens": [ + -4.761842250823975 + ], + "logprob_reward": -0.011842305772006512, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011842305772006512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18761169910430908, + 0.09647665172815323, + 0.056852515786886215, + 0.09190432727336884, + 0.1343103051185608, + 0.08214201778173447, + 0.07459020614624023, + 0.09703334420919418, + 0.09862740337848663, + 0.08045154809951782 + ], + "credit_map": [ + 0.03773079812526703, + 0.014393696561455727, + 0.01797494664788246, + 0.02641683630645275, + 0.015838678926229477, + 0.020127035677433014, + 0.028856080025434494, + 0.026273613795638084, + 0.013796992599964142, + 0.009479026310145855, + 0.011782708577811718, + 0.013324991799890995, + 0.010938075371086597, + 0.013090391643345356, + 0.012216843664646149, + 0.011847611516714096, + 0.00973593257367611, + 0.005631154403090477, + 0.007247650530189276, + 0.004792520776391029, + 0.01047772541642189, + 0.008567649871110916, + 0.00499181542545557, + 0.005408062133938074, + 0.01842292957007885, + 0.009305902756750584, + 0.008428060449659824, + 0.010585865937173367, + 0.009065002202987671, + 0.008402171544730663, + 0.016704900190234184, + 0.010989487171173096, + 0.021265361458063126, + 0.012458933517336845, + 0.025682177394628525, + 0.015438144095242023, + 0.019700361415743828, + 0.015016913414001465, + 0.012560841627418995, + 0.012187574990093708, + 0.009079738520085812, + 0.008164185099303722, + 0.010477591305971146, + 0.009026244282722473, + 0.015373239293694496, + 0.009129497222602367, + 0.011296037584543228, + 0.0095954779535532, + 0.010695695877075195, + 0.007049368228763342, + 0.011850296519696712, + 0.008306591771543026, + 0.012009494937956333, + 0.008692762814462185, + 0.007582198362797499, + 0.008403787389397621, + 0.011253013275563717, + 0.010546562261879444, + 0.011151491664350033, + 0.011624139733612537, + 0.02318567782640457, + 0.012863822281360626, + 0.008814002387225628, + 0.007594626862555742, + 0.01115861814469099, + 0.007316636387258768, + 0.014540533535182476, + 0.01598321832716465, + 0.014553443528711796, + 0.012328093871474266, + 0.01244446448981762, + 0.010302381590008736, + 0.009815122932195663, + 0.0048491209745407104, + 0.015822023153305054, + 0.0070789894089102745, + 0.008210071362555027, + 0.006581083871424198, + 0.016630586236715317, + 0.011464551091194153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.011842305772006512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001719545223750174 + ], + "opposite_logprob_tokens": [ + -6.3767194747924805 + ], + "logprob_reward": -0.001719545223750174, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001719545223750174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15186664462089539, + 0.1564616560935974, + 0.13743767142295837, + 0.050764184445142746, + 0.09116236865520477, + 0.0806361585855484, + 0.0586366131901741, + 0.10868857055902481, + 0.05609019100666046, + 0.10825596004724503 + ], + "credit_map": [ + 0.02005397155880928, + 0.015118224546313286, + 0.019663255661725998, + 0.016943885013461113, + 0.01956556737422943, + 0.02145463414490223, + 0.020526906475424767, + 0.018540211021900177, + 0.019947974011301994, + 0.014458144083619118, + 0.02488275244832039, + 0.027065090835094452, + 0.01757616363465786, + 0.016457512974739075, + 0.01879337802529335, + 0.017280641943216324, + 0.018624870106577873, + 0.012896757572889328, + 0.016135402023792267, + 0.021113010123372078, + 0.014247938990592957, + 0.020559612661600113, + 0.01712174341082573, + 0.01673833280801773, + 0.00787378940731287, + 0.0038944375701248646, + 0.007860341109335423, + 0.0072804102674126625, + 0.005100220907479525, + 0.0047685145400464535, + 0.008338210172951221, + 0.005648261401802301, + 0.011374115943908691, + 0.009394906461238861, + 0.013491972349584103, + 0.012661933898925781, + 0.012215475551784039, + 0.007710539270192385, + 0.013520913198590279, + 0.01079251617193222, + 0.011047267355024815, + 0.007252892479300499, + 0.012347565032541752, + 0.010660378262400627, + 0.010516841895878315, + 0.007670675870031118, + 0.008051943965256214, + 0.013088596053421497, + 0.008860318921506405, + 0.005632383748888969, + 0.008969896472990513, + 0.0060517736710608006, + 0.010421739891171455, + 0.007349515799432993, + 0.005093303043395281, + 0.006257686298340559, + 0.013746650889515877, + 0.00929450523108244, + 0.020942876115441322, + 0.014043242670595646, + 0.01526232436299324, + 0.012586272321641445, + 0.012279306538403034, + 0.010533392429351807, + 0.005872947163879871, + 0.0045214444398880005, + 0.007624316029250622, + 0.009100215509533882, + 0.008582329377532005, + 0.006572687532752752, + 0.008522133342921734, + 0.005294119007885456, + 0.02735096402466297, + 0.006368206813931465, + 0.03426141291856766, + 0.006563802249729633, + 0.010531533509492874, + 0.009701952338218689, + 0.007435614243149757, + 0.006042473949491978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.001719545223750174 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036191582679749 + ], + "opposite_logprob_tokens": [ + -1.910361886024475 + ], + "logprob_reward": -0.16036191582679749, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036191582679749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10482002794742584, + 0.10883297771215439, + 0.07344698160886765, + 0.29257932305336, + 0.12644922733306885, + 0.03753459081053734, + 0.04611450806260109, + 0.14681001007556915, + 0.019933415576815605, + 0.04347899183630943 + ], + "credit_map": [ + 0.012527411803603172, + 0.010636701248586178, + 0.010396392084658146, + 0.01419147476553917, + 0.01422341912984848, + 0.013889732770621777, + 0.015092701651155949, + 0.013862177729606628, + 0.013625475578010082, + 0.009460683912038803, + 0.01128922589123249, + 0.023033566772937775, + 0.01062417309731245, + 0.010972965508699417, + 0.015955347567796707, + 0.013871531002223492, + 0.012977108359336853, + 0.004470675718039274, + 0.006495988462120295, + 0.0108566889539361, + 0.00997028686106205, + 0.007486571557819843, + 0.011777549050748348, + 0.00941210426390171, + 0.0349569134414196, + 0.018583500757813454, + 0.030665535479784012, + 0.0587499774992466, + 0.02659863419830799, + 0.030434858053922653, + 0.05618707835674286, + 0.03640278801321983, + 0.017708461731672287, + 0.009958289563655853, + 0.022718163207173347, + 0.01915993168950081, + 0.013809021562337875, + 0.013822397217154503, + 0.014666319824755192, + 0.014606627635657787, + 0.0048479619435966015, + 0.0046194810420274734, + 0.005483255255967379, + 0.00322012254036963, + 0.005477764178067446, + 0.0035217185504734516, + 0.0037843186873942614, + 0.006579962093383074, + 0.006192744709551334, + 0.004396925680339336, + 0.006431852001696825, + 0.006240367889404297, + 0.0065123760141432285, + 0.00578736700117588, + 0.004678248893469572, + 0.005874615162611008, + 0.016509689390659332, + 0.010253910906612873, + 0.023266330361366272, + 0.023171495646238327, + 0.01618274673819542, + 0.019273845478892326, + 0.019816188141703606, + 0.018335796892642975, + 0.002622941741719842, + 0.0015446884790435433, + 0.002008080016821623, + 0.0025017859879881144, + 0.0028946995735168457, + 0.0023699949961155653, + 0.0034793384838849306, + 0.0025118838530033827, + 0.004531101789325476, + 0.004287982825189829, + 0.006056872196495533, + 0.00518845347687602, + 0.006488158367574215, + 0.005287571810185909, + 0.00687046255916357, + 0.004768386483192444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.16036191582679749 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252065509557724 + ], + "opposite_logprob_tokens": [ + -1.5020655393600464 + ], + "logprob_reward": -0.252065509557724, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252065509557724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10850462317466736, + 0.098556287586689, + 0.0882280096411705, + 0.18159876763820648, + 0.23865503072738647, + 0.16593559086322784, + 0.03572763130068779, + 0.03224816545844078, + 0.021575244143605232, + 0.028970586135983467 + ], + "credit_map": [ + 0.01551215723156929, + 0.008985861204564571, + 0.01252475380897522, + 0.01718587800860405, + 0.011011977680027485, + 0.01330804917961359, + 0.014983383007347584, + 0.014992566779255867, + 0.011471625417470932, + 0.007574901450425386, + 0.013229873962700367, + 0.01949942857027054, + 0.011828151531517506, + 0.008343664929270744, + 0.01519569382071495, + 0.011412947438657284, + 0.012915704399347305, + 0.006769656203687191, + 0.011424839496612549, + 0.013945961371064186, + 0.009326338768005371, + 0.008494261652231216, + 0.014708493836224079, + 0.0106427650898695, + 0.023640207946300507, + 0.013297471217811108, + 0.017801955342292786, + 0.039043426513671875, + 0.018787259235978127, + 0.018535545095801353, + 0.03221455216407776, + 0.01827835850417614, + 0.02828630991280079, + 0.02207847125828266, + 0.035820718854665756, + 0.028791505843400955, + 0.022602559998631477, + 0.023493649438023567, + 0.04241035878658295, + 0.035171475261449814, + 0.016613619402050972, + 0.010875307023525238, + 0.02448534406721592, + 0.017089154571294785, + 0.01679655723273754, + 0.014081932604312897, + 0.04434967786073685, + 0.02164400927722454, + 0.0054361228831112385, + 0.0033341010566800833, + 0.006483817007392645, + 0.004149576183408499, + 0.004721181467175484, + 0.004544900264590979, + 0.0031671763863414526, + 0.003890756983309984, + 0.004095630254596472, + 0.002221873961389065, + 0.004859469830989838, + 0.0052931057289242744, + 0.0030710939317941666, + 0.004392029717564583, + 0.004159718286246061, + 0.004155245143920183, + 0.0024236806202679873, + 0.0021195511799305677, + 0.0033950351644307375, + 0.00258560455404222, + 0.002723218873143196, + 0.0024873523507267237, + 0.0028941347263753414, + 0.0029466685373336077, + 0.003999494016170502, + 0.0037488420493900776, + 0.003590209875255823, + 0.0028867973014712334, + 0.005328372586518526, + 0.004602967761456966, + 0.0024968052748590708, + 0.0023170995991677046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.252065509557724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328423038125038 + ], + "opposite_logprob_tokens": [ + -3.7732841968536377 + ], + "logprob_reward": -0.02328423038125038, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328423038125038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32093939185142517, + 0.0939464122056961, + 0.060824908316135406, + 0.09029409289360046, + 0.09210861474275589, + 0.06179209426045418, + 0.05204063653945923, + 0.13759535551071167, + 0.06119106709957123, + 0.0292674470692873 + ], + "credit_map": [ + 0.05152496322989464, + 0.03350134938955307, + 0.03493507578969002, + 0.035011328756809235, + 0.05559913441538811, + 0.035687655210494995, + 0.03927212208509445, + 0.035407744348049164, + 0.01237956341356039, + 0.00965223927050829, + 0.010034559294581413, + 0.010213128291070461, + 0.017600232735276222, + 0.01194300502538681, + 0.010502006858587265, + 0.011621673591434956, + 0.01240492332726717, + 0.005956342443823814, + 0.005451193079352379, + 0.006585677154362202, + 0.012305104173719883, + 0.007370755076408386, + 0.005132111255079508, + 0.0056188032031059265, + 0.016666874289512634, + 0.010945392772555351, + 0.009835735894739628, + 0.012479172088205814, + 0.011768282391130924, + 0.009973730891942978, + 0.010061590000987053, + 0.008563312701880932, + 0.010529831983149052, + 0.008677897043526173, + 0.014733154326677322, + 0.010727016255259514, + 0.012509731575846672, + 0.008363334462046623, + 0.015904482454061508, + 0.010663166642189026, + 0.0075090425089001656, + 0.0059960284270346165, + 0.009291623719036579, + 0.006251963321119547, + 0.009672330692410469, + 0.0055835191160440445, + 0.00994991697371006, + 0.007537672761827707, + 0.006735288072377443, + 0.005824723280966282, + 0.006783721037209034, + 0.005772814154624939, + 0.010105271823704243, + 0.005918944254517555, + 0.004539973568171263, + 0.0063598984852433205, + 0.013982732780277729, + 0.011055883020162582, + 0.01752609945833683, + 0.019061872735619545, + 0.021924324333667755, + 0.01862870529294014, + 0.019892293959856033, + 0.015523438341915607, + 0.006224741227924824, + 0.004797172266989946, + 0.008850812911987305, + 0.008955991826951504, + 0.007057186681777239, + 0.005153809208422899, + 0.013635789975523949, + 0.0065155611373484135, + 0.003992045763880014, + 0.0037981164641678333, + 0.003929521888494492, + 0.0027856824453920126, + 0.005984160117805004, + 0.0038987454026937485, + 0.0025422400794923306, + 0.002336937002837658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02328423038125038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot clinically surveys the interior of a vast, biomechanical, Giger-esque chamber, its walls a nightmarish amalgam of vast petrified ribcages and weeping black metal pipes. At the center of the frame, a grotesque rat king lies on a damp, grated industrial floor. This horrific amalgam consists of the carcasses of numerous rats, their gray fur and pale, decaying flesh intertwined with and pierced by a network of smooth, polished black metal sinews that gleam in the cold light. The entire body begins to experience a sickening, hyper-attenuated vibration; a low-frequency, visceral shudder that causes metallic ligaments to creak and individual carcasses to twitch in grotesquely synchronized spasms. The tremors intensify into violent convulsions, the entire body contracting tightly before movement ceases with an abrupt, jarring finality. The creature lapses into an unnatural, frozen silence, a monstrous sculpture of flesh and steel. The scene is lit by harsh, sterile top-down lighting, creating sharp, damp highlights on the creatures’ metallic parts and the polished floor, while casting the depths of the room in deep, inky shadows, reinforcing the cold, detached, and deeply unsettling atmosphere.", + "seed": 18005, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_0/index_3239_seed_18005_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21244252477772535, + "vlm_logprob": -0.21244252477772535, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.21244252477772535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028188548982143402 + ], + "opposite_logprob_tokens": [ + -3.903188467025757 + ], + "logprob_reward": -0.028188548982143402, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.028188548982143402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23399990797042847, + 0.03340638801455498, + 0.4212605357170105, + 0.020652176812291145, + 0.03647206351161003, + 0.009614442475140095, + 0.04795822873711586, + 0.0477498397231102, + 0.07398998737335205, + 0.07489649951457977 + ], + "credit_map": [ + 0.023642513900995255, + 0.038302741944789886, + 0.02537561021745205, + 0.05680050328373909, + 0.012377392500638962, + 0.02200908586382866, + 0.031057314947247505, + 0.024434734135866165, + 0.003920373506844044, + 0.0051066153682768345, + 0.0035717796999961138, + 0.004868119489401579, + 0.005705417599529028, + 0.0036801951937377453, + 0.003434237092733383, + 0.0031196463387459517, + 0.07653216272592545, + 0.013326691463589668, + 0.015558833256363869, + 0.025570010766386986, + 0.01991894841194153, + 0.10607171058654785, + 0.14165104925632477, + 0.022631097584962845, + 0.002892141928896308, + 0.0024862554855644703, + 0.003250730223953724, + 0.002648201771080494, + 0.0021620672196149826, + 0.003402722766622901, + 0.0018751394236460328, + 0.0019349170615896583, + 0.004141999874264002, + 0.004684231709688902, + 0.004855376668274403, + 0.005941206589341164, + 0.003236232092604041, + 0.0042970916256308556, + 0.005059586837887764, + 0.004256334155797958, + 0.0012790693435817957, + 0.0012690804433077574, + 0.0011605179170146585, + 0.0008278992027044296, + 0.0010338618885725737, + 0.001729959505610168, + 0.0011157302651554346, + 0.001198322861455381, + 0.005873567424714565, + 0.004800658673048019, + 0.004774064291268587, + 0.012402496300637722, + 0.0046061621978878975, + 0.0056776078417897224, + 0.006146819330751896, + 0.0036768459249287844, + 0.009806819260120392, + 0.005319648887962103, + 0.007186648435890675, + 0.003507208777591586, + 0.0035525618586689234, + 0.010908410884439945, + 0.003363188588991761, + 0.004105349536985159, + 0.008991939947009087, + 0.0070455120876431465, + 0.008824371732771397, + 0.01543490681797266, + 0.006671152077615261, + 0.0082857059314847, + 0.011154095642268658, + 0.007582296617329121, + 0.019232414662837982, + 0.006288126111030579, + 0.006789836101233959, + 0.00611152034252882, + 0.003839619690552354, + 0.021346956491470337, + 0.005360110197216272, + 0.005927911028265953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.028188548982143402 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028299542609602213 + ], + "opposite_logprob_tokens": [ + -5.8778300285339355 + ], + "logprob_reward": -0.0028299542609602213, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028299542609602213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13527527451515198, + 0.07825052738189697, + 0.18724705278873444, + 0.036489229649305344, + 0.03697529435157776, + 0.04101474583148956, + 0.27632081508636475, + 0.04760511592030525, + 0.08421792089939117, + 0.07660408318042755 + ], + "credit_map": [ + 0.013783152215182781, + 0.014726940542459488, + 0.01925138756632805, + 0.0323224812746048, + 0.010748435743153095, + 0.014092429541051388, + 0.017531707882881165, + 0.012818729504942894, + 0.007763324771076441, + 0.00903723668307066, + 0.01076550129801035, + 0.017509236931800842, + 0.010631873272359371, + 0.007147993426769972, + 0.008646414615213871, + 0.006748937536031008, + 0.028625331819057465, + 0.005917346570640802, + 0.0065025151707232, + 0.012550744228065014, + 0.009167158976197243, + 0.05176954343914986, + 0.06152898445725441, + 0.01118540670722723, + 0.005088839679956436, + 0.0025436419527977705, + 0.005776224657893181, + 0.007940773852169514, + 0.0027249520644545555, + 0.005408002063632011, + 0.004073807969689369, + 0.002932985546067357, + 0.0056777046993374825, + 0.0023734953720122576, + 0.005367660894989967, + 0.007969173602759838, + 0.0022322332952171564, + 0.004608171060681343, + 0.004131602589040995, + 0.004615247715264559, + 0.005738598760217428, + 0.005112872924655676, + 0.007494347635656595, + 0.0055263107642531395, + 0.0034756551031023264, + 0.004008792340755463, + 0.004568179603666067, + 0.005089985206723213, + 0.041626714169979095, + 0.022836104035377502, + 0.03530092537403107, + 0.06290942430496216, + 0.024213822558522224, + 0.03500071167945862, + 0.03451276943087578, + 0.019920332357287407, + 0.008485354483127594, + 0.005012981593608856, + 0.008566402830183506, + 0.005097911227494478, + 0.004188173916190863, + 0.00723528116941452, + 0.005284437909722328, + 0.0037345688324421644, + 0.00998520664870739, + 0.006955565419048071, + 0.010357609018683434, + 0.014361243695020676, + 0.009093238040804863, + 0.009946784004569054, + 0.013346920721232891, + 0.010171341709792614, + 0.014931097626686096, + 0.009726005606353283, + 0.008319840766489506, + 0.00912898126989603, + 0.006070597097277641, + 0.012965643778443336, + 0.008122583851218224, + 0.0073393345810472965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.0028299542609602213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897710800170898 + ], + "opposite_logprob_tokens": [ + -2.578977108001709 + ], + "logprob_reward": -0.07897710800170898, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897710800170898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17103013396263123, + 0.07272184640169144, + 0.2201976329088211, + 0.05981912463903427, + 0.0609382763504982, + 0.10136278718709946, + 0.15628518164157867, + 0.03836030885577202, + 0.023530932143330574, + 0.09575387090444565 + ], + "credit_map": [ + 0.014443940483033657, + 0.015591629780828953, + 0.02201845683157444, + 0.05122379958629608, + 0.008935591205954552, + 0.019116822630167007, + 0.022888820618391037, + 0.016811061650514603, + 0.007668947335332632, + 0.009262839332222939, + 0.00804184377193451, + 0.015380565077066422, + 0.007530559785664082, + 0.009238205850124359, + 0.008342349901795387, + 0.007256535347551107, + 0.046258579939603806, + 0.00867378618568182, + 0.010268925689160824, + 0.014723226428031921, + 0.00994561705738306, + 0.047183021903038025, + 0.07202047109603882, + 0.011124012060463428, + 0.00891463179141283, + 0.00502593070268631, + 0.0094332629814744, + 0.01201677042990923, + 0.0057036541402339935, + 0.0064624096266925335, + 0.00712942099198699, + 0.005133043043315411, + 0.011052485555410385, + 0.005905205383896828, + 0.008359745144844055, + 0.011515887454152107, + 0.00253856647759676, + 0.006945432163774967, + 0.006151990499347448, + 0.008468964137136936, + 0.017790233716368675, + 0.007023159880191088, + 0.01946297287940979, + 0.01418185792863369, + 0.0074135116301476955, + 0.00965023785829544, + 0.015201111324131489, + 0.010639702901244164, + 0.021658651530742645, + 0.012490333057940006, + 0.022021200507879257, + 0.039033450186252594, + 0.012312511913478374, + 0.01957165263593197, + 0.018915381282567978, + 0.010281992144882679, + 0.005510492250323296, + 0.0036413706839084625, + 0.0067151617258787155, + 0.0023470495361834764, + 0.002511899918317795, + 0.011357930488884449, + 0.0041469307616353035, + 0.0021294730249792337, + 0.003287917235866189, + 0.0026402450166642666, + 0.0028491124976426363, + 0.004651769995689392, + 0.0022538118064403534, + 0.0024752505123615265, + 0.0030891234055161476, + 0.0022837005089968443, + 0.021724527701735497, + 0.00812266580760479, + 0.006666986271739006, + 0.007827596738934517, + 0.005393821746110916, + 0.029079649597406387, + 0.010352971032261848, + 0.006585649214684963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.07897710800170898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233262181282043 + ], + "opposite_logprob_tokens": [ + -0.8983262181282043 + ], + "logprob_reward": -0.5233262181282043, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233262181282043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058549486100673676, + 0.03796699270606041, + 0.4688483476638794, + 0.021327542141079903, + 0.051025405526161194, + 0.03450176492333412, + 0.034706179052591324, + 0.029581120237708092, + 0.240955650806427, + 0.022537529468536377 + ], + "credit_map": [ + 0.006564467679709196, + 0.007310816086828709, + 0.007044439669698477, + 0.013373375870287418, + 0.004590333439409733, + 0.006985028274357319, + 0.0064648594707250595, + 0.006216166540980339, + 0.004230338614434004, + 0.006428628694266081, + 0.005022476892918348, + 0.00856042467057705, + 0.0028255186043679714, + 0.003955119755119085, + 0.0038710851222276688, + 0.0030734024476259947, + 0.17346300184726715, + 0.019937176257371902, + 0.02506687305867672, + 0.021411798894405365, + 0.020933132618665695, + 0.03299712762236595, + 0.15055732429027557, + 0.024481922388076782, + 0.003237716155126691, + 0.002194553380832076, + 0.003518893150612712, + 0.0033905336167663336, + 0.0018802221165969968, + 0.002535267500206828, + 0.0027372378390282393, + 0.0018331207102164626, + 0.006256961729377508, + 0.004541872069239616, + 0.006064580753445625, + 0.008597485721111298, + 0.004176088608801365, + 0.0047495998442173, + 0.00871354341506958, + 0.007925277575850487, + 0.005206002853810787, + 0.0033216788433492184, + 0.005807927343994379, + 0.00441970769315958, + 0.002898271894082427, + 0.0036309738643467426, + 0.00412730174139142, + 0.005089901387691498, + 0.004629054572433233, + 0.003225566353648901, + 0.0048318887129426, + 0.006824642885476351, + 0.0031672094482928514, + 0.0038827068638056517, + 0.004119212739169598, + 0.0040258984081447124, + 0.005212764255702496, + 0.0037616391200572252, + 0.004498845431953669, + 0.0021187684033066034, + 0.002144177909940481, + 0.006635373458266258, + 0.0025284194853156805, + 0.0026811331044882536, + 0.02870398759841919, + 0.023742975667119026, + 0.027743499726057053, + 0.056057799607515335, + 0.018122367560863495, + 0.03082035295665264, + 0.033921532332897186, + 0.02184314839541912, + 0.004938632715493441, + 0.00200562528334558, + 0.0020678683649748564, + 0.0025015014689415693, + 0.0013793432153761387, + 0.006064699962735176, + 0.001996260369196534, + 0.0015835996018722653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.5233262181282043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42889079451560974 + ], + "opposite_logprob_tokens": [ + -1.0538908243179321 + ], + "logprob_reward": -0.42889079451560974, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42889079451560974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17122194170951843, + 0.0597204752266407, + 0.3459138870239258, + 0.04142022877931595, + 0.08126142621040344, + 0.041542764753103256, + 0.06484054028987885, + 0.040994446724653244, + 0.037366047501564026, + 0.1157183051109314 + ], + "credit_map": [ + 0.01848389022052288, + 0.01801992952823639, + 0.02367554046213627, + 0.043431270867586136, + 0.009366092272102833, + 0.01866430602967739, + 0.023298298940062523, + 0.016282616183161736, + 0.005586298182606697, + 0.005450159776955843, + 0.00883147306740284, + 0.006857889238744974, + 0.01588362269103527, + 0.007129551377147436, + 0.005435931496322155, + 0.004545547068119049, + 0.058986470103263855, + 0.011928966268897057, + 0.012173431925475597, + 0.02047460339963436, + 0.0173079501837492, + 0.08917254954576492, + 0.11644642800092697, + 0.019423484802246094, + 0.005357155576348305, + 0.00580875389277935, + 0.006287703290581703, + 0.0037345171440392733, + 0.005604282952845097, + 0.00544329546391964, + 0.004463727120310068, + 0.004720793571323156, + 0.015262799337506294, + 0.009975595399737358, + 0.009371427819132805, + 0.009358253329992294, + 0.011021786369383335, + 0.010284127667546272, + 0.00969848781824112, + 0.006288949865847826, + 0.008299540728330612, + 0.005126178730279207, + 0.0047387960366904736, + 0.0041627357713878155, + 0.00394406309351325, + 0.008499649353325367, + 0.003924502991139889, + 0.0028472961857914925, + 0.017512740567326546, + 0.004698324948549271, + 0.00800746027380228, + 0.009245708584785461, + 0.005313639063388109, + 0.010256868787109852, + 0.00602783914655447, + 0.0037779537960886955, + 0.007415372412651777, + 0.005174812860786915, + 0.0058828117325901985, + 0.002694571390748024, + 0.00275815324857831, + 0.010312676429748535, + 0.0026992398779839277, + 0.0040568094700574875, + 0.004488549195230007, + 0.004095593933016062, + 0.004665052983909845, + 0.006716128438711166, + 0.0034626175183802843, + 0.005066961515694857, + 0.005311952903866768, + 0.0035591928754001856, + 0.0251830592751503, + 0.012398089282214642, + 0.0101836659014225, + 0.009041469544172287, + 0.007226039655506611, + 0.032769493758678436, + 0.009722619317471981, + 0.009193874895572662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.42889079451560974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot clinically surveys the interior of a vast, biomechanical, Giger-esque chamber, its walls a nightmarish amalgam of vast petrified ribcages and weeping black metal pipes. At the center of the frame, a grotesque rat king lies on a damp, grated industrial floor. This horrific amalgam consists of the carcasses of numerous rats, their gray fur and pale, decaying flesh intertwined with and pierced by a network of smooth, polished black metal sinews that gleam in the cold light. The entire body begins to experience a sickening, hyper-attenuated vibration; a low-frequency, visceral shudder that causes metallic ligaments to creak and individual carcasses to twitch in grotesquely synchronized spasms. The tremors intensify into violent convulsions, the entire body contracting tightly before movement ceases with an abrupt, jarring finality. The creature lapses into an unnatural, frozen silence, a monstrous sculpture of flesh and steel. The scene is lit by harsh, sterile top-down lighting, creating sharp, damp highlights on the creatures’ metallic parts and the polished floor, while casting the depths of the room in deep, inky shadows, reinforcing the cold, detached, and deeply unsettling atmosphere.", + "seed": 18006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_0/index_3239_seed_18006_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.20125775495544077, + "vlm_logprob": -0.20125775495544077, + "vlm_yes_no_margin": 2.4750000119209288, + "vlm_token_credit": -0.20125775495544077, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026547741144895554 + ], + "opposite_logprob_tokens": [ + -3.90154767036438 + ], + "logprob_reward": -0.026547741144895554, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.026547741144895554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2724984884262085, + 0.11201062798500061, + 0.081930972635746, + 0.0589137002825737, + 0.04028039798140526, + 0.04804135859012604, + 0.11231435835361481, + 0.051284439861774445, + 0.07766611129045486, + 0.1450594961643219 + ], + "credit_map": [ + 0.025071797892451286, + 0.01432510744780302, + 0.03375054523348808, + 0.08742573112249374, + 0.016347205266356468, + 0.045623403042554855, + 0.03446023166179657, + 0.015494482591748238, + 0.01076607033610344, + 0.006625911220908165, + 0.013911361806094646, + 0.011880435049533844, + 0.011138031259179115, + 0.018095143139362335, + 0.02479034662246704, + 0.01480333972722292, + 0.01233399286866188, + 0.006037469021975994, + 0.009902619756758213, + 0.00875512883067131, + 0.011719364672899246, + 0.010499028488993645, + 0.01224770862609148, + 0.010435669682919979, + 0.008329260163009167, + 0.00456519890576601, + 0.0069320788607001305, + 0.006946604698896408, + 0.009133758954703808, + 0.008816435001790524, + 0.008032341487705708, + 0.00615803012624383, + 0.0038120984099805355, + 0.002589623909443617, + 0.004230649676173925, + 0.009467663243412971, + 0.0035831783898174763, + 0.006434234790503979, + 0.007396374829113483, + 0.0027665747329592705, + 0.005177060142159462, + 0.0029378265608102083, + 0.007504611276090145, + 0.005808309651911259, + 0.00374454353004694, + 0.006086742505431175, + 0.010392487049102783, + 0.006389779970049858, + 0.011437784880399704, + 0.006467917468398809, + 0.010706216096878052, + 0.03534083440899849, + 0.005541568156331778, + 0.0197439081966877, + 0.015132726170122623, + 0.0079434122890234, + 0.0052030389197170734, + 0.004119393415749073, + 0.006014969665557146, + 0.004402923863381147, + 0.005527505185455084, + 0.008058067411184311, + 0.012142092920839787, + 0.005816448479890823, + 0.008342606946825981, + 0.005997139494866133, + 0.00716329924762249, + 0.014091821387410164, + 0.01042982842773199, + 0.011119617149233818, + 0.013093727640807629, + 0.007428073789924383, + 0.01345670223236084, + 0.0060958280228078365, + 0.043089449405670166, + 0.01892077922821045, + 0.008393137715756893, + 0.02381184510886669, + 0.014040051959455013, + 0.017251716926693916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.026547741144895554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410511763766408 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.00410511763766408, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410511763766408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2517525553703308, + 0.15234564244747162, + 0.11248889565467834, + 0.04827006161212921, + 0.0876741036772728, + 0.05223041772842407, + 0.04361527040600777, + 0.04532875865697861, + 0.04500679299235344, + 0.1612875759601593 + ], + "credit_map": [ + 0.02428501285612583, + 0.01765560731291771, + 0.026318412274122238, + 0.05732891336083412, + 0.02053779549896717, + 0.037714481353759766, + 0.043033722788095474, + 0.02487858384847641, + 0.011916686780750751, + 0.011168401688337326, + 0.017155049368739128, + 0.02409435249865055, + 0.010516231879591942, + 0.0247044637799263, + 0.03690338134765625, + 0.015887051820755005, + 0.013071163557469845, + 0.007655233144760132, + 0.01264845859259367, + 0.02619956061244011, + 0.00719734700396657, + 0.01753976382315159, + 0.014906257390975952, + 0.013271108269691467, + 0.005174118094146252, + 0.0032557272352278233, + 0.005681552924215794, + 0.011450580321252346, + 0.0040704612620174885, + 0.007760148029774427, + 0.007174442522227764, + 0.003703027730807662, + 0.007662805262953043, + 0.0047292327508330345, + 0.009698648005723953, + 0.02306608110666275, + 0.004943763371556997, + 0.013636198826134205, + 0.0164145790040493, + 0.007522789761424065, + 0.005298301577568054, + 0.003348744474351406, + 0.006076524034142494, + 0.005468904040753841, + 0.0047582415863871574, + 0.007540411315858364, + 0.01232103817164898, + 0.007418245542794466, + 0.004152273293584585, + 0.002695963252335787, + 0.003948736935853958, + 0.0100031029433012, + 0.004811517428606749, + 0.006358070764690638, + 0.008272572420537472, + 0.0033730301074683666, + 0.0046763732098042965, + 0.0034647765569388866, + 0.005992073565721512, + 0.006551091093569994, + 0.004487206693738699, + 0.006194263231009245, + 0.010051203891634941, + 0.003911769017577171, + 0.005238773301243782, + 0.005227255169302225, + 0.004035627469420433, + 0.006878157611936331, + 0.004452919587492943, + 0.005692227277904749, + 0.008725842460989952, + 0.0047559901140630245, + 0.016162311658263206, + 0.00853207241743803, + 0.04530584439635277, + 0.012037793174386024, + 0.01221860758960247, + 0.02381281368434429, + 0.021225474774837494, + 0.021992651745676994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.00410511763766408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38706305623054504 + ], + "opposite_logprob_tokens": [ + -1.1370630264282227 + ], + "logprob_reward": -0.38706305623054504, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38706305623054504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17294181883335114, + 0.15649470686912537, + 0.053518567234277725, + 0.04733718931674957, + 0.08737517893314362, + 0.052154265344142914, + 0.2230546623468399, + 0.06177767738699913, + 0.07269048690795898, + 0.07265543937683105 + ], + "credit_map": [ + 0.018169770017266273, + 0.015338568016886711, + 0.025027459487318993, + 0.01707194373011589, + 0.01732349395751953, + 0.02351430244743824, + 0.03331374749541283, + 0.023182528093457222, + 0.01548326201736927, + 0.008596783503890038, + 0.026233987882733345, + 0.014641685411334038, + 0.013491688296198845, + 0.020295949652791023, + 0.0393708236515522, + 0.018380532041192055, + 0.0062559181824326515, + 0.0041313255205750465, + 0.006061994470655918, + 0.008579457178711891, + 0.004960807505995035, + 0.0059732431545853615, + 0.00984541792422533, + 0.0077104042284190655, + 0.00559064419940114, + 0.0026515100616961718, + 0.006232249550521374, + 0.011119137518107891, + 0.004259350243955851, + 0.007190647069364786, + 0.00611184211447835, + 0.004181806929409504, + 0.007697885390371084, + 0.005095374304801226, + 0.00882971752434969, + 0.018152376636862755, + 0.007111432496458292, + 0.012685979716479778, + 0.019832255318760872, + 0.00797016080468893, + 0.004789640661329031, + 0.004037884995341301, + 0.007476061582565308, + 0.006628039292991161, + 0.004511537496000528, + 0.006967510096728802, + 0.011908359825611115, + 0.005835228133946657, + 0.02343219518661499, + 0.013305636122822762, + 0.021526562049984932, + 0.06682868301868439, + 0.013449432328343391, + 0.0401676706969738, + 0.03168049082159996, + 0.012663987465202808, + 0.005244406405836344, + 0.0043005491606891155, + 0.006995475385338068, + 0.008169715292751789, + 0.005609486717730761, + 0.010254018008708954, + 0.013879992999136448, + 0.0073240348137915134, + 0.006742921657860279, + 0.004732741042971611, + 0.006406690459698439, + 0.02215251885354519, + 0.004312199540436268, + 0.0114931371062994, + 0.011228790506720543, + 0.005621491465717554, + 0.006352277938276529, + 0.0035709550138562918, + 0.02024681493639946, + 0.01003019418567419, + 0.004424228332936764, + 0.011377775110304356, + 0.007639296352863312, + 0.009013896808028221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.38706305623054504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870238661766052 + ], + "opposite_logprob_tokens": [ + -1.13702392578125 + ], + "logprob_reward": -0.3870238661766052, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870238661766052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18028905987739563, + 0.06586097925901413, + 0.07295434921979904, + 0.03965714946389198, + 0.053672678768634796, + 0.03012034296989441, + 0.10432028025388718, + 0.054322030395269394, + 0.050712257623672485, + 0.34809088706970215 + ], + "credit_map": [ + 0.017277127131819725, + 0.011451352387666702, + 0.01878049224615097, + 0.03370905667543411, + 0.021134749054908752, + 0.02893475815653801, + 0.034315094351768494, + 0.01468642707914114, + 0.006444043945521116, + 0.004231425933539867, + 0.008547809906303883, + 0.0058898464776575565, + 0.007386623416095972, + 0.008524535223841667, + 0.01801677606999874, + 0.006819917354732752, + 0.010338273830711842, + 0.00590936653316021, + 0.008371710777282715, + 0.010183928534388542, + 0.00907676387578249, + 0.009700191207230091, + 0.010040942579507828, + 0.00933317095041275, + 0.004333228804171085, + 0.003523556049913168, + 0.004974596668034792, + 0.005081637762486935, + 0.004668224602937698, + 0.004736420698463917, + 0.008300750516355038, + 0.004038735292851925, + 0.004799194633960724, + 0.002895335666835308, + 0.005016960669308901, + 0.01845490373671055, + 0.002894320758059621, + 0.0099337724968791, + 0.00660224724560976, + 0.0030759454239159822, + 0.0036750431172549725, + 0.0015729108126834035, + 0.004640122875571251, + 0.002664251020178199, + 0.0023489478044211864, + 0.004304463043808937, + 0.005446021910756826, + 0.005468582268804312, + 0.010327823460102081, + 0.005248029250651598, + 0.010323675349354744, + 0.03359740972518921, + 0.005637594498693943, + 0.018913784995675087, + 0.013472993858158588, + 0.006798977963626385, + 0.0043517169542610645, + 0.0030705961398780346, + 0.007416235748678446, + 0.005074190907180309, + 0.004977880511432886, + 0.007401819806545973, + 0.01541728712618351, + 0.006612301804125309, + 0.0047144815325737, + 0.0040378873236477375, + 0.004498268943279982, + 0.01500971894711256, + 0.0034572246950119734, + 0.007755322847515345, + 0.006855687126517296, + 0.00438366923481226, + 0.03455209732055664, + 0.017661260440945625, + 0.10409700870513916, + 0.028218287974596024, + 0.025479791685938835, + 0.05494440346956253, + 0.03357528895139694, + 0.049562741070985794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.3870238661766052 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast chamber with walls made of petrified ribcages and black metal pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a grotesque creature composed of intertwined rat carcasses and black metal sinews visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature undergo a sequence of vibrating, convulsing, and then freezing into a static state?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at eye-level in a static medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by harsh, sterile top-down lighting creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015489935874939 + ], + "opposite_logprob_tokens": [ + -1.7015490531921387 + ], + "logprob_reward": -0.2015489935874939, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015489935874939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36909598112106323, + 0.03630804643034935, + 0.22949036955833435, + 0.09117760509252548, + 0.027012383565306664, + 0.024435970932245255, + 0.061461590230464935, + 0.013289445079863071, + 0.1329977959394455, + 0.014730791561305523 + ], + "credit_map": [ + 0.03912397101521492, + 0.0220571868121624, + 0.036586564034223557, + 0.09226109087467194, + 0.04244018346071243, + 0.059489037841558456, + 0.049085766077041626, + 0.02805216796696186, + 0.00391570245847106, + 0.0022945404052734375, + 0.0034955674782395363, + 0.003936782944947481, + 0.003972399048507214, + 0.005022791679948568, + 0.007741074543446302, + 0.00592918461188674, + 0.029563751071691513, + 0.014488618820905685, + 0.024242078885436058, + 0.03936977684497833, + 0.02060750126838684, + 0.03151502460241318, + 0.036998216062784195, + 0.0327053964138031, + 0.013578006066381931, + 0.006742990575730801, + 0.00919156614691019, + 0.01606093719601631, + 0.012755616568028927, + 0.012085067108273506, + 0.013767403550446033, + 0.006996015086770058, + 0.0026789966505020857, + 0.0012812328059226274, + 0.0032073045149445534, + 0.006353206466883421, + 0.002351221861317754, + 0.0036282995715737343, + 0.004559400957077742, + 0.0029527193401008844, + 0.002173815853893757, + 0.001432792516425252, + 0.0022824748884886503, + 0.003884604200720787, + 0.0019174306653439999, + 0.0038864959497004747, + 0.005496690049767494, + 0.003361667972058058, + 0.006286480464041233, + 0.003267072606831789, + 0.005988811608403921, + 0.01927771605551243, + 0.003224751679226756, + 0.011072907596826553, + 0.00861027929931879, + 0.0037335685919970274, + 0.001356263062916696, + 0.0012054960243403912, + 0.0013291975483298302, + 0.001998696243390441, + 0.001474797143600881, + 0.001730297110043466, + 0.0029701171442866325, + 0.0012245801044628024, + 0.016602659597992897, + 0.008495510555803776, + 0.012282656505703926, + 0.035498421639204025, + 0.013034979812800884, + 0.018851161003112793, + 0.015697116032242775, + 0.012535287998616695, + 0.0018072751117870212, + 0.0010779600124806166, + 0.0025548222474753857, + 0.0015694338362663984, + 0.0015914221294224262, + 0.002129915403202176, + 0.00249886536039412, + 0.0015010967617854476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.2015489935874939 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_1/index_tensor([621, 621])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_1/index_tensor([621, 621])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..72101bff675d6419d6a24a028005e64aa245af43 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_1/index_tensor([621, 621])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a vertical, hand-painted image, captured in a medium close-up shot at eye level, an owner's hand, off-screen, slowly and playfully lures a golden Labrador puppy with a bone-shaped biscuit. The puppy's fluffy, warm-gold fur glows softly under the bright, warm, diffuse light. Its large, round, dark eyes gleam with desire, its pupils dilating as the biscuit moves, completely captivated by the treat. Its head sways with precise precision as the light-gold biscuit, delicately held between its fingers, slowly and precisely shifts left and right across the frame. Its pink tongue lolls out, its saliva glistening from the corner of its mouth, its greedy expression both innocent and comical. A warm, softly blurred interior background focuses the attention on this anticipatory interaction. The visual style of the entire picture resembles a classic children's book illustration, with soft lines, saturated colors and a watercolor-like texture. Every detail, from the fluffy feel of the dog's hair to the crispy texture of the biscuits, is vividly depicted, creating an atmosphere that is rich in details, full of childlike fun and warm and healing.", + "seed": 30120, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_1/index_621_seed_30120_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.022891136654652654, + "vlm_logprob": -0.022891136654652654, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.022891136654652654, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08148442953824997 + ], + "opposite_logprob_tokens": [ + -2.581484317779541 + ], + "logprob_reward": -0.08148442953824997, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08148442953824997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23023584485054016, + 0.0717034861445427, + 0.04025818407535553, + 0.0436297208070755, + 0.050697337836027145, + 0.0684991404414177, + 0.18407541513442993, + 0.09260008484125137, + 0.03936716914176941, + 0.17893363535404205 + ], + "credit_map": [ + 0.031804777681827545, + 0.0217082891613245, + 0.019639844074845314, + 0.02612278051674366, + 0.026369517669081688, + 0.030226444825530052, + 0.038668397814035416, + 0.035695791244506836, + 0.00942192878574133, + 0.005958376452326775, + 0.01065111719071865, + 0.005758774001151323, + 0.007862735539674759, + 0.009395845234394073, + 0.010272957384586334, + 0.012381751090288162, + 0.00654801819473505, + 0.0029365664813667536, + 0.007772600743919611, + 0.0034234863705933094, + 0.0030731551814824343, + 0.004885467234998941, + 0.0063341232016682625, + 0.005284765735268593, + 0.005773957818746567, + 0.004147826693952084, + 0.005580798722803593, + 0.004147473257035017, + 0.009064732119441032, + 0.0047615100629627705, + 0.005405553616583347, + 0.004747866652905941, + 0.005916179623454809, + 0.004573979414999485, + 0.01062456052750349, + 0.004644810687750578, + 0.0069986493326723576, + 0.004543202929198742, + 0.00792920682579279, + 0.005466754548251629, + 0.007573272101581097, + 0.005994163453578949, + 0.010033453814685345, + 0.01214542891830206, + 0.0068121314980089664, + 0.007016750052571297, + 0.008613074198365211, + 0.010310870595276356, + 0.024374252185225487, + 0.020873205736279488, + 0.012622971087694168, + 0.01466356124728918, + 0.022591985762119293, + 0.025037534534931183, + 0.026447027921676636, + 0.03746486455202103, + 0.013238008134067059, + 0.010204332880675793, + 0.007715769577771425, + 0.007398196961730719, + 0.012862415984272957, + 0.0112486332654953, + 0.011474208906292915, + 0.018458526581525803, + 0.004757366608828306, + 0.003155669430270791, + 0.006911857984960079, + 0.003289323067292571, + 0.004151500761508942, + 0.004236716777086258, + 0.004769427701830864, + 0.00809530820697546, + 0.024472055956721306, + 0.0178727637976408, + 0.027604887261986732, + 0.022593708708882332, + 0.015083221718668938, + 0.020714230835437775, + 0.019609494134783745, + 0.030983272939920425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.08148442953824997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006526962388306856 + ], + "opposite_logprob_tokens": [ + -7.37565279006958 + ], + "logprob_reward": -0.0006526962388306856, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006526962388306856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18806138634681702, + 0.06475577503442764, + 0.0811188742518425, + 0.048344213515520096, + 0.07152872532606125, + 0.05868903174996376, + 0.1255733072757721, + 0.07852533459663391, + 0.07497964054346085, + 0.20842374861240387 + ], + "credit_map": [ + 0.023784201592206955, + 0.01890278421342373, + 0.0327780544757843, + 0.01836695894598961, + 0.019644664600491524, + 0.021968776360154152, + 0.02749340608716011, + 0.025122545659542084, + 0.009277780540287495, + 0.006095679476857185, + 0.010990114882588387, + 0.006223300006240606, + 0.008118120953440666, + 0.0075377062894403934, + 0.008977224119007587, + 0.007535850163549185, + 0.0138755664229393, + 0.0057913255877792835, + 0.014245456084609032, + 0.007450325880199671, + 0.009111124090850353, + 0.005657052155584097, + 0.014996289275586605, + 0.009991739876568317, + 0.0076245819218456745, + 0.004620480351150036, + 0.005360821727663279, + 0.003980834502726793, + 0.006329802796244621, + 0.006077520549297333, + 0.005887478590011597, + 0.008462689816951752, + 0.00830104760825634, + 0.005888666957616806, + 0.0076072923839092255, + 0.008588496595621109, + 0.012119892984628677, + 0.008828401565551758, + 0.00819442793726921, + 0.01200050488114357, + 0.00672202417626977, + 0.005630122497677803, + 0.007279593963176012, + 0.010577215813100338, + 0.007905501872301102, + 0.006180103402584791, + 0.008994163013994694, + 0.0054003088735044, + 0.014919868670403957, + 0.013321627862751484, + 0.011137103661894798, + 0.0112962881103158, + 0.01682768017053604, + 0.01488692220300436, + 0.015110847540199757, + 0.028072966262698174, + 0.010978997685015202, + 0.007963179610669613, + 0.018391046673059464, + 0.00820112507790327, + 0.006357811391353607, + 0.007432207465171814, + 0.00930444709956646, + 0.009896521456539631, + 0.009232827462255955, + 0.0065659936517477036, + 0.014289221726357937, + 0.004511486738920212, + 0.008980054408311844, + 0.011173449456691742, + 0.00622407766059041, + 0.014002534560859203, + 0.02758619748055935, + 0.018972907215356827, + 0.038240619003772736, + 0.029374413192272186, + 0.018588639795780182, + 0.02469119429588318, + 0.02049945294857025, + 0.030470313504338264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0006526962388306856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590364933013916 + ], + "opposite_logprob_tokens": [ + -3.895590305328369 + ], + "logprob_reward": -0.020590364933013916, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590364933013916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14735101163387299, + 0.04838035628199577, + 0.11439648270606995, + 0.054276928305625916, + 0.05317704752087593, + 0.04595813900232315, + 0.072787806391716, + 0.33117619156837463, + 0.09785734862089157, + 0.034638676792383194 + ], + "credit_map": [ + 0.01732480712234974, + 0.012879731133580208, + 0.03975095972418785, + 0.020702892914414406, + 0.012603051029145718, + 0.011378264985978603, + 0.019162258133292198, + 0.013549048453569412, + 0.006212878040969372, + 0.0033781742677092552, + 0.011335261166095734, + 0.0055275470949709415, + 0.005356261041015387, + 0.004148933105170727, + 0.007420733571052551, + 0.0050005726516246796, + 0.01672506146132946, + 0.00730083417147398, + 0.032479602843523026, + 0.012652897275984287, + 0.008286910131573677, + 0.01117266621440649, + 0.013980261981487274, + 0.011798256076872349, + 0.007558657322078943, + 0.0041829366236925125, + 0.016428187489509583, + 0.004844342824071646, + 0.005027721170336008, + 0.004143720492720604, + 0.006237377412617207, + 0.005853986833244562, + 0.007028747349977493, + 0.0047506061382591724, + 0.013226151466369629, + 0.00550333596765995, + 0.006160830147564411, + 0.004894872196018696, + 0.006385037209838629, + 0.005227472633123398, + 0.0054090749472379684, + 0.004686976317316294, + 0.008265502750873566, + 0.005272264592349529, + 0.00500596733763814, + 0.005551428068429232, + 0.00571820605546236, + 0.00604872265830636, + 0.010055300779640675, + 0.00723250862210989, + 0.008810114115476608, + 0.005649307277053595, + 0.006498868111521006, + 0.009668007493019104, + 0.008695695549249649, + 0.016178013756871223, + 0.04445758834481239, + 0.028302138671278954, + 0.0747360959649086, + 0.06833134591579437, + 0.023490650579333305, + 0.026663441210985184, + 0.02605578862130642, + 0.039139196276664734, + 0.01116044633090496, + 0.008697450160980225, + 0.02523157186806202, + 0.007985379546880722, + 0.009665298275649548, + 0.012064137496054173, + 0.007906587794423103, + 0.01514649111777544, + 0.004791026469320059, + 0.0027526665944606066, + 0.00410419050604105, + 0.003458543913438916, + 0.0029282046016305685, + 0.004748526029288769, + 0.0035965468268841505, + 0.008258974179625511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.020590364933013916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785657748579979 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785657748579979, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785657748579979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1676342785358429, + 0.06263750046491623, + 0.09519866108894348, + 0.04302355647087097, + 0.07355041056871414, + 0.19028694927692413, + 0.05358951911330223, + 0.07312172651290894, + 0.07934024930000305, + 0.1616172194480896 + ], + "credit_map": [ + 0.01867680251598358, + 0.013470723293721676, + 0.031845685094594955, + 0.023714065551757812, + 0.02005414478480816, + 0.01219094730913639, + 0.02414640784263611, + 0.023535482585430145, + 0.009114980697631836, + 0.005575400311499834, + 0.010393308475613594, + 0.0051148091442883015, + 0.007520516403019428, + 0.006248156540095806, + 0.009062446653842926, + 0.009607871994376183, + 0.012677703984081745, + 0.0069891358725726604, + 0.022015903145074844, + 0.007948086597025394, + 0.011469931341707706, + 0.00806331541389227, + 0.013458656147122383, + 0.01257591973990202, + 0.006995535455644131, + 0.003276228439062834, + 0.009028375148773193, + 0.003504436928778887, + 0.005074887070804834, + 0.003830997971817851, + 0.005361797288060188, + 0.005951296538114548, + 0.00823226384818554, + 0.008480099961161613, + 0.008763779886066914, + 0.008117957971990108, + 0.013269095681607723, + 0.007065482437610626, + 0.009872164577245712, + 0.009749560616910458, + 0.020459968596696854, + 0.01574750617146492, + 0.014342623762786388, + 0.04172155261039734, + 0.020892787724733353, + 0.018750030547380447, + 0.026871291920542717, + 0.0315011702477932, + 0.005968111101537943, + 0.0043190415017306805, + 0.005493510514497757, + 0.005292871966958046, + 0.009605945087969303, + 0.0058100842870771885, + 0.00833977572619915, + 0.008760171942412853, + 0.009695327840745449, + 0.006657863035798073, + 0.012637249194085598, + 0.007992581464350224, + 0.006900706794112921, + 0.008495721966028214, + 0.008422703482210636, + 0.012319566681981087, + 0.009750301018357277, + 0.005805304739624262, + 0.02256755530834198, + 0.00790709350258112, + 0.006204175762832165, + 0.007556508295238018, + 0.0076803239062428474, + 0.011868981644511223, + 0.0230665635317564, + 0.013454913161695004, + 0.016671808436512947, + 0.016695119440555573, + 0.012637513689696789, + 0.02523917891085148, + 0.01609557867050171, + 0.0377565361559391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.009785657748579979 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019425348145887256 + ], + "opposite_logprob_tokens": [ + -6.2519426345825195 + ], + "logprob_reward": -0.0019425348145887256, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019425348145887256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15679700672626495, + 0.08940195292234421, + 0.051066987216472626, + 0.09875234216451645, + 0.16387687623500824, + 0.08430658280849457, + 0.055500589311122894, + 0.0403011180460453, + 0.07724067568778992, + 0.18275588750839233 + ], + "credit_map": [ + 0.019838569685816765, + 0.015490995720028877, + 0.02155325375497341, + 0.03358767181634903, + 0.020527204498648643, + 0.012011454440653324, + 0.01977439597249031, + 0.014013472013175488, + 0.01128991600126028, + 0.006998490542173386, + 0.01862027868628502, + 0.009973309934139252, + 0.011919928714632988, + 0.008185749873518944, + 0.012129578739404678, + 0.010284698568284512, + 0.007261582184582949, + 0.004684599116444588, + 0.011632864363491535, + 0.0050125024281442165, + 0.005830278620123863, + 0.004291938152164221, + 0.007044794037938118, + 0.005308427382260561, + 0.010643337853252888, + 0.0074172718450427055, + 0.021599050611257553, + 0.00918874517083168, + 0.013250946998596191, + 0.013133746571838856, + 0.012268690392374992, + 0.011250556446611881, + 0.022318372502923012, + 0.012124159373342991, + 0.05047529190778732, + 0.01654636114835739, + 0.011464779265224934, + 0.012210201472043991, + 0.01903834007680416, + 0.01969936303794384, + 0.0093887560069561, + 0.00867743231356144, + 0.010166340507566929, + 0.013852518983185291, + 0.013549013063311577, + 0.007052585482597351, + 0.01225154846906662, + 0.009368385188281536, + 0.0072129652835428715, + 0.0037875270936638117, + 0.011011261492967606, + 0.004569568671286106, + 0.005793726071715355, + 0.006403140723705292, + 0.004906456917524338, + 0.011815942823886871, + 0.004927425179630518, + 0.003551644505932927, + 0.006729584652930498, + 0.004076805897057056, + 0.005081477575004101, + 0.004653737880289555, + 0.006541075650602579, + 0.004739366006106138, + 0.009506037458777428, + 0.008080806583166122, + 0.02035829797387123, + 0.005763245280832052, + 0.008033321239054203, + 0.00928531214594841, + 0.005866489838808775, + 0.010347168892621994, + 0.027694545686244965, + 0.01626565307378769, + 0.029093241319060326, + 0.025895453989505768, + 0.013660354539752007, + 0.021192770451307297, + 0.013184494338929653, + 0.03576938062906265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.0019425348145887256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a vertical, hand-painted image, captured in a medium close-up shot at eye level, an owner's hand, off-screen, slowly and playfully lures a golden Labrador puppy with a bone-shaped biscuit. The puppy's fluffy, warm-gold fur glows softly under the bright, warm, diffuse light. Its large, round, dark eyes gleam with desire, its pupils dilating as the biscuit moves, completely captivated by the treat. Its head sways with precise precision as the light-gold biscuit, delicately held between its fingers, slowly and precisely shifts left and right across the frame. Its pink tongue lolls out, its saliva glistening from the corner of its mouth, its greedy expression both innocent and comical. A warm, softly blurred interior background focuses the attention on this anticipatory interaction. The visual style of the entire picture resembles a classic children's book illustration, with soft lines, saturated colors and a watercolor-like texture. Every detail, from the fluffy feel of the dog's hair to the crispy texture of the biscuits, is vividly depicted, creating an atmosphere that is rich in details, full of childlike fun and warm and healing.", + "seed": 30121, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_1/index_621_seed_30121_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008071500947698951, + "vlm_logprob": -0.008071500947698951, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.008071500947698951, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016439933329820633 + ], + "opposite_logprob_tokens": [ + -4.266439914703369 + ], + "logprob_reward": -0.016439933329820633, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016439933329820633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1800965964794159, + 0.06950646638870239, + 0.036131083965301514, + 0.03660272806882858, + 0.09308883547782898, + 0.1362277716398239, + 0.18671299517154694, + 0.22176747024059296, + 0.009577091783285141, + 0.03028901293873787 + ], + "credit_map": [ + 0.024676667526364326, + 0.012331872247159481, + 0.031427301466464996, + 0.023038523271679878, + 0.023278098553419113, + 0.02018413506448269, + 0.02288687601685524, + 0.02227311208844185, + 0.009680192917585373, + 0.004865037743002176, + 0.00827255379408598, + 0.012714475393295288, + 0.007427028845995665, + 0.008361735381186008, + 0.00908011756837368, + 0.009105313569307327, + 0.0038606394082307816, + 0.0024885234888643026, + 0.006265409756451845, + 0.0065669468604028225, + 0.004942108877003193, + 0.00283970613963902, + 0.005174477584660053, + 0.003993266727775335, + 0.003682703012600541, + 0.004073068965226412, + 0.003994053229689598, + 0.005188234616070986, + 0.0060358839109539986, + 0.0043203141540288925, + 0.005223351530730724, + 0.0040851132944226265, + 0.009539843536913395, + 0.0069777341559529305, + 0.01596764475107193, + 0.018963739275932312, + 0.00945720262825489, + 0.008674530312418938, + 0.008919628337025642, + 0.014588501304388046, + 0.020424604415893555, + 0.011057422496378422, + 0.02727483958005905, + 0.020818548277020454, + 0.014731657691299915, + 0.013156714849174023, + 0.013142869807779789, + 0.015621107071638107, + 0.02285093441605568, + 0.013615820556879044, + 0.028889315202832222, + 0.038214679807424545, + 0.02213902771472931, + 0.01642179861664772, + 0.02156578190624714, + 0.023015625774860382, + 0.028001897037029266, + 0.020654847845435143, + 0.045240290462970734, + 0.0391630157828331, + 0.015332263894379139, + 0.02251739613711834, + 0.02242250181734562, + 0.028435232117772102, + 0.001410583732649684, + 0.0007100064540281892, + 0.0018355073407292366, + 0.0017584423767402768, + 0.0007018993492238224, + 0.0011136679677292705, + 0.0009096993599087, + 0.0011372842127457261, + 0.004188244231045246, + 0.0021393774077296257, + 0.006745640654116869, + 0.004844217095524073, + 0.0022970044519752264, + 0.0029071124736219645, + 0.002411748981103301, + 0.004755664151161909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.016439933329820633 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005728749092668295 + ], + "opposite_logprob_tokens": [ + -7.500572681427002 + ], + "logprob_reward": -0.0005728749092668295, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005728749092668295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14993634819984436, + 0.07504171133041382, + 0.06820265203714371, + 0.03276952728629112, + 0.10163939744234085, + 0.1335797905921936, + 0.17311124503612518, + 0.19778352975845337, + 0.028815757483243942, + 0.03912004455924034 + ], + "credit_map": [ + 0.022194543853402138, + 0.011302158236503601, + 0.017776988446712494, + 0.014098621904850006, + 0.023190531879663467, + 0.02132587507367134, + 0.018561221659183502, + 0.021486405283212662, + 0.007469603326171637, + 0.005848710425198078, + 0.01920757070183754, + 0.009803698398172855, + 0.0070515647530555725, + 0.008363042958080769, + 0.006146558094769716, + 0.011150958016514778, + 0.007077150512486696, + 0.007314995862543583, + 0.011374677531421185, + 0.007337166462093592, + 0.008609470911324024, + 0.01003983709961176, + 0.008146321401000023, + 0.008303032256662846, + 0.0032402751967310905, + 0.003174210200086236, + 0.0036903780419379473, + 0.005840769037604332, + 0.004876095335930586, + 0.004054703749716282, + 0.004335527308285236, + 0.0035575679503381252, + 0.00971955806016922, + 0.00901392474770546, + 0.015383780002593994, + 0.014810453169047832, + 0.013950553722679615, + 0.012220016680657864, + 0.010729547590017319, + 0.01581157185137272, + 0.02037489041686058, + 0.011631431989371777, + 0.024225397035479546, + 0.020494922995567322, + 0.014642530120909214, + 0.012220199219882488, + 0.014963203109800816, + 0.015027224086225033, + 0.02153603546321392, + 0.011537706479430199, + 0.026171158999204636, + 0.035812780261039734, + 0.01987297832965851, + 0.016558464616537094, + 0.02057710476219654, + 0.021045012399554253, + 0.02541794255375862, + 0.018980685621500015, + 0.03976890444755554, + 0.033815156668424606, + 0.013691515661776066, + 0.019582360982894897, + 0.02099531888961792, + 0.025531644001603127, + 0.003475737292319536, + 0.0028218335937708616, + 0.0030026123858988285, + 0.0037854788824915886, + 0.002890202682465315, + 0.004609908442944288, + 0.004762894008308649, + 0.0034670911263674498, + 0.005262947175651789, + 0.00265232240781188, + 0.008404972031712532, + 0.005985046271234751, + 0.0026182078290730715, + 0.004628423601388931, + 0.003456351114436984, + 0.0061117736622691154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0005728749092668295 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200941659510136 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014200941659510136, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200941659510136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2857389450073242, + 0.11992926895618439, + 0.1312735676765442, + 0.07792304456233978, + 0.08583822101354599, + 0.12676183879375458, + 0.07271966338157654, + 0.03651576489210129, + 0.029690565541386604, + 0.033609017729759216 + ], + "credit_map": [ + 0.027862317860126495, + 0.015636032447218895, + 0.08506398648023605, + 0.05755235254764557, + 0.01745004765689373, + 0.02326282300055027, + 0.026316486299037933, + 0.03259490802884102, + 0.01472452376037836, + 0.007752061355859041, + 0.024333883076906204, + 0.02708105929195881, + 0.007606019265949726, + 0.012031962163746357, + 0.011788466013967991, + 0.014611302874982357, + 0.01676797866821289, + 0.006708920001983643, + 0.02976854145526886, + 0.025988681241869926, + 0.011118058115243912, + 0.00929195061326027, + 0.013996891677379608, + 0.017632566392421722, + 0.008639643900096416, + 0.006115502677857876, + 0.010847768746316433, + 0.01927373558282852, + 0.007439404726028442, + 0.005927305202931166, + 0.01149735413491726, + 0.008182343095541, + 0.008437618613243103, + 0.005826421082019806, + 0.015035253018140793, + 0.018823212012648582, + 0.008331581950187683, + 0.0072378357872366905, + 0.008239001035690308, + 0.013907310552895069, + 0.013640420511364937, + 0.008660176768898964, + 0.028882086277008057, + 0.02030712552368641, + 0.010928750969469547, + 0.012871883809566498, + 0.014672932215034962, + 0.01679847203195095, + 0.008482984267175198, + 0.006954228039830923, + 0.009108537808060646, + 0.008540538139641285, + 0.007562951650470495, + 0.012090854346752167, + 0.009540832601487637, + 0.010438741184771061, + 0.005261064041405916, + 0.003846656298264861, + 0.004667513072490692, + 0.005587683990597725, + 0.0037609904538840055, + 0.003885420970618725, + 0.005128353368490934, + 0.004378089215606451, + 0.004650890361517668, + 0.0016212278278544545, + 0.006016142666339874, + 0.005835270509123802, + 0.0019008463714271784, + 0.0024884496815502644, + 0.0028568992856889963, + 0.004320841748267412, + 0.0036440957337617874, + 0.0018789819441735744, + 0.008356679230928421, + 0.006141009274870157, + 0.0023875797633081675, + 0.0030571138486266136, + 0.0034657756332308054, + 0.0046777864918112755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.014200941659510136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950592458248138 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950592458248138, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950592458248138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.296745628118515, + 0.09223377704620361, + 0.10793003439903259, + 0.06021929159760475, + 0.08822770416736603, + 0.09349559992551804, + 0.10191317647695541, + 0.0802980586886406, + 0.026722373440861702, + 0.05221432447433472 + ], + "credit_map": [ + 0.027621161192655563, + 0.017426885664463043, + 0.07580452412366867, + 0.04614553973078728, + 0.02797313965857029, + 0.03134939447045326, + 0.03308175504207611, + 0.03734324127435684, + 0.011576217599213123, + 0.006063795182853937, + 0.01416785642504692, + 0.017978012561798096, + 0.010100106708705425, + 0.009080610238015652, + 0.013308671303093433, + 0.00995851680636406, + 0.011531546711921692, + 0.010467424057424068, + 0.012410200200974941, + 0.01754540577530861, + 0.011563828215003014, + 0.016033906489610672, + 0.013743224553763866, + 0.014634507708251476, + 0.006898721680045128, + 0.0034064287319779396, + 0.010726068168878555, + 0.01270142663270235, + 0.007221266161650419, + 0.005639601033180952, + 0.006767468061298132, + 0.006858317647129297, + 0.0083868857473135, + 0.006129844579845667, + 0.016519751399755478, + 0.011955518275499344, + 0.012831415981054306, + 0.0077114226296544075, + 0.011686055921018124, + 0.013006815686821938, + 0.009878144599497318, + 0.006929585710167885, + 0.014891032129526138, + 0.010365482419729233, + 0.010493436828255653, + 0.012317094951868057, + 0.018993664532899857, + 0.009627167135477066, + 0.013860361650586128, + 0.008486655540764332, + 0.010823637247085571, + 0.01620115339756012, + 0.012128370814025402, + 0.016505148261785507, + 0.012630008161067963, + 0.01127785537391901, + 0.010791006498038769, + 0.0062935492023825645, + 0.014326036907732487, + 0.010858496651053429, + 0.00765977380797267, + 0.007966551929712296, + 0.011801696382462978, + 0.010600948706269264, + 0.0031162742525339127, + 0.0025474182330071926, + 0.003333903616294265, + 0.004065667279064655, + 0.0028894792776554823, + 0.003971786238253117, + 0.004189749713987112, + 0.002608098089694977, + 0.006167922168970108, + 0.003713051090016961, + 0.013539436273276806, + 0.0069527518935501575, + 0.0043112533167004585, + 0.0034164278768002987, + 0.006392528768628836, + 0.007720959838479757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.005950592458248138 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031931623816490173 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.0031931623816490173, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031931623816490173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19842803478240967, + 0.24673089385032654, + 0.07209648936986923, + 0.0735795795917511, + 0.09576104581356049, + 0.06091691181063652, + 0.08821903169155121, + 0.08574490994215012, + 0.022114770486950874, + 0.056408319622278214 + ], + "credit_map": [ + 0.03305912762880325, + 0.012203003279864788, + 0.04756747558712959, + 0.02037372998893261, + 0.015804802998900414, + 0.023299895226955414, + 0.021038858219981194, + 0.025081142783164978, + 0.024404563009738922, + 0.014944834634661674, + 0.0748286098241806, + 0.04867088794708252, + 0.012843770906329155, + 0.016362490132451057, + 0.020445361733436584, + 0.03423038125038147, + 0.006603706162422895, + 0.006713853683322668, + 0.00885173212736845, + 0.012252870947122574, + 0.008400395512580872, + 0.009625082835555077, + 0.011188413947820663, + 0.008460434153676033, + 0.008161214180290699, + 0.008163261227309704, + 0.01171021070331335, + 0.01051757950335741, + 0.007742095738649368, + 0.010017680004239082, + 0.009242584928870201, + 0.008024954237043858, + 0.01167208794504404, + 0.009825625456869602, + 0.011620616540312767, + 0.014900627546012402, + 0.011197571642696857, + 0.011525417678058147, + 0.01409870944917202, + 0.010920400731265545, + 0.009982011280953884, + 0.005226755514740944, + 0.01058506965637207, + 0.00811216700822115, + 0.007073269225656986, + 0.006011242978274822, + 0.006541340611875057, + 0.007385058328509331, + 0.010606935247778893, + 0.009652736596763134, + 0.013376868329942226, + 0.010267317295074463, + 0.009706479497253895, + 0.012312152422964573, + 0.011594944633543491, + 0.010701602324843407, + 0.011222307570278645, + 0.007339237257838249, + 0.018712906166911125, + 0.015056866221129894, + 0.005592494271695614, + 0.00793471559882164, + 0.008834776468575, + 0.011051611974835396, + 0.002580083906650543, + 0.0021769588347524405, + 0.00344693916849792, + 0.0030216185841709375, + 0.0021254809107631445, + 0.0035509217996150255, + 0.0026478322688490152, + 0.0025649364106357098, + 0.007805922068655491, + 0.0038600151892751455, + 0.01551025453954935, + 0.007567345164716244, + 0.0039035805966705084, + 0.004737377632409334, + 0.0050810170359909534, + 0.007942805998027325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0031931623816490173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a vertical, hand-painted image, captured in a medium close-up shot at eye level, an owner's hand, off-screen, slowly and playfully lures a golden Labrador puppy with a bone-shaped biscuit. The puppy's fluffy, warm-gold fur glows softly under the bright, warm, diffuse light. Its large, round, dark eyes gleam with desire, its pupils dilating as the biscuit moves, completely captivated by the treat. Its head sways with precise precision as the light-gold biscuit, delicately held between its fingers, slowly and precisely shifts left and right across the frame. Its pink tongue lolls out, its saliva glistening from the corner of its mouth, its greedy expression both innocent and comical. A warm, softly blurred interior background focuses the attention on this anticipatory interaction. The visual style of the entire picture resembles a classic children's book illustration, with soft lines, saturated colors and a watercolor-like texture. Every detail, from the fluffy feel of the dog's hair to the crispy texture of the biscuits, is vividly depicted, creating an atmosphere that is rich in details, full of childlike fun and warm and healing.", + "seed": 30122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_1/index_621_seed_30122_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03570824294583872, + "vlm_logprob": -0.03570824294583872, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.03570824294583872, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14451318979263306 + ], + "opposite_logprob_tokens": [ + -2.0195131301879883 + ], + "logprob_reward": -0.14451318979263306, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14451318979263306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23207426071166992, + 0.09913475066423416, + 0.06834077835083008, + 0.06802014261484146, + 0.0837276354432106, + 0.06916175782680511, + 0.12209126353263855, + 0.09981874376535416, + 0.11216460168361664, + 0.04546605795621872 + ], + "credit_map": [ + 0.03741692006587982, + 0.013417691923677921, + 0.015018684789538383, + 0.015136852860450745, + 0.028911542147397995, + 0.021811898797750473, + 0.05179765075445175, + 0.04856303334236145, + 0.013673524372279644, + 0.006913239136338234, + 0.01046673208475113, + 0.013210012577474117, + 0.01013063546270132, + 0.011938737705349922, + 0.01541205681860447, + 0.017389817163348198, + 0.007593910675495863, + 0.00493032019585371, + 0.00824617687612772, + 0.009556901641190052, + 0.00977383367717266, + 0.009613477624952793, + 0.010619403794407845, + 0.00800674594938755, + 0.007977301254868507, + 0.005983018781989813, + 0.009089592844247818, + 0.010077600367367268, + 0.0062795160338282585, + 0.007595480419695377, + 0.013555796816945076, + 0.0074618360958993435, + 0.007539425045251846, + 0.007755369413644075, + 0.01430511474609375, + 0.01274046115577221, + 0.00830613449215889, + 0.006789403036236763, + 0.012596337124705315, + 0.013695389032363892, + 0.009255501441657543, + 0.005000504665076733, + 0.009626702405512333, + 0.0125571358948946, + 0.006272331345826387, + 0.00661937752738595, + 0.011232500895857811, + 0.008597702719271183, + 0.013791725970804691, + 0.0100023178383708, + 0.0205683670938015, + 0.013159387744963169, + 0.012275959365069866, + 0.01170746423304081, + 0.018007973209023476, + 0.02257806807756424, + 0.01419229619204998, + 0.01031457632780075, + 0.01543867401778698, + 0.011593193747103214, + 0.01077971514314413, + 0.010874246247112751, + 0.0112609276548028, + 0.0153651162981987, + 0.01147372554987669, + 0.010950802825391293, + 0.01630578190088272, + 0.011595625430345535, + 0.018195847049355507, + 0.015391599386930466, + 0.01615871861577034, + 0.012092494405806065, + 0.004620494786649942, + 0.0036991641391068697, + 0.005457814317196608, + 0.006746511906385422, + 0.005410522688180208, + 0.005447627976536751, + 0.007646011654287577, + 0.006437909789383411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.14451318979263306 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008291144040413201 + ], + "opposite_logprob_tokens": [ + -7.125829219818115 + ], + "logprob_reward": -0.0008291144040413201, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008291144040413201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28618323802948, + 0.08112448453903198, + 0.06284789741039276, + 0.07077423483133316, + 0.079799585044384, + 0.07399136573076248, + 0.11817547678947449, + 0.09096205234527588, + 0.08834905177354813, + 0.04779256507754326 + ], + "credit_map": [ + 0.05183738097548485, + 0.013019421137869358, + 0.021028639748692513, + 0.01924877054989338, + 0.029283206909894943, + 0.02043355628848076, + 0.04786721616983414, + 0.08346506208181381, + 0.013187605887651443, + 0.005476209335029125, + 0.0074892244301736355, + 0.008876164443790913, + 0.008164968341588974, + 0.011703073047101498, + 0.013416764326393604, + 0.0128104779869318, + 0.0062951380386948586, + 0.004363333340734243, + 0.007010837085545063, + 0.0075369300320744514, + 0.01204692479223013, + 0.008473849855363369, + 0.00977126695215702, + 0.007349617779254913, + 0.007067383266985416, + 0.005703303497284651, + 0.012847055681049824, + 0.0073370542377233505, + 0.007210331968963146, + 0.008034620434045792, + 0.011827380396425724, + 0.01074710488319397, + 0.008711353875696659, + 0.007465751841664314, + 0.013924378901720047, + 0.01061295997351408, + 0.006806755438446999, + 0.0074357762932777405, + 0.012333264574408531, + 0.012509345076978207, + 0.007933035492897034, + 0.005349205806851387, + 0.010798193514347076, + 0.01595625840127468, + 0.006098324898630381, + 0.0072015984915196896, + 0.010884225368499756, + 0.009770526550710201, + 0.01355421170592308, + 0.008830239996314049, + 0.016336187720298767, + 0.013118279166519642, + 0.012705275788903236, + 0.013399260118603706, + 0.01669137366116047, + 0.023540645837783813, + 0.012469310313463211, + 0.009318382479250431, + 0.014826563186943531, + 0.010487312451004982, + 0.00989669468253851, + 0.009693974629044533, + 0.010246984660625458, + 0.014022842980921268, + 0.010310763493180275, + 0.007590964436531067, + 0.014424525201320648, + 0.009243066422641277, + 0.01303137093782425, + 0.012851333245635033, + 0.011054165661334991, + 0.009842864237725735, + 0.004181730095297098, + 0.005055616609752178, + 0.004017027094960213, + 0.008772878907620907, + 0.007079745177179575, + 0.006407800130546093, + 0.006828721147030592, + 0.00544904638081789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0008291144040413201 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020593050867319107 + ], + "opposite_logprob_tokens": [ + -3.8955931663513184 + ], + "logprob_reward": -0.020593050867319107, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020593050867319107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1450657695531845, + 0.1013125330209732, + 0.10409078001976013, + 0.07583271712064743, + 0.10436228662729263, + 0.1043248325586319, + 0.14111551642417908, + 0.06719311326742172, + 0.10863364487886429, + 0.04806872084736824 + ], + "credit_map": [ + 0.019601857289671898, + 0.009830931201577187, + 0.023062538355588913, + 0.012246433645486832, + 0.014441321603953838, + 0.012205991894006729, + 0.028441717848181725, + 0.025234999135136604, + 0.013284902088344097, + 0.007104079704731703, + 0.017766853794455528, + 0.019164780154824257, + 0.00916006788611412, + 0.009433410130441189, + 0.015103242360055447, + 0.010295207612216473, + 0.014190415851771832, + 0.006552749313414097, + 0.018702950328588486, + 0.017894355580210686, + 0.009738934226334095, + 0.009675403125584126, + 0.01235663890838623, + 0.014979345723986626, + 0.008159308694303036, + 0.00529378792271018, + 0.015567787922918797, + 0.009437738917768002, + 0.00717322388663888, + 0.008011998608708382, + 0.010072809644043446, + 0.01211607363075018, + 0.009348511695861816, + 0.009200581349432468, + 0.02182403765618801, + 0.01624738983809948, + 0.009150130674242973, + 0.009909226559102535, + 0.01353201363235712, + 0.015150405466556549, + 0.010468288324773312, + 0.00814513023942709, + 0.01634102873504162, + 0.024031518027186394, + 0.008882599882781506, + 0.011086119338870049, + 0.01348321232944727, + 0.011886942200362682, + 0.01426253467798233, + 0.008926969952881336, + 0.030480803921818733, + 0.017283326014876366, + 0.010296549648046494, + 0.01837785914540291, + 0.017787596210837364, + 0.023699888959527016, + 0.010477351024746895, + 0.006018599029630423, + 0.011375761590898037, + 0.007912627421319485, + 0.009326628409326077, + 0.00810407567769289, + 0.007184031419456005, + 0.006794045213609934, + 0.01359971147030592, + 0.009976925328373909, + 0.018565097823739052, + 0.01332853827625513, + 0.013374226167798042, + 0.015252755023539066, + 0.011444318108260632, + 0.013092076405882835, + 0.0052882228046655655, + 0.0036609575618058443, + 0.010578298941254616, + 0.004982606507837772, + 0.004545620642602444, + 0.006739869248121977, + 0.005667494609951973, + 0.006605654489248991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.020593050867319107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978624727576971 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.00978624727576971, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978624727576971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15174654126167297, + 0.11050791293382645, + 0.10818292200565338, + 0.08616766333580017, + 0.05285271629691124, + 0.05938814580440521, + 0.12420292943716049, + 0.13367140293121338, + 0.08215851336717606, + 0.09112123399972916 + ], + "credit_map": [ + 0.016875142231583595, + 0.012965750880539417, + 0.01594538241624832, + 0.01808135211467743, + 0.02524542808532715, + 0.014990400522947311, + 0.02701425738632679, + 0.020628832280635834, + 0.013361205346882343, + 0.009122645482420921, + 0.014676529914140701, + 0.01451799925416708, + 0.012817911803722382, + 0.010286909528076649, + 0.021538712084293365, + 0.014185997657477856, + 0.013485603034496307, + 0.0071725063025951385, + 0.010910006240010262, + 0.01549495104700327, + 0.014786450192332268, + 0.013907013460993767, + 0.01751214638352394, + 0.014914239756762981, + 0.011277243494987488, + 0.008209684863686562, + 0.013852686621248722, + 0.010759606026113033, + 0.0075189294293522835, + 0.007936101406812668, + 0.011751163750886917, + 0.014862245880067348, + 0.005814609583467245, + 0.004779302515089512, + 0.007416343782097101, + 0.008891098201274872, + 0.007303329650312662, + 0.004178683273494244, + 0.007791989948600531, + 0.006677358411252499, + 0.0064353253692388535, + 0.0049315388314425945, + 0.005298336502164602, + 0.010272464714944363, + 0.007582452613860369, + 0.006793714128434658, + 0.01027283351868391, + 0.007801479659974575, + 0.012323572300374508, + 0.007593621965497732, + 0.02124212682247162, + 0.015657620504498482, + 0.012019832618534565, + 0.01679171249270439, + 0.017291031777858734, + 0.021283410489559174, + 0.017626993358135223, + 0.009562955237925053, + 0.01143631525337696, + 0.01554829627275467, + 0.020573899149894714, + 0.018950384110212326, + 0.024380741640925407, + 0.015591810457408428, + 0.011874349787831306, + 0.006955849938094616, + 0.021410569548606873, + 0.008064513094723225, + 0.009852694347500801, + 0.007616697810590267, + 0.006831080187112093, + 0.009552756324410439, + 0.00836521852761507, + 0.01035681739449501, + 0.006364044267684221, + 0.012363116256892681, + 0.011361772194504738, + 0.01721968501806259, + 0.01494620181620121, + 0.010144378989934921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.00978624727576971 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028196123894304037 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.0028196123894304037, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028196123894304037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1977720856666565, + 0.11172731220722198, + 0.043861404061317444, + 0.05756637081503868, + 0.04356664419174194, + 0.04310457408428192, + 0.0641399472951889, + 0.05524347722530365, + 0.27061763405799866, + 0.11240062862634659 + ], + "credit_map": [ + 0.03539431840181351, + 0.008185167796909809, + 0.012032447382807732, + 0.01041818130761385, + 0.01395986508578062, + 0.01359852310270071, + 0.03471582010388374, + 0.06946776807308197, + 0.01130764838308096, + 0.007616163231432438, + 0.017846444621682167, + 0.019117167219519615, + 0.009009144268929958, + 0.012982946820557117, + 0.016715675592422485, + 0.017132120206952095, + 0.005897361785173416, + 0.0027718418277800083, + 0.00799211859703064, + 0.004539357032626867, + 0.006316394545137882, + 0.0039712125435471535, + 0.005358693189918995, + 0.007014425005763769, + 0.007335176691412926, + 0.004545851610600948, + 0.008543203584849834, + 0.007941357791423798, + 0.0055696237832307816, + 0.005841424688696861, + 0.010602270253002644, + 0.007187461014837027, + 0.00493924506008625, + 0.0048358505591750145, + 0.0069905491545796394, + 0.005689759738743305, + 0.004715743940323591, + 0.0039484878070652485, + 0.005778857506811619, + 0.006668152287602425, + 0.005106612108647823, + 0.004055878147482872, + 0.006425343919545412, + 0.005733540281653404, + 0.004993460141122341, + 0.004270222503691912, + 0.006789819803088903, + 0.005729695316404104, + 0.008005499839782715, + 0.005480530206114054, + 0.010186564177274704, + 0.008716740645468235, + 0.007971150800585747, + 0.005589943844825029, + 0.008916368708014488, + 0.009273149073123932, + 0.00840295385569334, + 0.004946283530443907, + 0.007764851674437523, + 0.00589752895757556, + 0.007199712563306093, + 0.006001845933496952, + 0.008510797284543514, + 0.00651950016617775, + 0.0324227437376976, + 0.013794275000691414, + 0.09681127220392227, + 0.026142168790102005, + 0.018325291574001312, + 0.01884935237467289, + 0.020495213568210602, + 0.04377730190753937, + 0.014202329330146313, + 0.008307136595249176, + 0.030878210440278053, + 0.012502439320087433, + 0.01009149570018053, + 0.01052856631577015, + 0.010973417200148106, + 0.01491702999919653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0028196123894304037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a vertical, hand-painted image, captured in a medium close-up shot at eye level, an owner's hand, off-screen, slowly and playfully lures a golden Labrador puppy with a bone-shaped biscuit. The puppy's fluffy, warm-gold fur glows softly under the bright, warm, diffuse light. Its large, round, dark eyes gleam with desire, its pupils dilating as the biscuit moves, completely captivated by the treat. Its head sways with precise precision as the light-gold biscuit, delicately held between its fingers, slowly and precisely shifts left and right across the frame. Its pink tongue lolls out, its saliva glistening from the corner of its mouth, its greedy expression both innocent and comical. A warm, softly blurred interior background focuses the attention on this anticipatory interaction. The visual style of the entire picture resembles a classic children's book illustration, with soft lines, saturated colors and a watercolor-like texture. Every detail, from the fluffy feel of the dog's hair to the crispy texture of the biscuits, is vividly depicted, creating an atmosphere that is rich in details, full of childlike fun and warm and healing.", + "seed": 30123, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_1/index_621_seed_30123_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012843234033789485, + "vlm_logprob": -0.012843234033789485, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.012843234033789485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02921323850750923 + ], + "opposite_logprob_tokens": [ + -3.6542131900787354 + ], + "logprob_reward": -0.02921323850750923, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02921323850750923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29634860157966614, + 0.10569334775209427, + 0.10140068829059601, + 0.04894904047250748, + 0.08423591405153275, + 0.10117921233177185, + 0.0451803095638752, + 0.08369225263595581, + 0.10237492620944977, + 0.030945751816034317 + ], + "credit_map": [ + 0.0379367433488369, + 0.022243298590183258, + 0.040624137967824936, + 0.042173851281404495, + 0.04506227374076843, + 0.034457702189683914, + 0.03745303675532341, + 0.0363975465297699, + 0.015971871092915535, + 0.007745963986963034, + 0.01603407971560955, + 0.017094042152166367, + 0.012399963103234768, + 0.009047883562743664, + 0.01355094276368618, + 0.013848599046468735, + 0.01259576715528965, + 0.0076476577669382095, + 0.013182583265006542, + 0.011740507557988167, + 0.013016844168305397, + 0.010419932194054127, + 0.015461335889995098, + 0.01733606494963169, + 0.005631113890558481, + 0.0029976465739309788, + 0.008053161203861237, + 0.006013971753418446, + 0.006690850481390953, + 0.006771547254174948, + 0.004796246532350779, + 0.00799450371414423, + 0.014406259171664715, + 0.00533271674066782, + 0.01934254728257656, + 0.012754427269101143, + 0.007249444257467985, + 0.007024960592389107, + 0.008865167386829853, + 0.009260392747819424, + 0.013977482914924622, + 0.008576503954827785, + 0.017835410311818123, + 0.010520955547690392, + 0.011243925429880619, + 0.011526152491569519, + 0.010138062760233879, + 0.017360717058181763, + 0.006122052203863859, + 0.003743367735296488, + 0.006630515679717064, + 0.004525864031165838, + 0.00590938376262784, + 0.00503965187817812, + 0.006318436004221439, + 0.006891040131449699, + 0.012217645533382893, + 0.0056278398260474205, + 0.014531020075082779, + 0.007525705266743898, + 0.009380490519106388, + 0.012322146445512772, + 0.00961765181273222, + 0.012469757348299026, + 0.011444971896708012, + 0.00844933558255434, + 0.016441550105810165, + 0.020047996193170547, + 0.010802540928125381, + 0.01202433742582798, + 0.01080942340195179, + 0.012354766950011253, + 0.003830373752862215, + 0.0024352215696126223, + 0.0037256258074194193, + 0.002868108218535781, + 0.004344839137047529, + 0.003370273159816861, + 0.004192539025098085, + 0.006178771145641804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.02921323850750923 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005106813623569906 + ], + "opposite_logprob_tokens": [ + -7.6255106925964355 + ], + "logprob_reward": -0.0005106813623569906, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0005106813623569906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16921426355838776, + 0.09936924278736115, + 0.07130790501832962, + 0.06479459255933762, + 0.05852788686752319, + 0.05938171595335007, + 0.07534079253673553, + 0.09570761770009995, + 0.2533181309700012, + 0.053037866950035095 + ], + "credit_map": [ + 0.0271752942353487, + 0.012617780826985836, + 0.022926252335309982, + 0.032096389681100845, + 0.014473964460194111, + 0.02243109419941902, + 0.019882213324308395, + 0.017611289396882057, + 0.017322400584816933, + 0.007068904582411051, + 0.015455486252903938, + 0.01806444674730301, + 0.008924555033445358, + 0.009622403420507908, + 0.011310623027384281, + 0.011600421741604805, + 0.009323335252702236, + 0.005199091043323278, + 0.0076047866605222225, + 0.009815778583288193, + 0.00786343403160572, + 0.008175544440746307, + 0.00666276877745986, + 0.016663167625665665, + 0.007303805556148291, + 0.004383009858429432, + 0.012606605887413025, + 0.005425211042165756, + 0.00687955878674984, + 0.008464324288070202, + 0.005538035184144974, + 0.014194034971296787, + 0.009079117327928543, + 0.004782976116985083, + 0.012886427342891693, + 0.008672846481204033, + 0.0054420181550085545, + 0.004863616544753313, + 0.005569612141698599, + 0.007231266237795353, + 0.007930053398013115, + 0.004672677256166935, + 0.013571861200034618, + 0.007038434501737356, + 0.005553795024752617, + 0.00556548684835434, + 0.005534554366022348, + 0.009514855220913887, + 0.009632954373955727, + 0.0076531111262738705, + 0.007372591178864241, + 0.005328075494617224, + 0.011667224578559399, + 0.010358767583966255, + 0.010660612024366856, + 0.01266746036708355, + 0.012002767994999886, + 0.0074982792139053345, + 0.008691041730344296, + 0.006095761898905039, + 0.012320403009653091, + 0.012930902652442455, + 0.014050542376935482, + 0.022117912769317627, + 0.040927544236183167, + 0.023242127150297165, + 0.032144539058208466, + 0.018642304465174675, + 0.03308093547821045, + 0.030127720907330513, + 0.03303973376750946, + 0.04211323335766792, + 0.007778884377330542, + 0.004987986758351326, + 0.005214834585785866, + 0.004291383549571037, + 0.007020364515483379, + 0.005596623290330172, + 0.006521439645439386, + 0.0116263497620821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0005106813623569906 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059258334338665 + ], + "opposite_logprob_tokens": [ + -3.89559268951416 + ], + "logprob_reward": -0.02059258334338665, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059258334338665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11287521570920944, + 0.11399198323488235, + 0.17131410539150238, + 0.09656686335802078, + 0.08136580139398575, + 0.06761486083269119, + 0.08746711164712906, + 0.0640205442905426, + 0.17435672879219055, + 0.030426837503910065 + ], + "credit_map": [ + 0.017506977543234825, + 0.007658855523914099, + 0.018836496397852898, + 0.02507142350077629, + 0.010335967876017094, + 0.010458742268383503, + 0.010912220925092697, + 0.012094534002244473, + 0.020820776000618935, + 0.007256345823407173, + 0.02283325232565403, + 0.01962667517364025, + 0.007769981399178505, + 0.008418078534305096, + 0.01168350875377655, + 0.015583359636366367, + 0.028600342571735382, + 0.01319427601993084, + 0.02569606900215149, + 0.035547658801078796, + 0.015500104054808617, + 0.014061901718378067, + 0.018693290650844574, + 0.020020466297864914, + 0.012926933355629444, + 0.006267067044973373, + 0.017503604292869568, + 0.014137768186628819, + 0.010313276201486588, + 0.009629176929593086, + 0.009965386241674423, + 0.01582365296781063, + 0.01388197485357523, + 0.005076671950519085, + 0.019983628764748573, + 0.011455412022769451, + 0.007265088148415089, + 0.006906630005687475, + 0.008475011214613914, + 0.008321378380060196, + 0.01003890112042427, + 0.005460742861032486, + 0.009557665325701237, + 0.00625062920153141, + 0.007964354008436203, + 0.008425603620707989, + 0.007509015966206789, + 0.01240795012563467, + 0.01027894951403141, + 0.007351109758019447, + 0.01413378119468689, + 0.012596051208674908, + 0.009266600012779236, + 0.0072295949794352055, + 0.012357788160443306, + 0.014253230765461922, + 0.007674690335988998, + 0.004664493724703789, + 0.014761088415980339, + 0.007192154880613089, + 0.005672104191035032, + 0.00751337269321084, + 0.007120081223547459, + 0.009422562085092068, + 0.024924319237470627, + 0.01701236702501774, + 0.018582195043563843, + 0.01729484461247921, + 0.023530947044491768, + 0.0215973611921072, + 0.023358719423413277, + 0.028055990114808083, + 0.004446349106729031, + 0.0023525722790509462, + 0.004430736880749464, + 0.004045186098664999, + 0.004597536753863096, + 0.002843906870111823, + 0.003444207366555929, + 0.0042663440108299255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.02059258334338665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107910368591547 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.01107910368591547, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107910368591547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18089361488819122, + 0.07245153188705444, + 0.10771813243627548, + 0.10742559283971786, + 0.07886169105768204, + 0.08633054047822952, + 0.07440563291311264, + 0.05079619958996773, + 0.1756911724805832, + 0.06542588025331497 + ], + "credit_map": [ + 0.023576442152261734, + 0.012658472172915936, + 0.017472460865974426, + 0.017127269878983498, + 0.03447072580456734, + 0.021495386958122253, + 0.03380078449845314, + 0.020292088389396667, + 0.011755204759538174, + 0.005991281475871801, + 0.010158919729292393, + 0.010573211126029491, + 0.00877293199300766, + 0.008058042265474796, + 0.007738102693110704, + 0.009403839707374573, + 0.014325926080346107, + 0.00860295444726944, + 0.015038973651826382, + 0.013597573153674603, + 0.012566424906253815, + 0.014514824375510216, + 0.01450406201183796, + 0.014567396603524685, + 0.01478684600442648, + 0.006766775157302618, + 0.0211130753159523, + 0.015478434972465038, + 0.012704385444521904, + 0.01116215344518423, + 0.010403983294963837, + 0.015009944327175617, + 0.011531162075698376, + 0.004486915655434132, + 0.01695832423865795, + 0.008039399981498718, + 0.00985593069344759, + 0.006034829653799534, + 0.011652885936200619, + 0.010302242822945118, + 0.012309221550822258, + 0.005113222636282444, + 0.01293835137039423, + 0.010740009136497974, + 0.010607742704451084, + 0.008501010946929455, + 0.010572480969130993, + 0.015548508614301682, + 0.009438523091375828, + 0.006509057711809874, + 0.010966536588966846, + 0.010568048804998398, + 0.00942890252918005, + 0.006876515224575996, + 0.01075801346451044, + 0.009860033169388771, + 0.005872518755495548, + 0.002913104835897684, + 0.009426087141036987, + 0.00824553333222866, + 0.004628474824130535, + 0.006018354091793299, + 0.006638337392359972, + 0.007053791079670191, + 0.02614663541316986, + 0.01828456111252308, + 0.01730457693338394, + 0.01582344062626362, + 0.025736138224601746, + 0.021543877199292183, + 0.02308795601129532, + 0.02776399813592434, + 0.010846303775906563, + 0.00532554043456912, + 0.009888284839689732, + 0.007165956310927868, + 0.010545426048338413, + 0.0063261836767196655, + 0.006900625303387642, + 0.008427564054727554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.01107910368591547 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a vertical, hand-painted image resembling a classic children's book illustration with soft lines and watercolor-like texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a golden Labrador puppy with fluffy, warm-gold fur visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bone-shaped biscuit held by an off-screen hand visible, moving slowly and precisely left and right across the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the puppy exhibit specific behaviors including swaying its head, having dilating pupils, a lolling pink tongue with glistening saliva, and a greedy expression?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot as a medium close-up at eye level with a warm, softly blurred interior background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820563269779086 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.002820563269779086, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820563269779086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2064690887928009, + 0.07218189537525177, + 0.07503895461559296, + 0.08291751891374588, + 0.04863881692290306, + 0.09617619961500168, + 0.08980855345726013, + 0.08302736282348633, + 0.13266010582447052, + 0.11308149993419647 + ], + "credit_map": [ + 0.03404642641544342, + 0.01792476698756218, + 0.032281555235385895, + 0.021914532408118248, + 0.01974177174270153, + 0.020054779946804047, + 0.020988434553146362, + 0.03951682895421982, + 0.013177894987165928, + 0.007010364904999733, + 0.010761402547359467, + 0.010757391341030598, + 0.008614370599389076, + 0.005946480203419924, + 0.00794270634651184, + 0.00797128863632679, + 0.011496349237859249, + 0.011302269995212555, + 0.007716172840446234, + 0.006461175158619881, + 0.009039221331477165, + 0.008093627169728279, + 0.010186375118792057, + 0.01074376329779625, + 0.008875252678990364, + 0.005945185665041208, + 0.008603819645941257, + 0.00723752286285162, + 0.011141190305352211, + 0.011246699839830399, + 0.009324964135885239, + 0.020542887970805168, + 0.006207793485373259, + 0.003881134558469057, + 0.007957767695188522, + 0.007098909001797438, + 0.0049836826510727406, + 0.00566561333835125, + 0.006100549828261137, + 0.006743369624018669, + 0.015751631930470467, + 0.007701560389250517, + 0.010618213564157486, + 0.009900826960802078, + 0.010924444533884525, + 0.010477123782038689, + 0.01158925425261259, + 0.019213145598769188, + 0.012107746675610542, + 0.006688612047582865, + 0.01374247670173645, + 0.012759203091263771, + 0.011044853366911411, + 0.009936898946762085, + 0.012394286692142487, + 0.011134475469589233, + 0.010991912335157394, + 0.006462486926466227, + 0.016162212938070297, + 0.007596369367092848, + 0.009769100695848465, + 0.011033810675144196, + 0.008861423470079899, + 0.01215003989636898, + 0.02015542984008789, + 0.011148212477564812, + 0.028187869116663933, + 0.014890201389789581, + 0.012533030472695827, + 0.016274383291602135, + 0.014033980667591095, + 0.015437017194926739, + 0.019787294790148735, + 0.009365159086883068, + 0.015945063903927803, + 0.009243759326636791, + 0.01846090704202652, + 0.011042722500860691, + 0.011737614870071411, + 0.01749897375702858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.002820563269779086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_10/index_tensor([7127, 7127])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_10/index_tensor([7127, 7127])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dec43488c5c60c99027f8facc4fcdfcfeb2f0f5e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_10/index_tensor([7127, 7127])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static medium shot captures an athletic woman, her dark hair pulled back into a tight, high ponytail, her body brimming with potential energy. With explosive force, she swings her right arm forward, propelling her entire body upward from a hinged position. The movement is a single, fluid, and powerful motion; her hips thrust forward decisively, lifting the heavy, dark gray kettlebell to chest height. Her face, lightly beaded with sweat, is a mask of pure concentration, her chin lowered, her eyes staring forward. She wears a charcoal gray sports bra and black, high-waisted performance leggings, accentuating her muscular physique. The background is a functional gym, with scattered weights, a dumbbell rack, and various equipment lightly blurred to create a palpable sense of depth. The scene is illuminated by cool, overhead fluorescent lights, casting crisp, well-defined shadows, yet the entire frame takes on a cinematic, slightly grainy look, infusing this moment of raw, focused effort with texture, grit, and an aura of intense power.", + "seed": 12251, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_10/index_7127_seed_12251_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.29016395248472693, + "vlm_logprob": -0.29016395248472693, + "vlm_yes_no_margin": 2.175000011920929, + "vlm_token_credit": -0.29016395248472693, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4599990248680115 + ], + "opposite_logprob_tokens": [ + -1.2099990844726562 + ], + "logprob_reward": -0.4599990248680115, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.4599990248680115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10197760164737701, + 0.08636365085840225, + 0.054877761751413345, + 0.22533003985881805, + 0.3272339999675751, + 0.06188543885946274, + 0.031209168955683708, + 0.024654189124703407, + 0.03367910534143448, + 0.052789051085710526 + ], + "credit_map": [ + 0.012134287506341934, + 0.011457497254014015, + 0.01700654812157154, + 0.011931695975363255, + 0.01321602612733841, + 0.009659422561526299, + 0.015100701712071896, + 0.01147142332047224, + 0.008972818031907082, + 0.007464472204446793, + 0.018357589840888977, + 0.008937454782426357, + 0.009717087261378765, + 0.010458775795996189, + 0.007664875593036413, + 0.014790582470595837, + 0.00403764471411705, + 0.006316422950476408, + 0.008220764808356762, + 0.0074835484847426414, + 0.0066261314786970615, + 0.006569953635334969, + 0.008823312819004059, + 0.006799985188990831, + 0.01048702746629715, + 0.03531139716506004, + 0.021458638831973076, + 0.055202506482601166, + 0.02620813623070717, + 0.022173212841153145, + 0.035386454313993454, + 0.019102683290839195, + 0.023782478645443916, + 0.06364092230796814, + 0.027744917199015617, + 0.035907402634620667, + 0.07072064280509949, + 0.044500842690467834, + 0.032500363886356354, + 0.028436407446861267, + 0.004216983914375305, + 0.007862203754484653, + 0.007873354479670525, + 0.007427845615893602, + 0.01010496262460947, + 0.007064332719892263, + 0.010651064105331898, + 0.006684692110866308, + 0.003067339537665248, + 0.0032425804529339075, + 0.005266503430902958, + 0.0052979327738285065, + 0.0041902996599674225, + 0.0033389106392860413, + 0.0035131594631820917, + 0.003292444860562682, + 0.001973326550796628, + 0.0025271312333643436, + 0.006751941982656717, + 0.0025855053681880236, + 0.00249504204839468, + 0.0032684272155165672, + 0.0022668212186545134, + 0.0027859953697770834, + 0.002047824440523982, + 0.004750488325953484, + 0.004690583795309067, + 0.003194432705640793, + 0.007067743688821793, + 0.004159270785748959, + 0.004649647977203131, + 0.00311911478638649, + 0.00269376952201128, + 0.0033778881188482046, + 0.003953992854803801, + 0.0049454215914011, + 0.005466192960739136, + 0.017967909574508667, + 0.004499155096709728, + 0.009884720668196678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.4599990248680115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983891606330872 + ], + "opposite_logprob_tokens": [ + -0.5233891606330872 + ], + "logprob_reward": -0.8983891606330872, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983891606330872, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08053918182849884, + 0.061911579221487045, + 0.06365728378295898, + 0.06716331094503403, + 0.32663053274154663, + 0.25974294543266296, + 0.041412871330976486, + 0.023207765072584152, + 0.051366373896598816, + 0.024368129670619965 + ], + "credit_map": [ + 0.009116047993302345, + 0.011668632738292217, + 0.01429078634828329, + 0.011823417618870735, + 0.008089200593531132, + 0.008261105045676231, + 0.009197697974741459, + 0.008092294447124004, + 0.005079320166260004, + 0.009021066129207611, + 0.006834651343524456, + 0.009088154882192612, + 0.008890029042959213, + 0.006551194470375776, + 0.00931521225720644, + 0.007131956052035093, + 0.005627736449241638, + 0.008617921732366085, + 0.010198417119681835, + 0.00972351711243391, + 0.006512015126645565, + 0.006160741671919823, + 0.008603150956332684, + 0.008213781751692295, + 0.0044366163201630116, + 0.011737005785107613, + 0.00810500793159008, + 0.010679569095373154, + 0.007884211838245392, + 0.006527822930365801, + 0.00969447661191225, + 0.008098608814179897, + 0.01703883707523346, + 0.06272201240062714, + 0.03568144515156746, + 0.03680708259344101, + 0.0691223293542862, + 0.04208086431026459, + 0.03267814591526985, + 0.03049984946846962, + 0.015381112694740295, + 0.031167233362793922, + 0.03481395170092583, + 0.0331893116235733, + 0.038143597543239594, + 0.03463408350944519, + 0.040774255990982056, + 0.03163941577076912, + 0.0026850637514144182, + 0.004481507930904627, + 0.006175860296934843, + 0.009420198388397694, + 0.0046310825273394585, + 0.007166610099375248, + 0.003426130162551999, + 0.0034264156129211187, + 0.0016440863255411386, + 0.004061080049723387, + 0.002394347684457898, + 0.003789510577917099, + 0.004243260715156794, + 0.002616587094962597, + 0.0024433752987533808, + 0.0020155159290879965, + 0.0034743810538202524, + 0.008303330279886723, + 0.005374026019126177, + 0.005067802034318447, + 0.010821336880326271, + 0.0065133534371852875, + 0.007462490350008011, + 0.004349658731371164, + 0.002469967119395733, + 0.0031655204948037863, + 0.0029584476724267006, + 0.0045616962015628815, + 0.0027589979581534863, + 0.0027729200664907694, + 0.0036423574201762676, + 0.0020382232032716274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.8983891606330872 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813757747411728 + ], + "opposite_logprob_tokens": [ + -3.288137674331665 + ], + "logprob_reward": -0.03813757747411728, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813757747411728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11499249190092087, + 0.07129783183336258, + 0.047568827867507935, + 0.15960703790187836, + 0.3138962686061859, + 0.08928252756595612, + 0.04716634750366211, + 0.0633501335978508, + 0.057124290615320206, + 0.03571416437625885 + ], + "credit_map": [ + 0.009111997671425343, + 0.013575190678238869, + 0.012412217445671558, + 0.01560988649725914, + 0.01705196313560009, + 0.01406280416995287, + 0.015023334883153439, + 0.018145110458135605, + 0.0056228116154670715, + 0.00822001975029707, + 0.008076868019998074, + 0.008622152730822563, + 0.015486589632928371, + 0.00839541107416153, + 0.009304795414209366, + 0.007569191046059132, + 0.003161034546792507, + 0.006283084861934185, + 0.005907383747398853, + 0.008043697103857994, + 0.006039318162947893, + 0.004877475090324879, + 0.006388718727976084, + 0.006868119817227125, + 0.008191260509192944, + 0.023405924439430237, + 0.020060310140252113, + 0.03573281690478325, + 0.01977294124662876, + 0.014803501777350903, + 0.022254090756177902, + 0.015386207960546017, + 0.018682202324271202, + 0.055916719138622284, + 0.036154232919216156, + 0.035581957548856735, + 0.06768760085105896, + 0.03710664436221123, + 0.03497060388326645, + 0.027796339243650436, + 0.00638180784881115, + 0.011423462070524693, + 0.010839981026947498, + 0.010476985014975071, + 0.015820544213056564, + 0.010228066705167294, + 0.012803899124264717, + 0.011307794600725174, + 0.00452466681599617, + 0.004076456185430288, + 0.012882660143077374, + 0.0040266066789627075, + 0.0063742888160049915, + 0.006540738046169281, + 0.002980786142870784, + 0.0057601467706263065, + 0.005560620687901974, + 0.005788449198007584, + 0.011094899848103523, + 0.006719384808093309, + 0.0073570371605455875, + 0.011790229938924313, + 0.004567267838865519, + 0.010472248308360577, + 0.003092460799962282, + 0.008656652644276619, + 0.0062682610005140305, + 0.005533762741833925, + 0.013150286860764027, + 0.00701861921697855, + 0.008422776125371456, + 0.004981476813554764, + 0.0036883093416690826, + 0.002476171590387821, + 0.007799041923135519, + 0.004231119528412819, + 0.003117903368547559, + 0.006638368126004934, + 0.003692728467285633, + 0.0040705264545977116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.03813757747411728 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033712275326251984 + ], + "opposite_logprob_tokens": [ + -3.408712387084961 + ], + "logprob_reward": -0.033712275326251984, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033712275326251984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12466199696063995, + 0.09833496809005737, + 0.2146092653274536, + 0.19574156403541565, + 0.04997147247195244, + 0.1687842458486557, + 0.08760672062635422, + 0.017743529751896858, + 0.020294716581702232, + 0.022251464426517487 + ], + "credit_map": [ + 0.008271011523902416, + 0.019343415275216103, + 0.013277183286845684, + 0.015992071479558945, + 0.02665889821946621, + 0.012624036520719528, + 0.012387926690280437, + 0.016107451170682907, + 0.007701452821493149, + 0.010631471872329712, + 0.009487572126090527, + 0.018095647916197777, + 0.012319136410951614, + 0.01081500668078661, + 0.01787377893924713, + 0.011410892941057682, + 0.014807281084358692, + 0.02825501747429371, + 0.020724091678857803, + 0.03606855124235153, + 0.020578645169734955, + 0.020611723884940147, + 0.044054415076971054, + 0.0295095257461071, + 0.00934869796037674, + 0.02959618717432022, + 0.017255041748285294, + 0.05083142966032028, + 0.022972779348492622, + 0.019610416144132614, + 0.029447384178638458, + 0.016679629683494568, + 0.003211306408047676, + 0.008638895116746426, + 0.00620273407548666, + 0.006350490264594555, + 0.007793894968926907, + 0.0062778061255812645, + 0.005962279159575701, + 0.00553406635299325, + 0.00932244025170803, + 0.024987010285258293, + 0.023247554898262024, + 0.02944781258702278, + 0.013046053238213062, + 0.015234683640301228, + 0.02635773830115795, + 0.027140965685248375, + 0.005121314432471991, + 0.010180034674704075, + 0.01391957513988018, + 0.023956559598445892, + 0.007099061738699675, + 0.008433000184595585, + 0.010190923698246479, + 0.008706255815923214, + 0.0013990638544782996, + 0.003034894587472081, + 0.0020824039820581675, + 0.0030606407672166824, + 0.002280724933370948, + 0.001848118961788714, + 0.0016786213964223862, + 0.002359061036258936, + 0.0017143215518444777, + 0.0019334633834660053, + 0.004198961891233921, + 0.0016509325942024589, + 0.0033775819465517998, + 0.0026128932368010283, + 0.002801343100145459, + 0.0020052201580256224, + 0.0010457716416567564, + 0.0022803544998168945, + 0.003013408510014415, + 0.002982092322781682, + 0.002776481444016099, + 0.0039299228228628635, + 0.002186110010370612, + 0.004037322010844946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.033712275326251984 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581724122166634 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581724122166634, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581724122166634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17191825807094574, + 0.06077160686254501, + 0.040663398802280426, + 0.043424151837825775, + 0.07372593879699707, + 0.22653736174106598, + 0.10902907699346542, + 0.06881299614906311, + 0.07939285039901733, + 0.12572439014911652 + ], + "credit_map": [ + 0.018993334844708443, + 0.019631147384643555, + 0.02926611714065075, + 0.017662109807133675, + 0.025355644524097443, + 0.018119489774107933, + 0.01723085157573223, + 0.02565954066812992, + 0.0077089243568480015, + 0.006813291925936937, + 0.010235586203634739, + 0.007757443469017744, + 0.006922964472323656, + 0.006964895874261856, + 0.006219368893653154, + 0.008149124681949615, + 0.0032526860013604164, + 0.005586255807429552, + 0.00634565157815814, + 0.005110163241624832, + 0.005242364015430212, + 0.003804894397035241, + 0.006157180294394493, + 0.005164199508726597, + 0.0038589395117014647, + 0.00772130349650979, + 0.006999125238507986, + 0.006813389714807272, + 0.004661837592720985, + 0.003742640372365713, + 0.004042862448841333, + 0.005584050435572863, + 0.004167554434388876, + 0.0140830148011446, + 0.008242745883762836, + 0.009385533630847931, + 0.011260631494224072, + 0.00940343365073204, + 0.0098322294652462, + 0.007350789848715067, + 0.02348996140062809, + 0.02765941247344017, + 0.026500847190618515, + 0.022966472432017326, + 0.04123689979314804, + 0.028682593256235123, + 0.029539506882429123, + 0.02646164782345295, + 0.009489189833402634, + 0.007958679459989071, + 0.014120813459157944, + 0.012653203681111336, + 0.016386626288294792, + 0.02939736284315586, + 0.007054598536342382, + 0.011968585662543774, + 0.005367637146264315, + 0.008945518173277378, + 0.010177727788686752, + 0.00822168868035078, + 0.00680417288094759, + 0.012113962322473526, + 0.005756767932325602, + 0.01142551377415657, + 0.008951010182499886, + 0.008150960318744183, + 0.014102112501859665, + 0.007279775571078062, + 0.00991599541157484, + 0.015910005196928978, + 0.008006430231034756, + 0.007076554466038942, + 0.019481033086776733, + 0.00958105269819498, + 0.02454817295074463, + 0.012585951946675777, + 0.01363525353372097, + 0.013642393052577972, + 0.012312497943639755, + 0.0199380274862051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.020581724122166634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static medium shot captures an athletic woman, her dark hair pulled back into a tight, high ponytail, her body brimming with potential energy. With explosive force, she swings her right arm forward, propelling her entire body upward from a hinged position. The movement is a single, fluid, and powerful motion; her hips thrust forward decisively, lifting the heavy, dark gray kettlebell to chest height. Her face, lightly beaded with sweat, is a mask of pure concentration, her chin lowered, her eyes staring forward. She wears a charcoal gray sports bra and black, high-waisted performance leggings, accentuating her muscular physique. The background is a functional gym, with scattered weights, a dumbbell rack, and various equipment lightly blurred to create a palpable sense of depth. The scene is illuminated by cool, overhead fluorescent lights, casting crisp, well-defined shadows, yet the entire frame takes on a cinematic, slightly grainy look, infusing this moment of raw, focused effort with texture, grit, and an aura of intense power.", + "seed": 12252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_10/index_7127_seed_12252_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3638841973617673, + "vlm_logprob": -0.3638841973617673, + "vlm_yes_no_margin": 1.9500000119209289, + "vlm_token_credit": -0.3638841973617673, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5313848853111267 + ], + "opposite_logprob_tokens": [ + -1.1563849449157715 + ], + "logprob_reward": -0.5313848853111267, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.5313848853111267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31558430194854736, + 0.05803201347589493, + 0.08785200119018555, + 0.09459447860717773, + 0.08204367011785507, + 0.05796387419104576, + 0.18647311627864838, + 0.04817761108279228, + 0.030342938378453255, + 0.038935985416173935 + ], + "credit_map": [ + 0.017066676169633865, + 0.026062052696943283, + 0.03233850747346878, + 0.10294850170612335, + 0.018017100170254707, + 0.026215704157948494, + 0.05035415291786194, + 0.04258161783218384, + 0.004964169580489397, + 0.0070373560301959515, + 0.006520798429846764, + 0.007001761347055435, + 0.007660665549337864, + 0.008828943595290184, + 0.009990289807319641, + 0.006028033327311277, + 0.009257619269192219, + 0.010376284830272198, + 0.01308570895344019, + 0.012485223822295666, + 0.012460478581488132, + 0.01283341646194458, + 0.009058988653123379, + 0.008294275030493736, + 0.008573013357818127, + 0.006910244002938271, + 0.014924601651728153, + 0.02131682261824608, + 0.008865566924214363, + 0.009179853834211826, + 0.013415021821856499, + 0.011409356258809566, + 0.007492982782423496, + 0.00912106316536665, + 0.007723384536802769, + 0.007637841161340475, + 0.01450442336499691, + 0.00880101416260004, + 0.012006282806396484, + 0.014756673015654087, + 0.005402951966971159, + 0.005190778523683548, + 0.007077869027853012, + 0.007698076777160168, + 0.009997555986046791, + 0.006917435675859451, + 0.0074670882895588875, + 0.008212114684283733, + 0.009901725687086582, + 0.007402812130749226, + 0.06334701925516129, + 0.009961182251572609, + 0.009766955859959126, + 0.06844127178192139, + 0.005901269149035215, + 0.01175086759030819, + 0.004191236104816198, + 0.0029073820915073156, + 0.015610825270414352, + 0.01201032567769289, + 0.003667074954137206, + 0.003444553120061755, + 0.0035439464263617992, + 0.0028022651094943285, + 0.0023269811645150185, + 0.0024197620805352926, + 0.0026520797982811928, + 0.0030099519062787294, + 0.00807764008641243, + 0.003944878466427326, + 0.004213598556816578, + 0.003698046552017331, + 0.003839644370600581, + 0.006695037242025137, + 0.004636943340301514, + 0.003621984040364623, + 0.005616729613393545, + 0.0059753539972007275, + 0.004398905206471682, + 0.004151387605816126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.5313848853111267 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1371136903762817 + ], + "opposite_logprob_tokens": [ + -0.3871137201786041 + ], + "logprob_reward": -1.1371136903762817, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371136903762817, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.23946601152420044, + 0.05502616614103317, + 0.06844280660152435, + 0.053773876279592514, + 0.15177083015441895, + 0.0920705795288086, + 0.0844111293554306, + 0.14377091825008392, + 0.06079695746302605, + 0.05047066882252693 + ], + "credit_map": [ + 0.013296867720782757, + 0.019596299156546593, + 0.028110848739743233, + 0.07380769401788712, + 0.0163930244743824, + 0.02036101743578911, + 0.03430940583348274, + 0.03359084948897362, + 0.007440716493874788, + 0.005841976497322321, + 0.0056984261609613895, + 0.008456520736217499, + 0.005474446807056665, + 0.007234878372400999, + 0.00980297103524208, + 0.005076227243989706, + 0.005872071720659733, + 0.0075421710498631, + 0.010638147592544556, + 0.010387254878878593, + 0.012194566428661346, + 0.010229179635643959, + 0.0059402985498309135, + 0.005639120005071163, + 0.009830247610807419, + 0.0060682171024382114, + 0.007505153305828571, + 0.008742262609302998, + 0.004901960492134094, + 0.005018705502152443, + 0.00476126279681921, + 0.0069460682570934296, + 0.010009446181356907, + 0.015602133236825466, + 0.013669301755726337, + 0.014925998635590076, + 0.024967072531580925, + 0.01685473322868347, + 0.023211151361465454, + 0.03253098949790001, + 0.01875043474137783, + 0.017533501610159874, + 0.015503018163144588, + 0.01031570602208376, + 0.008206082507967949, + 0.007137061562389135, + 0.0068556866608560085, + 0.007769088260829449, + 0.005809969734400511, + 0.01030226331204176, + 0.008548729121685028, + 0.005463165231049061, + 0.005135297775268555, + 0.0220480989664793, + 0.005257839802652597, + 0.021845770999789238, + 0.01293936837464571, + 0.009316662326455116, + 0.05091221630573273, + 0.04107620567083359, + 0.005815188866108656, + 0.008031166158616543, + 0.006614291109144688, + 0.00906581524759531, + 0.005192241165786982, + 0.006876369938254356, + 0.00873631238937378, + 0.0063145882450044155, + 0.00681710010394454, + 0.008104348555207253, + 0.0066373939625918865, + 0.012118600308895111, + 0.007248146925121546, + 0.005207984242588282, + 0.006373049225658178, + 0.005497362464666367, + 0.008146856911480427, + 0.006063961423933506, + 0.006615179590880871, + 0.00531812896952033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -1.1371136903762817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062117744237184525 + ], + "opposite_logprob_tokens": [ + -2.8121178150177 + ], + "logprob_reward": -0.062117744237184525, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062117744237184525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25599920749664307, + 0.09220811724662781, + 0.0828685387969017, + 0.1336902230978012, + 0.09049218147993088, + 0.07430393993854523, + 0.11987818032503128, + 0.04502038285136223, + 0.06615208834409714, + 0.039387114346027374 + ], + "credit_map": [ + 0.014364124275743961, + 0.01532979216426611, + 0.03264110907912254, + 0.08165887743234634, + 0.013943348079919815, + 0.018586527556180954, + 0.03494124487042427, + 0.044534165412187576, + 0.014203451573848724, + 0.012318366207182407, + 0.013945365324616432, + 0.0121992202475667, + 0.0083369305357337, + 0.011169977486133575, + 0.012248845770955086, + 0.0077859582379460335, + 0.008138778619468212, + 0.011517452076077461, + 0.01096657570451498, + 0.008719502948224545, + 0.01608124189078808, + 0.011220350861549377, + 0.007828597910702229, + 0.008396041579544544, + 0.016918372362852097, + 0.010534143075346947, + 0.01969825103878975, + 0.027378449216485023, + 0.013419200666248798, + 0.012897158972918987, + 0.014773202128708363, + 0.018071450293064117, + 0.006883952766656876, + 0.0071498556062579155, + 0.008638341911137104, + 0.008478307165205479, + 0.015601146034896374, + 0.010750011540949345, + 0.014313810504972935, + 0.0186767615377903, + 0.005168163683265448, + 0.004439684096723795, + 0.006874662823975086, + 0.00846942514181137, + 0.013600626029074192, + 0.008619317784905434, + 0.014625857584178448, + 0.012506205588579178, + 0.007235153578221798, + 0.005928056314587593, + 0.04501238465309143, + 0.008075541816651821, + 0.00666390173137188, + 0.03735636919736862, + 0.00459522008895874, + 0.005011558998376131, + 0.008512458764016628, + 0.005916399881243706, + 0.011165346950292587, + 0.006145142018795013, + 0.0030863997526466846, + 0.003307002130895853, + 0.003287276951596141, + 0.003600355703383684, + 0.005061900708824396, + 0.005231838673353195, + 0.007941246964037418, + 0.0076125310733914375, + 0.012784230522811413, + 0.005575822666287422, + 0.009188017807900906, + 0.01275650318711996, + 0.0028686493169516325, + 0.004664691165089607, + 0.005605058744549751, + 0.00516605656594038, + 0.006463022902607918, + 0.004931837785989046, + 0.005045157857239246, + 0.004642638843506575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.062117744237184525 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901664823293686 + ], + "opposite_logprob_tokens": [ + -2.57901668548584 + ], + "logprob_reward": -0.07901664823293686, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901664823293686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2545326054096222, + 0.07953955978155136, + 0.0772387906908989, + 0.036981940269470215, + 0.13737764954566956, + 0.06473180651664734, + 0.18877747654914856, + 0.05377617105841637, + 0.054617591202259064, + 0.052426356822252274 + ], + "credit_map": [ + 0.016350209712982178, + 0.024814261123538017, + 0.03633620962500572, + 0.0664147362112999, + 0.022747976705431938, + 0.018230905756354332, + 0.027437934651970863, + 0.04220038652420044, + 0.006948343012481928, + 0.010912779718637466, + 0.011213450692594051, + 0.009657145477831364, + 0.009479010477662086, + 0.011355402879416943, + 0.009369730949401855, + 0.010603692382574081, + 0.008197593502700329, + 0.007772450335323811, + 0.010349689051508904, + 0.007919651456177235, + 0.017205743119120598, + 0.01247469149529934, + 0.006253431551158428, + 0.007065542973577976, + 0.002430426422506571, + 0.0018162901978939772, + 0.006653018295764923, + 0.009178074076771736, + 0.003947098273783922, + 0.003382174065336585, + 0.00499378377571702, + 0.004581073764711618, + 0.00808733981102705, + 0.012749707326292992, + 0.01195315457880497, + 0.015607482753694057, + 0.0273776613175869, + 0.015717821195721626, + 0.025843927636742592, + 0.02004055678844452, + 0.005747291259467602, + 0.004652401432394981, + 0.008009344339370728, + 0.007104335352778435, + 0.011408126913011074, + 0.006021192297339439, + 0.012702961452305317, + 0.009086154401302338, + 0.0157302375882864, + 0.019063983112573624, + 0.048391591757535934, + 0.01675649918615818, + 0.0146036921069026, + 0.03171071410179138, + 0.013141469098627567, + 0.029379278421401978, + 0.0030087202321738005, + 0.0028417715802788734, + 0.018281158059835434, + 0.014092311263084412, + 0.0035516321659088135, + 0.0036083341110497713, + 0.004144648555666208, + 0.004247595556080341, + 0.005324196070432663, + 0.005240684375166893, + 0.007346689701080322, + 0.004949024412781, + 0.011227150447666645, + 0.007964127697050571, + 0.006876225583255291, + 0.005689491517841816, + 0.004924717824906111, + 0.004020174499601126, + 0.004201272502541542, + 0.008637473918497562, + 0.009519649669528008, + 0.00552326999604702, + 0.008988607674837112, + 0.006611193064600229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.07901664823293686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978801865130663 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.00978801865130663, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978801865130663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17977501451969147, + 0.08823782205581665, + 0.05571090802550316, + 0.05375358462333679, + 0.10818292945623398, + 0.19196534156799316, + 0.09237510710954666, + 0.09294486790895462, + 0.046964216977357864, + 0.09009015560150146 + ], + "credit_map": [ + 0.01375965029001236, + 0.012410237453877926, + 0.022955963388085365, + 0.024139203131198883, + 0.019681885838508606, + 0.014729959890246391, + 0.04459431767463684, + 0.027503803372383118, + 0.009617882780730724, + 0.013616588898003101, + 0.012894626706838608, + 0.010694803670048714, + 0.009111125953495502, + 0.01009819284081459, + 0.011602276004850864, + 0.010602328926324844, + 0.005364710930734873, + 0.0064305453561246395, + 0.008019119501113892, + 0.007512756623327732, + 0.005263354629278183, + 0.009687483310699463, + 0.005953095853328705, + 0.0074798474088311195, + 0.00833460595458746, + 0.0046460991725325584, + 0.010557634755969048, + 0.007137805223464966, + 0.005722002126276493, + 0.0048324065282940865, + 0.00614131148904562, + 0.006381720770150423, + 0.008969190530478954, + 0.017172906547784805, + 0.01101171225309372, + 0.019255634397268295, + 0.013628331944346428, + 0.010093326680362225, + 0.013783207163214684, + 0.01426862645894289, + 0.02926008217036724, + 0.02466060407459736, + 0.033126700669527054, + 0.020720476284623146, + 0.02293948270380497, + 0.02352905459702015, + 0.01564759761095047, + 0.022081349045038223, + 0.006949578877538443, + 0.007249503396451473, + 0.017198292538523674, + 0.004597774241119623, + 0.00623671617358923, + 0.02720857784152031, + 0.004750862717628479, + 0.018183806911110878, + 0.009180554188787937, + 0.0060722436755895615, + 0.02935374341905117, + 0.02337626926600933, + 0.004511396400630474, + 0.008235263638198376, + 0.005107511300593615, + 0.007107892073690891, + 0.00575319305062294, + 0.0035500505473464727, + 0.0059134759940207005, + 0.006079998333007097, + 0.007835038006305695, + 0.00561994593590498, + 0.004937482997775078, + 0.007275036070495844, + 0.007627725601196289, + 0.013668789528310299, + 0.015554003417491913, + 0.011019140481948853, + 0.011635704897344112, + 0.011951160617172718, + 0.009578746743500233, + 0.009054886177182198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.00978801865130663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static medium shot captures an athletic woman, her dark hair pulled back into a tight, high ponytail, her body brimming with potential energy. With explosive force, she swings her right arm forward, propelling her entire body upward from a hinged position. The movement is a single, fluid, and powerful motion; her hips thrust forward decisively, lifting the heavy, dark gray kettlebell to chest height. Her face, lightly beaded with sweat, is a mask of pure concentration, her chin lowered, her eyes staring forward. She wears a charcoal gray sports bra and black, high-waisted performance leggings, accentuating her muscular physique. The background is a functional gym, with scattered weights, a dumbbell rack, and various equipment lightly blurred to create a palpable sense of depth. The scene is illuminated by cool, overhead fluorescent lights, casting crisp, well-defined shadows, yet the entire frame takes on a cinematic, slightly grainy look, infusing this moment of raw, focused effort with texture, grit, and an aura of intense power.", + "seed": 12253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_10/index_7127_seed_12253_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.43647627606987954, + "vlm_logprob": -0.43647627606987954, + "vlm_yes_no_margin": 1.675, + "vlm_token_credit": -0.43647627606987954, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3150720000267029 + ], + "opposite_logprob_tokens": [ + -1.5650720596313477 + ], + "logprob_reward": -0.3150720000267029, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.3150720000267029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08893496543169022, + 0.4000943601131439, + 0.15793757140636444, + 0.06803208589553833, + 0.08314480632543564, + 0.03874271363019943, + 0.046502742916345596, + 0.049946293234825134, + 0.03793490678071976, + 0.02872951701283455 + ], + "credit_map": [ + 0.00664559006690979, + 0.009715060703456402, + 0.01001402921974659, + 0.011762436479330063, + 0.015426336787641048, + 0.013835422694683075, + 0.007785622496157885, + 0.013750473037362099, + 0.06915047764778137, + 0.01781177520751953, + 0.07617774605751038, + 0.011648442596197128, + 0.015749895945191383, + 0.03133533149957657, + 0.01794045977294445, + 0.1602802276611328, + 0.01670985296368599, + 0.01786724478006363, + 0.019016066566109657, + 0.021480439230799675, + 0.010480202734470367, + 0.021415363997220993, + 0.01773190125823021, + 0.03323650360107422, + 0.009847749955952168, + 0.009968049824237823, + 0.0075194560922682285, + 0.006801560055464506, + 0.008827687241137028, + 0.009245750494301319, + 0.007290754932910204, + 0.008531075902283192, + 0.008863059803843498, + 0.01871267706155777, + 0.009373391047120094, + 0.009635892696678638, + 0.007936662994325161, + 0.011471167206764221, + 0.007374999579042196, + 0.009776956401765347, + 0.004437347874045372, + 0.0030816784128546715, + 0.0045371451415121555, + 0.008883106522262096, + 0.0033576481509953737, + 0.004921765066683292, + 0.00552810775116086, + 0.003995916806161404, + 0.005341795738786459, + 0.004658904857933521, + 0.0109549630433321, + 0.006442577112466097, + 0.003429235890507698, + 0.006540916860103607, + 0.004143475089222193, + 0.004990878514945507, + 0.008042637258768082, + 0.006161738187074661, + 0.006035010330379009, + 0.008061137050390244, + 0.0038286736235022545, + 0.008067309856414795, + 0.0049451920203864574, + 0.004804591182619333, + 0.0029462554957717657, + 0.0030207554809749126, + 0.005088320001959801, + 0.004393703304231167, + 0.003930853214114904, + 0.00906310509890318, + 0.0037803668528795242, + 0.00571154709905386, + 0.004489119164645672, + 0.0026898544747382402, + 0.005197939928621054, + 0.0019700455013662577, + 0.003158373525366187, + 0.003066624980419874, + 0.0036962274461984634, + 0.004461330361664295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.3150720000267029 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015610933303833 + ], + "opposite_logprob_tokens": [ + -0.2015610784292221 + ], + "logprob_reward": -1.7015610933303833, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015610933303833, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0940866768360138, + 0.32817307114601135, + 0.06738850474357605, + 0.0557958222925663, + 0.03877725079655647, + 0.04317115992307663, + 0.11889149248600006, + 0.136856809258461, + 0.08135563135147095, + 0.03550361841917038 + ], + "credit_map": [ + 0.007904521189630032, + 0.009122956544160843, + 0.013114331290125847, + 0.01768023893237114, + 0.01661030203104019, + 0.011368797160685062, + 0.007158044725656509, + 0.011127484031021595, + 0.05692450329661369, + 0.019147632643580437, + 0.06263155490159988, + 0.01669221557676792, + 0.013253418728709221, + 0.023006532341241837, + 0.021099653095006943, + 0.11541752517223358, + 0.006451784633100033, + 0.007625516504049301, + 0.007525658700615168, + 0.011260162107646465, + 0.005538192577660084, + 0.005299909971654415, + 0.007590819615870714, + 0.016096459701657295, + 0.007383034098893404, + 0.0070970687083899975, + 0.007577406242489815, + 0.005819177255034447, + 0.007622188422828913, + 0.0076058404520154, + 0.00663401884958148, + 0.006057079881429672, + 0.005891168490052223, + 0.005176260601729155, + 0.007220841944217682, + 0.004416944459080696, + 0.0038270496297627687, + 0.0034737263340502977, + 0.0036993389949202538, + 0.005071915686130524, + 0.005747300572693348, + 0.0036908420734107494, + 0.00404035858809948, + 0.008974535390734673, + 0.004750001709908247, + 0.0042601763270795345, + 0.007208700757473707, + 0.004499238915741444, + 0.022300440818071365, + 0.008537444286048412, + 0.013360784389078617, + 0.009588008746504784, + 0.007253264542669058, + 0.024374453350901604, + 0.007648529019206762, + 0.025828562676906586, + 0.018277592957019806, + 0.015403148718178272, + 0.015498455613851547, + 0.014700910076498985, + 0.011186905205249786, + 0.028622867539525032, + 0.015406413935124874, + 0.017760517075657845, + 0.007876350544393063, + 0.007647169753909111, + 0.01218860037624836, + 0.009955435059964657, + 0.010282053612172604, + 0.01433661300688982, + 0.009676232933998108, + 0.009393169544637203, + 0.0035131981130689383, + 0.0025047266390174627, + 0.007315013557672501, + 0.0019900286570191383, + 0.0019554048776626587, + 0.0038230265490710735, + 0.004232275299727917, + 0.010169941000640392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -1.7015610933303833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05501748248934746 + ], + "opposite_logprob_tokens": [ + -2.9300174713134766 + ], + "logprob_reward": -0.05501748248934746, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05501748248934746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07167482376098633, + 0.6672776341438293, + 0.03361911699175835, + 0.03326921910047531, + 0.03242664784193039, + 0.016410106793045998, + 0.04281378164887428, + 0.03450678288936615, + 0.04733831435441971, + 0.020663512870669365 + ], + "credit_map": [ + 0.0069388519041240215, + 0.007826452143490314, + 0.007323913276195526, + 0.013555489480495453, + 0.010788102634251118, + 0.008786684833467007, + 0.007472386583685875, + 0.008982935920357704, + 0.1242922767996788, + 0.024512751027941704, + 0.13355708122253418, + 0.018490374088287354, + 0.02529628574848175, + 0.04259765148162842, + 0.02813718281686306, + 0.2703940272331238, + 0.004099288489669561, + 0.006772561930119991, + 0.00477195717394352, + 0.004638533107936382, + 0.0037012463435530663, + 0.0020594564266502857, + 0.004219396039843559, + 0.0033566777128726244, + 0.005643817130476236, + 0.004695797339081764, + 0.004087171051651239, + 0.0036829998716712, + 0.003414132632315159, + 0.004503797739744186, + 0.0034359036944806576, + 0.0038056019693613052, + 0.0036057375837117434, + 0.005856740288436413, + 0.0026909862644970417, + 0.003985194489359856, + 0.0033989690709859133, + 0.0036988304927945137, + 0.004135345108807087, + 0.005054844077676535, + 0.002340471837669611, + 0.0016893952852115035, + 0.001871295622549951, + 0.0015325299464166164, + 0.002126841340214014, + 0.0015470680082216859, + 0.0034050627145916224, + 0.0018974423874169588, + 0.0054679857566952705, + 0.004327333066612482, + 0.006615511607378721, + 0.005982641596347094, + 0.003874958725646138, + 0.006548914592713118, + 0.005083373747766018, + 0.0049130613915622234, + 0.003830843837931752, + 0.004385956563055515, + 0.00479305163025856, + 0.006607570219784975, + 0.002393757225945592, + 0.005587354768067598, + 0.003980385605245829, + 0.0029278623405843973, + 0.00401635468006134, + 0.003956693224608898, + 0.00805409625172615, + 0.004686005413532257, + 0.004820236470550299, + 0.010531817562878132, + 0.004373074974864721, + 0.006900034379214048, + 0.0030212795827537775, + 0.0014946522423997521, + 0.0042403689585626125, + 0.0013629214372485876, + 0.0018082383321598172, + 0.0026976359076797962, + 0.0021863547153770924, + 0.0038520616944879293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.05501748248934746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867691919207573 + ], + "opposite_logprob_tokens": [ + -3.0486769676208496 + ], + "logprob_reward": -0.04867691919207573, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867691919207573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22115004062652588, + 0.21345962584018707, + 0.1282428354024887, + 0.032468970865011215, + 0.039158668369054794, + 0.02931819297373295, + 0.14116279780864716, + 0.15937665104866028, + 0.012340027838945389, + 0.023322273045778275 + ], + "credit_map": [ + 0.018989725038409233, + 0.022571846842765808, + 0.02308734506368637, + 0.046434711664915085, + 0.022528303787112236, + 0.04758236184716225, + 0.01863059215247631, + 0.02132513001561165, + 0.027501437813043594, + 0.01958025060594082, + 0.040598656982183456, + 0.016444630920886993, + 0.012320678681135178, + 0.017322072759270668, + 0.02025328204035759, + 0.05943859741091728, + 0.01231722254306078, + 0.016434507444500923, + 0.012823800556361675, + 0.01696941629052162, + 0.027073753997683525, + 0.01341735664755106, + 0.016245441511273384, + 0.012961323373019695, + 0.003995549399405718, + 0.003971101250499487, + 0.004102999810129404, + 0.0034319981932640076, + 0.005288411863148212, + 0.004293781239539385, + 0.003844901919364929, + 0.0035402243956923485, + 0.005090152379125357, + 0.010236457921564579, + 0.00512831611558795, + 0.0045549324713647366, + 0.004321295768022537, + 0.0035776053555309772, + 0.0034927998203784227, + 0.002757103182375431, + 0.002993227681145072, + 0.0025092875584959984, + 0.00298888236284256, + 0.006662496365606785, + 0.0030914293602108955, + 0.003294315654784441, + 0.004620853345841169, + 0.003157696919515729, + 0.02955096773803234, + 0.011600120924413204, + 0.017250876873731613, + 0.011989749036729336, + 0.008901662193238735, + 0.01942175254225731, + 0.011559825390577316, + 0.030887840315699577, + 0.011261015199124813, + 0.01859268918633461, + 0.010471913032233715, + 0.031156623736023903, + 0.024051019921898842, + 0.03994837775826454, + 0.013411462306976318, + 0.01048352848738432, + 0.0014866344863548875, + 0.0011413234751671553, + 0.0022513512521982193, + 0.0014104910660535097, + 0.001434224657714367, + 0.00168693158775568, + 0.0014714660355821252, + 0.0014576043467968702, + 0.0017063402337953448, + 0.0028098998591303825, + 0.0038961274549365044, + 0.0017000576481223106, + 0.0035397277679294348, + 0.002617561025545001, + 0.0032148913014680147, + 0.003837667405605316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.04867691919207573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205388531088829 + ], + "opposite_logprob_tokens": [ + -2.812053918838501 + ], + "logprob_reward": -0.06205388531088829, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205388531088829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21117642521858215, + 0.16991597414016724, + 0.0772082731127739, + 0.05298970267176628, + 0.055123042315244675, + 0.09743107855319977, + 0.04792163148522377, + 0.14600972831249237, + 0.03468640148639679, + 0.10753781348466873 + ], + "credit_map": [ + 0.018356529995799065, + 0.02294563502073288, + 0.03155079483985901, + 0.01372344046831131, + 0.0413341261446476, + 0.03502112999558449, + 0.021468503400683403, + 0.026776248589158058, + 0.01846790872514248, + 0.020134398713707924, + 0.031163280829787254, + 0.01769823022186756, + 0.01187902595847845, + 0.02094886265695095, + 0.021337291225790977, + 0.028286965563893318, + 0.009312543086707592, + 0.008303121663630009, + 0.015140212140977383, + 0.012610691599547863, + 0.006070631556212902, + 0.009756237268447876, + 0.007772344164550304, + 0.00824248231947422, + 0.009917063638567924, + 0.006731627508997917, + 0.006701275706291199, + 0.006567690055817366, + 0.0076840780675411224, + 0.00596710154786706, + 0.004525352269411087, + 0.004895510617643595, + 0.010575574822723866, + 0.008974015712738037, + 0.008614950813353062, + 0.006779986433684826, + 0.004117369651794434, + 0.0052392855286598206, + 0.005442546680569649, + 0.005379308480769396, + 0.015714174136519432, + 0.008100745268166065, + 0.013327536173164845, + 0.01630682311952114, + 0.007186569273471832, + 0.011151697486639023, + 0.01671130210161209, + 0.008932230994105339, + 0.00940865557640791, + 0.004652142059057951, + 0.005681752227246761, + 0.004530141595751047, + 0.0034983621444553137, + 0.009687698446214199, + 0.003981202375143766, + 0.006481674499809742, + 0.01967567205429077, + 0.016182681545615196, + 0.02422489784657955, + 0.011221309192478657, + 0.011891224421560764, + 0.027523351833224297, + 0.01540483720600605, + 0.019885743036866188, + 0.0025606020353734493, + 0.0023661074228584766, + 0.005559543147683144, + 0.0026120245456695557, + 0.002553759841248393, + 0.01073465682566166, + 0.002065913053229451, + 0.006233792752027512, + 0.015100231394171715, + 0.00817342009395361, + 0.020266924053430557, + 0.006541111972182989, + 0.011042664758861065, + 0.011980264447629452, + 0.014035348780453205, + 0.020397834479808807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.06205388531088829 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static medium shot captures an athletic woman, her dark hair pulled back into a tight, high ponytail, her body brimming with potential energy. With explosive force, she swings her right arm forward, propelling her entire body upward from a hinged position. The movement is a single, fluid, and powerful motion; her hips thrust forward decisively, lifting the heavy, dark gray kettlebell to chest height. Her face, lightly beaded with sweat, is a mask of pure concentration, her chin lowered, her eyes staring forward. She wears a charcoal gray sports bra and black, high-waisted performance leggings, accentuating her muscular physique. The background is a functional gym, with scattered weights, a dumbbell rack, and various equipment lightly blurred to create a palpable sense of depth. The scene is illuminated by cool, overhead fluorescent lights, casting crisp, well-defined shadows, yet the entire frame takes on a cinematic, slightly grainy look, infusing this moment of raw, focused effort with texture, grit, and an aura of intense power.", + "seed": 12254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_10/index_7127_seed_12254_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.31997095635160805, + "vlm_logprob": -0.31997095635160805, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.31997095635160805, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5198845863342285 + ], + "opposite_logprob_tokens": [ + -1.1448845863342285 + ], + "logprob_reward": -0.5198845863342285, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.5198845863342285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10639877617359161, + 0.05606484040617943, + 0.055750880390405655, + 0.43591225147247314, + 0.06566665321588516, + 0.0542144812643528, + 0.08345136046409607, + 0.04028107970952988, + 0.046238888055086136, + 0.05602075532078743 + ], + "credit_map": [ + 0.009068110957741737, + 0.01398105826228857, + 0.01224230881780386, + 0.011844352819025517, + 0.01430271752178669, + 0.01673087477684021, + 0.01235982496291399, + 0.01586952805519104, + 0.00426148995757103, + 0.008118723519146442, + 0.006457151845097542, + 0.007963999174535275, + 0.008821378462016582, + 0.006385338492691517, + 0.008244653232395649, + 0.005812101997435093, + 0.004381776787340641, + 0.0035978108644485474, + 0.006460719276219606, + 0.0050893183797597885, + 0.005251907277852297, + 0.012851638719439507, + 0.0062536862678825855, + 0.01186402142047882, + 0.016323938965797424, + 0.02081357128918171, + 0.026758767664432526, + 0.022412380203604698, + 0.022253643721342087, + 0.14880867302417755, + 0.03197634592652321, + 0.14656493067741394, + 0.005825641565024853, + 0.012424679473042488, + 0.007992604747414589, + 0.008069926872849464, + 0.01104897353798151, + 0.005683631170541048, + 0.008041475899517536, + 0.006579720880836248, + 0.007002802565693855, + 0.005209602881222963, + 0.004230030346661806, + 0.010622432455420494, + 0.00902075320482254, + 0.0033892644569277763, + 0.008928956463932991, + 0.005810640286654234, + 0.006467586848884821, + 0.005595251452177763, + 0.01263489294797182, + 0.010898500680923462, + 0.008688196539878845, + 0.019192690029740334, + 0.006909141317009926, + 0.013065099716186523, + 0.005669230129569769, + 0.003495391458272934, + 0.0050976211205124855, + 0.00456276023760438, + 0.004915005527436733, + 0.0031758807599544525, + 0.009001072496175766, + 0.004364117048680782, + 0.005425487644970417, + 0.006729682441800833, + 0.006579936947673559, + 0.004760939162224531, + 0.007209454663097858, + 0.005337177310138941, + 0.00542221125215292, + 0.004773998633027077, + 0.0055195908062160015, + 0.00682140002027154, + 0.004164045210927725, + 0.005018384195864201, + 0.01198583748191595, + 0.011032002046704292, + 0.004522930830717087, + 0.006956565193831921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.5198845863342285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743528962135315 + ], + "opposite_logprob_tokens": [ + -0.4743528962135315 + ], + "logprob_reward": -0.9743528962135315, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743528962135315, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10657918453216553, + 0.08580196648836136, + 0.0439748652279377, + 0.2843860387802124, + 0.06590941548347473, + 0.08273135125637054, + 0.11155880242586136, + 0.07527891546487808, + 0.07815329730510712, + 0.06562622636556625 + ], + "credit_map": [ + 0.014485611580312252, + 0.01032550074160099, + 0.02258315682411194, + 0.012648357078433037, + 0.008711355738341808, + 0.01665981486439705, + 0.00844019278883934, + 0.012725192122161388, + 0.005834443494677544, + 0.013185386545956135, + 0.011581933125853539, + 0.010418497957289219, + 0.017848771065473557, + 0.010605416260659695, + 0.007704812567681074, + 0.008622711524367332, + 0.0047296080738306046, + 0.004941616207361221, + 0.007077651098370552, + 0.005366729572415352, + 0.006049336865544319, + 0.004953125491738319, + 0.004803560674190521, + 0.006053232122212648, + 0.010158775374293327, + 0.01800849847495556, + 0.01741500198841095, + 0.015298395417630672, + 0.015427718870341778, + 0.09415393322706223, + 0.020932471379637718, + 0.09299122542142868, + 0.008647589012980461, + 0.006336697842925787, + 0.009893380105495453, + 0.006973343435674906, + 0.012002656236290932, + 0.006555911619216204, + 0.008246563374996185, + 0.0072532715275883675, + 0.010517907328903675, + 0.00991162657737732, + 0.0039857905358076096, + 0.02069683186709881, + 0.008934758603572845, + 0.0038258417043834925, + 0.014172953553497791, + 0.01068564597517252, + 0.017495468258857727, + 0.00864946749061346, + 0.016066255047917366, + 0.019265485927462578, + 0.008744029328227043, + 0.0164381954818964, + 0.008701768703758717, + 0.01619812846183777, + 0.008400172926485538, + 0.008234536275267601, + 0.009345714002847672, + 0.011277984827756882, + 0.00943378172814846, + 0.00427134009078145, + 0.017373017966747284, + 0.006942360196262598, + 0.008879809640347958, + 0.010130013339221478, + 0.012349432334303856, + 0.006973042152822018, + 0.010790213011205196, + 0.009688387624919415, + 0.010558846406638622, + 0.008783550001680851, + 0.009236814454197884, + 0.007382720243185759, + 0.006467496044933796, + 0.006193113978952169, + 0.008663608692586422, + 0.010842506773769855, + 0.00855050515383482, + 0.0082894591614604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.9743528962135315 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07906324416399002 + ], + "opposite_logprob_tokens": [ + -2.5790631771087646 + ], + "logprob_reward": -0.07906324416399002, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07906324416399002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10173496603965759, + 0.05682584270834923, + 0.05762839317321777, + 0.2118375450372696, + 0.2486506849527359, + 0.09831203520298004, + 0.07424440234899521, + 0.04630659520626068, + 0.05919669568538666, + 0.045262888073921204 + ], + "credit_map": [ + 0.011311122216284275, + 0.006836641579866409, + 0.018055859953165054, + 0.012200592085719109, + 0.007633386179804802, + 0.022200526669621468, + 0.011217750608921051, + 0.01227908581495285, + 0.0070703583769500256, + 0.007678363472223282, + 0.007203175686299801, + 0.007270526606589556, + 0.008605760522186756, + 0.005796241573989391, + 0.00735178729519248, + 0.0058496310375630856, + 0.003844026243314147, + 0.005543623585253954, + 0.0054773567244410515, + 0.006092292722314596, + 0.0076752919703722, + 0.011179942637681961, + 0.007317839656025171, + 0.010498023591935635, + 0.011152364313602448, + 0.010927136987447739, + 0.015288333408534527, + 0.009687157347798347, + 0.011735581792891026, + 0.07159235328435898, + 0.010983472689986229, + 0.0704711377620697, + 0.025935713201761246, + 0.03366908058524132, + 0.040779709815979004, + 0.028919678181409836, + 0.03461715579032898, + 0.026642270386219025, + 0.032201137393713, + 0.025885945186018944, + 0.009633054956793785, + 0.01284591294825077, + 0.0049369605258107185, + 0.030229909345507622, + 0.011815683916211128, + 0.004273144528269768, + 0.016136977821588516, + 0.008440383709967136, + 0.006768661085516214, + 0.006802372168749571, + 0.01144372671842575, + 0.01939847134053707, + 0.004818632733076811, + 0.008322187699377537, + 0.007607008796185255, + 0.00908333994448185, + 0.0060043600387871265, + 0.0046861604787409306, + 0.007569978013634682, + 0.005072294734418392, + 0.005132331047207117, + 0.004612209741026163, + 0.007853873074054718, + 0.005375383887439966, + 0.005842158105224371, + 0.008163700811564922, + 0.005402108654379845, + 0.005897744558751583, + 0.010305523872375488, + 0.006864392664283514, + 0.007301975507289171, + 0.009419090114533901, + 0.004581547807902098, + 0.006147660315036774, + 0.004486380610615015, + 0.007792648859322071, + 0.0035410583950579166, + 0.007474290207028389, + 0.005250949412584305, + 0.005988352466374636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.07906324416399002 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602159202098846 + ], + "opposite_logprob_tokens": [ + -3.895602226257324 + ], + "logprob_reward": -0.020602159202098846, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602159202098846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08516965061426163, + 0.08880302309989929, + 0.07457108050584793, + 0.10296229273080826, + 0.1320028305053711, + 0.03063393570482731, + 0.0629352405667305, + 0.3499142527580261, + 0.03665850684046745, + 0.036349207162857056 + ], + "credit_map": [ + 0.006082864478230476, + 0.007378918118774891, + 0.00864077266305685, + 0.007229031063616276, + 0.0061052460223436356, + 0.02105116844177246, + 0.0071768867783248425, + 0.021504759788513184, + 0.006187623832374811, + 0.011607140302658081, + 0.011999347247183323, + 0.014851506799459457, + 0.01359447930008173, + 0.01033811829984188, + 0.01325027272105217, + 0.006974537391215563, + 0.004097732249647379, + 0.004484385717660189, + 0.005757717881351709, + 0.005994367878884077, + 0.004954014904797077, + 0.022991832345724106, + 0.00585364643484354, + 0.02043737657368183, + 0.004282490815967321, + 0.009112607687711716, + 0.005484443623572588, + 0.007042150944471359, + 0.007555659394711256, + 0.029353681951761246, + 0.010211782529950142, + 0.02991947904229164, + 0.015356088057160378, + 0.020094560459256172, + 0.02177497185766697, + 0.014369563199579716, + 0.018086712807416916, + 0.012403312139213085, + 0.016854004934430122, + 0.013063616119325161, + 0.0032416051253676414, + 0.0035358346067368984, + 0.0010372325778007507, + 0.008721338585019112, + 0.0037952931597828865, + 0.0014606983168050647, + 0.005937820300459862, + 0.00290411408059299, + 0.0044677285477519035, + 0.0056724026799201965, + 0.010227294638752937, + 0.01767859421670437, + 0.004641090519726276, + 0.006069933529943228, + 0.00729278614744544, + 0.006885412614792585, + 0.025960657745599747, + 0.029088180512189865, + 0.04787138104438782, + 0.051395051181316376, + 0.03976215049624443, + 0.02449633926153183, + 0.0854824110865593, + 0.04585806652903557, + 0.0031151850707829, + 0.004217016510665417, + 0.0037815277464687824, + 0.003827942069619894, + 0.006157636642456055, + 0.005170978605747223, + 0.00501300860196352, + 0.005375209264457226, + 0.005293410271406174, + 0.0031290471088141203, + 0.004112929105758667, + 0.003054717555642128, + 0.005547821056097746, + 0.0075103845447301865, + 0.0033398012164980173, + 0.004361097700893879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.020602159202098846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman swing her right arm forward to propel her body upward from a hinged position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a charcoal gray sports bra and black high-waisted performance leggings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video have a cinematic, slightly grainy look with cool overhead fluorescent lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951895844191313 + ], + "opposite_logprob_tokens": [ + -5.130951881408691 + ], + "logprob_reward": -0.005951895844191313, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951895844191313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18812111020088196, + 0.04272199422121048, + 0.04895243048667908, + 0.05127817392349243, + 0.12230365723371506, + 0.11940138787031174, + 0.31354671716690063, + 0.026654046028852463, + 0.0341421514749527, + 0.05287828668951988 + ], + "credit_map": [ + 0.009485306218266487, + 0.010193072259426117, + 0.014784345403313637, + 0.014459216967225075, + 0.009050888009369373, + 0.05850108712911606, + 0.012567083351314068, + 0.05908012390136719, + 0.0038350974209606647, + 0.0050147841684520245, + 0.005121162161231041, + 0.0072595239616930485, + 0.004966622684150934, + 0.0049149710685014725, + 0.0064285327680408955, + 0.0051813009195029736, + 0.005162072367966175, + 0.005418749526143074, + 0.009894502349197865, + 0.005035661160945892, + 0.005747621413320303, + 0.006513851694762707, + 0.00529753603041172, + 0.005882438272237778, + 0.008710560388863087, + 0.00609615258872509, + 0.008375024423003197, + 0.0056371307000517845, + 0.006205213721841574, + 0.004336395300924778, + 0.0059633683413267136, + 0.005954330787062645, + 0.013685179874300957, + 0.017821619287133217, + 0.019394978880882263, + 0.01457077544182539, + 0.015911398455500603, + 0.012214966118335724, + 0.016785182058811188, + 0.011919571086764336, + 0.029549021273851395, + 0.009363955818116665, + 0.01390947587788105, + 0.013244304805994034, + 0.012675954960286617, + 0.014317742548882961, + 0.013523624278604984, + 0.012817316688597202, + 0.027496078982949257, + 0.025580793619155884, + 0.05247388780117035, + 0.08278824388980865, + 0.01941733807325363, + 0.03831108286976814, + 0.028158104047179222, + 0.03932122141122818, + 0.004279191140085459, + 0.002541095018386841, + 0.004088923335075378, + 0.0033661266788840294, + 0.002651092130690813, + 0.0028932339046150446, + 0.00416692066937685, + 0.0026674645487219095, + 0.00400638347491622, + 0.0033192539121955633, + 0.006769182626157999, + 0.00348983914591372, + 0.0038159897085279226, + 0.004406392108649015, + 0.004187851212918758, + 0.004147263243794441, + 0.005198889411985874, + 0.006388458423316479, + 0.004361873026937246, + 0.004774155560880899, + 0.013860359787940979, + 0.007548539433628321, + 0.005585215520113707, + 0.005160798784345388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.005951895844191313 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_11/index_tensor([8211, 8211])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_11/index_tensor([8211, 8211])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..741ce0e6e7d18758f825ac86c4b45d60d709cc17 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_11/index_tensor([8211, 8211])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows a majestic black horse at full speed along a dusty ancient road. The camera moves in sync with the horse, emphasizing its thunderous speed. Atop his horse, a man in ancient costume, long hair tied in a crown, sits as steady as a mountain. His raven black hair is held high by an exquisite white jade crown, the remaining strands dancing wildly behind him along with his wide, deep teal silk robe in the wind. In a slightly yellowish, retro light, the afterglow of the setting sun casts a warm, timeless hue over the entire image, precisely outlining his resolute facial features—his tightly pursed lips, his straight nose, and his piercing eyes fixed intently ahead. His robe is so intricately textured that even the finest cloud patterns embroidered in gold thread on the flowing hem are visible, fluttering violently with every leap of the horse. The horse's muscles bulge, its glossy coat dripping with beads of sweat, gleaming in the golden light. Its hooves churn, powerfully pawing the ground and stirring up a cloud of dust mixed with gravel. The backdrop, a backdrop of rolling barren hills and sparse, dead trees, appears desolate and vast in the elongated shadows, creating an epic sense of urgency and poignancy, a lone figure dashing toward a fate.", + "seed": 17344, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_11/index_8211_seed_17344_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07843452710658312, + "vlm_logprob": -0.07843452710658312, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.07843452710658312, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12386410683393478 + ], + "opposite_logprob_tokens": [ + -2.74886417388916 + ], + "logprob_reward": -0.12386410683393478, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.12386410683393478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15998612344264984, + 0.0442994087934494, + 0.02093423902988434, + 0.11822205036878586, + 0.28253260254859924, + 0.018742982298135757, + 0.03439978137612343, + 0.05456311255693436, + 0.054412078112363815, + 0.21190759539604187 + ], + "credit_map": [ + 0.02634594403207302, + 0.0168197862803936, + 0.011407209560275078, + 0.007506547961384058, + 0.013131653890013695, + 0.03449524566531181, + 0.011472734622657299, + 0.03880700469017029, + 0.006178814452141523, + 0.005846083629876375, + 0.004215525928884745, + 0.0029678354039788246, + 0.0050542233511805534, + 0.010128567926585674, + 0.00500447815284133, + 0.004903883207589388, + 0.002846116665750742, + 0.0023584465961903334, + 0.0021307324059307575, + 0.0016740034334361553, + 0.0018969865050166845, + 0.0045516351237893105, + 0.002564396243542433, + 0.002911922987550497, + 0.04512472078204155, + 0.0053097764030098915, + 0.007614556699991226, + 0.004882537294179201, + 0.00538589246571064, + 0.017025722190737724, + 0.005145986098796129, + 0.027732867747545242, + 0.056139569729566574, + 0.019780876114964485, + 0.026181183755397797, + 0.034762367606163025, + 0.029719946905970573, + 0.05237148329615593, + 0.02008671499788761, + 0.043490465730428696, + 0.0025959026534110308, + 0.00291456189006567, + 0.0024201085325330496, + 0.001734493998810649, + 0.001935219974257052, + 0.0029237885028123856, + 0.0018891522195190191, + 0.0023297553416341543, + 0.007035044953227043, + 0.0029325198847800493, + 0.006078583188354969, + 0.002244858304038644, + 0.0028031293768435717, + 0.005891889333724976, + 0.0023084792774170637, + 0.005105280317366123, + 0.008331375196576118, + 0.006117841694504023, + 0.006616234313696623, + 0.006227116100490093, + 0.007085298653692007, + 0.0088694728910923, + 0.005194841884076595, + 0.006120935082435608, + 0.007602005731314421, + 0.006203983444720507, + 0.00627737445756793, + 0.003250586101785302, + 0.006314306519925594, + 0.013497619889676571, + 0.0038216854445636272, + 0.007444525137543678, + 0.07004470378160477, + 0.006837459746748209, + 0.0071548274718225, + 0.003472318407148123, + 0.0068453033454716206, + 0.08846069127321243, + 0.003953626379370689, + 0.025138666853308678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.12386410683393478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043117180466651917 + ], + "opposite_logprob_tokens": [ + -3.1681172847747803 + ], + "logprob_reward": -0.043117180466651917, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043117180466651917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11805867403745651, + 0.0666959285736084, + 0.05623804032802582, + 0.07232634723186493, + 0.1302315890789032, + 0.03306850418448448, + 0.0929332822561264, + 0.15094974637031555, + 0.09457534551620483, + 0.18492253124713898 + ], + "credit_map": [ + 0.01406295970082283, + 0.017835592851042747, + 0.00974629819393158, + 0.0077911363914608955, + 0.012336457148194313, + 0.025413276627659798, + 0.007692095823585987, + 0.02318086475133896, + 0.009131889790296555, + 0.00949100498110056, + 0.0077603100799024105, + 0.005982865113765001, + 0.006621245294809341, + 0.011491489596664906, + 0.008854212239384651, + 0.007362913805991411, + 0.008330524899065495, + 0.0077353292144834995, + 0.004492348991334438, + 0.004327324219048023, + 0.007347577717155218, + 0.011952277272939682, + 0.005284735467284918, + 0.006767920218408108, + 0.01765727810561657, + 0.0047781397588551044, + 0.005400754977017641, + 0.004349336493760347, + 0.008766986429691315, + 0.00908733531832695, + 0.004587426315993071, + 0.01769908331334591, + 0.02521681971848011, + 0.009787440299987793, + 0.012775260023772717, + 0.015999674797058105, + 0.01227252185344696, + 0.022482970729470253, + 0.009160959161818027, + 0.02253594622015953, + 0.00539218308404088, + 0.004597445018589497, + 0.00491737388074398, + 0.003178076818585396, + 0.003220453392714262, + 0.005552075803279877, + 0.0025409923400729895, + 0.003669903613626957, + 0.0187503881752491, + 0.00686994194984436, + 0.010430766269564629, + 0.005496114958077669, + 0.00583674618974328, + 0.018499046564102173, + 0.007267652079463005, + 0.019782623276114464, + 0.024644184857606888, + 0.015206637792289257, + 0.015818046405911446, + 0.010579091496765614, + 0.016560737043619156, + 0.030705653131008148, + 0.012164830230176449, + 0.025270551443099976, + 0.01287854090332985, + 0.009524487890303135, + 0.01120885368436575, + 0.00809804443269968, + 0.01516764797270298, + 0.01867690682411194, + 0.008650877512991428, + 0.010369986295700073, + 0.07383235543966293, + 0.007039620075374842, + 0.009153628721833229, + 0.00345628522336483, + 0.006626143120229244, + 0.05993138998746872, + 0.004442911595106125, + 0.02044021338224411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.043117180466651917 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157724618911743 + ], + "opposite_logprob_tokens": [ + -1.7015771865844727 + ], + "logprob_reward": -0.20157724618911743, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157724618911743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25641945004463196, + 0.07492675632238388, + 0.03477497026324272, + 0.04502048343420029, + 0.2189621776342392, + 0.017045307904481888, + 0.05801552161574364, + 0.11391455680131912, + 0.05021851137280464, + 0.13070233166217804 + ], + "credit_map": [ + 0.04715438559651375, + 0.022557783871889114, + 0.014127915725111961, + 0.012088841758668423, + 0.025886129587888718, + 0.04673773795366287, + 0.016954330727458, + 0.07091230899095535, + 0.012387514114379883, + 0.006854674778878689, + 0.004928786773234606, + 0.003921427298337221, + 0.007017532829195261, + 0.025429479777812958, + 0.004139664117246866, + 0.010247667320072651, + 0.004881139844655991, + 0.0026783945504575968, + 0.0035790977999567986, + 0.0026288290973752737, + 0.0035551341716200113, + 0.005769351497292519, + 0.005262164864689112, + 0.006420858204364777, + 0.011290004476904869, + 0.0029784501530230045, + 0.004258410073816776, + 0.0025919226463884115, + 0.005264211446046829, + 0.00682104704901576, + 0.0037232469767332077, + 0.008093184791505337, + 0.032497864216566086, + 0.017988339066505432, + 0.021128349006175995, + 0.03169404715299606, + 0.02580287680029869, + 0.036547690629959106, + 0.0217889491468668, + 0.03151404485106468, + 0.0027682806830853224, + 0.0018632652936503291, + 0.0024289407301694155, + 0.001758186612278223, + 0.001995469443500042, + 0.0025617992505431175, + 0.0015932261012494564, + 0.002076139207929373, + 0.011853962205350399, + 0.005883030593395233, + 0.0062821791507303715, + 0.003013294655829668, + 0.005098221357911825, + 0.01017860695719719, + 0.003957059700042009, + 0.011749167926609516, + 0.015809711068868637, + 0.010958232916891575, + 0.019179584458470345, + 0.008028416894376278, + 0.015786772593855858, + 0.018337832763791084, + 0.00893351249396801, + 0.01688048429787159, + 0.007138191256672144, + 0.004970645532011986, + 0.005370388738811016, + 0.0030650305561721325, + 0.005208216607570648, + 0.011846743524074554, + 0.0033052607905119658, + 0.009314031340181828, + 0.035269685089588165, + 0.004592224024236202, + 0.00523004587739706, + 0.002794465981423855, + 0.005030830856412649, + 0.05408526957035065, + 0.0033432457130402327, + 0.0203565564006567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.20157724618911743 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077453382313251 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077453382313251, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077453382313251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2312830686569214, + 0.02384801022708416, + 0.012956438586115837, + 0.07684046030044556, + 0.23414131999015808, + 0.016597742214798927, + 0.010219756513834, + 0.059633564203977585, + 0.040895115584135056, + 0.29358458518981934 + ], + "credit_map": [ + 0.03929782286286354, + 0.02648092806339264, + 0.015489774756133556, + 0.01792067475616932, + 0.01832270622253418, + 0.03552519530057907, + 0.015849439427256584, + 0.06239650771021843, + 0.0037031301762908697, + 0.0031201501842588186, + 0.0021832063794136047, + 0.0022135116159915924, + 0.002605259884148836, + 0.004960567690432072, + 0.0027043933514505625, + 0.0023577881511300802, + 0.0017394113820046186, + 0.0015104911290109158, + 0.002175448229536414, + 0.0012481447774916887, + 0.0011552998330444098, + 0.0020813269075006247, + 0.0015919201541692019, + 0.0014543957076966763, + 0.0270785354077816, + 0.0031314596999436617, + 0.0060287956148386, + 0.0024034101516008377, + 0.0037278563249856234, + 0.013823225162923336, + 0.0023199329152703285, + 0.018327239900827408, + 0.0448012575507164, + 0.01725776307284832, + 0.016020633280277252, + 0.03257530555129051, + 0.026396894827485085, + 0.04211277887225151, + 0.017801815643906593, + 0.03717485070228577, + 0.0030907916370779276, + 0.0013220134424045682, + 0.001414427999407053, + 0.0010894741863012314, + 0.0018841858254745603, + 0.0031880130991339684, + 0.002281484892591834, + 0.00232735020108521, + 0.0021333969198167324, + 0.0009350780746899545, + 0.0011871318565681577, + 0.0006878588465042412, + 0.0010084925452247262, + 0.0017827249830588698, + 0.0006388752954080701, + 0.001846196479164064, + 0.010739076882600784, + 0.008162621408700943, + 0.006780036725103855, + 0.004079013597220182, + 0.007100293412804604, + 0.011063056997954845, + 0.004935824312269688, + 0.006773639470338821, + 0.0049933274276554585, + 0.005009193904697895, + 0.00509097333997488, + 0.004443611018359661, + 0.005461010616272688, + 0.005924444645643234, + 0.004120935685932636, + 0.005851616617292166, + 0.09632561355829239, + 0.008998739533126354, + 0.00886966660618782, + 0.0037252618931233883, + 0.007565523963421583, + 0.12624774873256683, + 0.005340951960533857, + 0.03651105612516403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.011077453382313251 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536648660898209 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536648660898209, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536648660898209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1995391547679901, + 0.11115850508213043, + 0.0946512520313263, + 0.04101148620247841, + 0.2863847315311432, + 0.03203217312693596, + 0.022180858999490738, + 0.08357010036706924, + 0.024889038875699043, + 0.10458271950483322 + ], + "credit_map": [ + 0.04023601487278938, + 0.02448444627225399, + 0.012743431143462658, + 0.014681120403110981, + 0.020925018936395645, + 0.03496885672211647, + 0.013018656522035599, + 0.038481611758470535, + 0.019519556313753128, + 0.008684043772518635, + 0.00681145628914237, + 0.0046438188292086124, + 0.011630477383732796, + 0.040318530052900314, + 0.0063611711375415325, + 0.013189448975026608, + 0.016180992126464844, + 0.00855198409408331, + 0.009739511646330357, + 0.006508930120617151, + 0.007281912490725517, + 0.02434864267706871, + 0.009879295714199543, + 0.01215998362749815, + 0.006344177294522524, + 0.00336030893959105, + 0.004232520237565041, + 0.002109808847308159, + 0.004416255746036768, + 0.008417218923568726, + 0.002850235439836979, + 0.009280961006879807, + 0.057121146470308304, + 0.02132842130959034, + 0.019503505900502205, + 0.03896057978272438, + 0.030911752954125404, + 0.05323660373687744, + 0.020762190222740173, + 0.0445605106651783, + 0.004458299372345209, + 0.0021713990718126297, + 0.003449669573456049, + 0.002099552657455206, + 0.0036938197445124388, + 0.006773673463612795, + 0.005509922746568918, + 0.0038758364971727133, + 0.004527485929429531, + 0.0015403160359710455, + 0.0023197492118924856, + 0.0008707617525942624, + 0.0012710331939160824, + 0.003129748161882162, + 0.0017547145253047347, + 0.006767048966139555, + 0.0125214122235775, + 0.00790010578930378, + 0.010557944886386395, + 0.006972630973905325, + 0.008098093792796135, + 0.012152038514614105, + 0.00934619177132845, + 0.016021687537431717, + 0.003110684920102358, + 0.0025593587197363377, + 0.003168434603139758, + 0.0025588429998606443, + 0.003036783542484045, + 0.004565647337585688, + 0.001987669849768281, + 0.0039016162045300007, + 0.0367841012775898, + 0.0032268334180116653, + 0.0037880439776927233, + 0.0015002276049926877, + 0.0029708787333220243, + 0.04323422163724899, + 0.0019067188259214163, + 0.011171686463057995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.012536648660898209 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a majestic black horse at full speed along a dusty ancient road. The camera moves in sync with the horse, emphasizing its thunderous speed. Atop his horse, a man in ancient costume, long hair tied in a crown, sits as steady as a mountain. His raven black hair is held high by an exquisite white jade crown, the remaining strands dancing wildly behind him along with his wide, deep teal silk robe in the wind. In a slightly yellowish, retro light, the afterglow of the setting sun casts a warm, timeless hue over the entire image, precisely outlining his resolute facial features—his tightly pursed lips, his straight nose, and his piercing eyes fixed intently ahead. His robe is so intricately textured that even the finest cloud patterns embroidered in gold thread on the flowing hem are visible, fluttering violently with every leap of the horse. The horse's muscles bulge, its glossy coat dripping with beads of sweat, gleaming in the golden light. Its hooves churn, powerfully pawing the ground and stirring up a cloud of dust mixed with gravel. The backdrop, a backdrop of rolling barren hills and sparse, dead trees, appears desolate and vast in the elongated shadows, creating an epic sense of urgency and poignancy, a lone figure dashing toward a fate.", + "seed": 17345, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_11/index_8211_seed_17345_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06898717656731605, + "vlm_logprob": -0.06898717656731605, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.06898717656731605, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11220429837703705 + ], + "opposite_logprob_tokens": [ + -2.7372043132781982 + ], + "logprob_reward": -0.11220429837703705, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.11220429837703705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16031618416309357, + 0.10233528167009354, + 0.044663239270448685, + 0.1962241679430008, + 0.06536579132080078, + 0.12566645443439484, + 0.058320093899965286, + 0.06509415060281754, + 0.04750865697860718, + 0.13450606167316437 + ], + "credit_map": [ + 0.04519842565059662, + 0.030971422791481018, + 0.018215090036392212, + 0.017903564497828484, + 0.011877995915710926, + 0.01375353429466486, + 0.010360040701925755, + 0.012036076746881008, + 0.01324229035526514, + 0.015211468562483788, + 0.012610938400030136, + 0.01286546140909195, + 0.00933351181447506, + 0.015232972800731659, + 0.01001980621367693, + 0.013818817213177681, + 0.009561525657773018, + 0.011876349337399006, + 0.006010565906763077, + 0.003117006504908204, + 0.004306051880121231, + 0.004366522189229727, + 0.0023158336989581585, + 0.0031093796715140343, + 0.0668741762638092, + 0.027627117931842804, + 0.05036218464374542, + 0.012389468029141426, + 0.01640324853360653, + 0.008419174700975418, + 0.007711307145655155, + 0.006437463685870171, + 0.016905058175325394, + 0.005418263841420412, + 0.003137837164103985, + 0.005140396300703287, + 0.004767989739775658, + 0.016659684479236603, + 0.005400961264967918, + 0.00793558917939663, + 0.02539282664656639, + 0.014588838443160057, + 0.01566985994577408, + 0.014500590972602367, + 0.017495809122920036, + 0.01451198197901249, + 0.01247984729707241, + 0.011026683263480663, + 0.010240893810987473, + 0.0056404974311590195, + 0.008113018237054348, + 0.006598759908229113, + 0.007517298683524132, + 0.008099395781755447, + 0.00559162488207221, + 0.006518599111586809, + 0.009339266456663609, + 0.008774099871516228, + 0.006222852040082216, + 0.004712376743555069, + 0.0057522328570485115, + 0.01586640067398548, + 0.005650606472045183, + 0.008776308037340641, + 0.006320859305560589, + 0.005865732207894325, + 0.007832896895706654, + 0.004095969256013632, + 0.00566075649112463, + 0.005856189411133528, + 0.006194062530994415, + 0.005682183429598808, + 0.011900101788341999, + 0.03492511808872223, + 0.03795631602406502, + 0.009859778918325901, + 0.009464876726269722, + 0.009563036262989044, + 0.009196583181619644, + 0.01164024043828249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.11220429837703705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043092403560876846 + ], + "opposite_logprob_tokens": [ + -3.1680924892425537 + ], + "logprob_reward": -0.043092403560876846, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043092403560876846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19568762183189392, + 0.07207607477903366, + 0.025994881987571716, + 0.017744319513440132, + 0.07622502744197845, + 0.03631482645869255, + 0.1664542257785797, + 0.1085495725274086, + 0.03712617605924606, + 0.26382723450660706 + ], + "credit_map": [ + 0.05552581325173378, + 0.04119833931326866, + 0.023562315851449966, + 0.01553335040807724, + 0.0145421689376235, + 0.019161175936460495, + 0.012809183448553085, + 0.013355277478694916, + 0.008673242293298244, + 0.012380518019199371, + 0.007975708693265915, + 0.009688083082437515, + 0.006313278339803219, + 0.010239063762128353, + 0.007517074700444937, + 0.009289100766181946, + 0.0034144807141274214, + 0.004357721656560898, + 0.003401244292035699, + 0.004238837398588657, + 0.0038837981410324574, + 0.002725450322031975, + 0.00218393187969923, + 0.0017894168850034475, + 0.003155139274895191, + 0.0024861174169927835, + 0.002673507435247302, + 0.002322488697245717, + 0.0014695521676912904, + 0.0025080470368266106, + 0.00107018300332129, + 0.0020592843648046255, + 0.014204943552613258, + 0.006731582339853048, + 0.00579668628051877, + 0.007341653574258089, + 0.009562465362250805, + 0.015016012825071812, + 0.006148345768451691, + 0.011423339135944843, + 0.005409256089478731, + 0.0033985069021582603, + 0.0038750898092985153, + 0.004265145864337683, + 0.0039027128368616104, + 0.008287536911666393, + 0.0036059750709682703, + 0.0035706008784472942, + 0.03413993865251541, + 0.0160308089107275, + 0.01854325830936432, + 0.01454897690564394, + 0.018955787643790245, + 0.02711845561861992, + 0.016452239826321602, + 0.020664768293499947, + 0.014777974225580692, + 0.011161958798766136, + 0.011088167317211628, + 0.012511368840932846, + 0.015646925196051598, + 0.01823608949780464, + 0.009450217708945274, + 0.01567687839269638, + 0.00503073213621974, + 0.0043737562373280525, + 0.0040019783191382885, + 0.0033461377024650574, + 0.008576150052249432, + 0.004774865694344044, + 0.003151377197355032, + 0.0038711796514689922, + 0.07609483599662781, + 0.01957358419895172, + 0.08476680517196655, + 0.013267043977975845, + 0.016495345160365105, + 0.021218907088041306, + 0.012095767073333263, + 0.02031497284770012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.043092403560876846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037927567958832 + ], + "opposite_logprob_tokens": [ + -1.9103792905807495 + ], + "logprob_reward": -0.16037927567958832, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037927567958832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14719179272651672, + 0.0808187872171402, + 0.061178337782621384, + 0.15702557563781738, + 0.05480187013745308, + 0.12890960276126862, + 0.14305227994918823, + 0.052545804530382156, + 0.09552397578954697, + 0.07895202934741974 + ], + "credit_map": [ + 0.027023201808333397, + 0.03502655774354935, + 0.018490150570869446, + 0.01610611006617546, + 0.010924406349658966, + 0.012617187574505806, + 0.013906083069741726, + 0.013098079711198807, + 0.021009938791394234, + 0.01419703010469675, + 0.012092133052647114, + 0.007397244684398174, + 0.005864454433321953, + 0.007960837334394455, + 0.005272329784929752, + 0.007024814374744892, + 0.0132394228130579, + 0.01604234054684639, + 0.0083714434877038, + 0.004386963788419962, + 0.006036273669451475, + 0.00484564108774066, + 0.003432951169088483, + 0.00482329772785306, + 0.05743701383471489, + 0.020102230831980705, + 0.05102604627609253, + 0.007510210853070021, + 0.0070896632969379425, + 0.005747165996581316, + 0.00437367707490921, + 0.0037395444232970476, + 0.0076262918300926685, + 0.007732773199677467, + 0.007243394386023283, + 0.010320300236344337, + 0.004888246767222881, + 0.006353812757879496, + 0.00540448073297739, + 0.005232563242316246, + 0.025317519903182983, + 0.018923964351415634, + 0.021280884742736816, + 0.01114546600729227, + 0.015993312001228333, + 0.013716070912778378, + 0.010399061255156994, + 0.012133307754993439, + 0.022539954632520676, + 0.016321126371622086, + 0.0189980436116457, + 0.012906408868730068, + 0.0226559080183506, + 0.017954261973500252, + 0.014345787465572357, + 0.017330780625343323, + 0.007853833958506584, + 0.005899522453546524, + 0.007570536807179451, + 0.00417111674323678, + 0.005627527367323637, + 0.0107123963534832, + 0.004320491570979357, + 0.006390377879142761, + 0.01825273223221302, + 0.014202680438756943, + 0.017081882804632187, + 0.007665249519050121, + 0.0076940590515732765, + 0.01305911224335432, + 0.009480176493525505, + 0.008088071830570698, + 0.009602200239896774, + 0.018346888944506645, + 0.031919870525598526, + 0.003824380924925208, + 0.00239196908660233, + 0.005374264437705278, + 0.002196205547079444, + 0.0052962396293878555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.16037927567958832 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818976178765297 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.01818976178765297, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818976178765297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09000793099403381, + 0.072489894926548, + 0.2957459092140198, + 0.09953712671995163, + 0.0585516095161438, + 0.09583276510238647, + 0.046260859817266464, + 0.1263474076986313, + 0.05387690290808678, + 0.06134955957531929 + ], + "credit_map": [ + 0.016468647867441177, + 0.014789631590247154, + 0.01323737297207117, + 0.009280351921916008, + 0.009352995082736015, + 0.008184963837265968, + 0.009020493365824223, + 0.009673474356532097, + 0.01719738356769085, + 0.015689387917518616, + 0.010183568112552166, + 0.0072194975800812244, + 0.005006439983844757, + 0.007639019750058651, + 0.004219283349812031, + 0.005335322581231594, + 0.08125472068786621, + 0.06609257310628891, + 0.05245471000671387, + 0.01749310828745365, + 0.03141722455620766, + 0.019427642226219177, + 0.016176026314496994, + 0.011429891921579838, + 0.029623204842209816, + 0.0234481543302536, + 0.009732881560921669, + 0.007696831598877907, + 0.013299547135829926, + 0.006332992576062679, + 0.005417636130005121, + 0.003985878545790911, + 0.011849060654640198, + 0.005100773647427559, + 0.005549815949052572, + 0.0054425387643277645, + 0.006160478573292494, + 0.012985181994736195, + 0.00516965426504612, + 0.006294103804975748, + 0.01855406165122986, + 0.01154235191643238, + 0.016420478001236916, + 0.007923677563667297, + 0.010049407370388508, + 0.01211605966091156, + 0.009475587867200375, + 0.00975114107131958, + 0.010400247760117054, + 0.00531476316973567, + 0.004084093030542135, + 0.006021532230079174, + 0.00703730434179306, + 0.0067115239799022675, + 0.0035305186174809933, + 0.0031608727294951677, + 0.0168252382427454, + 0.014913320541381836, + 0.012040083296597004, + 0.012914121150970459, + 0.013753406703472137, + 0.024799002334475517, + 0.012183502316474915, + 0.018918735906481743, + 0.012154272757470608, + 0.009920015931129456, + 0.008905966766178608, + 0.004068813286721706, + 0.004041073843836784, + 0.0068633584305644035, + 0.003968989942222834, + 0.003954414278268814, + 0.0036034800577908754, + 0.02134047821164131, + 0.021019773557782173, + 0.0031736993696540594, + 0.0021645596716552973, + 0.004208914004266262, + 0.002479765797033906, + 0.003358883783221245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.01818976178765297 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070143431425095 + ], + "opposite_logprob_tokens": [ + -4.511070251464844 + ], + "logprob_reward": -0.011070143431425095, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070143431425095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.344422310590744, + 0.09466006606817245, + 0.12418846786022186, + 0.154326394200325, + 0.031904786825180054, + 0.021444927901029587, + 0.06163341924548149, + 0.055995792150497437, + 0.018363753333687782, + 0.09306003898382187 + ], + "credit_map": [ + 0.09569104015827179, + 0.07631725072860718, + 0.04432803764939308, + 0.025366973131895065, + 0.02272658422589302, + 0.035932812839746475, + 0.02221536450088024, + 0.0218442864716053, + 0.012939706444740295, + 0.01976832188665867, + 0.010982884094119072, + 0.011195911094546318, + 0.007993046194314957, + 0.012534617446362972, + 0.008909225463867188, + 0.010336354374885559, + 0.022791745141148567, + 0.03143700212240219, + 0.016315698623657227, + 0.014530179090797901, + 0.01291962992399931, + 0.011426826007664204, + 0.008271093480288982, + 0.006496300920844078, + 0.024475106969475746, + 0.04716920480132103, + 0.025866372510790825, + 0.013671495951712132, + 0.011440882459282875, + 0.012473270297050476, + 0.010464465245604515, + 0.008765620179474354, + 0.008319010026752949, + 0.002198253059759736, + 0.0013789773220196366, + 0.0020758977625519037, + 0.002924540778622031, + 0.008575626648962498, + 0.0020411978475749493, + 0.0043912832625210285, + 0.003339221468195319, + 0.0019306835019961, + 0.0028059319593012333, + 0.0021985950879752636, + 0.0019776292610913515, + 0.005544270854443312, + 0.0019277642713859677, + 0.0017208342906087637, + 0.008076873607933521, + 0.004994741640985012, + 0.00902252271771431, + 0.0053534735925495625, + 0.009659435600042343, + 0.009584803134202957, + 0.006338687147945166, + 0.008602886460721493, + 0.008483663201332092, + 0.007758504711091518, + 0.004580105189234018, + 0.003544793464243412, + 0.004666447173804045, + 0.014962052926421165, + 0.004461375530809164, + 0.007538852281868458, + 0.004581957124173641, + 0.0028679906390607357, + 0.0034942824859172106, + 0.0009864199673756957, + 0.0013022123603150249, + 0.0028271712362766266, + 0.0010310938814654946, + 0.0012726255226880312, + 0.02447625622153282, + 0.00874079205095768, + 0.038851313292980194, + 0.0031309262849390507, + 0.0017264146590605378, + 0.0073195453733205795, + 0.002119580050930381, + 0.006695214658975601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.011070143431425095 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a majestic black horse at full speed along a dusty ancient road. The camera moves in sync with the horse, emphasizing its thunderous speed. Atop his horse, a man in ancient costume, long hair tied in a crown, sits as steady as a mountain. His raven black hair is held high by an exquisite white jade crown, the remaining strands dancing wildly behind him along with his wide, deep teal silk robe in the wind. In a slightly yellowish, retro light, the afterglow of the setting sun casts a warm, timeless hue over the entire image, precisely outlining his resolute facial features—his tightly pursed lips, his straight nose, and his piercing eyes fixed intently ahead. His robe is so intricately textured that even the finest cloud patterns embroidered in gold thread on the flowing hem are visible, fluttering violently with every leap of the horse. The horse's muscles bulge, its glossy coat dripping with beads of sweat, gleaming in the golden light. Its hooves churn, powerfully pawing the ground and stirring up a cloud of dust mixed with gravel. The backdrop, a backdrop of rolling barren hills and sparse, dead trees, appears desolate and vast in the elongated shadows, creating an epic sense of urgency and poignancy, a lone figure dashing toward a fate.", + "seed": 17346, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_11/index_8211_seed_17346_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.036668424308300016, + "vlm_logprob": -0.036668424308300016, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.036668424308300016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06416545063257217 + ], + "opposite_logprob_tokens": [ + -3.3141653537750244 + ], + "logprob_reward": -0.06416545063257217, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06416545063257217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10460031777620316, + 0.437435120344162, + 0.032929208129644394, + 0.1920827180147171, + 0.034416936337947845, + 0.02777782641351223, + 0.03346981480717659, + 0.031091613695025444, + 0.0666620060801506, + 0.039534490555524826 + ], + "credit_map": [ + 0.01890598051249981, + 0.008421102538704872, + 0.01058349572122097, + 0.006701343227177858, + 0.01248998288065195, + 0.015605763532221317, + 0.007189252879470587, + 0.024703385308384895, + 0.06153620779514313, + 0.009816124103963375, + 0.012227646075189114, + 0.00505237327888608, + 0.009909063577651978, + 0.19816336035728455, + 0.009776975959539413, + 0.1309533268213272, + 0.004553532227873802, + 0.002600037958472967, + 0.0034487927332520485, + 0.0027669044211506844, + 0.003986475523561239, + 0.004076450131833553, + 0.0033933829981833696, + 0.008103631436824799, + 0.06093869358301163, + 0.007528620772063732, + 0.011313153430819511, + 0.004672132898122072, + 0.008782735094428062, + 0.07102235406637192, + 0.0063376640900969505, + 0.021487349644303322, + 0.007794487755745649, + 0.0032264110632240772, + 0.003367625642567873, + 0.0034885755740106106, + 0.0030102035962045193, + 0.006578722037374973, + 0.0031206905841827393, + 0.0038302158936858177, + 0.004303190857172012, + 0.002154625253751874, + 0.002656175522133708, + 0.0024649621918797493, + 0.0024809911847114563, + 0.005464725662022829, + 0.002729198196902871, + 0.0055239563807845116, + 0.00873508770018816, + 0.002209295751526952, + 0.003573419526219368, + 0.001992603065446019, + 0.002650485374033451, + 0.007567303255200386, + 0.0029930928722023964, + 0.0037485251668840647, + 0.0025595095939934254, + 0.00831980537623167, + 0.007775155361741781, + 0.0016201954567804933, + 0.002635799115523696, + 0.003423430258408189, + 0.002281160093843937, + 0.0024765562266111374, + 0.01933855190873146, + 0.010687458328902721, + 0.010136224329471588, + 0.003811730071902275, + 0.004638946149498224, + 0.004801630042493343, + 0.0054453294724226, + 0.007802129723131657, + 0.006672453600913286, + 0.0064542461186647415, + 0.011092324741184711, + 0.003067552112042904, + 0.0032181486021727324, + 0.003942183218896389, + 0.0023665446788072586, + 0.0027210377156734467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.06416545063257217 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038132987916469574 + ], + "opposite_logprob_tokens": [ + -3.288132905960083 + ], + "logprob_reward": -0.038132987916469574, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038132987916469574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3190124034881592, + 0.06219492480158806, + 0.036468662321567535, + 0.4203729033470154, + 0.03480838984251022, + 0.011702332645654678, + 0.030536595731973648, + 0.021086081862449646, + 0.026559287682175636, + 0.03725844994187355 + ], + "credit_map": [ + 0.030793029814958572, + 0.01742626167833805, + 0.03790874406695366, + 0.014887007884681225, + 0.017081420868635178, + 0.09114477783441544, + 0.012386462651193142, + 0.09738466888666153, + 0.007665459532290697, + 0.006479169242084026, + 0.00729145435616374, + 0.005646164994686842, + 0.007867838256061077, + 0.013261509127914906, + 0.0037307722959667444, + 0.010252549313008785, + 0.006040997337549925, + 0.004758997820317745, + 0.003945320378988981, + 0.005253154318779707, + 0.003376738401129842, + 0.004130440764129162, + 0.003439530497416854, + 0.005523477215319872, + 0.12808141112327576, + 0.015205946750938892, + 0.021514885127544403, + 0.009365351870656013, + 0.018215591087937355, + 0.16324888169765472, + 0.014282356016337872, + 0.050458427518606186, + 0.005787591449916363, + 0.0028235381469130516, + 0.004532691556960344, + 0.003969086799770594, + 0.003584462683647871, + 0.006712025497108698, + 0.003262606915086508, + 0.004136383067816496, + 0.002161275362595916, + 0.0010369876399636269, + 0.0013961740769445896, + 0.0009470952209085226, + 0.001374698244035244, + 0.002491088816896081, + 0.001089495257474482, + 0.0012055167462676764, + 0.006432741414755583, + 0.0022447535302489996, + 0.002364221727475524, + 0.0016892343992367387, + 0.0036968772765249014, + 0.007097449619323015, + 0.002784718293696642, + 0.004226596560329199, + 0.0020421925000846386, + 0.005326309707015753, + 0.00259981001727283, + 0.0015733048785477877, + 0.002673741662874818, + 0.0024026224855333567, + 0.0017991288332268596, + 0.002668969100341201, + 0.00753784691914916, + 0.004686238709837198, + 0.005433622281998396, + 0.0014817483024671674, + 0.0015587463276460767, + 0.0017006066627800465, + 0.00156350526958704, + 0.002596972044557333, + 0.004862708039581776, + 0.0044050440192222595, + 0.005669912323355675, + 0.0036169944796711206, + 0.006017283536493778, + 0.004986231215298176, + 0.0037623560056090355, + 0.003937915898859501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.038132987916469574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865988716483116 + ], + "opposite_logprob_tokens": [ + -3.0486598014831543 + ], + "logprob_reward": -0.04865988716483116, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865988716483116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15078777074813843, + 0.4907416105270386, + 0.13679957389831543, + 0.11783938109874725, + 0.023311948403716087, + 0.01054860558360815, + 0.01616595685482025, + 0.0194894727319479, + 0.0191862303763628, + 0.015129397623240948 + ], + "credit_map": [ + 0.013936336152255535, + 0.00751072820276022, + 0.015394317917525768, + 0.007767423056066036, + 0.008890427649021149, + 0.04199700802564621, + 0.006741003133356571, + 0.04855053126811981, + 0.06164421886205673, + 0.009801400825381279, + 0.013941336423158646, + 0.004865196067839861, + 0.010497109033167362, + 0.22240501642227173, + 0.010888190940022469, + 0.15669916570186615, + 0.02345358394086361, + 0.011047115549445152, + 0.01854192093014717, + 0.008349847979843616, + 0.011860017664730549, + 0.019381675869226456, + 0.012477128766477108, + 0.031688276678323746, + 0.03595519810914993, + 0.004566503688693047, + 0.005899105221033096, + 0.002658949000760913, + 0.005440515000373125, + 0.044681742787361145, + 0.004013483412563801, + 0.014623885042965412, + 0.003534038318321109, + 0.002014120342209935, + 0.0021876629907637835, + 0.0017351813148707151, + 0.0026507615111768246, + 0.0051260413601994514, + 0.0017008522991091013, + 0.004363290034234524, + 0.002421521581709385, + 0.000832975609228015, + 0.00097888323944062, + 0.0010095498291775584, + 0.0010645895963534713, + 0.002239271067082882, + 0.0009849208872765303, + 0.001016894355416298, + 0.002468063961714506, + 0.0019973258022218943, + 0.0026867117267102003, + 0.0017610325012356043, + 0.0019415982533246279, + 0.002172677544876933, + 0.0016442430205643177, + 0.001494304509833455, + 0.0018148355884477496, + 0.004297406878322363, + 0.00620872201398015, + 0.0013429670361801982, + 0.0011569648049771786, + 0.001020735944621265, + 0.0013512765290215611, + 0.0022965630050748587, + 0.003743883688002825, + 0.003917418420314789, + 0.00465735699981451, + 0.0015566966030746698, + 0.0011116608511656523, + 0.0011207519564777613, + 0.001336558721959591, + 0.001741903368383646, + 0.0025199269875884056, + 0.002294979291036725, + 0.0028216626960784197, + 0.0011935873189941049, + 0.0014094681246206164, + 0.0023010638542473316, + 0.0011760223424062133, + 0.001412685844115913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04865988716483116 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200706034898758 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200706034898758, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200706034898758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11920894682407379, + 0.09121581166982651, + 0.027686096727848053, + 0.592982292175293, + 0.06308634579181671, + 0.007777787279337645, + 0.008121606893837452, + 0.016252171248197556, + 0.020687691867351532, + 0.052981238812208176 + ], + "credit_map": [ + 0.0101859662681818, + 0.005122988950461149, + 0.010053485631942749, + 0.005648217163980007, + 0.005854540504515171, + 0.03516242653131485, + 0.004553577397018671, + 0.04262774437665939, + 0.012332280166447163, + 0.0027392241172492504, + 0.0043463148176670074, + 0.0021344097331166267, + 0.002645838772878051, + 0.038362838327884674, + 0.002133098663762212, + 0.026521798223257065, + 0.004741524811834097, + 0.002633561147376895, + 0.003639271715655923, + 0.0021946029737591743, + 0.002602019812911749, + 0.0037240018136799335, + 0.0026416510809212923, + 0.0055094631388783455, + 0.1856149137020111, + 0.021701034158468246, + 0.031087912619113922, + 0.011536994017660618, + 0.02508665807545185, + 0.22873951494693756, + 0.0189875029027462, + 0.07022779434919357, + 0.010520501993596554, + 0.005416664760559797, + 0.005590884480625391, + 0.005662869196385145, + 0.005226614885032177, + 0.016139917075634003, + 0.005556201096624136, + 0.008972689509391785, + 0.0011432924075052142, + 0.0005719902692362666, + 0.0008515446679666638, + 0.0006976122967898846, + 0.000993994646705687, + 0.0016805080231279135, + 0.0007306855404749513, + 0.0011081587290391326, + 0.001739533618092537, + 0.000744204968214035, + 0.001440045889467001, + 0.0007044047233648598, + 0.0012076327111572027, + 0.0009018530836328864, + 0.0007968810386955738, + 0.0005870502209290862, + 0.0017886679852381349, + 0.0024519041180610657, + 0.004875372163951397, + 0.0007821176550351083, + 0.0010178028605878353, + 0.0010561372619122267, + 0.0013109719147905707, + 0.0029691969975829124, + 0.005384956486523151, + 0.003814021358266473, + 0.002880098996683955, + 0.001243161503225565, + 0.0013499013148248196, + 0.0017909178277477622, + 0.0017690338427200913, + 0.0024555986747145653, + 0.011048298329114914, + 0.009613464586436749, + 0.01610245741903782, + 0.0032329827081412077, + 0.003048371057957411, + 0.00421899976208806, + 0.002453883644193411, + 0.003262776881456375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.014200706034898758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183089792728424 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183089792728424, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183089792728424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1656501740217209, + 0.1519155204296112, + 0.07159914821386337, + 0.2757662534713745, + 0.04960425943136215, + 0.06406398117542267, + 0.029518360272049904, + 0.025200679898262024, + 0.07444017380475998, + 0.09224148839712143 + ], + "credit_map": [ + 0.018104247748851776, + 0.012164977379143238, + 0.0140443816781044, + 0.0112778814509511, + 0.01405242457985878, + 0.04304492101073265, + 0.008915958926081657, + 0.044045377522706985, + 0.025729453191161156, + 0.005427355412393808, + 0.005936626344919205, + 0.0027629851829260588, + 0.00473983958363533, + 0.06366818398237228, + 0.005002261139452457, + 0.0386488139629364, + 0.012146967463195324, + 0.007236524950712919, + 0.007757223676890135, + 0.006425605155527592, + 0.006201798561960459, + 0.010312584228813648, + 0.00686236284673214, + 0.014656079933047295, + 0.07040023058652878, + 0.011815794743597507, + 0.015602520667016506, + 0.00925594661384821, + 0.015184934251010418, + 0.10236488282680511, + 0.010427175089716911, + 0.040714774280786514, + 0.008590622805058956, + 0.004282982554286718, + 0.0067489054054021835, + 0.0032458011992275715, + 0.0046867486089468, + 0.009049504064023495, + 0.004251814913004637, + 0.008747880347073078, + 0.01806916669011116, + 0.004529996775090694, + 0.005210279021412134, + 0.004664915148168802, + 0.0039950767531991005, + 0.01514842826873064, + 0.005089275073260069, + 0.00735684297978878, + 0.006832584738731384, + 0.0027672932483255863, + 0.003577094990760088, + 0.0029776189476251602, + 0.0022667963057756424, + 0.00621984526515007, + 0.0029113965574651957, + 0.001965730218216777, + 0.0028395229019224644, + 0.006840537302196026, + 0.003588989609852433, + 0.0017515005310997367, + 0.0027224018704146147, + 0.0022457547020167112, + 0.0020998341497033834, + 0.0031121380161494017, + 0.022527696564793587, + 0.012293314561247826, + 0.008048124611377716, + 0.005613289773464203, + 0.004997728858143091, + 0.005872911773622036, + 0.007065560668706894, + 0.00802154652774334, + 0.01510905846953392, + 0.014651535078883171, + 0.026695219799876213, + 0.00680666696280241, + 0.0089501291513443, + 0.008610059507191181, + 0.005772611126303673, + 0.005646208766847849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.018183089792728424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a majestic black horse at full speed along a dusty ancient road. The camera moves in sync with the horse, emphasizing its thunderous speed. Atop his horse, a man in ancient costume, long hair tied in a crown, sits as steady as a mountain. His raven black hair is held high by an exquisite white jade crown, the remaining strands dancing wildly behind him along with his wide, deep teal silk robe in the wind. In a slightly yellowish, retro light, the afterglow of the setting sun casts a warm, timeless hue over the entire image, precisely outlining his resolute facial features—his tightly pursed lips, his straight nose, and his piercing eyes fixed intently ahead. His robe is so intricately textured that even the finest cloud patterns embroidered in gold thread on the flowing hem are visible, fluttering violently with every leap of the horse. The horse's muscles bulge, its glossy coat dripping with beads of sweat, gleaming in the golden light. Its hooves churn, powerfully pawing the ground and stirring up a cloud of dust mixed with gravel. The backdrop, a backdrop of rolling barren hills and sparse, dead trees, appears desolate and vast in the elongated shadows, creating an epic sense of urgency and poignancy, a lone figure dashing toward a fate.", + "seed": 17347, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_11/index_8211_seed_17347_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0464020567946136, + "vlm_logprob": -0.0464020567946136, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.0464020567946136, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10283675044775009 + ], + "opposite_logprob_tokens": [ + -2.977836847305298 + ], + "logprob_reward": -0.10283675044775009, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.10283675044775009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2216547727584839, + 0.13094232976436615, + 0.050996679812669754, + 0.10570520907640457, + 0.06847620755434036, + 0.12505480647087097, + 0.09104534983634949, + 0.03914663940668106, + 0.0771041065454483, + 0.08987385779619217 + ], + "credit_map": [ + 0.02368003875017166, + 0.018772117793560028, + 0.03500957414507866, + 0.019266227260231972, + 0.030301567167043686, + 0.02707853354513645, + 0.025023477151989937, + 0.042523231357336044, + 0.0096987159922719, + 0.011640423908829689, + 0.028633877635002136, + 0.011606241576373577, + 0.013634533621370792, + 0.018834566697478294, + 0.01285911351442337, + 0.02403486706316471, + 0.005897576920688152, + 0.004256839863955975, + 0.008244959637522697, + 0.005747687071561813, + 0.00764099508523941, + 0.0071377502754330635, + 0.005310105159878731, + 0.006760763470083475, + 0.010137497447431087, + 0.011779619380831718, + 0.011052191257476807, + 0.0067868949845433235, + 0.02630922943353653, + 0.01415741816163063, + 0.016764996573328972, + 0.008717372082173824, + 0.005345799028873444, + 0.009646086022257805, + 0.009527532383799553, + 0.006456043571233749, + 0.014383107423782349, + 0.009499241597950459, + 0.006228047888725996, + 0.007390356156975031, + 0.015300940722227097, + 0.01625300757586956, + 0.018485398963093758, + 0.010829501785337925, + 0.021426156163215637, + 0.018149403855204582, + 0.009921018034219742, + 0.014689384028315544, + 0.010955276899039745, + 0.007790892850607634, + 0.01425599493086338, + 0.006402686238288879, + 0.007993682287633419, + 0.0224623903632164, + 0.009035511873662472, + 0.012148909270763397, + 0.00658818893134594, + 0.0042174337431788445, + 0.004307292867451906, + 0.002652994357049465, + 0.007140409201383591, + 0.006426718086004257, + 0.004305307753384113, + 0.003508293069899082, + 0.010317914187908173, + 0.007983392104506493, + 0.010003398172557354, + 0.00448880298063159, + 0.012141856364905834, + 0.010225593112409115, + 0.013227192685008049, + 0.00871595274657011, + 0.0069482349790632725, + 0.004751277156174183, + 0.007200825028121471, + 0.004973533097654581, + 0.015116102993488312, + 0.01683652028441429, + 0.026363231241703033, + 0.0076841372065246105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.10283675044775009 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868429899215698 + ], + "opposite_logprob_tokens": [ + -3.0486843585968018 + ], + "logprob_reward": -0.04868429899215698, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868429899215698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07659083604812622, + 0.1014653742313385, + 0.01728205382823944, + 0.0371934249997139, + 0.030147133395075798, + 0.18026164174079895, + 0.046979036182165146, + 0.095206618309021, + 0.34603142738342285, + 0.06884244084358215 + ], + "credit_map": [ + 0.005728587973862886, + 0.008493123576045036, + 0.010851720348000526, + 0.012455353513360023, + 0.009851068258285522, + 0.009445806965231895, + 0.009181422181427479, + 0.010583756491541862, + 0.01187660451978445, + 0.0070997183211147785, + 0.01993715763092041, + 0.013062872923910618, + 0.008212883025407791, + 0.014610527083277702, + 0.0116947703063488, + 0.01497084740549326, + 0.0020083002746105194, + 0.0014263488119468093, + 0.002292558318004012, + 0.0015704904217272997, + 0.002590187592431903, + 0.0022624777629971504, + 0.0025271703489124775, + 0.0026045197155326605, + 0.0043159485794603825, + 0.00418308237567544, + 0.004331739619374275, + 0.0026879108045250177, + 0.008286829106509686, + 0.005285768304020166, + 0.00451233284547925, + 0.0035898154601454735, + 0.0029707890935242176, + 0.003198804333806038, + 0.0037313492503017187, + 0.003269321285188198, + 0.007128805387765169, + 0.004040727391839027, + 0.003833862952888012, + 0.0019734736997634172, + 0.019500385969877243, + 0.018991684541106224, + 0.037527259439229965, + 0.025738894939422607, + 0.020915746688842773, + 0.018563399091362953, + 0.02141399122774601, + 0.01761028543114662, + 0.011825197376310825, + 0.004315793979912996, + 0.004601082764565945, + 0.003842817386612296, + 0.004237492568790913, + 0.008719333447515965, + 0.0041282279416918755, + 0.005309090483933687, + 0.013159310445189476, + 0.012323753908276558, + 0.01549008209258318, + 0.008123982697725296, + 0.010874775238335133, + 0.017587384209036827, + 0.009029722772538662, + 0.008617604151368141, + 0.031582433730363846, + 0.03281093388795853, + 0.03184240683913231, + 0.018091026693582535, + 0.0696449726819992, + 0.04675782471895218, + 0.08150561153888702, + 0.033796221017837524, + 0.00802199449390173, + 0.009057419374585152, + 0.007822256535291672, + 0.009064344689249992, + 0.010934773832559586, + 0.008778424933552742, + 0.007902064360678196, + 0.007261160295456648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.04868429899215698 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306568205356598 + ], + "opposite_logprob_tokens": [ + -3.1680657863616943 + ], + "logprob_reward": -0.04306568205356598, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306568205356598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33421429991722107, + 0.1779240071773529, + 0.03924892842769623, + 0.13679490983486176, + 0.04092993959784508, + 0.15226295590400696, + 0.02610713057219982, + 0.018220599740743637, + 0.05238352343440056, + 0.021913684904575348 + ], + "credit_map": [ + 0.03638873249292374, + 0.025338158011436462, + 0.06344285607337952, + 0.027792224660515785, + 0.048523835837841034, + 0.0438961386680603, + 0.04080912098288536, + 0.04802325740456581, + 0.01616736873984337, + 0.011500241234898567, + 0.04358692467212677, + 0.01064992044121027, + 0.015123216435313225, + 0.025588996708393097, + 0.011844865046441555, + 0.043462470173835754, + 0.0033464219886809587, + 0.0036372181493788958, + 0.007607178296893835, + 0.003335100132972002, + 0.00599716417491436, + 0.006370663642883301, + 0.004422803875058889, + 0.004532379563897848, + 0.010624824091792107, + 0.0156125882640481, + 0.01835988648235798, + 0.012851708568632603, + 0.026994463056325912, + 0.02231074683368206, + 0.01276337169110775, + 0.017277320846915245, + 0.004783031065016985, + 0.007096725050359964, + 0.006554703693836927, + 0.004060631617903709, + 0.005366996396332979, + 0.0052967495284974575, + 0.003441387787461281, + 0.0043297139927744865, + 0.019814062863588333, + 0.022844022139906883, + 0.020295459777116776, + 0.015017544850707054, + 0.027316132560372353, + 0.02275192178785801, + 0.010975529439747334, + 0.013248289935290813, + 0.004049227572977543, + 0.002335252007469535, + 0.0037154406309127808, + 0.0029431001748889685, + 0.0043582608923316, + 0.002871997654438019, + 0.0027539071161299944, + 0.0030799449887126684, + 0.0026265813503414392, + 0.0016673627542331815, + 0.003446318209171295, + 0.0012266976991668344, + 0.002479894319549203, + 0.0023991840425878763, + 0.0015274951001629233, + 0.0028470675460994244, + 0.006084509193897247, + 0.005867852363735437, + 0.008065042085945606, + 0.003871972905471921, + 0.009204409085214138, + 0.005733134690672159, + 0.008685367181897163, + 0.004871234763413668, + 0.0024549956433475018, + 0.0018525870982557535, + 0.002077121753245592, + 0.0017157726688310504, + 0.002483188873156905, + 0.004121905192732811, + 0.004945201799273491, + 0.002262913156300783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.04306568205356598 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980063669383526 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980063669383526, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980063669383526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26974397897720337, + 0.08738973736763, + 0.03990840166807175, + 0.09478927403688431, + 0.04081932082772255, + 0.14069387316703796, + 0.06745799630880356, + 0.045978400856256485, + 0.07099024951457977, + 0.1422288417816162 + ], + "credit_map": [ + 0.03432929515838623, + 0.01899690181016922, + 0.0456700474023819, + 0.023250557482242584, + 0.041187189519405365, + 0.037084732204675674, + 0.026874538511037827, + 0.04235069453716278, + 0.006067384500056505, + 0.006401723716408014, + 0.017039835453033447, + 0.008580487221479416, + 0.008282479830086231, + 0.011232750490307808, + 0.008821309544146061, + 0.0209637563675642, + 0.0041264197789132595, + 0.003974413964897394, + 0.009713754057884216, + 0.003934318199753761, + 0.00527270557358861, + 0.005278880707919598, + 0.0032507472205907106, + 0.0043571628630161285, + 0.01212545856833458, + 0.010774945840239525, + 0.012940434738993645, + 0.009258605539798737, + 0.012974978424608707, + 0.018360506743192673, + 0.007422967813909054, + 0.01093137264251709, + 0.006254475563764572, + 0.007655961439013481, + 0.0057218619622290134, + 0.003432357218116522, + 0.00536107923835516, + 0.004444231279194355, + 0.0032908532302826643, + 0.00465850206092, + 0.016596496105194092, + 0.017065931111574173, + 0.024924561381340027, + 0.017613358795642853, + 0.018218349665403366, + 0.020628344267606735, + 0.012336974032223225, + 0.013309850357472897, + 0.0072693959809839725, + 0.005897894501686096, + 0.01317967101931572, + 0.006062251981347799, + 0.009282463230192661, + 0.011869753710925579, + 0.00804021768271923, + 0.005856345873326063, + 0.007030890788882971, + 0.004910912364721298, + 0.009312083944678307, + 0.002007843926548958, + 0.005291325505822897, + 0.007736436557024717, + 0.004532283172011375, + 0.0051566231995821, + 0.008280721493065357, + 0.007393574807792902, + 0.008958322927355766, + 0.004005803260952234, + 0.012243586592376232, + 0.008884171023964882, + 0.014804657548666, + 0.006419409066438675, + 0.014885598793625832, + 0.0057808589190244675, + 0.01143497321754694, + 0.008824964053928852, + 0.018791064620018005, + 0.01931859366595745, + 0.0497460812330246, + 0.013446707278490067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.02980063669383526 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a man wearing an ancient costume with a white jade crown and a teal silk robe riding a black horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse moving at full speed, causing dust and gravel to churn up from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting yellowish and retro, simulating the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of rolling barren hills and sparse dead trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622915785759687 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622915785759687, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622915785759687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15829645097255707, + 0.27722445130348206, + 0.09280038625001907, + 0.11691224575042725, + 0.033296551555395126, + 0.07623371481895447, + 0.06064861640334129, + 0.11768020689487457, + 0.04676203429698944, + 0.020145289599895477 + ], + "credit_map": [ + 0.013589000329375267, + 0.013780658133327961, + 0.022762106731534004, + 0.012782451696693897, + 0.029084965586662292, + 0.02025521546602249, + 0.01758178509771824, + 0.02846027910709381, + 0.02714325301349163, + 0.01840914972126484, + 0.071640744805336, + 0.016161056235432625, + 0.021331848576664925, + 0.038473520427942276, + 0.018715836107730865, + 0.0653490498661995, + 0.00936213880777359, + 0.007786933798342943, + 0.024227846413850784, + 0.008528674021363258, + 0.008971682749688625, + 0.014185870066285133, + 0.008049155585467815, + 0.011688091792166233, + 0.01472618617117405, + 0.010423332452774048, + 0.013425616547465324, + 0.007951797917485237, + 0.019177015870809555, + 0.019629573449492455, + 0.014185050502419472, + 0.017393678426742554, + 0.006056540179997683, + 0.007907718420028687, + 0.004888725932687521, + 0.002786995843052864, + 0.002948024310171604, + 0.0029670847579836845, + 0.002435064408928156, + 0.0033063984010368586, + 0.011206081137061119, + 0.012575958855450153, + 0.009371326304972172, + 0.0053715514950454235, + 0.012838173657655716, + 0.011089304462075233, + 0.0048716748133301735, + 0.008909647352993488, + 0.010914921760559082, + 0.003567870706319809, + 0.0067123365588486195, + 0.00380452792160213, + 0.00502025056630373, + 0.012819367460906506, + 0.004092569928616285, + 0.013716774061322212, + 0.019795536994934082, + 0.008325818926095963, + 0.023220455273985863, + 0.007532279472798109, + 0.013298236764967442, + 0.022817494347691536, + 0.008595169521868229, + 0.014095230028033257, + 0.005235089920461178, + 0.005206088535487652, + 0.004785466939210892, + 0.002621137071400881, + 0.007309851702302694, + 0.006457962561398745, + 0.010301657021045685, + 0.0048447828739881516, + 0.0023780358023941517, + 0.0023254724219441414, + 0.00220824358984828, + 0.002217125380411744, + 0.0029047965072095394, + 0.0033286619000136852, + 0.002995692193508148, + 0.0017872629687190056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.007622915785759687 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_12/index_tensor([8425, 8425])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_12/index_tensor([8425, 8425])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..510864c791e36f3e87e876c934e70a754b8c48b4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_12/index_tensor([8425, 8425])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The shot begins with a high-angle perspective, piercing the churning belly of a stormy sky filled with scarred charcoal-gray and deep purple clouds, roiling with unspoken fury. As the camera executes a slow, deliberate crane descent, a magnificent Gothic mansion gradually emerges, its jagged spires and towers first tearing through the low-hanging mist. The building looms, constructed of dark, weathered stone mortar, threatened by the rain, its walls partially devoured by creeping, dark-green ivy. Tall lancet windows, like hollow eyes, stare out from the facade, their blackened glass reflecting the turbulent sky, while eerie-faced stone gargoyles perch on the ledges, frozen in silent screams. The camera continues its descent, gliding smoothly across the imposing facade, ultimately landing at a low angle at the manor's ornate front gate, a towering masterpiece of black wrought iron twisted into intricate patterns of thorns, vines, and skeletons, its surface revealing a rust-colored decay. The entire scene is bathed in the cold, dim light of an impending storm, punctuated by occasional bursts of silent, sheet lightning that briefly cast harsh shadows and illuminate the textures of the damp stone, creating an atmosphere thick with foreboding, decaying grandeur, and a profound sense of isolation.", + "seed": 9541, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_12/index_8425_seed_9541_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0224198205396533, + "vlm_logprob": -0.0224198205396533, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.0224198205396533, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029154423624277115 + ], + "opposite_logprob_tokens": [ + -3.6541545391082764 + ], + "logprob_reward": -0.029154423624277115, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.029154423624277115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0507957823574543, + 0.02757486328482628, + 0.41594284772872925, + 0.06232088804244995, + 0.04090000316500664, + 0.023544078692793846, + 0.04859771206974983, + 0.042920131236314774, + 0.24312640726566315, + 0.04427725449204445 + ], + "credit_map": [ + 0.004857331048697233, + 0.007201189175248146, + 0.004270724020898342, + 0.00738917663693428, + 0.0036144349724054337, + 0.007848584093153477, + 0.010767624713480473, + 0.004846717230975628, + 0.0028665768913924694, + 0.002571478020399809, + 0.003065696218982339, + 0.0037829310167580843, + 0.002936136908829212, + 0.0035712169483304024, + 0.004742164630442858, + 0.004038663115352392, + 0.02039952017366886, + 0.16703230142593384, + 0.02027437463402748, + 0.013478495180606842, + 0.022964952513575554, + 0.15555518865585327, + 0.007793379947543144, + 0.008444654755294323, + 0.0054715522564947605, + 0.0061803534626960754, + 0.010999001562595367, + 0.011090201325714588, + 0.008038843050599098, + 0.006477852817624807, + 0.006758559495210648, + 0.007304523140192032, + 0.00491831311956048, + 0.006108762230724096, + 0.004172673914581537, + 0.005228050053119659, + 0.005638657137751579, + 0.006526910699903965, + 0.004051470197737217, + 0.004255164880305529, + 0.002709131920710206, + 0.003511850256472826, + 0.0031511627603322268, + 0.0026187002658843994, + 0.0029100484680384398, + 0.003002077341079712, + 0.003240335499867797, + 0.0024007740430533886, + 0.005954476539045572, + 0.0044886465184390545, + 0.005120522808283567, + 0.005640378687530756, + 0.007526540197432041, + 0.009371970780193806, + 0.0050915009342134, + 0.005403672344982624, + 0.003982917405664921, + 0.003489711321890354, + 0.004617858212441206, + 0.00865329708904028, + 0.0033042074646800756, + 0.0030683395452797413, + 0.007114089094102383, + 0.008689716458320618, + 0.013807862997055054, + 0.020392123609781265, + 0.026999225839972496, + 0.05644401162862778, + 0.017482729628682137, + 0.03002743236720562, + 0.03914952278137207, + 0.03882349655032158, + 0.0034897171426564455, + 0.0035303726326674223, + 0.007897156290709972, + 0.0061209299601614475, + 0.0034570866264402866, + 0.005428625270724297, + 0.007921669632196426, + 0.006431694608181715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.029154423624277115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370694860816002 + ], + "opposite_logprob_tokens": [ + -3.6513707637786865 + ], + "logprob_reward": -0.026370694860816002, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370694860816002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04122835397720337, + 0.027481362223625183, + 0.1981278359889984, + 0.03211169317364693, + 0.04091056063771248, + 0.0564948208630085, + 0.11508018523454666, + 0.02869301661849022, + 0.33327266573905945, + 0.12659944593906403 + ], + "credit_map": [ + 0.003528542583808303, + 0.006431873422116041, + 0.003856597002595663, + 0.005058337468653917, + 0.0035505888517946005, + 0.006187770050019026, + 0.008314925245940685, + 0.004299720749258995, + 0.0023559383116662502, + 0.0031084863003343344, + 0.005388782825320959, + 0.0027591767720878124, + 0.004392017610371113, + 0.002624532440677285, + 0.004643579013645649, + 0.0022088505793362856, + 0.0097291748970747, + 0.07811388373374939, + 0.009489699266850948, + 0.008467077277600765, + 0.010762092657387257, + 0.0731399655342102, + 0.003953387960791588, + 0.004472577478736639, + 0.004145551938563585, + 0.0033881717827171087, + 0.003962650429457426, + 0.007063237018883228, + 0.0037946952506899834, + 0.0034993840381503105, + 0.003535935655236244, + 0.00272207148373127, + 0.005678130313754082, + 0.007598292548209429, + 0.003323795972391963, + 0.005719357635825872, + 0.005652423482388258, + 0.0053180791437625885, + 0.00405135378241539, + 0.0035691349767148495, + 0.004320158157497644, + 0.00553237646818161, + 0.006111583206802607, + 0.005049642641097307, + 0.010154119692742825, + 0.011496205814182758, + 0.004551345948129892, + 0.009279392659664154, + 0.01362127996981144, + 0.012086980044841766, + 0.009451794438064098, + 0.013843812048435211, + 0.026715941727161407, + 0.018657037988305092, + 0.01064265239983797, + 0.010060694999992847, + 0.0029118689708411694, + 0.002241536509245634, + 0.004064491949975491, + 0.006989221088588238, + 0.002425034996122122, + 0.0023012328892946243, + 0.004231709521263838, + 0.003527922322973609, + 0.017673801630735397, + 0.022845376282930374, + 0.042783066630363464, + 0.06618023663759232, + 0.021245429292321205, + 0.04481250047683716, + 0.05195748060941696, + 0.06577477604150772, + 0.010974347591400146, + 0.01272896584123373, + 0.01869697868824005, + 0.02110203728079796, + 0.01033175177872181, + 0.012113275937736034, + 0.019855275750160217, + 0.020796818658709526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.026370694860816002 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059472741559147835 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.0059472741559147835, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059472741559147835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18379011750221252, + 0.1377267986536026, + 0.17180317640304565, + 0.034071702510118484, + 0.09585478901863098, + 0.0826098844408989, + 0.10910411179065704, + 0.04485054314136505, + 0.07127086073160172, + 0.06891805678606033 + ], + "credit_map": [ + 0.014650451019406319, + 0.02777808904647827, + 0.013797168619930744, + 0.0250187199562788, + 0.017589479684829712, + 0.028736133128404617, + 0.030783984810113907, + 0.025436080992221832, + 0.011065857484936714, + 0.013422952964901924, + 0.02350775897502899, + 0.015233517624437809, + 0.017227401956915855, + 0.013179785571992397, + 0.026538437232375145, + 0.017551086843013763, + 0.007920477539300919, + 0.06032314896583557, + 0.009815100580453873, + 0.01567249186336994, + 0.010473133064806461, + 0.05675920099020004, + 0.005379126872867346, + 0.0054604895412921906, + 0.0037737395614385605, + 0.004704523365944624, + 0.0037856465205550194, + 0.00797702744603157, + 0.004227687604725361, + 0.004417078103870153, + 0.003266399260610342, + 0.0019195990171283484, + 0.017204634845256805, + 0.017431296408176422, + 0.009245788678526878, + 0.010804733261466026, + 0.012708965688943863, + 0.014067312702536583, + 0.007784107234328985, + 0.006607947871088982, + 0.008703982457518578, + 0.006686714943498373, + 0.00651942566037178, + 0.008456318639218807, + 0.013967212289571762, + 0.01648815907537937, + 0.007140283938497305, + 0.014647788368165493, + 0.012118455953896046, + 0.011200960725545883, + 0.008582813665270805, + 0.013254533521831036, + 0.027453714981675148, + 0.01817026548087597, + 0.009029271081089973, + 0.009294087067246437, + 0.003294366179034114, + 0.0035256787668913603, + 0.003822777420282364, + 0.010422705672681332, + 0.0035582222044467926, + 0.0034325490705668926, + 0.011811697855591774, + 0.004982542246580124, + 0.003627728670835495, + 0.006563553586602211, + 0.01175673957914114, + 0.01662282459437847, + 0.0035807506646960974, + 0.0073164221830666065, + 0.010942983441054821, + 0.010859861969947815, + 0.004809985868632793, + 0.007991486229002476, + 0.012674421072006226, + 0.01166880689561367, + 0.003929996397346258, + 0.007132828235626221, + 0.009824593551456928, + 0.010885938070714474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0059472741559147835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540768831968307 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540768831968307, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540768831968307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07746440172195435, + 0.0650646835565567, + 0.2134828269481659, + 0.05294060707092285, + 0.04432392865419388, + 0.11001847684383392, + 0.19506217539310455, + 0.052268486469984055, + 0.1332481950521469, + 0.0561261847615242 + ], + "credit_map": [ + 0.0070374091155827045, + 0.009450403042137623, + 0.009021387435495853, + 0.010412531904876232, + 0.009834357537329197, + 0.007430754601955414, + 0.012773255817592144, + 0.011504299007356167, + 0.0047205123119056225, + 0.006337399128824472, + 0.009791512042284012, + 0.009094394743442535, + 0.007063786964863539, + 0.00937237124890089, + 0.010674601420760155, + 0.008010105229914188, + 0.009367408230900764, + 0.07768462598323822, + 0.013991373591125011, + 0.016565321013331413, + 0.012356053106486797, + 0.07014474272727966, + 0.0064216554164886475, + 0.006951653864234686, + 0.005952491424977779, + 0.00802204292267561, + 0.0057811732403934, + 0.010393430478870869, + 0.007922765798866749, + 0.00601833313703537, + 0.004703656770288944, + 0.00414671003818512, + 0.004895905498415232, + 0.006212982814759016, + 0.004372823052108288, + 0.006889625918120146, + 0.005326473619788885, + 0.00543179688975215, + 0.006163538433611393, + 0.0050307828933000565, + 0.015115742571651936, + 0.010511086322367191, + 0.008911817334592342, + 0.011265700682997704, + 0.017140254378318787, + 0.021057860925793648, + 0.009714878164231777, + 0.016301138326525688, + 0.030778422951698303, + 0.013461494818329811, + 0.015134159475564957, + 0.021895617246627808, + 0.044806383550167084, + 0.02514992468059063, + 0.01957264170050621, + 0.0242635365575552, + 0.002867282135412097, + 0.003114008344709873, + 0.004263630136847496, + 0.014737306162714958, + 0.003787847701460123, + 0.0034686981234699488, + 0.01531624048948288, + 0.0047134761698544025, + 0.00782320648431778, + 0.01142182108014822, + 0.015291813760995865, + 0.026120033115148544, + 0.0090419240295887, + 0.017764724791049957, + 0.018669243901968002, + 0.027115432545542717, + 0.004084835760295391, + 0.004868124611675739, + 0.011811218224465847, + 0.007641131989657879, + 0.005222713574767113, + 0.0061890240758657455, + 0.009001913480460644, + 0.007307224441319704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.012540768831968307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380859412252903 + ], + "opposite_logprob_tokens": [ + -3.2880859375 + ], + "logprob_reward": -0.0380859412252903, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380859412252903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049367550760507584, + 0.04235074296593666, + 0.3005906045436859, + 0.030857209116220474, + 0.028636468574404716, + 0.02923688292503357, + 0.10223104804754257, + 0.07796132564544678, + 0.09468310326337814, + 0.24408508837223053 + ], + "credit_map": [ + 0.00484608905389905, + 0.005246215965598822, + 0.005224039312452078, + 0.006059657782316208, + 0.007570508401840925, + 0.006806814577430487, + 0.00734816724434495, + 0.006266061682254076, + 0.003331237705424428, + 0.003494087839499116, + 0.006891156081110239, + 0.004852721933275461, + 0.005225271452218294, + 0.004318814724683762, + 0.008965114131569862, + 0.005272336304187775, + 0.014104140922427177, + 0.11791665107011795, + 0.014566643163561821, + 0.015059471130371094, + 0.016450833529233932, + 0.10968916863203049, + 0.006222525145858526, + 0.006581190042197704, + 0.0028923279605805874, + 0.004980813711881638, + 0.003824749728664756, + 0.004353531636297703, + 0.0037102645728737116, + 0.005273797549307346, + 0.002435215748846531, + 0.0033865105360746384, + 0.0037291639018803835, + 0.005282633006572723, + 0.0028599516954272985, + 0.003674502484500408, + 0.003936173859983683, + 0.003281253855675459, + 0.002877591410651803, + 0.002995198592543602, + 0.0026887652929872274, + 0.003166426904499531, + 0.0033137428108602762, + 0.004038849379867315, + 0.003684070659801364, + 0.004124116152524948, + 0.002850966528058052, + 0.005369945429265499, + 0.012057228945195675, + 0.009612924419343472, + 0.006801124662160873, + 0.012437882833182812, + 0.02757248282432556, + 0.016307169571518898, + 0.009153933264315128, + 0.00828829687088728, + 0.004343318287283182, + 0.004478133749216795, + 0.006450561806559563, + 0.022082380950450897, + 0.003621740499511361, + 0.003767424263060093, + 0.023373229429125786, + 0.009844539687037468, + 0.006344999652355909, + 0.011423822492361069, + 0.013924521394073963, + 0.01567246951162815, + 0.00753118796274066, + 0.013978403992950916, + 0.012482010759413242, + 0.01332568284124136, + 0.018629727885127068, + 0.020636217668652534, + 0.05783240869641304, + 0.03906022012233734, + 0.01727703958749771, + 0.021338317543268204, + 0.03468308597803116, + 0.03462808579206467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.0380859412252903 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a high-angle perspective, piercing the churning belly of a stormy sky filled with scarred charcoal-gray and deep purple clouds, roiling with unspoken fury. As the camera executes a slow, deliberate crane descent, a magnificent Gothic mansion gradually emerges, its jagged spires and towers first tearing through the low-hanging mist. The building looms, constructed of dark, weathered stone mortar, threatened by the rain, its walls partially devoured by creeping, dark-green ivy. Tall lancet windows, like hollow eyes, stare out from the facade, their blackened glass reflecting the turbulent sky, while eerie-faced stone gargoyles perch on the ledges, frozen in silent screams. The camera continues its descent, gliding smoothly across the imposing facade, ultimately landing at a low angle at the manor's ornate front gate, a towering masterpiece of black wrought iron twisted into intricate patterns of thorns, vines, and skeletons, its surface revealing a rust-colored decay. The entire scene is bathed in the cold, dim light of an impending storm, punctuated by occasional bursts of silent, sheet lightning that briefly cast harsh shadows and illuminate the textures of the damp stone, creating an atmosphere thick with foreboding, decaying grandeur, and a profound sense of isolation.", + "seed": 9542, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_12/index_8425_seed_9542_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.004277385445311666, + "vlm_logprob": -0.004277385445311666, + "vlm_yes_no_margin": 5.55, + "vlm_token_credit": -0.004277385445311666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0043703047558665276 + ], + "opposite_logprob_tokens": [ + -5.754370212554932 + ], + "logprob_reward": -0.0043703047558665276, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0043703047558665276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25816279649734497, + 0.05788878723978996, + 0.1988440304994583, + 0.06045789271593094, + 0.11496668308973312, + 0.0550665557384491, + 0.07038859277963638, + 0.08261540532112122, + 0.03994445502758026, + 0.06166481971740723 + ], + "credit_map": [ + 0.017072131857275963, + 0.06170232966542244, + 0.017697609961032867, + 0.027327274903655052, + 0.05668153613805771, + 0.03252249211072922, + 0.02657369151711464, + 0.018585724756121635, + 0.0036253693979233503, + 0.004332583863288164, + 0.005128825549036264, + 0.008161148056387901, + 0.011700833216309547, + 0.01302874181419611, + 0.005965893622487783, + 0.005945393815636635, + 0.006118672899901867, + 0.007190102711319923, + 0.005177977029234171, + 0.008760911412537098, + 0.07352428138256073, + 0.07783346623182297, + 0.008954602293670177, + 0.011284012347459793, + 0.004281220957636833, + 0.0053574140183627605, + 0.0059026796370744705, + 0.010865751653909683, + 0.008555132895708084, + 0.008195159025490284, + 0.0066399783827364445, + 0.010660558007657528, + 0.007338545750826597, + 0.010014256462454796, + 0.006844461429864168, + 0.019150791689753532, + 0.012455279007554054, + 0.026189828291535378, + 0.014065484516322613, + 0.01890803687274456, + 0.00334282242693007, + 0.006959761492908001, + 0.004824841860681772, + 0.011398137547075748, + 0.005581142380833626, + 0.007191885262727737, + 0.008263777010142803, + 0.007504188921302557, + 0.006305214483290911, + 0.005931934807449579, + 0.009306937456130981, + 0.012253567576408386, + 0.009684382937848568, + 0.009391863830387592, + 0.009625783190131187, + 0.007888901978731155, + 0.011718415655195713, + 0.0055319517850875854, + 0.01241831574589014, + 0.008242899551987648, + 0.005915960296988487, + 0.010041281580924988, + 0.012851572595536709, + 0.01589500717818737, + 0.0050887162797153, + 0.004467243328690529, + 0.006719578057527542, + 0.004525248892605305, + 0.006051029544323683, + 0.004482886753976345, + 0.004075133241713047, + 0.004534618929028511, + 0.006800907664000988, + 0.008930991403758526, + 0.009891302324831486, + 0.012912225909531116, + 0.005061204545199871, + 0.0061897337436676025, + 0.006820553448051214, + 0.005057903006672859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0043703047558665276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643134772777557 + ], + "opposite_logprob_tokens": [ + -5.379642963409424 + ], + "logprob_reward": -0.004643134772777557, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643134772777557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12387876957654953, + 0.03373708575963974, + 0.06589949131011963, + 0.08871309459209442, + 0.08659499883651733, + 0.07504386454820633, + 0.07805608212947845, + 0.15827620029449463, + 0.18122780323028564, + 0.10857260227203369 + ], + "credit_map": [ + 0.008080211468040943, + 0.026780681684613228, + 0.006845836061984301, + 0.015424760058522224, + 0.026887957006692886, + 0.01623019203543663, + 0.012725505977869034, + 0.010903624817728996, + 0.0025956041645258665, + 0.002141609787940979, + 0.002167438855394721, + 0.002400066703557968, + 0.00887264683842659, + 0.010007469914853573, + 0.0030875145457684994, + 0.0024647300597280264, + 0.004149287939071655, + 0.0071416995488107204, + 0.004381599836051464, + 0.007222650106996298, + 0.01613239198923111, + 0.016833260655403137, + 0.004977779928594828, + 0.0050608208402991295, + 0.005780587904155254, + 0.008410893380641937, + 0.004874468315392733, + 0.014973497949540615, + 0.014235053211450577, + 0.020292680710554123, + 0.009175616316497326, + 0.010970302857458591, + 0.005287928506731987, + 0.006034240126609802, + 0.005386393517255783, + 0.010928701609373093, + 0.009227976202964783, + 0.02613949403166771, + 0.007437596097588539, + 0.01615266315639019, + 0.005036485381424427, + 0.01111095491796732, + 0.005836342461407185, + 0.014236380346119404, + 0.010372191667556763, + 0.008733713999390602, + 0.010437022894620895, + 0.009280769154429436, + 0.0058174775913357735, + 0.005997061729431152, + 0.008342050015926361, + 0.013405395671725273, + 0.014854203909635544, + 0.01072657760232687, + 0.010584809817373753, + 0.008328504860401154, + 0.02870826981961727, + 0.01125896256417036, + 0.019728511571884155, + 0.01696532592177391, + 0.01256554014980793, + 0.02245424874126911, + 0.02159578539431095, + 0.024999579414725304, + 0.02964567020535469, + 0.018064621835947037, + 0.02379273809492588, + 0.028535297140479088, + 0.0165463387966156, + 0.0202461089938879, + 0.022512629628181458, + 0.021884402260184288, + 0.009745635092258453, + 0.016644196584820747, + 0.01515743974596262, + 0.021506570279598236, + 0.010330772958695889, + 0.011584668420255184, + 0.013783521018922329, + 0.00981980375945568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.004643134772777557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816283842548728 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.002816283842548728, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816283842548728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37068092823028564, + 0.05826117470860481, + 0.09834203124046326, + 0.08254237473011017, + 0.08033955097198486, + 0.046721380203962326, + 0.06806184351444244, + 0.0530848503112793, + 0.0720633789896965, + 0.0699024572968483 + ], + "credit_map": [ + 0.02846633829176426, + 0.09235556423664093, + 0.027951311320066452, + 0.038828253746032715, + 0.0714406669139862, + 0.04553613439202309, + 0.037611063569784164, + 0.028491592034697533, + 0.004194879904389381, + 0.006062804721295834, + 0.004559239372611046, + 0.006487707141786814, + 0.013832674361765385, + 0.011328431777656078, + 0.0072074574418365955, + 0.004587977193295956, + 0.005457353312522173, + 0.005824709311127663, + 0.007726613897830248, + 0.007366152014583349, + 0.028077267110347748, + 0.029391828924417496, + 0.0065312921069562435, + 0.007966818287968636, + 0.00507099786773324, + 0.007629239931702614, + 0.005857828073203564, + 0.014805478043854237, + 0.011328631080687046, + 0.013455836102366447, + 0.010003697127103806, + 0.014390665106475353, + 0.0048133921809494495, + 0.008817988447844982, + 0.006518205162137747, + 0.010013805702328682, + 0.00941397249698639, + 0.01905905269086361, + 0.009264068678021431, + 0.012439057230949402, + 0.001990135759115219, + 0.004509081598371267, + 0.0043266513384878635, + 0.006569704972207546, + 0.010800047777593136, + 0.0060202740132808685, + 0.007892546243965626, + 0.004612937569618225, + 0.006081257946789265, + 0.006368301343172789, + 0.009011014364659786, + 0.009885387495160103, + 0.011713801883161068, + 0.007508659735321999, + 0.00973002053797245, + 0.007763396017253399, + 0.006045149639248848, + 0.004036550410091877, + 0.006621113047003746, + 0.009623272344470024, + 0.005458558909595013, + 0.007799899205565453, + 0.006624636240303516, + 0.006875669118016958, + 0.005552254617214203, + 0.005405769683420658, + 0.010687745176255703, + 0.011035807430744171, + 0.0050757648423314095, + 0.008787347935140133, + 0.015279395505785942, + 0.010239296592772007, + 0.007915301248431206, + 0.012029004283249378, + 0.007398978807032108, + 0.010078875347971916, + 0.00769799342378974, + 0.009846996515989304, + 0.008970599621534348, + 0.005964709911495447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.002816283842548728 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946918856352568 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005946918856352568, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946918856352568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20950493216514587, + 0.03999657183885574, + 0.054661065340042114, + 0.16264238953590393, + 0.15124396979808807, + 0.0842510238289833, + 0.06086684390902519, + 0.13749291002750397, + 0.06111758574843407, + 0.038222696632146835 + ], + "credit_map": [ + 0.01601000502705574, + 0.05136576294898987, + 0.013034988194704056, + 0.022467002272605896, + 0.044317252933979034, + 0.026587778702378273, + 0.02136562205851078, + 0.014356526546180248, + 0.004237852059304714, + 0.004688108805567026, + 0.0023256917484104633, + 0.005418595392256975, + 0.007192566059529781, + 0.007884045131504536, + 0.0035586210433393717, + 0.004691088106483221, + 0.004472087137401104, + 0.00516405189409852, + 0.008206075057387352, + 0.004197578877210617, + 0.009866274893283844, + 0.013030133210122585, + 0.004792564082890749, + 0.0049322983250021935, + 0.011059539392590523, + 0.016197284683585167, + 0.011894169263541698, + 0.03459988161921501, + 0.017677491530776024, + 0.027769647538661957, + 0.01702873408794403, + 0.026415644213557243, + 0.008444208651781082, + 0.010142997838556767, + 0.007738768123090267, + 0.02044527977705002, + 0.016399400308728218, + 0.044070545583963394, + 0.015022896230220795, + 0.02897987700998783, + 0.005958139896392822, + 0.009424888528883457, + 0.006613478064537048, + 0.015498735941946507, + 0.014275748282670975, + 0.010083005763590336, + 0.011106097139418125, + 0.011290932074189186, + 0.003515227697789669, + 0.006173795089125633, + 0.006911966484040022, + 0.008826040662825108, + 0.0075470865704119205, + 0.012059573084115982, + 0.0062761465087533, + 0.009557010605931282, + 0.01032748632133007, + 0.009755500592291355, + 0.01986224576830864, + 0.01944972388446331, + 0.010882725939154625, + 0.02100430056452751, + 0.01871052198112011, + 0.02750040590763092, + 0.004372831899672747, + 0.004907782655209303, + 0.01187987346202135, + 0.008819964714348316, + 0.007273840717971325, + 0.006467959377914667, + 0.007587770000100136, + 0.00980756338685751, + 0.004949179477989674, + 0.008427013643085957, + 0.003450850024819374, + 0.005027540028095245, + 0.0036223495844751596, + 0.005552249029278755, + 0.0037014340050518513, + 0.0034920794423669577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.005946918856352568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003610284999012947 + ], + "opposite_logprob_tokens": [ + -5.628610134124756 + ], + "logprob_reward": -0.003610284999012947, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003610284999012947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07203774154186249, + 0.039024535566568375, + 0.05230060592293739, + 0.1220451071858406, + 0.06229157745838165, + 0.028061609715223312, + 0.10414337366819382, + 0.14632095396518707, + 0.15123474597930908, + 0.2225397378206253 + ], + "credit_map": [ + 0.005060140509158373, + 0.014289889484643936, + 0.0046085333451628685, + 0.009715264663100243, + 0.014953900128602982, + 0.00964128877967596, + 0.008149899542331696, + 0.005618834402412176, + 0.0020402639638632536, + 0.0026210639625787735, + 0.002091102534905076, + 0.003996444866061211, + 0.010477552190423012, + 0.011143788695335388, + 0.0032952609471976757, + 0.00335906189866364, + 0.0031037458684295416, + 0.004494671244174242, + 0.004636041820049286, + 0.005582118406891823, + 0.010977322235703468, + 0.013056600466370583, + 0.0046523078344762325, + 0.00579780014231801, + 0.007723518647253513, + 0.009446842595934868, + 0.008984536863863468, + 0.023099124431610107, + 0.01563887670636177, + 0.026220086961984634, + 0.01211177185177803, + 0.01882036216557026, + 0.006845977623015642, + 0.009180644527077675, + 0.006777028553187847, + 0.006780534517019987, + 0.007344426587224007, + 0.010088182985782623, + 0.007531809154897928, + 0.007742978632450104, + 0.001641144510358572, + 0.0028245942667126656, + 0.0025076407473534346, + 0.00526687316596508, + 0.004884793423116207, + 0.0027276130858808756, + 0.004808703903108835, + 0.0034002482425421476, + 0.00679373275488615, + 0.012340460903942585, + 0.013625083491206169, + 0.012014872394502163, + 0.00900738500058651, + 0.022872207686305046, + 0.012085994705557823, + 0.015403642319142818, + 0.013742422685027122, + 0.011321141384541988, + 0.02025541290640831, + 0.02100416272878647, + 0.015304537490010262, + 0.017508940771222115, + 0.020311428233981133, + 0.026872914284467697, + 0.012204190716147423, + 0.018468178808689117, + 0.02406660094857216, + 0.02145443856716156, + 0.020021557807922363, + 0.02022537961602211, + 0.019723042845726013, + 0.015071362257003784, + 0.017106156796216965, + 0.027797756716609, + 0.020862838253378868, + 0.027472667396068573, + 0.019737139344215393, + 0.03861851617693901, + 0.046533919870853424, + 0.024410776793956757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.003610284999012947 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a high-angle perspective, piercing the churning belly of a stormy sky filled with scarred charcoal-gray and deep purple clouds, roiling with unspoken fury. As the camera executes a slow, deliberate crane descent, a magnificent Gothic mansion gradually emerges, its jagged spires and towers first tearing through the low-hanging mist. The building looms, constructed of dark, weathered stone mortar, threatened by the rain, its walls partially devoured by creeping, dark-green ivy. Tall lancet windows, like hollow eyes, stare out from the facade, their blackened glass reflecting the turbulent sky, while eerie-faced stone gargoyles perch on the ledges, frozen in silent screams. The camera continues its descent, gliding smoothly across the imposing facade, ultimately landing at a low angle at the manor's ornate front gate, a towering masterpiece of black wrought iron twisted into intricate patterns of thorns, vines, and skeletons, its surface revealing a rust-colored decay. The entire scene is bathed in the cold, dim light of an impending storm, punctuated by occasional bursts of silent, sheet lightning that briefly cast harsh shadows and illuminate the textures of the damp stone, creating an atmosphere thick with foreboding, decaying grandeur, and a profound sense of isolation.", + "seed": 9543, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_12/index_8425_seed_9543_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01034165769815445, + "vlm_logprob": -0.01034165769815445, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.01034165769815445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004426917992532253 + ], + "opposite_logprob_tokens": [ + -6.004426956176758 + ], + "logprob_reward": -0.004426917992532253, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.004426917992532253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12095404416322708, + 0.04633954539895058, + 0.05324637144804001, + 0.03845507279038429, + 0.05434732884168625, + 0.07277345657348633, + 0.2348508983850479, + 0.23028650879859924, + 0.07677775621414185, + 0.07196901738643646 + ], + "credit_map": [ + 0.009745010174810886, + 0.01314444188028574, + 0.011021822690963745, + 0.005591643042862415, + 0.04593385010957718, + 0.01981310546398163, + 0.007214819546788931, + 0.008489349856972694, + 0.0032806498929858208, + 0.004131213296204805, + 0.004365372937172651, + 0.0019149862928315997, + 0.012202299199998379, + 0.014012639410793781, + 0.0034454658161848783, + 0.002986919367685914, + 0.007714900653809309, + 0.016651106998324394, + 0.0024149895180016756, + 0.002050700131803751, + 0.004782252945005894, + 0.012306738644838333, + 0.004686953965574503, + 0.0026387223042547703, + 0.0038416306488215923, + 0.00430650357156992, + 0.004378216806799173, + 0.0025980931241065264, + 0.008390909992158413, + 0.007860098965466022, + 0.004262913018465042, + 0.0028167071286588907, + 0.003727592993527651, + 0.005679742433130741, + 0.003654654137790203, + 0.00492330314591527, + 0.008287956938147545, + 0.012221232987940311, + 0.012531405314803123, + 0.003321436932310462, + 0.010584315285086632, + 0.009447729215025902, + 0.00715018343180418, + 0.006906167138367891, + 0.008730640634894371, + 0.008910026401281357, + 0.009385500103235245, + 0.011658893898129463, + 0.01723041944205761, + 0.058134399354457855, + 0.01630874164402485, + 0.02269245870411396, + 0.031425002962350845, + 0.04220559075474739, + 0.02154168300330639, + 0.025312604382634163, + 0.0202791728079319, + 0.013870932161808014, + 0.026279272511601448, + 0.02795562520623207, + 0.021008161827921867, + 0.012084158137440681, + 0.03170425817370415, + 0.07710491120815277, + 0.011944335885345936, + 0.008612006902694702, + 0.009068288840353489, + 0.007314893417060375, + 0.016072029247879982, + 0.005171024706214666, + 0.010906707495450974, + 0.00768847344443202, + 0.003976576495915651, + 0.004071742761880159, + 0.005624806974083185, + 0.00724981352686882, + 0.0058984835632145405, + 0.006222078111022711, + 0.02409646101295948, + 0.014829056337475777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.004426917992532253 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085587553679943 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.011085587553679943, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085587553679943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1194278821349144, + 0.030656689777970314, + 0.03191274031996727, + 0.02729589119553566, + 0.05230562016367912, + 0.04148401692509651, + 0.18035317957401276, + 0.28503066301345825, + 0.024992894381284714, + 0.20654042065143585 + ], + "credit_map": [ + 0.009896681644022465, + 0.014033114537596703, + 0.008043304085731506, + 0.005173755809664726, + 0.046543676406145096, + 0.020249279215931892, + 0.00738134840503335, + 0.008106717839837074, + 0.0029391238931566477, + 0.006628677248954773, + 0.004007560666650534, + 0.002214146079495549, + 0.004899311810731888, + 0.002743193181231618, + 0.004933894611895084, + 0.002290779957547784, + 0.003845648607239127, + 0.006593169178813696, + 0.0026050941087305546, + 0.0024816542863845825, + 0.0049623288214206696, + 0.003584828693419695, + 0.005520579405128956, + 0.00231943535618484, + 0.003082824172452092, + 0.0027457536198198795, + 0.0028652579057961702, + 0.0018480124417692423, + 0.005774467717856169, + 0.006262448616325855, + 0.0026888020802289248, + 0.0020283234771341085, + 0.0037383020389825106, + 0.0064672487787902355, + 0.0028519320767372847, + 0.004026195500046015, + 0.011180467903614044, + 0.011899183504283428, + 0.008735419251024723, + 0.0034068699460476637, + 0.005999892950057983, + 0.005371524021029472, + 0.00388436671346426, + 0.005326079670339823, + 0.005283541977405548, + 0.005303016863763332, + 0.004626898560672998, + 0.00568869523704052, + 0.01469817478209734, + 0.04616807401180267, + 0.013008170761168003, + 0.01792619377374649, + 0.020632538944482803, + 0.03475407138466835, + 0.013407174497842789, + 0.0197587963193655, + 0.03846177086234093, + 0.014429050497710705, + 0.02673090063035488, + 0.02999112382531166, + 0.020426340401172638, + 0.013312368653714657, + 0.03138840198516846, + 0.11029071360826492, + 0.002576688304543495, + 0.0034158267080783844, + 0.003274492919445038, + 0.0033241198398172855, + 0.0036814899649471045, + 0.0024124409537762403, + 0.0037111237179487944, + 0.0025967108085751534, + 0.010159723460674286, + 0.014189100824296474, + 0.013780257664620876, + 0.021527916193008423, + 0.017671823501586914, + 0.01243579387664795, + 0.0692145898938179, + 0.047561224550008774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.011085587553679943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945141427218914 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.005945141427218914, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945141427218914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08525285124778748, + 0.05111197382211685, + 0.1469697505235672, + 0.024198656901717186, + 0.023853644728660583, + 0.07360885292291641, + 0.3434433341026306, + 0.1822376549243927, + 0.027258096262812614, + 0.04206513613462448 + ], + "credit_map": [ + 0.0072234803810715675, + 0.010346279479563236, + 0.004706148058176041, + 0.003582349279895425, + 0.034038785845041275, + 0.014504307880997658, + 0.004686364904046059, + 0.006165143568068743, + 0.004397666547447443, + 0.008202703669667244, + 0.007975783199071884, + 0.0034343451261520386, + 0.0063911364413797855, + 0.010276401415467262, + 0.006999736186116934, + 0.0034342037979513407, + 0.023531559854745865, + 0.0526418499648571, + 0.003934173379093409, + 0.004059135913848877, + 0.01299471128731966, + 0.036963388323783875, + 0.00755962124094367, + 0.005285314749926329, + 0.0027588617522269487, + 0.0028142668306827545, + 0.0023424928076565266, + 0.0015662865480408072, + 0.005702745169401169, + 0.005338867660611868, + 0.0020140830893069506, + 0.0016610526945441961, + 0.002604009350761771, + 0.002216135384514928, + 0.002826756564900279, + 0.0027377759106457233, + 0.0032726083882153034, + 0.003965271636843681, + 0.0036689734552055597, + 0.002562115201726556, + 0.011155487969517708, + 0.009254449978470802, + 0.0075093964114785194, + 0.007242986466735601, + 0.008662547916173935, + 0.008947628550231457, + 0.009950175881385803, + 0.010886184871196747, + 0.025995319709181786, + 0.08647911995649338, + 0.025034304708242416, + 0.0337132103741169, + 0.041649676859378815, + 0.0633031353354454, + 0.02823813259601593, + 0.03903045505285263, + 0.011547284200787544, + 0.01082129031419754, + 0.02486027218401432, + 0.0243001040071249, + 0.01657646708190441, + 0.010129550471901894, + 0.029150832444429398, + 0.054851874709129333, + 0.0033893499057739973, + 0.0027581520844250917, + 0.0029423709493130445, + 0.003241020953282714, + 0.004239431116729975, + 0.001920690992847085, + 0.0051993378438055515, + 0.003567745676264167, + 0.003176238154992461, + 0.003305007703602314, + 0.003922514151781797, + 0.004648375324904919, + 0.0038785506039857864, + 0.0040902988985180855, + 0.012635793536901474, + 0.006408360321074724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.005945141427218914 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192714355885983 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192714355885983, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192714355885983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04906875640153885, + 0.029818423092365265, + 0.11332488059997559, + 0.04500285163521767, + 0.05072973296046257, + 0.07960046082735062, + 0.13042336702346802, + 0.2312110811471939, + 0.24496501684188843, + 0.02585538476705551 + ], + "credit_map": [ + 0.0041357832960784435, + 0.005682670045644045, + 0.006411869078874588, + 0.003968473523855209, + 0.013192699290812016, + 0.007142985239624977, + 0.0036706910468637943, + 0.004863585811108351, + 0.003098232438787818, + 0.007157198619097471, + 0.004536117427051067, + 0.0022628705482929945, + 0.004296541679650545, + 0.0031521341297775507, + 0.0031532715074718, + 0.0021620555780828, + 0.017160508781671524, + 0.03844159096479416, + 0.004465789068490267, + 0.003844519378617406, + 0.009732554666697979, + 0.028105512261390686, + 0.007332739420235157, + 0.004241662565618753, + 0.0040375227108597755, + 0.005717399995774031, + 0.0037727104499936104, + 0.0029873927123844624, + 0.011118649505078793, + 0.010564683936536312, + 0.0037277706433087587, + 0.0030767221469432116, + 0.0039391592144966125, + 0.00471586873754859, + 0.004927486646920443, + 0.0047266921028494835, + 0.011310973204672337, + 0.007786101195961237, + 0.008835277520120144, + 0.004488174803555012, + 0.01163776870816946, + 0.01112245861440897, + 0.008258620277047157, + 0.006272740196436644, + 0.009391643106937408, + 0.010550525039434433, + 0.010622897185385227, + 0.011743803508579731, + 0.009402673691511154, + 0.026457972824573517, + 0.010165422223508358, + 0.015598172321915627, + 0.02329760231077671, + 0.015315678901970387, + 0.01715865358710289, + 0.013027207925915718, + 0.014420033432543278, + 0.01929948665201664, + 0.027413493022322655, + 0.030105510726571083, + 0.02862376533448696, + 0.01476344931870699, + 0.03534916788339615, + 0.0612361840903759, + 0.041134312748909, + 0.026787860319018364, + 0.030059456825256348, + 0.023262422531843185, + 0.04859904572367668, + 0.016254592686891556, + 0.03407084196805954, + 0.02479648031294346, + 0.0016968684503808618, + 0.0027648364193737507, + 0.0033230597618967295, + 0.002859697910025716, + 0.0039206561632454395, + 0.00394394202157855, + 0.00396891450509429, + 0.0033774101175367832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.014192714355885983 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016057927161455154 + ], + "opposite_logprob_tokens": [ + -4.141057968139648 + ], + "logprob_reward": -0.016057927161455154, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016057927161455154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057700831443071365, + 0.02037476748228073, + 0.04949791356921196, + 0.01649542525410652, + 0.045751139521598816, + 0.03534555807709694, + 0.3751152753829956, + 0.12455858290195465, + 0.07156599313020706, + 0.20359446108341217 + ], + "credit_map": [ + 0.004542885348200798, + 0.0059850444085896015, + 0.0036336746998131275, + 0.003537632990628481, + 0.022134829312562943, + 0.008897582069039345, + 0.004047054331749678, + 0.004922129213809967, + 0.001995697384700179, + 0.004209100268781185, + 0.0037637220229953527, + 0.0015869708731770515, + 0.002053428441286087, + 0.0024265965912491083, + 0.002585108857601881, + 0.0017541443230584264, + 0.007686987519264221, + 0.017003396525979042, + 0.0022778252605348825, + 0.0017482878174632788, + 0.004580605309456587, + 0.011198969557881355, + 0.0034451449755579233, + 0.0015566996298730373, + 0.0022850779350847006, + 0.0023811254650354385, + 0.0015418257098644972, + 0.0018759380327537656, + 0.002377115422859788, + 0.002102077938616276, + 0.0020352823194116354, + 0.0018969819648191333, + 0.003674326464533806, + 0.006205765530467033, + 0.0031289844773709774, + 0.0034763943403959274, + 0.008033827878534794, + 0.010138800367712975, + 0.008219435811042786, + 0.002873603953048587, + 0.005055968649685383, + 0.004368118941783905, + 0.003665837226435542, + 0.0028380374424159527, + 0.0043277558870613575, + 0.0038033961318433285, + 0.005673221778124571, + 0.005613225046545267, + 0.03249138593673706, + 0.09450819343328476, + 0.032179512083530426, + 0.03530927747488022, + 0.03614392876625061, + 0.06732045114040375, + 0.028756476938724518, + 0.04840604588389397, + 0.013468433171510696, + 0.010563385672867298, + 0.012504366226494312, + 0.012740932404994965, + 0.01599816232919693, + 0.008023933507502079, + 0.013838245533406734, + 0.037421125918626785, + 0.009538812562823296, + 0.006981755141168833, + 0.008461394347250462, + 0.009959840215742588, + 0.009094011969864368, + 0.0043715606443583965, + 0.013813409954309464, + 0.009345210157334805, + 0.015884310007095337, + 0.019457083195447922, + 0.018202347680926323, + 0.026433318853378296, + 0.01836586557328701, + 0.019374174997210503, + 0.05576031655073166, + 0.030117060989141464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.016057927161455154 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a high-angle perspective, piercing the churning belly of a stormy sky filled with scarred charcoal-gray and deep purple clouds, roiling with unspoken fury. As the camera executes a slow, deliberate crane descent, a magnificent Gothic mansion gradually emerges, its jagged spires and towers first tearing through the low-hanging mist. The building looms, constructed of dark, weathered stone mortar, threatened by the rain, its walls partially devoured by creeping, dark-green ivy. Tall lancet windows, like hollow eyes, stare out from the facade, their blackened glass reflecting the turbulent sky, while eerie-faced stone gargoyles perch on the ledges, frozen in silent screams. The camera continues its descent, gliding smoothly across the imposing facade, ultimately landing at a low angle at the manor's ornate front gate, a towering masterpiece of black wrought iron twisted into intricate patterns of thorns, vines, and skeletons, its surface revealing a rust-colored decay. The entire scene is bathed in the cold, dim light of an impending storm, punctuated by occasional bursts of silent, sheet lightning that briefly cast harsh shadows and illuminate the textures of the damp stone, creating an atmosphere thick with foreboding, decaying grandeur, and a profound sense of isolation.", + "seed": 9544, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_12/index_8425_seed_9544_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010071518085896968, + "vlm_logprob": -0.010071518085896968, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.010071518085896968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007707263343036175 + ], + "opposite_logprob_tokens": [ + -5.632707118988037 + ], + "logprob_reward": -0.007707263343036175, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.007707263343036175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21916985511779785, + 0.05763932317495346, + 0.056191641837358475, + 0.11032034456729889, + 0.15931496024131775, + 0.12333012372255325, + 0.14001567661762238, + 0.055003393441438675, + 0.04774066433310509, + 0.0312739834189415 + ], + "credit_map": [ + 0.014486856758594513, + 0.023608364164829254, + 0.04856155067682266, + 0.04075419902801514, + 0.015478714369237423, + 0.04636227339506149, + 0.01936054602265358, + 0.010557369329035282, + 0.004264559131115675, + 0.005186763592064381, + 0.008463277481496334, + 0.0056917183101177216, + 0.005957954563200474, + 0.009598801843822002, + 0.010483309626579285, + 0.007992939092218876, + 0.004492673557251692, + 0.004773879423737526, + 0.010620142333209515, + 0.010047031566500664, + 0.003956352826207876, + 0.009672769345343113, + 0.0069349221885204315, + 0.005693871062248945, + 0.01240047812461853, + 0.010473518632352352, + 0.013552235439419746, + 0.010579907335340977, + 0.016205715015530586, + 0.01969386450946331, + 0.013571112416684628, + 0.013843516819179058, + 0.01673831231892109, + 0.03366569057106972, + 0.014695915393531322, + 0.019178414717316628, + 0.027995526790618896, + 0.012200817465782166, + 0.019898269325494766, + 0.014942025765776634, + 0.02418854646384716, + 0.01385396346449852, + 0.009251503273844719, + 0.01821230724453926, + 0.008592513389885426, + 0.012049398384988308, + 0.016575980931520462, + 0.02060592733323574, + 0.013262815773487091, + 0.027089936658740044, + 0.011666343547403812, + 0.018607771024107933, + 0.020364172756671906, + 0.016244366765022278, + 0.01895042136311531, + 0.013829858042299747, + 0.004472638014703989, + 0.005277509801089764, + 0.009277217090129852, + 0.007305737119168043, + 0.004918376915156841, + 0.0061742146499454975, + 0.007562385872006416, + 0.010015315376222134, + 0.004423895385116339, + 0.006668651942163706, + 0.006034996826201677, + 0.007995796389877796, + 0.006314082071185112, + 0.006237905006855726, + 0.005039415787905455, + 0.005025921855121851, + 0.0023225753102451563, + 0.004418646916747093, + 0.0025342688895761967, + 0.0065129585564136505, + 0.0038880701176822186, + 0.004833954386413097, + 0.0033465346787124872, + 0.0034169757273048162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.007707263343036175 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084998026490211 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.011084998026490211, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084998026490211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09858923405408859, + 0.0369810089468956, + 0.04947587102651596, + 0.07991846650838852, + 0.08965476602315903, + 0.12035099416971207, + 0.3052402436733246, + 0.037935998290777206, + 0.1004648432135582, + 0.08138857781887054 + ], + "credit_map": [ + 0.0065048811957240105, + 0.006892198231071234, + 0.022130494937300682, + 0.014146994799375534, + 0.008062274195253849, + 0.028241882100701332, + 0.006647717673331499, + 0.005962800234556198, + 0.0028053184505552053, + 0.0037997765466570854, + 0.004415963776409626, + 0.003535954747349024, + 0.005083656869828701, + 0.004994192160665989, + 0.006660234183073044, + 0.0056859166361391544, + 0.0039677987806499004, + 0.004837139043956995, + 0.006430281326174736, + 0.007160812616348267, + 0.006789376027882099, + 0.006380586884915829, + 0.008347601629793644, + 0.005562275182455778, + 0.008547280915081501, + 0.006089114584028721, + 0.00943051278591156, + 0.010342895053327084, + 0.00806450005620718, + 0.012864750809967518, + 0.01220824383199215, + 0.012371174059808254, + 0.00866398960351944, + 0.015995189547538757, + 0.008454481139779091, + 0.009932558052241802, + 0.015198673121631145, + 0.007870578207075596, + 0.014894937165081501, + 0.008644365705549717, + 0.011541183106601238, + 0.021048514172434807, + 0.007743945810943842, + 0.017586365342140198, + 0.023093150928616524, + 0.010996052995324135, + 0.016541004180908203, + 0.011800790205597878, + 0.03550753369927406, + 0.017378782853484154, + 0.02519865892827511, + 0.054191891103982925, + 0.01741699129343033, + 0.018547197803854942, + 0.09178057312965393, + 0.04521859809756279, + 0.0026089942548424006, + 0.003939123824238777, + 0.0031332899816334248, + 0.006469921208918095, + 0.0059618866071105, + 0.004679033998399973, + 0.006389884743839502, + 0.0047538671642541885, + 0.0076128290966153145, + 0.014310573227703571, + 0.01046891137957573, + 0.01912592351436615, + 0.015391524881124496, + 0.01584395207464695, + 0.008731456473469734, + 0.0089796781539917, + 0.004745047073811293, + 0.011732246726751328, + 0.006602409295737743, + 0.013800488784909248, + 0.013036026619374752, + 0.014536915346980095, + 0.008957667276263237, + 0.007977781817317009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.011084998026490211 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613373264670372 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.003613373264670372, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613373264670372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24775168299674988, + 0.05340442806482315, + 0.09068898111581802, + 0.09977038949728012, + 0.09434167295694351, + 0.04859621450304985, + 0.1501493602991104, + 0.035822559148073196, + 0.11955107748508453, + 0.05992366373538971 + ], + "credit_map": [ + 0.015574722550809383, + 0.019046803936362267, + 0.03924083337187767, + 0.021276069805026054, + 0.034192126244306564, + 0.0728483498096466, + 0.02928011864423752, + 0.016292652115225792, + 0.003740512067452073, + 0.009267859160900116, + 0.00652105500921607, + 0.004844190087169409, + 0.007208108901977539, + 0.006588444579392672, + 0.009340937249362469, + 0.005893320310860872, + 0.006005850620567799, + 0.010714694857597351, + 0.018959233537316322, + 0.010285262018442154, + 0.015073518268764019, + 0.0141210425645113, + 0.008084552362561226, + 0.00744482921436429, + 0.010422594845294952, + 0.01223006658256054, + 0.011981423012912273, + 0.010359097272157669, + 0.015958748757839203, + 0.01309243869036436, + 0.010986906476318836, + 0.01473911665380001, + 0.009353247471153736, + 0.014877703040838242, + 0.008978137746453285, + 0.011353012174367905, + 0.016298139467835426, + 0.008407620713114738, + 0.015069056302309036, + 0.010004754178225994, + 0.009652669541537762, + 0.006906370632350445, + 0.004203916061669588, + 0.006075773388147354, + 0.004713150672614574, + 0.004971885588020086, + 0.005702611058950424, + 0.0063698370940983295, + 0.017856892198324203, + 0.005041114520281553, + 0.01102998573333025, + 0.028709182515740395, + 0.007756417617201805, + 0.009045680984854698, + 0.0479418970644474, + 0.022768182680010796, + 0.002547857351601124, + 0.0035790333058685064, + 0.0025715711526572704, + 0.007003590930253267, + 0.005056740716099739, + 0.005403191316872835, + 0.004291656892746687, + 0.0053689186461269855, + 0.011765366420149803, + 0.012318601831793785, + 0.012246853671967983, + 0.022863954305648804, + 0.018978819251060486, + 0.01446507778018713, + 0.0143762631341815, + 0.012536137364804745, + 0.004993453621864319, + 0.007482149172574282, + 0.007845824584364891, + 0.012595652602612972, + 0.0070579443126916885, + 0.009004911407828331, + 0.005454522557556629, + 0.005489202681928873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.003613373264670372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779401123523712 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.009779401123523712, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779401123523712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2531016170978546, + 0.03222197666764259, + 0.055041443556547165, + 0.041509147733449936, + 0.06336291879415512, + 0.1468198448419571, + 0.3251951336860657, + 0.023557869717478752, + 0.04570051282644272, + 0.013489493168890476 + ], + "credit_map": [ + 0.018326139077544212, + 0.01467062532901764, + 0.04325368255376816, + 0.017224770039319992, + 0.03895982727408409, + 0.08062872290611267, + 0.02524578385055065, + 0.014792080037295818, + 0.0025451509281992912, + 0.0036385359708219767, + 0.003139158710837364, + 0.002911640563979745, + 0.004078974947333336, + 0.0040910555981099606, + 0.007549710106104612, + 0.004267747979611158, + 0.0033953385427594185, + 0.0036551624070852995, + 0.0154128298163414, + 0.005633816588670015, + 0.004326252732425928, + 0.011520125903189182, + 0.00478504691272974, + 0.006312875542789698, + 0.00709177739918232, + 0.0030931495130062103, + 0.0066194538958370686, + 0.003621046431362629, + 0.004569109529256821, + 0.0043143401853740215, + 0.004797973204404116, + 0.007402294781059027, + 0.006743740290403366, + 0.01624220795929432, + 0.004092863295227289, + 0.0057497876696288586, + 0.013187022879719734, + 0.004459514748305082, + 0.008140850812196732, + 0.004746933467686176, + 0.0146949402987957, + 0.02846788614988327, + 0.00918542593717575, + 0.019204745069146156, + 0.028378233313560486, + 0.013220097869634628, + 0.019181998446583748, + 0.014486514031887054, + 0.032565124332904816, + 0.05373686924576759, + 0.017532754689455032, + 0.04959118738770485, + 0.04052400588989258, + 0.03306726738810539, + 0.07081080228090286, + 0.027367131784558296, + 0.0017782411305233836, + 0.002091736998409033, + 0.00270170159637928, + 0.003712278325110674, + 0.002111193258315325, + 0.002672256901860237, + 0.0047417497262358665, + 0.0037487109657377005, + 0.005206120200455189, + 0.004854200873523951, + 0.006917954888194799, + 0.0072273313999176025, + 0.006053612567484379, + 0.005387449637055397, + 0.005263906437903643, + 0.004789934493601322, + 0.0012994169956073165, + 0.0018080959562212229, + 0.0011169039644300938, + 0.002496688859537244, + 0.0018229283159598708, + 0.0023417039774358273, + 0.0013022104976698756, + 0.001301544369198382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.009779401123523712 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Gothic mansion with jagged spires and towers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the building constructed of dark, weathered stone with creeping dark-green ivy on the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a stormy sky with charcoal-gray and deep purple clouds and sheet lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle perspective and execute a slow crane descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera end at a low angle focusing on an ornate black wrought iron gate with rust-colored decay?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018172554671764374 + ], + "opposite_logprob_tokens": [ + -4.018172740936279 + ], + "logprob_reward": -0.018172554671764374, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018172554671764374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04226196929812431, + 0.015720272436738014, + 0.025218011811375618, + 0.0264054536819458, + 0.020516371354460716, + 0.06498824805021286, + 0.5493072271347046, + 0.02540062740445137, + 0.20083585381507874, + 0.02934597246348858 + ], + "credit_map": [ + 0.003168958006426692, + 0.005331960506737232, + 0.008067824877798557, + 0.008542990311980247, + 0.0033741514198482037, + 0.0068776607513427734, + 0.004370912443846464, + 0.002527509815990925, + 0.0015108442166820168, + 0.0014649134827777743, + 0.0018898840062320232, + 0.001594196306541562, + 0.0017989546759054065, + 0.0021645298693329096, + 0.0034278251696377993, + 0.001869124243967235, + 0.0014201037120074034, + 0.0016789737856015563, + 0.007348131854087114, + 0.002672637114301324, + 0.0029856404289603233, + 0.004716519266366959, + 0.002929168287664652, + 0.0014668385265395045, + 0.003342477837577462, + 0.002386997686699033, + 0.0027155254501849413, + 0.003207844914868474, + 0.0027016487438231707, + 0.0046831825748085976, + 0.003505873493850231, + 0.003861901583150029, + 0.002575905993580818, + 0.0044573284685611725, + 0.0018014856614172459, + 0.0017666418571025133, + 0.003976619802415371, + 0.0016622398979961872, + 0.0027596279978752136, + 0.0015165233053267002, + 0.008429432287812233, + 0.009595640935003757, + 0.00410411087796092, + 0.009363149292767048, + 0.011523939669132233, + 0.0054345098324120045, + 0.009580862708389759, + 0.006956601049751043, + 0.06381580978631973, + 0.03985808789730072, + 0.039894990622997284, + 0.0955701544880867, + 0.03575626015663147, + 0.0347207672894001, + 0.16309767961502075, + 0.07659344375133514, + 0.0018909081118181348, + 0.002871075877919793, + 0.0019293786026537418, + 0.004284824710339308, + 0.0033693714067339897, + 0.0028973831795156, + 0.005484872497618198, + 0.002672812668606639, + 0.02075367234647274, + 0.02158239483833313, + 0.02405448444187641, + 0.03560176119208336, + 0.024726584553718567, + 0.025234337896108627, + 0.027228662744164467, + 0.02165394462645054, + 0.0027779226657003164, + 0.004041519481688738, + 0.003705007489770651, + 0.0042890398763120174, + 0.004646613262593746, + 0.003822754370048642, + 0.0029099327512085438, + 0.0031531830318272114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.018172554671764374 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_13/index_tensor([3953, 3953])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_13/index_tensor([3953, 3953])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dff2c6cb199649750d6c537c195a8d6afba8e94c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_13/index_tensor([3953, 3953])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level wide-angle shot captures the vast, eerie emptiness of a nighttime indoor swimming pool, creating a visceral, suspenseful atmosphere. The Olympic-sized pool, filled with unnatural turquoise water, its surface a perfect mirror, reflects the cold, even glow of industrial fluorescent lights mounted high in the cavernous ceiling. The pool is tiled with pale blue tiles, dark lane lines stretching into the gloom of the deep end, while the surrounding wet concrete deck lies empty. Suddenly, without warning, the overhead lights begin to flicker erratically, a violent, chaotic stutter that plunges the entire scene into jarring moments of absolute darkness. Each plunge into darkness is perfectly synchronized with a deep, resonant thunder—a visceral, low-frequency rumble that seems to shake the entire room, causing subtle vibrations in the camera and temporarily disturbing the tranquil surface of the water. This sensory assault shatters the tense silence, creating a claustrophobic feeling of disorientation and deep unease.", + "seed": 11252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_13/index_3953_seed_11252_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19464323222637175, + "vlm_logprob": -0.19464323222637175, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.19464323222637175, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008774415589869022 + ], + "opposite_logprob_tokens": [ + -5.883774280548096 + ], + "logprob_reward": -0.008774415589869022, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.008774415589869022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09471962600946426, + 0.07844487577676773, + 0.032738883048295975, + 0.047544319182634354, + 0.04477856308221817, + 0.045871712267398834, + 0.05893620848655701, + 0.39164164662361145, + 0.10236239433288574, + 0.1029617190361023 + ], + "credit_map": [ + 0.009419605135917664, + 0.015169434249401093, + 0.011271001771092415, + 0.012955433689057827, + 0.012126398272812366, + 0.014796237461268902, + 0.0095972316339612, + 0.009384281001985073, + 0.005454779602587223, + 0.010049751028418541, + 0.007947253063321114, + 0.01651657000184059, + 0.008706772699952126, + 0.00900507252663374, + 0.00861981138586998, + 0.012144866399466991, + 0.003060078714042902, + 0.0036918397527188063, + 0.003973342012614012, + 0.006117044482380152, + 0.003611119231209159, + 0.0037915075663477182, + 0.003933016210794449, + 0.0045609320513904095, + 0.004990919027477503, + 0.006704389583319426, + 0.003883874509483576, + 0.007574788760393858, + 0.004590771626681089, + 0.007613732013851404, + 0.006815149448812008, + 0.00537069421261549, + 0.002961221383884549, + 0.0053695994429290295, + 0.005252030678093433, + 0.009773763827979565, + 0.004606227390468121, + 0.00541725242510438, + 0.005893381778150797, + 0.005505089182406664, + 0.0031043076887726784, + 0.006132190115749836, + 0.0048728762194514275, + 0.009971944615244865, + 0.003600791562348604, + 0.004878039937466383, + 0.007292883470654488, + 0.00601867726072669, + 0.005852638743817806, + 0.0063044968992471695, + 0.005565950646996498, + 0.008455854840576649, + 0.00773404398933053, + 0.008470102213323116, + 0.010268843732774258, + 0.006284274626523256, + 0.05910620093345642, + 0.011999533511698246, + 0.018289459869265556, + 0.020544646307826042, + 0.03424893692135811, + 0.14473406970500946, + 0.08226463943719864, + 0.020454181358218193, + 0.014727029949426651, + 0.009643889032304287, + 0.011593601666390896, + 0.015513204969465733, + 0.012247845530509949, + 0.012873179279267788, + 0.016974637284874916, + 0.0087890038266778, + 0.015853455290198326, + 0.009624175727367401, + 0.013953154906630516, + 0.012753813527524471, + 0.012957247905433178, + 0.013021852821111679, + 0.009773172438144684, + 0.015024842694401741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.008774415589869022 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206979602575302 + ], + "opposite_logprob_tokens": [ + -2.812069892883301 + ], + "logprob_reward": -0.06206979602575302, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206979602575302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08683297783136368, + 0.07803063839673996, + 0.028111813589930534, + 0.02225988358259201, + 0.03324571251869202, + 0.027461767196655273, + 0.02383478172123432, + 0.32773539423942566, + 0.26897621154785156, + 0.1035107970237732 + ], + "credit_map": [ + 0.010059427469968796, + 0.01130648609250784, + 0.009157028049230576, + 0.010819969698786736, + 0.013812897726893425, + 0.013074863702058792, + 0.009509949944913387, + 0.009092356078326702, + 0.006166227627545595, + 0.009604696184396744, + 0.007584518752992153, + 0.015373315662145615, + 0.007911654189229012, + 0.008381915278732777, + 0.011743237264454365, + 0.011265077628195286, + 0.0032051645684987307, + 0.00308176688849926, + 0.0034902459010481834, + 0.004115748219192028, + 0.0036625219509005547, + 0.004032983910292387, + 0.0036541956942528486, + 0.0028691869229078293, + 0.00266889575868845, + 0.002433579880744219, + 0.0019897695165127516, + 0.0024564587511122227, + 0.0027714543975889683, + 0.0034404746256768703, + 0.004116323310881853, + 0.002382925944402814, + 0.003643674310296774, + 0.0043004644103348255, + 0.003217835910618305, + 0.004962535575032234, + 0.004827079828828573, + 0.004369922913610935, + 0.003942397888749838, + 0.0039818002842366695, + 0.002118393313139677, + 0.003441037843003869, + 0.0034017174039036036, + 0.004761161282658577, + 0.0023618547711521387, + 0.0032914818730205297, + 0.003056308953091502, + 0.0050298115238547325, + 0.0017937442753463984, + 0.003048430662602186, + 0.0025579696521162987, + 0.004642576444894075, + 0.002620201325044036, + 0.002968236105516553, + 0.0035743904300034046, + 0.0026292309630662203, + 0.04012829810380936, + 0.009823991917073727, + 0.027561446651816368, + 0.02281903848052025, + 0.02252098359167576, + 0.08977722376585007, + 0.0972336083650589, + 0.017870815470814705, + 0.037208616733551025, + 0.018269523978233337, + 0.03543096408247948, + 0.034893572330474854, + 0.029566971585154533, + 0.028581205755472183, + 0.03994697704911232, + 0.04507838562130928, + 0.016698244959115982, + 0.008466651663184166, + 0.011043556965887547, + 0.011417879723012447, + 0.012644890695810318, + 0.016788793727755547, + 0.013167298398911953, + 0.01328347623348236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.06206979602575302 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261144161224365 + ], + "opposite_logprob_tokens": [ + -0.5761144161224365 + ], + "logprob_reward": -0.8261144161224365, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261144161224365, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.036729197949171066, + 0.013424943201243877, + 0.016629617661237717, + 0.018247587606310844, + 0.01578109711408615, + 0.019382556900382042, + 0.024621769785881042, + 0.7072530388832092, + 0.06592820584774017, + 0.08200207352638245 + ], + "credit_map": [ + 0.0041441242210567, + 0.005193204153329134, + 0.0036910271737724543, + 0.00355852791108191, + 0.004686291329562664, + 0.00438218517228961, + 0.007269695866852999, + 0.0038041407242417336, + 0.0010866717202588916, + 0.0016548536950722337, + 0.0014431464951485395, + 0.002169928979128599, + 0.0016894089058041573, + 0.001749369315803051, + 0.0018565310165286064, + 0.0017750305123627186, + 0.0014633691171184182, + 0.0023913956247270107, + 0.001660530106164515, + 0.002804284216836095, + 0.002060983330011368, + 0.0023145992308855057, + 0.0020792100112885237, + 0.0018552441615611315, + 0.0016128490678966045, + 0.0022812392562627792, + 0.0017531884368509054, + 0.0025075492449104786, + 0.0032557968515902758, + 0.0019051857525482774, + 0.003170258365571499, + 0.0017615160904824734, + 0.0014802864752709866, + 0.0021938078571110964, + 0.0016799371223896742, + 0.002449319465085864, + 0.0015050377696752548, + 0.002043450018391013, + 0.0019529415294528008, + 0.0024763161782175303, + 0.0016429862007498741, + 0.002757728798314929, + 0.002169254934415221, + 0.0032366642262786627, + 0.002050983952358365, + 0.002347630215808749, + 0.0023057127837091684, + 0.00287159183062613, + 0.0019617548678070307, + 0.003315516049042344, + 0.0030048068147152662, + 0.004449170082807541, + 0.0021927570924162865, + 0.002871366683393717, + 0.003507800167426467, + 0.0033185931388288736, + 0.0787055641412735, + 0.023755887523293495, + 0.06961584836244583, + 0.05186585336923599, + 0.06156807020306587, + 0.16165128350257874, + 0.22197377681732178, + 0.03811665251851082, + 0.007749371696263552, + 0.004293125122785568, + 0.008491613902151585, + 0.010733602568507195, + 0.0062488107942044735, + 0.0089812520891428, + 0.0077196466736495495, + 0.01171076949685812, + 0.011135580018162727, + 0.005777688231319189, + 0.009435846470296383, + 0.01184913795441389, + 0.008346066810190678, + 0.015291751362383366, + 0.011723790317773819, + 0.008442193269729614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.8261144161224365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206822767853737 + ], + "opposite_logprob_tokens": [ + -2.812068223953247 + ], + "logprob_reward": -0.06206822767853737, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206822767853737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15477026998996735, + 0.030181890353560448, + 0.04083374887704849, + 0.03490853309631348, + 0.025264745578169823, + 0.01712365634739399, + 0.028178278356790543, + 0.5462110638618469, + 0.07034764438867569, + 0.05218019336462021 + ], + "credit_map": [ + 0.012625998817384243, + 0.02220076322555542, + 0.016382796689867973, + 0.03098975121974945, + 0.014465407468378544, + 0.015924232080578804, + 0.019380975514650345, + 0.02280033752322197, + 0.00310406438075006, + 0.004009647760540247, + 0.003610387910157442, + 0.003880258882418275, + 0.003302685683593154, + 0.004180144052952528, + 0.0047512780874967575, + 0.0033434228971600533, + 0.003959312103688717, + 0.005329933483153582, + 0.0037273175548762083, + 0.009129645302891731, + 0.004245575983077288, + 0.004930805414915085, + 0.003942684270441532, + 0.005568475928157568, + 0.0043161483481526375, + 0.0045985993929207325, + 0.003339751623570919, + 0.0037978170439600945, + 0.003749113529920578, + 0.0067804050631821156, + 0.004870476201176643, + 0.0034562197979539633, + 0.002413459587842226, + 0.003265440696850419, + 0.002448761835694313, + 0.0048193661496043205, + 0.0027003404684364796, + 0.003231553826481104, + 0.0027257453184574842, + 0.00366007792763412, + 0.0016155252233147621, + 0.002081601181998849, + 0.002095802454277873, + 0.00242699240334332, + 0.0016681983834132552, + 0.0021226657554507256, + 0.002398837823420763, + 0.0027140332385897636, + 0.0026643644087016582, + 0.003938858397305012, + 0.0030334910843521357, + 0.003958267625421286, + 0.002557844389230013, + 0.00431803148239851, + 0.003881257027387619, + 0.0038261637091636658, + 0.07042083889245987, + 0.018891580402851105, + 0.04220649600028992, + 0.036144379526376724, + 0.03970928117632866, + 0.1579381376504898, + 0.153410866856575, + 0.027489494532346725, + 0.008835460990667343, + 0.0055502657778561115, + 0.009392280131578445, + 0.008344659581780434, + 0.006219248287379742, + 0.011919603683054447, + 0.009958872571587563, + 0.010127256624400616, + 0.006854323670268059, + 0.006381988991051912, + 0.0066688451915979385, + 0.005825227126479149, + 0.006237169727683067, + 0.006350050214678049, + 0.006876134313642979, + 0.006986456457525492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.06206822767853737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418930571526289 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.01418930571526289, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418930571526289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20071075856685638, + 0.030809441581368446, + 0.067292720079422, + 0.06672508269548416, + 0.10246057063341141, + 0.04361165687441826, + 0.020898612216114998, + 0.2593054473400116, + 0.09211596846580505, + 0.11606985330581665 + ], + "credit_map": [ + 0.018884560093283653, + 0.033598560839891434, + 0.02084144949913025, + 0.02825896628201008, + 0.02397807687520981, + 0.028797760605812073, + 0.019532592967152596, + 0.026818772777915, + 0.0025165278930217028, + 0.005435157101601362, + 0.0029585673473775387, + 0.004013017751276493, + 0.0033668314572423697, + 0.004133137408643961, + 0.0031599111389368773, + 0.005226289853453636, + 0.005818815436214209, + 0.008860831148922443, + 0.007583723869174719, + 0.012274667620658875, + 0.005977344233542681, + 0.008560722693800926, + 0.008912181481719017, + 0.009304425679147243, + 0.006047114264219999, + 0.008920976892113686, + 0.0072138505056500435, + 0.011032315902411938, + 0.007081619929522276, + 0.007188024930655956, + 0.010764779523015022, + 0.00847639236599207, + 0.00857841782271862, + 0.014472735114395618, + 0.011984240263700485, + 0.015455489978194237, + 0.009882066398859024, + 0.011145371943712234, + 0.01402073074132204, + 0.0169215090572834, + 0.003404192626476288, + 0.005485464818775654, + 0.004940485581755638, + 0.009539682418107986, + 0.004105843137949705, + 0.00447445223107934, + 0.006755144335329533, + 0.004906387068331242, + 0.002465336350724101, + 0.0027506789192557335, + 0.0020461922977119684, + 0.0023380948696285486, + 0.0026406203396618366, + 0.0025367094203829765, + 0.004183205775916576, + 0.0019377723801881075, + 0.028555480763316154, + 0.009004996158182621, + 0.02475467324256897, + 0.01969854161143303, + 0.020783357322216034, + 0.06017352640628815, + 0.08183296769857407, + 0.014501887373626232, + 0.019013777375221252, + 0.007898837327957153, + 0.009916865266859531, + 0.010339695028960705, + 0.012224748730659485, + 0.014601391740143299, + 0.0076302457600831985, + 0.010490396991372108, + 0.014615779742598534, + 0.012273256666958332, + 0.013915250077843666, + 0.015204083174467087, + 0.013935319148004055, + 0.01569635607302189, + 0.014505534432828426, + 0.01592426374554634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.01418930571526289 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures the vast, eerie emptiness of a nighttime indoor swimming pool, creating a visceral, suspenseful atmosphere. The Olympic-sized pool, filled with unnatural turquoise water, its surface a perfect mirror, reflects the cold, even glow of industrial fluorescent lights mounted high in the cavernous ceiling. The pool is tiled with pale blue tiles, dark lane lines stretching into the gloom of the deep end, while the surrounding wet concrete deck lies empty. Suddenly, without warning, the overhead lights begin to flicker erratically, a violent, chaotic stutter that plunges the entire scene into jarring moments of absolute darkness. Each plunge into darkness is perfectly synchronized with a deep, resonant thunder—a visceral, low-frequency rumble that seems to shake the entire room, causing subtle vibrations in the camera and temporarily disturbing the tranquil surface of the water. This sensory assault shatters the tense silence, creating a claustrophobic feeling of disorientation and deep unease.", + "seed": 11253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_13/index_3953_seed_11253_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21401226222515107, + "vlm_logprob": -0.21401226222515107, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.21401226222515107, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013560089282691479 + ], + "opposite_logprob_tokens": [ + -6.0135602951049805 + ], + "logprob_reward": -0.013560089282691479, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.013560089282691479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3216554820537567, + 0.04479455575346947, + 0.05460486561059952, + 0.06007470190525055, + 0.21367549896240234, + 0.062117189168930054, + 0.04243304580450058, + 0.025308413431048393, + 0.12202317267656326, + 0.053313106298446655 + ], + "credit_map": [ + 0.06475327163934708, + 0.018581414595246315, + 0.01498543843626976, + 0.048672497272491455, + 0.019329501315951347, + 0.01652907393872738, + 0.10805221647024155, + 0.03075205162167549, + 0.007341302465647459, + 0.0042149657383561134, + 0.004875179845839739, + 0.004885263275355101, + 0.004456505179405212, + 0.006162820849567652, + 0.008188004605472088, + 0.004670512862503529, + 0.009574787691235542, + 0.0028951261192560196, + 0.0033957550767809153, + 0.008363435976207256, + 0.0037605396937578917, + 0.0031189676374197006, + 0.01829867623746395, + 0.005197575781494379, + 0.012450583279132843, + 0.003284142119809985, + 0.0036404812708497047, + 0.00826813094317913, + 0.0038670895155519247, + 0.0029311974067240953, + 0.019106511026620865, + 0.006526571698486805, + 0.039283763617277145, + 0.01623881421983242, + 0.011320902965962887, + 0.028347086161375046, + 0.01445682067424059, + 0.014694172888994217, + 0.06769518554210663, + 0.021638739854097366, + 0.008248599246144295, + 0.005765133071690798, + 0.003595533547922969, + 0.009567664004862309, + 0.006045087240636349, + 0.005207114852964878, + 0.01791643723845482, + 0.005771618336439133, + 0.00828507449477911, + 0.0064175729639828205, + 0.0039649889804422855, + 0.0038259646389633417, + 0.005118097178637981, + 0.005152887664735317, + 0.0048830825835466385, + 0.004785376600921154, + 0.003354673972353339, + 0.004090746399015188, + 0.0024457774125039577, + 0.002508033998310566, + 0.004093995317816734, + 0.0026111810002475977, + 0.0038989675231277943, + 0.0023050401359796524, + 0.022025568410754204, + 0.0194291640073061, + 0.007913060486316681, + 0.007766342256218195, + 0.02739707939326763, + 0.02032092958688736, + 0.008275464177131653, + 0.008895572274923325, + 0.006767221726477146, + 0.007870471104979515, + 0.005978223867714405, + 0.00645851157605648, + 0.007817485369741917, + 0.0069977035745978355, + 0.006626739166676998, + 0.004796752706170082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.013560089282691479 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890510231256485 + ], + "opposite_logprob_tokens": [ + -2.4640510082244873 + ], + "logprob_reward": -0.0890510231256485, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890510231256485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17157939076423645, + 0.12958724796772003, + 0.22951914370059967, + 0.108819879591465, + 0.08786232024431229, + 0.021480904892086983, + 0.08913961797952652, + 0.03523777797818184, + 0.07233166694641113, + 0.05444207787513733 + ], + "credit_map": [ + 0.03683483973145485, + 0.012167012318968773, + 0.010118584148585796, + 0.025323616340756416, + 0.014955336228013039, + 0.010325689800083637, + 0.04750797897577286, + 0.01434633880853653, + 0.01627149060368538, + 0.023742709308862686, + 0.014525871723890305, + 0.01212275493890047, + 0.018361199647188187, + 0.0223393514752388, + 0.012791326269507408, + 0.009432541206479073, + 0.04580714553594589, + 0.010606865398585796, + 0.011961699463427067, + 0.03571948781609535, + 0.012626776471734047, + 0.009160250425338745, + 0.08052662014961243, + 0.023110302165150642, + 0.021816886961460114, + 0.005682071205228567, + 0.00542382150888443, + 0.015233240090310574, + 0.006520440801978111, + 0.004585978575050831, + 0.0378439761698246, + 0.01171346940100193, + 0.017180467024445534, + 0.004694495815783739, + 0.0049020336009562016, + 0.012281217612326145, + 0.005396014545112848, + 0.004866385366767645, + 0.029056621715426445, + 0.009485079906880856, + 0.0036957392003387213, + 0.0032974870409816504, + 0.0025955219753086567, + 0.0035315128043293953, + 0.0019193128682672977, + 0.0017332545248791575, + 0.002724318765103817, + 0.0019837573636323214, + 0.015020324848592281, + 0.00946445856243372, + 0.0072653042152523994, + 0.01204105094075203, + 0.010701077990233898, + 0.007856528274714947, + 0.01926574669778347, + 0.007525128312408924, + 0.004655591212213039, + 0.004116866271942854, + 0.004514781292527914, + 0.004530422855168581, + 0.0048758708871901035, + 0.003902269061654806, + 0.004896730184555054, + 0.0037452485412359238, + 0.012724264524877071, + 0.00877415668219328, + 0.00478141475468874, + 0.008423076942563057, + 0.011789127252995968, + 0.010581680573523045, + 0.008709470741450787, + 0.0065484726801514626, + 0.009574750438332558, + 0.008248893544077873, + 0.003960159607231617, + 0.005719177890568972, + 0.006490249652415514, + 0.010477062314748764, + 0.0049885972402989864, + 0.004983187653124332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0890510231256485 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8982995748519897 + ], + "opposite_logprob_tokens": [ + -0.5232995748519897 + ], + "logprob_reward": -0.8982995748519897, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8982995748519897, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4779432415962219, + 0.033004872500896454, + 0.03436458483338356, + 0.10776316374540329, + 0.060637857764959335, + 0.040227919816970825, + 0.029680723324418068, + 0.02325291559100151, + 0.03888227790594101, + 0.1542424112558365 + ], + "credit_map": [ + 0.09691116213798523, + 0.022737132385373116, + 0.021523965522646904, + 0.07134947925806046, + 0.027393195778131485, + 0.022738313302397728, + 0.16555193066596985, + 0.049738094210624695, + 0.00396021269261837, + 0.0047331745736300945, + 0.002955046249553561, + 0.0037484299391508102, + 0.004484970588237047, + 0.003390737809240818, + 0.006960045080631971, + 0.002772255800664425, + 0.006513924337923527, + 0.002228563418611884, + 0.0020289216190576553, + 0.004783636890351772, + 0.0031401501037180424, + 0.002429401036351919, + 0.01026616245508194, + 0.002973827300593257, + 0.02033572643995285, + 0.005439684726297855, + 0.005601411685347557, + 0.014966906048357487, + 0.006100453902035952, + 0.004438724368810654, + 0.037879180163145065, + 0.013001088984310627, + 0.01074974611401558, + 0.0055630444549024105, + 0.0033594267442822456, + 0.007557370699942112, + 0.004680032376199961, + 0.005021045450121164, + 0.01791095919907093, + 0.0057962373830378056, + 0.005454044323414564, + 0.0029872367158532143, + 0.002012417186051607, + 0.0065567344427108765, + 0.003559119999408722, + 0.00347775686532259, + 0.0124067272990942, + 0.0037738848477602005, + 0.00538249546661973, + 0.004321116488426924, + 0.00234963302500546, + 0.003032559994608164, + 0.002892580348998308, + 0.002470343606546521, + 0.00654916325584054, + 0.002682835329324007, + 0.003139680251479149, + 0.0035050311125814915, + 0.0020857963245362043, + 0.0021074428223073483, + 0.004291060380637646, + 0.0034313586074858904, + 0.002421742305159569, + 0.0022708054166287184, + 0.00663453945890069, + 0.005199293605983257, + 0.002833643229678273, + 0.003965259995311499, + 0.005733632016927004, + 0.005842011421918869, + 0.004352000541985035, + 0.004321900196373463, + 0.019953545182943344, + 0.020661361515522003, + 0.017665205523371696, + 0.01920311711728573, + 0.021425088867545128, + 0.019159862771630287, + 0.019972221925854683, + 0.01620202511548996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.8982995748519897 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496637150645256 + ], + "opposite_logprob_tokens": [ + -2.929966449737549 + ], + "logprob_reward": -0.05496637150645256, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496637150645256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3174566328525543, + 0.028612973168492317, + 0.23226548731327057, + 0.03621426969766617, + 0.14991292357444763, + 0.0946597084403038, + 0.0105811832472682, + 0.011580906808376312, + 0.059067051857709885, + 0.059648822993040085 + ], + "credit_map": [ + 0.06561360508203506, + 0.021359790116548538, + 0.014453641138970852, + 0.044176846742630005, + 0.022920958697795868, + 0.017663801088929176, + 0.1018630787730217, + 0.029404932633042336, + 0.005336404778063297, + 0.0024807490408420563, + 0.003108302131295204, + 0.0030489666387438774, + 0.0029985213186591864, + 0.002867052098736167, + 0.0058250343427062035, + 0.002947944914922118, + 0.04645291343331337, + 0.010797967202961445, + 0.011659199371933937, + 0.034607429057359695, + 0.013598454184830189, + 0.011728348210453987, + 0.08032284677028656, + 0.02309834212064743, + 0.005018504802137613, + 0.003627597354352474, + 0.002326914109289646, + 0.004953986965119839, + 0.0025688682217150927, + 0.003524293191730976, + 0.009789151139557362, + 0.004404955077916384, + 0.02602771669626236, + 0.00969582051038742, + 0.010776473209261894, + 0.020219964906573296, + 0.01033374760299921, + 0.010301146656274796, + 0.04741370305418968, + 0.0151443500071764, + 0.015886420384049416, + 0.004594910889863968, + 0.004399572033435106, + 0.015285388566553593, + 0.0066762520000338554, + 0.004335812292993069, + 0.03284377604722977, + 0.010637572035193443, + 0.0015303690452128649, + 0.0012282775714993477, + 0.0010720603168010712, + 0.0008684430504217744, + 0.00169316993560642, + 0.0012438609264791012, + 0.0014556216774508357, + 0.001489381305873394, + 0.0016383685870096087, + 0.0019412977853789926, + 0.0011418602662160993, + 0.0010756098199635744, + 0.002222381765022874, + 0.001577307702973485, + 0.0010307385819032788, + 0.0009533423581160605, + 0.009624181315302849, + 0.008721486665308475, + 0.004305236507207155, + 0.003881748067215085, + 0.014408973976969719, + 0.010744809173047543, + 0.0037892807740718126, + 0.003591336542740464, + 0.007509617600589991, + 0.0077803852036595345, + 0.007532636169344187, + 0.007567515596747398, + 0.008588663302361965, + 0.00782858021557331, + 0.008031862787902355, + 0.00480956444516778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.05496637150645256 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014184252358973026 + ], + "opposite_logprob_tokens": [ + -4.264184474945068 + ], + "logprob_reward": -0.014184252358973026, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014184252358973026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2800307273864746, + 0.1647690385580063, + 0.03408455848693848, + 0.0562080442905426, + 0.09923149645328522, + 0.10732322186231613, + 0.069241002202034, + 0.0394827276468277, + 0.0743652805685997, + 0.0752638727426529 + ], + "credit_map": [ + 0.038697805255651474, + 0.03177643567323685, + 0.017653755843639374, + 0.03460339829325676, + 0.02532629668712616, + 0.026355020701885223, + 0.07556845247745514, + 0.030049555003643036, + 0.032024282962083817, + 0.011202193796634674, + 0.01034348364919424, + 0.02335103042423725, + 0.009957064874470234, + 0.007597127929329872, + 0.05236482247710228, + 0.01792903244495392, + 0.0064519732259213924, + 0.003441833658143878, + 0.004023021087050438, + 0.004861845634877682, + 0.00359887327067554, + 0.003342116717249155, + 0.004740037489682436, + 0.0036248585674911737, + 0.011349211446940899, + 0.003946069628000259, + 0.003848830470815301, + 0.007510860916227102, + 0.0040012551471591, + 0.0028586597181856632, + 0.016893668100237846, + 0.005799496546387672, + 0.016316231340169907, + 0.007812219206243753, + 0.006434140261262655, + 0.013591679744422436, + 0.007266289088875055, + 0.006397397723048925, + 0.031011218205094337, + 0.010402328334748745, + 0.014720000326633453, + 0.009436131455004215, + 0.00617807125672698, + 0.018246782943606377, + 0.008490819483995438, + 0.006201347801834345, + 0.03338790684938431, + 0.010662168264389038, + 0.011693472974002361, + 0.00820282194763422, + 0.005697376560419798, + 0.008573874831199646, + 0.00842142105102539, + 0.005823819898068905, + 0.014985148794949055, + 0.005843070801347494, + 0.0054731410928070545, + 0.0063970196060836315, + 0.004128417000174522, + 0.0035179490223526955, + 0.006725580431520939, + 0.004212772939354181, + 0.006308943964540958, + 0.0027189080137759447, + 0.014116710983216763, + 0.007858900353312492, + 0.005577901378273964, + 0.0067429072223603725, + 0.010494520887732506, + 0.012071380391716957, + 0.010939988307654858, + 0.006562983617186546, + 0.014311646111309528, + 0.010243041440844536, + 0.0073994859121739864, + 0.007742435671389103, + 0.009936481714248657, + 0.011429102160036564, + 0.006608578376471996, + 0.00759310694411397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.014184252358973026 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures the vast, eerie emptiness of a nighttime indoor swimming pool, creating a visceral, suspenseful atmosphere. The Olympic-sized pool, filled with unnatural turquoise water, its surface a perfect mirror, reflects the cold, even glow of industrial fluorescent lights mounted high in the cavernous ceiling. The pool is tiled with pale blue tiles, dark lane lines stretching into the gloom of the deep end, while the surrounding wet concrete deck lies empty. Suddenly, without warning, the overhead lights begin to flicker erratically, a violent, chaotic stutter that plunges the entire scene into jarring moments of absolute darkness. Each plunge into darkness is perfectly synchronized with a deep, resonant thunder—a visceral, low-frequency rumble that seems to shake the entire room, causing subtle vibrations in the camera and temporarily disturbing the tranquil surface of the water. This sensory assault shatters the tense silence, creating a claustrophobic feeling of disorientation and deep unease.", + "seed": 11254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_13/index_3953_seed_11254_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2117597281932831, + "vlm_logprob": -0.2117597281932831, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.2117597281932831, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017410656437277794 + ], + "opposite_logprob_tokens": [ + -4.642410755157471 + ], + "logprob_reward": -0.017410656437277794, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.017410656437277794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08722923696041107, + 0.08067183941602707, + 0.036463674157857895, + 0.04120718687772751, + 0.21817518770694733, + 0.16190019249916077, + 0.037131235003471375, + 0.15550900995731354, + 0.14877694845199585, + 0.03293551877140999 + ], + "credit_map": [ + 0.007239473983645439, + 0.01514879334717989, + 0.009733119048178196, + 0.007411598693579435, + 0.019806889817118645, + 0.010877635329961777, + 0.008217236027121544, + 0.008794479072093964, + 0.007775988429784775, + 0.01640152372419834, + 0.005715721286833286, + 0.006529388017952442, + 0.0258432999253273, + 0.007983818650245667, + 0.006569539196789265, + 0.003852555528283119, + 0.003474406199529767, + 0.007600834127515554, + 0.0028008592780679464, + 0.0024472048971801996, + 0.011357396841049194, + 0.003973058424890041, + 0.00268541369587183, + 0.002124497899785638, + 0.0037675860803574324, + 0.0069376020692288876, + 0.00404524989426136, + 0.00437995046377182, + 0.011127492412924767, + 0.0034140893258154392, + 0.005000808276236057, + 0.002534405095502734, + 0.015773527324199677, + 0.07955341041088104, + 0.00902443565428257, + 0.010274378582835197, + 0.04011811316013336, + 0.0051279133185744286, + 0.04679553210735321, + 0.011507874354720116, + 0.010950003750622272, + 0.06164569780230522, + 0.006189029663801193, + 0.006191852502524853, + 0.03636851906776428, + 0.004056242760270834, + 0.02905336767435074, + 0.007445465307682753, + 0.0026521654799580574, + 0.005422678310424089, + 0.003529129084199667, + 0.004000494256615639, + 0.006761833094060421, + 0.0028781176079064608, + 0.008912455290555954, + 0.002974357921630144, + 0.022844288498163223, + 0.025826487690210342, + 0.010740945115685463, + 0.01577247492969036, + 0.03753547742962837, + 0.012723315507173538, + 0.017650557681918144, + 0.012415463104844093, + 0.015352185815572739, + 0.024943172931671143, + 0.027016524225473404, + 0.010900676250457764, + 0.028709951788187027, + 0.016301151365041733, + 0.01442896481603384, + 0.011124313808977604, + 0.0031423016916960478, + 0.005683836527168751, + 0.002491624327376485, + 0.003995790146291256, + 0.006836531218141317, + 0.004157757852226496, + 0.003976831678301096, + 0.002650841837748885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.017410656437277794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706391513347626 + ], + "opposite_logprob_tokens": [ + -2.1270639896392822 + ], + "logprob_reward": -0.12706391513347626, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706391513347626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0734919086098671, + 0.059441592544317245, + 0.03380506858229637, + 0.03105846419930458, + 0.23421765863895416, + 0.18586128950119019, + 0.15127313137054443, + 0.12439443916082382, + 0.035746507346630096, + 0.07070986926555634 + ], + "credit_map": [ + 0.006601982284337282, + 0.009518357925117016, + 0.008341672830283642, + 0.007494449615478516, + 0.008806249126791954, + 0.005093735177069902, + 0.01862466149032116, + 0.009010805748403072, + 0.00530985975638032, + 0.01015847735106945, + 0.00478246808052063, + 0.005859341472387314, + 0.0183979794383049, + 0.005144556984305382, + 0.006913382560014725, + 0.0028755273669958115, + 0.0036237582098692656, + 0.008553624153137207, + 0.0018115753773599863, + 0.0021376439835876226, + 0.011174633167684078, + 0.0029915147460997105, + 0.0020725505892187357, + 0.001439770800061524, + 0.0030060254503041506, + 0.006697086151689291, + 0.003861646167933941, + 0.003818916156888008, + 0.006093979813158512, + 0.0021376600489020348, + 0.003455402795225382, + 0.0019877494778484106, + 0.01681738719344139, + 0.08637901395559311, + 0.00995560921728611, + 0.01050576288253069, + 0.042943477630615234, + 0.005022653378546238, + 0.050153475254774094, + 0.012440293096005917, + 0.012452169321477413, + 0.07141265273094177, + 0.007490417920053005, + 0.0069465297274291515, + 0.04499119147658348, + 0.004111615475267172, + 0.02994862012565136, + 0.008508116938173771, + 0.010338875465095043, + 0.060250259935855865, + 0.004445806611329317, + 0.005905740428715944, + 0.03443790599703789, + 0.0046837725676596165, + 0.025407953187823296, + 0.0058028362691402435, + 0.008168229833245277, + 0.03630302846431732, + 0.00658173905685544, + 0.00600505480542779, + 0.04622110351920128, + 0.006013902369886637, + 0.009859020821750164, + 0.00524236261844635, + 0.0040323734283447266, + 0.005318185780197382, + 0.006507123354822397, + 0.0035749259404838085, + 0.0066698710434138775, + 0.002879765583202243, + 0.004054933786392212, + 0.0027093319222331047, + 0.005867404397577047, + 0.014590307138860226, + 0.006097869481891394, + 0.007357662543654442, + 0.017883069813251495, + 0.00834847241640091, + 0.006155202630907297, + 0.00440987991169095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.12706391513347626 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261111378669739 + ], + "opposite_logprob_tokens": [ + -0.5761111378669739 + ], + "logprob_reward": -0.8261111378669739, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261111378669739, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06334453076124191, + 0.021238049492239952, + 0.015894556418061256, + 0.019846322014927864, + 0.04773359373211861, + 0.27400436997413635, + 0.1875336468219757, + 0.13157536089420319, + 0.050713542848825455, + 0.18811605870723724 + ], + "credit_map": [ + 0.005558050703257322, + 0.017364518716931343, + 0.00309379817917943, + 0.003708950709551573, + 0.021067334339022636, + 0.004752872046083212, + 0.004795074462890625, + 0.0030039274133741856, + 0.0017546871677041054, + 0.004108835011720657, + 0.002159425523132086, + 0.0023938396479934454, + 0.005235439166426659, + 0.0018478144193068147, + 0.0022427833173424006, + 0.0014952258206903934, + 0.0016937204636633396, + 0.003243246115744114, + 0.0013104003155604005, + 0.001698629348538816, + 0.002926570363342762, + 0.0016714301891624928, + 0.0020762463100254536, + 0.0012743143597617745, + 0.0021223730873316526, + 0.00394868291914463, + 0.001443960820324719, + 0.0021769492886960506, + 0.0037705583963543177, + 0.002087789122015238, + 0.002554020145907998, + 0.001741988817229867, + 0.0034202400129288435, + 0.014849035069346428, + 0.002399530727416277, + 0.0026118496898561716, + 0.007114793639630079, + 0.0019698587711900473, + 0.011744643561542034, + 0.0036236427258700132, + 0.013258733786642551, + 0.08702543377876282, + 0.0055121625773608685, + 0.00578073738142848, + 0.12071498483419418, + 0.010972741059958935, + 0.020863743498921394, + 0.009875809773802757, + 0.013661275617778301, + 0.07624559104442596, + 0.005672219675034285, + 0.005986148025840521, + 0.04102269560098648, + 0.00704524852335453, + 0.030928174033761024, + 0.006972290109843016, + 0.011643072590231895, + 0.034842781722545624, + 0.009082389995455742, + 0.008908607065677643, + 0.0391426756978035, + 0.009858633391559124, + 0.01235898770391941, + 0.005738221574574709, + 0.0042780926451087, + 0.014514602720737457, + 0.0035592070780694485, + 0.004043084569275379, + 0.013725184835493565, + 0.003443049965426326, + 0.004968208260834217, + 0.00218211323954165, + 0.01916489377617836, + 0.036731015890836716, + 0.02636806108057499, + 0.025231849402189255, + 0.03818538039922714, + 0.01607252098619938, + 0.015251604840159416, + 0.011110730469226837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.8261111378669739 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003604620695114 + ], + "opposite_logprob_tokens": [ + -2.695035934448242 + ], + "logprob_reward": -0.07003604620695114, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003604620695114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06273243576288223, + 0.03583709895610809, + 0.027104271575808525, + 0.05868413671851158, + 0.16396847367286682, + 0.15191905200481415, + 0.08643520623445511, + 0.14721958339214325, + 0.07794461399316788, + 0.18815509974956512 + ], + "credit_map": [ + 0.008433815091848373, + 0.011623506434261799, + 0.005615612026304007, + 0.007035866845399141, + 0.008609460666775703, + 0.009166404604911804, + 0.006434058304876089, + 0.005813714116811752, + 0.0034640415105968714, + 0.006946173962205648, + 0.002547327894717455, + 0.004666229709982872, + 0.008108265697956085, + 0.0024079803843051195, + 0.0053049735724925995, + 0.0023921073880046606, + 0.002874220721423626, + 0.004440061282366514, + 0.0021860385313630104, + 0.0035535276401787996, + 0.006487400270998478, + 0.0029395506717264652, + 0.0028409347869455814, + 0.0017825373215600848, + 0.0061899228021502495, + 0.011759040877223015, + 0.004121419042348862, + 0.0047514610923826694, + 0.01799912005662918, + 0.0056324852630496025, + 0.005010537337511778, + 0.0032201518770307302, + 0.011457793414592743, + 0.05849402770400047, + 0.006555897183716297, + 0.007103676442056894, + 0.03254557400941849, + 0.004953927360475063, + 0.03250556439161301, + 0.01035202108323574, + 0.008835946209728718, + 0.03575361147522926, + 0.004447637125849724, + 0.004489653743803501, + 0.06367754191160202, + 0.006721679586917162, + 0.022142702713608742, + 0.005850289016962051, + 0.0058675240725278854, + 0.03184184804558754, + 0.003646573983132839, + 0.004364128224551678, + 0.018733559176325798, + 0.004111892078071833, + 0.013961135409772396, + 0.0039085522294044495, + 0.015278133563697338, + 0.03279328718781471, + 0.011138420552015305, + 0.012983876280486584, + 0.036653246730566025, + 0.012998359277844429, + 0.01749923825263977, + 0.007875015027821064, + 0.008546270430088043, + 0.010203774087131023, + 0.013082272373139858, + 0.007463790010660887, + 0.012126320973038673, + 0.010517364367842674, + 0.009755191393196583, + 0.0062496243044734, + 0.014863220043480396, + 0.0379348061978817, + 0.010378032922744751, + 0.027614668011665344, + 0.04884443059563637, + 0.01826726272702217, + 0.01695009507238865, + 0.013302590698003769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.07003604620695114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176885321736336 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.018176885321736336, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176885321736336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08032144606113434, + 0.04709753766655922, + 0.012595300562679768, + 0.026827936992049217, + 0.12339485436677933, + 0.2516038417816162, + 0.267924964427948, + 0.10730385035276413, + 0.038081418722867966, + 0.04484875500202179 + ], + "credit_map": [ + 0.007068335544317961, + 0.01907450519502163, + 0.004443011712282896, + 0.0058653708547353745, + 0.024006584659218788, + 0.007162771187722683, + 0.007611599285155535, + 0.005089275073260069, + 0.004594267345964909, + 0.011124718002974987, + 0.002333278302103281, + 0.003042025025933981, + 0.015841256827116013, + 0.004356458317488432, + 0.003856079885736108, + 0.0019494572188705206, + 0.0010845899814739823, + 0.0022237016819417477, + 0.0014520023250952363, + 0.00146552047226578, + 0.0021647063549607992, + 0.0011005616979673505, + 0.0018091611564159393, + 0.001295058405958116, + 0.0022142301313579082, + 0.006409392226487398, + 0.0014660201268270612, + 0.0018163541099056602, + 0.009538853541016579, + 0.002378738485276699, + 0.0018502884777262807, + 0.0011540610576048493, + 0.010141176171600819, + 0.04155593737959862, + 0.00677367951720953, + 0.007127970457077026, + 0.01602361351251602, + 0.004475046414881945, + 0.029782265424728394, + 0.007515172008424997, + 0.011679187417030334, + 0.09634628891944885, + 0.005769270937889814, + 0.0048232306726276875, + 0.10709074884653091, + 0.010203800164163113, + 0.00606899056583643, + 0.009622358717024326, + 0.01931474171578884, + 0.10710153728723526, + 0.008101040497422218, + 0.010108829475939274, + 0.05805132910609245, + 0.009571249596774578, + 0.04549044743180275, + 0.010185813531279564, + 0.0065676188096404076, + 0.040278349071741104, + 0.003878627670928836, + 0.005337385460734367, + 0.036492183804512024, + 0.004751081578433514, + 0.006567633710801601, + 0.003430980257689953, + 0.003930471837520599, + 0.008488934487104416, + 0.004449443891644478, + 0.0034302950371056795, + 0.005653871688991785, + 0.004067159257829189, + 0.004743080120533705, + 0.0033181689213961363, + 0.00396891450509429, + 0.008155380375683308, + 0.004662897903472185, + 0.0037651285529136658, + 0.010638088919222355, + 0.004640401341021061, + 0.0056632282212376595, + 0.003354723332449794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.018176885321736336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle shot captures the vast, eerie emptiness of a nighttime indoor swimming pool, creating a visceral, suspenseful atmosphere. The Olympic-sized pool, filled with unnatural turquoise water, its surface a perfect mirror, reflects the cold, even glow of industrial fluorescent lights mounted high in the cavernous ceiling. The pool is tiled with pale blue tiles, dark lane lines stretching into the gloom of the deep end, while the surrounding wet concrete deck lies empty. Suddenly, without warning, the overhead lights begin to flicker erratically, a violent, chaotic stutter that plunges the entire scene into jarring moments of absolute darkness. Each plunge into darkness is perfectly synchronized with a deep, resonant thunder—a visceral, low-frequency rumble that seems to shake the entire room, causing subtle vibrations in the camera and temporarily disturbing the tranquil surface of the water. This sensory assault shatters the tense silence, creating a claustrophobic feeling of disorientation and deep unease.", + "seed": 11255, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_13/index_3953_seed_11255_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.21041520033031702, + "vlm_logprob": -0.21041520033031702, + "vlm_yes_no_margin": 3.6500000059604645, + "vlm_token_credit": -0.21041520033031702, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007128874771296978 + ], + "opposite_logprob_tokens": [ + -6.007128715515137 + ], + "logprob_reward": -0.007128874771296978, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.007128874771296978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18447816371917725, + 0.19665850698947906, + 0.044963937252759933, + 0.025422440841794014, + 0.03607723489403725, + 0.14070969820022583, + 0.13596105575561523, + 0.09773764759302139, + 0.03620633855462074, + 0.10178498923778534 + ], + "credit_map": [ + 0.010765626095235348, + 0.06447524577379227, + 0.00840106513351202, + 0.010818330571055412, + 0.014588342048227787, + 0.06407994776964188, + 0.006792697124183178, + 0.004556913860142231, + 0.007865002378821373, + 0.07943891733884811, + 0.0037580104544758797, + 0.008795704692602158, + 0.009932825341820717, + 0.07888631522655487, + 0.004127355758100748, + 0.003854373237118125, + 0.0044738552533090115, + 0.009252818301320076, + 0.0038229357451200485, + 0.004903979599475861, + 0.005615520291030407, + 0.012316043488681316, + 0.00267750583589077, + 0.0019012772245332599, + 0.003141473513096571, + 0.003125521820038557, + 0.003852513385936618, + 0.0036506394390016794, + 0.0028748472686856985, + 0.004665226209908724, + 0.002205147175118327, + 0.0019070712151005864, + 0.0037864684127271175, + 0.00771382637321949, + 0.0032080630771815777, + 0.004679532255977392, + 0.005906796082854271, + 0.006443836726248264, + 0.002281507011502981, + 0.0020572063513100147, + 0.007440298795700073, + 0.054471202194690704, + 0.0037536448799073696, + 0.006488246843218803, + 0.007224283181130886, + 0.05385535955429077, + 0.0039707752875983715, + 0.0035058893263339996, + 0.0075985342264175415, + 0.04937802255153656, + 0.005298903677612543, + 0.007789304479956627, + 0.008302992209792137, + 0.048807404935359955, + 0.005846749991178513, + 0.0029391481075435877, + 0.0063063837587833405, + 0.032239533960819244, + 0.004228186793625355, + 0.007044121157377958, + 0.0068939924240112305, + 0.03467420116066933, + 0.00360640836879611, + 0.002744820900261402, + 0.003486509434878826, + 0.0070030721835792065, + 0.0029309948440641165, + 0.006257108878344297, + 0.005585296545177698, + 0.006746800150722265, + 0.002473455388098955, + 0.0017230991506949067, + 0.009985776618123055, + 0.022597242146730423, + 0.009036844596266747, + 0.011499465443193913, + 0.014962372370064259, + 0.0229372289031744, + 0.0069525474682450294, + 0.003813507966697216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.007128874771296978 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088416911661625 + ], + "opposite_logprob_tokens": [ + -4.5110883712768555 + ], + "logprob_reward": -0.011088416911661625, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088416911661625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24688473343849182, + 0.2509004771709442, + 0.2614189684391022, + 0.05538255721330643, + 0.019091520458459854, + 0.02239665389060974, + 0.04391473904252052, + 0.03410624340176582, + 0.019119087606668472, + 0.04678508639335632 + ], + "credit_map": [ + 0.01480895560234785, + 0.09301704168319702, + 0.009101826697587967, + 0.009016178548336029, + 0.015870671719312668, + 0.09250480681657791, + 0.006078104954212904, + 0.006487125530838966, + 0.01008617039769888, + 0.10524208843708038, + 0.004468060564249754, + 0.01043047197163105, + 0.011186463758349419, + 0.10112223029136658, + 0.003954337444156408, + 0.004410634282976389, + 0.015218407846987247, + 0.09700959175825119, + 0.007234151940792799, + 0.013928155414760113, + 0.016141189262270927, + 0.09984289109706879, + 0.006209925282746553, + 0.005834611598402262, + 0.005636492278426886, + 0.012031816877424717, + 0.004888462834060192, + 0.00544138764962554, + 0.00800198782235384, + 0.011785006150603294, + 0.0043090456165373325, + 0.0032883496023714542, + 0.0017204262549057603, + 0.003777537727728486, + 0.0015790227334946394, + 0.0021864676382392645, + 0.0027210633270442486, + 0.003596986411139369, + 0.00238978723064065, + 0.0011202257592231035, + 0.0022716692183166742, + 0.004864756017923355, + 0.0013867159141227603, + 0.0031160477083176374, + 0.003440782194957137, + 0.005127346608787775, + 0.001285230042412877, + 0.0009041031007654965, + 0.002530616009607911, + 0.015804963186383247, + 0.0013349448563531041, + 0.003023742698132992, + 0.0025553335435688496, + 0.016219062730669975, + 0.0013515462633222342, + 0.0010945227695629, + 0.0025251826737076044, + 0.01094962377101183, + 0.0013573429314419627, + 0.0020724383648484945, + 0.005033473484218121, + 0.010032601654529572, + 0.0011341803474351764, + 0.0010013944702222943, + 0.00180822832044214, + 0.002502505900338292, + 0.002594914985820651, + 0.003200077684596181, + 0.0027723978273570538, + 0.003607623279094696, + 0.0016765821492299438, + 0.0009567555971443653, + 0.004172350745648146, + 0.00973437074571848, + 0.0035914636682718992, + 0.007247899658977985, + 0.006747812032699585, + 0.01153093483299017, + 0.0021373629570007324, + 0.00162288558203727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.011088416911661625 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742610454559326 + ], + "opposite_logprob_tokens": [ + -0.474261075258255 + ], + "logprob_reward": -0.9742610454559326, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9742610454559326, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14122571051120758, + 0.07184004783630371, + 0.048399604856967926, + 0.086024209856987, + 0.013428578153252602, + 0.019535599276423454, + 0.06379075348377228, + 0.09706085175275803, + 0.0373423807322979, + 0.4213522672653198 + ], + "credit_map": [ + 0.010852329432964325, + 0.04364665970206261, + 0.010213499888777733, + 0.010013293474912643, + 0.01522057130932808, + 0.039206430315971375, + 0.006394663825631142, + 0.00567827932536602, + 0.0043685296550393105, + 0.026415079832077026, + 0.002470784354954958, + 0.004241867922246456, + 0.004242045804858208, + 0.02589246816933155, + 0.0025434603448957205, + 0.0016658136155456305, + 0.003267543390393257, + 0.012127716094255447, + 0.0021779476664960384, + 0.006150250323116779, + 0.004889118950814009, + 0.015382936224341393, + 0.0025903901550918818, + 0.001813705894164741, + 0.005785869434475899, + 0.030378488823771477, + 0.0028283996507525444, + 0.003794875228777528, + 0.009265000931918621, + 0.027566729113459587, + 0.0032425611279904842, + 0.0031622799579054117, + 0.0016043573850765824, + 0.002311545889824629, + 0.0016222369158640504, + 0.001411294681020081, + 0.0024669368285685778, + 0.001910627819597721, + 0.0014532837085425854, + 0.0006482950411736965, + 0.001547743217088282, + 0.003969973884522915, + 0.001474503194913268, + 0.0022475318983197212, + 0.002548519754782319, + 0.005592821631580591, + 0.0014022247632965446, + 0.0007522836094722152, + 0.004074519034475088, + 0.022676926106214523, + 0.0025069061666727066, + 0.0030898298136889935, + 0.006077003665268421, + 0.021678905934095383, + 0.0017829671269282699, + 0.0019036983139812946, + 0.005106801632791758, + 0.03575366362929344, + 0.003306684084236622, + 0.006179865449666977, + 0.0060117593966424465, + 0.0358150415122509, + 0.0026940719690173864, + 0.0021929689683020115, + 0.00398495327681303, + 0.00894523411989212, + 0.0023630375508219004, + 0.0044538467191159725, + 0.003876209259033203, + 0.009012667462229729, + 0.0033123630564659834, + 0.0013940705684944987, + 0.03478076681494713, + 0.10129258036613464, + 0.030191829428076744, + 0.05802425369620323, + 0.041061628609895706, + 0.11682026833295822, + 0.025493256747722626, + 0.013687714003026485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.9742610454559326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054963547736406326 + ], + "opposite_logprob_tokens": [ + -2.9299635887145996 + ], + "logprob_reward": -0.054963547736406326, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054963547736406326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07785548269748688, + 0.1446743905544281, + 0.03687504306435585, + 0.07559750229120255, + 0.029602468013763428, + 0.0641142874956131, + 0.0898907482624054, + 0.11879225075244904, + 0.03693512827157974, + 0.3256627321243286 + ], + "credit_map": [ + 0.010391220450401306, + 0.02044784650206566, + 0.00594363221898675, + 0.0059260353446006775, + 0.010647296905517578, + 0.016623780131340027, + 0.0043293447233736515, + 0.0035463226959109306, + 0.008127950131893158, + 0.05577056482434273, + 0.006769644096493721, + 0.005959568545222282, + 0.006480277981609106, + 0.05434447526931763, + 0.003714374965056777, + 0.003507543122395873, + 0.005821623839437962, + 0.005901301745325327, + 0.0034974568989127874, + 0.006494065281003714, + 0.004845373332500458, + 0.0053883884102106094, + 0.0028724647127091885, + 0.0020543686114251614, + 0.005216510966420174, + 0.025736218318343163, + 0.003348099533468485, + 0.00505886971950531, + 0.00685154041275382, + 0.022380134090781212, + 0.004245146177709103, + 0.002760986564680934, + 0.002736393827944994, + 0.007018986623734236, + 0.002398722106590867, + 0.002322613960132003, + 0.004015014972537756, + 0.0069683147594332695, + 0.002169848419725895, + 0.0019725700840353966, + 0.0043550641275942326, + 0.021784968674182892, + 0.002672893460839987, + 0.004666693974286318, + 0.004711998160928488, + 0.0219414122402668, + 0.002123146317899227, + 0.0018581015756353736, + 0.00704811979085207, + 0.029619503766298294, + 0.005471513140946627, + 0.007681443355977535, + 0.005327544175088406, + 0.02813698910176754, + 0.004269426688551903, + 0.002336203120648861, + 0.010021974332630634, + 0.035875026136636734, + 0.006542994640767574, + 0.00894281454384327, + 0.013244902715086937, + 0.035672225058078766, + 0.004969003144651651, + 0.0035233013331890106, + 0.0035448004491627216, + 0.008284296840429306, + 0.0024579542223364115, + 0.005594999995082617, + 0.003316539106890559, + 0.009272607043385506, + 0.0033069129567593336, + 0.0011570171918720007, + 0.026075728237628937, + 0.07479695230722427, + 0.023489156737923622, + 0.048583388328552246, + 0.03553213179111481, + 0.08821498602628708, + 0.01893927901983261, + 0.010031111538410187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.054963547736406326 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a nighttime indoor swimming pool with pale blue tiles and dark lane lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water turquoise and acting as a perfect mirror reflecting industrial fluorescent lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the overhead lights flicker erratically, causing the scene to plunge into absolute darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera exhibit subtle vibrations synchronized with the light flickers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the surrounding wet concrete deck empty with no people or objects present?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634116776287556 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004634116776287556, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634116776287556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4000268876552582, + 0.19848628342151642, + 0.025234799832105637, + 0.018139559775590897, + 0.02930636703968048, + 0.0267616119235754, + 0.0434243306517601, + 0.05932973697781563, + 0.01618729718029499, + 0.18310309946537018 + ], + "credit_map": [ + 0.02098945528268814, + 0.1628706306219101, + 0.009228640235960484, + 0.019595246762037277, + 0.021068593487143517, + 0.14934058487415314, + 0.008049553260207176, + 0.008884187787771225, + 0.008206686936318874, + 0.08340436965227127, + 0.0035427934490144253, + 0.008654394187033176, + 0.00884542427957058, + 0.0792899802327156, + 0.00308584189042449, + 0.0034568216651678085, + 0.002837022067978978, + 0.004140167497098446, + 0.0026632053777575493, + 0.0041941432282328606, + 0.0031117296312004328, + 0.005207991227507591, + 0.0018855836242437363, + 0.00119496020488441, + 0.0022795211989432573, + 0.003152400953695178, + 0.00220222189091146, + 0.0020495483186095953, + 0.0030948997009545565, + 0.003246476175263524, + 0.0011263353517279029, + 0.0009881585137918591, + 0.0032561623957008123, + 0.004381965380162001, + 0.0036638786550611258, + 0.004036235623061657, + 0.005669682286679745, + 0.004783124662935734, + 0.001898731803521514, + 0.001616586814634502, + 0.0015609452966600657, + 0.009397617541253567, + 0.0009066417114809155, + 0.001563298748806119, + 0.0018283803947269917, + 0.008900173008441925, + 0.0015365324215963483, + 0.001068026409484446, + 0.0023065549321472645, + 0.01677313633263111, + 0.001267291372641921, + 0.0029094889760017395, + 0.0022497009485960007, + 0.015514091588556767, + 0.0014195244293659925, + 0.000984550453722477, + 0.0034966133534908295, + 0.021187299862504005, + 0.0016960327047854662, + 0.003507128683850169, + 0.004716768395155668, + 0.02144724503159523, + 0.001834510243497789, + 0.0014441367238759995, + 0.002016000682488084, + 0.0023572829086333513, + 0.0024173781275749207, + 0.0020460865925997496, + 0.002858323510736227, + 0.0021185169462114573, + 0.0014362287474796176, + 0.0009374819928780198, + 0.010115198791027069, + 0.059147752821445465, + 0.008430439047515392, + 0.01770637184381485, + 0.015930362045764923, + 0.061774980276823044, + 0.00555659644305706, + 0.004441406112164259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.004634116776287556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_14/index_tensor([1521, 1521])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_14/index_tensor([1521, 1521])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7b2e05e16896ff151dbe9af0cd8f8ea75d7e64bd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_14/index_tensor([1521, 1521])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angle close-up captures a portion of a clean, light-colored laminate desk, illuminated by the bright, even glow of realistic overhead office lighting. A pair of hands firmly grasps a matte black foam tube, its dense, slightly porous texture clearly visible. The tube is forcibly bent into a tight, compressed zigzag shape, the material showing visible strain and tension along its creases. In a swift and decisive action, the hands release. With dynamic, whip-like speed, the foam tube snaps instantly and violently in a straight line, its form blurred for a fraction of a second by the rapid motion. It falls to the desk surface with a soft, muffled thud, forming a perfectly straight line. Its dark, non-reflective form contrasts sharply with the pale, linear surface of its final state, creating a focused, almost clinical atmosphere.", + "seed": 15901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_14/index_1521_seed_15901_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.42256126143038275, + "vlm_logprob": -0.42256126143038275, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.42256126143038275, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01743701472878456 + ], + "opposite_logprob_tokens": [ + -4.642436981201172 + ], + "logprob_reward": -0.01743701472878456, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01743701472878456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15774373710155487, + 0.1828550100326538, + 0.16055768728256226, + 0.05488438904285431, + 0.04686436057090759, + 0.053997259587049484, + 0.08759236335754395, + 0.10494314134120941, + 0.10217808187007904, + 0.048383940011262894 + ], + "credit_map": [ + 0.04163641855120659, + 0.009541325271129608, + 0.014946854673326015, + 0.011677474714815617, + 0.008575625717639923, + 0.017162742093205452, + 0.015536979772150517, + 0.03866632655262947, + 0.015182680450379848, + 0.026747655123472214, + 0.01474385429173708, + 0.02693876251578331, + 0.029738696292042732, + 0.02055361494421959, + 0.032530900090932846, + 0.01641884632408619, + 0.014493070542812347, + 0.02855723164975643, + 0.009864014573395252, + 0.05766582116484642, + 0.016508104279637337, + 0.00718498881906271, + 0.0173252634704113, + 0.008959188126027584, + 0.006485812831670046, + 0.011136255227029324, + 0.006155300885438919, + 0.0075714364647865295, + 0.007671013940125704, + 0.0044844443909823895, + 0.0055878120474517345, + 0.005792311392724514, + 0.004498261027038097, + 0.009306952357292175, + 0.006379517260938883, + 0.006455988623201847, + 0.006336696445941925, + 0.0038842165376991034, + 0.0037702713161706924, + 0.0062324577011168, + 0.008608800359070301, + 0.009087993763387203, + 0.004597918130457401, + 0.005839291960000992, + 0.006222310476005077, + 0.006955339573323727, + 0.008280952461063862, + 0.004404651001095772, + 0.006486482918262482, + 0.013960259035229683, + 0.013116690330207348, + 0.009079153649508953, + 0.011453395709395409, + 0.010650817304849625, + 0.013719414360821247, + 0.009126147255301476, + 0.017308250069618225, + 0.022331656888127327, + 0.014272591099143028, + 0.008014130406081676, + 0.013474258594214916, + 0.008382485248148441, + 0.00879143737256527, + 0.01236832793802023, + 0.01693568378686905, + 0.012604636140167713, + 0.010425586253404617, + 0.009793485514819622, + 0.029344571754336357, + 0.008843748830258846, + 0.004461723379790783, + 0.00976863969117403, + 0.00766358058899641, + 0.008804013952612877, + 0.006031946744769812, + 0.006028193514794111, + 0.006767423823475838, + 0.005564000923186541, + 0.002983243204653263, + 0.004541539587080479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.01743701472878456 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009825674816966057 + ], + "opposite_logprob_tokens": [ + -4.634825706481934 + ], + "logprob_reward": -0.009825674816966057, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009825674816966057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3307494521141052, + 0.20009148120880127, + 0.1005077213048935, + 0.0546875, + 0.03252658247947693, + 0.05200813710689545, + 0.03392898291349411, + 0.06986620277166367, + 0.09263487160205841, + 0.03299902006983757 + ], + "credit_map": [ + 0.09528758376836777, + 0.020945703610777855, + 0.028552643954753876, + 0.024650374427437782, + 0.018656283617019653, + 0.02490130066871643, + 0.034874700009822845, + 0.0828808844089508, + 0.020882081240415573, + 0.026283500716090202, + 0.02839934453368187, + 0.026155008003115654, + 0.02822606824338436, + 0.02230045571923256, + 0.02836501970887184, + 0.019480004906654358, + 0.008088670670986176, + 0.017813825979828835, + 0.005215198267251253, + 0.03528454527258873, + 0.0063008819706737995, + 0.004658637568354607, + 0.01088707335293293, + 0.012258889153599739, + 0.006131586618721485, + 0.0078106168657541275, + 0.005256208591163158, + 0.006969482637941837, + 0.007526567671447992, + 0.004950748756527901, + 0.009121911600232124, + 0.006920375861227512, + 0.003995236940681934, + 0.004350368864834309, + 0.005679571069777012, + 0.004083674866706133, + 0.004879415035247803, + 0.0024224405642598867, + 0.0029200855642557144, + 0.004195792600512505, + 0.006298364605754614, + 0.012389296665787697, + 0.004563695285469294, + 0.004698398523032665, + 0.006548185832798481, + 0.0033461833372712135, + 0.005520676728338003, + 0.008643335662782192, + 0.0029308584053069353, + 0.0038313481491059065, + 0.003397759748622775, + 0.0028563402593135834, + 0.005122620612382889, + 0.004263318143785, + 0.006810292601585388, + 0.004716443829238415, + 0.010375267826020718, + 0.013977669179439545, + 0.008960774168372154, + 0.004002117086201906, + 0.010155442170798779, + 0.008656694553792477, + 0.004811367020010948, + 0.008926869370043278, + 0.012490321882069111, + 0.012615887448191643, + 0.01124567911028862, + 0.010140524245798588, + 0.019755136221647263, + 0.013338101096451283, + 0.005989976227283478, + 0.007059239782392979, + 0.004960023798048496, + 0.0037086689844727516, + 0.006932428572326899, + 0.00516866147518158, + 0.003830620553344488, + 0.003048175247386098, + 0.0021816748194396496, + 0.0031687654554843903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009825674816966057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256375402212143 + ], + "opposite_logprob_tokens": [ + -1.6006375551223755 + ], + "logprob_reward": -0.2256375402212143, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256375402212143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06137251481413841, + 0.08408857882022858, + 0.6892011165618896, + 0.027225330471992493, + 0.05305487662553787, + 0.020335106179118156, + 0.02236234024167061, + 0.013893421739339828, + 0.012257709167897701, + 0.01620897650718689 + ], + "credit_map": [ + 0.012784447520971298, + 0.006546696648001671, + 0.0074052647687494755, + 0.005075317807495594, + 0.006287929601967335, + 0.005417587235569954, + 0.006420160178095102, + 0.011435109190642834, + 0.008905552327632904, + 0.010062050074338913, + 0.006737973541021347, + 0.011118830181658268, + 0.011591257527470589, + 0.006989519577473402, + 0.012487273663282394, + 0.016196127980947495, + 0.05532477796077728, + 0.13361606001853943, + 0.03173943608999252, + 0.26987355947494507, + 0.02241428755223751, + 0.02233850583434105, + 0.07835263013839722, + 0.07554184645414352, + 0.002477841917425394, + 0.0040780543349683285, + 0.0023613551165908575, + 0.0029816278256475925, + 0.004379097372293472, + 0.0035794670693576336, + 0.0046060592867434025, + 0.0027618289459496737, + 0.005682077258825302, + 0.01072578877210617, + 0.006860885303467512, + 0.005467709619551897, + 0.009162885136902332, + 0.003941225819289684, + 0.005227785557508469, + 0.005986519157886505, + 0.002580048516392708, + 0.005206363275647163, + 0.0011777697363868356, + 0.0018135581631213427, + 0.0021770240273326635, + 0.002154483227059245, + 0.0031068623065948486, + 0.0021189951803535223, + 0.0029596786480396986, + 0.002868643729016185, + 0.003260314930230379, + 0.002027777722105384, + 0.0029036751948297024, + 0.002417973941192031, + 0.003780728206038475, + 0.0021435481030493975, + 0.0014890461461618543, + 0.0018047254998236895, + 0.001995448488742113, + 0.001604220480658114, + 0.0019216765649616718, + 0.0018402282148599625, + 0.0013394722482189536, + 0.0018986037466675043, + 0.0015848339535295963, + 0.0013252871576696634, + 0.0021877314429730177, + 0.0015069050714373589, + 0.0027322364039719105, + 0.0011345876846462488, + 0.000895294826477766, + 0.0008908328018151224, + 0.0023267781361937523, + 0.0035651568323373795, + 0.0016447539674118161, + 0.002284055110067129, + 0.0028277274686843157, + 0.0009263334213756025, + 0.0016591098392382264, + 0.0009750603348948061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.2256375402212143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.804888367652893 + ], + "opposite_logprob_tokens": [ + -0.17988835275173187 + ], + "logprob_reward": -1.804888367652893, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.804888367652893, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19561830163002014, + 0.14888077974319458, + 0.18288089334964752, + 0.07040321081876755, + 0.040650174021720886, + 0.11941151320934296, + 0.061098452657461166, + 0.09098517149686813, + 0.04827320575714111, + 0.04179834946990013 + ], + "credit_map": [ + 0.05996991693973541, + 0.013971807435154915, + 0.02061789110302925, + 0.013419118709862232, + 0.007914230227470398, + 0.011714943684637547, + 0.014263647608458996, + 0.053746748715639114, + 0.015700340270996094, + 0.017207367345690727, + 0.020717496052384377, + 0.019974563270807266, + 0.017728805541992188, + 0.016797294840216637, + 0.01830650493502617, + 0.022448407486081123, + 0.013623720966279507, + 0.03568989038467407, + 0.007507081143558025, + 0.07359355688095093, + 0.008246620185673237, + 0.006076178513467312, + 0.021158622577786446, + 0.016985220834612846, + 0.009311131201684475, + 0.013201880268752575, + 0.0072682141326367855, + 0.008020158857107162, + 0.011429570615291595, + 0.0066229416988790035, + 0.008299725130200386, + 0.006249584723263979, + 0.003854125738143921, + 0.005694463383406401, + 0.008354919031262398, + 0.004741617012768984, + 0.0058000278659164906, + 0.003298151073977351, + 0.00404720613732934, + 0.004859666805714369, + 0.012003541924059391, + 0.02472730726003647, + 0.009239108301699162, + 0.011423604562878609, + 0.015089055523276329, + 0.014927119947969913, + 0.01700107753276825, + 0.015000693500041962, + 0.007088643033057451, + 0.007857891730964184, + 0.00998806394636631, + 0.005151718854904175, + 0.006962374318391085, + 0.008512432686984539, + 0.008861007168889046, + 0.006676321383565664, + 0.01263811718672514, + 0.018880028277635574, + 0.010007222183048725, + 0.005820025224238634, + 0.015013517811894417, + 0.01406713854521513, + 0.005387103650718927, + 0.009172016754746437, + 0.0068270363844931126, + 0.0066959732212126255, + 0.006293368525803089, + 0.004769566003233194, + 0.011357483454048634, + 0.004755704663693905, + 0.0024537246208637953, + 0.005120350979268551, + 0.00692875636741519, + 0.0075660571455955505, + 0.005811810493469238, + 0.0041296714916825294, + 0.006599539890885353, + 0.00462033087387681, + 0.0030470264609903097, + 0.003095159539952874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -1.804888367652893 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055017709732055664 + ], + "opposite_logprob_tokens": [ + -2.9300177097320557 + ], + "logprob_reward": -0.055017709732055664, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055017709732055664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15340249240398407, + 0.19888952374458313, + 0.2830791175365448, + 0.02700481377542019, + 0.036429502069950104, + 0.07742041349411011, + 0.04380399361252785, + 0.06957975029945374, + 0.04920099303126335, + 0.06118932366371155 + ], + "credit_map": [ + 0.04056483507156372, + 0.014279386959969997, + 0.018000761047005653, + 0.01032596081495285, + 0.010858544148504734, + 0.00974922813475132, + 0.013637789525091648, + 0.03598598763346672, + 0.020464792847633362, + 0.01834455505013466, + 0.020639516413211823, + 0.026712428778409958, + 0.02487022802233696, + 0.022413460537791252, + 0.027847979217767715, + 0.03759656101465225, + 0.022045711055397987, + 0.055128809064626694, + 0.011957026086747646, + 0.11208701133728027, + 0.008608558215200901, + 0.009206163696944714, + 0.032429929822683334, + 0.03161592409014702, + 0.002708503045141697, + 0.0038893092423677444, + 0.002972677582874894, + 0.0041923122480511665, + 0.0032945056445896626, + 0.0019882055930793285, + 0.004144036211073399, + 0.0038152646739035845, + 0.004076105076819658, + 0.005907709244638681, + 0.0040585678070783615, + 0.005042403470724821, + 0.005752843338996172, + 0.002570392331108451, + 0.004233364015817642, + 0.004788113757967949, + 0.014294929802417755, + 0.015176202170550823, + 0.006687820889055729, + 0.008463336154818535, + 0.008081289008259773, + 0.007780545856803656, + 0.009384376928210258, + 0.007551913149654865, + 0.0033345615956932306, + 0.004291952587664127, + 0.005256746429949999, + 0.004068442154675722, + 0.004804265219718218, + 0.005545237101614475, + 0.011250531300902367, + 0.005252256523817778, + 0.010211008600890636, + 0.010466029867529869, + 0.009927919134497643, + 0.006517970468848944, + 0.007223275024443865, + 0.010892857797443867, + 0.005237363278865814, + 0.009103327989578247, + 0.009187866002321243, + 0.005712324753403664, + 0.007454555016011, + 0.005579781718552113, + 0.008683079853653908, + 0.005609031300991774, + 0.002962216269224882, + 0.004012140911072493, + 0.008334580808877945, + 0.005293936002999544, + 0.011433196254074574, + 0.009876315481960773, + 0.007917365059256554, + 0.0060103884898126125, + 0.006280213128775358, + 0.006043329834938049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.055017709732055664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up captures a portion of a clean, light-colored laminate desk, illuminated by the bright, even glow of realistic overhead office lighting. A pair of hands firmly grasps a matte black foam tube, its dense, slightly porous texture clearly visible. The tube is forcibly bent into a tight, compressed zigzag shape, the material showing visible strain and tension along its creases. In a swift and decisive action, the hands release. With dynamic, whip-like speed, the foam tube snaps instantly and violently in a straight line, its form blurred for a fraction of a second by the rapid motion. It falls to the desk surface with a soft, muffled thud, forming a perfectly straight line. Its dark, non-reflective form contrasts sharply with the pale, linear surface of its final state, creating a focused, almost clinical atmosphere.", + "seed": 15902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_14/index_1521_seed_15902_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3629021676722914, + "vlm_logprob": -0.3629021676722914, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.3629021676722914, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022037621587514877 + ], + "opposite_logprob_tokens": [ + -4.147037506103516 + ], + "logprob_reward": -0.022037621587514877, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022037621587514877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.134752094745636, + 0.24567420780658722, + 0.034385938197374344, + 0.09351084381341934, + 0.07007351517677307, + 0.14929915964603424, + 0.08654460310935974, + 0.06766334176063538, + 0.08967605978250504, + 0.02842019312083721 + ], + "credit_map": [ + 0.01676798053085804, + 0.018605254590511322, + 0.015195329673588276, + 0.015637626871466637, + 0.013773834332823753, + 0.019311977550387383, + 0.025535322725772858, + 0.009924765676259995, + 0.03877343237400055, + 0.033741116523742676, + 0.026290135458111763, + 0.024376068264245987, + 0.027942461892962456, + 0.04736950621008873, + 0.027654075995087624, + 0.019527394324541092, + 0.0039103226736187935, + 0.0045473151840269566, + 0.0026873047463595867, + 0.005492100957781076, + 0.003937847912311554, + 0.004168475978076458, + 0.006680578924715519, + 0.0029619927518069744, + 0.010655593127012253, + 0.012345788069069386, + 0.009102337062358856, + 0.01376656349748373, + 0.008834329433739185, + 0.013400360941886902, + 0.013355376198887825, + 0.012050501070916653, + 0.007566136308014393, + 0.009009904228150845, + 0.007209848612546921, + 0.009660182520747185, + 0.007934044115245342, + 0.01072586327791214, + 0.008770015090703964, + 0.009197519160807133, + 0.02175784297287464, + 0.008977697230875492, + 0.014155237935483456, + 0.01185399666428566, + 0.016586320474743843, + 0.04407656937837601, + 0.014651834033429623, + 0.017239661887288094, + 0.00843234732747078, + 0.010386627167463303, + 0.010378248989582062, + 0.012432617135345936, + 0.011999058537185192, + 0.009549147449433804, + 0.009102562442421913, + 0.014263991266489029, + 0.009642085060477257, + 0.009335611015558243, + 0.010958594270050526, + 0.004943149629980326, + 0.010566714219748974, + 0.00859033688902855, + 0.00466808769851923, + 0.008958759717643261, + 0.011103564873337746, + 0.01640540547668934, + 0.009177327156066895, + 0.008395278826355934, + 0.012763998471200466, + 0.014239095151424408, + 0.007081976626068354, + 0.010509411804378033, + 0.0032034646719694138, + 0.0041924030520021915, + 0.00364875141531229, + 0.0018608022946864367, + 0.002924987580627203, + 0.006898889318108559, + 0.002055560704320669, + 0.0036353333853185177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.022037621587514877 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022128161508589983 + ], + "opposite_logprob_tokens": [ + -6.127213001251221 + ], + "logprob_reward": -0.0022128161508589983, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022128161508589983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16163812577724457, + 0.15809907019138336, + 0.12446636706590652, + 0.15058094263076782, + 0.04816744849085808, + 0.13066953420639038, + 0.060144536197185516, + 0.04661121591925621, + 0.0473225973546505, + 0.07230019569396973 + ], + "credit_map": [ + 0.024743584915995598, + 0.020087657496333122, + 0.019169436767697334, + 0.017936548218131065, + 0.015117937698960304, + 0.024491610005497932, + 0.026198802515864372, + 0.01389254815876484, + 0.02348259463906288, + 0.022369323298335075, + 0.016218379139900208, + 0.016813891008496284, + 0.020050542429089546, + 0.026528965681791306, + 0.01727530173957348, + 0.015360085293650627, + 0.020290866494178772, + 0.014472946524620056, + 0.010151929222047329, + 0.013911686837673187, + 0.013076765462756157, + 0.023377995938062668, + 0.017537010833621025, + 0.011647158302366734, + 0.017851483076810837, + 0.02044104039669037, + 0.013694277964532375, + 0.02208453230559826, + 0.01512774545699358, + 0.022449444979429245, + 0.021548723801970482, + 0.017383690923452377, + 0.0049066306091845036, + 0.007265476509928703, + 0.0037857170682400465, + 0.005501309409737587, + 0.009209197014570236, + 0.005393441300839186, + 0.007308457046747208, + 0.004797220695763826, + 0.014398647472262383, + 0.01316503994166851, + 0.022372573614120483, + 0.018871454522013664, + 0.01337507925927639, + 0.022718161344528198, + 0.013458185829222202, + 0.012310403399169445, + 0.005853842943906784, + 0.009019455872476101, + 0.006096510216593742, + 0.009869367815554142, + 0.009110771119594574, + 0.005831005517393351, + 0.005554267670959234, + 0.008809316903352737, + 0.006823414005339146, + 0.0065728286281228065, + 0.006046645808964968, + 0.0033022111747413874, + 0.0061875395476818085, + 0.005191556178033352, + 0.003743839217349887, + 0.008743183687329292, + 0.00776359997689724, + 0.009720097295939922, + 0.006036464124917984, + 0.0034367286134511232, + 0.005428859032690525, + 0.0046416642144322395, + 0.0032618173863738775, + 0.007033366244286299, + 0.009706309996545315, + 0.012832777574658394, + 0.007810025010257959, + 0.006414937786757946, + 0.006842931732535362, + 0.00741396052762866, + 0.006031469441950321, + 0.015247786417603493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0022128161508589983 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497866868972778 + ], + "opposite_logprob_tokens": [ + -2.929978609085083 + ], + "logprob_reward": -0.05497866868972778, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497866868972778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08433753252029419, + 0.09516826272010803, + 0.05098038911819458, + 0.2713358700275421, + 0.05405222624540329, + 0.22626644372940063, + 0.059530049562454224, + 0.04303788021206856, + 0.08971551060676575, + 0.025575799867510796 + ], + "credit_map": [ + 0.011824268847703934, + 0.008746194653213024, + 0.011445771902799606, + 0.007889002561569214, + 0.008988095447421074, + 0.013175317086279392, + 0.011873696930706501, + 0.010395186021924019, + 0.01464689988642931, + 0.01276822667568922, + 0.011212234385311604, + 0.008255538530647755, + 0.011057588271796703, + 0.018174979835748672, + 0.011094367131590843, + 0.0079584289342165, + 0.005334154237061739, + 0.007750351447612047, + 0.004301511216908693, + 0.006938524078577757, + 0.006745487451553345, + 0.007185815833508968, + 0.008681343868374825, + 0.004043205175548792, + 0.019084392115473747, + 0.042801011353731155, + 0.023730875924229622, + 0.06188470497727394, + 0.029612768441438675, + 0.020820170640945435, + 0.05336730182170868, + 0.0200346652418375, + 0.004974746145308018, + 0.006966176442801952, + 0.005503189750015736, + 0.00856711808592081, + 0.0081693846732378, + 0.005369020160287619, + 0.010113499127328396, + 0.004389097448438406, + 0.023285672068595886, + 0.019604559987783432, + 0.044832661747932434, + 0.031021710485219955, + 0.021567264571785927, + 0.04087004065513611, + 0.02210686169564724, + 0.022977687418460846, + 0.005161127541214228, + 0.004927672445774078, + 0.006215204019099474, + 0.004867504816502333, + 0.01796792820096016, + 0.006249573547393084, + 0.004004030022770166, + 0.010137009434401989, + 0.007642571348696947, + 0.004515399225056171, + 0.009290889836847782, + 0.0045468686148524284, + 0.004280488006770611, + 0.004750047344714403, + 0.0025027089286595583, + 0.005508906207978725, + 0.015103749930858612, + 0.01320407073944807, + 0.010812308639287949, + 0.007885930128395557, + 0.011929025873541832, + 0.011489924043416977, + 0.005245247855782509, + 0.014045252464711666, + 0.004414158407598734, + 0.005549875553697348, + 0.0031936666928231716, + 0.0021142223849892616, + 0.0026666554622352123, + 0.0029306092765182257, + 0.0013932426227256656, + 0.0033133712131530046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.05497866868972778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015421390533447 + ], + "opposite_logprob_tokens": [ + -0.20154216885566711 + ], + "logprob_reward": -1.7015421390533447, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015421390533447, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09244862198829651, + 0.08222614228725433, + 0.06022455170750618, + 0.033927712589502335, + 0.10422073304653168, + 0.09562136977910995, + 0.21210363507270813, + 0.13605362176895142, + 0.11259429156780243, + 0.07057932019233704 + ], + "credit_map": [ + 0.014297396875917912, + 0.007742678746581078, + 0.012073910795152187, + 0.009094459936022758, + 0.010945162735879421, + 0.017105109989643097, + 0.009490474127233028, + 0.011699426919221878, + 0.012994278222322464, + 0.010503790341317654, + 0.009042543359100819, + 0.008377980440855026, + 0.008715743198990822, + 0.01383297611027956, + 0.011265727691352367, + 0.007493098732084036, + 0.005295175593346357, + 0.008435459807515144, + 0.0059840683825314045, + 0.011503087356686592, + 0.00670321611687541, + 0.0068494416773319244, + 0.01092781312763691, + 0.0045262896455824375, + 0.004177910275757313, + 0.004770485684275627, + 0.0029486974235624075, + 0.004652298055589199, + 0.004386672284454107, + 0.0043386840261518955, + 0.0049615963362157345, + 0.003691371064633131, + 0.009970853105187416, + 0.012992117553949356, + 0.011543345637619495, + 0.018766168504953384, + 0.012241357006132603, + 0.013020156882703304, + 0.016350938007235527, + 0.009335794486105442, + 0.012762108817696571, + 0.008177544921636581, + 0.01582629792392254, + 0.01157413050532341, + 0.00940400455147028, + 0.019852573052048683, + 0.010722856968641281, + 0.007301847450435162, + 0.020630493760108948, + 0.024226604029536247, + 0.02428426966071129, + 0.024886105209589005, + 0.03299188241362572, + 0.02757728286087513, + 0.01988767273724079, + 0.03761933371424675, + 0.020314369350671768, + 0.01806161180138588, + 0.02035006508231163, + 0.009403296746313572, + 0.01882304809987545, + 0.016684837639331818, + 0.009864913299679756, + 0.022551482543349266, + 0.013881267979741096, + 0.023121550679206848, + 0.011686866171658039, + 0.010019189678132534, + 0.016023455187678337, + 0.01574907824397087, + 0.008477799594402313, + 0.013635084964334965, + 0.009261835366487503, + 0.012627791613340378, + 0.008501912467181683, + 0.006314356345683336, + 0.007108430378139019, + 0.009489284828305244, + 0.006770661100745201, + 0.010505051352083683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -1.7015421390533447 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033739592880010605 + ], + "opposite_logprob_tokens": [ + -3.4087395668029785 + ], + "logprob_reward": -0.033739592880010605, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033739592880010605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07605965435504913, + 0.04708392545580864, + 0.04221394285559654, + 0.11956773698329926, + 0.09811312705278397, + 0.08093438297510147, + 0.2012728750705719, + 0.17807623744010925, + 0.07507859170436859, + 0.08159951120615005 + ], + "credit_map": [ + 0.012575908564031124, + 0.005640578921884298, + 0.009461096487939358, + 0.008044098503887653, + 0.008342412300407887, + 0.014856114983558655, + 0.009049415588378906, + 0.00809002947062254, + 0.007834883406758308, + 0.004388778004795313, + 0.005034000612795353, + 0.004984772764146328, + 0.0051734368316829205, + 0.009538956917822361, + 0.005037054885178804, + 0.005092042963951826, + 0.005092451814562082, + 0.005414572544395924, + 0.00370151549577713, + 0.006899133790284395, + 0.004782143980264664, + 0.00599364098161459, + 0.006779571995139122, + 0.00355091062374413, + 0.009199710562825203, + 0.017423179000616074, + 0.011250137351453304, + 0.02457798644900322, + 0.01109798438847065, + 0.00939476490020752, + 0.022825730964541435, + 0.013798251748085022, + 0.008490527980029583, + 0.011312291026115417, + 0.009430844336748123, + 0.018794383853673935, + 0.009506018832325935, + 0.012188696302473545, + 0.01756976917386055, + 0.010820597410202026, + 0.01053702924400568, + 0.0064633493311703205, + 0.012826423160731792, + 0.009080459363758564, + 0.008147227577865124, + 0.018697919324040413, + 0.008222177624702454, + 0.006959796883165836, + 0.01607382297515869, + 0.02170700579881668, + 0.015527772717177868, + 0.02399253472685814, + 0.053522467613220215, + 0.02160458080470562, + 0.013744398951530457, + 0.035100311040878296, + 0.026348808780312538, + 0.02507658116519451, + 0.026402726769447327, + 0.010766384191811085, + 0.02873009629547596, + 0.020511334761977196, + 0.012818343006074429, + 0.02742195315659046, + 0.009690597653388977, + 0.016022326424717903, + 0.007663022261112928, + 0.006440558936446905, + 0.00872118305414915, + 0.011522764340043068, + 0.005304089747369289, + 0.009714046493172646, + 0.010095048695802689, + 0.013347846455872059, + 0.009332329966127872, + 0.00678100623190403, + 0.007571046240627766, + 0.010887689888477325, + 0.006974636111408472, + 0.016609907150268555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.033739592880010605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up captures a portion of a clean, light-colored laminate desk, illuminated by the bright, even glow of realistic overhead office lighting. A pair of hands firmly grasps a matte black foam tube, its dense, slightly porous texture clearly visible. The tube is forcibly bent into a tight, compressed zigzag shape, the material showing visible strain and tension along its creases. In a swift and decisive action, the hands release. With dynamic, whip-like speed, the foam tube snaps instantly and violently in a straight line, its form blurred for a fraction of a second by the rapid motion. It falls to the desk surface with a soft, muffled thud, forming a perfectly straight line. Its dark, non-reflective form contrasts sharply with the pale, linear surface of its final state, creating a focused, almost clinical atmosphere.", + "seed": 15903, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_14/index_1521_seed_15903_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12356069926172494, + "vlm_logprob": -0.12356069926172494, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.12356069926172494, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014378863386809826 + ], + "opposite_logprob_tokens": [ + -5.014379024505615 + ], + "logprob_reward": -0.014378863386809826, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.014378863386809826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12228240072727203, + 0.04606145992875099, + 0.04086986556649208, + 0.0905505120754242, + 0.27063101530075073, + 0.10763085633516312, + 0.1308388113975525, + 0.04083412140607834, + 0.027542395517230034, + 0.12275850027799606 + ], + "credit_map": [ + 0.018669765442609787, + 0.015269311144948006, + 0.022683778777718544, + 0.01249273493885994, + 0.009762221947312355, + 0.01445178221911192, + 0.013775991275906563, + 0.015176822431385517, + 0.005459538195282221, + 0.005724688060581684, + 0.010382757522165775, + 0.00362871284596622, + 0.005139857996255159, + 0.0032823923975229263, + 0.005926959682255983, + 0.006516553927212954, + 0.0037062102928757668, + 0.004448374267667532, + 0.009269540198147297, + 0.0038519459776580334, + 0.006460703443735838, + 0.0026675937697291374, + 0.003960789181292057, + 0.006504710763692856, + 0.008693801239132881, + 0.013298735953867435, + 0.013554440811276436, + 0.009321141988039017, + 0.01589157246053219, + 0.008079231716692448, + 0.010657270438969135, + 0.011054318398237228, + 0.022180261090397835, + 0.027918459847569466, + 0.05156835913658142, + 0.034083057194948196, + 0.03649967163801193, + 0.0218487661331892, + 0.03709862381219864, + 0.03943381831049919, + 0.008165990933775902, + 0.008221779949963093, + 0.015317087061703205, + 0.025151923298835754, + 0.00872564222663641, + 0.010850689373910427, + 0.019272463396191597, + 0.0119252884760499, + 0.017553962767124176, + 0.01650061644613743, + 0.02039429359138012, + 0.012724092230200768, + 0.00956533569842577, + 0.01623082160949707, + 0.01488270703703165, + 0.022986983880400658, + 0.005717935040593147, + 0.0056694382801651955, + 0.005317076575011015, + 0.005454761441797018, + 0.004230402875691652, + 0.0035738423466682434, + 0.004488937556743622, + 0.0063817258924245834, + 0.0043349312618374825, + 0.004021821543574333, + 0.0038423941005021334, + 0.004190226551145315, + 0.0023014529142528772, + 0.0037009352818131447, + 0.0028870836831629276, + 0.002263549482449889, + 0.014665788970887661, + 0.009273134171962738, + 0.023595526814460754, + 0.015415750443935394, + 0.006364982109516859, + 0.012373839505016804, + 0.015759125351905823, + 0.02531035803258419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.014378863386809826 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02639647014439106 + ], + "opposite_logprob_tokens": [ + -3.6513965129852295 + ], + "logprob_reward": -0.02639647014439106, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02639647014439106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18255773186683655, + 0.2745008170604706, + 0.07976613193750381, + 0.11386604607105255, + 0.04712231084704399, + 0.11135198175907135, + 0.062280673533678055, + 0.03913913294672966, + 0.027103792876005173, + 0.062311384826898575 + ], + "credit_map": [ + 0.025206860154867172, + 0.03176692873239517, + 0.032261986285448074, + 0.019932324066758156, + 0.013456220738589764, + 0.015268204733729362, + 0.022176295518875122, + 0.022488925606012344, + 0.03550524264574051, + 0.036471135914325714, + 0.057255860418081284, + 0.020843863487243652, + 0.03631404787302017, + 0.0225317794829607, + 0.02718275785446167, + 0.03839612752199173, + 0.01162767130881548, + 0.012324834242463112, + 0.01734229177236557, + 0.0055029126815497875, + 0.00926251895725727, + 0.006586178205907345, + 0.006084216292947531, + 0.011035513132810593, + 0.012998530641198158, + 0.013432018458843231, + 0.016439039260149002, + 0.010835466906428337, + 0.01604696549475193, + 0.012516659684479237, + 0.014489780180156231, + 0.017107592895627022, + 0.003798013785853982, + 0.004363557323813438, + 0.008325408212840557, + 0.005914272740483284, + 0.005639560986310244, + 0.004314463119953871, + 0.007080957759171724, + 0.0076860785484313965, + 0.009654711000621319, + 0.01260074507445097, + 0.015774361789226532, + 0.025061648339033127, + 0.012042645364999771, + 0.009472084231674671, + 0.015963058918714523, + 0.010782727040350437, + 0.008434399031102657, + 0.007757274899631739, + 0.010362431406974792, + 0.009677205234766006, + 0.0038272691890597343, + 0.006656890735030174, + 0.007350744679570198, + 0.00821446068584919, + 0.004239863716065884, + 0.003461596556007862, + 0.004489923361688852, + 0.004772622603923082, + 0.0031624080147594213, + 0.007550547365099192, + 0.0052783130668103695, + 0.006183859892189503, + 0.004099484998732805, + 0.0017877775244414806, + 0.005548570305109024, + 0.00267052766866982, + 0.0014495247742161155, + 0.0031314664520323277, + 0.0027126451022922993, + 0.005703795701265335, + 0.008159133605659008, + 0.0029815114103257656, + 0.014053537510335445, + 0.00666635949164629, + 0.002642789389938116, + 0.005744619760662317, + 0.007123672869056463, + 0.014939761720597744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02639647014439106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004816085100174 + ], + "opposite_logprob_tokens": [ + -2.6950480937957764 + ], + "logprob_reward": -0.07004816085100174, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004816085100174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18937519192695618, + 0.11710336804389954, + 0.06655732542276382, + 0.1584487110376358, + 0.1307554692029953, + 0.11633024364709854, + 0.0753934383392334, + 0.016959208995103836, + 0.047794148325920105, + 0.08128287643194199 + ], + "credit_map": [ + 0.021674538031220436, + 0.02714146301150322, + 0.029674533754587173, + 0.025189220905303955, + 0.023804636672139168, + 0.020103629678487778, + 0.020407849922776222, + 0.021379318088293076, + 0.011467848904430866, + 0.01101894211024046, + 0.025925520807504654, + 0.01147694792598486, + 0.011282660998404026, + 0.010350602678954601, + 0.019436318427324295, + 0.016144519671797752, + 0.007488534785807133, + 0.008867238648235798, + 0.01225157082080841, + 0.006233449559658766, + 0.006374250631779432, + 0.005013189744204283, + 0.010545817203819752, + 0.009783273562788963, + 0.02117033302783966, + 0.018412603065371513, + 0.02524220198392868, + 0.015814946964383125, + 0.024633651599287987, + 0.013241086155176163, + 0.01913050375878811, + 0.02080337516963482, + 0.011728381738066673, + 0.01078051794320345, + 0.029124490916728973, + 0.014911109581589699, + 0.019007110968232155, + 0.012255231849849224, + 0.014998333528637886, + 0.017950300127267838, + 0.008993756957352161, + 0.013986189849674702, + 0.01595061644911766, + 0.023500902578234673, + 0.010331979021430016, + 0.007467355579137802, + 0.020066959783434868, + 0.016032477840781212, + 0.009786756709218025, + 0.012945071794092655, + 0.008729997090995312, + 0.009697978384792805, + 0.005704683251678944, + 0.009774530306458473, + 0.009710131213068962, + 0.009044288657605648, + 0.0023396045435220003, + 0.0020396800246089697, + 0.0030991537496447563, + 0.0020662506576627493, + 0.001266740495339036, + 0.002258381573483348, + 0.0015209941193461418, + 0.002368404995650053, + 0.006226478144526482, + 0.006621113512665033, + 0.009423458948731422, + 0.004818929824978113, + 0.003493857802823186, + 0.003898342838510871, + 0.004883136600255966, + 0.008428829722106457, + 0.010748219676315784, + 0.00316111883148551, + 0.01821739971637726, + 0.008608863689005375, + 0.003199036931619048, + 0.008500048890709877, + 0.009833459742367268, + 0.01901473104953766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.07004816085100174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562041878700256 + ], + "opposite_logprob_tokens": [ + -1.6006203889846802 + ], + "logprob_reward": -0.22562041878700256, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562041878700256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11399874091148376, + 0.046442873775959015, + 0.03794952854514122, + 0.057117167860269547, + 0.053472667932510376, + 0.34502917528152466, + 0.2547987997531891, + 0.029596487060189247, + 0.04193253442645073, + 0.01966206729412079 + ], + "credit_map": [ + 0.012315357103943825, + 0.01586092822253704, + 0.0227455236017704, + 0.015596672892570496, + 0.008877360261976719, + 0.007981888018548489, + 0.014970664866268635, + 0.015650345012545586, + 0.005791186820715666, + 0.005564623046666384, + 0.008227203972637653, + 0.004501434043049812, + 0.004886317066848278, + 0.004068627022206783, + 0.008100263774394989, + 0.005303218960762024, + 0.005128106102347374, + 0.004165121354162693, + 0.007725575473159552, + 0.003574509872123599, + 0.0035067193675786257, + 0.0026138925459235907, + 0.00543243857100606, + 0.005803166422992945, + 0.005623037461191416, + 0.005964591633528471, + 0.008767028339207172, + 0.005942325573414564, + 0.009495931677520275, + 0.004591249395161867, + 0.009450187906622887, + 0.0072828191332519054, + 0.004849477205425501, + 0.0038633309304714203, + 0.012431059032678604, + 0.006214668974280357, + 0.005592187400907278, + 0.005268142558634281, + 0.007234959863126278, + 0.008018838241696358, + 0.024051694199442863, + 0.046709973365068436, + 0.054266296327114105, + 0.09488337486982346, + 0.03259240835905075, + 0.028490189462900162, + 0.0346878282725811, + 0.029347408562898636, + 0.03142872825264931, + 0.016918053850531578, + 0.040597617626190186, + 0.02509087324142456, + 0.015891525894403458, + 0.04016603156924248, + 0.026274235919117928, + 0.05843173712491989, + 0.0019253811333328485, + 0.0026609604246914387, + 0.0032168105244636536, + 0.0034598836209625006, + 0.0017479208763688803, + 0.003115220693871379, + 0.006386935710906982, + 0.0070833731442689896, + 0.004816862754523754, + 0.002476983703672886, + 0.007748524192720652, + 0.004801950883120298, + 0.0020290629472583532, + 0.005017067771404982, + 0.00640478590503335, + 0.00863729603588581, + 0.0018991553224623203, + 0.0016068186378106475, + 0.0028714986983686686, + 0.002759653842076659, + 0.0013295363169163465, + 0.002704218728467822, + 0.0027405470609664917, + 0.003750638337805867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.22562041878700256 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135958313941956 + ], + "opposite_logprob_tokens": [ + -1.4063595533370972 + ], + "logprob_reward": -0.28135958313941956, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135958313941956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1374133974313736, + 0.054835278540849686, + 0.05347580462694168, + 0.10520382970571518, + 0.1348615437746048, + 0.1972401738166809, + 0.21342192590236664, + 0.032160282135009766, + 0.051632314920425415, + 0.019755380228161812 + ], + "credit_map": [ + 0.013281633146107197, + 0.017450088635087013, + 0.02421579323709011, + 0.019831165671348572, + 0.011353489011526108, + 0.013814112171530724, + 0.02215404249727726, + 0.015313095413148403, + 0.006457657553255558, + 0.006134363356977701, + 0.0109580522403121, + 0.005126051604747772, + 0.005259918048977852, + 0.006737822201102972, + 0.00655682198703289, + 0.007604597602039576, + 0.0062888083048164845, + 0.007477349601686001, + 0.009177451021969318, + 0.004472767934203148, + 0.00850505381822586, + 0.004907762631773949, + 0.006213554181158543, + 0.0064330631867051125, + 0.01207131426781416, + 0.01574563980102539, + 0.012757530435919762, + 0.009480980224907398, + 0.01180555485188961, + 0.010760325007140636, + 0.017025893554091454, + 0.015556604601442814, + 0.011020381934940815, + 0.013244234025478363, + 0.025048211216926575, + 0.01623527705669403, + 0.017780834808945656, + 0.017676109448075294, + 0.015723425894975662, + 0.01813308335840702, + 0.012738129124045372, + 0.034143079072237015, + 0.029218584299087524, + 0.050544194877147675, + 0.01751248724758625, + 0.014144782908260822, + 0.020381756126880646, + 0.018557172268629074, + 0.03265998885035515, + 0.01543005183339119, + 0.03941008448600769, + 0.024229608476161957, + 0.010210199281573296, + 0.030485600233078003, + 0.0170874185860157, + 0.04390901327133179, + 0.0033752440940588713, + 0.0033331532031297684, + 0.00357771385461092, + 0.004146632272750139, + 0.002872389042750001, + 0.0037066976074129343, + 0.006201990880072117, + 0.004946463741362095, + 0.006180283147841692, + 0.006958201061934233, + 0.006019412074238062, + 0.008763037621974945, + 0.0038245872128754854, + 0.008180474862456322, + 0.007046693004667759, + 0.004659631755203009, + 0.0032245556358247995, + 0.0041162967681884766, + 0.0021783967968076468, + 0.002659356454387307, + 0.0018852768698707223, + 0.0018870207713916898, + 0.0022729572374373674, + 0.0015315207419916987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.28135958313941956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up captures a portion of a clean, light-colored laminate desk, illuminated by the bright, even glow of realistic overhead office lighting. A pair of hands firmly grasps a matte black foam tube, its dense, slightly porous texture clearly visible. The tube is forcibly bent into a tight, compressed zigzag shape, the material showing visible strain and tension along its creases. In a swift and decisive action, the hands release. With dynamic, whip-like speed, the foam tube snaps instantly and violently in a straight line, its form blurred for a fraction of a second by the rapid motion. It falls to the desk surface with a soft, muffled thud, forming a perfectly straight line. Its dark, non-reflective form contrasts sharply with the pale, linear surface of its final state, creating a focused, almost clinical atmosphere.", + "seed": 15904, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_14/index_1521_seed_15904_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1565378088504076, + "vlm_logprob": -0.1565378088504076, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.1565378088504076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011263711377978325 + ], + "opposite_logprob_tokens": [ + -5.636263847351074 + ], + "logprob_reward": -0.011263711377978325, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.011263711377978325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08192475140094757, + 0.1971200555562973, + 0.08152809739112854, + 0.037002552300691605, + 0.030760781839489937, + 0.01691257581114769, + 0.035342250019311905, + 0.06590663641691208, + 0.28919073939323425, + 0.16431161761283875 + ], + "credit_map": [ + 0.0178256556391716, + 0.007891755551099777, + 0.01189655251801014, + 0.008560734800994396, + 0.008177582174539566, + 0.011838127858936787, + 0.005988608114421368, + 0.009745730087161064, + 0.017342161387205124, + 0.018215151503682137, + 0.03787388652563095, + 0.04855591058731079, + 0.00890028290450573, + 0.01854514144361019, + 0.014768362045288086, + 0.03291916102170944, + 0.015895767137408257, + 0.00834883376955986, + 0.014235722832381725, + 0.01328266505151987, + 0.004799790680408478, + 0.0099358344450593, + 0.004084586165845394, + 0.010944897308945656, + 0.006786756217479706, + 0.0029379657935351133, + 0.005919863004237413, + 0.005038256756961346, + 0.0033800178207457066, + 0.0035172172356396914, + 0.0029934276826679707, + 0.00642904918640852, + 0.0043454887345433235, + 0.003424793714657426, + 0.004988781176507473, + 0.004641421604901552, + 0.002780379494652152, + 0.004081120248883963, + 0.0026616251561790705, + 0.003837172407656908, + 0.003562299534678459, + 0.001367194578051567, + 0.002281631575897336, + 0.002260584617033601, + 0.0019736834801733494, + 0.001530829700641334, + 0.0014091251650825143, + 0.00252722785808146, + 0.005743367597460747, + 0.0052666934207081795, + 0.004658827092498541, + 0.004362636245787144, + 0.0037515226285904646, + 0.0035977307707071304, + 0.0029812140855938196, + 0.0049802567809820175, + 0.008027960546314716, + 0.011768827214837074, + 0.010494519025087357, + 0.011074277572333813, + 0.00844547525048256, + 0.005904925987124443, + 0.004862971603870392, + 0.005327673628926277, + 0.029517529532313347, + 0.03979458287358284, + 0.04618197679519653, + 0.020948832854628563, + 0.07653932273387909, + 0.04543933644890785, + 0.015698442235589027, + 0.015070748515427113, + 0.016812577843666077, + 0.017062677070498466, + 0.011058512143790722, + 0.0075530605390667915, + 0.03475131094455719, + 0.03818456456065178, + 0.012895193882286549, + 0.025993723422288895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.011263711377978325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04314012825489044 + ], + "opposite_logprob_tokens": [ + -3.168140172958374 + ], + "logprob_reward": -0.04314012825489044, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04314012825489044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4973297715187073, + 0.15894672274589539, + 0.10050149261951447, + 0.01731191761791706, + 0.02217346616089344, + 0.013048102147877216, + 0.027022769674658775, + 0.05962611362338066, + 0.05429544672369957, + 0.04974416643381119 + ], + "credit_map": [ + 0.0714106559753418, + 0.05674012005329132, + 0.06900560855865479, + 0.07337455451488495, + 0.03452357277274132, + 0.09452087432146072, + 0.034898705780506134, + 0.06285569816827774, + 0.01580187678337097, + 0.013343598693609238, + 0.02864883467555046, + 0.03474047780036926, + 0.008069535717368126, + 0.01855691522359848, + 0.01198718510568142, + 0.027798321098089218, + 0.013804846443235874, + 0.010280681774020195, + 0.012730630114674568, + 0.007860400713980198, + 0.014278935268521309, + 0.017565840855240822, + 0.012905608862638474, + 0.011074559763073921, + 0.0026873971801251173, + 0.0014264206402003765, + 0.002328475471585989, + 0.001909172278828919, + 0.0018890106584876776, + 0.002773246495053172, + 0.0015309769660234451, + 0.0027672199066728354, + 0.00232866033911705, + 0.0027581313624978065, + 0.0032193816732615232, + 0.003255401272326708, + 0.0024756486527621746, + 0.002487011021003127, + 0.0026435572654008865, + 0.0030056762043386698, + 0.0020783564541488886, + 0.0012449801433831453, + 0.001702236128039658, + 0.0018085731426253915, + 0.0017871891614049673, + 0.0012351003242656589, + 0.0013327435590326786, + 0.0018589243991300464, + 0.0038565208669751883, + 0.0029447823762893677, + 0.00406886450946331, + 0.004097522236406803, + 0.0030895250383764505, + 0.002261290792375803, + 0.0026010109577327967, + 0.004103254992514849, + 0.007680549751967192, + 0.007969401776790619, + 0.00952544342726469, + 0.00783947017043829, + 0.006030181888490915, + 0.005743901710957289, + 0.0054343207739293575, + 0.009402859024703503, + 0.008097777143120766, + 0.010454345494508743, + 0.006809327285736799, + 0.004924590699374676, + 0.007521503139287233, + 0.0065146847628057, + 0.004323888104408979, + 0.005649331025779247, + 0.004030808340758085, + 0.00569722568616271, + 0.004122746642678976, + 0.00466814124956727, + 0.0180182084441185, + 0.008697232231497765, + 0.0025011731777340174, + 0.002008636947721243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04314012825489044 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208951398730278 + ], + "opposite_logprob_tokens": [ + -2.812089443206787 + ], + "logprob_reward": -0.06208951398730278, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208951398730278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14288738369941711, + 0.21811142563819885, + 0.08186062425374985, + 0.019031360745429993, + 0.08631321042776108, + 0.08564271032810211, + 0.03678170591592789, + 0.11589063704013824, + 0.13847917318344116, + 0.0750017836689949 + ], + "credit_map": [ + 0.019044065847992897, + 0.015050693415105343, + 0.020931722596287727, + 0.01823519542813301, + 0.01658591628074646, + 0.019310256466269493, + 0.015675680711865425, + 0.01805386133491993, + 0.019369076937437057, + 0.02098678983747959, + 0.041120827198028564, + 0.05116182565689087, + 0.011698476038873196, + 0.0217482578009367, + 0.017898162826895714, + 0.03412799537181854, + 0.010900188237428665, + 0.009639368392527103, + 0.0126833850517869, + 0.00984784122556448, + 0.008713856339454651, + 0.009277488104999065, + 0.011210198514163494, + 0.009588298387825489, + 0.0016702477587386966, + 0.001841469667851925, + 0.0026279068551957607, + 0.0020363249350339174, + 0.0021232585422694683, + 0.00258682225830853, + 0.0030872218776494265, + 0.0030581094324588776, + 0.01413656398653984, + 0.005891481414437294, + 0.013607674278318882, + 0.010217341594398022, + 0.011076659895479679, + 0.012964189052581787, + 0.009235898032784462, + 0.009183402173221111, + 0.009533624164760113, + 0.007552760187536478, + 0.010103522799909115, + 0.011084577068686485, + 0.0182460006326437, + 0.011287257075309753, + 0.007981989532709122, + 0.009852987714111805, + 0.006996872369199991, + 0.004237717483192682, + 0.005421826615929604, + 0.005420126020908356, + 0.004656870383769274, + 0.0025844955816864967, + 0.002971404930576682, + 0.004492396023124456, + 0.015159405767917633, + 0.020269805565476418, + 0.018247270956635475, + 0.01457956712692976, + 0.012127935886383057, + 0.012056050822138786, + 0.012207316234707832, + 0.011243280954658985, + 0.0123540498316288, + 0.017582949250936508, + 0.019624395295977592, + 0.010155205614864826, + 0.04222508892416954, + 0.022038253024220467, + 0.008285650983452797, + 0.006213570944964886, + 0.006596041843295097, + 0.01804923079907894, + 0.007149543147534132, + 0.0035937682259827852, + 0.019364165142178535, + 0.012587092816829681, + 0.0036994286347180605, + 0.0039625102654099464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.06208951398730278 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233559608459473 + ], + "opposite_logprob_tokens": [ + -0.8983559608459473 + ], + "logprob_reward": -0.5233559608459473, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233559608459473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05125369131565094, + 0.028323782607913017, + 0.014521750621497631, + 0.024272536858916283, + 0.03365176543593407, + 0.019359534606337547, + 0.133306086063385, + 0.20483466982841492, + 0.26582038402557373, + 0.22465580701828003 + ], + "credit_map": [ + 0.008390660397708416, + 0.004687704611569643, + 0.007471752353012562, + 0.005568951368331909, + 0.006340751424431801, + 0.006488092243671417, + 0.004806663375347853, + 0.007499115541577339, + 0.0027078220155090094, + 0.0032831679563969374, + 0.004596434533596039, + 0.005605536047369242, + 0.002539644483476877, + 0.003139922395348549, + 0.0026562653947621584, + 0.0037949909456074238, + 0.002303812652826309, + 0.0013227710733190179, + 0.001835060422308743, + 0.0015838579274713993, + 0.001669270102865994, + 0.0027365190908312798, + 0.0015501214656978846, + 0.001520337536931038, + 0.0024908145423978567, + 0.0025309445336461067, + 0.0035383119247853756, + 0.0026239482685923576, + 0.002650056965649128, + 0.0028927517123520374, + 0.0030464809387922287, + 0.0044992282055318356, + 0.004290411248803139, + 0.003534995950758457, + 0.004440919030457735, + 0.004479975905269384, + 0.004390331916511059, + 0.004169158171862364, + 0.00346136256121099, + 0.0048846094869077206, + 0.002390993060544133, + 0.0015627964166924357, + 0.0021996712312102318, + 0.0025360011495649815, + 0.0034419812727719545, + 0.0025420738384127617, + 0.002074926858767867, + 0.0026110901962965727, + 0.019150083884596825, + 0.016539691016077995, + 0.018651148304343224, + 0.019203010946512222, + 0.019118960946798325, + 0.016407759860157967, + 0.011866563931107521, + 0.012368867173790932, + 0.02925463579595089, + 0.05201897770166397, + 0.037472132593393326, + 0.023462576791644096, + 0.017242446541786194, + 0.015858909115195274, + 0.014987119473516941, + 0.014537866227328777, + 0.032904159277677536, + 0.031214095652103424, + 0.060502152889966965, + 0.018864234909415245, + 0.045188892632722855, + 0.044584307819604874, + 0.017955755814909935, + 0.014606776647269726, + 0.01217347290366888, + 0.0862186998128891, + 0.008739803917706013, + 0.005965726915746927, + 0.06611515581607819, + 0.011254800483584404, + 0.02078547328710556, + 0.013402687385678291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.5233559608459473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a clean, light-colored laminate desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands visible firmly grasping a matte black foam tube?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the foam tube bent into a tight, compressed zigzag shape showing visible strain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the foam tube snap instantly and violently into a straight line with motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the foam tube fall to the desk surface and form a perfectly straight line?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428397297859192 + ], + "opposite_logprob_tokens": [ + -2.0178396701812744 + ], + "logprob_reward": -0.1428397297859192, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428397297859192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05865906551480293, + 0.0686761811375618, + 0.023847734555602074, + 0.02276545763015747, + 0.04115000739693642, + 0.024853285402059555, + 0.055045612156391144, + 0.054351504892110825, + 0.12385000288486481, + 0.5268011689186096 + ], + "credit_map": [ + 0.008095882833003998, + 0.005007010418921709, + 0.009215480647981167, + 0.007478047627955675, + 0.003976707346737385, + 0.0091771325096488, + 0.004979210440069437, + 0.010729587636888027, + 0.006350518204271793, + 0.004764888435602188, + 0.014206932857632637, + 0.018828406929969788, + 0.0027597572188824415, + 0.00514568155631423, + 0.004737467970699072, + 0.011882521212100983, + 0.005374326836317778, + 0.0022059129551053047, + 0.0036500105634331703, + 0.0029816487804055214, + 0.002750291023403406, + 0.0021124619524925947, + 0.0022284910082817078, + 0.00254459073767066, + 0.0019280359847471118, + 0.0025793390814214945, + 0.0029279489535838366, + 0.0022809437941759825, + 0.002979069948196411, + 0.003268561791628599, + 0.002697901101782918, + 0.004103655926883221, + 0.005532894283533096, + 0.004451190587133169, + 0.006036372389644384, + 0.006057577207684517, + 0.0034438741859048605, + 0.004826965741813183, + 0.0033992731478065252, + 0.00740185659378767, + 0.0039053724613040686, + 0.0021022981964051723, + 0.0035466295666992664, + 0.003906261641532183, + 0.0029337673913687468, + 0.0020611851941794157, + 0.0020383954979479313, + 0.00435937475413084, + 0.007642112206667662, + 0.008655997924506664, + 0.007237485609948635, + 0.00859812181442976, + 0.006627501454204321, + 0.005353729706257582, + 0.005362273193895817, + 0.0055683874525129795, + 0.007839168421924114, + 0.01087398175150156, + 0.008736655116081238, + 0.008655622601509094, + 0.005231585819274187, + 0.0049097128212451935, + 0.004113743081688881, + 0.003991033881902695, + 0.013020693324506283, + 0.014081570319831371, + 0.022442122921347618, + 0.008670804090797901, + 0.02918936498463154, + 0.019692368805408478, + 0.010045629926025867, + 0.006707441993057728, + 0.024795571342110634, + 0.22750787436962128, + 0.017672523856163025, + 0.0134359672665596, + 0.14447344839572906, + 0.05587036907672882, + 0.012669695541262627, + 0.030375707894563675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.1428397297859192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_15/index_tensor([4134, 4134])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_15/index_tensor([4134, 4134])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9c9f50db92331ce817affb0ba2d64db04fadc9e0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_15/index_tensor([4134, 4134])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This dynamic, eye-level medium-long shot, captured by a sleek onboard camera, perfectly parallels a young woman with striking biracial features carving a vast, pristine snow slope. Her warm olive skin contrasts starkly against the bright white surroundings, her focus absolute, her dark almond-shaped eyes visible behind large, iridescent goggles reflecting the deep blue of the sky. A long, dark braid tucks down her back from beneath a matte black helmet. She wears a vibrant teal high-performance snowboard jacket and stylish black snow pants, adding a pop of color against the monochrome landscape. With a low, athletic stance and powerful precision, she fluidly shifts her weight from heel edge to toe edge, her body a study in controlled grace. As the edge of her ski slices into the untouched powder, it sends a spectacular plume of fine, crystalline snow billowing into the fresh mountain air, gleaming intensely in the bright, high-altitude sunlight. The scene is bathed in the harsh yet beautiful light of a clear day in the mountains, creating sharp and dramatic shadows on the surface of the snow and an exhilarating atmosphere of speed, freedom and pure, unadulterated joy.", + "seed": 167, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_15/index_4134_seed_167_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13737371019087732, + "vlm_logprob": -0.13737371019087732, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.13737371019087732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1841856837272644 + ], + "opposite_logprob_tokens": [ + -2.559185743331909 + ], + "logprob_reward": -0.1841856837272644, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1841856837272644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17183387279510498, + 0.07099796086549759, + 0.12212690711021423, + 0.12301714718341827, + 0.06035056710243225, + 0.08809268474578857, + 0.12276410311460495, + 0.08240856975317001, + 0.09366367757320404, + 0.06474454700946808 + ], + "credit_map": [ + 0.02471894770860672, + 0.019916146993637085, + 0.01998455636203289, + 0.033388007432222366, + 0.01913408748805523, + 0.014807596802711487, + 0.01803290657699108, + 0.02185163088142872, + 0.008769385516643524, + 0.008952216245234013, + 0.008425725623965263, + 0.009715646505355835, + 0.013740935362875462, + 0.007777864579111338, + 0.006767585873603821, + 0.006848600693047047, + 0.012249957770109177, + 0.014886977151036263, + 0.013501298613846302, + 0.01935729756951332, + 0.0191674642264843, + 0.015692811459302902, + 0.015394017100334167, + 0.011877081356942654, + 0.014395270496606827, + 0.019351553171873093, + 0.016310496255755424, + 0.015144004486501217, + 0.013313706032931805, + 0.019428491592407227, + 0.010215658694505692, + 0.014857969246804714, + 0.0072824349626898766, + 0.009276334196329117, + 0.005918333772569895, + 0.0035609148908406496, + 0.011824147775769234, + 0.009380050003528595, + 0.007710401434451342, + 0.005397948436439037, + 0.007619406096637249, + 0.010595534928143024, + 0.010846436023712158, + 0.0068254307843744755, + 0.022195586934685707, + 0.011455336585640907, + 0.009049767628312111, + 0.009505181573331356, + 0.009884103201329708, + 0.01618497632443905, + 0.028888480737805367, + 0.019472183659672737, + 0.015111141838133335, + 0.009392631240189075, + 0.013779602013528347, + 0.010050979442894459, + 0.009246142581105232, + 0.01102607510983944, + 0.009840200655162334, + 0.010942857712507248, + 0.010352949611842632, + 0.011970555409789085, + 0.012784955091774464, + 0.006244831718504429, + 0.013740298338234425, + 0.011874021030962467, + 0.01561846025288105, + 0.015556203201413155, + 0.00993411522358656, + 0.010035993531346321, + 0.008142979815602303, + 0.008761605247855186, + 0.008195707574486732, + 0.008024212904274464, + 0.014157777652144432, + 0.007542594335973263, + 0.008336178958415985, + 0.006366466637700796, + 0.005241472274065018, + 0.006880137138068676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.1841856837272644 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135247230529785 + ], + "opposite_logprob_tokens": [ + -1.3135247230529785 + ], + "logprob_reward": -0.3135247230529785, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135247230529785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.136399507522583, + 0.08463925868272781, + 0.14813752472400665, + 0.17136545479297638, + 0.09543340653181076, + 0.09125380963087082, + 0.029205596074461937, + 0.11022704839706421, + 0.041164543479681015, + 0.09217389672994614 + ], + "credit_map": [ + 0.013326479122042656, + 0.01756935752928257, + 0.013562003150582314, + 0.0206266138702631, + 0.025429172441363335, + 0.012828798033297062, + 0.017353663221001625, + 0.015703434124588966, + 0.01088488008826971, + 0.011331181041896343, + 0.005275042727589607, + 0.0077773514203727245, + 0.022590262815356255, + 0.011752563528716564, + 0.007392251864075661, + 0.0076357233338057995, + 0.014676145277917385, + 0.01919226534664631, + 0.013942666351795197, + 0.020017948001623154, + 0.02817639894783497, + 0.01569397933781147, + 0.023277245461940765, + 0.013160877861082554, + 0.01740410551428795, + 0.022055745124816895, + 0.018784306943416595, + 0.01666736789047718, + 0.025489144027233124, + 0.02713690884411335, + 0.019521821290254593, + 0.024306057021021843, + 0.008897315710783005, + 0.015210800804197788, + 0.008290346711874008, + 0.008751711808145046, + 0.01768006570637226, + 0.010888983495533466, + 0.01393087301403284, + 0.011783304624259472, + 0.00889729242771864, + 0.015762755647301674, + 0.010101775638759136, + 0.005751519929617643, + 0.018875962123274803, + 0.011881551705300808, + 0.010517367161810398, + 0.009465586394071579, + 0.002771936357021332, + 0.003675491316244006, + 0.005508867558091879, + 0.0034986704122275114, + 0.005325163248926401, + 0.002355733420699835, + 0.003638086374849081, + 0.002431647852063179, + 0.012482612393796444, + 0.017815234139561653, + 0.009061136282980442, + 0.012277746573090553, + 0.015264471061527729, + 0.01698244735598564, + 0.016802506521344185, + 0.009540896862745285, + 0.004044889472424984, + 0.004061838611960411, + 0.008545679040253162, + 0.007393188774585724, + 0.004920192528516054, + 0.0030476863030344248, + 0.004606789909303188, + 0.004544278606772423, + 0.009563534520566463, + 0.01565222255885601, + 0.016272006556391716, + 0.00983442459255457, + 0.015855757519602776, + 0.011467020958662033, + 0.005842143669724464, + 0.007686789147555828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.3135247230529785 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024921807926148176 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.0024921807926148176, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024921807926148176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.169011652469635, + 0.054789818823337555, + 0.06827731430530548, + 0.13326457142829895, + 0.04319066181778908, + 0.1253558248281479, + 0.03263477981090546, + 0.1312445104122162, + 0.10203128308057785, + 0.14019961655139923 + ], + "credit_map": [ + 0.017741156741976738, + 0.024383515119552612, + 0.01398441568017006, + 0.02540496364235878, + 0.030597420409321785, + 0.015328052453696728, + 0.023118628188967705, + 0.018453499302268028, + 0.006923974491655827, + 0.008249553851783276, + 0.004531522747129202, + 0.00409997534006834, + 0.014392356388270855, + 0.006670793052762747, + 0.0042432802729308605, + 0.005678359419107437, + 0.010469788685441017, + 0.009361584670841694, + 0.0052988058887422085, + 0.005438872147351503, + 0.015069695189595222, + 0.00877713318914175, + 0.008138640783727169, + 0.005722790025174618, + 0.013659374788403511, + 0.015369588509202003, + 0.015592537820339203, + 0.012466914020478725, + 0.020805541425943375, + 0.017863379791378975, + 0.01877979002892971, + 0.01872744970023632, + 0.00444054463878274, + 0.005238817073404789, + 0.003989244345575571, + 0.004817495588213205, + 0.007672890089452267, + 0.004919826984405518, + 0.00696770241484046, + 0.005144134163856506, + 0.011856894008815289, + 0.019740963354706764, + 0.01604374498128891, + 0.009262915700674057, + 0.028421586379408836, + 0.014365299604833126, + 0.015652082860469818, + 0.01001233235001564, + 0.0030348668806254864, + 0.0044085197150707245, + 0.0053168125450611115, + 0.003965850919485092, + 0.005284586921334267, + 0.0029041145462542772, + 0.004328915849328041, + 0.003391109872609377, + 0.014274397864937782, + 0.020936299115419388, + 0.011425002478063107, + 0.01411400455981493, + 0.0201457217335701, + 0.01950732432305813, + 0.018488917499780655, + 0.012352834455668926, + 0.007993258535861969, + 0.010882810689508915, + 0.022987086325883865, + 0.017194489017128944, + 0.013501842506229877, + 0.010300961323082447, + 0.009447307325899601, + 0.009723521769046783, + 0.015868021175265312, + 0.016826706007122993, + 0.023916173726320267, + 0.017775725573301315, + 0.026186184957623482, + 0.016377484425902367, + 0.009856796823441982, + 0.013392522931098938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0024921807926148176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338066905736923 + ], + "opposite_logprob_tokens": [ + -3.6513381004333496 + ], + "logprob_reward": -0.026338066905736923, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338066905736923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12982383370399475, + 0.07249040901660919, + 0.08935786783695221, + 0.06477892398834229, + 0.03574584424495697, + 0.09435340017080307, + 0.08699454367160797, + 0.07437336444854736, + 0.21910670399665833, + 0.13297511637210846 + ], + "credit_map": [ + 0.009389493614435196, + 0.012810780666768551, + 0.014236720278859138, + 0.027856944128870964, + 0.01691279374063015, + 0.007454244419932365, + 0.02273869514465332, + 0.01842416077852249, + 0.010758799500763416, + 0.009290451183915138, + 0.00642507104203105, + 0.007687289733439684, + 0.016999078914523125, + 0.0076824030838906765, + 0.004960613325238228, + 0.008686699904501438, + 0.010261236689984798, + 0.009906201623380184, + 0.014003906399011612, + 0.016662942245602608, + 0.009680775925517082, + 0.010329453274607658, + 0.012295307591557503, + 0.006218048743903637, + 0.006149744614958763, + 0.008871921338140965, + 0.0076032476499676704, + 0.006401546765118837, + 0.010442806407809258, + 0.010347927920520306, + 0.008703877218067646, + 0.006257855799049139, + 0.0033934032544493675, + 0.004472276195883751, + 0.00378473661839962, + 0.0033760303631424904, + 0.006692891474813223, + 0.005576686467975378, + 0.005366557277739048, + 0.003083260962739587, + 0.009693054482340813, + 0.013469308614730835, + 0.017758453264832497, + 0.008388751186430454, + 0.01706213317811489, + 0.009659296832978725, + 0.010279192589223385, + 0.00804321002215147, + 0.006503374315798283, + 0.011656239628791809, + 0.014633565209805965, + 0.010146641172468662, + 0.017419403418898582, + 0.009337169118225574, + 0.009500531479716301, + 0.007797622121870518, + 0.0071535720489919186, + 0.01033089216798544, + 0.011002583429217339, + 0.012072047218680382, + 0.006888401694595814, + 0.008487232960760593, + 0.010316179133951664, + 0.008122449740767479, + 0.016887538135051727, + 0.022823654115200043, + 0.03871345520019531, + 0.032424114644527435, + 0.034045737236738205, + 0.02133891172707081, + 0.027140038087964058, + 0.025733262300491333, + 0.013925474137067795, + 0.021005187183618546, + 0.030027616769075394, + 0.018474813550710678, + 0.01934230700135231, + 0.011598670855164528, + 0.008290037512779236, + 0.010311010293662548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.026338066905736923 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032789647579193 + ], + "opposite_logprob_tokens": [ + -1.9103279113769531 + ], + "logprob_reward": -0.16032789647579193, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032789647579193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15207284688949585, + 0.061413586139678955, + 0.04879026114940643, + 0.05065273866057396, + 0.029005374759435654, + 0.05385875701904297, + 0.05704022943973541, + 0.06846887618303299, + 0.32953962683677673, + 0.14915777742862701 + ], + "credit_map": [ + 0.017962947487831116, + 0.016247661784291267, + 0.017432060092687607, + 0.024522876366972923, + 0.021619047969579697, + 0.014253356494009495, + 0.02014204114675522, + 0.019892843440175056, + 0.007921764627099037, + 0.009192521683871746, + 0.005128615070134401, + 0.00574106490239501, + 0.012698673643171787, + 0.008728262968361378, + 0.006328223738819361, + 0.005674455780535936, + 0.006764879450201988, + 0.005881821271032095, + 0.005589617416262627, + 0.005544311832636595, + 0.009711978025734425, + 0.005881350953131914, + 0.004745667800307274, + 0.0046706306748092175, + 0.005188412498682737, + 0.005976465530693531, + 0.006330614443868399, + 0.006112783681601286, + 0.008460596203804016, + 0.006278857588768005, + 0.0055607822723686695, + 0.006744219921529293, + 0.002649389673024416, + 0.0035682907328009605, + 0.0032987911254167557, + 0.0032018106430768967, + 0.004214753862470388, + 0.004317545797675848, + 0.0029654353857040405, + 0.004789355210959911, + 0.006483281962573528, + 0.006906153634190559, + 0.012009337544441223, + 0.006168833002448082, + 0.007511321920901537, + 0.004324732348322868, + 0.004733920097351074, + 0.005721171386539936, + 0.004700522869825363, + 0.008216881193220615, + 0.011283593252301216, + 0.0077523463405668736, + 0.007509093265980482, + 0.005526065360754728, + 0.006806003395467997, + 0.005245715379714966, + 0.007554815150797367, + 0.008719372563064098, + 0.011990691535174847, + 0.009122902527451515, + 0.006828084122389555, + 0.007193516939878464, + 0.010102716274559498, + 0.006956768222153187, + 0.037471890449523926, + 0.035909172147512436, + 0.09018915146589279, + 0.06627175956964493, + 0.020744873210787773, + 0.020724421367049217, + 0.026352865621447563, + 0.03187546133995056, + 0.016648583114147186, + 0.01792527362704277, + 0.052949439734220505, + 0.025442833080887794, + 0.010106871835887432, + 0.009330661967396736, + 0.007321625482290983, + 0.009432459250092506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.16032789647579193 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This dynamic, eye-level medium-long shot, captured by a sleek onboard camera, perfectly parallels a young woman with striking biracial features carving a vast, pristine snow slope. Her warm olive skin contrasts starkly against the bright white surroundings, her focus absolute, her dark almond-shaped eyes visible behind large, iridescent goggles reflecting the deep blue of the sky. A long, dark braid tucks down her back from beneath a matte black helmet. She wears a vibrant teal high-performance snowboard jacket and stylish black snow pants, adding a pop of color against the monochrome landscape. With a low, athletic stance and powerful precision, she fluidly shifts her weight from heel edge to toe edge, her body a study in controlled grace. As the edge of her ski slices into the untouched powder, it sends a spectacular plume of fine, crystalline snow billowing into the fresh mountain air, gleaming intensely in the bright, high-altitude sunlight. The scene is bathed in the harsh yet beautiful light of a clear day in the mountains, creating sharp and dramatic shadows on the surface of the snow and an exhilarating atmosphere of speed, freedom and pure, unadulterated joy.", + "seed": 168, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_15/index_4134_seed_168_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13459940669126808, + "vlm_logprob": -0.13459940669126808, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.13459940669126808, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1593143194913864 + ], + "opposite_logprob_tokens": [ + -2.9093143939971924 + ], + "logprob_reward": -0.1593143194913864, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.1593143194913864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09923255443572998, + 0.05897524580359459, + 0.06128251925110817, + 0.11216866970062256, + 0.0628163069486618, + 0.06610400974750519, + 0.028485380113124847, + 0.09978418797254562, + 0.33430615067481995, + 0.07684493064880371 + ], + "credit_map": [ + 0.014268292114138603, + 0.01481590699404478, + 0.014425667002797127, + 0.013181278482079506, + 0.01177817303687334, + 0.008884048089385033, + 0.010049229487776756, + 0.011829952709376812, + 0.008013314567506313, + 0.00873086042702198, + 0.00699985958635807, + 0.00756444176658988, + 0.0097153689712286, + 0.00576768396422267, + 0.0062363650649785995, + 0.005947353318333626, + 0.004195086658000946, + 0.006923613604158163, + 0.010974863544106483, + 0.010829146951436996, + 0.007133319973945618, + 0.008615192025899887, + 0.004953988827764988, + 0.007657305337488651, + 0.016679540276527405, + 0.01901547610759735, + 0.012576881796121597, + 0.009210037998855114, + 0.018317850306630135, + 0.014251316897571087, + 0.010271395556628704, + 0.011846178211271763, + 0.006812995299696922, + 0.006534579209983349, + 0.009740348905324936, + 0.008024710230529308, + 0.008357936516404152, + 0.00859841424971819, + 0.007358592003583908, + 0.007388736121356487, + 0.007977793924510479, + 0.008333142846822739, + 0.009912184439599514, + 0.006877150386571884, + 0.009951641783118248, + 0.009401203133165836, + 0.005962382536381483, + 0.007688510697335005, + 0.0045964159071445465, + 0.003418817650526762, + 0.0029346365481615067, + 0.0038166993763297796, + 0.00391364423558116, + 0.0038250379730015993, + 0.0026535638608038425, + 0.0033265631645917892, + 0.012346256524324417, + 0.010147198103368282, + 0.012574625201523304, + 0.013755876570940018, + 0.01020735688507557, + 0.019674474373459816, + 0.006344987079501152, + 0.014733412303030491, + 0.0156293623149395, + 0.12775875627994537, + 0.009004466235637665, + 0.01150451973080635, + 0.012923862785100937, + 0.13889653980731964, + 0.0061543588526546955, + 0.012434294447302818, + 0.005859195254743099, + 0.016173314303159714, + 0.006651391740888357, + 0.0064596994780004025, + 0.014754385687410831, + 0.013939137570559978, + 0.007265270687639713, + 0.005742539186030626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.1593143194913864 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31347155570983887 + ], + "opposite_logprob_tokens": [ + -1.3134715557098389 + ], + "logprob_reward": -0.31347155570983887, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31347155570983887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13304999470710754, + 0.13650082051753998, + 0.09403954446315765, + 0.06694727391004562, + 0.07768082618713379, + 0.09353611618280411, + 0.11991750448942184, + 0.10391148179769516, + 0.1265663206577301, + 0.04785013943910599 + ], + "credit_map": [ + 0.019464289769530296, + 0.019608011469244957, + 0.014081205241382122, + 0.01302138902246952, + 0.024005955085158348, + 0.011426065117120743, + 0.01679341495037079, + 0.014649661257863045, + 0.01736079342663288, + 0.020088644698262215, + 0.01619945466518402, + 0.016325974836945534, + 0.023109734058380127, + 0.01675753854215145, + 0.012927353382110596, + 0.013731315732002258, + 0.007464922033250332, + 0.011457659304141998, + 0.013564223423600197, + 0.010696657933294773, + 0.015060659497976303, + 0.01352127268910408, + 0.012698483653366566, + 0.009575659409165382, + 0.005453227087855339, + 0.009499069303274155, + 0.010892875492572784, + 0.006208418402820826, + 0.012648802250623703, + 0.007316906470805407, + 0.0072490861639380455, + 0.007678881287574768, + 0.008133228868246078, + 0.008165059611201286, + 0.008416361175477505, + 0.010054112412035465, + 0.013880012556910515, + 0.007226888090372086, + 0.01109004020690918, + 0.010715126059949398, + 0.012653563171625137, + 0.018419766798615456, + 0.01156041119247675, + 0.008546795696020126, + 0.01894293539226055, + 0.008286349475383759, + 0.006421296391636133, + 0.008704992942512035, + 0.011522716842591763, + 0.014500219374895096, + 0.017606308683753014, + 0.019840408116579056, + 0.016200363636016846, + 0.014716031961143017, + 0.011126255616545677, + 0.014405196532607079, + 0.010478843934834003, + 0.014778155833482742, + 0.013384979218244553, + 0.012331347912549973, + 0.017588842660188675, + 0.01459837332367897, + 0.01025870069861412, + 0.0104922354221344, + 0.0044840420596301556, + 0.05249893665313721, + 0.0027888629119843245, + 0.002342505846172571, + 0.005323078483343124, + 0.05370023474097252, + 0.002282171742990613, + 0.0031464744824916124, + 0.0068589262664318085, + 0.006745802704244852, + 0.0056990101002156734, + 0.004743504803627729, + 0.00835394486784935, + 0.00692002335563302, + 0.0035697349812835455, + 0.004959189333021641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.31347155570983887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031960143242031336 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.0031960143242031336, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031960143242031336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09680328518152237, + 0.08118891716003418, + 0.10618384182453156, + 0.0793415978550911, + 0.06056489795446396, + 0.0837029293179512, + 0.08021441847085953, + 0.041516147553920746, + 0.2613104283809662, + 0.10917353630065918 + ], + "credit_map": [ + 0.010803469456732273, + 0.013533473014831543, + 0.01554405689239502, + 0.015778344124555588, + 0.012365901842713356, + 0.00910602044314146, + 0.009454404003918171, + 0.01021761354058981, + 0.009527352638542652, + 0.011493274010717869, + 0.00914990808814764, + 0.008542835712432861, + 0.013635070063173771, + 0.009088931605219841, + 0.010745787061750889, + 0.009005755186080933, + 0.008878632448613644, + 0.015049053356051445, + 0.01467723585665226, + 0.013295522890985012, + 0.017562201246619225, + 0.014075476676225662, + 0.013077341951429844, + 0.009568379260599613, + 0.00812916923314333, + 0.012488413602113724, + 0.012135456316173077, + 0.006376658566296101, + 0.012917486019432545, + 0.009016497991979122, + 0.00949000846594572, + 0.0087879104539752, + 0.004831181839108467, + 0.006789465434849262, + 0.006860589142888784, + 0.007466393057256937, + 0.010171189904212952, + 0.005078716669231653, + 0.00939225684851408, + 0.009975103661417961, + 0.010822266340255737, + 0.0145339947193861, + 0.010220693424344063, + 0.005922388751059771, + 0.018799476325511932, + 0.007806980516761541, + 0.007111141923815012, + 0.008485984988510609, + 0.009899459779262543, + 0.011939885094761848, + 0.006990945432335138, + 0.00932015385478735, + 0.009332418441772461, + 0.01490840595215559, + 0.008950527757406235, + 0.008872617036104202, + 0.004389715380966663, + 0.006171727553009987, + 0.005171176977455616, + 0.005313209258019924, + 0.005973885767161846, + 0.00582938501611352, + 0.004110381007194519, + 0.0045566679909825325, + 0.011628614738583565, + 0.0977892056107521, + 0.00783984549343586, + 0.00957639142870903, + 0.010439887642860413, + 0.10805123299360275, + 0.004712556954473257, + 0.011272708885371685, + 0.015808578580617905, + 0.03135097399353981, + 0.0064906529150903225, + 0.006188242230564356, + 0.009820040315389633, + 0.0280270017683506, + 0.004447400104254484, + 0.007040639873594046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.0031960143242031336 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000236958265305 + ], + "opposite_logprob_tokens": [ + -2.695002317428589 + ], + "logprob_reward": -0.07000236958265305, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000236958265305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09639569371938705, + 0.0510200597345829, + 0.09041357785463333, + 0.09133017808198929, + 0.03960246592760086, + 0.03796698525547981, + 0.1773267239332199, + 0.04408617317676544, + 0.1250137835741043, + 0.2468443512916565 + ], + "credit_map": [ + 0.010522295720875263, + 0.013001530431210995, + 0.011291008442640305, + 0.0143775949254632, + 0.012792783789336681, + 0.010908657684922218, + 0.013941830024123192, + 0.0095599927008152, + 0.005205989815294743, + 0.0069848960265517235, + 0.006630319636315107, + 0.006330739241093397, + 0.009231609292328358, + 0.00437042536213994, + 0.006608746014535427, + 0.005657332483679056, + 0.007797445170581341, + 0.0124812051653862, + 0.010305339470505714, + 0.01413626316934824, + 0.017517533153295517, + 0.009164286777377129, + 0.0108100026845932, + 0.00820150226354599, + 0.011139729991555214, + 0.014403481967747211, + 0.00930935051292181, + 0.007685456424951553, + 0.018657710403203964, + 0.011613653041422367, + 0.009986123070120811, + 0.008534674532711506, + 0.005212710704654455, + 0.005315398797392845, + 0.003549317829310894, + 0.004006967414170504, + 0.009008802473545074, + 0.004061214160174131, + 0.004536770284175873, + 0.003911284264177084, + 0.0050385999493300915, + 0.00524324644356966, + 0.005734323989599943, + 0.0030394464265555143, + 0.005641022697091103, + 0.006523245945572853, + 0.002773809479549527, + 0.003973287530243397, + 0.022498037666082382, + 0.02636834606528282, + 0.01893802545964718, + 0.02066877856850624, + 0.0168831255286932, + 0.03802892938256264, + 0.014725800603628159, + 0.01921568438410759, + 0.00393393449485302, + 0.006345883943140507, + 0.004399011377245188, + 0.004831523634493351, + 0.008442393504083157, + 0.008733894675970078, + 0.002878493396565318, + 0.004521036986261606, + 0.008395034819841385, + 0.04958977922797203, + 0.00463789701461792, + 0.0038080334197729826, + 0.006397238466888666, + 0.0427069365978241, + 0.005034542642533779, + 0.004444329533725977, + 0.029472166672348976, + 0.06837160140275955, + 0.010109915398061275, + 0.012337889522314072, + 0.03441844880580902, + 0.05720595270395279, + 0.017472002655267715, + 0.01745636761188507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.07000236958265305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12701277434825897 + ], + "opposite_logprob_tokens": [ + -2.1270127296447754 + ], + "logprob_reward": -0.12701277434825897, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12701277434825897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1496793031692505, + 0.06987947970628738, + 0.07202420383691788, + 0.0801362544298172, + 0.034662842750549316, + 0.0593172051012516, + 0.06188608705997467, + 0.13211393356323242, + 0.28176841139793396, + 0.058532316237688065 + ], + "credit_map": [ + 0.02022484876215458, + 0.02243468165397644, + 0.019796419888734818, + 0.022619254887104034, + 0.017705820500850677, + 0.02224298194050789, + 0.01366212498396635, + 0.010993151925504208, + 0.009538008831441402, + 0.009637420997023582, + 0.009271630086004734, + 0.00930179562419653, + 0.011204570531845093, + 0.006522087380290031, + 0.007826427929103374, + 0.006577534135431051, + 0.011257327161729336, + 0.013774449005723, + 0.009324069134891033, + 0.0074014305137097836, + 0.009874801151454449, + 0.007280325051397085, + 0.006929265800863504, + 0.006182523909956217, + 0.013232405297458172, + 0.013587383553385735, + 0.00874332059174776, + 0.006881866604089737, + 0.011371903121471405, + 0.011265468783676624, + 0.006972923409193754, + 0.008080972358584404, + 0.004114439710974693, + 0.0037626060657203197, + 0.004432550631463528, + 0.004218918737024069, + 0.005507539492100477, + 0.005124503280967474, + 0.003488987684249878, + 0.00401329668238759, + 0.0070777335204184055, + 0.0072726900689303875, + 0.009321230463683605, + 0.00806141272187233, + 0.005246671382337809, + 0.009882507845759392, + 0.0046197352930903435, + 0.00783521682024002, + 0.005667663644999266, + 0.004875474143773317, + 0.010303419083356857, + 0.013230997137725353, + 0.0053542787209153175, + 0.007667578291147947, + 0.005449801217764616, + 0.009336872957646847, + 0.016080930829048157, + 0.013828746974468231, + 0.02310945838689804, + 0.015468478202819824, + 0.009255384095013142, + 0.027199236676096916, + 0.006460606586188078, + 0.020711075514554977, + 0.01116753090173006, + 0.11560650169849396, + 0.007300254423171282, + 0.007472851313650608, + 0.010772842913866043, + 0.1178031638264656, + 0.003771515330299735, + 0.007873709313571453, + 0.012124643661081791, + 0.012116829864680767, + 0.004822429269552231, + 0.004702257923781872, + 0.006735370960086584, + 0.008244447410106659, + 0.003615931374952197, + 0.006170402280986309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.12701277434825897 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This dynamic, eye-level medium-long shot, captured by a sleek onboard camera, perfectly parallels a young woman with striking biracial features carving a vast, pristine snow slope. Her warm olive skin contrasts starkly against the bright white surroundings, her focus absolute, her dark almond-shaped eyes visible behind large, iridescent goggles reflecting the deep blue of the sky. A long, dark braid tucks down her back from beneath a matte black helmet. She wears a vibrant teal high-performance snowboard jacket and stylish black snow pants, adding a pop of color against the monochrome landscape. With a low, athletic stance and powerful precision, she fluidly shifts her weight from heel edge to toe edge, her body a study in controlled grace. As the edge of her ski slices into the untouched powder, it sends a spectacular plume of fine, crystalline snow billowing into the fresh mountain air, gleaming intensely in the bright, high-altitude sunlight. The scene is bathed in the harsh yet beautiful light of a clear day in the mountains, creating sharp and dramatic shadows on the surface of the snow and an exhilarating atmosphere of speed, freedom and pure, unadulterated joy.", + "seed": 169, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_15/index_4134_seed_169_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12461204933933914, + "vlm_logprob": -0.12461204933933914, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.12461204933933914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16693228483200073 + ], + "opposite_logprob_tokens": [ + -2.7919323444366455 + ], + "logprob_reward": -0.16693228483200073, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.16693228483200073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1321675330400467, + 0.030016399919986725, + 0.26512160897254944, + 0.07866208255290985, + 0.08795241266489029, + 0.07612956315279007, + 0.056853000074625015, + 0.07088306546211243, + 0.13877174258232117, + 0.0634426474571228 + ], + "credit_map": [ + 0.013232924975454807, + 0.016233326867222786, + 0.016733495518565178, + 0.018336497247219086, + 0.024343211203813553, + 0.014266233891248703, + 0.013447613455355167, + 0.015574219636619091, + 0.0035124137066304684, + 0.0045976052060723305, + 0.0018709487048909068, + 0.0028076288290321827, + 0.007715261075645685, + 0.0022228755988180637, + 0.003984446637332439, + 0.0033052191138267517, + 0.06111295521259308, + 0.08289599418640137, + 0.00716549064964056, + 0.020102282986044884, + 0.018517866730690002, + 0.011583002284169197, + 0.01621079631149769, + 0.047533195465803146, + 0.01563693769276142, + 0.015380213968455791, + 0.010669630952179432, + 0.00850505568087101, + 0.008589477278292179, + 0.00756382429972291, + 0.004643221851438284, + 0.00767372315749526, + 0.010566066019237041, + 0.0161112230271101, + 0.007407281082123518, + 0.007470956537872553, + 0.016043463721871376, + 0.010420050472021103, + 0.007188796065747738, + 0.01274457573890686, + 0.009388196282088757, + 0.013986959122121334, + 0.004453515168279409, + 0.005613196175545454, + 0.01659139059484005, + 0.012799317017197609, + 0.007371113635599613, + 0.0059258718974888325, + 0.005669925827533007, + 0.006089323200285435, + 0.00593243632465601, + 0.006240624934434891, + 0.012683851644396782, + 0.008445633575320244, + 0.00871453806757927, + 0.003076663240790367, + 0.01287384144961834, + 0.00826321728527546, + 0.005634910427033901, + 0.00485254917293787, + 0.01673346571624279, + 0.011902747675776482, + 0.007067891303449869, + 0.003554434049874544, + 0.02990487590432167, + 0.024613292887806892, + 0.021075792610645294, + 0.010543886572122574, + 0.014406728558242321, + 0.016117116436362267, + 0.012737822718918324, + 0.009372211061418056, + 0.006633552256971598, + 0.01087193563580513, + 0.0048853931948542595, + 0.005578990560024977, + 0.015622210688889027, + 0.009719819761812687, + 0.007205831818282604, + 0.002924910979345441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.16693228483200073 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28136733174324036 + ], + "opposite_logprob_tokens": [ + -1.406367301940918 + ], + "logprob_reward": -0.28136733174324036, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28136733174324036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1066124439239502, + 0.044669490307569504, + 0.3809265196323395, + 0.0521569587290287, + 0.0220668725669384, + 0.040824294090270996, + 0.03116103634238243, + 0.033126480877399445, + 0.09646862000226974, + 0.1919872760772705 + ], + "credit_map": [ + 0.009320931509137154, + 0.015029973350465298, + 0.012967982329428196, + 0.01594998501241207, + 0.01733614131808281, + 0.012359733693301678, + 0.010518747381865978, + 0.01312894094735384, + 0.00789313018321991, + 0.00833189394325018, + 0.0025007601361721754, + 0.0031463196501135826, + 0.011072195135056973, + 0.003143724985420704, + 0.004750227089971304, + 0.0038312342949211597, + 0.0784815102815628, + 0.11010625958442688, + 0.007890641689300537, + 0.03511713445186615, + 0.026261165738105774, + 0.014945227652788162, + 0.027493000030517578, + 0.08063157647848129, + 0.007388518191874027, + 0.009543297812342644, + 0.0061348807066679, + 0.004877208266407251, + 0.008973798714578152, + 0.005774285644292831, + 0.004797010682523251, + 0.004667961969971657, + 0.0027296016924083233, + 0.004040399566292763, + 0.0020061044488102198, + 0.0022395651321858168, + 0.003912126645445824, + 0.0026209403295069933, + 0.001672000391408801, + 0.002846133429557085, + 0.0050157783553004265, + 0.005774755030870438, + 0.003449993673712015, + 0.004540431313216686, + 0.005839139223098755, + 0.006245238706469536, + 0.005004140082746744, + 0.004954823292791843, + 0.0034469657111912966, + 0.004601669032126665, + 0.002094859955832362, + 0.0028993680607527494, + 0.006485958583652973, + 0.004991788882762194, + 0.00456966832280159, + 0.0020707566291093826, + 0.003598039271309972, + 0.003338966052979231, + 0.0026717332657426596, + 0.002995243528857827, + 0.008731012232601643, + 0.0044982703402638435, + 0.005194688681513071, + 0.0020985272713005543, + 0.010795965790748596, + 0.010051772929728031, + 0.006960729602724314, + 0.006996861193329096, + 0.027899887412786484, + 0.019839251413941383, + 0.008067164570093155, + 0.005856988951563835, + 0.020470043644309044, + 0.03376707807183266, + 0.0156398992985487, + 0.015568623319268227, + 0.03911586478352547, + 0.0359986238181591, + 0.02110464870929718, + 0.010322493501007557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.28136733174324036 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021999699529260397 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.0021999699529260397, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021999699529260397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17271177470684052, + 0.05067992955446243, + 0.12082859873771667, + 0.07027681916952133, + 0.05853298306465149, + 0.066954605281353, + 0.085467129945755, + 0.02370251715183258, + 0.17730003595352173, + 0.17354562878608704 + ], + "credit_map": [ + 0.015995647758245468, + 0.02034047245979309, + 0.02660369500517845, + 0.022582881152629852, + 0.03260992839932442, + 0.014440525323152542, + 0.01852366514503956, + 0.021614933386445045, + 0.005367577075958252, + 0.0073339613154530525, + 0.003315424546599388, + 0.0046002063900232315, + 0.012493965215981007, + 0.004697601776570082, + 0.007222088519483805, + 0.005649100057780743, + 0.017677921801805496, + 0.025014478713274002, + 0.00772639038041234, + 0.013055911287665367, + 0.014099128544330597, + 0.009964332915842533, + 0.011621092446148396, + 0.021669326350092888, + 0.009517362341284752, + 0.011385091580450535, + 0.00785071775317192, + 0.006449589971452951, + 0.013478673994541168, + 0.007232983596622944, + 0.007927539758384228, + 0.0064348564483225346, + 0.0056657628156244755, + 0.009256260469555855, + 0.006705719977617264, + 0.006655006669461727, + 0.00843000691384077, + 0.00513469660654664, + 0.0052870833314955235, + 0.011398443020880222, + 0.006583519279956818, + 0.007037610746920109, + 0.007173131685703993, + 0.005491363815963268, + 0.017341671511530876, + 0.009748350828886032, + 0.007722152397036552, + 0.0058567943051457405, + 0.009426696226000786, + 0.014171400107443333, + 0.004942920990288258, + 0.005342502612620592, + 0.022763865068554878, + 0.016105996444821358, + 0.008380196988582611, + 0.004333543125540018, + 0.0028551907744258642, + 0.002744492143392563, + 0.0017856486374512315, + 0.0023708029184490442, + 0.005359988193958998, + 0.0035324785858392715, + 0.003873730544000864, + 0.0011801844229921699, + 0.020474368706345558, + 0.02106643095612526, + 0.009968849830329418, + 0.010579890571534634, + 0.054138973355293274, + 0.03742198273539543, + 0.015222134999930859, + 0.008427374996244907, + 0.017643719911575317, + 0.030676409602165222, + 0.012722458690404892, + 0.015789823606610298, + 0.037626683712005615, + 0.03096308931708336, + 0.02029295265674591, + 0.007830468006432056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0021999699529260397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793692752718925 + ], + "opposite_logprob_tokens": [ + -3.5297937393188477 + ], + "logprob_reward": -0.029793692752718925, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793692752718925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2467338591814041, + 0.08186191320419312, + 0.07563147693872452, + 0.10973836481571198, + 0.09261377900838852, + 0.15549048781394958, + 0.04981967806816101, + 0.09232966601848602, + 0.05694179981946945, + 0.038838960230350494 + ], + "credit_map": [ + 0.02073640190064907, + 0.028531236574053764, + 0.03699270635843277, + 0.03360488638281822, + 0.04700550436973572, + 0.024587616324424744, + 0.02356487512588501, + 0.031710635870695114, + 0.015153948217630386, + 0.015707504004240036, + 0.0051469579339027405, + 0.007546869106590748, + 0.017580296844244003, + 0.004983443766832352, + 0.008889956399798393, + 0.00685293972492218, + 0.011240449734032154, + 0.010808574967086315, + 0.005650022998452187, + 0.0070356219075620174, + 0.015012129209935665, + 0.005019723903387785, + 0.009486912749707699, + 0.011378050781786442, + 0.014396466314792633, + 0.01880384050309658, + 0.009926814585924149, + 0.008989032357931137, + 0.022962303832173347, + 0.01458113081753254, + 0.012377483770251274, + 0.007701304275542498, + 0.009403582662343979, + 0.013394198380410671, + 0.007240593433380127, + 0.009950562380254269, + 0.016128158196806908, + 0.009339326061308384, + 0.00838819332420826, + 0.018769165500998497, + 0.018404066562652588, + 0.01862013339996338, + 0.011256324127316475, + 0.009033992886543274, + 0.04497620835900307, + 0.027102138847112656, + 0.01699243113398552, + 0.009105201810598373, + 0.004901987966150045, + 0.005708618555217981, + 0.0032720351591706276, + 0.003675904357805848, + 0.015339147299528122, + 0.007565656211227179, + 0.006710990332067013, + 0.0026453386526554823, + 0.008913051337003708, + 0.011005326174199581, + 0.005429323296993971, + 0.006473771296441555, + 0.01952960342168808, + 0.01222602091729641, + 0.024431327357888222, + 0.004321248270571232, + 0.00919269397854805, + 0.011575207114219666, + 0.0065474859438836575, + 0.0049522449262440205, + 0.009548533707857132, + 0.007916392758488655, + 0.0040505011565983295, + 0.0031587425619363785, + 0.00709930807352066, + 0.005314959213137627, + 0.004837193060666323, + 0.0035759638994932175, + 0.005187924951314926, + 0.006214832421392202, + 0.0030928300693631172, + 0.003515950869768858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.029793692752718925 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14276696741580963 + ], + "opposite_logprob_tokens": [ + -2.0177669525146484 + ], + "logprob_reward": -0.14276696741580963, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14276696741580963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.230809286236763, + 0.03942442312836647, + 0.21226471662521362, + 0.0429045744240284, + 0.11699889600276947, + 0.03779681399464607, + 0.033579789102077484, + 0.08731810003519058, + 0.10564830899238586, + 0.09325504302978516 + ], + "credit_map": [ + 0.020959392189979553, + 0.023214565590023994, + 0.03816530853509903, + 0.03238370269536972, + 0.040179040282964706, + 0.017016857862472534, + 0.027496356517076492, + 0.03139406070113182, + 0.005513757932931185, + 0.006091604940593243, + 0.0039734747260808945, + 0.004442618694156408, + 0.006981929764151573, + 0.004114281386137009, + 0.004534619860351086, + 0.0037721353583037853, + 0.04156147688627243, + 0.058959826827049255, + 0.006770308129489422, + 0.02132391557097435, + 0.012715383432805538, + 0.01100431103259325, + 0.013866386376321316, + 0.046063102781772614, + 0.005451520439237356, + 0.0068417200818657875, + 0.0038850123528391123, + 0.005039106588810682, + 0.00803196057677269, + 0.004806514363735914, + 0.004812097176909447, + 0.0040366388857364655, + 0.01860293187201023, + 0.024964531883597374, + 0.011763248592615128, + 0.012254550121724606, + 0.010371290147304535, + 0.012719584628939629, + 0.00823348481208086, + 0.01808926835656166, + 0.005087044555693865, + 0.00470426632091403, + 0.006535715889185667, + 0.0038323188200592995, + 0.0056182933039963245, + 0.005071159452199936, + 0.0036618714220821857, + 0.0032861419022083282, + 0.005899807438254356, + 0.005780595820397139, + 0.0028139320202171803, + 0.002541391644626856, + 0.005947155877947807, + 0.004753347486257553, + 0.0033864201977849007, + 0.0024571367539465427, + 0.015260295942425728, + 0.01461762748658657, + 0.006761530879884958, + 0.00756973447278142, + 0.017268439754843712, + 0.012844860553741455, + 0.008531102910637856, + 0.004464508034288883, + 0.01669257879257202, + 0.018361564725637436, + 0.01325974240899086, + 0.006455538794398308, + 0.018257971853017807, + 0.01703263260424137, + 0.007690445985645056, + 0.007897831499576569, + 0.018252111971378326, + 0.012980274856090546, + 0.010544446296989918, + 0.010722845792770386, + 0.013232764787971973, + 0.013464816845953465, + 0.007909436710178852, + 0.006148342974483967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.14276696741580963 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This dynamic, eye-level medium-long shot, captured by a sleek onboard camera, perfectly parallels a young woman with striking biracial features carving a vast, pristine snow slope. Her warm olive skin contrasts starkly against the bright white surroundings, her focus absolute, her dark almond-shaped eyes visible behind large, iridescent goggles reflecting the deep blue of the sky. A long, dark braid tucks down her back from beneath a matte black helmet. She wears a vibrant teal high-performance snowboard jacket and stylish black snow pants, adding a pop of color against the monochrome landscape. With a low, athletic stance and powerful precision, she fluidly shifts her weight from heel edge to toe edge, her body a study in controlled grace. As the edge of her ski slices into the untouched powder, it sends a spectacular plume of fine, crystalline snow billowing into the fresh mountain air, gleaming intensely in the bright, high-altitude sunlight. The scene is bathed in the harsh yet beautiful light of a clear day in the mountains, creating sharp and dramatic shadows on the surface of the snow and an exhilarating atmosphere of speed, freedom and pure, unadulterated joy.", + "seed": 170, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_15/index_4134_seed_170_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1423427059315145, + "vlm_logprob": -0.1423427059315145, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.1423427059315145, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2325405478477478 + ], + "opposite_logprob_tokens": [ + -2.2325406074523926 + ], + "logprob_reward": -0.2325405478477478, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.2325405478477478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1628410369157791, + 0.1144665777683258, + 0.07115191966295242, + 0.13951082527637482, + 0.08213598281145096, + 0.10428699851036072, + 0.09331829100847244, + 0.09044075012207031, + 0.10058767348527908, + 0.04125984013080597 + ], + "credit_map": [ + 0.025431374087929726, + 0.02031020075082779, + 0.024402176961302757, + 0.025956079363822937, + 0.012657529674470425, + 0.019077831879258156, + 0.015185357071459293, + 0.019820477813482285, + 0.011820655316114426, + 0.017679555341601372, + 0.013732247054576874, + 0.010629835538566113, + 0.02588808722794056, + 0.01135912723839283, + 0.010919135063886642, + 0.012437940575182438, + 0.010104548186063766, + 0.008864887058734894, + 0.00859038531780243, + 0.007463438902050257, + 0.01033039204776287, + 0.00984430406242609, + 0.008711070753633976, + 0.007242897059768438, + 0.026327483355998993, + 0.03114980459213257, + 0.01305992342531681, + 0.011681336909532547, + 0.017639663070440292, + 0.011996710672974586, + 0.014566125348210335, + 0.013089788146317005, + 0.01915065012872219, + 0.013552531599998474, + 0.007595578208565712, + 0.006266060285270214, + 0.013253491371870041, + 0.006401804741472006, + 0.009699766524136066, + 0.00621610414236784, + 0.009898430667817593, + 0.013560108840465546, + 0.008544797077775002, + 0.010919782333076, + 0.01642223447561264, + 0.019863311201334, + 0.015559091232717037, + 0.009519246406853199, + 0.011865375563502312, + 0.012477303855121136, + 0.012172981165349483, + 0.009389955550432205, + 0.018289705738425255, + 0.008525197394192219, + 0.011619010008871555, + 0.008978762663900852, + 0.010601620189845562, + 0.01497400552034378, + 0.008439584635198116, + 0.005170599091798067, + 0.028964480385184288, + 0.009068408980965614, + 0.00786673929542303, + 0.005355309695005417, + 0.01304594986140728, + 0.011977000162005424, + 0.0204760879278183, + 0.008161216974258423, + 0.017729492858052254, + 0.011931114830076694, + 0.007944653742015362, + 0.009322158992290497, + 0.005206162575632334, + 0.004850311204791069, + 0.005865364335477352, + 0.004677736200392246, + 0.005178547464311123, + 0.006893632933497429, + 0.003923255950212479, + 0.00466482900083065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.2325405478477478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521604001522064 + ], + "opposite_logprob_tokens": [ + -1.5021604299545288 + ], + "logprob_reward": -0.2521604001522064, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521604001522064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.140042245388031, + 0.1171928346157074, + 0.06362862139940262, + 0.14972126483917236, + 0.08991517871618271, + 0.04828287661075592, + 0.01988835446536541, + 0.29266557097435, + 0.03761810064315796, + 0.041044920682907104 + ], + "credit_map": [ + 0.014673915691673756, + 0.018031533807516098, + 0.02062186598777771, + 0.02123342826962471, + 0.018185894936323166, + 0.017887640744447708, + 0.012092942371964455, + 0.01731501892209053, + 0.011122840456664562, + 0.01778259128332138, + 0.013200235553085804, + 0.010690049268305302, + 0.027251699939370155, + 0.010844669304788113, + 0.012906068935990334, + 0.013394688256084919, + 0.00825678464025259, + 0.009001658298075199, + 0.008524075150489807, + 0.003993375226855278, + 0.012563252821564674, + 0.0069147441536188126, + 0.008812813088297844, + 0.005561919417232275, + 0.016843318939208984, + 0.0175517275929451, + 0.013930502347648144, + 0.016648897901177406, + 0.02712542749941349, + 0.015775438398122787, + 0.023966919630765915, + 0.017879022285342216, + 0.017751656472682953, + 0.013069197535514832, + 0.008995951153337955, + 0.006698030978441238, + 0.015409816987812519, + 0.00680265948176384, + 0.012673929333686829, + 0.008513938635587692, + 0.005435536615550518, + 0.00622295867651701, + 0.005375251639634371, + 0.005209955386817455, + 0.00924836378544569, + 0.0058142622001469135, + 0.005763309076428413, + 0.005213238298892975, + 0.002782159950584173, + 0.0028148924466222525, + 0.002270790748298168, + 0.0019004442729055882, + 0.0035488547291606665, + 0.0019295514794066548, + 0.0030451950151473284, + 0.001596464542672038, + 0.03519674763083458, + 0.051644641906023026, + 0.02661765366792679, + 0.015589476563036442, + 0.09744277596473694, + 0.02925638109445572, + 0.02121768519282341, + 0.015700196847319603, + 0.0050116335041821, + 0.004704525228589773, + 0.006516698282212019, + 0.0035047768615186214, + 0.0058544655330479145, + 0.0048014516942203045, + 0.003271938068792224, + 0.003952608909457922, + 0.005311322398483753, + 0.0047101168893277645, + 0.007284839171916246, + 0.006177651230245829, + 0.004051127005368471, + 0.0033096750266849995, + 0.004155084956437349, + 0.006045101676136255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.2521604001522064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200921531766653 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.002200921531766653, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200921531766653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1745332032442093, + 0.0931403711438179, + 0.08519069105386734, + 0.1469840109348297, + 0.07381454110145569, + 0.05212996527552605, + 0.029603205621242523, + 0.2697000801563263, + 0.033712368458509445, + 0.04119153693318367 + ], + "credit_map": [ + 0.023555133491754532, + 0.024012655019760132, + 0.02668423391878605, + 0.02439666911959648, + 0.01457556989043951, + 0.021334046497941017, + 0.015542814508080482, + 0.024432094767689705, + 0.009763138368725777, + 0.013573221862316132, + 0.012012898921966553, + 0.008743835613131523, + 0.016827790066599846, + 0.010332551784813404, + 0.011040517129004002, + 0.010846422053873539, + 0.011772336438298225, + 0.012139777652919292, + 0.009900104254484177, + 0.005698144435882568, + 0.018038207665085793, + 0.009809658862650394, + 0.010801248252391815, + 0.007031215820461512, + 0.019612981006503105, + 0.0228924211114645, + 0.017121320590376854, + 0.013595093972980976, + 0.03232557699084282, + 0.013816386461257935, + 0.013178392313420773, + 0.014441845007240772, + 0.012987404130399227, + 0.01150311529636383, + 0.007156298961490393, + 0.005433755926787853, + 0.014278926886618137, + 0.007452747784554958, + 0.008713479153811932, + 0.006288811564445496, + 0.004902963060885668, + 0.0067778038792312145, + 0.005375915206968784, + 0.005573015660047531, + 0.010769106447696686, + 0.006679513957351446, + 0.006618316285312176, + 0.005433334037661552, + 0.00402598362416029, + 0.0037372782826423645, + 0.00464656250551343, + 0.0028980509378015995, + 0.0040816981345415115, + 0.004215161316096783, + 0.003170010633766651, + 0.0028284615837037563, + 0.03203074261546135, + 0.047330595552921295, + 0.025455530732870102, + 0.014019926078617573, + 0.0903945192694664, + 0.026426056399941444, + 0.01967630535364151, + 0.014366403222084045, + 0.006324132438749075, + 0.005294278729707003, + 0.005228513386100531, + 0.0024594541173428297, + 0.003616477595642209, + 0.0034862016327679157, + 0.0031944196671247482, + 0.004108894616365433, + 0.006063830107450485, + 0.004999889992177486, + 0.00567619688808918, + 0.0048108696937561035, + 0.006717542186379433, + 0.0040967632085084915, + 0.0042227813974022865, + 0.004603669513016939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.002200921531766653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327992208302021 + ], + "opposite_logprob_tokens": [ + -3.773279905319214 + ], + "logprob_reward": -0.02327992208302021, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327992208302021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1907176822423935, + 0.13681571185588837, + 0.1020302101969719, + 0.06225264444947243, + 0.06675444543361664, + 0.06436650454998016, + 0.046591904014348984, + 0.05958656594157219, + 0.11437623202800751, + 0.15650810301303864 + ], + "credit_map": [ + 0.02993316575884819, + 0.026730027049779892, + 0.02973511815071106, + 0.022066717967391014, + 0.022287337109446526, + 0.01840941235423088, + 0.016487209126353264, + 0.025068702176213264, + 0.015460983850061893, + 0.016482019796967506, + 0.018457286059856415, + 0.011262863874435425, + 0.022422976791858673, + 0.018206363543868065, + 0.01854710280895233, + 0.01597611792385578, + 0.01643659919500351, + 0.014707561582326889, + 0.014878208748996258, + 0.009403417818248272, + 0.017291447147727013, + 0.00937463715672493, + 0.0135654266923666, + 0.006372906733304262, + 0.00952285248786211, + 0.010874290950596333, + 0.005422065034508705, + 0.0045790900476276875, + 0.011082639917731285, + 0.006229816470295191, + 0.008938047103583813, + 0.005603842902928591, + 0.015389950945973396, + 0.01033176202327013, + 0.008007900789380074, + 0.004037389066070318, + 0.012234577909111977, + 0.0062162503600120544, + 0.005342873744666576, + 0.005193737801164389, + 0.0072998590767383575, + 0.009110610000789165, + 0.005127224139869213, + 0.007501160725951195, + 0.013475657440721989, + 0.006369178183376789, + 0.009477897547185421, + 0.006004914175719023, + 0.005881724413484335, + 0.006458731833845377, + 0.00649300729855895, + 0.00416185799986124, + 0.009338976815342903, + 0.004651992116123438, + 0.005505224224179983, + 0.004100385122001171, + 0.008345795795321465, + 0.008709942921996117, + 0.007329245563596487, + 0.004414853174239397, + 0.012871580198407173, + 0.006575382314622402, + 0.005998606793582439, + 0.0053411563858389854, + 0.019646674394607544, + 0.019084837287664413, + 0.01442463044077158, + 0.0073274411261081696, + 0.0189139973372221, + 0.013799767941236496, + 0.009482498280704021, + 0.01169638242572546, + 0.023231344297528267, + 0.022011496126651764, + 0.014573323540389538, + 0.0124730309471488, + 0.019711961969733238, + 0.025886625051498413, + 0.014935287646949291, + 0.023685041815042496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.02327992208302021 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from an eye-level medium-long perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young woman with biracial features wearing a matte black helmet and large iridescent goggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a vibrant teal snowboard jacket and black snow pants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman carving a snow slope while shifting her weight from heel edge to toe edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene show a plume of fine crystalline snow billowing from the ski edge into the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153173804283142 + ], + "opposite_logprob_tokens": [ + -1.7015317678451538 + ], + "logprob_reward": -0.20153173804283142, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153173804283142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2617914378643036, + 0.13806182146072388, + 0.04961966723203659, + 0.05573165789246559, + 0.07596033066511154, + 0.10217374563217163, + 0.13045009970664978, + 0.03283831104636192, + 0.07700039446353912, + 0.07637257128953934 + ], + "credit_map": [ + 0.04773232713341713, + 0.04001271352171898, + 0.04440813511610031, + 0.03083747625350952, + 0.024124588817358017, + 0.024662083014845848, + 0.023440634831786156, + 0.026573503389954567, + 0.02104814350605011, + 0.020588573068380356, + 0.025290463119745255, + 0.013113802298903465, + 0.015368792228400707, + 0.01613818295300007, + 0.012720800004899502, + 0.013793068006634712, + 0.006318297702819109, + 0.007478039711713791, + 0.006716287229210138, + 0.003865310689434409, + 0.008240173570811749, + 0.005092337261885405, + 0.007495034486055374, + 0.004414186347275972, + 0.008652232587337494, + 0.009629768319427967, + 0.0050016106106340885, + 0.006544320844113827, + 0.007329663261771202, + 0.004141026642173529, + 0.007923668250441551, + 0.006509366445243359, + 0.014450496062636375, + 0.01234650332480669, + 0.008057170547544956, + 0.005499379243701696, + 0.012813683599233627, + 0.008174791932106018, + 0.008241372182965279, + 0.006376932375133038, + 0.012813480570912361, + 0.012582879513502121, + 0.011954334564507008, + 0.011363882571458817, + 0.01781761832535267, + 0.013362064957618713, + 0.011988294310867786, + 0.010291188955307007, + 0.02654825709760189, + 0.017776869237422943, + 0.023851139470934868, + 0.013389751315116882, + 0.01269590388983488, + 0.012909558601677418, + 0.010975520126521587, + 0.012303103692829609, + 0.006087217014282942, + 0.004790615290403366, + 0.005773097276687622, + 0.002687737112864852, + 0.005446532741189003, + 0.00282747158780694, + 0.0026628454215824604, + 0.002562795765697956, + 0.009345644153654575, + 0.009342867881059647, + 0.011350240558385849, + 0.005879840813577175, + 0.01596781425178051, + 0.010712677612900734, + 0.00547073595225811, + 0.008930581621825695, + 0.009342549368739128, + 0.008301923982799053, + 0.011599179357290268, + 0.006001475732773542, + 0.009741422720253468, + 0.010160334408283234, + 0.006997394375503063, + 0.014228284358978271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.20153173804283142 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_2/index_tensor([2758, 2758])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_2/index_tensor([2758, 2758])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b011a1ce46890a1f5a2213a869b4085d5a0424dd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_2/index_tensor([2758, 2758])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up portrait focuses with intimate intensity on a distinctively stylish man, his features rendered in sharp detail. He wears a vibrant brown wool felt hat, its chiseled brim casting a soft shadow across his forehead. His face, accented by a carefully groomed mustache and goatee, is stern. His attire is simple yet deliberate: a matte black cotton shirt that absorbs light, providing a stark canvas for a large silver curb chain necklace clinging to his collarbone. The background melts into a soft-focus bokeh of blurred interior space, dominated by a cool, cinematic blue light spilling from an off-camera source, depicting a sharp, electric rim light along his chin and shoulders. This contrasts sharply with the warm, soft key light that models the planes of his face. His dark eyes stare intently at a point just beyond the frame, and his entire being seems to pause as he listens to something beyond the frame or wrestles with deep inner thoughts, creating an atmosphere of quiet, contemplative, and powerful concentration.", + "seed": 20884, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_2/index_2758_seed_20884_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07326195486821234, + "vlm_logprob": -0.07326195486821234, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.07326195486821234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018525319173932076 + ], + "opposite_logprob_tokens": [ + -5.143525123596191 + ], + "logprob_reward": -0.018525319173932076, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.018525319173932076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15656037628650665, + 0.07314343005418777, + 0.0668674185872078, + 0.06431244313716888, + 0.29018551111221313, + 0.047777581959962845, + 0.05529426410794258, + 0.11730925738811493, + 0.06113759055733681, + 0.0674121156334877 + ], + "credit_map": [ + 0.013913173228502274, + 0.016225123777985573, + 0.02599969133734703, + 0.03523916378617287, + 0.01693880558013916, + 0.01145236287266016, + 0.019710542634129524, + 0.0170814860612154, + 0.0067114210687577724, + 0.004091982264071703, + 0.014819574542343616, + 0.004360003396868706, + 0.006408467423170805, + 0.010037792846560478, + 0.006538501475006342, + 0.020175686106085777, + 0.007838922552764416, + 0.007672312669456005, + 0.011786561459302902, + 0.005447093863040209, + 0.008877494372427464, + 0.006723559461534023, + 0.006492140702903271, + 0.012029323726892471, + 0.006514557637274265, + 0.004682985600084066, + 0.010876741260290146, + 0.007886103354394436, + 0.00817795842885971, + 0.005009193904697895, + 0.006290718447417021, + 0.014874180778861046, + 0.02448379434645176, + 0.012193430215120316, + 0.09313270449638367, + 0.013654887676239014, + 0.010302111506462097, + 0.022867970168590546, + 0.016972295939922333, + 0.09657831490039825, + 0.004594700876623392, + 0.0022761474829167128, + 0.013189863413572311, + 0.0028737043030560017, + 0.003558105556294322, + 0.004326824564486742, + 0.003372650593519211, + 0.013585580512881279, + 0.006251619663089514, + 0.0032444444950670004, + 0.016146982088685036, + 0.00321357324719429, + 0.00435583246871829, + 0.003553737886250019, + 0.0056263417936861515, + 0.012901724316179752, + 0.029852285981178284, + 0.006132620386779308, + 0.022065503522753716, + 0.009942690841853619, + 0.007121989503502846, + 0.00859496183693409, + 0.008558879606425762, + 0.025040319189429283, + 0.009617757983505726, + 0.0033721786458045244, + 0.009565530344843864, + 0.004771874751895666, + 0.00459074042737484, + 0.005661761853843927, + 0.00652628717944026, + 0.01703144982457161, + 0.01400609128177166, + 0.003623449243605137, + 0.015745216980576515, + 0.005468349438160658, + 0.0038128928281366825, + 0.006553126964718103, + 0.004098216537386179, + 0.014104769565165043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.018525319173932076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16040192544460297 + ], + "opposite_logprob_tokens": [ + -1.9104019403457642 + ], + "logprob_reward": -0.16040192544460297, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16040192544460297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17545413970947266, + 0.15706133842468262, + 0.1321960836648941, + 0.04327024891972542, + 0.0734078660607338, + 0.04863126948475838, + 0.054510150104761124, + 0.1557772159576416, + 0.028367040678858757, + 0.1313246339559555 + ], + "credit_map": [ + 0.018113218247890472, + 0.024518193677067757, + 0.026569046080112457, + 0.014917946420609951, + 0.024767465889453888, + 0.0196710005402565, + 0.017232583835721016, + 0.029664700850844383, + 0.014155592769384384, + 0.014731718227267265, + 0.03640533611178398, + 0.008948003873229027, + 0.017358049750328064, + 0.018880555406212807, + 0.014649312011897564, + 0.0319327637553215, + 0.014173664152622223, + 0.014727959409356117, + 0.02271616831421852, + 0.014643221162259579, + 0.010392528027296066, + 0.015691671520471573, + 0.008495182730257511, + 0.031355686485767365, + 0.004248020704835653, + 0.004032494034618139, + 0.007018186617642641, + 0.0059840865433216095, + 0.0036329778376966715, + 0.0037192164454609156, + 0.003377241548150778, + 0.011258026584982872, + 0.007704542018473148, + 0.007760079111903906, + 0.011850506067276001, + 0.011500086635351181, + 0.0066661410965025425, + 0.006049265619367361, + 0.009997716173529625, + 0.011879531666636467, + 0.005242689047008753, + 0.004279358312487602, + 0.006633896846324205, + 0.008268316276371479, + 0.005391385406255722, + 0.00443112151697278, + 0.00539065757766366, + 0.008993848226964474, + 0.005890045780688524, + 0.0037321383133530617, + 0.015207997523248196, + 0.004347021691501141, + 0.003810030873864889, + 0.004561632871627808, + 0.005092849489301443, + 0.011868435889482498, + 0.029763398692011833, + 0.011016866192221642, + 0.03756740689277649, + 0.021660203114151955, + 0.013391595333814621, + 0.011384814977645874, + 0.014104221947491169, + 0.01688871718943119, + 0.005090313497930765, + 0.0017408613348379731, + 0.007076298352330923, + 0.0026690433733165264, + 0.0037101467605680227, + 0.0018267801497131586, + 0.0034402806777507067, + 0.002813318045809865, + 0.03191819041967392, + 0.006329947151243687, + 0.04008619114756584, + 0.005215680226683617, + 0.006158359814435244, + 0.01166630256921053, + 0.006143357139080763, + 0.023806612938642502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.16040192544460297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991763353347778 + ], + "opposite_logprob_tokens": [ + -1.804917573928833 + ], + "logprob_reward": -0.17991763353347778, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991763353347778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17090585827827454, + 0.11311770230531693, + 0.0989457294344902, + 0.03153074532747269, + 0.032215822488069534, + 0.09284575283527374, + 0.056665413081645966, + 0.37195295095443726, + 0.014693533070385456, + 0.017126452177762985 + ], + "credit_map": [ + 0.015501808375120163, + 0.015447915531694889, + 0.0302178505808115, + 0.016194118186831474, + 0.018953975290060043, + 0.020273609086871147, + 0.014659402891993523, + 0.03965718671679497, + 0.009966464713215828, + 0.00899400282651186, + 0.024710137397050858, + 0.007326940074563026, + 0.014199771918356419, + 0.013043052516877651, + 0.011887535452842712, + 0.022989794611930847, + 0.011175177060067654, + 0.007606810424476862, + 0.014506036415696144, + 0.013699730858206749, + 0.011831182986497879, + 0.007103954441845417, + 0.011256322264671326, + 0.02176651917397976, + 0.0023661060258746147, + 0.002855077851563692, + 0.005690998863428831, + 0.0023674138356000185, + 0.0030471982900053263, + 0.003843967104330659, + 0.0019508992554619908, + 0.009409083053469658, + 0.004676934331655502, + 0.0029184904415160418, + 0.004645167849957943, + 0.002965897088870406, + 0.0032508899457752705, + 0.0032266161870211363, + 0.002916261088103056, + 0.007615565788000822, + 0.01589873619377613, + 0.004987383261322975, + 0.03171441704034805, + 0.006962109822779894, + 0.0066896844655275345, + 0.006756584160029888, + 0.005525514483451843, + 0.014311322011053562, + 0.004495526198297739, + 0.002622013445943594, + 0.016568608582019806, + 0.0039223008789122105, + 0.0036324658431112766, + 0.004606707952916622, + 0.003782168962061405, + 0.017035622149705887, + 0.08064558357000351, + 0.01663350872695446, + 0.05387737601995468, + 0.015500964596867561, + 0.02033916860818863, + 0.03440552204847336, + 0.022499755024909973, + 0.1280510574579239, + 0.001309254439547658, + 0.0011559377890080214, + 0.002250714460387826, + 0.0015332818729802966, + 0.002544583287090063, + 0.0013359462609514594, + 0.0019004250643774867, + 0.002663389081135392, + 0.002766774268820882, + 0.0009540412575006485, + 0.00264549907296896, + 0.0014428481226786971, + 0.00151739118155092, + 0.0018573913257569075, + 0.0012476789997890592, + 0.004694827366620302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.17991763353347778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637913778424263 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637913778424263, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637913778424263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20020614564418793, + 0.354190856218338, + 0.09578680247068405, + 0.029592406004667282, + 0.15262949466705322, + 0.04149283841252327, + 0.022850964218378067, + 0.049879446625709534, + 0.025076916441321373, + 0.028294134885072708 + ], + "credit_map": [ + 0.019076578319072723, + 0.018755249679088593, + 0.028840672224760056, + 0.03723285347223282, + 0.028094958513975143, + 0.017087724059820175, + 0.02505956031382084, + 0.026058539748191833, + 0.04639207571744919, + 0.026573745533823967, + 0.07085059583187103, + 0.025823935866355896, + 0.04202855005860329, + 0.038216732442379, + 0.035463590174913406, + 0.06884165853261948, + 0.010615730658173561, + 0.009270887821912766, + 0.017084477469325066, + 0.011118847876787186, + 0.010329884476959705, + 0.010443292558193207, + 0.008267291821539402, + 0.01865638606250286, + 0.0029380787163972855, + 0.0024144104681909084, + 0.004960223566740751, + 0.0037468825466930866, + 0.0025358463171869516, + 0.002703601960092783, + 0.0027721712831407785, + 0.0075211916118860245, + 0.013585112057626247, + 0.0065448652021586895, + 0.05061229690909386, + 0.007031225599348545, + 0.005460795480757952, + 0.011572279967367649, + 0.0072671896778047085, + 0.05055573210120201, + 0.005805591121315956, + 0.0023094762582331896, + 0.012072554789483547, + 0.0034214414190500975, + 0.004830272868275642, + 0.0029568204190582037, + 0.003152620978653431, + 0.006944061256945133, + 0.0022820497397333384, + 0.002001130487769842, + 0.003339709248393774, + 0.003961486276239157, + 0.0025712877977639437, + 0.0020204803440719843, + 0.0033446375746279955, + 0.003330184845253825, + 0.008982065133750439, + 0.0034640401136130095, + 0.012861269526183605, + 0.00294734095223248, + 0.006357192061841488, + 0.0035898254718631506, + 0.00525812478736043, + 0.0064195869490504265, + 0.0037446559872478247, + 0.0017855264013633132, + 0.004694589413702488, + 0.001986517570912838, + 0.003658595494925976, + 0.0019792732782661915, + 0.0030545229092240334, + 0.0041732341051101685, + 0.0055465903133153915, + 0.0015606258530169725, + 0.007668477948755026, + 0.0018969657830893993, + 0.0026579301338642836, + 0.0022876134607940912, + 0.0022858306765556335, + 0.0043901000171899796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.004637913778424263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028269824106246233 + ], + "opposite_logprob_tokens": [ + -5.877827167510986 + ], + "logprob_reward": -0.0028269824106246233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028269824106246233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14749428629875183, + 0.08714768290519714, + 0.032797232270240784, + 0.042988210916519165, + 0.1488291174173355, + 0.11230769753456116, + 0.023163259029388428, + 0.18573923408985138, + 0.14711378514766693, + 0.07241949439048767 + ], + "credit_map": [ + 0.01457716803997755, + 0.012396630831062794, + 0.022662997245788574, + 0.015866754576563835, + 0.02226129174232483, + 0.01235251035541296, + 0.02013842575252056, + 0.02723851427435875, + 0.009411242790520191, + 0.007975218817591667, + 0.014943692833185196, + 0.006807622034102678, + 0.011142110452055931, + 0.009085984900593758, + 0.007588300388306379, + 0.02019350789487362, + 0.002974631730467081, + 0.003324205055832863, + 0.005029734224081039, + 0.005610357504338026, + 0.0044491891749203205, + 0.0024695461615920067, + 0.0028109883423894644, + 0.006128582172095776, + 0.0040479376912117004, + 0.002508650068193674, + 0.00865740142762661, + 0.0029094433411955833, + 0.0039535569958388805, + 0.004585435148328543, + 0.0029708389192819595, + 0.013354948721826077, + 0.01259130984544754, + 0.008058932609856129, + 0.04446086660027504, + 0.011212592013180256, + 0.006625006441026926, + 0.010468294844031334, + 0.011034339666366577, + 0.04437778517603874, + 0.01936151273548603, + 0.006325685419142246, + 0.03434756025671959, + 0.008783443830907345, + 0.010130851529538631, + 0.007640140131115913, + 0.007599079981446266, + 0.01811942644417286, + 0.0028252520132809877, + 0.0019288392504677176, + 0.005994787905365229, + 0.0024040527641773224, + 0.002832049271091819, + 0.0016089251730591059, + 0.0028584199026226997, + 0.0027109328657388687, + 0.017466818913817406, + 0.007102514151483774, + 0.055660393089056015, + 0.005682841874659061, + 0.009914705529808998, + 0.015629690140485764, + 0.009988321922719479, + 0.06429394334554672, + 0.03556143492460251, + 0.006415399257093668, + 0.04288559406995773, + 0.00805768370628357, + 0.012366781942546368, + 0.00947516318410635, + 0.008678401820361614, + 0.023673327639698982, + 0.012593425810337067, + 0.004193239379674196, + 0.019551048055291176, + 0.009862171486020088, + 0.0067759063094854355, + 0.005234002135694027, + 0.004844750743359327, + 0.00936494953930378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.0028269824106246233 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up portrait focuses with intimate intensity on a distinctively stylish man, his features rendered in sharp detail. He wears a vibrant brown wool felt hat, its chiseled brim casting a soft shadow across his forehead. His face, accented by a carefully groomed mustache and goatee, is stern. His attire is simple yet deliberate: a matte black cotton shirt that absorbs light, providing a stark canvas for a large silver curb chain necklace clinging to his collarbone. The background melts into a soft-focus bokeh of blurred interior space, dominated by a cool, cinematic blue light spilling from an off-camera source, depicting a sharp, electric rim light along his chin and shoulders. This contrasts sharply with the warm, soft key light that models the planes of his face. His dark eyes stare intently at a point just beyond the frame, and his entire being seems to pause as he listens to something beyond the frame or wrestles with deep inner thoughts, creating an atmosphere of quiet, contemplative, and powerful concentration.", + "seed": 20885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_2/index_2758_seed_20885_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12701549688354136, + "vlm_logprob": -0.12701549688354136, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.12701549688354136, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02678452804684639 + ], + "opposite_logprob_tokens": [ + -4.401784420013428 + ], + "logprob_reward": -0.02678452804684639, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02678452804684639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15825478732585907, + 0.0983075425028801, + 0.05429167300462723, + 0.04447071999311447, + 0.1165260598063469, + 0.06653928011655807, + 0.06767943501472473, + 0.2317328006029129, + 0.08297941088676453, + 0.0792182981967926 + ], + "credit_map": [ + 0.02180912159383297, + 0.02174549549818039, + 0.019575567916035652, + 0.020720090717077255, + 0.022141996771097183, + 0.02014741115272045, + 0.014856376685202122, + 0.017258714884519577, + 0.011843879707157612, + 0.014707275666296482, + 0.012447217479348183, + 0.01000443659722805, + 0.013873957097530365, + 0.01223713904619217, + 0.009773388504981995, + 0.013420240953564644, + 0.004854938946664333, + 0.009488082490861416, + 0.005100619047880173, + 0.008262782357633114, + 0.007888225838541985, + 0.006492453161627054, + 0.005958856083452702, + 0.006245709024369717, + 0.005205439403653145, + 0.006497184280306101, + 0.006288033910095692, + 0.004617943428456783, + 0.0059284367598593235, + 0.00576285133138299, + 0.005120355170220137, + 0.005050471983850002, + 0.015007443726062775, + 0.018101181834936142, + 0.023636331781744957, + 0.014921031892299652, + 0.014424830675125122, + 0.01248562429100275, + 0.00978025235235691, + 0.008169356733560562, + 0.008321622386574745, + 0.013090431690216064, + 0.009315231814980507, + 0.009005937725305557, + 0.006928192917257547, + 0.005438057240098715, + 0.009061045944690704, + 0.005378757603466511, + 0.010229025967419147, + 0.007946726866066456, + 0.015015034936368465, + 0.007782679982483387, + 0.008807712234556675, + 0.006767874117940664, + 0.005466809496283531, + 0.00566356536000967, + 0.03013266623020172, + 0.018855588510632515, + 0.05406801402568817, + 0.019474869593977928, + 0.02304922044277191, + 0.04674254357814789, + 0.016654545441269875, + 0.0227553378790617, + 0.008375165984034538, + 0.005780261475592852, + 0.011491199024021626, + 0.007182530593127012, + 0.007917453534901142, + 0.01801268383860588, + 0.007278303615748882, + 0.016941800713539124, + 0.012006444856524467, + 0.008195647038519382, + 0.011675090529024601, + 0.010767635889351368, + 0.0070451791398227215, + 0.008696830831468105, + 0.008388908579945564, + 0.012442562729120255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.02678452804684639 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34865716099739075 + ], + "opposite_logprob_tokens": [ + -1.2236571311950684 + ], + "logprob_reward": -0.34865716099739075, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34865716099739075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09226235747337341, + 0.0533609502017498, + 0.03123397007584572, + 0.024442655965685844, + 0.12104541808366776, + 0.08931995928287506, + 0.49776485562324524, + 0.03783892095088959, + 0.030168525874614716, + 0.022562401369214058 + ], + "credit_map": [ + 0.008854576386511326, + 0.015419344417750835, + 0.011448205448687077, + 0.014157082885503769, + 0.011215494014322758, + 0.012361007742583752, + 0.009245593100786209, + 0.009561053477227688, + 0.004836492706090212, + 0.008813357912003994, + 0.006771388463675976, + 0.005303631536662579, + 0.00798905361443758, + 0.00883686263114214, + 0.005117049440741539, + 0.005693115759640932, + 0.0028370716609060764, + 0.00426102802157402, + 0.004638706333935261, + 0.00300753116607666, + 0.003616190981119871, + 0.005471828859299421, + 0.0023188500199466944, + 0.005082760937511921, + 0.0026030754670500755, + 0.0045003206469118595, + 0.003085906384512782, + 0.0031791420187801123, + 0.0026482040993869305, + 0.0027653598226606846, + 0.002651918213814497, + 0.0030087288469076157, + 0.016043690964579582, + 0.012663823552429676, + 0.023471100255846977, + 0.010463135316967964, + 0.016940142959356308, + 0.017035363242030144, + 0.014751345850527287, + 0.009676823392510414, + 0.011961687356233597, + 0.013698389753699303, + 0.013020658865571022, + 0.006923020351678133, + 0.017056334763765335, + 0.009316813200712204, + 0.012167680077254772, + 0.005175373051315546, + 0.07522007077932358, + 0.01538606733083725, + 0.19320259988307953, + 0.01916794292628765, + 0.029421085491776466, + 0.11941304057836533, + 0.014666172675788403, + 0.03128783777356148, + 0.006064827553927898, + 0.003895021276548505, + 0.00796208344399929, + 0.003577995579689741, + 0.004614819772541523, + 0.0042378599755465984, + 0.004086424130946398, + 0.0033998899161815643, + 0.004309303127229214, + 0.0030777002684772015, + 0.0030731684528291225, + 0.004126375075429678, + 0.004834406077861786, + 0.0034008745569735765, + 0.0037261545658111572, + 0.0036205421201884747, + 0.002617646474391222, + 0.002891865326091647, + 0.0026029907166957855, + 0.003567431354895234, + 0.0031782665755599737, + 0.0022185544949024916, + 0.0032753711566329002, + 0.0022102752700448036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.34865716099739075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25216901302337646 + ], + "opposite_logprob_tokens": [ + -1.5021690130233765 + ], + "logprob_reward": -0.25216901302337646, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25216901302337646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25142839550971985, + 0.08821416646242142, + 0.06526549160480499, + 0.043029289692640305, + 0.05003802478313446, + 0.057438164949417114, + 0.16902609169483185, + 0.112993024289608, + 0.09925415366888046, + 0.06331319361925125 + ], + "credit_map": [ + 0.021596750244498253, + 0.04608010873198509, + 0.032922934740781784, + 0.028912832960486412, + 0.045586299151182175, + 0.028405003249645233, + 0.026429908350110054, + 0.021494561806321144, + 0.008266407996416092, + 0.017558187246322632, + 0.007313164882361889, + 0.013334902003407478, + 0.013048513792455196, + 0.012842792086303234, + 0.008864126168191433, + 0.006986075080931187, + 0.007397953886538744, + 0.009167379699647427, + 0.009627812542021275, + 0.008826367557048798, + 0.008305570110678673, + 0.008477188646793365, + 0.007815612480044365, + 0.005647601094096899, + 0.0049518742598593235, + 0.006737223360687494, + 0.007004168350249529, + 0.005535957403481007, + 0.005127942189574242, + 0.005220527295023203, + 0.00338049721904099, + 0.005071103107184172, + 0.008155337534844875, + 0.0071070874109864235, + 0.007435091305524111, + 0.006385632324963808, + 0.005107209552079439, + 0.004774419590830803, + 0.004801752511411905, + 0.006271494552493095, + 0.0068408427760005, + 0.007789472118020058, + 0.008399194106459618, + 0.008859180845320225, + 0.006127612199634314, + 0.005035720765590668, + 0.006976642645895481, + 0.007409494370222092, + 0.023411685600876808, + 0.007733085658401251, + 0.06090078130364418, + 0.008010827004909515, + 0.011184760369360447, + 0.038798000663518906, + 0.0070841251872479916, + 0.011902823112905025, + 0.015143531374633312, + 0.0198186244815588, + 0.013465329073369503, + 0.0196071844547987, + 0.0136083560064435, + 0.006628571543842554, + 0.015720145776867867, + 0.009001275524497032, + 0.012292809784412384, + 0.00839680153876543, + 0.011202503927052021, + 0.010442791506648064, + 0.017714133486151695, + 0.016464952379465103, + 0.012988772243261337, + 0.009751391597092152, + 0.009688896127045155, + 0.006848577409982681, + 0.009813356213271618, + 0.007961707189679146, + 0.006510251201689243, + 0.00862155482172966, + 0.0066329315304756165, + 0.0072359140031039715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.25216901302337646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820206806063652 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820206806063652, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820206806063652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2225920557975769, + 0.0874398797750473, + 0.09954981505870819, + 0.09259738773107529, + 0.09087260067462921, + 0.06030972674489021, + 0.032969918102025986, + 0.24133221805095673, + 0.04477633908390999, + 0.027560019865632057 + ], + "credit_map": [ + 0.03250512853264809, + 0.0272208321839571, + 0.0346352644264698, + 0.02386324480175972, + 0.022415565326809883, + 0.03609306737780571, + 0.016800493001937866, + 0.029058458283543587, + 0.009705595672130585, + 0.01063231285661459, + 0.01397416740655899, + 0.011315630748867989, + 0.009460113942623138, + 0.010056653060019016, + 0.008905868977308273, + 0.013389541767537594, + 0.01049770973622799, + 0.010740747675299644, + 0.015669412910938263, + 0.010051372461020947, + 0.010907063260674477, + 0.01677844673395157, + 0.009904910810291767, + 0.015000149607658386, + 0.009830346331000328, + 0.01535998098552227, + 0.014572889544069767, + 0.010999598540365696, + 0.012361893430352211, + 0.008997676894068718, + 0.00886614341288805, + 0.011608858592808247, + 0.012548653408885002, + 0.01131406519562006, + 0.015163129195570946, + 0.006519685965031385, + 0.016058064997196198, + 0.009250938892364502, + 0.01160701084882021, + 0.008411056362092495, + 0.006760939955711365, + 0.007233999669551849, + 0.008179796859622002, + 0.0074440171010792255, + 0.008098960854113102, + 0.006918094586580992, + 0.006937989965081215, + 0.008735926821827888, + 0.004837394692003727, + 0.0037865357007831335, + 0.0030025248415768147, + 0.0041603087447583675, + 0.003201788756996393, + 0.00546605046838522, + 0.0026347190141677856, + 0.005880594253540039, + 0.030560117214918137, + 0.01202754769474268, + 0.08752040565013885, + 0.01197725534439087, + 0.016066476702690125, + 0.05104397237300873, + 0.014172540046274662, + 0.017963897436857224, + 0.00428124750033021, + 0.003100147470831871, + 0.00437633041292429, + 0.0035848182160407305, + 0.007935908623039722, + 0.008361904881894588, + 0.005081796552985907, + 0.008054184727370739, + 0.0029444973915815353, + 0.0027715968899428844, + 0.00408848887309432, + 0.0030803517438471317, + 0.0027983037289232016, + 0.003637197893112898, + 0.0033286355901509523, + 0.004910948686301708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.002820206806063652 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646575544029474 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646575544029474, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646575544029474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12765276432037354, + 0.042150575667619705, + 0.03473810479044914, + 0.06239129602909088, + 0.05395951122045517, + 0.08947374671697617, + 0.3737964332103729, + 0.15935949981212616, + 0.029054289683699608, + 0.027423761785030365 + ], + "credit_map": [ + 0.020520219579339027, + 0.01571071892976761, + 0.016629626974463463, + 0.012917925603687763, + 0.01675521396100521, + 0.015652064234018326, + 0.014760035090148449, + 0.01470696646720171, + 0.004744902718812227, + 0.007606833707541227, + 0.00518073933199048, + 0.00537029979750514, + 0.00433158315718174, + 0.005566333420574665, + 0.004017396364361048, + 0.0053324890322983265, + 0.004832323640584946, + 0.004602111876010895, + 0.0039042390417307615, + 0.003951100632548332, + 0.004114865325391293, + 0.004652968142181635, + 0.004569332581013441, + 0.004111165180802345, + 0.00692649045959115, + 0.010838220827281475, + 0.007876143790781498, + 0.007657716516405344, + 0.008446719497442245, + 0.007769722491502762, + 0.006182749755680561, + 0.00669353362172842, + 0.007093026768416166, + 0.005386615637689829, + 0.014746918343007565, + 0.005937206093221903, + 0.005025837104767561, + 0.006575658917427063, + 0.004823604598641396, + 0.004370644222944975, + 0.011497804895043373, + 0.01453470066189766, + 0.013942639343440533, + 0.012893295846879482, + 0.01060356292873621, + 0.005935224704444408, + 0.013906708918511868, + 0.006159808952361345, + 0.0557468980550766, + 0.012378698214888573, + 0.14341293275356293, + 0.014732718467712402, + 0.023719478398561478, + 0.0876697301864624, + 0.012526285834610462, + 0.02360968105494976, + 0.018534021452069283, + 0.011876868084073067, + 0.042938049882650375, + 0.010002262890338898, + 0.015324873849749565, + 0.03560701757669449, + 0.010293232277035713, + 0.01478318590670824, + 0.0031364108435809612, + 0.0022690061014145613, + 0.005416537635028362, + 0.004188896622508764, + 0.00350859179161489, + 0.004443594254553318, + 0.0022635755594819784, + 0.003827678272500634, + 0.00276393024250865, + 0.002275867387652397, + 0.003076002234593034, + 0.0027181070763617754, + 0.0031972001306712627, + 0.005148150026798248, + 0.0034911504480987787, + 0.0047533526085317135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.004646575544029474 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up portrait focuses with intimate intensity on a distinctively stylish man, his features rendered in sharp detail. He wears a vibrant brown wool felt hat, its chiseled brim casting a soft shadow across his forehead. His face, accented by a carefully groomed mustache and goatee, is stern. His attire is simple yet deliberate: a matte black cotton shirt that absorbs light, providing a stark canvas for a large silver curb chain necklace clinging to his collarbone. The background melts into a soft-focus bokeh of blurred interior space, dominated by a cool, cinematic blue light spilling from an off-camera source, depicting a sharp, electric rim light along his chin and shoulders. This contrasts sharply with the warm, soft key light that models the planes of his face. His dark eyes stare intently at a point just beyond the frame, and his entire being seems to pause as he listens to something beyond the frame or wrestles with deep inner thoughts, creating an atmosphere of quiet, contemplative, and powerful concentration.", + "seed": 20886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_2/index_2758_seed_20886_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06158125521615147, + "vlm_logprob": -0.06158125521615147, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.06158125521615147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015088239684700966 + ], + "opposite_logprob_tokens": [ + -5.265088081359863 + ], + "logprob_reward": -0.015088239684700966, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.015088239684700966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16753894090652466, + 0.3635817766189575, + 0.021507536992430687, + 0.023652540519833565, + 0.04874427989125252, + 0.02255888842046261, + 0.03500683605670929, + 0.036020901054143906, + 0.21421310305595398, + 0.06717520952224731 + ], + "credit_map": [ + 0.014521781355142593, + 0.012302880175411701, + 0.02199448272585869, + 0.02599147893488407, + 0.015524237416684628, + 0.022165054455399513, + 0.01497445534914732, + 0.04006456956267357, + 0.04591323807835579, + 0.025319308042526245, + 0.05485496670007706, + 0.03718757629394531, + 0.022353405132889748, + 0.04308561980724335, + 0.01927081309258938, + 0.11559683829545975, + 0.002156333765015006, + 0.0019460712792351842, + 0.003468794049695134, + 0.0029010630678385496, + 0.002472484018653631, + 0.0024145597126334906, + 0.0019685947336256504, + 0.004179635550826788, + 0.0025437851436436176, + 0.002110809786245227, + 0.0025458845775574446, + 0.0028541437350213528, + 0.003214254043996334, + 0.0025903931818902493, + 0.0023850034922361374, + 0.005408266559243202, + 0.004743333440274, + 0.003847074694931507, + 0.012845111079514027, + 0.006643385626375675, + 0.004499903414398432, + 0.005174887832254171, + 0.003729473799467087, + 0.007261104416102171, + 0.00247121206484735, + 0.0013405911158770323, + 0.007046125363558531, + 0.0028951221611350775, + 0.0025155143812298775, + 0.002031215000897646, + 0.0016540107317268848, + 0.002605095971375704, + 0.003061536466702819, + 0.0030387549195438623, + 0.008832198567688465, + 0.004962217062711716, + 0.0033805021084845066, + 0.0038096497301012278, + 0.0023836896289139986, + 0.00553828664124012, + 0.002742786193266511, + 0.00388318975456059, + 0.006257169879972935, + 0.003260009689256549, + 0.00514182960614562, + 0.004177829250693321, + 0.003645183052867651, + 0.006912903860211372, + 0.01695256121456623, + 0.022156422957777977, + 0.038927529007196426, + 0.022430308163166046, + 0.02242058329284191, + 0.030834652483463287, + 0.019851146265864372, + 0.04063990339636803, + 0.006717926822602749, + 0.003221594961360097, + 0.01877114735543728, + 0.008704611100256443, + 0.003978276159614325, + 0.0063748834654688835, + 0.004204231314361095, + 0.015202544629573822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.015088239684700966 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16041217744350433 + ], + "opposite_logprob_tokens": [ + -1.9104121923446655 + ], + "logprob_reward": -0.16041217744350433, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16041217744350433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24603858590126038, + 0.30066269636154175, + 0.11169616132974625, + 0.0655791163444519, + 0.020815948024392128, + 0.022911736741662025, + 0.10899180173873901, + 0.015167892910540104, + 0.08212850987911224, + 0.026007477194070816 + ], + "credit_map": [ + 0.027406221255660057, + 0.01836734265089035, + 0.05137563869357109, + 0.027682693675160408, + 0.01878873072564602, + 0.027296505868434906, + 0.014944585971534252, + 0.06017688661813736, + 0.036824580281972885, + 0.020366549491882324, + 0.04706835001707077, + 0.030222088098526, + 0.018902108073234558, + 0.03617331385612488, + 0.01677868328988552, + 0.09432700276374817, + 0.010985000990331173, + 0.008774000220000744, + 0.027029985561966896, + 0.012974272482097149, + 0.010916252620518208, + 0.01172325387597084, + 0.009171132929623127, + 0.020122265443205833, + 0.00655903946608305, + 0.005670802667737007, + 0.01782354526221752, + 0.0077308667823672295, + 0.005638516508042812, + 0.006901554297655821, + 0.004592243582010269, + 0.01066255196928978, + 0.002046880777925253, + 0.0015708826249465346, + 0.005692719481885433, + 0.0026601646095514297, + 0.0020778998732566833, + 0.0020996942184865475, + 0.0019494992448017001, + 0.0027182092890143394, + 0.002417081268504262, + 0.0018368249293416739, + 0.0067926254123449326, + 0.002525817835703492, + 0.002083518775179982, + 0.0025338754057884216, + 0.001826534396968782, + 0.0028954592999070883, + 0.010072227567434311, + 0.008862758986651897, + 0.028360025957226753, + 0.01375161949545145, + 0.011148417368531227, + 0.01162303239107132, + 0.0072986711747944355, + 0.01787503808736801, + 0.0013124903198331594, + 0.0013785307528451085, + 0.0034891224931925535, + 0.0016703560249879956, + 0.002405810635536909, + 0.0014360086061060429, + 0.0014630341902375221, + 0.0020125408191233873, + 0.006801135838031769, + 0.008245894685387611, + 0.015480262227356434, + 0.008491330780088902, + 0.00828123465180397, + 0.011090034618973732, + 0.00730975391343236, + 0.016428859904408455, + 0.0023617385886609554, + 0.0024988960940390825, + 0.005612744949758053, + 0.0027513825334608555, + 0.0024866650346666574, + 0.003708867123350501, + 0.001716712722554803, + 0.004870470147579908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.16041217744350433 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708574533462524 + ], + "opposite_logprob_tokens": [ + -2.1270856857299805 + ], + "logprob_reward": -0.12708574533462524, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708574533462524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17701305449008942, + 0.3268316984176636, + 0.03369872272014618, + 0.17948952317237854, + 0.06121785566210747, + 0.02715529315173626, + 0.034753117710351944, + 0.01834459975361824, + 0.036594826728105545, + 0.1049012690782547 + ], + "credit_map": [ + 0.018959423527121544, + 0.012103806249797344, + 0.03517179191112518, + 0.03637421131134033, + 0.015933651477098465, + 0.018837934359908104, + 0.01037135161459446, + 0.02926088683307171, + 0.03577851131558418, + 0.024024005979299545, + 0.03602518513798714, + 0.037332963198423386, + 0.02387716807425022, + 0.049324534833431244, + 0.019051676616072655, + 0.1014176681637764, + 0.002895886776968837, + 0.0024921302683651447, + 0.0033787242136895657, + 0.004955783486366272, + 0.005696222651749849, + 0.004053068347275257, + 0.004293324425816536, + 0.005933580454438925, + 0.016743900254368782, + 0.017431972548365593, + 0.037231288850307465, + 0.019816696643829346, + 0.019122429192066193, + 0.020559901371598244, + 0.014392889104783535, + 0.03419043868780136, + 0.005472669843584299, + 0.006450656335800886, + 0.007804650813341141, + 0.006775508634746075, + 0.008291100151836872, + 0.007422547321766615, + 0.005819767713546753, + 0.013180957175791264, + 0.002233153907582164, + 0.0024809904862195253, + 0.006177698262035847, + 0.004554465878754854, + 0.0029638647101819515, + 0.003127289703115821, + 0.002015350852161646, + 0.0036024777218699455, + 0.003035253379493952, + 0.0031977668404579163, + 0.007540186867117882, + 0.003793317824602127, + 0.004778340924531221, + 0.003917722497135401, + 0.0028578408528119326, + 0.005632686894387007, + 0.0021091175731271505, + 0.001792083727195859, + 0.0028912057168781757, + 0.00168830796610564, + 0.0027160688769072294, + 0.003041552845388651, + 0.0019021008629351854, + 0.002204161835834384, + 0.0026959094684571028, + 0.003745810594409704, + 0.0038529811426997185, + 0.0041624861769378185, + 0.004386608023196459, + 0.005834358278661966, + 0.0031970131676644087, + 0.008719658479094505, + 0.009685576893389225, + 0.009359919466078281, + 0.023415960371494293, + 0.011601739563047886, + 0.010015033185482025, + 0.012609998695552349, + 0.007663703989237547, + 0.020549334585666656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.12708574533462524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002490278100594878 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.002490278100594878, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002490278100594878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08743591606616974, + 0.4561064839363098, + 0.01485357154160738, + 0.06351344287395477, + 0.016750264912843704, + 0.011430060490965843, + 0.060195453464984894, + 0.05237415432929993, + 0.20468130707740784, + 0.032659292221069336 + ], + "credit_map": [ + 0.008100799284875393, + 0.007579886820167303, + 0.017797458916902542, + 0.012741966173052788, + 0.00850597396492958, + 0.010208958759903908, + 0.005604560486972332, + 0.01689632050693035, + 0.05761150270700455, + 0.03305521979928017, + 0.05705748498439789, + 0.04643247276544571, + 0.031827300786972046, + 0.05567365139722824, + 0.026894133538007736, + 0.14755472540855408, + 0.0017422096570953727, + 0.0011708452366292477, + 0.002307821763679385, + 0.002026581671088934, + 0.0018301816890016198, + 0.0013213468482717872, + 0.0016383681213483214, + 0.002816216554492712, + 0.006004637107253075, + 0.006093083415180445, + 0.014303509145975113, + 0.0072288308292627335, + 0.006437982432544231, + 0.0069424002431333065, + 0.005007030908018351, + 0.011495969258248806, + 0.001561653334647417, + 0.001618778333067894, + 0.00360623630695045, + 0.0019155668560415506, + 0.0016658831154927611, + 0.0019804283510893583, + 0.0013391741085797548, + 0.003062543924897909, + 0.0013440751936286688, + 0.0008180621080100536, + 0.0029214026872068644, + 0.0016182719264179468, + 0.0009447534685023129, + 0.0012861524010077119, + 0.0007311687804758549, + 0.0017661743331700563, + 0.006131661124527454, + 0.0037314973305910826, + 0.02014794386923313, + 0.008528477512300014, + 0.004423659760504961, + 0.005748711992055178, + 0.003856819588690996, + 0.007626682985574007, + 0.003617302281782031, + 0.005883465986698866, + 0.009167931042611599, + 0.005235330201685429, + 0.007841422222554684, + 0.006136399693787098, + 0.004979787860065699, + 0.009512510150671005, + 0.017740381881594658, + 0.018367793411016464, + 0.051531173288822174, + 0.023827126249670982, + 0.019434496760368347, + 0.02459949254989624, + 0.017330016940832138, + 0.03185082599520683, + 0.003509529633447528, + 0.0027475550305098295, + 0.00839553028345108, + 0.004344523418694735, + 0.003008042462170124, + 0.0034610868897289038, + 0.002371380105614662, + 0.004821646027266979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.002490278100594878 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002829835517331958 + ], + "opposite_logprob_tokens": [ + -5.8778300285339355 + ], + "logprob_reward": -0.002829835517331958, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002829835517331958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07921086996793747, + 0.4639085829257965, + 0.06990107148885727, + 0.1256827861070633, + 0.08709676563739777, + 0.03111160360276699, + 0.058207668364048004, + 0.03146635740995407, + 0.02499707043170929, + 0.02841712161898613 + ], + "credit_map": [ + 0.008185950107872486, + 0.005202121566981077, + 0.01735215075314045, + 0.011576900258660316, + 0.007299116346985102, + 0.008130930364131927, + 0.006016622297465801, + 0.015447072684764862, + 0.05852881446480751, + 0.03315494582056999, + 0.06235244497656822, + 0.04575050249695778, + 0.03118964284658432, + 0.056679580360651016, + 0.02646470256149769, + 0.14978796243667603, + 0.006791999097913504, + 0.005654898472130299, + 0.014987524598836899, + 0.008107254281640053, + 0.00671027647331357, + 0.00799471978098154, + 0.005779995117336512, + 0.013874403201043606, + 0.011854840442538261, + 0.011688526719808578, + 0.027808042243123055, + 0.014202204532921314, + 0.012783586047589779, + 0.013928293250501156, + 0.009816107340157032, + 0.023601187393069267, + 0.008243578486144543, + 0.007288734894245863, + 0.01982746459543705, + 0.011329486034810543, + 0.008439283818006516, + 0.009601072408258915, + 0.007116645574569702, + 0.015250500291585922, + 0.0038663388695567846, + 0.0024184384383261204, + 0.00789222028106451, + 0.0038735491689294577, + 0.0028181832749396563, + 0.00284273992292583, + 0.0026684855110943317, + 0.0047316462732851505, + 0.004472970496863127, + 0.006127486936748028, + 0.0069149574264883995, + 0.005091041326522827, + 0.008024699985980988, + 0.007728497497737408, + 0.005968575831502676, + 0.013879436068236828, + 0.002458017785102129, + 0.0035182139836251736, + 0.0036485143937170506, + 0.0037758515682071447, + 0.004587152041494846, + 0.004256706219166517, + 0.0038199061527848244, + 0.0054019964300096035, + 0.0029357115272432566, + 0.0013621398247778416, + 0.005633515305817127, + 0.0029232425149530172, + 0.0013567795976996422, + 0.002698800526559353, + 0.0014302416238933802, + 0.006656639277935028, + 0.0029133493080735207, + 0.0016816038405522704, + 0.007436797954142094, + 0.0036536608822643757, + 0.002323239343240857, + 0.002751376712694764, + 0.0017735707806423306, + 0.005883523263037205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.002829835517331958 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up portrait focuses with intimate intensity on a distinctively stylish man, his features rendered in sharp detail. He wears a vibrant brown wool felt hat, its chiseled brim casting a soft shadow across his forehead. His face, accented by a carefully groomed mustache and goatee, is stern. His attire is simple yet deliberate: a matte black cotton shirt that absorbs light, providing a stark canvas for a large silver curb chain necklace clinging to his collarbone. The background melts into a soft-focus bokeh of blurred interior space, dominated by a cool, cinematic blue light spilling from an off-camera source, depicting a sharp, electric rim light along his chin and shoulders. This contrasts sharply with the warm, soft key light that models the planes of his face. His dark eyes stare intently at a point just beyond the frame, and his entire being seems to pause as he listens to something beyond the frame or wrestles with deep inner thoughts, creating an atmosphere of quiet, contemplative, and powerful concentration.", + "seed": 20887, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_2/index_2758_seed_20887_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.037954798177815975, + "vlm_logprob": -0.037954798177815975, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.037954798177815975, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01632993295788765 + ], + "opposite_logprob_tokens": [ + -4.891329765319824 + ], + "logprob_reward": -0.01632993295788765, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01632993295788765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07014177739620209, + 0.572096049785614, + 0.041108716279268265, + 0.02267388068139553, + 0.01608414389193058, + 0.06529641151428223, + 0.05598687380552292, + 0.07772073149681091, + 0.03028716705739498, + 0.048604290932416916 + ], + "credit_map": [ + 0.006428336258977652, + 0.0053895991295576096, + 0.00921820942312479, + 0.01194493006914854, + 0.010138353332877159, + 0.008082285523414612, + 0.008071428164839745, + 0.01086863037198782, + 0.05637218430638313, + 0.03073805943131447, + 0.13706117868423462, + 0.19945643842220306, + 0.028671203181147575, + 0.014952268451452255, + 0.0219465009868145, + 0.08289819210767746, + 0.0035682234447449446, + 0.0029758731834590435, + 0.003984425216913223, + 0.01072169654071331, + 0.005834403447806835, + 0.003730467986315489, + 0.0036357964854687452, + 0.006657825782895088, + 0.0017162057338282466, + 0.0022192178294062614, + 0.001965645933523774, + 0.004535018000751734, + 0.004553930833935738, + 0.0016395293641835451, + 0.003053029766306281, + 0.002991301007568836, + 0.0012743620900437236, + 0.001355217769742012, + 0.0019636922515928745, + 0.0023361302446573973, + 0.0028121951036155224, + 0.002518543042242527, + 0.0018802449340000749, + 0.0019437578739598393, + 0.007661435753107071, + 0.006673391442745924, + 0.008034097030758858, + 0.016750866547226906, + 0.006203359924256802, + 0.0036082579754292965, + 0.007638273295015097, + 0.008726723492145538, + 0.0041361781768500805, + 0.004123773891478777, + 0.005335118621587753, + 0.01073935255408287, + 0.004681527614593506, + 0.004781518597155809, + 0.007630991283804178, + 0.01455841027200222, + 0.013514308258891106, + 0.007418992463499308, + 0.006904578302055597, + 0.021406786516308784, + 0.005647994577884674, + 0.0038467550184577703, + 0.0076103536412119865, + 0.011370958760380745, + 0.004242749884724617, + 0.0029388798866420984, + 0.003559085540473461, + 0.005842949729412794, + 0.0035181024577468634, + 0.002058741869404912, + 0.004224054515361786, + 0.0039026017766445875, + 0.004509133752435446, + 0.003954793326556683, + 0.00567880691960454, + 0.009055549278855324, + 0.004761712159961462, + 0.004274826962500811, + 0.007161443587392569, + 0.009208020754158497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.01632993295788765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903405278921127 + ], + "opposite_logprob_tokens": [ + -2.5790340900421143 + ], + "logprob_reward": -0.07903405278921127, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903405278921127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19986580312252045, + 0.5540093183517456, + 0.026726840063929558, + 0.019399605691432953, + 0.02560894936323166, + 0.06654775887727737, + 0.01499275304377079, + 0.02465786412358284, + 0.017787989228963852, + 0.050403155386447906 + ], + "credit_map": [ + 0.02245863899588585, + 0.010575580410659313, + 0.05226345732808113, + 0.030278295278549194, + 0.017301809042692184, + 0.018192127346992493, + 0.0276021771132946, + 0.021193716675043106, + 0.05225996673107147, + 0.030389003455638885, + 0.12723904848098755, + 0.19525326788425446, + 0.02769341506063938, + 0.01597673073410988, + 0.0230095237493515, + 0.08218837529420853, + 0.00228565139696002, + 0.0020058131776750088, + 0.00240185740403831, + 0.0051718950271606445, + 0.0053415447473526, + 0.0026060915552079678, + 0.0030439537949860096, + 0.003870029468089342, + 0.0017239677254110575, + 0.001847717328928411, + 0.002692063106223941, + 0.003431441029533744, + 0.002743460237979889, + 0.0016474269796162844, + 0.0025128645356744528, + 0.00280066579580307, + 0.0018721895758062601, + 0.0034668250009417534, + 0.002282888162881136, + 0.0038518381770700216, + 0.004276423715054989, + 0.0028966122772544622, + 0.004033071920275688, + 0.0029291012324392796, + 0.006375899538397789, + 0.006601070053875446, + 0.005139878951013088, + 0.014570201747119427, + 0.006305929273366928, + 0.005684477277100086, + 0.008606357499957085, + 0.0132639454677701, + 0.0011957045644521713, + 0.0014996720710769296, + 0.0014295921428129077, + 0.0023208209313452244, + 0.001874535228125751, + 0.0017073990311473608, + 0.0027129228692501783, + 0.0022521065548062325, + 0.002306124661117792, + 0.002112838439643383, + 0.002264206763356924, + 0.003968110773712397, + 0.0038546130526810884, + 0.002557882573455572, + 0.0035198477562516928, + 0.004074240569025278, + 0.0020390076097100973, + 0.0020407880656421185, + 0.0014340501511469483, + 0.0029870960861444473, + 0.0029822690412402153, + 0.001745613873936236, + 0.0024461557622998953, + 0.0021130091045051813, + 0.005620933137834072, + 0.0033638121094554663, + 0.004375713411718607, + 0.014706715941429138, + 0.0052899206057190895, + 0.0028777357656508684, + 0.005446720402687788, + 0.008721602149307728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.07903405278921127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907032757997513 + ], + "opposite_logprob_tokens": [ + -2.4640703201293945 + ], + "logprob_reward": -0.08907032757997513, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907032757997513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13372214138507843, + 0.5080652236938477, + 0.0430791899561882, + 0.030076563358306885, + 0.037495214492082596, + 0.05548984929919243, + 0.023215526714920998, + 0.026781661435961723, + 0.03793355077505112, + 0.10414116829633713 + ], + "credit_map": [ + 0.009717350825667381, + 0.006975685711950064, + 0.02094920724630356, + 0.03406262397766113, + 0.014821498654782772, + 0.011679536662995815, + 0.014568986371159554, + 0.02094723843038082, + 0.04073186591267586, + 0.030113551765680313, + 0.09599190205335617, + 0.18318821489810944, + 0.028874672949314117, + 0.015957241877913475, + 0.034531958401203156, + 0.07867573946714401, + 0.004326215945184231, + 0.0037903368938714266, + 0.0042246985249221325, + 0.008976612240076065, + 0.006602092646062374, + 0.004201867152005434, + 0.004543539602309465, + 0.006413823459297419, + 0.003503279760479927, + 0.0022573373280465603, + 0.0036723946686834097, + 0.008473583497107029, + 0.003154283855110407, + 0.0019725258462131023, + 0.003008956555277109, + 0.004034198820590973, + 0.0030731703154742718, + 0.004113766364753246, + 0.0037110974080860615, + 0.007108819670975208, + 0.00537379365414381, + 0.004112919792532921, + 0.004776259418576956, + 0.00522538460791111, + 0.00754070607945323, + 0.005712628830224276, + 0.005775367841124535, + 0.015126630663871765, + 0.004992165137082338, + 0.003426261944696307, + 0.004584288690239191, + 0.00833179336041212, + 0.0029136589728295803, + 0.00191719108261168, + 0.0020603144075721502, + 0.005934171844273806, + 0.002209104597568512, + 0.0015809707110747695, + 0.0027112511452287436, + 0.0038888617418706417, + 0.0038148469757288694, + 0.002419502241536975, + 0.0031370557844638824, + 0.003925804048776627, + 0.003342677606269717, + 0.002033272525295615, + 0.004523284733295441, + 0.0035852165892720222, + 0.005635127425193787, + 0.003471680451184511, + 0.0032122405245900154, + 0.010204752907156944, + 0.003727992530912161, + 0.0021534792613238096, + 0.0034811573568731546, + 0.006047115661203861, + 0.011715736240148544, + 0.007534177508205175, + 0.006659857463091612, + 0.03518558666110039, + 0.010058682411909103, + 0.004582834895700216, + 0.008687429130077362, + 0.01971685327589512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.08907032757997513 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017135950038209558 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017135950038209558, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017135950038209558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1057753711938858, + 0.23107992112636566, + 0.11471033096313477, + 0.04356371983885765, + 0.09827912598848343, + 0.024820426478981972, + 0.013987146317958832, + 0.02819029800593853, + 0.06948097050189972, + 0.2701127231121063 + ], + "credit_map": [ + 0.012078430503606796, + 0.0066971867345273495, + 0.029351048171520233, + 0.020027125254273415, + 0.009711997583508492, + 0.008654785342514515, + 0.00918128527700901, + 0.010073509067296982, + 0.019372904673218727, + 0.01329979207366705, + 0.04593059420585632, + 0.08705716580152512, + 0.01187971793115139, + 0.006391240283846855, + 0.010216256603598595, + 0.036932241171598434, + 0.012576894834637642, + 0.009049886837601662, + 0.01175379753112793, + 0.03629423305392265, + 0.011078245006501675, + 0.0067433807998895645, + 0.009080704301595688, + 0.018133187666535378, + 0.004905968438833952, + 0.0035954380873590708, + 0.005291776731610298, + 0.008441359736025333, + 0.007037946488708258, + 0.004672051873058081, + 0.004517578519880772, + 0.005101601593196392, + 0.012068980373442173, + 0.0087051372975111, + 0.010365325026214123, + 0.029502836987376213, + 0.00875681173056364, + 0.004620501771569252, + 0.008035493083298206, + 0.01622404344379902, + 0.0031034930143505335, + 0.0024215138982981443, + 0.0019646515138447285, + 0.00691545195877552, + 0.002412323374301195, + 0.0013799372827634215, + 0.0021011210046708584, + 0.004521935246884823, + 0.0014818190829828382, + 0.0010830867104232311, + 0.0014908022712916136, + 0.002998039359226823, + 0.0011653511319309473, + 0.0011441224487498403, + 0.001672485494054854, + 0.002951438771560788, + 0.0037336547393351793, + 0.0026742888148874044, + 0.004694823175668716, + 0.0049757990054786205, + 0.003577598836272955, + 0.0017929964233189821, + 0.003684757510200143, + 0.0030563792679458857, + 0.011400636285543442, + 0.007367216516286135, + 0.008714486844837666, + 0.01698639616370201, + 0.005471065640449524, + 0.004770514089614153, + 0.005499789025634527, + 0.00927086640149355, + 0.02901405841112137, + 0.019366826862096786, + 0.018762709572911263, + 0.08633653819561005, + 0.026773937046527863, + 0.01432893518358469, + 0.02554473467171192, + 0.04998497664928436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.0017135950038209558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level close-up portrait of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a vibrant brown wool felt hat with a chiseled brim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a matte black cotton shirt with a large silver curb chain necklace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus bokeh of blurred interior space with cool blue rim lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's face illuminated by a warm, soft key light while his eyes stare intently beyond the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003626082558184862 + ], + "opposite_logprob_tokens": [ + -5.628625869750977 + ], + "logprob_reward": -0.003626082558184862, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003626082558184862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10856018960475922, + 0.08993606269359589, + 0.0721254050731659, + 0.06748512387275696, + 0.061478141695261, + 0.06135354936122894, + 0.04052124544978142, + 0.07572250068187714, + 0.04199424758553505, + 0.3808235824108124 + ], + "credit_map": [ + 0.010495670139789581, + 0.007165166549384594, + 0.025638408958911896, + 0.016401953995227814, + 0.012686599977314472, + 0.010491491295397282, + 0.012414420023560524, + 0.013266484253108501, + 0.009598255157470703, + 0.007001781370490789, + 0.007066939026117325, + 0.03320421651005745, + 0.006705824285745621, + 0.003949597477912903, + 0.007737935986369848, + 0.014671510085463524, + 0.005912747234106064, + 0.0049042971804738045, + 0.008576533757150173, + 0.02030278742313385, + 0.007834259420633316, + 0.004853296559303999, + 0.00664021959528327, + 0.013101265765726566, + 0.005998983979225159, + 0.006002130452543497, + 0.0074001760222017765, + 0.015676334500312805, + 0.008733480237424374, + 0.006872247438877821, + 0.008258297108113766, + 0.008543473668396473, + 0.0074750385247170925, + 0.00574031937867403, + 0.006258675362914801, + 0.015282667241990566, + 0.007545372471213341, + 0.004848068580031395, + 0.006318456958979368, + 0.008009538054466248, + 0.005411825608462095, + 0.008426491171121597, + 0.005862388759851456, + 0.012704147957265377, + 0.00642230361700058, + 0.004541818518191576, + 0.007350341882556677, + 0.010634228587150574, + 0.004618029575794935, + 0.003190490184351802, + 0.004030696582049131, + 0.008603723719716072, + 0.004480176605284214, + 0.00329774571582675, + 0.004794774577021599, + 0.007505608256906271, + 0.010805118829011917, + 0.00740669434890151, + 0.007767864037305117, + 0.0155088622123003, + 0.009723746217787266, + 0.0048995111137628555, + 0.010101801715791225, + 0.00950890127569437, + 0.00667644664645195, + 0.004552793689072132, + 0.0038061507511883974, + 0.010511988773941994, + 0.004344730172306299, + 0.002686549909412861, + 0.0037186751142144203, + 0.0056969113647937775, + 0.04349428787827492, + 0.02842271886765957, + 0.020598817616701126, + 0.1305679976940155, + 0.037284791469573975, + 0.01664271019399166, + 0.02883986569941044, + 0.07497239857912064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.003626082558184862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_3/index_tensor([3312, 3312])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_3/index_tensor([3312, 3312])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a56d18e8e32a4d867fd5fec477ee075cd19fb2c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_3/index_tensor([3312, 3312])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot, placed the viewer at eye level before a majestic Siberian tiger, against a blurred backdrop of snow-covered coniferous forest. Under the cool, bright winter light, the tiger's dense, rich fur reveals every detail: vibrant orange interwoven with pure white, broken by broad, deep black stripes, each hair seemingly sharply defined. At first, its piercing amber eyes narrow slightly, its wiry white whiskers twitching with tension, its facial muscles tense, foreshadowing an impending explosion. Then, without warning, its jaws open to the limit, frozen in a silent yet powerful roar, revealing dagger-sharp canines, serrated molars, and a pink, barbed tongue. A moist sheen lingers on the surface of its teeth. Its facial skin is wrinkled due to extreme tension, forming powerful folds. The whole picture is full of primitive wildness and unquestionable majesty, conveying an awe-inspiring and soul-stirring atmosphere of tension.", + "seed": 7889, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_3/index_3312_seed_7889_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04103394648991525, + "vlm_logprob": -0.04103394648991525, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.04103394648991525, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019749430939555168 + ], + "opposite_logprob_tokens": [ + -4.519749641418457 + ], + "logprob_reward": -0.019749430939555168, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.019749430939555168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06205946207046509, + 0.03086758591234684, + 0.030317125841975212, + 0.06970662623643875, + 0.036808088421821594, + 0.15555164217948914, + 0.052303384989500046, + 0.13075166940689087, + 0.38597849011421204, + 0.04565587639808655 + ], + "credit_map": [ + 0.011238979175686836, + 0.00693815853446722, + 0.0058561572805047035, + 0.008205825462937355, + 0.008271138183772564, + 0.0069631412625312805, + 0.006997994612902403, + 0.007588067092001438, + 0.007032873574644327, + 0.0025765267200767994, + 0.003798616584390402, + 0.002769079292193055, + 0.004320321138948202, + 0.003621171461418271, + 0.0033906332682818174, + 0.0033583620097488165, + 0.003939156886190176, + 0.0031731477938592434, + 0.002879260340705514, + 0.004774111323058605, + 0.005162382032722235, + 0.0037349513731896877, + 0.0035306760109961033, + 0.003123440546914935, + 0.009403049014508724, + 0.00758006889373064, + 0.006581977009773254, + 0.0075036888010799885, + 0.011984270066022873, + 0.00822493527084589, + 0.010723577812314034, + 0.007705055642873049, + 0.006827061530202627, + 0.003223248291760683, + 0.003134670667350292, + 0.0030232060234993696, + 0.004597027320414782, + 0.004167633131146431, + 0.007411203812807798, + 0.004424037411808968, + 0.041929177939891815, + 0.008487817831337452, + 0.04794584587216377, + 0.007762255147099495, + 0.014165988191962242, + 0.011076359078288078, + 0.014135790057480335, + 0.010048404335975647, + 0.004723931662738323, + 0.003990424331277609, + 0.003823369974270463, + 0.005894930101931095, + 0.005758189596235752, + 0.010179286822676659, + 0.01105024665594101, + 0.0068830084055662155, + 0.011974754743278027, + 0.01666458509862423, + 0.0189592894166708, + 0.014683053828775883, + 0.011770796962082386, + 0.02597120963037014, + 0.015354754403233528, + 0.015373224392533302, + 0.022610442712903023, + 0.0121846254914999, + 0.008757886476814747, + 0.0060564810410141945, + 0.009462285786867142, + 0.17324820160865784, + 0.007429847493767738, + 0.14622873067855835, + 0.004263739101588726, + 0.0048154788091778755, + 0.004485882353037596, + 0.006857806351035833, + 0.00821119174361229, + 0.005037717986851931, + 0.006814524997025728, + 0.00516953831538558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.019749430939555168 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16044984757900238 + ], + "opposite_logprob_tokens": [ + -1.9104498624801636 + ], + "logprob_reward": -0.16044984757900238, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16044984757900238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10634220391511917, + 0.05720540136098862, + 0.0513312928378582, + 0.12058968842029572, + 0.09957300871610641, + 0.07643948495388031, + 0.2463677078485489, + 0.08827098459005356, + 0.06002308800816536, + 0.09385713189840317 + ], + "credit_map": [ + 0.026173057034611702, + 0.011685330420732498, + 0.010543297044932842, + 0.011162355542182922, + 0.01019149087369442, + 0.01082585472613573, + 0.012488530948758125, + 0.013272288255393505, + 0.007794905919581652, + 0.007834427990019321, + 0.007516778074204922, + 0.005013578571379185, + 0.009735940024256706, + 0.006495627574622631, + 0.007508662063628435, + 0.0053054774180054665, + 0.008447352796792984, + 0.004852934740483761, + 0.004691517911851406, + 0.008308090269565582, + 0.007934482768177986, + 0.005215718410909176, + 0.0064526451751589775, + 0.005428554490208626, + 0.012500346638262272, + 0.011883223429322243, + 0.011413082480430603, + 0.013710025697946548, + 0.022619543597102165, + 0.01307744812220335, + 0.02039944939315319, + 0.014986575581133366, + 0.02169480361044407, + 0.008739504963159561, + 0.008872587233781815, + 0.01014796458184719, + 0.012335844337940216, + 0.011668196879327297, + 0.016522863879799843, + 0.00959125068038702, + 0.010156712494790554, + 0.0055785165168344975, + 0.009821480140089989, + 0.007953622378408909, + 0.00797257013618946, + 0.009312410838901997, + 0.018424902111291885, + 0.007219275459647179, + 0.03275749459862709, + 0.011976917274296284, + 0.012383081950247288, + 0.01765383780002594, + 0.02380206249654293, + 0.07098676264286041, + 0.04645226523280144, + 0.030355293303728104, + 0.00807973276823759, + 0.008557579480111599, + 0.011012149974703789, + 0.010601711459457874, + 0.009236340411007404, + 0.016091598197817802, + 0.013272371143102646, + 0.011419502086937428, + 0.005759772844612598, + 0.003999367821961641, + 0.005446936469525099, + 0.005177829414606094, + 0.007519443053752184, + 0.013817189261317253, + 0.0066635748371481895, + 0.011638972908258438, + 0.007515168283134699, + 0.015886999666690826, + 0.010660545900464058, + 0.015117069706320763, + 0.011542150750756264, + 0.0114497821778059, + 0.01077363733202219, + 0.010911781340837479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.16044984757900238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052670761942863464 + ], + "opposite_logprob_tokens": [ + -5.255267143249512 + ], + "logprob_reward": -0.0052670761942863464, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052670761942863464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050213124603033066, + 0.06553252786397934, + 0.03899972140789032, + 0.025828996673226357, + 0.05129328742623329, + 0.036149609833955765, + 0.1539141684770584, + 0.02567363530397415, + 0.3089834153652191, + 0.24341146647930145 + ], + "credit_map": [ + 0.010948625393211842, + 0.004861748777329922, + 0.006209517363458872, + 0.004613999277353287, + 0.004746335558593273, + 0.0047227186150848866, + 0.004221067763864994, + 0.009889112785458565, + 0.02060997299849987, + 0.005546376574784517, + 0.008506388403475285, + 0.0056821247562766075, + 0.005279247183352709, + 0.005491908174008131, + 0.006789169739931822, + 0.007627346087247133, + 0.0071290102787315845, + 0.004005363676697016, + 0.004420457873493433, + 0.0036184731870889664, + 0.003276148810982704, + 0.004023229703307152, + 0.005365344695746899, + 0.0071616945788264275, + 0.005627316888421774, + 0.0021247463300824165, + 0.002700376557186246, + 0.0021805448923259974, + 0.003851833753287792, + 0.0027718576602637768, + 0.0026625506579875946, + 0.0039097703993320465, + 0.012578517198562622, + 0.003364491742104292, + 0.005154064390808344, + 0.004208980593830347, + 0.004192298278212547, + 0.0057585877366364, + 0.004594582598656416, + 0.01144176721572876, + 0.010249270126223564, + 0.001941454829648137, + 0.011908548884093761, + 0.002638888079673052, + 0.0030538381543010473, + 0.002075318479910493, + 0.0019263700814917684, + 0.002355922246351838, + 0.007675796281546354, + 0.004731100983917713, + 0.005755042191594839, + 0.007922767661511898, + 0.009012170135974884, + 0.05297718197107315, + 0.0191951896995306, + 0.04664492979645729, + 0.0025141816586256027, + 0.002885698340833187, + 0.002187594072893262, + 0.002957702148705721, + 0.0035954532213509083, + 0.004196100402623415, + 0.004162201192229986, + 0.0031747061293572187, + 0.017509235069155693, + 0.01123555563390255, + 0.007882881909608841, + 0.006615700200200081, + 0.011596580035984516, + 0.13317328691482544, + 0.008171822875738144, + 0.11279839277267456, + 0.028443915769457817, + 0.02119915932416916, + 0.026314808055758476, + 0.028884703293442726, + 0.025615325197577477, + 0.033383604139089584, + 0.03539793938398361, + 0.044172029942274094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.0052670761942863464 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015188835095614195 + ], + "opposite_logprob_tokens": [ + -6.501518726348877 + ], + "logprob_reward": -0.0015188835095614195, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015188835095614195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03309480473399162, + 0.013902395963668823, + 0.010822365991771221, + 0.010996049270033836, + 0.09363523125648499, + 0.09891175478696823, + 0.16135749220848083, + 0.03213384002447128, + 0.509166419506073, + 0.03597962111234665 + ], + "credit_map": [ + 0.0040178499184548855, + 0.003118991618975997, + 0.0035335845313966274, + 0.003376944921910763, + 0.004765122663229704, + 0.0055482834577560425, + 0.004895553458482027, + 0.003838476026430726, + 0.0024668823461979628, + 0.0017646059859544039, + 0.0014741503400728106, + 0.001331747742369771, + 0.0022264779545366764, + 0.0013126160483807325, + 0.0019871711265295744, + 0.0013387446524575353, + 0.0012262966483831406, + 0.0010591772152110934, + 0.0009460058063268661, + 0.0012181161437183619, + 0.0021462002769112587, + 0.0011852141469717026, + 0.001824524370022118, + 0.0012168321991339326, + 0.0016045277006924152, + 0.0011724710930138826, + 0.0010272995568811893, + 0.0014252703404054046, + 0.001253763330169022, + 0.0013240608386695385, + 0.0016661726403981447, + 0.0015224835369735956, + 0.006582077592611313, + 0.007033505477011204, + 0.00899091362953186, + 0.008686209097504616, + 0.013919690623879433, + 0.008255349472165108, + 0.03046032227575779, + 0.009707164019346237, + 0.015253006480634212, + 0.004435694310814142, + 0.027218306437134743, + 0.007389038801193237, + 0.010460161603987217, + 0.010995568707585335, + 0.015866508707404137, + 0.007293463218957186, + 0.008067903108894825, + 0.005066557787358761, + 0.005903612356632948, + 0.008033567108213902, + 0.010750834830105305, + 0.055409934371709824, + 0.021347248926758766, + 0.046777840703725815, + 0.0030829282477498055, + 0.002734759356826544, + 0.0025968141853809357, + 0.0037227675784379244, + 0.0036839593667536974, + 0.00656407605856657, + 0.0051263198256492615, + 0.004622216336429119, + 0.030957134440541267, + 0.019198112189769745, + 0.014670567587018013, + 0.011647304520010948, + 0.025127803906798363, + 0.21242570877075195, + 0.014386532828211784, + 0.18075327575206757, + 0.0021164421923458576, + 0.0028943498618900776, + 0.003240977879613638, + 0.0042673745192587376, + 0.006487532518804073, + 0.004766405560076237, + 0.008572175167500973, + 0.003634361084550619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.0015188835095614195 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184494227170944 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184494227170944, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184494227170944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009700090624392033, + 0.007624343037605286, + 0.004804686643183231, + 0.004190827254205942, + 0.01275057252496481, + 0.03368276357650757, + 0.07501351088285446, + 0.005763043649494648, + 0.8393236398696899, + 0.007146560121327639 + ], + "credit_map": [ + 0.00181162403896451, + 0.000804107403382659, + 0.0012327153235673904, + 0.0010529818246141076, + 0.0011640680022537708, + 0.0009928182698786259, + 0.0016486500389873981, + 0.0009931252570822835, + 0.0010815486311912537, + 0.0010541657684370875, + 0.0008717163582332432, + 0.0009878461714833975, + 0.0009556119330227375, + 0.0008243581978604198, + 0.0011562617728486657, + 0.000692834029905498, + 0.0006166762905195355, + 0.0005273153656162322, + 0.0005971370846964419, + 0.0005902954726479948, + 0.0007315241964533925, + 0.00048311444697901607, + 0.000768920115660876, + 0.0004897034377790987, + 0.0004216601955704391, + 0.0003848574124276638, + 0.00045901216799393296, + 0.00048667268129065633, + 0.000684253522194922, + 0.0004362950858194381, + 0.0007988903671503067, + 0.00051918561803177, + 0.000998343457467854, + 0.0008910027099773288, + 0.0012343706330284476, + 0.0011003419058397412, + 0.0016527881380170584, + 0.001159214531071484, + 0.0040894765406847, + 0.001625034143216908, + 0.007515349425375462, + 0.0017049908638000488, + 0.009672194719314575, + 0.002239476190879941, + 0.00305796111933887, + 0.003382724244147539, + 0.003903191303834319, + 0.00220687803812325, + 0.004014835227280855, + 0.0022875750437378883, + 0.0025592553429305553, + 0.003405585652217269, + 0.0050978525541722775, + 0.02606983110308647, + 0.011017774231731892, + 0.020560799166560173, + 0.0005422785179689527, + 0.0006413363735191524, + 0.0007143084076233208, + 0.0005361827206797898, + 0.0008830952574498951, + 0.0008125901222229004, + 0.0009242647211067379, + 0.0007089877035468817, + 0.04904526099562645, + 0.025674445554614067, + 0.01873287558555603, + 0.012093626894056797, + 0.021883206441998482, + 0.3777787983417511, + 0.016408538445830345, + 0.3177068531513214, + 0.0008865247946232557, + 0.0006883328314870596, + 0.000742135860491544, + 0.0008837307686917484, + 0.0009743259288370609, + 0.0009780001128092408, + 0.0009072109824046493, + 0.0010862986091524363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.018184494227170944 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, placed the viewer at eye level before a majestic Siberian tiger, against a blurred backdrop of snow-covered coniferous forest. Under the cool, bright winter light, the tiger's dense, rich fur reveals every detail: vibrant orange interwoven with pure white, broken by broad, deep black stripes, each hair seemingly sharply defined. At first, its piercing amber eyes narrow slightly, its wiry white whiskers twitching with tension, its facial muscles tense, foreshadowing an impending explosion. Then, without warning, its jaws open to the limit, frozen in a silent yet powerful roar, revealing dagger-sharp canines, serrated molars, and a pink, barbed tongue. A moist sheen lingers on the surface of its teeth. Its facial skin is wrinkled due to extreme tension, forming powerful folds. The whole picture is full of primitive wildness and unquestionable majesty, conveying an awe-inspiring and soul-stirring atmosphere of tension.", + "seed": 7890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_3/index_3312_seed_7890_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06806341372430325, + "vlm_logprob": -0.06806341372430325, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.06806341372430325, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039327386766672134 + ], + "opposite_logprob_tokens": [ + -3.789327383041382 + ], + "logprob_reward": -0.039327386766672134, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.039327386766672134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19827495515346527, + 0.051589593291282654, + 0.09447959810495377, + 0.13222575187683105, + 0.12199188023805618, + 0.04123396798968315, + 0.14416328072547913, + 0.08809661865234375, + 0.014265211299061775, + 0.11367914825677872 + ], + "credit_map": [ + 0.016660936176776886, + 0.008499899879097939, + 0.056602004915475845, + 0.012425962835550308, + 0.015230459161102772, + 0.014161601662635803, + 0.009519620798528194, + 0.06517448276281357, + 0.007660381495952606, + 0.005019072908908129, + 0.0062063164077699184, + 0.004650918301194906, + 0.008650561794638634, + 0.009874541312456131, + 0.003759036073461175, + 0.005768766161054373, + 0.007848474197089672, + 0.004536942578852177, + 0.03155051916837692, + 0.009383064694702625, + 0.005615822039544582, + 0.004754344467073679, + 0.004047280643135309, + 0.02674315869808197, + 0.011398539878427982, + 0.012862454168498516, + 0.03621569648385048, + 0.0076978327706456184, + 0.01564408279955387, + 0.011572771705687046, + 0.009369464591145515, + 0.027464909479022026, + 0.009573099203407764, + 0.014288520440459251, + 0.0387510247528553, + 0.006770083215087652, + 0.012616249732673168, + 0.007274934556335211, + 0.006942473351955414, + 0.025775492191314697, + 0.004090234637260437, + 0.005203642416745424, + 0.005792209878563881, + 0.004217673093080521, + 0.0070048547349870205, + 0.003980786539614201, + 0.0046479906886816025, + 0.0062965755350887775, + 0.016587942838668823, + 0.02586694248020649, + 0.016958434134721756, + 0.016551468521356583, + 0.022813616320490837, + 0.01665571518242359, + 0.01459732186049223, + 0.014131832867860794, + 0.0076460507698357105, + 0.009973401203751564, + 0.007620966527611017, + 0.013572774827480316, + 0.010115740820765495, + 0.006550167221575975, + 0.02485603839159012, + 0.007761481683701277, + 0.0014345626113936305, + 0.002184131648391485, + 0.0013681878335773945, + 0.0019146197009831667, + 0.0016737570986151695, + 0.0013954287860542536, + 0.002838949440047145, + 0.0014555742964148521, + 0.007591496221721172, + 0.007000864017754793, + 0.004397083073854446, + 0.00819997489452362, + 0.005906115286052227, + 0.03887741640210152, + 0.010427926667034626, + 0.031278274953365326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.039327386766672134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11301177740097046 + ], + "opposite_logprob_tokens": [ + -2.2380118370056152 + ], + "logprob_reward": -0.11301177740097046, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11301177740097046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4649399220943451, + 0.0402885302901268, + 0.170439675450325, + 0.04459312558174133, + 0.05984115973114967, + 0.021416420117020607, + 0.015046085231006145, + 0.041341107338666916, + 0.07102470844984055, + 0.07106925547122955 + ], + "credit_map": [ + 0.04534636437892914, + 0.01723485067486763, + 0.17056351900100708, + 0.03247102349996567, + 0.0297186728566885, + 0.028941186144948006, + 0.017196064814925194, + 0.12346825748682022, + 0.006003839895129204, + 0.003940758761018515, + 0.0037028538063168526, + 0.002608923939988017, + 0.007288213819265366, + 0.007010308559983969, + 0.002821110188961029, + 0.006912520155310631, + 0.014582191593945026, + 0.005742354318499565, + 0.0677817240357399, + 0.01295613031834364, + 0.010394689626991749, + 0.009485160000622272, + 0.005319943185895681, + 0.04417748376727104, + 0.0038660187274217606, + 0.005023024510592222, + 0.009497829712927341, + 0.003982252441346645, + 0.004773280117660761, + 0.0027354091871529818, + 0.0069473860785365105, + 0.007767926435917616, + 0.005279208533465862, + 0.009173762053251266, + 0.015621401369571686, + 0.0039105042815208435, + 0.008395002223551273, + 0.004583125468343496, + 0.004285544157028198, + 0.00859261117875576, + 0.0019317403202876449, + 0.00396380852907896, + 0.0028664455749094486, + 0.002375255571678281, + 0.0034632107708603144, + 0.0014680391177535057, + 0.002346054185181856, + 0.003001866163685918, + 0.0014041734393686056, + 0.003477956634014845, + 0.0014134289231151342, + 0.0013025581138208508, + 0.0029781197663396597, + 0.00149638915900141, + 0.0013965043472126126, + 0.001576955197378993, + 0.0035960888490080833, + 0.005648898426443338, + 0.004754208959639072, + 0.004478696268051863, + 0.005331871099770069, + 0.0024414374493062496, + 0.009152998216450214, + 0.005936904344707727, + 0.007663516793400049, + 0.0077002146281301975, + 0.007655118592083454, + 0.011168064549565315, + 0.008505715988576412, + 0.006406272295862436, + 0.013739974237978458, + 0.008185835555195808, + 0.006272926460951567, + 0.007227272726595402, + 0.006698973011225462, + 0.00836918130517006, + 0.006031498778611422, + 0.013995986431837082, + 0.010695933364331722, + 0.01177748292684555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.11301177740097046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963390693068504 + ], + "opposite_logprob_tokens": [ + -5.130963325500488 + ], + "logprob_reward": -0.005963390693068504, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963390693068504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37128326296806335, + 0.08231818675994873, + 0.1800292581319809, + 0.13953319191932678, + 0.07327374070882797, + 0.0323462039232254, + 0.01993609592318535, + 0.026289481669664383, + 0.04663972556591034, + 0.028350843116641045 + ], + "credit_map": [ + 0.03613727539777756, + 0.013835242949426174, + 0.13446111977100372, + 0.025403782725334167, + 0.025449015200138092, + 0.022889818996191025, + 0.013473328202962875, + 0.09963367134332657, + 0.012154964730143547, + 0.00612747622653842, + 0.012139419093728065, + 0.0060753715224564075, + 0.008841990493237972, + 0.016100266948342323, + 0.0055269706062972546, + 0.015351719222962856, + 0.016303082928061485, + 0.005616254173219204, + 0.07085402309894562, + 0.014480267651379108, + 0.009731844998896122, + 0.011488605290651321, + 0.004698147997260094, + 0.04685702547430992, + 0.014649653807282448, + 0.01782895065844059, + 0.02737484686076641, + 0.009351869113743305, + 0.021907970309257507, + 0.017068400979042053, + 0.010341478511691093, + 0.02101001888513565, + 0.0070441728457808495, + 0.007792093325406313, + 0.01725323498249054, + 0.005595842842012644, + 0.00885770097374916, + 0.00591185363009572, + 0.0064556230790913105, + 0.014363213442265987, + 0.00305650825612247, + 0.006004692986607552, + 0.005115481559187174, + 0.004109955858439207, + 0.003914256114512682, + 0.003204343607649207, + 0.0034740676637738943, + 0.0034668990410864353, + 0.0020604911260306835, + 0.0027915204409509897, + 0.0030811242759227753, + 0.0019392077811062336, + 0.0027629926335066557, + 0.0019006971269845963, + 0.0025458720047026873, + 0.002854190533980727, + 0.0035040073562413454, + 0.002481167670339346, + 0.002865417394787073, + 0.003472578478977084, + 0.0027020121924579144, + 0.0044136084616184235, + 0.0033673157449811697, + 0.0034833746030926704, + 0.00475304713472724, + 0.006175723858177662, + 0.004739535506814718, + 0.005931457504630089, + 0.004780068062245846, + 0.005085150711238384, + 0.008360789157450199, + 0.006813954561948776, + 0.003072593128308654, + 0.0024855711963027716, + 0.002407786203548312, + 0.0042044962756335735, + 0.0026893671602010727, + 0.00466876057907939, + 0.0064224665984511375, + 0.002399800345301628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.005963390693068504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495700612664223 + ], + "opposite_logprob_tokens": [ + -2.9299569129943848 + ], + "logprob_reward": -0.05495700612664223, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495700612664223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16441497206687927, + 0.09255586564540863, + 0.1682211309671402, + 0.055193327367305756, + 0.03281180188059807, + 0.008122584782540798, + 0.027716979384422302, + 0.3723694682121277, + 0.0050813923589885235, + 0.07351253181695938 + ], + "credit_map": [ + 0.014799743890762329, + 0.005246239714324474, + 0.05860815942287445, + 0.010405444540083408, + 0.010406594723463058, + 0.008241899311542511, + 0.005753881763666868, + 0.05095298960804939, + 0.015527919866144657, + 0.007871553301811218, + 0.011404948309063911, + 0.006259526126086712, + 0.011329795233905315, + 0.0214044488966465, + 0.005263487808406353, + 0.013494179584085941, + 0.01438901200890541, + 0.0046798717230558395, + 0.06701324135065079, + 0.013373189605772495, + 0.009053424000740051, + 0.008390381932258606, + 0.004399886820465326, + 0.04692210629582405, + 0.006917024962604046, + 0.007946877740323544, + 0.006801555398851633, + 0.004822948016226292, + 0.009151363745331764, + 0.008011753670871258, + 0.006169405300170183, + 0.005372394807636738, + 0.0026896509807556868, + 0.002340095117688179, + 0.011236228048801422, + 0.0016988330753520131, + 0.0025664945133030415, + 0.0017498537199571729, + 0.0018650650745257735, + 0.008665578439831734, + 0.0007524570100940764, + 0.001521559082902968, + 0.0008361933869309723, + 0.0009198287734761834, + 0.0013301311992108822, + 0.0006514457054436207, + 0.0008120790007524192, + 0.0012988900998607278, + 0.003169915173202753, + 0.004561859183013439, + 0.0035585006698966026, + 0.003742113010957837, + 0.0031035502906888723, + 0.0037446259520947933, + 0.0029934789054095745, + 0.002842934103682637, + 0.037427473813295364, + 0.030358947813510895, + 0.029263991862535477, + 0.06088864430785179, + 0.040611281991004944, + 0.034784432500600815, + 0.10348686575889587, + 0.03554778918623924, + 0.0005818239878863096, + 0.0007535530021414161, + 0.0005974944797344506, + 0.000493171566631645, + 0.0006629734998568892, + 0.0006231620791368186, + 0.0007849833345972002, + 0.0005842296523042023, + 0.00548686645925045, + 0.00591324595734477, + 0.0038265367038547993, + 0.011166715994477272, + 0.007082008291035891, + 0.009470888413488865, + 0.019638724625110626, + 0.010927543975412846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.05495700612664223 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270575076341629 + ], + "opposite_logprob_tokens": [ + -2.1270575523376465 + ], + "logprob_reward": -0.1270575076341629, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270575076341629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2572972774505615, + 0.14509227871894836, + 0.04879787936806679, + 0.1040007546544075, + 0.06379979848861694, + 0.056782741099596024, + 0.03240175172686577, + 0.1720253974199295, + 0.0234491266310215, + 0.0963529720902443 + ], + "credit_map": [ + 0.023243894800543785, + 0.010143712162971497, + 0.06945910304784775, + 0.015275098383426666, + 0.01912813074886799, + 0.023751650005578995, + 0.011749272234737873, + 0.08454644680023193, + 0.02514396235346794, + 0.011710669845342636, + 0.016941657289862633, + 0.010547835379838943, + 0.01880960911512375, + 0.03327689319849014, + 0.008491343818604946, + 0.020170317962765694, + 0.005390917882323265, + 0.0036166079808026552, + 0.009451285004615784, + 0.005283260717988014, + 0.006325734779238701, + 0.005474652163684368, + 0.0032545723952353, + 0.010000853799283504, + 0.011023073457181454, + 0.012997869402170181, + 0.02208508364856243, + 0.006852764170616865, + 0.01587076112627983, + 0.012696363963186741, + 0.007780326530337334, + 0.014694515615701675, + 0.005618414841592312, + 0.0046922266483306885, + 0.021379241719841957, + 0.0034071500413119793, + 0.005145803559571505, + 0.003967089578509331, + 0.0031643230468034744, + 0.01642555557191372, + 0.005220718216150999, + 0.0066860755905508995, + 0.009504654444754124, + 0.006145082879811525, + 0.007909045554697514, + 0.004180487245321274, + 0.005823470186442137, + 0.011313210241496563, + 0.002704255748540163, + 0.0071925013326108456, + 0.005157401319593191, + 0.003485152032226324, + 0.0046517010778188705, + 0.002743524732068181, + 0.002932619070634246, + 0.003534599207341671, + 0.022001754492521286, + 0.01394165214151144, + 0.017736023291945457, + 0.023713935166597366, + 0.017992449924349785, + 0.026779720559716225, + 0.03093072958290577, + 0.018929146230220795, + 0.0023826591204851866, + 0.002010532421991229, + 0.0020540831610560417, + 0.003342707408592105, + 0.0021798245143145323, + 0.0033352391328662634, + 0.005342287477105856, + 0.0028017954900860786, + 0.00786634162068367, + 0.004890107549726963, + 0.003642462659627199, + 0.005038118455559015, + 0.005079988390207291, + 0.03415593504905701, + 0.005588720552623272, + 0.030091309919953346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.1270575076341629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, placed the viewer at eye level before a majestic Siberian tiger, against a blurred backdrop of snow-covered coniferous forest. Under the cool, bright winter light, the tiger's dense, rich fur reveals every detail: vibrant orange interwoven with pure white, broken by broad, deep black stripes, each hair seemingly sharply defined. At first, its piercing amber eyes narrow slightly, its wiry white whiskers twitching with tension, its facial muscles tense, foreshadowing an impending explosion. Then, without warning, its jaws open to the limit, frozen in a silent yet powerful roar, revealing dagger-sharp canines, serrated molars, and a pink, barbed tongue. A moist sheen lingers on the surface of its teeth. Its facial skin is wrinkled due to extreme tension, forming powerful folds. The whole picture is full of primitive wildness and unquestionable majesty, conveying an awe-inspiring and soul-stirring atmosphere of tension.", + "seed": 7891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_3/index_3312_seed_7891_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030052110110409556, + "vlm_logprob": -0.030052110110409556, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.030052110110409556, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01872587390244007 + ], + "opposite_logprob_tokens": [ + -4.893725872039795 + ], + "logprob_reward": -0.01872587390244007, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01872587390244007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11559056490659714, + 0.05246410146355629, + 0.03313720226287842, + 0.052084241062402725, + 0.06251855939626694, + 0.39061281085014343, + 0.07522393018007278, + 0.05591416731476784, + 0.09520874917507172, + 0.06724569946527481 + ], + "credit_map": [ + 0.016017919406294823, + 0.01628025248646736, + 0.011330592446029186, + 0.016003068536520004, + 0.013285722583532333, + 0.012306226417422295, + 0.012192506343126297, + 0.018174272030591965, + 0.005622959695756435, + 0.005636758171021938, + 0.00489056296646595, + 0.007396314293146133, + 0.006768150720745325, + 0.0062753926031291485, + 0.006938929669559002, + 0.008935036137700081, + 0.002831772668287158, + 0.00833384320139885, + 0.003202559193596244, + 0.0036276073660701513, + 0.0034776758402585983, + 0.0031065118964761496, + 0.004672234877943993, + 0.003884995821863413, + 0.006987086031585932, + 0.007673843298107386, + 0.004055446945130825, + 0.006349996197968721, + 0.005668702069669962, + 0.0055679562501609325, + 0.007348949555307627, + 0.008432259783148766, + 0.004772443324327469, + 0.008769449777901173, + 0.005803974810987711, + 0.006136756390333176, + 0.004664409905672073, + 0.009079305455088615, + 0.008405929431319237, + 0.014886289834976196, + 0.02569054253399372, + 0.05773013457655907, + 0.031172065064311028, + 0.07151865214109421, + 0.03879036009311676, + 0.035799093544483185, + 0.09419147670269012, + 0.03572049364447594, + 0.008875000290572643, + 0.012929159216582775, + 0.008169570006430149, + 0.007107982877641916, + 0.010858610272407532, + 0.007954047061502934, + 0.011122473515570164, + 0.008207092061638832, + 0.005867999512702227, + 0.0053827217780053616, + 0.005808951333165169, + 0.005033750087022781, + 0.0103386789560318, + 0.005585735198110342, + 0.012468032538890839, + 0.0054282997734844685, + 0.008935913443565369, + 0.008633111603558064, + 0.008319756016135216, + 0.008122163824737072, + 0.024569926783442497, + 0.009810300543904305, + 0.015240909531712532, + 0.011576669290661812, + 0.006715197116136551, + 0.0075225443579256535, + 0.006513528060168028, + 0.006747633218765259, + 0.01459358911961317, + 0.007272390183061361, + 0.01032971777021885, + 0.007551099173724651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.01872587390244007 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790446326136589 + ], + "opposite_logprob_tokens": [ + -2.5790445804595947 + ], + "logprob_reward": -0.0790446326136589, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790446326136589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047154851257801056, + 0.0551958866417408, + 0.05047973617911339, + 0.10879989713430405, + 0.04832221195101738, + 0.4474358558654785, + 0.0814909115433693, + 0.043493080884218216, + 0.03648918867111206, + 0.08113839477300644 + ], + "credit_map": [ + 0.0065134447067976, + 0.007391463033854961, + 0.0038822030182927847, + 0.004930148832499981, + 0.005795910023152828, + 0.005049009807407856, + 0.005810869857668877, + 0.007781804073601961, + 0.006443328224122524, + 0.007104053162038326, + 0.005621808581054211, + 0.006951460149139166, + 0.007530071772634983, + 0.005566377192735672, + 0.007101709954440594, + 0.008877082727849483, + 0.004805692471563816, + 0.0062376162968575954, + 0.004852256737649441, + 0.008924325928092003, + 0.0054169208742678165, + 0.00662616454064846, + 0.0077725667506456375, + 0.005844198167324066, + 0.017088744789361954, + 0.014721698127686977, + 0.012368934229016304, + 0.010571236722171307, + 0.012637557461857796, + 0.01162452157586813, + 0.012479601427912712, + 0.017307618632912636, + 0.004072240088135004, + 0.004805445205420256, + 0.005499531049281359, + 0.005524123087525368, + 0.0052276114001870155, + 0.005610224325209856, + 0.008893150836229324, + 0.008689891546964645, + 0.034687235951423645, + 0.06900894641876221, + 0.03735638037323952, + 0.07864559441804886, + 0.0460280105471611, + 0.03857337683439255, + 0.10364221036434174, + 0.039494141936302185, + 0.01016050111502409, + 0.00824248231947422, + 0.006059425882995129, + 0.007935338653624058, + 0.011127112433314323, + 0.008111250586807728, + 0.019719358533620834, + 0.010135450400412083, + 0.004930625203996897, + 0.004456990398466587, + 0.004212859086692333, + 0.00467240484431386, + 0.006705035921186209, + 0.004859672859311104, + 0.0076532429084181786, + 0.006002255715429783, + 0.0024617724120616913, + 0.003963721916079521, + 0.003211564151570201, + 0.003512027207762003, + 0.006923638749867678, + 0.003467280650511384, + 0.008361859247088432, + 0.004587323870509863, + 0.009181574918329716, + 0.007382130715996027, + 0.007041449658572674, + 0.006221691146492958, + 0.018204793334007263, + 0.011733286082744598, + 0.011208191514015198, + 0.010165284387767315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.0790446326136589 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00250038574449718 + ], + "opposite_logprob_tokens": [ + -6.002500534057617 + ], + "logprob_reward": -0.00250038574449718, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00250038574449718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1827058494091034, + 0.0698470026254654, + 0.1444569230079651, + 0.09258303791284561, + 0.18318800628185272, + 0.17876660823822021, + 0.051068127155303955, + 0.03310336917638779, + 0.047252923250198364, + 0.017028160393238068 + ], + "credit_map": [ + 0.020757436752319336, + 0.019696835428476334, + 0.015846168622374535, + 0.01644749939441681, + 0.016724733635783195, + 0.02470015361905098, + 0.0185818150639534, + 0.04995119571685791, + 0.007555138319730759, + 0.00813630037009716, + 0.0071074324660003185, + 0.007306019309908152, + 0.00796432327479124, + 0.008689532987773418, + 0.007622621487826109, + 0.015465638600289822, + 0.013193314895033836, + 0.015683716163039207, + 0.012714888900518417, + 0.01679009385406971, + 0.014608370140194893, + 0.018396681174635887, + 0.02019028551876545, + 0.032879576086997986, + 0.011140571907162666, + 0.007583466824144125, + 0.008486402221024036, + 0.008391549810767174, + 0.007604310289025307, + 0.01377170067280531, + 0.011050986126065254, + 0.02455405332148075, + 0.01027120091021061, + 0.01948344148695469, + 0.012128233909606934, + 0.014781787060201168, + 0.01574072614312172, + 0.030274059623479843, + 0.02800220251083374, + 0.05250636115670204, + 0.013026890344917774, + 0.02690313570201397, + 0.014411154203116894, + 0.027024494484066963, + 0.016032852232456207, + 0.017386341467499733, + 0.03662736341357231, + 0.02735436148941517, + 0.004612055607140064, + 0.006589935161173344, + 0.0041472720913589, + 0.005631574429571629, + 0.006624947767704725, + 0.0059946742840111256, + 0.010874390602111816, + 0.006593278143554926, + 0.0025098109617829323, + 0.0024948709178715944, + 0.0033866255544126034, + 0.004090291913598776, + 0.004245676565915346, + 0.003737102961167693, + 0.006707089953124523, + 0.005931898485869169, + 0.003157601924613118, + 0.004379143938422203, + 0.0035548079758882523, + 0.0050706104375422, + 0.010292546823620796, + 0.004374172072857618, + 0.010676275938749313, + 0.005747761577367783, + 0.0020966308657079935, + 0.0015299514634534717, + 0.0016837473958730698, + 0.00189398683141917, + 0.0034649402368813753, + 0.0020731615368276834, + 0.0020420432556420565, + 0.0022436960134655237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.00250038574449718 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013404440833255649 + ], + "opposite_logprob_tokens": [ + -6.626340389251709 + ], + "logprob_reward": -0.0013404440833255649, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013404440833255649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059512700885534286, + 0.04398020729422569, + 0.05863195285201073, + 0.043997108936309814, + 0.053897347301244736, + 0.08649028837680817, + 0.19691063463687897, + 0.16606038808822632, + 0.15853528678417206, + 0.13198406994342804 + ], + "credit_map": [ + 0.005274003371596336, + 0.004841268993914127, + 0.0043114665895700455, + 0.006620748434215784, + 0.0077832406386733055, + 0.00981234572827816, + 0.00842984113842249, + 0.012439786456525326, + 0.004799060523509979, + 0.005289482418447733, + 0.004071989096701145, + 0.00584060326218605, + 0.005323964636772871, + 0.0054560028947889805, + 0.003968830220401287, + 0.00923027191311121, + 0.006673100404441357, + 0.0077700549736619, + 0.006855233572423458, + 0.006984354462474585, + 0.0061959619633853436, + 0.00613360945135355, + 0.008379213511943817, + 0.00964042916893959, + 0.0035192023497074842, + 0.004551470745354891, + 0.003478205529972911, + 0.006102423649281263, + 0.004122383426874876, + 0.005763811990618706, + 0.00636261748149991, + 0.010096992366015911, + 0.003970213700085878, + 0.006259267218410969, + 0.003360970877110958, + 0.009406407363712788, + 0.004761564079672098, + 0.005309377331286669, + 0.01405174471437931, + 0.00677780294790864, + 0.007653157226741314, + 0.007806285284459591, + 0.0053696478717029095, + 0.013779773376882076, + 0.008474703878164291, + 0.009217238053679466, + 0.022598059847950935, + 0.011591426096856594, + 0.018531955778598785, + 0.01891021616756916, + 0.012402260676026344, + 0.02247692085802555, + 0.028281431645154953, + 0.01954948902130127, + 0.05269978567957878, + 0.024058567360043526, + 0.012740413658320904, + 0.014880452305078506, + 0.013278002850711346, + 0.018080400303006172, + 0.023110592737793922, + 0.019432779401540756, + 0.03654605895280838, + 0.027991682291030884, + 0.014033050276339054, + 0.015318162739276886, + 0.012324118986725807, + 0.014674825593829155, + 0.03341768682003021, + 0.016663920134305954, + 0.03107699379324913, + 0.02102651447057724, + 0.01034168154001236, + 0.01516731083393097, + 0.009851940907537937, + 0.012882371433079243, + 0.03682025521993637, + 0.013258115388453007, + 0.01656487211585045, + 0.01709751784801483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.0013404440833255649 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864921420812607 + ], + "opposite_logprob_tokens": [ + -3.048649311065674 + ], + "logprob_reward": -0.04864921420812607, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864921420812607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08487498760223389, + 0.03867989778518677, + 0.03926698863506317, + 0.04615018516778946, + 0.06359352916479111, + 0.048299506306648254, + 0.1877780705690384, + 0.2927151918411255, + 0.1284928172826767, + 0.0701487809419632 + ], + "credit_map": [ + 0.008803490549325943, + 0.009354918263852596, + 0.008616882376372814, + 0.010902320966124535, + 0.00824664905667305, + 0.01113041490316391, + 0.01087256520986557, + 0.016947750002145767, + 0.0037927618250250816, + 0.003912052605301142, + 0.004174883011728525, + 0.007244188338518143, + 0.003693015081807971, + 0.005769757553935051, + 0.004124870523810387, + 0.0059683737345039845, + 0.003178548999130726, + 0.006283873226493597, + 0.0033971043303608894, + 0.00839262269437313, + 0.0036324039101600647, + 0.0046354723162949085, + 0.004551200196146965, + 0.005195763427764177, + 0.0049550035037100315, + 0.004996153060346842, + 0.0046235634945333, + 0.01004133839160204, + 0.004681746941059828, + 0.004768290091305971, + 0.005688236095011234, + 0.006395857315510511, + 0.007387630175799131, + 0.006285724230110645, + 0.005569359287619591, + 0.008395735174417496, + 0.006187606602907181, + 0.006445711478590965, + 0.013031244277954102, + 0.010290524922311306, + 0.0037938046734780073, + 0.005628583952784538, + 0.0047243679873645306, + 0.0060056899674236774, + 0.004901059903204441, + 0.0055643608793616295, + 0.009309028275310993, + 0.008372612297534943, + 0.016657382249832153, + 0.01639530621469021, + 0.012879308313131332, + 0.020346904173493385, + 0.02571328543126583, + 0.021887926384806633, + 0.04390297830104828, + 0.029994988813996315, + 0.018412776291370392, + 0.025913618505001068, + 0.028967062011361122, + 0.04252832010388374, + 0.03958780691027641, + 0.031696468591690063, + 0.0572584792971611, + 0.04835068807005882, + 0.008323028683662415, + 0.013262024149298668, + 0.009649205952882767, + 0.01071111112833023, + 0.029293118044734, + 0.013618509285151958, + 0.020052839070558548, + 0.023582996800541878, + 0.0059447879903018475, + 0.006946668494492769, + 0.005419162567704916, + 0.006167861167341471, + 0.017739826813340187, + 0.007730494718998671, + 0.008594617247581482, + 0.011605367995798588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.04864921420812607 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, placed the viewer at eye level before a majestic Siberian tiger, against a blurred backdrop of snow-covered coniferous forest. Under the cool, bright winter light, the tiger's dense, rich fur reveals every detail: vibrant orange interwoven with pure white, broken by broad, deep black stripes, each hair seemingly sharply defined. At first, its piercing amber eyes narrow slightly, its wiry white whiskers twitching with tension, its facial muscles tense, foreshadowing an impending explosion. Then, without warning, its jaws open to the limit, frozen in a silent yet powerful roar, revealing dagger-sharp canines, serrated molars, and a pink, barbed tongue. A moist sheen lingers on the surface of its teeth. Its facial skin is wrinkled due to extreme tension, forming powerful folds. The whole picture is full of primitive wildness and unquestionable majesty, conveying an awe-inspiring and soul-stirring atmosphere of tension.", + "seed": 7892, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_3/index_3312_seed_7892_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04858601111918688, + "vlm_logprob": -0.04858601111918688, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04858601111918688, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07083997130393982 + ], + "opposite_logprob_tokens": [ + -2.9458398818969727 + ], + "logprob_reward": -0.07083997130393982, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07083997130393982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07198598980903625, + 0.04230141267180443, + 0.16736553609371185, + 0.04003699868917465, + 0.03499537706375122, + 0.03499925881624222, + 0.21623890101909637, + 0.11955076456069946, + 0.20750519633293152, + 0.06502050161361694 + ], + "credit_map": [ + 0.008491480723023415, + 0.006577675696462393, + 0.009069116786122322, + 0.006872702855616808, + 0.010237427428364754, + 0.008744845166802406, + 0.010354596190154552, + 0.011638144962489605, + 0.006038064602762461, + 0.004130071494728327, + 0.004572730511426926, + 0.004519354086369276, + 0.007290168199688196, + 0.003549757180735469, + 0.007150723598897457, + 0.005050540901720524, + 0.030011771246790886, + 0.012830965220928192, + 0.01232787687331438, + 0.007288255728781223, + 0.022114835679531097, + 0.03804587945342064, + 0.015428045764565468, + 0.029317907989025116, + 0.008105372078716755, + 0.003300820477306843, + 0.0035165485460311174, + 0.0027295376639813185, + 0.0076143778860569, + 0.004438447300344706, + 0.005506501067429781, + 0.004825393669307232, + 0.00793072022497654, + 0.003761038649827242, + 0.0035520668607205153, + 0.002737062983214855, + 0.00429536821320653, + 0.00407422985881567, + 0.004507696721702814, + 0.004137193318456411, + 0.005741451401263475, + 0.0028947307728230953, + 0.004296737723052502, + 0.0031863597687333822, + 0.005023532081395388, + 0.0034558686893433332, + 0.006113570183515549, + 0.004287010990083218, + 0.026138456538319588, + 0.016852131113409996, + 0.04979383945465088, + 0.014408232644200325, + 0.02551587112247944, + 0.015479112043976784, + 0.01727483607828617, + 0.0507764033973217, + 0.015127371065318584, + 0.011803391389548779, + 0.013682258315384388, + 0.01069836039096117, + 0.026594601571559906, + 0.010320579633116722, + 0.013017455115914345, + 0.018306752666831017, + 0.022007642313838005, + 0.019043970853090286, + 0.02719588577747345, + 0.019387025386095047, + 0.0429534949362278, + 0.016041161492466927, + 0.02383444644510746, + 0.03704157471656799, + 0.005631791427731514, + 0.0039357952773571014, + 0.012615000829100609, + 0.005666714161634445, + 0.011203877627849579, + 0.005069734528660774, + 0.013895601965487003, + 0.007001986261457205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.07083997130393982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10041049122810364 + ], + "opposite_logprob_tokens": [ + -2.3504104614257812 + ], + "logprob_reward": -0.10041049122810364, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10041049122810364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032396528869867325, + 0.01479889266192913, + 0.11426424980163574, + 0.025215787813067436, + 0.021924952045083046, + 0.02978304959833622, + 0.1824970692396164, + 0.14846085011959076, + 0.38206857442855835, + 0.048590052872896194 + ], + "credit_map": [ + 0.0038031511940062046, + 0.0037794061936438084, + 0.0034995556343346834, + 0.0033572539687156677, + 0.0068554142490029335, + 0.0034522104542702436, + 0.003838558215647936, + 0.0038109757006168365, + 0.0016268203034996986, + 0.0017221991438418627, + 0.0014051280450075865, + 0.0012646880932152271, + 0.0023459477815777063, + 0.0012982484186068177, + 0.0027131757233291864, + 0.0024226848036050797, + 0.020817361772060394, + 0.008928832598030567, + 0.008327294141054153, + 0.004869450349360704, + 0.014362700283527374, + 0.026200320571660995, + 0.00936706643551588, + 0.02139122411608696, + 0.002798962639644742, + 0.0029178180266171694, + 0.002519060391932726, + 0.0025237849913537502, + 0.004815375432372093, + 0.002599294064566493, + 0.0030234798323363066, + 0.004018012899905443, + 0.0036955776158720255, + 0.002751173684373498, + 0.002655462361872196, + 0.0016589233418926597, + 0.0028816363774240017, + 0.002837180392816663, + 0.0029048265423625708, + 0.002540172776207328, + 0.005453946068882942, + 0.0018239550990983844, + 0.0027837171219289303, + 0.0024060984142124653, + 0.004301330074667931, + 0.003174860030412674, + 0.0039699021726846695, + 0.005869239568710327, + 0.023348724469542503, + 0.012676453217864037, + 0.033879879862070084, + 0.009311777539551258, + 0.016273345798254013, + 0.022092711180448532, + 0.01335194893181324, + 0.051562223583459854, + 0.018317120149731636, + 0.01547886710613966, + 0.0163678377866745, + 0.013092915527522564, + 0.03484085574746132, + 0.01267607044428587, + 0.015421235002577305, + 0.02226596139371395, + 0.05485007166862488, + 0.04175214096903801, + 0.04561322554945946, + 0.04270235076546669, + 0.050511013716459274, + 0.03533340245485306, + 0.04889203980565071, + 0.062414322048425674, + 0.004687295760959387, + 0.0029736345168203115, + 0.011569123715162277, + 0.0036671129055321217, + 0.007798308506608009, + 0.004435985814779997, + 0.006673024035990238, + 0.0067855664528906345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.10041049122810364 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014224916696548462 + ], + "opposite_logprob_tokens": [ + -4.264225006103516 + ], + "logprob_reward": -0.014224916696548462, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014224916696548462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13423478603363037, + 0.10830166190862656, + 0.06852546334266663, + 0.14984165132045746, + 0.1331084668636322, + 0.045954614877700806, + 0.05624304711818695, + 0.040942441672086716, + 0.13812489807605743, + 0.1247229054570198 + ], + "credit_map": [ + 0.02420942112803459, + 0.009482438676059246, + 0.013968528248369694, + 0.008953535929322243, + 0.011475143022835255, + 0.023356035351753235, + 0.013430596329271793, + 0.02935909479856491, + 0.02545054256916046, + 0.009117516689002514, + 0.011902675963938236, + 0.006005598232150078, + 0.01381780207157135, + 0.017929300665855408, + 0.009836667217314243, + 0.014241566881537437, + 0.010010997764766216, + 0.007477885577827692, + 0.005280274897813797, + 0.005390430800616741, + 0.011212064884603024, + 0.009245490655303001, + 0.010034671984612942, + 0.009873656556010246, + 0.04217346012592316, + 0.011524439789354801, + 0.014211316592991352, + 0.008002415299415588, + 0.019010812044143677, + 0.020475193858146667, + 0.01370982825756073, + 0.020734213292598724, + 0.03941282257437706, + 0.011665613390505314, + 0.013514234684407711, + 0.010833818465471268, + 0.012915248051285744, + 0.014999715611338615, + 0.014147954992949963, + 0.015619072131812572, + 0.009208988398313522, + 0.004166711587458849, + 0.005147326271981001, + 0.0034966710954904556, + 0.006564965937286615, + 0.004662254359573126, + 0.0064965286292135715, + 0.0062111737206578255, + 0.00749594671651721, + 0.005001529585570097, + 0.006346198730170727, + 0.0034093926660716534, + 0.010055753402411938, + 0.008953280746936798, + 0.00626108655706048, + 0.008719862438738346, + 0.005179530940949917, + 0.003911526873707771, + 0.005084459204226732, + 0.003532598027959466, + 0.008088978007435799, + 0.004383571911603212, + 0.005497677717357874, + 0.005264105275273323, + 0.019652605056762695, + 0.017969895154237747, + 0.015193707309663296, + 0.014749100431799889, + 0.018120089545845985, + 0.012739160098135471, + 0.018384413793683052, + 0.021315935999155045, + 0.015202735550701618, + 0.007446189410984516, + 0.012540564872324467, + 0.00904075801372528, + 0.014450537972152233, + 0.007646895479410887, + 0.022672444581985474, + 0.03572278842329979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.014224916696548462 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002500266768038273 + ], + "opposite_logprob_tokens": [ + -6.002500057220459 + ], + "logprob_reward": -0.002500266768038273, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002500266768038273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03909241408109665, + 0.03776801750063896, + 0.07662578672170639, + 0.024274785071611404, + 0.034814849495887756, + 0.02489614300429821, + 0.039602987468242645, + 0.18187963962554932, + 0.05455633997917175, + 0.48648908734321594 + ], + "credit_map": [ + 0.005820000544190407, + 0.004040881525725126, + 0.00554296700283885, + 0.0033780031371861696, + 0.006298809312283993, + 0.004961709957569838, + 0.004161184653639793, + 0.004888854455202818, + 0.006953095551580191, + 0.0024298159405589104, + 0.003475311677902937, + 0.0020831446163356304, + 0.003973358776420355, + 0.006741689518094063, + 0.003999692853540182, + 0.008111909963190556, + 0.01263562310487032, + 0.005677401553839445, + 0.006412649992853403, + 0.003418550593778491, + 0.009538982063531876, + 0.016448454931378365, + 0.007664295844733715, + 0.014829822815954685, + 0.003946925047785044, + 0.002267986536026001, + 0.0019728096667677164, + 0.0019962312653660774, + 0.0033334302715957165, + 0.0026260213926434517, + 0.004704936407506466, + 0.0034264440182596445, + 0.005064453464001417, + 0.002126604551449418, + 0.003011666936799884, + 0.0030235929880291224, + 0.004562973976135254, + 0.004437314812093973, + 0.008178317919373512, + 0.004409923683851957, + 0.00437846640124917, + 0.001671886071562767, + 0.0025460547767579556, + 0.002047936664894223, + 0.004436721559613943, + 0.002549185650423169, + 0.0025093050207942724, + 0.0047565847635269165, + 0.006120601668953896, + 0.003485019551590085, + 0.004023968707770109, + 0.0028075857553631067, + 0.004010115284472704, + 0.005518204066902399, + 0.004111939575523138, + 0.009525550529360771, + 0.02185753732919693, + 0.015387721359729767, + 0.018829135224223137, + 0.014298481866717339, + 0.04804832860827446, + 0.01846633478999138, + 0.022824497893452644, + 0.022167593240737915, + 0.006866435054689646, + 0.008920054882764816, + 0.007021417375653982, + 0.004145678132772446, + 0.009357678703963757, + 0.004535018466413021, + 0.006006268784403801, + 0.007703783921897411, + 0.046518608927726746, + 0.027475647628307343, + 0.06512756645679474, + 0.03692597150802612, + 0.05695144832134247, + 0.028984563425183296, + 0.09053374826908112, + 0.13397151231765747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.002500266768038273 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a Siberian tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and placed at eye level with the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a blurred snow-covered coniferous forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger open its jaws wide to reveal sharp canines, molars, and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's facial skin wrinkled with folds due to extreme tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549544095993042 + ], + "opposite_logprob_tokens": [ + -2.9299545288085938 + ], + "logprob_reward": -0.0549544095993042, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549544095993042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061291180551052094, + 0.026151636615395546, + 0.0878085047006607, + 0.03349713608622551, + 0.03096734546124935, + 0.06288813799619675, + 0.04193994402885437, + 0.11588549613952637, + 0.1280391812324524, + 0.4115314781665802 + ], + "credit_map": [ + 0.007455939892679453, + 0.006119292229413986, + 0.007690771482884884, + 0.006159726530313492, + 0.009902849793434143, + 0.007159180473536253, + 0.008226773701608181, + 0.00857663806527853, + 0.0032302618492394686, + 0.002171478932723403, + 0.0035279500298202038, + 0.0019467101665213704, + 0.003526533953845501, + 0.00310140335932374, + 0.0027758064679801464, + 0.0058714887127280235, + 0.014326915144920349, + 0.005276178941130638, + 0.006455143913626671, + 0.003217788180336356, + 0.009899049997329712, + 0.02209234982728958, + 0.006971193943172693, + 0.0195698793977499, + 0.005194688215851784, + 0.0029147069435566664, + 0.002816088031977415, + 0.002154221525415778, + 0.007422720082104206, + 0.0030410296749323606, + 0.005941647104918957, + 0.004012031946331263, + 0.0040589976124465466, + 0.0017972307978197932, + 0.0024244545493274927, + 0.0029486967250704765, + 0.00428218487650156, + 0.0035724793560802937, + 0.008343311958014965, + 0.0035399908665567636, + 0.008401656523346901, + 0.004588218405842781, + 0.00617754552513361, + 0.005385093856602907, + 0.009692985564470291, + 0.004975112620741129, + 0.01018994115293026, + 0.013477581553161144, + 0.004752573557198048, + 0.003006697865203023, + 0.005446086172014475, + 0.002943346044048667, + 0.00587483262643218, + 0.006890076212584972, + 0.004831880331039429, + 0.008194446563720703, + 0.013461951166391373, + 0.00962580181658268, + 0.011265215463936329, + 0.009409391321241856, + 0.02863764949142933, + 0.012875133194029331, + 0.017586268484592438, + 0.013024085201323032, + 0.016267679631710052, + 0.016794899478554726, + 0.01443525031208992, + 0.014134818688035011, + 0.019256042316555977, + 0.00993733387440443, + 0.01655433513224125, + 0.02065880596637726, + 0.051230017095804214, + 0.02477666363120079, + 0.02419000118970871, + 0.034232139587402344, + 0.0552111454308033, + 0.0397348552942276, + 0.08242500573396683, + 0.09973160177469254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0549544095993042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_4/index_tensor([2221, 2221])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_4/index_tensor([2221, 2221])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..716e0397e636ca9dafa3a36ebb58b9fba20e3566 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_4/index_tensor([2221, 2221])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "single, high-speed, dynamic tracking shot, starting from a low angle and spiraling upwards in unison with the subject's movements, weaves through a vast, mountainous scrapyard. At the center of the frame, a young man in a faded gray hoodie and oil-stained overalls, wearing heavy industrial boots, clutches a massive, industrial-grade energy weapon. The weapon is ruggedly constructed, crisscrossed with pipes and fins, its core emitting a dazzling blue arc of electricity when activated. With his legs firmly planted on the twisted pile of metal, he leans back to gather power, then, with a dramatic, explosive gesture, slams the weapon forward. A visible, fan-shaped wave of energy erupts from the weapon's front, instantly impacting the mountain of scrap metal ahead. Countless rusted car hulls, twisted rebar, and broken mechanical parts are instantly ripped apart and flung into the air by this invisible force, creating a spectacular storm of metal. The fragments gleam coldly in the glaring sunlight, scattering across the sky. The camera uses a first-person drone-like perspective, thrillingly shuttling among the flying debris, always locking the young man's focused and fanatical expression and his powerful core movements in the picture. The whole scene presents a high-contrast rough texture under strong sunlight, and the air is filled with dust and metal particles, creating a chaotic, wild and visually impactful apocalyptic industrial atmosphere.", + "seed": 422, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_4/index_2221_seed_422_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04399800533428788, + "vlm_logprob": -0.04399800533428788, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.04399800533428788, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1814832091331482 + ], + "opposite_logprob_tokens": [ + -1.806483268737793 + ], + "logprob_reward": -0.1814832091331482, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1814832091331482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2499125748872757, + 0.06858114898204803, + 0.056719161570072174, + 0.04585684463381767, + 0.04768538847565651, + 0.031038403511047363, + 0.06328510493040085, + 0.043119292706251144, + 0.23339687287807465, + 0.16040518879890442 + ], + "credit_map": [ + 0.052599769085645676, + 0.09102335572242737, + 0.021989211440086365, + 0.017926795408129692, + 0.0156400203704834, + 0.011114868335425854, + 0.026752324774861336, + 0.012866217643022537, + 0.00931962113827467, + 0.01065837312489748, + 0.007191039156168699, + 0.007361690513789654, + 0.007529168389737606, + 0.0074588521383702755, + 0.013969819992780685, + 0.005092589650303125, + 0.010014540515840054, + 0.008845167234539986, + 0.008330099284648895, + 0.007422002963721752, + 0.007038486190140247, + 0.004366594832390547, + 0.0055953338742256165, + 0.005106935743242502, + 0.00545491324737668, + 0.0059770639054477215, + 0.0051208371296525, + 0.008681737817823887, + 0.005431079771369696, + 0.005956931505352259, + 0.00411163829267025, + 0.00512264110147953, + 0.006135153584182262, + 0.007148220203816891, + 0.005890290252864361, + 0.008977822959423065, + 0.005122424103319645, + 0.0047757248394191265, + 0.004642954561859369, + 0.004992800764739513, + 0.003313044784590602, + 0.004451002925634384, + 0.003582591889426112, + 0.003755471669137478, + 0.004916694946587086, + 0.003674132050946355, + 0.003326562698930502, + 0.004018900915980339, + 0.007988461293280125, + 0.008366348221898079, + 0.007506124675273895, + 0.006326497998088598, + 0.008871952071785927, + 0.007761397864669561, + 0.008289748802781105, + 0.008174577727913857, + 0.005146723240613937, + 0.0049972315318882465, + 0.004260185174643993, + 0.005241110920906067, + 0.005488213151693344, + 0.003726622788235545, + 0.005062426906079054, + 0.009196780622005463, + 0.017226725816726685, + 0.08743001520633698, + 0.013680580072104931, + 0.014049642719328403, + 0.0726860985159874, + 0.011036290787160397, + 0.009136388078331947, + 0.008151127025485039, + 0.009159835986793041, + 0.06604164838790894, + 0.006997755728662014, + 0.006776182446628809, + 0.05359917879104614, + 0.006907350849360228, + 0.005929199513047934, + 0.004994036164134741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.1814832091331482 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005967537872493267 + ], + "opposite_logprob_tokens": [ + -5.130967617034912 + ], + "logprob_reward": -0.005967537872493267, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005967537872493267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3743513524532318, + 0.016125615686178207, + 0.045709267258644104, + 0.02757841907441616, + 0.06448131799697876, + 0.02874763496220112, + 0.016294805333018303, + 0.033665064722299576, + 0.20937538146972656, + 0.18367111682891846 + ], + "credit_map": [ + 0.06732809543609619, + 0.1439714878797531, + 0.06515704095363617, + 0.024004677310585976, + 0.019047189503908157, + 0.01541874185204506, + 0.02211111970245838, + 0.017312992364168167, + 0.0024424453731626272, + 0.00290504889562726, + 0.0016253507928922772, + 0.0015664156526327133, + 0.0018648370169103146, + 0.0016489463159814477, + 0.0027948697097599506, + 0.0012777006486430764, + 0.0049788737669587135, + 0.005203378386795521, + 0.006379825994372368, + 0.007299527060240507, + 0.0070073651149868965, + 0.006008933763951063, + 0.005300966557115316, + 0.003530393820255995, + 0.002419654978439212, + 0.0034034578129649162, + 0.003191852243617177, + 0.004498305730521679, + 0.0034239636734128, + 0.004035810008645058, + 0.0035598394460976124, + 0.0030455351807177067, + 0.00928686372935772, + 0.009649602696299553, + 0.0075034997425973415, + 0.00662188557907939, + 0.006124705541878939, + 0.013212992809712887, + 0.004863949958235025, + 0.007217812351882458, + 0.003300637472420931, + 0.003944909665733576, + 0.004085753578692675, + 0.003475064178928733, + 0.004305220674723387, + 0.003003503195941448, + 0.0032792785204946995, + 0.00335326767526567, + 0.0016667884774506092, + 0.00194824009668082, + 0.0021963410545140505, + 0.001802456215955317, + 0.002406690502539277, + 0.0016017525922507048, + 0.001937281689606607, + 0.002735255751758814, + 0.005131294950842857, + 0.005198074039071798, + 0.0037949082907289267, + 0.0035720199812203646, + 0.00574484933167696, + 0.0025352053344249725, + 0.0044426582753658295, + 0.003246053121984005, + 0.015020417980849743, + 0.07888717949390411, + 0.011985545977950096, + 0.009976697154343128, + 0.06867901980876923, + 0.009844829328358173, + 0.00785053800791502, + 0.007131156045943499, + 0.01460197288542986, + 0.05964581295847893, + 0.03702942654490471, + 0.015913890674710274, + 0.033486682921648026, + 0.006529076024889946, + 0.008539712987840176, + 0.007924539037048817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.005967537872493267 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076280031353235245 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.0076280031353235245, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076280031353235245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1794031411409378, + 0.01763242296874523, + 0.124349445104599, + 0.011381040327250957, + 0.08509381860494614, + 0.019131703302264214, + 0.01406871061772108, + 0.026335394009947777, + 0.1013558879494667, + 0.4212484657764435 + ], + "credit_map": [ + 0.030761029571294785, + 0.05602659285068512, + 0.017647627741098404, + 0.015028833411633968, + 0.01189359836280346, + 0.008956174366176128, + 0.027282925322651863, + 0.011806340888142586, + 0.0019736678805202246, + 0.0022485593799501657, + 0.00255186902359128, + 0.001987659139558673, + 0.0021553533151745796, + 0.002114824950695038, + 0.003021481679752469, + 0.001579006784595549, + 0.013951646164059639, + 0.01458780188113451, + 0.017460310831665993, + 0.020283598452806473, + 0.019214104861021042, + 0.015539096668362617, + 0.014012742787599564, + 0.009300137870013714, + 0.0009906087070703506, + 0.0015384875005111098, + 0.001269233413040638, + 0.0016262736171483994, + 0.0013027454260736704, + 0.0020266275387257338, + 0.000861996493767947, + 0.0017650664085522294, + 0.010480083525180817, + 0.011134415864944458, + 0.012666940689086914, + 0.01712823286652565, + 0.0073010362684726715, + 0.011416934430599213, + 0.0052300612442195415, + 0.009736109524965286, + 0.002557320287451148, + 0.002785139251500368, + 0.0031607509590685368, + 0.0023563215509057045, + 0.002092512557283044, + 0.0019421834731474519, + 0.0022283843718469143, + 0.002009088173508644, + 0.001968957018107176, + 0.001916280249133706, + 0.0022566253319382668, + 0.0015749664744362235, + 0.00204727565869689, + 0.0013963687233626842, + 0.00154189916793257, + 0.0013663360150530934, + 0.00338513171300292, + 0.00342286447994411, + 0.002814409788697958, + 0.003190036164596677, + 0.00347558269277215, + 0.0025056558661162853, + 0.002900900086387992, + 0.004640810191631317, + 0.006472015753388405, + 0.03834548220038414, + 0.0061294883489608765, + 0.004637173842638731, + 0.03311212360858917, + 0.0038884314708411694, + 0.00473286397755146, + 0.004038297105580568, + 0.030025605112314224, + 0.1678197830915451, + 0.05979834869503975, + 0.023655960336327553, + 0.10966178774833679, + 0.008032532408833504, + 0.012641849927604198, + 0.00961257517337799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.0076280031353235245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732400041073561 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732400041073561, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732400041073561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13532212376594543, + 0.041227441281080246, + 0.024295108392834663, + 0.010937762446701527, + 0.026276232674717903, + 0.01520135160535574, + 0.01602434553205967, + 0.015941333025693893, + 0.043466124683618546, + 0.6713082194328308 + ], + "credit_map": [ + 0.031106363981962204, + 0.05337958410382271, + 0.009540844708681107, + 0.008500210009515285, + 0.0067015597596764565, + 0.004966918844729662, + 0.015089496970176697, + 0.006037138402462006, + 0.004252101760357618, + 0.003678674343973398, + 0.0037173673044890165, + 0.004319211468100548, + 0.005099805071949959, + 0.006470407824963331, + 0.010820457711815834, + 0.00286941509693861, + 0.0027931786607950926, + 0.004704263061285019, + 0.0034064012579619884, + 0.0033382342662662268, + 0.003078936133533716, + 0.002482327865436673, + 0.0023701677564531565, + 0.0021215956658124924, + 0.0010541772935539484, + 0.0009392874781042337, + 0.001661870046518743, + 0.0015805696602910757, + 0.0013329009525477886, + 0.002066546818241477, + 0.0008707894012331963, + 0.001431620097719133, + 0.004049423150718212, + 0.00575614720582962, + 0.003368520177900791, + 0.003047615522518754, + 0.002236583735793829, + 0.003165740519762039, + 0.0017407475970685482, + 0.0029114510398358107, + 0.0013275580713525414, + 0.0018975501880049706, + 0.0019153531175106764, + 0.0020154023077338934, + 0.002423027064651251, + 0.001623612828552723, + 0.001663542934693396, + 0.0023353032302111387, + 0.0015111751854419708, + 0.001956371357664466, + 0.0015083916950970888, + 0.0015986911021173, + 0.0025736771058291197, + 0.0019765214528888464, + 0.00222988985478878, + 0.002669626148417592, + 0.0020515676587820053, + 0.0020900305826216936, + 0.0020509532187134027, + 0.0017530914628878236, + 0.002310013398528099, + 0.0014175057876855135, + 0.0016394745325669646, + 0.0026286940556019545, + 0.004386840388178825, + 0.012040112167596817, + 0.005077152047306299, + 0.003665781347081065, + 0.011007473804056644, + 0.0030342864338308573, + 0.0017817538464441895, + 0.002472721040248871, + 0.05578877404332161, + 0.22177034616470337, + 0.13797923922538757, + 0.054042380303144455, + 0.1215343102812767, + 0.02502295933663845, + 0.02620236761868, + 0.028967764228582382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.006732400041073561 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178876489400864 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178876489400864, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178876489400864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09131781756877899, + 0.0271125640720129, + 0.06415940821170807, + 0.004641388077288866, + 0.022364433854818344, + 0.013353921473026276, + 0.026215936988592148, + 0.02176668681204319, + 0.15096202492713928, + 0.5781058073043823 + ], + "credit_map": [ + 0.015232473611831665, + 0.03113517165184021, + 0.007666192948818207, + 0.006737027782946825, + 0.006249648053199053, + 0.004558445420116186, + 0.014311394654214382, + 0.005427471362054348, + 0.0031194102484732866, + 0.002257243962958455, + 0.003164738416671753, + 0.00215915497392416, + 0.0037233619950711727, + 0.0035060876980423927, + 0.00699293939396739, + 0.002189632272347808, + 0.006419481243938208, + 0.0070002139545977116, + 0.009035064838826656, + 0.010366778820753098, + 0.009326782077550888, + 0.009399889968335629, + 0.007451531942933798, + 0.0051596686244010925, + 0.0004354258126113564, + 0.0005856202333234251, + 0.0005909460014663637, + 0.0007865618681535125, + 0.0005434786435216665, + 0.0006238245405256748, + 0.0004727198393084109, + 0.0006028112838976085, + 0.003428627969697118, + 0.0037251096218824387, + 0.0027964934706687927, + 0.002233133651316166, + 0.0023021004162728786, + 0.003999180626124144, + 0.0014292977284640074, + 0.0024504929315298796, + 0.0012166101951152086, + 0.001531795714981854, + 0.0018414409132674336, + 0.0018874339293688536, + 0.001961990725249052, + 0.0016772930976003408, + 0.0016848266823217273, + 0.0015525309136137366, + 0.004209023900330067, + 0.0036777600180357695, + 0.003772979835048318, + 0.0028519874904304743, + 0.003058589994907379, + 0.002362613333389163, + 0.0032745879143476486, + 0.0030083968304097652, + 0.002644461812451482, + 0.002765465062111616, + 0.002119723241776228, + 0.002149055013433099, + 0.0038412658032029867, + 0.0022534518502652645, + 0.002647173358127475, + 0.003346092300489545, + 0.010233516804873943, + 0.06262785196304321, + 0.006838275585323572, + 0.004537625703960657, + 0.054023321717977524, + 0.005028253421187401, + 0.004670321475714445, + 0.0030028684996068478, + 0.0405721589922905, + 0.23904865980148315, + 0.07284460961818695, + 0.02678525820374489, + 0.16192185878753662, + 0.010708906687796116, + 0.015338401310145855, + 0.01088602002710104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.018178876489400864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, high-speed, dynamic tracking shot, starting from a low angle and spiraling upwards in unison with the subject's movements, weaves through a vast, mountainous scrapyard. At the center of the frame, a young man in a faded gray hoodie and oil-stained overalls, wearing heavy industrial boots, clutches a massive, industrial-grade energy weapon. The weapon is ruggedly constructed, crisscrossed with pipes and fins, its core emitting a dazzling blue arc of electricity when activated. With his legs firmly planted on the twisted pile of metal, he leans back to gather power, then, with a dramatic, explosive gesture, slams the weapon forward. A visible, fan-shaped wave of energy erupts from the weapon's front, instantly impacting the mountain of scrap metal ahead. Countless rusted car hulls, twisted rebar, and broken mechanical parts are instantly ripped apart and flung into the air by this invisible force, creating a spectacular storm of metal. The fragments gleam coldly in the glaring sunlight, scattering across the sky. The camera uses a first-person drone-like perspective, thrillingly shuttling among the flying debris, always locking the young man's focused and fanatical expression and his powerful core movements in the picture. The whole scene presents a high-contrast rough texture under strong sunlight, and the air is filled with dust and metal particles, creating a chaotic, wild and visually impactful apocalyptic industrial atmosphere.", + "seed": 423, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_4/index_2221_seed_423_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12143615772947669, + "vlm_logprob": -0.12143615772947669, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.12143615772947669, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5246098041534424 + ], + "opposite_logprob_tokens": [ + -0.8996098041534424 + ], + "logprob_reward": -0.5246098041534424, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5246098041534424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03296046331524849, + 0.027686335146427155, + 0.04230977222323418, + 0.02214653603732586, + 0.7282716035842896, + 0.02792358212172985, + 0.041314467787742615, + 0.024035975337028503, + 0.023200247436761856, + 0.030150996521115303 + ], + "credit_map": [ + 0.0036865323781967163, + 0.005037266761064529, + 0.0031111242715269327, + 0.00308435782790184, + 0.006489390507340431, + 0.00453332532197237, + 0.0037985367234796286, + 0.0032199297565966845, + 0.0028385224286466837, + 0.0032333023846149445, + 0.009573780931532383, + 0.0030484977178275585, + 0.002156761009246111, + 0.0031627165153622627, + 0.0017953580245375633, + 0.0018773990450426936, + 0.005021518096327782, + 0.006327374838292599, + 0.00407944992184639, + 0.006231666076928377, + 0.008744800463318825, + 0.00380459101870656, + 0.00499644223600626, + 0.0031039363238960505, + 0.0032514864578843117, + 0.0038192416541278362, + 0.0015448308549821377, + 0.0023833829909563065, + 0.0033894020598381758, + 0.002656067255884409, + 0.003119564615190029, + 0.001982560846954584, + 0.05116508901119232, + 0.29472047090530396, + 0.013746397569775581, + 0.23142790794372559, + 0.024822598323225975, + 0.05384388193488121, + 0.04628618061542511, + 0.012259121052920818, + 0.0045991698279976845, + 0.004350078757852316, + 0.002720166929066181, + 0.004141134209930897, + 0.004531980492174625, + 0.0024108956567943096, + 0.0031264189165085554, + 0.002043740591034293, + 0.005456680431962013, + 0.0027438343968242407, + 0.010400986298918724, + 0.006001489236950874, + 0.003468381240963936, + 0.003720964305102825, + 0.0035873698070645332, + 0.0059347632341086864, + 0.0033696501050144434, + 0.0030323262326419353, + 0.005889930296689272, + 0.0022584570106118917, + 0.0019550540018826723, + 0.003023922210559249, + 0.002580423839390278, + 0.001926212222315371, + 0.003388503100723028, + 0.004001184366643429, + 0.0022545126266777515, + 0.0025173432659357786, + 0.0025434517301619053, + 0.0031758544500917196, + 0.0024766530841588974, + 0.0028427457436919212, + 0.005279570817947388, + 0.004321484360843897, + 0.002983622718602419, + 0.002661255421116948, + 0.006497948430478573, + 0.0033728922717273235, + 0.00285717542283237, + 0.002177049173042178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.5246098041534424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964338313788176 + ], + "opposite_logprob_tokens": [ + -5.130964279174805 + ], + "logprob_reward": -0.005964338313788176, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964338313788176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10614225268363953, + 0.18083296716213226, + 0.06923326104879379, + 0.040557727217674255, + 0.03876321017742157, + 0.15539349615573883, + 0.039895109832286835, + 0.11442594975233078, + 0.2074059247970581, + 0.04735012724995613 + ], + "credit_map": [ + 0.01945999450981617, + 0.011343903839588165, + 0.008506011217832565, + 0.007037385366857052, + 0.015496794134378433, + 0.018103372305631638, + 0.012349625118076801, + 0.013845153152942657, + 0.01624293625354767, + 0.025354823097586632, + 0.022524597123265266, + 0.01457860879600048, + 0.041527364403009415, + 0.0261989738792181, + 0.01738433912396431, + 0.017021313309669495, + 0.00913005881011486, + 0.009640702977776527, + 0.004821479320526123, + 0.006079588085412979, + 0.015536105260252953, + 0.0062095606699585915, + 0.008289842866361141, + 0.009525910019874573, + 0.004758128896355629, + 0.009359381161630154, + 0.003250034525990486, + 0.004602381959557533, + 0.004442197736352682, + 0.005443034693598747, + 0.0055198827758431435, + 0.00318268034607172, + 0.0033033895306289196, + 0.013670812360942364, + 0.0017364269588142633, + 0.010459930635988712, + 0.003041898598894477, + 0.0027866424061357975, + 0.002393066417425871, + 0.0013710393104702234, + 0.023620668798685074, + 0.023061498999595642, + 0.009810812771320343, + 0.030842246487736702, + 0.027719533070921898, + 0.014310934580862522, + 0.017007697373628616, + 0.009020080789923668, + 0.005777140613645315, + 0.004007210023701191, + 0.00618072459474206, + 0.006421517580747604, + 0.00571390800178051, + 0.0050839330069720745, + 0.004302315879613161, + 0.0024083552416414022, + 0.01707102172076702, + 0.016635768115520477, + 0.012365582399070263, + 0.015086589381098747, + 0.013172926381230354, + 0.015460459515452385, + 0.015715906396508217, + 0.00891768466681242, + 0.02288742922246456, + 0.023501984775066376, + 0.02048386260867119, + 0.028189899399876595, + 0.029183248057961464, + 0.028580524027347565, + 0.03560863807797432, + 0.018970321863889694, + 0.007983354851603508, + 0.00615726737305522, + 0.007286055013537407, + 0.006561223417520523, + 0.006786223966628313, + 0.004199394024908543, + 0.00465159909799695, + 0.0037250048480927944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.005964338313788176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07341061532497406, + 0.17875246703624725, + 0.049445975571870804, + 0.05645401030778885, + 0.042950063943862915, + 0.04787388816475868, + 0.06722155958414078, + 0.038111474364995956, + 0.4029912054538727, + 0.04278874769806862 + ], + "credit_map": [ + 0.010502670891582966, + 0.010457837022840977, + 0.006874114740639925, + 0.006192270200699568, + 0.010282794944941998, + 0.010772612877190113, + 0.007721981965005398, + 0.010606334544718266, + 0.01701773889362812, + 0.021914474666118622, + 0.039823323488235474, + 0.01450959499925375, + 0.030929729342460632, + 0.023818759247660637, + 0.015998659655451775, + 0.014740188606083393, + 0.008044512942433357, + 0.007012148853391409, + 0.005104727111756802, + 0.005399961490184069, + 0.005927840247750282, + 0.005718972999602556, + 0.003903374308720231, + 0.008334439247846603, + 0.007018673699349165, + 0.011106409132480621, + 0.004690573550760746, + 0.006711974274367094, + 0.005854133516550064, + 0.009229627437889576, + 0.007111497223377228, + 0.004731119144707918, + 0.002935514086857438, + 0.016487451270222664, + 0.0011612395755946636, + 0.01278060581535101, + 0.0021282380912452936, + 0.003468879498541355, + 0.003019727533683181, + 0.0009684100514277816, + 0.007352007087320089, + 0.007705820724368095, + 0.002749864710494876, + 0.008765259757637978, + 0.008457153104245663, + 0.004470008425414562, + 0.005468118470162153, + 0.002905655652284622, + 0.00924075860530138, + 0.005262462422251701, + 0.010895777493715286, + 0.011594878509640694, + 0.006369190756231546, + 0.00679742218926549, + 0.006689123343676329, + 0.010371943935751915, + 0.006330430041998625, + 0.005520136095583439, + 0.0060137733817100525, + 0.004451211076229811, + 0.004301905632019043, + 0.005047299433499575, + 0.0038272447418421507, + 0.0026194732636213303, + 0.044115979224443436, + 0.04639528691768646, + 0.04022849723696709, + 0.05005979537963867, + 0.05649366229772568, + 0.05740776285529137, + 0.0702468603849411, + 0.03804336488246918, + 0.007552051916718483, + 0.0062466999515891075, + 0.006024786736816168, + 0.006049841642379761, + 0.006380865350365639, + 0.0039466191083192825, + 0.0036359564401209354, + 0.002951929112896323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.008636740036308765 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943245254456997 + ], + "opposite_logprob_tokens": [ + -5.130943298339844 + ], + "logprob_reward": -0.005943245254456997, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943245254456997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06056729331612587, + 0.034053217619657516, + 0.05412814021110535, + 0.06688877195119858, + 0.39917099475860596, + 0.02017616294324398, + 0.02976064383983612, + 0.023239700123667717, + 0.23138578236103058, + 0.08062923699617386 + ], + "credit_map": [ + 0.008720090612769127, + 0.009772025048732758, + 0.006847242824733257, + 0.004337027668952942, + 0.004236060660332441, + 0.013527455739676952, + 0.005191618110984564, + 0.007935783825814724, + 0.0038123824633657932, + 0.004867111332714558, + 0.007272427435964346, + 0.003171468386426568, + 0.004643288441002369, + 0.004413721151649952, + 0.0028520545456558466, + 0.003020771313458681, + 0.007340956479310989, + 0.005016515031456947, + 0.007303636986762285, + 0.0069687520153820515, + 0.006775791756808758, + 0.007296864874660969, + 0.007816417142748833, + 0.005609212908893824, + 0.00800428818911314, + 0.012642627581954002, + 0.006827556993812323, + 0.007784606888890266, + 0.006637220270931721, + 0.010014321655035019, + 0.008589057251811028, + 0.0063890935853123665, + 0.028827987611293793, + 0.16132429242134094, + 0.007329307962208986, + 0.12582655251026154, + 0.013929874636232853, + 0.029896866530179977, + 0.025093430653214455, + 0.006942732259631157, + 0.002835770370438695, + 0.003359707770869136, + 0.002623273292556405, + 0.0033020339906215668, + 0.0028295048978179693, + 0.0016147716669365764, + 0.0018074255203828216, + 0.0018036763649433851, + 0.003890338120982051, + 0.0024425687734037638, + 0.005583633668720722, + 0.0035515164490789175, + 0.0024985945783555508, + 0.0029700309969484806, + 0.002744843252003193, + 0.0060791210271418095, + 0.0037095199804753065, + 0.0028432852122932673, + 0.0050561316311359406, + 0.0018571738619357347, + 0.002396670635789633, + 0.00336335226893425, + 0.0019321092404425144, + 0.0020814603194594383, + 0.03526753932237625, + 0.03817538172006607, + 0.030708305537700653, + 0.02384823001921177, + 0.03144611045718193, + 0.03294267877936363, + 0.02028643898665905, + 0.018711131066083908, + 0.017313141375780106, + 0.015412013977766037, + 0.011387825012207031, + 0.009967590682208538, + 0.007531614508479834, + 0.008437068201601505, + 0.005616762675344944, + 0.004963221959769726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.005943245254456997 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202666088938713 + ], + "opposite_logprob_tokens": [ + -2.8120267391204834 + ], + "logprob_reward": -0.06202666088938713, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202666088938713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10012519359588623, + 0.13660916686058044, + 0.1298496276140213, + 0.0618053562939167, + 0.2795613408088684, + 0.036551717668771744, + 0.041500601917505264, + 0.03720169514417648, + 0.09989892691373825, + 0.07689634710550308 + ], + "credit_map": [ + 0.016339385882019997, + 0.012527238577604294, + 0.00897202454507351, + 0.009342189878225327, + 0.014375214464962482, + 0.013601932674646378, + 0.011955653317272663, + 0.013011555187404156, + 0.017366928979754448, + 0.014534221962094307, + 0.04295370355248451, + 0.013753426261246204, + 0.011193973943591118, + 0.01690848544239998, + 0.00929534062743187, + 0.01060307677835226, + 0.013643495738506317, + 0.007196684833616018, + 0.008415243588387966, + 0.036557912826538086, + 0.01102925930172205, + 0.0320226289331913, + 0.014018814079463482, + 0.006965584121644497, + 0.00637779338285327, + 0.011384706944227219, + 0.006265800911933184, + 0.005796179641038179, + 0.008692536503076553, + 0.01001790538430214, + 0.009052326902747154, + 0.004218106623739004, + 0.02015903778374195, + 0.10983768105506897, + 0.0062436191365122795, + 0.08710043132305145, + 0.01050676591694355, + 0.02087482251226902, + 0.018898703157901764, + 0.005940274801105261, + 0.005284366197884083, + 0.006613995414227247, + 0.002997404895722866, + 0.006633751560002565, + 0.004142304882407188, + 0.004034969490021467, + 0.004060495179146528, + 0.0027844335418194532, + 0.005179722327739, + 0.004386496730148792, + 0.007865183986723423, + 0.006992599461227655, + 0.003747158218175173, + 0.004242512863129377, + 0.005203800275921822, + 0.0038831259589642286, + 0.005608081351965666, + 0.005013352259993553, + 0.009305648505687714, + 0.003189622424542904, + 0.0036980186123400927, + 0.004852499812841415, + 0.002418958581984043, + 0.0031155142933130264, + 0.01728763058781624, + 0.012990938499569893, + 0.010023174807429314, + 0.010637419298291206, + 0.011798259802162647, + 0.013237266801297665, + 0.009977943263947964, + 0.013946295715868473, + 0.01669936254620552, + 0.011389602907001972, + 0.012280041351914406, + 0.011420042254030704, + 0.007505676243454218, + 0.006934854201972485, + 0.006093043368309736, + 0.004573717247694731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.06202666088938713 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, high-speed, dynamic tracking shot, starting from a low angle and spiraling upwards in unison with the subject's movements, weaves through a vast, mountainous scrapyard. At the center of the frame, a young man in a faded gray hoodie and oil-stained overalls, wearing heavy industrial boots, clutches a massive, industrial-grade energy weapon. The weapon is ruggedly constructed, crisscrossed with pipes and fins, its core emitting a dazzling blue arc of electricity when activated. With his legs firmly planted on the twisted pile of metal, he leans back to gather power, then, with a dramatic, explosive gesture, slams the weapon forward. A visible, fan-shaped wave of energy erupts from the weapon's front, instantly impacting the mountain of scrap metal ahead. Countless rusted car hulls, twisted rebar, and broken mechanical parts are instantly ripped apart and flung into the air by this invisible force, creating a spectacular storm of metal. The fragments gleam coldly in the glaring sunlight, scattering across the sky. The camera uses a first-person drone-like perspective, thrillingly shuttling among the flying debris, always locking the young man's focused and fanatical expression and his powerful core movements in the picture. The whole scene presents a high-contrast rough texture under strong sunlight, and the air is filled with dust and metal particles, creating a chaotic, wild and visually impactful apocalyptic industrial atmosphere.", + "seed": 424, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_4/index_2221_seed_424_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03792701680213213, + "vlm_logprob": -0.03792701680213213, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03792701680213213, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12831895053386688 + ], + "opposite_logprob_tokens": [ + -2.128319025039673 + ], + "logprob_reward": -0.12831895053386688, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12831895053386688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12983669340610504, + 0.20671933889389038, + 0.08373340964317322, + 0.03311244770884514, + 0.03441360592842102, + 0.04527414217591286, + 0.07268936932086945, + 0.2306065708398819, + 0.08537998050451279, + 0.07823442667722702 + ], + "credit_map": [ + 0.02865997888147831, + 0.013824564404785633, + 0.019732369109988213, + 0.01203187182545662, + 0.016137707978487015, + 0.011111892759799957, + 0.018246907740831375, + 0.010091407224535942, + 0.021917639300227165, + 0.01868998073041439, + 0.030401242896914482, + 0.02498268149793148, + 0.034824810922145844, + 0.021090947091579437, + 0.03561561927199364, + 0.019196413457393646, + 0.023565908893942833, + 0.01086578331887722, + 0.008089171722531319, + 0.009411314502358437, + 0.009983467869460583, + 0.006420724093914032, + 0.010647466406226158, + 0.004749569110572338, + 0.0035746826324611902, + 0.0031551122665405273, + 0.0035301796160638332, + 0.0034622640814632177, + 0.0056739263236522675, + 0.004827702417969704, + 0.004338448867201805, + 0.004550132900476456, + 0.004147135186940432, + 0.004030797630548477, + 0.003235150594264269, + 0.003819923847913742, + 0.004850831348448992, + 0.005364452488720417, + 0.00369099504314363, + 0.005274318624287844, + 0.004145611077547073, + 0.008143552578985691, + 0.0036307622212916613, + 0.004941610619425774, + 0.006219396833330393, + 0.010389750823378563, + 0.004054561723023653, + 0.0037488944362848997, + 0.008906506933271885, + 0.009512397460639477, + 0.012921838089823723, + 0.008757566101849079, + 0.010560842230916023, + 0.008748835884034634, + 0.007092641666531563, + 0.006188744679093361, + 0.028131995350122452, + 0.08258223533630371, + 0.010896196588873863, + 0.009337948635220528, + 0.019098909571766853, + 0.017214829102158546, + 0.01600227691233158, + 0.04734218120574951, + 0.019867608323693275, + 0.01972276158630848, + 0.014399939216673374, + 0.004771921318024397, + 0.007787176873534918, + 0.00665618572384119, + 0.0052746254950761795, + 0.0068997666239738464, + 0.009963970631361008, + 0.011526950635015965, + 0.00877715926617384, + 0.005216390360146761, + 0.015284902416169643, + 0.0089390454813838, + 0.008811604231595993, + 0.009714409708976746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.12831895053386688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005966945551335812 + ], + "opposite_logprob_tokens": [ + -5.130967140197754 + ], + "logprob_reward": -0.005966945551335812, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005966945551335812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.077578105032444, + 0.17114877700805664, + 0.10168338567018509, + 0.09567707031965256, + 0.14138337969779968, + 0.05009153112769127, + 0.06802300363779068, + 0.06581179797649384, + 0.11518921703100204, + 0.11341378837823868 + ], + "credit_map": [ + 0.012881511822342873, + 0.007963018491864204, + 0.012916015461087227, + 0.008145001716911793, + 0.008110587485134602, + 0.007482566405087709, + 0.011516604572534561, + 0.008562787435948849, + 0.018790120258927345, + 0.017902199178934097, + 0.028451286256313324, + 0.022509025409817696, + 0.03102339617908001, + 0.0188723336905241, + 0.023140747100114822, + 0.0104596596211195, + 0.02239081636071205, + 0.011963557451963425, + 0.011755396611988544, + 0.012408554553985596, + 0.010412133298814297, + 0.011041615158319473, + 0.01295868493616581, + 0.008752617053687572, + 0.011369060724973679, + 0.011084629222750664, + 0.011972177773714066, + 0.012064578011631966, + 0.013991397805511951, + 0.010208080522716045, + 0.013385118916630745, + 0.011602014303207397, + 0.03288581594824791, + 0.015078938566148281, + 0.010372593067586422, + 0.008308935910463333, + 0.020297782495617867, + 0.0379873663187027, + 0.00749243563041091, + 0.008959508500993252, + 0.005735605955123901, + 0.007950905710458755, + 0.006749392952769995, + 0.0059479097835719585, + 0.005143608897924423, + 0.010556784458458424, + 0.0034964382648468018, + 0.004510883241891861, + 0.007267484441399574, + 0.010521337389945984, + 0.006782095413655043, + 0.007340462412685156, + 0.008229950442910194, + 0.007553781848400831, + 0.01054074801504612, + 0.009787144139409065, + 0.011502563953399658, + 0.012875263579189777, + 0.0055825961753726006, + 0.00351255526766181, + 0.008245078846812248, + 0.0054383897222578526, + 0.010778339579701424, + 0.007877005264163017, + 0.02310783788561821, + 0.02188291773200035, + 0.014845927245914936, + 0.00683332234621048, + 0.014852255582809448, + 0.01169857569038868, + 0.009769773110747337, + 0.01219860278069973, + 0.012381864711642265, + 0.017871160060167313, + 0.011913144029676914, + 0.010856696404516697, + 0.019022060558199883, + 0.015429203398525715, + 0.011463161557912827, + 0.01447648461908102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.005966945551335812 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636386133730412 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636386133730412, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636386133730412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08652561157941818, + 0.09116309136152267, + 0.05487382039427757, + 0.08966648578643799, + 0.040078867226839066, + 0.06474075466394424, + 0.11059710383415222, + 0.13039728999137878, + 0.06297535449266434, + 0.2689816951751709 + ], + "credit_map": [ + 0.0168446097522974, + 0.012136150151491165, + 0.0126626156270504, + 0.008945944719016552, + 0.009255806915462017, + 0.007467571645975113, + 0.010819802060723305, + 0.008393100462853909, + 0.008556325919926167, + 0.010161906480789185, + 0.01591041497886181, + 0.011140123009681702, + 0.01440373994410038, + 0.009001540020108223, + 0.01513933390378952, + 0.006849699188023806, + 0.013301506638526917, + 0.0057741026394069195, + 0.0050396546721458435, + 0.006212698295712471, + 0.006935587152838707, + 0.006334477104246616, + 0.006784310098737478, + 0.004491479601711035, + 0.010577993467450142, + 0.00919827725738287, + 0.010560615919530392, + 0.011724594980478287, + 0.012557963840663433, + 0.01027853973209858, + 0.013086045160889626, + 0.011682447046041489, + 0.0059622772969305515, + 0.0034318133257329464, + 0.00455472432076931, + 0.004235194530338049, + 0.005340806674212217, + 0.0064967116340994835, + 0.005923927295953035, + 0.004133408889174461, + 0.007350768893957138, + 0.008546607568860054, + 0.0145486518740654, + 0.011081009171903133, + 0.006473443936556578, + 0.006907325703650713, + 0.0045676203444600105, + 0.0052653211168944836, + 0.012916576117277145, + 0.01303964201360941, + 0.015598876401782036, + 0.009652023203670979, + 0.01379144936800003, + 0.01381122786551714, + 0.017549313604831696, + 0.014237984083592892, + 0.01701233722269535, + 0.03499683737754822, + 0.012346888892352581, + 0.014296196401119232, + 0.012790818698704243, + 0.014538002200424671, + 0.010881075635552406, + 0.013535127975046635, + 0.010816537775099277, + 0.013332960195839405, + 0.00709777744486928, + 0.004508695099502802, + 0.00711212819442153, + 0.006432805210351944, + 0.005765906069427729, + 0.007908540777862072, + 0.03400418907403946, + 0.0468570776283741, + 0.025158192962408066, + 0.024288009852170944, + 0.04730020835995674, + 0.04295653477311134, + 0.027173997834324837, + 0.021243467926979065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.008636386133730412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632485754787922 + ], + "opposite_logprob_tokens": [ + -4.758632659912109 + ], + "logprob_reward": -0.008632485754787922, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632485754787922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.094463050365448, + 0.07563064992427826, + 0.0486551858484745, + 0.10417157411575317, + 0.18801291286945343, + 0.1684630960226059, + 0.06450110673904419, + 0.09344400465488434, + 0.051188502460718155, + 0.11146993190050125 + ], + "credit_map": [ + 0.008816385641694069, + 0.00804510060697794, + 0.01416691392660141, + 0.01621817611157894, + 0.013251558877527714, + 0.010179014876484871, + 0.015311945229768753, + 0.008473956026136875, + 0.00835812371224165, + 0.007353436201810837, + 0.015317434445023537, + 0.009874409064650536, + 0.011073993518948555, + 0.006859940011054277, + 0.011593650095164776, + 0.00519966147840023, + 0.010993307456374168, + 0.00521504832431674, + 0.004198937211185694, + 0.005131502635776997, + 0.007048754021525383, + 0.004918248858302832, + 0.007436798885464668, + 0.0037125926464796066, + 0.012924766167998314, + 0.013731452636420727, + 0.011978489346802235, + 0.013676755130290985, + 0.019055016338825226, + 0.011368623934686184, + 0.012236791662871838, + 0.00919968169182539, + 0.05025812238454819, + 0.0172731913626194, + 0.011314089410007, + 0.009491553530097008, + 0.025679180398583412, + 0.05586565285921097, + 0.009390058927237988, + 0.008741075173020363, + 0.021532339975237846, + 0.017708584666252136, + 0.04520381987094879, + 0.03148097172379494, + 0.01821199245750904, + 0.01465645432472229, + 0.009128455072641373, + 0.010540477931499481, + 0.006272846832871437, + 0.01387293916195631, + 0.0052232625894248486, + 0.005320527590811253, + 0.010919180698692799, + 0.006986735854297876, + 0.008494541980326176, + 0.007411074358969927, + 0.014612605795264244, + 0.030046477913856506, + 0.006010277196764946, + 0.0055616446770727634, + 0.007907569408416748, + 0.00703959446400404, + 0.006547973491251469, + 0.015717852860689163, + 0.008388672024011612, + 0.007997562177479267, + 0.007276913616806269, + 0.004110605921596289, + 0.007392414379864931, + 0.005468766670674086, + 0.004464165307581425, + 0.006089407484978437, + 0.012363418005406857, + 0.014170167967677116, + 0.010838943533599377, + 0.007615766022354364, + 0.024025579914450645, + 0.016929931938648224, + 0.012668664567172527, + 0.01285746693611145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.008632485754787922 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808031603693962 + ], + "opposite_logprob_tokens": [ + -3.2880802154541016 + ], + "logprob_reward": -0.03808031603693962, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808031603693962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0885181576013565, + 0.03485112264752388, + 0.046905968338251114, + 0.07583983987569809, + 0.1248423382639885, + 0.09222151339054108, + 0.08567731827497482, + 0.21668478846549988, + 0.14547155797481537, + 0.08898739516735077 + ], + "credit_map": [ + 0.020686810836195946, + 0.012929528020322323, + 0.011795608326792717, + 0.007944121956825256, + 0.0085761658847332, + 0.007239087484776974, + 0.010950715281069279, + 0.008396121673285961, + 0.0041719647124409676, + 0.003382100025191903, + 0.006773373112082481, + 0.00348289986141026, + 0.004883230663836002, + 0.003279797499999404, + 0.005222732666879892, + 0.003655023407191038, + 0.0075819301418960094, + 0.004940155893564224, + 0.003947672434151173, + 0.006268157623708248, + 0.0076700798235833645, + 0.005434843245893717, + 0.007813422940671444, + 0.003249706234782934, + 0.008061741478741169, + 0.0073333559557795525, + 0.0094516072422266, + 0.008495712652802467, + 0.015123998746275902, + 0.009154930710792542, + 0.00917877722531557, + 0.009039713069796562, + 0.030685722827911377, + 0.012038392946124077, + 0.008477865718305111, + 0.009305927902460098, + 0.013917158357799053, + 0.03581685572862625, + 0.008160457015037537, + 0.006439946126192808, + 0.011437410488724709, + 0.012661175802350044, + 0.019274050369858742, + 0.015570471994578838, + 0.014083381742238998, + 0.009237460792064667, + 0.004452359862625599, + 0.00550520233809948, + 0.009498621337115765, + 0.015388070605695248, + 0.01127719134092331, + 0.008466014638543129, + 0.012156732380390167, + 0.007929205894470215, + 0.012705995701253414, + 0.00825548730790615, + 0.028197303414344788, + 0.08365046232938766, + 0.009886182844638824, + 0.011289946734905243, + 0.01677883230149746, + 0.013961379416286945, + 0.010753192007541656, + 0.042167484760284424, + 0.02217092737555504, + 0.02898121438920498, + 0.018039679154753685, + 0.013570819981396198, + 0.017868604511022568, + 0.014222778379917145, + 0.01399005763232708, + 0.016627470031380653, + 0.017257599160075188, + 0.016053101047873497, + 0.011196129955351353, + 0.0073999157175421715, + 0.011381149291992188, + 0.009510793723165989, + 0.008513051085174084, + 0.007675657514482737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.03808031603693962 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, high-speed, dynamic tracking shot, starting from a low angle and spiraling upwards in unison with the subject's movements, weaves through a vast, mountainous scrapyard. At the center of the frame, a young man in a faded gray hoodie and oil-stained overalls, wearing heavy industrial boots, clutches a massive, industrial-grade energy weapon. The weapon is ruggedly constructed, crisscrossed with pipes and fins, its core emitting a dazzling blue arc of electricity when activated. With his legs firmly planted on the twisted pile of metal, he leans back to gather power, then, with a dramatic, explosive gesture, slams the weapon forward. A visible, fan-shaped wave of energy erupts from the weapon's front, instantly impacting the mountain of scrap metal ahead. Countless rusted car hulls, twisted rebar, and broken mechanical parts are instantly ripped apart and flung into the air by this invisible force, creating a spectacular storm of metal. The fragments gleam coldly in the glaring sunlight, scattering across the sky. The camera uses a first-person drone-like perspective, thrillingly shuttling among the flying debris, always locking the young man's focused and fanatical expression and his powerful core movements in the picture. The whole scene presents a high-contrast rough texture under strong sunlight, and the air is filled with dust and metal particles, creating a chaotic, wild and visually impactful apocalyptic industrial atmosphere.", + "seed": 425, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_4/index_2221_seed_425_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024118355149403213, + "vlm_logprob": -0.024118355149403213, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.024118355149403213, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09002046287059784 + ], + "opposite_logprob_tokens": [ + -2.4650204181671143 + ], + "logprob_reward": -0.09002046287059784, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09002046287059784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14354415237903595, + 0.061926208436489105, + 0.12127254903316498, + 0.17700032889842987, + 0.0403803214430809, + 0.04532444104552269, + 0.06596839427947998, + 0.1368705928325653, + 0.1007077768445015, + 0.10700520873069763 + ], + "credit_map": [ + 0.017650702968239784, + 0.018930457532405853, + 0.019304612651467323, + 0.0212966687977314, + 0.011912682093679905, + 0.021410716697573662, + 0.018072428181767464, + 0.014965889044106007, + 0.006493631284683943, + 0.008011913858354092, + 0.007971297949552536, + 0.011865876615047455, + 0.009318121708929539, + 0.006428974214941263, + 0.007491739466786385, + 0.004344649612903595, + 0.009326759725809097, + 0.01590132899582386, + 0.008977754041552544, + 0.017852982506155968, + 0.02202189899981022, + 0.01532249990850687, + 0.015201851725578308, + 0.016667479649186134, + 0.017796410247683525, + 0.02836911752820015, + 0.01675737090408802, + 0.020632540807127953, + 0.03129304200410843, + 0.01861196756362915, + 0.021760081872344017, + 0.02177978865802288, + 0.006220778916031122, + 0.00505399564281106, + 0.004261539783328772, + 0.004894035402685404, + 0.006578319706022739, + 0.0035346236545592546, + 0.0055588711984455585, + 0.004278156440705061, + 0.009574887342751026, + 0.004399725701659918, + 0.006162446457892656, + 0.005191291216760874, + 0.005395797081291676, + 0.004454744514077902, + 0.005783807020634413, + 0.004361744038760662, + 0.006412717048078775, + 0.008133742958307266, + 0.00707031786441803, + 0.010092363692820072, + 0.009796367958188057, + 0.008381318300962448, + 0.007091099862009287, + 0.008990462869405746, + 0.012097839266061783, + 0.01570283994078636, + 0.015144075267016888, + 0.021851178258657455, + 0.013832462951540947, + 0.014141320250928402, + 0.024778230115771294, + 0.019322652369737625, + 0.009956901893019676, + 0.017371268942952156, + 0.00881541520357132, + 0.012154299765825272, + 0.01213542278856039, + 0.01786588691174984, + 0.00963745266199112, + 0.012771128676831722, + 0.009345490485429764, + 0.013344762846827507, + 0.009419240988790989, + 0.026141682639718056, + 0.013614782132208347, + 0.013066638261079788, + 0.013712293468415737, + 0.008360316045582294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.09002046287059784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003625251119956374 + ], + "opposite_logprob_tokens": [ + -5.628625392913818 + ], + "logprob_reward": -0.003625251119956374, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003625251119956374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09705976396799088, + 0.06698253750801086, + 0.06260918825864792, + 0.2813417911529541, + 0.10839886218309402, + 0.04251862317323685, + 0.10505443811416626, + 0.12185750901699066, + 0.06211262196302414, + 0.052064601331949234 + ], + "credit_map": [ + 0.012882724404335022, + 0.011939785443246365, + 0.014601747505366802, + 0.014255791902542114, + 0.006722903810441494, + 0.013190110214054585, + 0.012735926546156406, + 0.010730774141848087, + 0.007945992983877659, + 0.008013559505343437, + 0.009135136380791664, + 0.011864998377859592, + 0.008911354467272758, + 0.006990743800997734, + 0.008834880776703358, + 0.00528587494045496, + 0.008573572151362896, + 0.006495608016848564, + 0.005713664926588535, + 0.009229723364114761, + 0.008147385902702808, + 0.01099633239209652, + 0.007271702401340008, + 0.006181198172271252, + 0.02687063254415989, + 0.025382718071341515, + 0.023224052041769028, + 0.08940084278583527, + 0.025526683777570724, + 0.04490477591753006, + 0.029193779453635216, + 0.016838299110531807, + 0.01883101463317871, + 0.03316741809248924, + 0.011407487094402313, + 0.01172459777444601, + 0.009367072023451328, + 0.007960811257362366, + 0.008376050740480423, + 0.007564408704638481, + 0.007209838833659887, + 0.004301041830331087, + 0.0049491263926029205, + 0.004902909975498915, + 0.007792866323143244, + 0.004501023329794407, + 0.004716937430202961, + 0.004144877195358276, + 0.009645112790167332, + 0.009417900815606117, + 0.012433678843080997, + 0.01071863528341055, + 0.017018409445881844, + 0.013438427820801735, + 0.019892461597919464, + 0.012489809654653072, + 0.01013488695025444, + 0.010866709053516388, + 0.012740001082420349, + 0.016288505867123604, + 0.015833092853426933, + 0.013754584826529026, + 0.021022005006670952, + 0.021217722445726395, + 0.008196321316063404, + 0.010968002490699291, + 0.00441909721121192, + 0.0050010960549116135, + 0.007971826009452343, + 0.010620507411658764, + 0.009676840156316757, + 0.00525892898440361, + 0.0037792755756527185, + 0.009513655677437782, + 0.0032590089831501245, + 0.00601576454937458, + 0.008352670818567276, + 0.010595651343464851, + 0.005464805290102959, + 0.005083770025521517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.003625251119956374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190429415553808 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.003190429415553808, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190429415553808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1004730835556984, + 0.04402035102248192, + 0.05051358416676521, + 0.11113918572664261, + 0.09286296367645264, + 0.03567586466670036, + 0.18201591074466705, + 0.16403183341026306, + 0.17819340527057648, + 0.04107385501265526 + ], + "credit_map": [ + 0.013839936815202236, + 0.013892658054828644, + 0.013990480452775955, + 0.0118410037830472, + 0.006565284449607134, + 0.01477611344307661, + 0.012842354364693165, + 0.01272524707019329, + 0.004724101163446903, + 0.005946550518274307, + 0.006394037511199713, + 0.005331207532435656, + 0.00523762172088027, + 0.004761986434459686, + 0.006327473092824221, + 0.0052973744459450245, + 0.0067552304826676846, + 0.005243516527116299, + 0.006154399365186691, + 0.00671724509447813, + 0.005265884101390839, + 0.007270305883139372, + 0.008412245661020279, + 0.004694760777056217, + 0.015461769886314869, + 0.01219246070832014, + 0.009197363629937172, + 0.02828168123960495, + 0.011271937750279903, + 0.013107274658977985, + 0.014435703866183758, + 0.00719098886474967, + 0.01625613309442997, + 0.02917919121682644, + 0.009377828799188137, + 0.01024986058473587, + 0.008270650170743465, + 0.006583587732166052, + 0.00642524054273963, + 0.0065204668790102005, + 0.00417630560696125, + 0.0043365745805203915, + 0.005080890841782093, + 0.0059327431954443455, + 0.003681400790810585, + 0.0028855972923338413, + 0.004596204496920109, + 0.004986146464943886, + 0.018776509910821915, + 0.021851498633623123, + 0.021166343241930008, + 0.014626710675656796, + 0.027960749343037605, + 0.024141579866409302, + 0.033928535878658295, + 0.01956396922469139, + 0.0150242755189538, + 0.02110789529979229, + 0.02032262273132801, + 0.01898171380162239, + 0.02050204388797283, + 0.01999448984861374, + 0.03151959180831909, + 0.016579212620854378, + 0.025054389610886574, + 0.02787623181939125, + 0.015259145759046078, + 0.017572661861777306, + 0.021972395479679108, + 0.02626173011958599, + 0.02216425910592079, + 0.02203257754445076, + 0.003198616672307253, + 0.00618456257507205, + 0.003045712597668171, + 0.004791952669620514, + 0.0066093928180634975, + 0.010340690612792969, + 0.003360372968018055, + 0.003542557591572404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.003190429415553808 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031873397529125214 + ], + "opposite_logprob_tokens": [ + -5.75318717956543 + ], + "logprob_reward": -0.0031873397529125214, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031873397529125214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13606032729148865, + 0.05322933942079544, + 0.03396676108241081, + 0.14421266317367554, + 0.06966190785169601, + 0.04608050361275673, + 0.09219831973314285, + 0.08148901909589767, + 0.26635727286338806, + 0.07674384117126465 + ], + "credit_map": [ + 0.016840867698192596, + 0.019192572683095932, + 0.019650207832455635, + 0.02238033339381218, + 0.00856701284646988, + 0.01738583855330944, + 0.014105084352195263, + 0.017938416451215744, + 0.005137193948030472, + 0.010528765618801117, + 0.005591932218521833, + 0.006775547284632921, + 0.005236612632870674, + 0.008808230981230736, + 0.005984988994896412, + 0.00516606867313385, + 0.006030193064361811, + 0.005022596567869186, + 0.0033995008561760187, + 0.003779700957238674, + 0.003885860089212656, + 0.004671507980674505, + 0.003804016625508666, + 0.0033733851741999388, + 0.01504971832036972, + 0.012492514215409756, + 0.012245621532201767, + 0.05096551775932312, + 0.008038694970309734, + 0.022352201864123344, + 0.013824690133333206, + 0.00924371462315321, + 0.011203764006495476, + 0.020175589248538017, + 0.00817894097417593, + 0.007498414721339941, + 0.007600888609886169, + 0.004929972812533379, + 0.005766986403614283, + 0.0043073492124676704, + 0.006529945880174637, + 0.005744503811001778, + 0.006966469809412956, + 0.0056434073485434055, + 0.005151562858372927, + 0.004048624075949192, + 0.0060318647883832455, + 0.005964123643934727, + 0.00921113695949316, + 0.008653263561427593, + 0.012410569004714489, + 0.013583066873252392, + 0.013613492250442505, + 0.010454987175762653, + 0.014107259921729565, + 0.010164542123675346, + 0.007107363548129797, + 0.007643474265933037, + 0.010925798676908016, + 0.009317928925156593, + 0.010238147340714931, + 0.009047439321875572, + 0.01324600912630558, + 0.013962856493890285, + 0.02480364218354225, + 0.040875550359487534, + 0.03218083456158638, + 0.04908602312207222, + 0.03045736439526081, + 0.03606110438704491, + 0.030956925824284554, + 0.021935826167464256, + 0.007278346456587315, + 0.009021345525979996, + 0.004660440143197775, + 0.013112412765622139, + 0.009442184120416641, + 0.012113967910408974, + 0.014468193985521793, + 0.006646950263530016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.0031873397529125214 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a young man wearing a faded gray hoodie and oil-stained overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man holding a massive industrial-grade energy weapon with pipes, fins, and a core emitting a blue electric arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man slam the weapon forward to create a fan-shaped wave of energy that rips apart rusted car hulls and metal debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a vast mountainous scrapyard filled with rusted car hulls, twisted rebar, and broken mechanical parts under glaring sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a first-person drone-like perspective that spirals upwards while tracking the subject amidst flying debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02056829258799553 + ], + "opposite_logprob_tokens": [ + -3.895568370819092 + ], + "logprob_reward": -0.02056829258799553, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02056829258799553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12430017441511154, + 0.04072793945670128, + 0.045499421656131744, + 0.07684360444545746, + 0.026547517627477646, + 0.11072926223278046, + 0.07882419973611832, + 0.16134068369865417, + 0.28510817885398865, + 0.050079040229320526 + ], + "credit_map": [ + 0.01446146983653307, + 0.011260736733675003, + 0.024198131635785103, + 0.012943356297910213, + 0.009536169469356537, + 0.021500756964087486, + 0.011549217626452446, + 0.018850333988666534, + 0.004643975757062435, + 0.00847578328102827, + 0.004681549966335297, + 0.004590210504829884, + 0.003751707961782813, + 0.0060158902779221535, + 0.004913816228508949, + 0.003655005944892764, + 0.005318090785294771, + 0.005631694104522467, + 0.0036394738126546144, + 0.004981129430234432, + 0.00916166789829731, + 0.006415122654289007, + 0.006162615027278662, + 0.004189623985439539, + 0.006732166279107332, + 0.010415636003017426, + 0.005695938132703304, + 0.016285737976431847, + 0.009494214318692684, + 0.010280872695147991, + 0.01163579523563385, + 0.006303239148110151, + 0.003950350917875767, + 0.0073289223946630955, + 0.002895686309784651, + 0.002673177281394601, + 0.0030113710090517998, + 0.0016741015715524554, + 0.002940781880170107, + 0.0020731270778924227, + 0.012638963758945465, + 0.017428098246455193, + 0.012058411724865437, + 0.019486399367451668, + 0.008968676440417767, + 0.008297910913825035, + 0.0146928196772933, + 0.01715797930955887, + 0.007801194209605455, + 0.007993007078766823, + 0.010631402023136616, + 0.010262930765748024, + 0.011003506369888783, + 0.0094754658639431, + 0.01289596688002348, + 0.008760727941989899, + 0.014372747391462326, + 0.022453848272562027, + 0.015192818827927113, + 0.01775362901389599, + 0.021760540083050728, + 0.019135503098368645, + 0.03526074066758156, + 0.015410850755870342, + 0.02685445547103882, + 0.04321267455816269, + 0.03112664632499218, + 0.04672544077038765, + 0.03375992551445961, + 0.03871914744377136, + 0.03796369582414627, + 0.026746174320578575, + 0.003737478284165263, + 0.00578720960766077, + 0.004216382745653391, + 0.0067547098733484745, + 0.007751667406409979, + 0.011348399333655834, + 0.005765734240412712, + 0.004717457573860884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.02056829258799553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_5/index_tensor([5562, 5562])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_5/index_tensor([5562, 5562])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eb67e644002e902dc87c7ab04f7bbbcdf481ac65 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_5/index_tensor([5562, 5562])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle shot, the frame trembling violently, captures the tragic final moments of a magnificent, ancient, white-walled, black-tiled Hui-style building. Against a backdrop of a hazy, suspended dusty sky, the building's mottled white walls and towering horse-head walls appear remarkably solemn. Suddenly, with a resounding, muffled roar from the ground, the building's foundations begin to tremble violently. The thick, dark brown wooden beams crackle with a piercing sound under the overwhelming pressure, instantly twisting and snapping. Immediately afterwards, the layers of dark gray tiles on the roof lose their grip and are hurled in all directions like black hail, hurled in a destructive force, leaving countless shattered paths in the air. After losing its core support, the main body of the building tilted catastrophically to one side and finally disintegrated completely with a devastating roar, turning into a pile of tumbling bricks, stones and wood. An even larger, denser column of dust shot up into the sky, instantly engulfing all the wreckage. The entire scene was filled with a sense of destructive power and an atmosphere of irreversible end.", + "seed": 21163, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_5/index_5562_seed_21163_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15346725843846798, + "vlm_logprob": -0.15346725843846798, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.15346725843846798, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09852573275566101 + ], + "opposite_logprob_tokens": [ + -2.4735257625579834 + ], + "logprob_reward": -0.09852573275566101, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09852573275566101, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2477223128080368, + 0.1385098397731781, + 0.050152670592069626, + 0.048944536596536636, + 0.07526672631502151, + 0.06369797140359879, + 0.062472812831401825, + 0.07249219715595245, + 0.18342489004135132, + 0.05731603503227234 + ], + "credit_map": [ + 0.035434648394584656, + 0.02881699427962303, + 0.02995825558900833, + 0.039597369730472565, + 0.036975517868995667, + 0.036736443638801575, + 0.021472513675689697, + 0.018730580806732178, + 0.016133274883031845, + 0.015625307336449623, + 0.02992032654583454, + 0.026006381958723068, + 0.012775520794093609, + 0.019413968548178673, + 0.01090435404330492, + 0.007730699609965086, + 0.006792566739022732, + 0.007936348207294941, + 0.007568372879177332, + 0.006575122941285372, + 0.006373336538672447, + 0.005124560557305813, + 0.005574264097958803, + 0.004208101890981197, + 0.00739561440423131, + 0.004269415512681007, + 0.013278951868414879, + 0.005646214820444584, + 0.004247875418514013, + 0.005615112371742725, + 0.004481274168938398, + 0.004010071977972984, + 0.008902760222554207, + 0.009571455419063568, + 0.006486274767667055, + 0.007408459670841694, + 0.010386986657977104, + 0.012492689304053783, + 0.009956425055861473, + 0.010061673820018768, + 0.0059769777581095695, + 0.00751016940921545, + 0.006524545606225729, + 0.008824161253869534, + 0.0079198544844985, + 0.005520954262465239, + 0.011532740667462349, + 0.009888571687042713, + 0.00741464551538229, + 0.007177132647484541, + 0.007621515076607466, + 0.008702756837010384, + 0.0063483938574790955, + 0.004700863733887672, + 0.013627517968416214, + 0.006879987195134163, + 0.004287237301468849, + 0.005032114684581757, + 0.003786787623539567, + 0.007185208611190319, + 0.012832957319915295, + 0.014345340430736542, + 0.015132525004446507, + 0.009890026412904263, + 0.022547174245119095, + 0.01201090682297945, + 0.01026281900703907, + 0.010193918831646442, + 0.036662254482507706, + 0.022423595190048218, + 0.03246542811393738, + 0.03685878962278366, + 0.00704605458304286, + 0.005463678855448961, + 0.0052361320704221725, + 0.005112555343657732, + 0.00676870858296752, + 0.005361837334930897, + 0.012308366596698761, + 0.010018703527748585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.09852573275566101 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048687707632780075 + ], + "opposite_logprob_tokens": [ + -3.048687696456909 + ], + "logprob_reward": -0.048687707632780075, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048687707632780075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13342107832431793, + 0.03644166141748428, + 0.020181862637400627, + 0.019658256322145462, + 0.03375159949064255, + 0.03614715486764908, + 0.09027331322431564, + 0.09005262702703476, + 0.38412439823150635, + 0.15594802796840668 + ], + "credit_map": [ + 0.02598084881901741, + 0.023508453741669655, + 0.015987006947398186, + 0.012777039781212807, + 0.013974124565720558, + 0.01880797930061817, + 0.014057448133826256, + 0.008328178897500038, + 0.005077839829027653, + 0.002804590854793787, + 0.007946181111037731, + 0.007799179293215275, + 0.002069433219730854, + 0.0050091976299881935, + 0.0035433422308415174, + 0.00219189771451056, + 0.0023670268710702658, + 0.0023157414980232716, + 0.0021661128848791122, + 0.002563892863690853, + 0.002102564787492156, + 0.0019566004630178213, + 0.004102989099919796, + 0.0026069346349686384, + 0.003108417149633169, + 0.0021473276428878307, + 0.0025924856308847666, + 0.0023751510307192802, + 0.0021502780728042126, + 0.0038671211805194616, + 0.001983899623155594, + 0.0014335763407871127, + 0.005076600704342127, + 0.0035478754434734583, + 0.0023375379387289286, + 0.0031982692889869213, + 0.003319792915135622, + 0.00866626389324665, + 0.004934681113809347, + 0.002670577960088849, + 0.004406065214425325, + 0.0034690715838223696, + 0.004256716929376125, + 0.0030386322177946568, + 0.003667452372610569, + 0.006314732600003481, + 0.005492719821631908, + 0.005501763429492712, + 0.010548243299126625, + 0.010426552034914494, + 0.011595581658184528, + 0.011346746236085892, + 0.008509300649166107, + 0.007975215092301369, + 0.019882479682564735, + 0.009989196434617043, + 0.005757853388786316, + 0.007225141394883394, + 0.005090377759188414, + 0.007956678047776222, + 0.016992295160889626, + 0.02009938657283783, + 0.016151033341884613, + 0.010779867880046368, + 0.04436010122299194, + 0.024085791781544685, + 0.021138647571206093, + 0.023955779150128365, + 0.07201331853866577, + 0.056147802621126175, + 0.06791691482067108, + 0.07450607419013977, + 0.01840806007385254, + 0.014569406397640705, + 0.010063844732940197, + 0.01293248776346445, + 0.02420281432569027, + 0.015387903898954391, + 0.026532603427767754, + 0.033850912004709244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.048687707632780075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762366056442261 + ], + "opposite_logprob_tokens": [ + -0.8262366056442261 + ], + "logprob_reward": -0.5762366056442261, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762366056442261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11203479021787643, + 0.0347614511847496, + 0.08010781556367874, + 0.029200740158557892, + 0.05738068372011185, + 0.06659410148859024, + 0.09275916963815689, + 0.24864067137241364, + 0.14428173005580902, + 0.1342388093471527 + ], + "credit_map": [ + 0.020167088136076927, + 0.021727977320551872, + 0.016415808349847794, + 0.011539910919964314, + 0.011136494576931, + 0.014706923626363277, + 0.009523800574243069, + 0.00681679230183363, + 0.005862914025783539, + 0.0044360077008605, + 0.004645771346986294, + 0.003863480407744646, + 0.004198231268674135, + 0.00539562851190567, + 0.003952436149120331, + 0.0024069887585937977, + 0.013149335980415344, + 0.01353391632437706, + 0.007361413445323706, + 0.007223428692668676, + 0.010396215133368969, + 0.00896314438432455, + 0.012912149541079998, + 0.006568221375346184, + 0.0043132067658007145, + 0.002923954976722598, + 0.0037788625340908766, + 0.004572737962007523, + 0.0025599042419344187, + 0.004449855536222458, + 0.003333308268338442, + 0.0032689112704247236, + 0.00769887026399374, + 0.005630568601191044, + 0.0045760562643408775, + 0.007070929277688265, + 0.005823029205203056, + 0.009836309589445591, + 0.009656229987740517, + 0.0070886919274926186, + 0.008137340657413006, + 0.006661440711468458, + 0.004704453516751528, + 0.006295057479292154, + 0.00843661930412054, + 0.01258776243776083, + 0.010365103371441364, + 0.00940632726997137, + 0.011747628450393677, + 0.009703402407467365, + 0.012637457810342312, + 0.011817596852779388, + 0.007920274510979652, + 0.008412136696279049, + 0.019049616530537605, + 0.011471062898635864, + 0.017225170508027077, + 0.014223417267203331, + 0.016487782821059227, + 0.030954286456108093, + 0.03284174203872681, + 0.03618372976779938, + 0.06353385001420975, + 0.037190694361925125, + 0.015002069063484669, + 0.010105839930474758, + 0.009778919629752636, + 0.012123595923185349, + 0.01870829239487648, + 0.02238871157169342, + 0.028544362634420395, + 0.027629956603050232, + 0.016924146562814713, + 0.01712139882147312, + 0.006609143689274788, + 0.009407766163349152, + 0.02225697785615921, + 0.010482228361070156, + 0.027636196464300156, + 0.023800963535904884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.5762366056442261 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294713348150253 + ], + "opposite_logprob_tokens": [ + -3.773294687271118 + ], + "logprob_reward": -0.023294713348150253, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294713348150253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04610874131321907, + 0.02227550558745861, + 0.020346608012914658, + 0.019123943522572517, + 0.030167873948812485, + 0.026603706181049347, + 0.10535898059606552, + 0.4239320158958435, + 0.11866438388824463, + 0.18741822242736816 + ], + "credit_map": [ + 0.006428597029298544, + 0.005575770977884531, + 0.007950663566589355, + 0.006287842523306608, + 0.004965122323483229, + 0.005990371573716402, + 0.005143263842910528, + 0.003767111338675022, + 0.0029619832057505846, + 0.0026993684004992247, + 0.0036781986709684134, + 0.0031792516820132732, + 0.0020202624145895243, + 0.0027887229807674885, + 0.002443592296913266, + 0.0025041259359568357, + 0.002991053042933345, + 0.002489826874807477, + 0.0021819458343088627, + 0.0023697614669799805, + 0.0023653798270970583, + 0.0022672987543046474, + 0.0034964801743626595, + 0.002184861106798053, + 0.0033757800702005625, + 0.001439487561583519, + 0.002853277139365673, + 0.002332334639504552, + 0.00159004470333457, + 0.002974424511194229, + 0.0022590807639062405, + 0.002299515763297677, + 0.004603832960128784, + 0.002586271846666932, + 0.0026683509349823, + 0.0026552388444542885, + 0.003492240561172366, + 0.007121787406504154, + 0.00389322848059237, + 0.0031469224486500025, + 0.0027095878031104803, + 0.003650385420769453, + 0.0033879471011459827, + 0.0028955196030437946, + 0.0026304288767278194, + 0.0021059317514300346, + 0.004533963743597269, + 0.004689940717071295, + 0.006381147541105747, + 0.016289426013827324, + 0.00486625824123621, + 0.018866300582885742, + 0.008264442905783653, + 0.007555691059678793, + 0.03352241963148117, + 0.009613286703824997, + 0.028488928452134132, + 0.027432221919298172, + 0.027621502056717873, + 0.04478290677070618, + 0.05844295397400856, + 0.07248782366514206, + 0.09549303352832794, + 0.0691826269030571, + 0.013375191017985344, + 0.006535968743264675, + 0.005484741181135178, + 0.0072302017360925674, + 0.027267921715974808, + 0.01463342271745205, + 0.03166741505265236, + 0.012469531036913395, + 0.019715066999197006, + 0.01686805486679077, + 0.013840785250067711, + 0.014795028604567051, + 0.03224946931004524, + 0.020805589854717255, + 0.04048265516757965, + 0.0286615751683712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.023294713348150253 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591532811522484 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591532811522484, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591532811522484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1633242666721344, + 0.07783205807209015, + 0.06587819010019302, + 0.056866735219955444, + 0.05415977165102959, + 0.05506006255745888, + 0.22474569082260132, + 0.09713725000619888, + 0.07217930257320404, + 0.1328166425228119 + ], + "credit_map": [ + 0.03593280911445618, + 0.023650670424103737, + 0.02866014651954174, + 0.014775642193853855, + 0.013410612940788269, + 0.025858178734779358, + 0.01251241285353899, + 0.0085237892344594, + 0.009314463473856449, + 0.011059047654271126, + 0.01661272533237934, + 0.010203881189227104, + 0.00632129842415452, + 0.010255293920636177, + 0.007716018240898848, + 0.006349335890263319, + 0.01001411210745573, + 0.01120967511087656, + 0.006463479716330767, + 0.008484832011163235, + 0.009634760208427906, + 0.006643298547714949, + 0.00773503165692091, + 0.005693006794899702, + 0.007882661186158657, + 0.004771523177623749, + 0.013466092757880688, + 0.007227757479995489, + 0.00610562227666378, + 0.005659724585711956, + 0.007750732358545065, + 0.00400262326002121, + 0.007084083277732134, + 0.006004186347126961, + 0.004619809798896313, + 0.005400087684392929, + 0.0069128829054534435, + 0.010020808316767216, + 0.008235308341681957, + 0.005882606375962496, + 0.00600013229995966, + 0.006344777997583151, + 0.007099739275872707, + 0.005927016492933035, + 0.006683265790343285, + 0.006890897639095783, + 0.00839760061353445, + 0.0077166324481368065, + 0.02322489768266678, + 0.026630913838744164, + 0.02863527275621891, + 0.033333927392959595, + 0.019437162205576897, + 0.023490678519010544, + 0.04739711061120033, + 0.022595738992094994, + 0.00755758211016655, + 0.004793132189661264, + 0.0049077547155320644, + 0.009655573405325413, + 0.011729007586836815, + 0.014312991872429848, + 0.03093586675822735, + 0.01324534136801958, + 0.008008340373635292, + 0.007003661245107651, + 0.00782718975096941, + 0.005749400705099106, + 0.008157776668667793, + 0.011697443202137947, + 0.009672843851149082, + 0.01406264491379261, + 0.015570350922644138, + 0.01559598371386528, + 0.012182839214801788, + 0.014647575095295906, + 0.02165100909769535, + 0.014782111160457134, + 0.014151918701827526, + 0.024234866723418236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.020591532811522484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot, the frame trembling violently, captures the tragic final moments of a magnificent, ancient, white-walled, black-tiled Hui-style building. Against a backdrop of a hazy, suspended dusty sky, the building's mottled white walls and towering horse-head walls appear remarkably solemn. Suddenly, with a resounding, muffled roar from the ground, the building's foundations begin to tremble violently. The thick, dark brown wooden beams crackle with a piercing sound under the overwhelming pressure, instantly twisting and snapping. Immediately afterwards, the layers of dark gray tiles on the roof lose their grip and are hurled in all directions like black hail, hurled in a destructive force, leaving countless shattered paths in the air. After losing its core support, the main body of the building tilted catastrophically to one side and finally disintegrated completely with a devastating roar, turning into a pile of tumbling bricks, stones and wood. An even larger, denser column of dust shot up into the sky, instantly engulfing all the wreckage. The entire scene was filled with a sense of destructive power and an atmosphere of irreversible end.", + "seed": 21164, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_5/index_5562_seed_21164_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15479835495352745, + "vlm_logprob": -0.15479835495352745, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.15479835495352745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0795801430940628 + ], + "opposite_logprob_tokens": [ + -2.704580068588257 + ], + "logprob_reward": -0.0795801430940628, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0795801430940628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06898380070924759, + 0.025448232889175415, + 0.1472756266593933, + 0.016507791355252266, + 0.12402606755495071, + 0.4286026060581207, + 0.04478960111737251, + 0.03229974955320358, + 0.055176183581352234, + 0.05689031258225441 + ], + "credit_map": [ + 0.010570758022367954, + 0.006405675783753395, + 0.011241508647799492, + 0.007203615736216307, + 0.008671514689922333, + 0.011080065742135048, + 0.006665801629424095, + 0.007144863251596689, + 0.0037531631533056498, + 0.002209549071267247, + 0.003529180306941271, + 0.0030475680250674486, + 0.0027180525939911604, + 0.004541304428130388, + 0.002352983457967639, + 0.003296430455520749, + 0.05346224829554558, + 0.005536472890526056, + 0.006413829978555441, + 0.004960756748914719, + 0.004014190752059221, + 0.06105475872755051, + 0.006099959835410118, + 0.005733394529670477, + 0.0023841767106205225, + 0.0016073682345449924, + 0.0016029721591621637, + 0.00221998686902225, + 0.002255020197480917, + 0.0021887884940952063, + 0.002564564812928438, + 0.0016849141102284193, + 0.04080688953399658, + 0.006159122567623854, + 0.00608042161911726, + 0.00600418820977211, + 0.005448672454804182, + 0.048421554267406464, + 0.006093318574130535, + 0.005011907313019037, + 0.1512286365032196, + 0.020263483747839928, + 0.011379131115972996, + 0.012048866599798203, + 0.012323633767664433, + 0.19274815917015076, + 0.016342604532837868, + 0.012268107384443283, + 0.004254248924553394, + 0.004212769214063883, + 0.004158910363912582, + 0.008384319953620434, + 0.004302035551518202, + 0.006362562067806721, + 0.0066840467043221, + 0.00643070787191391, + 0.003090046579018235, + 0.0033994060941040516, + 0.005511952098459005, + 0.00458748172968626, + 0.003568793646991253, + 0.004291955381631851, + 0.0038633348885923624, + 0.003986777272075415, + 0.007530005648732185, + 0.006884480360895395, + 0.004452118184417486, + 0.005775192752480507, + 0.008793243207037449, + 0.007125843781977892, + 0.008807296864688396, + 0.0058079990558326244, + 0.004474518354982138, + 0.007061943411827087, + 0.004972105845808983, + 0.0046857548877596855, + 0.00761395925655961, + 0.011902784928679466, + 0.007991406135261059, + 0.008187840692698956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.0795801430940628 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332965098321438 + ], + "opposite_logprob_tokens": [ + -3.773329734802246 + ], + "logprob_reward": -0.02332965098321438, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332965098321438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01263065543025732, + 0.009259101934731007, + 0.2754696011543274, + 0.003400262678042054, + 0.029893601313233376, + 0.47730493545532227, + 0.024280736222863197, + 0.01239458005875349, + 0.049454446882009506, + 0.10591201484203339 + ], + "credit_map": [ + 0.0037887783255428076, + 0.0005855800700373948, + 0.0011283046333119273, + 0.000806110561825335, + 0.0006552544655278325, + 0.0036810473538935184, + 0.0007494944729842246, + 0.0012360854307189584, + 0.0019337175181135535, + 0.0006445246981456876, + 0.0012711299350485206, + 0.0005318028852343559, + 0.0008838308276608586, + 0.0025193083565682173, + 0.00065707036992535, + 0.0008177173440344632, + 0.11048107594251633, + 0.008804620243608952, + 0.005890266969799995, + 0.004142804071307182, + 0.004917587153613567, + 0.12572357058525085, + 0.009192969650030136, + 0.006316730286926031, + 0.0006654096068814397, + 0.00018438143888488412, + 0.0004294048121664673, + 0.0004554048355203122, + 0.00028621082310564816, + 0.00047385116340592504, + 0.0004942225641570985, + 0.00041137757943943143, + 0.010145129635930061, + 0.001551333931274712, + 0.0013061502249911427, + 0.0013066896935924888, + 0.001156786223873496, + 0.011903039179742336, + 0.0013303000014275312, + 0.0011941733537241817, + 0.17204177379608154, + 0.021759746596217155, + 0.011087862774729729, + 0.008499841205775738, + 0.013722972944378853, + 0.21911415457725525, + 0.016650140285491943, + 0.014428462833166122, + 0.0024525581393390894, + 0.0023743000347167253, + 0.003336499212309718, + 0.004419270902872086, + 0.002945647342130542, + 0.0032058535143733025, + 0.0027862940914928913, + 0.0027603094931691885, + 0.0013109981082379818, + 0.0013851842377334833, + 0.001202515559270978, + 0.0014764376683160663, + 0.0017471556784585118, + 0.0013554457109421492, + 0.0020216144621372223, + 0.0018952293321490288, + 0.008005520328879356, + 0.00770212197676301, + 0.003388406941667199, + 0.0038884698878973722, + 0.00870270561426878, + 0.005948190577328205, + 0.00743095763027668, + 0.004388073924928904, + 0.008345573209226131, + 0.015276821330189705, + 0.0064834062941372395, + 0.006831644102931023, + 0.014378348365426064, + 0.024282846599817276, + 0.014110459014773369, + 0.016202915459871292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.02332965098321438 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234429836273193 + ], + "opposite_logprob_tokens": [ + -0.8984429836273193 + ], + "logprob_reward": -0.5234429836273193, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234429836273193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05092720314860344, + 0.018461868166923523, + 0.3161640465259552, + 0.014065477065742016, + 0.03234925866127014, + 0.37486547231674194, + 0.03440573439002037, + 0.043470874428749084, + 0.05866403877735138, + 0.056626029312610626 + ], + "credit_map": [ + 0.01700921729207039, + 0.0018386038718745112, + 0.004126426298171282, + 0.0020311567932367325, + 0.0018038350390270352, + 0.01837850920855999, + 0.002168449107557535, + 0.003571006003767252, + 0.0022309995256364346, + 0.0013190731406211853, + 0.003932110499590635, + 0.0017240406014025211, + 0.0022190676536411047, + 0.0028593053575605154, + 0.0020928685553371906, + 0.002084402833133936, + 0.12629276514053345, + 0.010259994305670261, + 0.006781851407140493, + 0.004900981672108173, + 0.005905367434024811, + 0.14406313002109528, + 0.010490907356142998, + 0.007469037547707558, + 0.003719244385138154, + 0.0009746733121573925, + 0.0010619348613545299, + 0.0011163994204252958, + 0.0013208239106461406, + 0.004134702496230602, + 0.0011006249114871025, + 0.0006370734190568328, + 0.009142154827713966, + 0.001978571992367506, + 0.0020614759996533394, + 0.002159020397812128, + 0.0019357872661203146, + 0.010907846502959728, + 0.001912074745632708, + 0.002252329373732209, + 0.13461540639400482, + 0.017370840534567833, + 0.008877399377524853, + 0.007322662975639105, + 0.010797672905027866, + 0.1712845116853714, + 0.013430831022560596, + 0.0111661646515131, + 0.0034712376073002815, + 0.0028547660913318396, + 0.003252895548939705, + 0.0049772425554692745, + 0.003996570594608784, + 0.005268200766295195, + 0.004652873612940311, + 0.005931946914643049, + 0.0036425760481506586, + 0.004378180019557476, + 0.006235762964934111, + 0.007103915326297283, + 0.005212686490267515, + 0.005191620904952288, + 0.006843596696853638, + 0.0048625352792441845, + 0.0072323232889175415, + 0.007148108910769224, + 0.004918191581964493, + 0.006942854728549719, + 0.009564967826008797, + 0.006267948541790247, + 0.010568025521934032, + 0.0060216160491108894, + 0.004846439231187105, + 0.008160531520843506, + 0.005280497018247843, + 0.005412839353084564, + 0.010908138006925583, + 0.009851353242993355, + 0.007064764853566885, + 0.0051014618948102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.5234429836273193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704606354236603 + ], + "opposite_logprob_tokens": [ + -2.1270461082458496 + ], + "logprob_reward": -0.12704606354236603, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704606354236603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02826075814664364, + 0.006451512686908245, + 0.428291380405426, + 0.00314663746394217, + 0.08553925901651382, + 0.3220692574977875, + 0.03697384148836136, + 0.02013484202325344, + 0.05122222378849983, + 0.017910290509462357 + ], + "credit_map": [ + 0.010334576480090618, + 0.0010235877707600594, + 0.0015631737187504768, + 0.0008526108576916158, + 0.0008501454722136259, + 0.011113087646663189, + 0.0011891578324139118, + 0.0013344193575903773, + 0.001229538000188768, + 0.0005714567378163338, + 0.0006231056759133935, + 0.0007082444499246776, + 0.0007490142015740275, + 0.0014686065260320902, + 0.0006729682791046798, + 0.0004285791073925793, + 0.17173460125923157, + 0.01369759626686573, + 0.009186995215713978, + 0.0064344038255512714, + 0.00767745403572917, + 0.19543786346912384, + 0.014152592048048973, + 0.009969898499548435, + 0.000359516532626003, + 0.00044489873107522726, + 0.00040194473695009947, + 0.00034927736851386726, + 0.0006203859229572117, + 0.0003119682369288057, + 0.0004252455837558955, + 0.00023340062762144953, + 0.03112037293612957, + 0.0037913736887276173, + 0.0020560601260513067, + 0.0037096580490469933, + 0.002289769472554326, + 0.036995187401771545, + 0.0033345124684274197, + 0.0022423285990953445, + 0.11582588404417038, + 0.014975282363593578, + 0.007573815993964672, + 0.005951133091002703, + 0.009249967522919178, + 0.1475982666015625, + 0.011345564387738705, + 0.009549340233206749, + 0.0039154053665697575, + 0.003912519663572311, + 0.00469613540917635, + 0.006342426408082247, + 0.0046190242283046246, + 0.005209971684962511, + 0.004392910748720169, + 0.0038854528684169054, + 0.0018948977813124657, + 0.002165153855457902, + 0.0032781537156552076, + 0.0030051828362047672, + 0.002213961910456419, + 0.0024894913658499718, + 0.0027393028140068054, + 0.0023486982099711895, + 0.007528620772063732, + 0.006410255096852779, + 0.003318101167678833, + 0.004680653102695942, + 0.009808477945625782, + 0.0053911274299025536, + 0.009963881224393845, + 0.004121110308915377, + 0.0015644049271941185, + 0.0025800939183682203, + 0.0016147330170497298, + 0.0016445055371150374, + 0.003667077049612999, + 0.002518127439543605, + 0.002251504920423031, + 0.0020698457956314087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.12704606354236603 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592933520674706 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592933520674706, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592933520674706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045877858996391296, + 0.02570144273340702, + 0.16195687651634216, + 0.013972737826406956, + 0.38089945912361145, + 0.28372934460639954, + 0.018191754817962646, + 0.009455026127398014, + 0.029931215569376945, + 0.03028431162238121 + ], + "credit_map": [ + 0.01056189276278019, + 0.0028011430986225605, + 0.007069275714457035, + 0.003846404841169715, + 0.004003102425485849, + 0.010426116175949574, + 0.003845304250717163, + 0.00332461716607213, + 0.004287119954824448, + 0.0015847372123971581, + 0.002901292871683836, + 0.0021058316342532635, + 0.003341862466186285, + 0.005738369654864073, + 0.0018680677749216557, + 0.0038741605821996927, + 0.06455153971910477, + 0.005071916617453098, + 0.003655390813946724, + 0.002739501418545842, + 0.0029312698170542717, + 0.07381768524646759, + 0.00528082437813282, + 0.003908755257725716, + 0.003403985872864723, + 0.000767018529586494, + 0.0007857575546950102, + 0.001291573280468583, + 0.001048461184836924, + 0.003783112158998847, + 0.001387599972076714, + 0.001505230087786913, + 0.1369481384754181, + 0.017002642154693604, + 0.012721146456897259, + 0.014868177473545074, + 0.010315008461475372, + 0.16278615593910217, + 0.014937304891645908, + 0.011320884339511395, + 0.10113632678985596, + 0.01348341815173626, + 0.007307499647140503, + 0.0052881003357470036, + 0.00872145127505064, + 0.12881578505039215, + 0.010257557965815067, + 0.00871919747442007, + 0.0020240372978150845, + 0.0019075177842751145, + 0.002184817101806402, + 0.0029391064308583736, + 0.002501643728464842, + 0.002521930728107691, + 0.0022151346784085035, + 0.0018975675338879228, + 0.00089450174709782, + 0.0009227950358763337, + 0.00120857625734061, + 0.001394971739500761, + 0.0010900559136644006, + 0.0011218175059184432, + 0.0015261825174093246, + 0.001296125235967338, + 0.0053672874346375465, + 0.004101084545254707, + 0.002080043777823448, + 0.0020561995916068554, + 0.0061532845720648766, + 0.003254308830946684, + 0.00460728257894516, + 0.0023117244709283113, + 0.002560388995334506, + 0.004066332243382931, + 0.0028661391697824, + 0.0028773832600563765, + 0.004520551301538944, + 0.0045105768367648125, + 0.004755356814712286, + 0.004127581603825092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.020592933520674706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot, the frame trembling violently, captures the tragic final moments of a magnificent, ancient, white-walled, black-tiled Hui-style building. Against a backdrop of a hazy, suspended dusty sky, the building's mottled white walls and towering horse-head walls appear remarkably solemn. Suddenly, with a resounding, muffled roar from the ground, the building's foundations begin to tremble violently. The thick, dark brown wooden beams crackle with a piercing sound under the overwhelming pressure, instantly twisting and snapping. Immediately afterwards, the layers of dark gray tiles on the roof lose their grip and are hurled in all directions like black hail, hurled in a destructive force, leaving countless shattered paths in the air. After losing its core support, the main body of the building tilted catastrophically to one side and finally disintegrated completely with a devastating roar, turning into a pile of tumbling bricks, stones and wood. An even larger, denser column of dust shot up into the sky, instantly engulfing all the wreckage. The entire scene was filled with a sense of destructive power and an atmosphere of irreversible end.", + "seed": 21165, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_5/index_5562_seed_21165_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10672632344067097, + "vlm_logprob": -0.10672632344067097, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.10672632344067097, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027910824865102768 + ], + "opposite_logprob_tokens": [ + -3.7779107093811035 + ], + "logprob_reward": -0.027910824865102768, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027910824865102768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10542244464159012, + 0.05925684794783592, + 0.024076731875538826, + 0.06477090716362, + 0.028607038781046867, + 0.37403246760368347, + 0.10268139094114304, + 0.11178033798933029, + 0.09196341037750244, + 0.037408407777547836 + ], + "credit_map": [ + 0.01428946666419506, + 0.014057078398764133, + 0.01855257712304592, + 0.009214435704052448, + 0.016847608610987663, + 0.01090586930513382, + 0.010049157775938511, + 0.011506257578730583, + 0.009873753413558006, + 0.00669963750988245, + 0.012136664241552353, + 0.00557367829605937, + 0.005342120770365, + 0.007061515934765339, + 0.0053884307853877544, + 0.007181047927588224, + 0.002902258886024356, + 0.0027925404720008373, + 0.004047024995088577, + 0.0019805270712822676, + 0.004314206540584564, + 0.0028299614787101746, + 0.0026993732899427414, + 0.002510838909074664, + 0.008312252350151539, + 0.006364724133163691, + 0.010318073444068432, + 0.009679596871137619, + 0.004670750815421343, + 0.007714792154729366, + 0.008477971889078617, + 0.00923275388777256, + 0.002903144108131528, + 0.002839641412720084, + 0.004078129306435585, + 0.005050803069025278, + 0.003564820159226656, + 0.003711029188707471, + 0.0036995750851929188, + 0.002759896917268634, + 0.023221489042043686, + 0.10879003256559372, + 0.015966352075338364, + 0.030672909691929817, + 0.1423453539609909, + 0.017212023958563805, + 0.020509768277406693, + 0.015314524061977863, + 0.012794683687388897, + 0.02119619771838188, + 0.007388510275632143, + 0.024168889969587326, + 0.0064371004700660706, + 0.005700225941836834, + 0.01942945271730423, + 0.005566325504332781, + 0.015580935403704643, + 0.017062753438949585, + 0.014683757908642292, + 0.013377606868743896, + 0.013030169531702995, + 0.00888351071625948, + 0.015484157018363476, + 0.013677438721060753, + 0.010033284313976765, + 0.012763513252139091, + 0.009278325363993645, + 0.00892568938434124, + 0.012794792652130127, + 0.018030036240816116, + 0.012257118709385395, + 0.007880656979978085, + 0.005523374769836664, + 0.00609615258872509, + 0.0038230533245950937, + 0.003309563035145402, + 0.007274434436112642, + 0.005030197557061911, + 0.002908320166170597, + 0.0034433137625455856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.027910824865102768 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362914592027664 + ], + "opposite_logprob_tokens": [ + -3.651362895965576 + ], + "logprob_reward": -0.026362914592027664, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362914592027664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04476369917392731, + 0.02998744137585163, + 0.01337022427469492, + 0.014964714646339417, + 0.03769589960575104, + 0.2324318140745163, + 0.08519239723682404, + 0.04802924394607544, + 0.3100064694881439, + 0.1835581213235855 + ], + "credit_map": [ + 0.008329509757459164, + 0.005667822901159525, + 0.007812797091901302, + 0.003901470685377717, + 0.005008297506719828, + 0.004502840805798769, + 0.003998466767370701, + 0.005542493425309658, + 0.004730140324681997, + 0.0028966479003429413, + 0.003972976468503475, + 0.003391968784853816, + 0.003019588766619563, + 0.0032931456808000803, + 0.004245043732225895, + 0.004437930416315794, + 0.0017313247080892324, + 0.001496830489486456, + 0.003042171709239483, + 0.0012795054353773594, + 0.0011457494692876935, + 0.002083594910800457, + 0.001266902661882341, + 0.0013241455890238285, + 0.0018434040248394012, + 0.0015997504815459251, + 0.0022928346879780293, + 0.002058343030512333, + 0.0015453989617526531, + 0.0015989512903615832, + 0.0023809331469237804, + 0.0016450990224257112, + 0.0025776131078600883, + 0.002094112802296877, + 0.006537080742418766, + 0.005976079497486353, + 0.003680408000946045, + 0.0059080845676362514, + 0.007616592105478048, + 0.0033059297129511833, + 0.01564544439315796, + 0.06848806142807007, + 0.009769349358975887, + 0.018379662185907364, + 0.0867505818605423, + 0.010824494995176792, + 0.01283798273652792, + 0.009736239910125732, + 0.006779429968446493, + 0.00858552660793066, + 0.008208098821341991, + 0.025851504877209663, + 0.0037329320330172777, + 0.004411463160067797, + 0.0242586862295866, + 0.0033647629898041487, + 0.006623216904699802, + 0.010310260578989983, + 0.004950706847012043, + 0.003771667368710041, + 0.007661025505512953, + 0.004581249784678221, + 0.0049856677651405334, + 0.0051454524509608746, + 0.03672229126095772, + 0.06004646420478821, + 0.017924077808856964, + 0.0225495807826519, + 0.04224322363734245, + 0.08421166241168976, + 0.030488064512610435, + 0.01582108624279499, + 0.03121643327176571, + 0.027913613244891167, + 0.013794150203466415, + 0.013030817732214928, + 0.03902541100978851, + 0.03224293887615204, + 0.015715723857283592, + 0.010619032196700573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.026362914592027664 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289657175540924 + ], + "opposite_logprob_tokens": [ + -1.05396568775177 + ], + "logprob_reward": -0.4289657175540924, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289657175540924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030392324551939964, + 0.0277569517493248, + 0.012628759257495403, + 0.039166972041130066, + 0.03946531563997269, + 0.1180911511182785, + 0.24534645676612854, + 0.10194621235132217, + 0.22585971653461456, + 0.15934614837169647 + ], + "credit_map": [ + 0.0051537081599235535, + 0.0040938640013337135, + 0.005831782706081867, + 0.002862396650016308, + 0.0027698513586074114, + 0.0029313245322555304, + 0.0033296977635473013, + 0.003419698914512992, + 0.004702569451183081, + 0.0024095638655126095, + 0.005886924918740988, + 0.002553891623392701, + 0.0027127426583319902, + 0.00307592679746449, + 0.00286310538649559, + 0.0035522268153727055, + 0.0012312185717746615, + 0.0015484680188819766, + 0.002682320075109601, + 0.0015921818558126688, + 0.0014832253800705075, + 0.001768925110809505, + 0.0011584446765482426, + 0.0011639754520729184, + 0.0038404653314501047, + 0.004965947940945625, + 0.005207469221204519, + 0.005000180099159479, + 0.004260240588337183, + 0.003975857514888048, + 0.007315672934055328, + 0.00460113724693656, + 0.003085264004766941, + 0.0019737279508262873, + 0.00873807817697525, + 0.00561818340793252, + 0.003239111742004752, + 0.00888480432331562, + 0.004774655215442181, + 0.0031514917500317097, + 0.010189983993768692, + 0.03358849510550499, + 0.006971259601414204, + 0.010771239176392555, + 0.03820032626390457, + 0.0051994724199175835, + 0.006741134449839592, + 0.006429241504520178, + 0.029091445729136467, + 0.04912697523832321, + 0.012294389307498932, + 0.06687406450510025, + 0.011965267360210419, + 0.00941350869834423, + 0.05531320720911026, + 0.01126760896295309, + 0.013576637953519821, + 0.013195141218602657, + 0.013754921965301037, + 0.012358926236629486, + 0.012370282784104347, + 0.009923705831170082, + 0.014282767660915852, + 0.012483830563724041, + 0.027103908360004425, + 0.03453846275806427, + 0.01936834491789341, + 0.019745932891964912, + 0.02676873654127121, + 0.04898380488157272, + 0.03367431461811066, + 0.015676207840442657, + 0.024264799430966377, + 0.02356758713722229, + 0.01449707429856062, + 0.017806867137551308, + 0.0228742565959692, + 0.02704744227230549, + 0.015521823428571224, + 0.013766307383775711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.4289657175540924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801445081830025 + ], + "opposite_logprob_tokens": [ + -3.529801368713379 + ], + "logprob_reward": -0.029801445081830025, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801445081830025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02459542267024517, + 0.020914385095238686, + 0.00986933708190918, + 0.04328601434826851, + 0.059606343507766724, + 0.40362951159477234, + 0.036982461810112, + 0.06538321822881699, + 0.15405525267124176, + 0.18167805671691895 + ], + "credit_map": [ + 0.0034053106792271137, + 0.003004671772941947, + 0.004426930565387011, + 0.0023206346668303013, + 0.0031388001516461372, + 0.0025298073887825012, + 0.0034234074410051107, + 0.002345861867070198, + 0.003713623620569706, + 0.0022747409529983997, + 0.004680486861616373, + 0.0020510382018983364, + 0.001948236720636487, + 0.0024023642763495445, + 0.00233577867038548, + 0.0015081192832440138, + 0.0011560264974832535, + 0.0010667245369404554, + 0.0019297779072076082, + 0.000950454268604517, + 0.0014173219678923488, + 0.0013109487481415272, + 0.001029432169161737, + 0.0010086518013849854, + 0.005236614029854536, + 0.005429526325315237, + 0.005706579424440861, + 0.005800756625831127, + 0.004447943530976772, + 0.0047425501979887486, + 0.0063492655754089355, + 0.005572780501097441, + 0.0036976924166083336, + 0.005705046467483044, + 0.005933297332376242, + 0.009900492615997791, + 0.006859852001070976, + 0.010619422420859337, + 0.008751926943659782, + 0.008138622157275677, + 0.018163146451115608, + 0.09716715663671494, + 0.018773429095745087, + 0.0457354336977005, + 0.16014976799488068, + 0.017716798931360245, + 0.02667001076042652, + 0.019253768026828766, + 0.003028283594176173, + 0.00477735698223114, + 0.0027811971958726645, + 0.008904248476028442, + 0.0030475121457129717, + 0.003020724281668663, + 0.009415161795914173, + 0.0020079792011529207, + 0.009443171322345734, + 0.00961916521191597, + 0.008953792974352837, + 0.007413232233375311, + 0.008153153583407402, + 0.005652911029756069, + 0.008857293985784054, + 0.00729050487279892, + 0.015362652018666267, + 0.01595490425825119, + 0.024588171392679214, + 0.022141890600323677, + 0.01727762445807457, + 0.01756485179066658, + 0.0265357606112957, + 0.014629412442445755, + 0.03222838416695595, + 0.026339931413531303, + 0.015843922272324562, + 0.019087333232164383, + 0.03213704004883766, + 0.02827451005578041, + 0.01623486541211605, + 0.011532075703144073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.029801445081830025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059071511030197 + ], + "opposite_logprob_tokens": [ + -3.8955907821655273 + ], + "logprob_reward": -0.02059071511030197, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059071511030197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15627658367156982, + 0.09192398190498352, + 0.06348361819982529, + 0.05736880749464035, + 0.05544918775558472, + 0.13851307332515717, + 0.10441319644451141, + 0.04731586202979088, + 0.18714679777622223, + 0.09810890257358551 + ], + "credit_map": [ + 0.03135188668966293, + 0.0243173036724329, + 0.0286153182387352, + 0.016146138310432434, + 0.0156455859541893, + 0.017614029347896576, + 0.009476308710873127, + 0.01311000157147646, + 0.015273121185600758, + 0.00983376707881689, + 0.011288468725979328, + 0.008587252348661423, + 0.012128710746765137, + 0.011693919077515602, + 0.010754266753792763, + 0.012364471331238747, + 0.009027790278196335, + 0.007779309060424566, + 0.01248085405677557, + 0.005826367065310478, + 0.006021860521286726, + 0.009623851627111435, + 0.006479866337031126, + 0.006243720650672913, + 0.007303762249648571, + 0.006231066770851612, + 0.00729867210611701, + 0.008798609487712383, + 0.004909717943519354, + 0.007423109374940395, + 0.007269370369613171, + 0.008134501054883003, + 0.004404121078550816, + 0.0036436109803617, + 0.009585759602487087, + 0.010936882346868515, + 0.0038427580147981644, + 0.007885071448981762, + 0.009207470342516899, + 0.005943511612713337, + 0.007637362461537123, + 0.031829845160245895, + 0.006656039040535688, + 0.01456462126225233, + 0.05318985506892204, + 0.0067536053247749805, + 0.011786684393882751, + 0.0060950592160224915, + 0.006556216161698103, + 0.007161982823163271, + 0.01811404339969158, + 0.031955115497112274, + 0.00431863171979785, + 0.003173799254000187, + 0.029357999563217163, + 0.0037754084914922714, + 0.005681898444890976, + 0.008599738590419292, + 0.0068037426099181175, + 0.0049064247868955135, + 0.006184019148349762, + 0.0031823383178561926, + 0.005867568776011467, + 0.006090131122618914, + 0.022354856133461, + 0.0370122492313385, + 0.01236275490373373, + 0.016688933596014977, + 0.02358345314860344, + 0.0492396354675293, + 0.016058187931776047, + 0.009846734814345837, + 0.015069600194692612, + 0.015787649899721146, + 0.014360307715833187, + 0.01139985490590334, + 0.016653776168823242, + 0.011554498225450516, + 0.00697269756346941, + 0.006310515105724335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.02059071511030197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot, the frame trembling violently, captures the tragic final moments of a magnificent, ancient, white-walled, black-tiled Hui-style building. Against a backdrop of a hazy, suspended dusty sky, the building's mottled white walls and towering horse-head walls appear remarkably solemn. Suddenly, with a resounding, muffled roar from the ground, the building's foundations begin to tremble violently. The thick, dark brown wooden beams crackle with a piercing sound under the overwhelming pressure, instantly twisting and snapping. Immediately afterwards, the layers of dark gray tiles on the roof lose their grip and are hurled in all directions like black hail, hurled in a destructive force, leaving countless shattered paths in the air. After losing its core support, the main body of the building tilted catastrophically to one side and finally disintegrated completely with a devastating roar, turning into a pile of tumbling bricks, stones and wood. An even larger, denser column of dust shot up into the sky, instantly engulfing all the wreckage. The entire scene was filled with a sense of destructive power and an atmosphere of irreversible end.", + "seed": 21166, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_5/index_5562_seed_21166_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19873809441924095, + "vlm_logprob": -0.19873809441924095, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.19873809441924095, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07164081186056137 + ], + "opposite_logprob_tokens": [ + -2.821640729904175 + ], + "logprob_reward": -0.07164081186056137, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07164081186056137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10018067806959152, + 0.058337073773145676, + 0.024124884977936745, + 0.09418686479330063, + 0.08512747287750244, + 0.40274694561958313, + 0.02258267253637314, + 0.028362147510051727, + 0.10577347129583359, + 0.07857773452997208 + ], + "credit_map": [ + 0.014329041354358196, + 0.008705183863639832, + 0.020334968343377113, + 0.014787097461521626, + 0.008197598159313202, + 0.012759385630488396, + 0.013361021876335144, + 0.007706394884735346, + 0.00472485926002264, + 0.007659563794732094, + 0.005776398349553347, + 0.008857512846589088, + 0.007646504323929548, + 0.009710281156003475, + 0.006985507905483246, + 0.006976451259106398, + 0.0016800368903204799, + 0.0026856116019189358, + 0.0015861561987549067, + 0.004553914535790682, + 0.004432774148881435, + 0.004365835338830948, + 0.002175185363739729, + 0.0026453740429133177, + 0.01003552507609129, + 0.005473939701914787, + 0.010884081944823265, + 0.011126155033707619, + 0.008279385976493359, + 0.025771426036953926, + 0.013758048415184021, + 0.008858312852680683, + 0.009202327579259872, + 0.005044836085289717, + 0.004443490412086248, + 0.006389671005308628, + 0.03066067397594452, + 0.01133735105395317, + 0.011173153296113014, + 0.006875976920127869, + 0.042789798229932785, + 0.030938034877181053, + 0.026667268946766853, + 0.05596475303173065, + 0.034540943801403046, + 0.034553080797195435, + 0.060541294515132904, + 0.11675179749727249, + 0.001440907479263842, + 0.0016978010535240173, + 0.0020690010860562325, + 0.00410928949713707, + 0.0024960499722510576, + 0.001922778901644051, + 0.005426973104476929, + 0.0034198728390038013, + 0.0031037619337439537, + 0.0036370872985571623, + 0.0035535646602511406, + 0.005448468960821629, + 0.0038393656723201275, + 0.0025261954870074987, + 0.0036393655464053154, + 0.002614343073219061, + 0.007095797918736935, + 0.010682133957743645, + 0.013184623792767525, + 0.012292103841900826, + 0.018229590728878975, + 0.01686905138194561, + 0.012996306642889977, + 0.014423880726099014, + 0.00589013984426856, + 0.013564061373472214, + 0.008701024577021599, + 0.011351210996508598, + 0.01638614572584629, + 0.007495522033423185, + 0.009384596720337868, + 0.0058050379157066345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.07164081186056137 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263761505484581 + ], + "opposite_logprob_tokens": [ + -3.651376247406006 + ], + "logprob_reward": -0.0263761505484581, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263761505484581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044855646789073944, + 0.02663417160511017, + 0.03524255380034447, + 0.03340033441781998, + 0.041027795523405075, + 0.04375274106860161, + 0.10443190485239029, + 0.06984402239322662, + 0.4523898661136627, + 0.14842097461223602 + ], + "credit_map": [ + 0.005942076910287142, + 0.003086160169914365, + 0.0069899349473416805, + 0.008287707343697548, + 0.0030231073033064604, + 0.004966174717992544, + 0.008681228384375572, + 0.0038792577106505632, + 0.0017523153219372034, + 0.002120307646691799, + 0.0026077833026647568, + 0.004609988536685705, + 0.004472632892429829, + 0.004463039338588715, + 0.004077667370438576, + 0.002530437894165516, + 0.0019280500710010529, + 0.005280736368149519, + 0.0027695167809724808, + 0.004603942856192589, + 0.007797295227646828, + 0.005350017920136452, + 0.0033836248330771923, + 0.004129370674490929, + 0.003069409169256687, + 0.0021737527567893267, + 0.006807025987654924, + 0.004942938685417175, + 0.003172673285007477, + 0.004901495762169361, + 0.0047744507901370525, + 0.00355858588591218, + 0.005121749825775623, + 0.002918811747804284, + 0.0036004807334393263, + 0.006297783460468054, + 0.007251877337694168, + 0.0028598164208233356, + 0.007571659982204437, + 0.005405615549534559, + 0.004217073321342468, + 0.003919563256204128, + 0.002925430191680789, + 0.006523391231894493, + 0.004210003186017275, + 0.0036725413519889116, + 0.007230169605463743, + 0.01105457078665495, + 0.00945688970386982, + 0.010443794541060925, + 0.01088336668908596, + 0.01648261398077011, + 0.015240944921970367, + 0.009792091324925423, + 0.019324420019984245, + 0.012807788327336311, + 0.010639922693371773, + 0.010578088462352753, + 0.008321412838995457, + 0.008463666774332523, + 0.007793725933879614, + 0.0071958680637180805, + 0.009175879880785942, + 0.00767545634880662, + 0.030559368431568146, + 0.0563165619969368, + 0.05113603547215462, + 0.052444156259298325, + 0.07367508858442307, + 0.07087640464305878, + 0.059384725987911224, + 0.057997483760118484, + 0.00790393352508545, + 0.03651486337184906, + 0.01217866875231266, + 0.016439657658338547, + 0.038066815584897995, + 0.014985242858529091, + 0.013182505033910275, + 0.009149286895990372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.0263761505484581 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262364268302917 + ], + "opposite_logprob_tokens": [ + -0.5762364268302917 + ], + "logprob_reward": -0.8262364268302917, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262364268302917, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05420428141951561, + 0.036970846354961395, + 0.05788165330886841, + 0.0468372143805027, + 0.10197541862726212, + 0.027015117928385735, + 0.04710715636610985, + 0.0648719072341919, + 0.34291109442710876, + 0.22022533416748047 + ], + "credit_map": [ + 0.006843878887593746, + 0.006580380257219076, + 0.009663207456469536, + 0.00910637341439724, + 0.004400996491312981, + 0.005775533150881529, + 0.008750200271606445, + 0.003083709394559264, + 0.0026957436930388212, + 0.0036121781449764967, + 0.004526764620095491, + 0.0077994205057621, + 0.0039935423992574215, + 0.004251749254763126, + 0.005423918832093477, + 0.0046675303019583225, + 0.0030185088980942965, + 0.008586104959249496, + 0.0054295179434120655, + 0.0068255458027124405, + 0.012141386047005653, + 0.009720474481582642, + 0.005634041503071785, + 0.0065260715782642365, + 0.0038743328768759966, + 0.003113909624516964, + 0.005290825851261616, + 0.007863839156925678, + 0.005298777483403683, + 0.00898048933595419, + 0.008071660995483398, + 0.004343380220234394, + 0.009620199911296368, + 0.0051702288910746574, + 0.0052039604634046555, + 0.008704044856131077, + 0.0365329384803772, + 0.014686718583106995, + 0.013130633160471916, + 0.008926693350076675, + 0.0026299976743757725, + 0.0027677956968545914, + 0.0021208864636719227, + 0.003887435421347618, + 0.003558384021744132, + 0.002466036006808281, + 0.004378977231681347, + 0.005205604247748852, + 0.0032352854032069445, + 0.005024496000260115, + 0.004194169770926237, + 0.006794178392738104, + 0.00948720145970583, + 0.0049148439429700375, + 0.008033756166696548, + 0.005423227325081825, + 0.00790202897042036, + 0.010154429823160172, + 0.010234436951577663, + 0.006025160662829876, + 0.012442635372281075, + 0.007575123570859432, + 0.005390502512454987, + 0.005147586110979319, + 0.023955021053552628, + 0.03761070594191551, + 0.042815349996089935, + 0.05169802904129028, + 0.044873908162117004, + 0.04416881501674652, + 0.06311323493719101, + 0.03467603400349617, + 0.012568906880915165, + 0.06828425824642181, + 0.015221504494547844, + 0.017617512494325638, + 0.06529829651117325, + 0.013086465187370777, + 0.014721986837685108, + 0.013426396064460278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.8262364268302917 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307892918586731 + ], + "opposite_logprob_tokens": [ + -3.168078899383545 + ], + "logprob_reward": -0.04307892918586731, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307892918586731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043171465396881104, + 0.03669918328523636, + 0.020393693819642067, + 0.06856777518987656, + 0.06128276139497757, + 0.0155336270108819, + 0.12158328294754028, + 0.040109772235155106, + 0.30333733558654785, + 0.2893211841583252 + ], + "credit_map": [ + 0.006576294079422951, + 0.0047393906861543655, + 0.00601870147511363, + 0.0064348503947257996, + 0.0038543385453522205, + 0.005115065723657608, + 0.0066034626215696335, + 0.003829358844086528, + 0.0037009036168456078, + 0.003607526421546936, + 0.005525566637516022, + 0.005716014187783003, + 0.0036077045369893312, + 0.005862927529960871, + 0.0047551365569233894, + 0.003923399839550257, + 0.001970627810806036, + 0.0021478876005858183, + 0.0021754431072622538, + 0.0035115559585392475, + 0.002379019046202302, + 0.0036985864862799644, + 0.0019456663867458701, + 0.0025649061426520348, + 0.006587688811123371, + 0.004189011175185442, + 0.008593380451202393, + 0.00821094959974289, + 0.006143529899418354, + 0.018684085458517075, + 0.009131165221333504, + 0.007027961313724518, + 0.005869764368981123, + 0.003504929831251502, + 0.004470559768378735, + 0.006890479009598494, + 0.01968873105943203, + 0.00796282198280096, + 0.007390221580862999, + 0.005505251232534647, + 0.0015660007484257221, + 0.0011695324210450053, + 0.0014674278208985925, + 0.001951643149368465, + 0.0022357311099767685, + 0.0014201472513377666, + 0.0027396774385124445, + 0.00298346602357924, + 0.010166177526116371, + 0.011996564455330372, + 0.010570116341114044, + 0.020476339384913445, + 0.012496882118284702, + 0.010058672167360783, + 0.027303291484713554, + 0.01851522922515869, + 0.00412305211648345, + 0.005015229806303978, + 0.0038112998008728027, + 0.005081844050437212, + 0.007570511661469936, + 0.005086303222924471, + 0.0046724192798137665, + 0.004749107174575329, + 0.018464719876646996, + 0.03436259552836418, + 0.054552629590034485, + 0.04535747319459915, + 0.032549988478422165, + 0.03659910336136818, + 0.04751887917518616, + 0.03393193334341049, + 0.021613672375679016, + 0.060270894318819046, + 0.02552366815507412, + 0.03522377461194992, + 0.050133686512708664, + 0.030695760622620583, + 0.03999977931380272, + 0.025859905406832695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.04307892918586731 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the building white-walled with black tiles and features horse-head walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dark gray tiles fly off the roof in all directions like black hail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do thick dark brown wooden beams twist and snap under pressure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the building tilt catastrophically to one side before disintegrating into a pile of rubble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a low-angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635815367102623 + ], + "opposite_logprob_tokens": [ + -3.651358127593994 + ], + "logprob_reward": -0.02635815367102623, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635815367102623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19120097160339355, + 0.031504493206739426, + 0.055462177842855453, + 0.1397777944803238, + 0.042302828282117844, + 0.05661443993449211, + 0.17575587332248688, + 0.05419239029288292, + 0.1282798945903778, + 0.12490913271903992 + ], + "credit_map": [ + 0.026503123342990875, + 0.01576000265777111, + 0.04741126671433449, + 0.03144008293747902, + 0.010331329889595509, + 0.022157615050673485, + 0.02517220377922058, + 0.012425354681909084, + 0.0037110564298927784, + 0.0020786665845662355, + 0.004772937390953302, + 0.004077990539371967, + 0.005397013854235411, + 0.005631760228425264, + 0.003281121840700507, + 0.002553946105763316, + 0.00367177277803421, + 0.006973737850785255, + 0.0037061600014567375, + 0.006858282256871462, + 0.012384532950818539, + 0.007045913953334093, + 0.0060846381820738316, + 0.008737143129110336, + 0.016161896288394928, + 0.008616190403699875, + 0.017176711931824684, + 0.01745612919330597, + 0.011140059679746628, + 0.03595234826207161, + 0.019390646368265152, + 0.013883807696402073, + 0.004603582434356213, + 0.002291368553414941, + 0.002705655060708523, + 0.004363680258393288, + 0.013549694791436195, + 0.004298718646168709, + 0.006355918478220701, + 0.004134209360927343, + 0.005676348693668842, + 0.0048940968699753284, + 0.004062385763972998, + 0.008977017365396023, + 0.004841160494834185, + 0.004976659081876278, + 0.00894701573997736, + 0.014239749871194363, + 0.018445245921611786, + 0.01633065566420555, + 0.017309272661805153, + 0.02852514199912548, + 0.026035014539957047, + 0.016681190580129623, + 0.02782093733549118, + 0.024608397856354713, + 0.007339831907302141, + 0.009406918659806252, + 0.007708619348704815, + 0.004947705660015345, + 0.00980775710195303, + 0.006198363844305277, + 0.004036654252558947, + 0.004746544174849987, + 0.008944695815443993, + 0.012680324725806713, + 0.0238017700612545, + 0.013099814765155315, + 0.02072814665734768, + 0.021357044577598572, + 0.013751224614679813, + 0.01391687337309122, + 0.0067195529118180275, + 0.03453543409705162, + 0.0077059571631252766, + 0.011491287499666214, + 0.038546156138181686, + 0.008178085088729858, + 0.01026376336812973, + 0.007468895986676216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.02635815367102623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_6/index_tensor([7044, 7044])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_6/index_tensor([7044, 7044])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7c17a7e58252cbd67624882687647f2e901d50b6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_6/index_tensor([7044, 7044])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, slow-motion shot, shot from a slightly overhead angle, completely immerses the viewer in a black cast-iron pot. Within, a deep, rich, dark red broth simmers fiercely and continuously, its surface coated in a layer of translucent red oil that reflects off alluring highlights under the warm indoor lighting. Countless large, orange-red bubbles burst forth from the pot's depths, darting, colliding, and bursting among the thin, instantly curled and discolored slices of beef, the broth-soaked, verdant vegetable leaves, and the cross-cut shiitake mushrooms. As the bubbles rise, plumes of thick, white steam, mingled with a pungent aroma, rise, forming a fleeting, hazy mist in front of the camera, blurring the background and imbuing the entire scene with a vibrant, hearty, and appetizing dynamic beauty, creating a warm and vibrant dining atmosphere.", + "seed": 25563, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_6/index_7044_seed_25563_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.014532650029286743, + "vlm_logprob": -0.014532650029286743, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.014532650029286743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028286824002861977 + ], + "opposite_logprob_tokens": [ + -3.653286933898926 + ], + "logprob_reward": -0.028286824002861977, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028286824002861977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16886822879314423, + 0.13408046960830688, + 0.040344126522541046, + 0.04336633160710335, + 0.12005914002656937, + 0.23189584910869598, + 0.03561853989958763, + 0.1609066277742386, + 0.03212694451212883, + 0.03273371234536171 + ], + "credit_map": [ + 0.028624558821320534, + 0.02575177326798439, + 0.03320383280515671, + 0.013939806260168552, + 0.016042008996009827, + 0.016357410699129105, + 0.01852315478026867, + 0.01642567664384842, + 0.01889614388346672, + 0.017162714153528214, + 0.032035768032073975, + 0.0171525739133358, + 0.010385049507021904, + 0.013264203444123268, + 0.011988673359155655, + 0.013195334933698177, + 0.0057205623015761375, + 0.007298281881958246, + 0.006058017257601023, + 0.004140533972531557, + 0.003693807404488325, + 0.0037897173315286636, + 0.00434638699516654, + 0.005296815652400255, + 0.006117627955973148, + 0.004920800216495991, + 0.008063002489507198, + 0.005895993672311306, + 0.004212833940982819, + 0.004095062147825956, + 0.004781270865350962, + 0.005279743578284979, + 0.01623726822435856, + 0.021502859890460968, + 0.009423406794667244, + 0.013158186338841915, + 0.012839283794164658, + 0.015093817375600338, + 0.016316311433911324, + 0.01548799965530634, + 0.025620248168706894, + 0.007636914029717445, + 0.006650057155638933, + 0.009822340682148933, + 0.016197508201003075, + 0.09078064560890198, + 0.06374169141054153, + 0.011446458287537098, + 0.0073659420013427734, + 0.005465557333081961, + 0.0042941072024405, + 0.002728352788835764, + 0.0027000647969543934, + 0.006065931171178818, + 0.0034204113762825727, + 0.00357817392796278, + 0.05515133589506149, + 0.01061583124101162, + 0.0072140321135520935, + 0.0044294698163867, + 0.01013303454965353, + 0.06066472828388214, + 0.0054913475178182125, + 0.007206861395388842, + 0.004535383079200983, + 0.0039333440363407135, + 0.004997036885470152, + 0.003017629496753216, + 0.0025044686626642942, + 0.007645133417099714, + 0.003420265158638358, + 0.0020736842416226864, + 0.0047876788303256035, + 0.005976583808660507, + 0.0029719348531216383, + 0.003680236404761672, + 0.003915257751941681, + 0.003670034697279334, + 0.004322741646319628, + 0.0034092420246452093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.028286824002861977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020604612305760384 + ], + "opposite_logprob_tokens": [ + -3.8956046104431152 + ], + "logprob_reward": -0.020604612305760384, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020604612305760384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1480286866426468, + 0.13283394277095795, + 0.02309143729507923, + 0.06249949336051941, + 0.18099407851696014, + 0.2831706404685974, + 0.01881367340683937, + 0.11746285855770111, + 0.019340701401233673, + 0.01376454345881939 + ], + "credit_map": [ + 0.02423545904457569, + 0.029953638091683388, + 0.011248929426074028, + 0.014976647682487965, + 0.01744714565575123, + 0.015688737854361534, + 0.015226129442453384, + 0.019252007827162743, + 0.021686431020498276, + 0.04718110337853432, + 0.0048505826853215694, + 0.00620733667165041, + 0.009047378785908222, + 0.028398923575878143, + 0.007145870942622423, + 0.008316312916576862, + 0.003071186365559697, + 0.002673833165317774, + 0.005438491702079773, + 0.0023236656561493874, + 0.002299232641234994, + 0.002377638593316078, + 0.001750607043504715, + 0.0031567830592393875, + 0.00736186234280467, + 0.011806054972112179, + 0.006202680058777332, + 0.006417686585336924, + 0.008704896084964275, + 0.006158045958727598, + 0.00873730331659317, + 0.007110967300832272, + 0.023714901879429817, + 0.02285032905638218, + 0.01619568280875683, + 0.02367984876036644, + 0.0288184005767107, + 0.019083499908447266, + 0.02500743605196476, + 0.02164398320019245, + 0.03290534391999245, + 0.009094353765249252, + 0.006661109160631895, + 0.011060158722102642, + 0.019814766943454742, + 0.11394580453634262, + 0.07651236653327942, + 0.01317674107849598, + 0.0025491947308182716, + 0.0020490288734436035, + 0.003585432656109333, + 0.001988424453884363, + 0.002092694863677025, + 0.002020901534706354, + 0.002588731935247779, + 0.001939264708198607, + 0.03603145480155945, + 0.00750349648296833, + 0.005476323887705803, + 0.004374186974018812, + 0.008074630983173847, + 0.044057928025722504, + 0.006530814804136753, + 0.0054140230640769005, + 0.0028529600240290165, + 0.003124807495623827, + 0.004630306735634804, + 0.001674526371061802, + 0.001645341282710433, + 0.0017315906006842852, + 0.002191798994317651, + 0.0014893715269863605, + 0.002405014121904969, + 0.0023346184752881527, + 0.0013405353529378772, + 0.0015242410590872169, + 0.0013850631657987833, + 0.0012845799792557955, + 0.0023868305142968893, + 0.0011036607902497053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.020604612305760384 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036204999778419733 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036204999778419733, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036204999778419733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11814545094966888, + 0.062333155423402786, + 0.028253864496946335, + 0.04529739171266556, + 0.1878500133752823, + 0.30218154191970825, + 0.057464975863695145, + 0.14068713784217834, + 0.020404743030667305, + 0.037381723523139954 + ], + "credit_map": [ + 0.02314923144876957, + 0.02124212123453617, + 0.015191932208836079, + 0.015891583636403084, + 0.010424519889056683, + 0.009256675839424133, + 0.011874223127961159, + 0.011115162633359432, + 0.008820047602057457, + 0.013540862128138542, + 0.0045319292694330215, + 0.004159423988312483, + 0.005962352734059095, + 0.012821073643863201, + 0.005022895056754351, + 0.007474572863429785, + 0.004302238579839468, + 0.002950564958155155, + 0.006610262673348188, + 0.0026880991645157337, + 0.0025788352359086275, + 0.002446929458528757, + 0.0033653387799859047, + 0.003311595181003213, + 0.005034035071730614, + 0.00583898788318038, + 0.0030171661637723446, + 0.006013398990035057, + 0.00721252104267478, + 0.004809540696442127, + 0.00789683312177658, + 0.005474912468343973, + 0.021509716287255287, + 0.03507107123732567, + 0.01595633663237095, + 0.017536375671625137, + 0.023248203098773956, + 0.03335491940379143, + 0.02283896505832672, + 0.018334442749619484, + 0.04175912216305733, + 0.009357196278870106, + 0.01041853241622448, + 0.011073037050664425, + 0.021378610283136368, + 0.12183639407157898, + 0.07289116829633713, + 0.013467466458678246, + 0.00670449435710907, + 0.008716519922018051, + 0.0044632344506680965, + 0.008875079452991486, + 0.006343804765492678, + 0.008459021337330341, + 0.007126240525394678, + 0.006776578724384308, + 0.05064414441585541, + 0.009699303656816483, + 0.005244973115622997, + 0.004038108978420496, + 0.008761129342019558, + 0.052695825695991516, + 0.003344933735206723, + 0.006258724723011255, + 0.003100137459114194, + 0.002607486443594098, + 0.004580048844218254, + 0.0018152263946831226, + 0.0017834415193647146, + 0.0018214319134131074, + 0.002964985091239214, + 0.0017319859471172094, + 0.006567893084138632, + 0.005277375690639019, + 0.004665736109018326, + 0.0038133568596094847, + 0.0041616978123784065, + 0.004503172356635332, + 0.004218483809381723, + 0.004174009896814823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.0036204999778419733 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952488165348768 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952488165348768, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952488165348768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11600551754236221, + 0.26139089465141296, + 0.041842591017484665, + 0.08305338770151138, + 0.13392218947410583, + 0.10138294100761414, + 0.03649366274476051, + 0.18612980842590332, + 0.02206745557487011, + 0.017711564898490906 + ], + "credit_map": [ + 0.018184954300522804, + 0.013357489369809628, + 0.013558396138250828, + 0.01794292777776718, + 0.013531118631362915, + 0.010142609477043152, + 0.016448665410280228, + 0.012839358299970627, + 0.03632893040776253, + 0.08147682994604111, + 0.011034848168492317, + 0.014645946212112904, + 0.02056790143251419, + 0.06281538307666779, + 0.014953561127185822, + 0.019567500799894333, + 0.005227108020335436, + 0.004182581324130297, + 0.010165097191929817, + 0.00524577870965004, + 0.004732075612992048, + 0.003697835374623537, + 0.0047906916588544846, + 0.0038014191668480635, + 0.008563121780753136, + 0.010710354894399643, + 0.0055907126516103745, + 0.009501934982836246, + 0.013146044686436653, + 0.009448661468923092, + 0.013995762914419174, + 0.012096800841391087, + 0.019921377301216125, + 0.012873124331235886, + 0.012117188423871994, + 0.012518125586211681, + 0.014972060918807983, + 0.02884562686085701, + 0.01924987882375717, + 0.013424809090793133, + 0.018819335848093033, + 0.003896046429872513, + 0.004364731255918741, + 0.0037732699420303106, + 0.006968008819967508, + 0.04030178487300873, + 0.018341902643442154, + 0.004917866084724665, + 0.005958997178822756, + 0.004775373265147209, + 0.005438472144305706, + 0.00414537126198411, + 0.0025469677057117224, + 0.005794666241854429, + 0.003156428225338459, + 0.004677386488765478, + 0.06295274198055267, + 0.012220025062561035, + 0.007157598156481981, + 0.004801927600055933, + 0.012323056347668171, + 0.07091988623142242, + 0.007169783115386963, + 0.008584801107645035, + 0.0033137521240860224, + 0.003455128986388445, + 0.004668685141950846, + 0.0017409530701115727, + 0.0017652137903496623, + 0.0037722974084317684, + 0.0016757295234128833, + 0.0016756952973082662, + 0.003097888780757785, + 0.00263047544285655, + 0.0020465515553951263, + 0.002310109091922641, + 0.002122719306498766, + 0.0017504070419818163, + 0.0021812100894749165, + 0.0015722044045105577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.005952488165348768 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419882569462061 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.01419882569462061, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419882569462061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24894556403160095, + 0.22566647827625275, + 0.07332202792167664, + 0.040905844420194626, + 0.13653960824012756, + 0.07697989791631699, + 0.07252804934978485, + 0.03380003571510315, + 0.03232293203473091, + 0.05898952856659889 + ], + "credit_map": [ + 0.057007286697626114, + 0.06352650374174118, + 0.02303895354270935, + 0.0294558834284544, + 0.022771812975406647, + 0.01644083298742771, + 0.019815752282738686, + 0.01688855141401291, + 0.04205605387687683, + 0.08314972370862961, + 0.015287552028894424, + 0.0073048886843025684, + 0.01461352501064539, + 0.04554188251495361, + 0.005934101529419422, + 0.011778769083321095, + 0.008695388212800026, + 0.011103381402790546, + 0.014427095651626587, + 0.00653777876868844, + 0.010122719220817089, + 0.00850813090801239, + 0.006495113018900156, + 0.007432427257299423, + 0.0047668227925896645, + 0.006667986046522856, + 0.003742529544979334, + 0.006179735530167818, + 0.004035415593534708, + 0.0044580441899597645, + 0.005456077866256237, + 0.005599238444119692, + 0.018414707854390144, + 0.011160816065967083, + 0.009246988222002983, + 0.011306061409413815, + 0.016723478212952614, + 0.03660508617758751, + 0.018199235200881958, + 0.014883255586028099, + 0.02420656755566597, + 0.00406620604917407, + 0.008178171701729298, + 0.0035431187134236097, + 0.00490654818713665, + 0.02121327817440033, + 0.007413588464260101, + 0.0034524197690188885, + 0.010712617076933384, + 0.010413228534162045, + 0.009274806827306747, + 0.007730518002063036, + 0.006411598064005375, + 0.009451662190258503, + 0.009071596898138523, + 0.009462024085223675, + 0.007187689654529095, + 0.003995512146502733, + 0.003429173957556486, + 0.0025526878889650106, + 0.002715176437050104, + 0.007952938787639141, + 0.002915751188993454, + 0.003051107982173562, + 0.005301326978951693, + 0.004630235489457846, + 0.004804524127393961, + 0.0019328336929902434, + 0.002852988662198186, + 0.007820967584848404, + 0.003046836471185088, + 0.001933220773935318, + 0.010693538002669811, + 0.00826445035636425, + 0.009748191572725773, + 0.0062078023329377174, + 0.005720083601772785, + 0.0057235402055084705, + 0.006863036192953587, + 0.005768891423940659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01419882569462061 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, slow-motion shot, shot from a slightly overhead angle, completely immerses the viewer in a black cast-iron pot. Within, a deep, rich, dark red broth simmers fiercely and continuously, its surface coated in a layer of translucent red oil that reflects off alluring highlights under the warm indoor lighting. Countless large, orange-red bubbles burst forth from the pot's depths, darting, colliding, and bursting among the thin, instantly curled and discolored slices of beef, the broth-soaked, verdant vegetable leaves, and the cross-cut shiitake mushrooms. As the bubbles rise, plumes of thick, white steam, mingled with a pungent aroma, rise, forming a fleeting, hazy mist in front of the camera, blurring the background and imbuing the entire scene with a vibrant, hearty, and appetizing dynamic beauty, creating a warm and vibrant dining atmosphere.", + "seed": 25564, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_6/index_7044_seed_25564_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.088921142462641, + "vlm_logprob": -0.088921142462641, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.088921142462641, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38988062739372253 + ], + "opposite_logprob_tokens": [ + -1.139880657196045 + ], + "logprob_reward": -0.38988062739372253, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38988062739372253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1482062190771103, + 0.08023455739021301, + 0.05477091297507286, + 0.061606910079717636, + 0.13341642916202545, + 0.23342278599739075, + 0.03929826244711876, + 0.07409797608852386, + 0.03644789755344391, + 0.13849806785583496 + ], + "credit_map": [ + 0.021038567647337914, + 0.023855814710259438, + 0.01053952518850565, + 0.011814986355602741, + 0.026870302855968475, + 0.019328424707055092, + 0.018675485625863075, + 0.016083113849163055, + 0.006385567598044872, + 0.004177107475697994, + 0.011331252753734589, + 0.0037493000272661448, + 0.03200423717498779, + 0.015268610790371895, + 0.003704128321260214, + 0.003614358138293028, + 0.00464656762778759, + 0.002056208672001958, + 0.008874342776834965, + 0.0030526358168572187, + 0.02042563632130623, + 0.010185123421251774, + 0.003082508919760585, + 0.002447896869853139, + 0.005006903316825628, + 0.006032904610037804, + 0.0063497573137283325, + 0.012778589501976967, + 0.008224062621593475, + 0.006876753177493811, + 0.010576163418591022, + 0.005761780310422182, + 0.01433640532195568, + 0.01769491657614708, + 0.007042707875370979, + 0.008355849422514439, + 0.0424717515707016, + 0.026932889595627785, + 0.009969927370548248, + 0.006611991208046675, + 0.028463423252105713, + 0.027807805687189102, + 0.024517308920621872, + 0.026863742619752884, + 0.04462340474128723, + 0.02634943462908268, + 0.03417983278632164, + 0.020617837086319923, + 0.004229596350342035, + 0.006550148595124483, + 0.00181120028719306, + 0.002445405116304755, + 0.011808904819190502, + 0.006010481622070074, + 0.0035676793195307255, + 0.002874848898500204, + 0.006824650801718235, + 0.01165454089641571, + 0.0037265298888087273, + 0.00930760521441698, + 0.015792937949299812, + 0.00804950948804617, + 0.007452292367815971, + 0.011289915069937706, + 0.0032355254516005516, + 0.006099790334701538, + 0.0029817698523402214, + 0.0036033897195011377, + 0.005325815640389919, + 0.0059569054283201694, + 0.0066611324436962605, + 0.002583572408184409, + 0.01246435847133398, + 0.007040707860141993, + 0.028442347422242165, + 0.007716997526586056, + 0.019494108855724335, + 0.006494802888482809, + 0.048625655472278595, + 0.0082190977409482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.38988062739372253 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020607998594641685 + ], + "opposite_logprob_tokens": [ + -3.8956079483032227 + ], + "logprob_reward": -0.020607998594641685, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020607998594641685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11264868080615997, + 0.15864033997058868, + 0.04145641624927521, + 0.04904977232217789, + 0.2632482647895813, + 0.09729830920696259, + 0.02393929846584797, + 0.07349817454814911, + 0.0369030125439167, + 0.14331777393817902 + ], + "credit_map": [ + 0.013220548629760742, + 0.015987252816557884, + 0.012163130566477776, + 0.012603343464434147, + 0.017738385125994682, + 0.00929916650056839, + 0.015076974406838417, + 0.016559869050979614, + 0.010121691040694714, + 0.006712835282087326, + 0.028029942885041237, + 0.017735593020915985, + 0.04940977692604065, + 0.02654714323580265, + 0.011366356164216995, + 0.008716999553143978, + 0.004015770275145769, + 0.0032308094669133425, + 0.009013088420033455, + 0.004068868234753609, + 0.007835504598915577, + 0.0036981620360165834, + 0.00671337591484189, + 0.0028808359056711197, + 0.004316515754908323, + 0.004821179434657097, + 0.004849423188716173, + 0.007956035435199738, + 0.007135114166885614, + 0.006304388400167227, + 0.008741859346628189, + 0.004925255198031664, + 0.018735401332378387, + 0.01713811233639717, + 0.013483645394444466, + 0.02535838820040226, + 0.07257726788520813, + 0.055164236575365067, + 0.04303007572889328, + 0.0177611131221056, + 0.01183296274393797, + 0.011748574674129486, + 0.01037557702511549, + 0.010968731716275215, + 0.01786377839744091, + 0.011175369843840599, + 0.014225010760128498, + 0.009108294732868671, + 0.0020674660336226225, + 0.002735875779762864, + 0.0012307899305596948, + 0.0017431634478271008, + 0.008257806301116943, + 0.0034804048482328653, + 0.0028179390355944633, + 0.0016058525070548058, + 0.006693569011986256, + 0.009854570031166077, + 0.003852071473374963, + 0.009459491819143295, + 0.013775075785815716, + 0.00760973896831274, + 0.008234962821006775, + 0.014018683694303036, + 0.0032979724928736687, + 0.004388647619634867, + 0.004198252689093351, + 0.004185469355434179, + 0.006706060376018286, + 0.003800635924562812, + 0.006926727946847677, + 0.0033992468379437923, + 0.012760152108967304, + 0.005733920261263847, + 0.03561043739318848, + 0.007758647203445435, + 0.017998220399022102, + 0.005201310850679874, + 0.05051388591527939, + 0.007741181179881096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.020607998594641685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017218063585460186 + ], + "opposite_logprob_tokens": [ + -6.3767218589782715 + ], + "logprob_reward": -0.0017218063585460186, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017218063585460186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11927516013383865, + 0.11712971329689026, + 0.077248714864254, + 0.06752277165651321, + 0.11608559638261795, + 0.25382477045059204, + 0.020774291828274727, + 0.0784679502248764, + 0.0546434223651886, + 0.09502760320901871 + ], + "credit_map": [ + 0.014718731865286827, + 0.0167560912668705, + 0.011198196560144424, + 0.01633191667497158, + 0.01785142719745636, + 0.011665367521345615, + 0.016756786033511162, + 0.013996642082929611, + 0.01014227420091629, + 0.0066541638225317, + 0.02100876346230507, + 0.00803421437740326, + 0.041137732565402985, + 0.017514267936348915, + 0.0073294672183692455, + 0.005308831576257944, + 0.007963676005601883, + 0.006058008410036564, + 0.012712160125374794, + 0.007447441574186087, + 0.019425276666879654, + 0.008625173009932041, + 0.008857144974172115, + 0.0061598364263772964, + 0.006017016712576151, + 0.007393510080873966, + 0.006228466983884573, + 0.01132498774677515, + 0.009225049987435341, + 0.007367372512817383, + 0.012637038715183735, + 0.007329326588660479, + 0.01602374017238617, + 0.01743064448237419, + 0.0088607557117939, + 0.014323999173939228, + 0.014748590998351574, + 0.017942633479833603, + 0.016774635761976242, + 0.009980599395930767, + 0.030920838937163353, + 0.03083563782274723, + 0.025580300018191338, + 0.028780460357666016, + 0.04975590109825134, + 0.030035605654120445, + 0.03599650785326958, + 0.021919526159763336, + 0.002076180884614587, + 0.003492048243060708, + 0.0020381093490868807, + 0.0026621657889336348, + 0.0034714306239038706, + 0.0016904175281524658, + 0.0032153278589248657, + 0.0021286108531057835, + 0.007189062424004078, + 0.011492391116917133, + 0.003894731169566512, + 0.00885800551623106, + 0.01845272071659565, + 0.00924717914313078, + 0.00848502665758133, + 0.010848834179341793, + 0.005903927143663168, + 0.010392792522907257, + 0.003923931159079075, + 0.005797308869659901, + 0.008014556020498276, + 0.00826278980821371, + 0.008195999078452587, + 0.004152118694037199, + 0.008647143840789795, + 0.0048004756681621075, + 0.020412687212228775, + 0.0059170834720134735, + 0.013642651960253716, + 0.003775538643822074, + 0.0316951647400856, + 0.006136863958090544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0017218063585460186 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203527010977268 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203527010977268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203527010977268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09804506599903107, + 0.112819142639637, + 0.30097126960754395, + 0.05496637895703316, + 0.09276662021875381, + 0.06004665046930313, + 0.05871926620602608, + 0.06779562681913376, + 0.052457284182310104, + 0.10141263902187347 + ], + "credit_map": [ + 0.014858664013445377, + 0.010949447751045227, + 0.011178744956851006, + 0.008027710020542145, + 0.014624855481088161, + 0.011348491534590721, + 0.012636617757380009, + 0.014420544728636742, + 0.008327583782374859, + 0.006503344513475895, + 0.010554849170148373, + 0.016293855383992195, + 0.037645839154720306, + 0.016516301780939102, + 0.01001712866127491, + 0.006960253231227398, + 0.020481375977396965, + 0.012011253274977207, + 0.011739934794604778, + 0.009985974058508873, + 0.12229906022548676, + 0.08279692381620407, + 0.030812164768576622, + 0.010844610631465912, + 0.004250054247677326, + 0.005271898116916418, + 0.006108211353421211, + 0.009967326186597347, + 0.008148152381181717, + 0.006091500166803598, + 0.010335566475987434, + 0.004793672356754541, + 0.006961377803236246, + 0.007351191248744726, + 0.003444286296144128, + 0.005048489198088646, + 0.03349069878458977, + 0.02384893409907818, + 0.008116967044770718, + 0.0045046755112707615, + 0.007594255730509758, + 0.007802652195096016, + 0.006428202148526907, + 0.006836357060819864, + 0.009130134247243404, + 0.007793662138283253, + 0.00857058446854353, + 0.0058908117935061455, + 0.0054925596341490746, + 0.0037782127037644386, + 0.002946617780253291, + 0.005695938132703304, + 0.018296103924512863, + 0.00737124914303422, + 0.011848222464323044, + 0.00329037057235837, + 0.0065791974775493145, + 0.010726856999099255, + 0.006485816091299057, + 0.007951021194458008, + 0.013416597619652748, + 0.006049834191799164, + 0.007618097122758627, + 0.008968218229711056, + 0.004516120068728924, + 0.007809063885360956, + 0.004976258613169193, + 0.00555392075330019, + 0.012174350209534168, + 0.006863507442176342, + 0.005266545806080103, + 0.005297521129250526, + 0.009243813343346119, + 0.0037209081929177046, + 0.02285931631922722, + 0.005881554447114468, + 0.013393383473157883, + 0.004525964614003897, + 0.035904645919799805, + 0.005883061792701483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.014203527010977268 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191752955317497 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.018191752955317497, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191752955317497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09870954602956772, + 0.11813931912183762, + 0.08085349202156067, + 0.07649679481983185, + 0.16016551852226257, + 0.20977677404880524, + 0.025552257895469666, + 0.07650896906852722, + 0.06703610718250275, + 0.08676119893789291 + ], + "credit_map": [ + 0.015376216731965542, + 0.015491130761802197, + 0.008899509906768799, + 0.009483478032052517, + 0.013322007842361927, + 0.013537117280066013, + 0.012154069729149342, + 0.010446028783917427, + 0.009286721237003803, + 0.0051721446216106415, + 0.014399648644030094, + 0.0062703643925487995, + 0.04991024732589722, + 0.023432103917002678, + 0.004940030165016651, + 0.00472806952893734, + 0.005589053966104984, + 0.0037544043734669685, + 0.004925215151160955, + 0.0038907777052372694, + 0.03130797669291496, + 0.02265801653265953, + 0.005780218634754419, + 0.002947825938463211, + 0.005837662145495415, + 0.007062932010740042, + 0.007878667674958706, + 0.015562922693789005, + 0.010737824253737926, + 0.008972721174359322, + 0.013568803668022156, + 0.006875263527035713, + 0.017396263778209686, + 0.021930471062660217, + 0.007127443794161081, + 0.011462199501693249, + 0.051336612552404404, + 0.03048759512603283, + 0.013784199021756649, + 0.006640744861215353, + 0.02511652186512947, + 0.0247941967099905, + 0.02240845561027527, + 0.023773496970534325, + 0.04023626819252968, + 0.023845810443162918, + 0.030561672523617744, + 0.019040366634726524, + 0.0027488521300256252, + 0.0021876152604818344, + 0.0014255306450650096, + 0.002595186699181795, + 0.006309436168521643, + 0.003150071483105421, + 0.004974599462002516, + 0.002160967094823718, + 0.007162775844335556, + 0.01126936636865139, + 0.003974270541220903, + 0.00897390116006136, + 0.01577598601579666, + 0.00848919153213501, + 0.008660304360091686, + 0.012203178368508816, + 0.005997446831315756, + 0.009763524867594242, + 0.006154555361717939, + 0.007681726478040218, + 0.013242928311228752, + 0.0056161354295909405, + 0.011776539497077465, + 0.006803249008953571, + 0.008054343983530998, + 0.005888982210308313, + 0.01524218637496233, + 0.005595030263066292, + 0.01394728198647499, + 0.004151803907006979, + 0.027710532769560814, + 0.006171039771288633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.018191752955317497 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, slow-motion shot, shot from a slightly overhead angle, completely immerses the viewer in a black cast-iron pot. Within, a deep, rich, dark red broth simmers fiercely and continuously, its surface coated in a layer of translucent red oil that reflects off alluring highlights under the warm indoor lighting. Countless large, orange-red bubbles burst forth from the pot's depths, darting, colliding, and bursting among the thin, instantly curled and discolored slices of beef, the broth-soaked, verdant vegetable leaves, and the cross-cut shiitake mushrooms. As the bubbles rise, plumes of thick, white steam, mingled with a pungent aroma, rise, forming a fleeting, hazy mist in front of the camera, blurring the background and imbuing the entire scene with a vibrant, hearty, and appetizing dynamic beauty, creating a warm and vibrant dining atmosphere.", + "seed": 25565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_6/index_7044_seed_25565_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.00888866763561964, + "vlm_logprob": -0.00888866763561964, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.00888866763561964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018274277448654175 + ], + "opposite_logprob_tokens": [ + -4.143274307250977 + ], + "logprob_reward": -0.018274277448654175, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018274277448654175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10148205608129501, + 0.4770088791847229, + 0.028858372941613197, + 0.021528447046875954, + 0.01417449489235878, + 0.023549841716885567, + 0.2128324657678604, + 0.016997648403048515, + 0.07728126645088196, + 0.026286527514457703 + ], + "credit_map": [ + 0.011645421385765076, + 0.012166250497102737, + 0.015759915113449097, + 0.01279571931809187, + 0.00988136138767004, + 0.015445522032678127, + 0.01555642206221819, + 0.008231455460190773, + 0.0489627867937088, + 0.045102816075086594, + 0.07161690294742584, + 0.05787966772913933, + 0.03167429566383362, + 0.03897910565137863, + 0.14477600157260895, + 0.03801731765270233, + 0.0025687271263450384, + 0.006797832436859608, + 0.0027771389577537775, + 0.003784634405747056, + 0.002963508013635874, + 0.0023828514385968447, + 0.004330036696046591, + 0.0032536471262574196, + 0.0017715675057843328, + 0.0027228044345974922, + 0.002733133500441909, + 0.0033113532699644566, + 0.002784373937174678, + 0.002188402693718672, + 0.003882703837007284, + 0.002134108915925026, + 0.001994506921619177, + 0.0017569743795320392, + 0.0014621757436543703, + 0.0021548799704760313, + 0.002393127651885152, + 0.0013465647352859378, + 0.0020881087984889746, + 0.0009781579719856381, + 0.002001230837777257, + 0.006328321062028408, + 0.002343062311410904, + 0.0042771268635988235, + 0.001639544265344739, + 0.0009611919522285461, + 0.0044296481646597385, + 0.0015697175404056907, + 0.025310423225164413, + 0.03057524375617504, + 0.016919897869229317, + 0.021363168954849243, + 0.03252511844038963, + 0.039710476994514465, + 0.023544400930404663, + 0.022883759811520576, + 0.0019381304737180471, + 0.0019528089324012399, + 0.0019543664529919624, + 0.0020689337980002165, + 0.003154592588543892, + 0.0011861093807965517, + 0.0034660750534385443, + 0.0012766311410814524, + 0.004489146173000336, + 0.01830446347594261, + 0.012021961621940136, + 0.011927743442356586, + 0.009884480386972427, + 0.003979995381087065, + 0.00940147414803505, + 0.007272004149854183, + 0.0022627771832048893, + 0.0031537911854684353, + 0.0030981525778770447, + 0.003979361616075039, + 0.003543006256222725, + 0.0023549143224954605, + 0.004437056835740805, + 0.0034574680030345917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.018274277448654175 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052657718770205975 + ], + "opposite_logprob_tokens": [ + -5.255265712738037 + ], + "logprob_reward": -0.0052657718770205975, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052657718770205975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15365983545780182, + 0.1568858027458191, + 0.18988625705242157, + 0.07046658545732498, + 0.052327293902635574, + 0.02489139698445797, + 0.04171407222747803, + 0.025447839871048927, + 0.20566530525684357, + 0.07905559241771698 + ], + "credit_map": [ + 0.013652149587869644, + 0.017582667991518974, + 0.020156733691692352, + 0.026716243475675583, + 0.012221205979585648, + 0.014450637623667717, + 0.03413849323987961, + 0.014741702005267143, + 0.019008612260222435, + 0.021591374650597572, + 0.016228286549448967, + 0.02762993425130844, + 0.020941877737641335, + 0.01713624596595764, + 0.018491139635443687, + 0.01585834100842476, + 0.017609158530831337, + 0.02220001257956028, + 0.031110573559999466, + 0.020496899262070656, + 0.016915084794163704, + 0.013412920758128166, + 0.03516292944550514, + 0.03297867625951767, + 0.005869302432984114, + 0.016474289819598198, + 0.007113955449312925, + 0.009308584034442902, + 0.008082224987447262, + 0.004960275255143642, + 0.013373532332479954, + 0.00528442719951272, + 0.005737556144595146, + 0.008538036607205868, + 0.004726048093289137, + 0.007163121365010738, + 0.007439593318849802, + 0.004745818674564362, + 0.008775540627539158, + 0.005201576743274927, + 0.0028796119149774313, + 0.004363194573670626, + 0.003162059001624584, + 0.003019993659108877, + 0.003026944352313876, + 0.0021033298689872026, + 0.003831904148682952, + 0.0025043594650924206, + 0.003590427339076996, + 0.005335366353392601, + 0.003992252983152866, + 0.005476778838783503, + 0.005956733599305153, + 0.007559026591479778, + 0.005019354168325663, + 0.004784134216606617, + 0.0023878272622823715, + 0.0051779961213469505, + 0.002284090965986252, + 0.0016781098674982786, + 0.008785009384155273, + 0.002007425297051668, + 0.0018744313856586814, + 0.0012529493542388082, + 0.012400350533425808, + 0.05037127062678337, + 0.03524753823876381, + 0.03224829211831093, + 0.023338546976447105, + 0.011836852878332138, + 0.02105233073234558, + 0.019170129671692848, + 0.006266850512474775, + 0.009363986551761627, + 0.011974859051406384, + 0.019653689116239548, + 0.006643910892307758, + 0.0042586978524923325, + 0.012787682935595512, + 0.008105916902422905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.0052657718770205975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011860248632729053 + ], + "opposite_logprob_tokens": [ + -6.751185894012451 + ], + "logprob_reward": -0.0011860248632729053, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011860248632729053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10657400637865067, + 0.1163223460316658, + 0.10418733209371567, + 0.19915886223316193, + 0.07412900030612946, + 0.10558536648750305, + 0.08531530946493149, + 0.029605235904455185, + 0.08188042789697647, + 0.09724216163158417 + ], + "credit_map": [ + 0.010488561354577541, + 0.01406968105584383, + 0.015139542520046234, + 0.018849797546863556, + 0.011402315460145473, + 0.007228536065667868, + 0.020671267062425613, + 0.008724306710064411, + 0.012319116853177547, + 0.015303942374885082, + 0.016520589590072632, + 0.014462553896009922, + 0.011836688034236431, + 0.015605168417096138, + 0.012151462957262993, + 0.018122820183634758, + 0.014231081120669842, + 0.017334138974547386, + 0.013340977020561695, + 0.01271271426230669, + 0.009583065286278725, + 0.00881852675229311, + 0.01860290952026844, + 0.009563922882080078, + 0.015774354338645935, + 0.027719656005501747, + 0.019751837477087975, + 0.0336172915995121, + 0.028169725090265274, + 0.015324411913752556, + 0.03696710616350174, + 0.02183447778224945, + 0.009916253387928009, + 0.014282546006143093, + 0.008525645360350609, + 0.01014501042664051, + 0.006852591875940561, + 0.005397567059844732, + 0.014295479282736778, + 0.004713901784271002, + 0.007958668284118176, + 0.026241274550557137, + 0.010580130852758884, + 0.021978802978992462, + 0.00719077605754137, + 0.00407805759459734, + 0.020450826734304428, + 0.007106829434633255, + 0.008369378745555878, + 0.01134322490543127, + 0.008796912617981434, + 0.011423332616686821, + 0.011212672106921673, + 0.01615823432803154, + 0.00874392781406641, + 0.00926763005554676, + 0.0031542242504656315, + 0.003256149822846055, + 0.003367592580616474, + 0.0039013694040477276, + 0.0055950721725821495, + 0.0028281216509640217, + 0.004610838368535042, + 0.002891868818551302, + 0.005383925512433052, + 0.019304988905787468, + 0.01257411390542984, + 0.013111493550240993, + 0.0099557563662529, + 0.005095191299915314, + 0.008850651793181896, + 0.007604304235428572, + 0.006454637739807367, + 0.011988786980509758, + 0.015481793321669102, + 0.026431189849972725, + 0.007396341767162085, + 0.004393036011606455, + 0.014082140289247036, + 0.011014237068593502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.0011860248632729053 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078160256147385 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.011078160256147385, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078160256147385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07995041459798813, + 0.4176960587501526, + 0.08936704695224762, + 0.0344814658164978, + 0.01950208470225334, + 0.09196504950523376, + 0.041908275336027145, + 0.01072214636951685, + 0.17688381671905518, + 0.03752367943525314 + ], + "credit_map": [ + 0.009174037724733353, + 0.011399336159229279, + 0.012475905939936638, + 0.013284332118928432, + 0.008781863376498222, + 0.006728370673954487, + 0.009767672047019005, + 0.00833890214562416, + 0.04568558558821678, + 0.03597886115312576, + 0.056195300072431564, + 0.05152209848165512, + 0.030084669589996338, + 0.03835471719503403, + 0.11863195151090622, + 0.04124286770820618, + 0.008974127471446991, + 0.01753311976790428, + 0.011372336186468601, + 0.012980126775801182, + 0.009873527102172375, + 0.008664170280098915, + 0.010217208415269852, + 0.009752437472343445, + 0.0036533637903630733, + 0.007816063240170479, + 0.003071892773732543, + 0.004800822585821152, + 0.003943952731788158, + 0.0026586304884403944, + 0.005789697170257568, + 0.002747046062722802, + 0.002640842692926526, + 0.0029597438406199217, + 0.0019025004003196955, + 0.002313794568181038, + 0.0021651117131114006, + 0.0028482526540756226, + 0.0026951429899781942, + 0.001976695843040943, + 0.007729232311248779, + 0.029726332053542137, + 0.005326674785465002, + 0.007741947658360004, + 0.008011805824935436, + 0.005152822006493807, + 0.021889813244342804, + 0.006386412773281336, + 0.003943548537790775, + 0.004763423930853605, + 0.00280841626226902, + 0.006192093249410391, + 0.006025880109518766, + 0.006709970999509096, + 0.006699081510305405, + 0.004765861667692661, + 0.0010843805503100157, + 0.001253326889127493, + 0.0008593973470851779, + 0.0011092047207057476, + 0.003198317950591445, + 0.0008203804609365761, + 0.0011824154062196612, + 0.0012147220550104976, + 0.011939997784793377, + 0.038208138197660446, + 0.0244536641985178, + 0.03283476084470749, + 0.021012423560023308, + 0.009733014740049839, + 0.020691588521003723, + 0.018010221421718597, + 0.004013505298644304, + 0.004518858622759581, + 0.0042665558867156506, + 0.004781723953783512, + 0.0061693680472671986, + 0.0031913432758301497, + 0.006300111301243305, + 0.004282214678823948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.011078160256147385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863910373300314 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.00863910373300314, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863910373300314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33811381459236145, + 0.10196714103221893, + 0.14590273797512054, + 0.04872797057032585, + 0.025210972875356674, + 0.03857455030083656, + 0.13291935622692108, + 0.0582469217479229, + 0.049852531403303146, + 0.060484059154987335 + ], + "credit_map": [ + 0.027459539473056793, + 0.05880046635866165, + 0.04256304353475571, + 0.06040038540959358, + 0.044373296201229095, + 0.03425510600209236, + 0.049562204629182816, + 0.02069975435733795, + 0.0067590526305139065, + 0.011744967661798, + 0.014682216569781303, + 0.021754717454314232, + 0.011655845679342747, + 0.010576366446912289, + 0.012336257845163345, + 0.012457695789635181, + 0.01298388559371233, + 0.05009905621409416, + 0.012668348848819733, + 0.016380220651626587, + 0.015776345506310463, + 0.009230184368789196, + 0.02107381634414196, + 0.007690854370594025, + 0.004232900682836771, + 0.007529725320637226, + 0.005448475480079651, + 0.0078350231051445, + 0.006622890941798687, + 0.004049662034958601, + 0.007880163379013538, + 0.005129124037921429, + 0.0034710646141320467, + 0.0038714564871042967, + 0.003069621743634343, + 0.0030473745428025723, + 0.0032493374310433865, + 0.0018616756424307823, + 0.0046107713133096695, + 0.002029667841270566, + 0.00333807454444468, + 0.005892057437449694, + 0.004472062923014164, + 0.007538106292486191, + 0.004472864791750908, + 0.0029884350951761007, + 0.0064876568503677845, + 0.003385290503501892, + 0.014741409569978714, + 0.017078377306461334, + 0.012992100790143013, + 0.014328202232718468, + 0.018856719136238098, + 0.02563222497701645, + 0.013842051848769188, + 0.015448247082531452, + 0.00581395672634244, + 0.00960338581353426, + 0.00469982111826539, + 0.004531608894467354, + 0.020798921585083008, + 0.004672994837164879, + 0.004837228450924158, + 0.0032889964058995247, + 0.0033797593787312508, + 0.007539308629930019, + 0.00672112125903368, + 0.010242974385619164, + 0.009002452716231346, + 0.002280036685988307, + 0.006325620226562023, + 0.004361245781183243, + 0.0064736888743937016, + 0.010724752210080624, + 0.009098073467612267, + 0.006632015574723482, + 0.007262094411998987, + 0.004371301271021366, + 0.009916706010699272, + 0.006005418952554464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.00863910373300314 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, slow-motion shot, shot from a slightly overhead angle, completely immerses the viewer in a black cast-iron pot. Within, a deep, rich, dark red broth simmers fiercely and continuously, its surface coated in a layer of translucent red oil that reflects off alluring highlights under the warm indoor lighting. Countless large, orange-red bubbles burst forth from the pot's depths, darting, colliding, and bursting among the thin, instantly curled and discolored slices of beef, the broth-soaked, verdant vegetable leaves, and the cross-cut shiitake mushrooms. As the bubbles rise, plumes of thick, white steam, mingled with a pungent aroma, rise, forming a fleeting, hazy mist in front of the camera, blurring the background and imbuing the entire scene with a vibrant, hearty, and appetizing dynamic beauty, creating a warm and vibrant dining atmosphere.", + "seed": 25566, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_6/index_7044_seed_25566_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015544129069894552, + "vlm_logprob": -0.015544129069894552, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.015544129069894552, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03692268580198288 + ], + "opposite_logprob_tokens": [ + -3.4119226932525635 + ], + "logprob_reward": -0.03692268580198288, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03692268580198288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0856279581785202, + 0.04400496929883957, + 0.051209669560194016, + 0.05618143081665039, + 0.05306312069296837, + 0.038611043244600296, + 0.5392832159996033, + 0.04032235965132713, + 0.03393881395459175, + 0.05775739997625351 + ], + "credit_map": [ + 0.008587315678596497, + 0.012231908738613129, + 0.007644854951649904, + 0.010461326688528061, + 0.017623526975512505, + 0.00963760819286108, + 0.011194156482815742, + 0.008247260004281998, + 0.00570465624332428, + 0.005669484846293926, + 0.004268134944140911, + 0.006238431204110384, + 0.005585110280662775, + 0.004941036459058523, + 0.005572850815951824, + 0.006025264505296946, + 0.007303531281650066, + 0.0061705997213721275, + 0.006993018090724945, + 0.0073351566679775715, + 0.00517360819503665, + 0.005349271930754185, + 0.007122686132788658, + 0.005761794280260801, + 0.010154915042221546, + 0.00602872995659709, + 0.006537886336445808, + 0.008482755161821842, + 0.005908655934035778, + 0.00628765719011426, + 0.007444217801094055, + 0.005336611531674862, + 0.005432403180748224, + 0.006254902109503746, + 0.0054855975322425365, + 0.005921178963035345, + 0.005908599589020014, + 0.009726696647703648, + 0.008563993498682976, + 0.005769752897322178, + 0.004786221776157618, + 0.00584620563313365, + 0.0031383924651890993, + 0.0052078887820243835, + 0.006141353864222765, + 0.004592019598931074, + 0.005697146058082581, + 0.003201815765351057, + 0.05223265290260315, + 0.06569316983222961, + 0.054537929594516754, + 0.0667971670627594, + 0.10104051232337952, + 0.06768366694450378, + 0.06439877301454544, + 0.06689934432506561, + 0.008781246840953827, + 0.005808527115732431, + 0.00508874049410224, + 0.0036758477799594402, + 0.0034235925413668156, + 0.0055202459916472435, + 0.003366457996889949, + 0.004657698329538107, + 0.0039765359833836555, + 0.004266695119440556, + 0.003392116166651249, + 0.004161941818892956, + 0.0069349645636975765, + 0.003385350340977311, + 0.004127784166485071, + 0.003693425329402089, + 0.010755198076367378, + 0.008611870929598808, + 0.007611970882862806, + 0.008312517777085304, + 0.00705764489248395, + 0.005373260006308556, + 0.00489004934206605, + 0.00514488760381937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.03692268580198288 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004659746307879686 + ], + "opposite_logprob_tokens": [ + -5.379659652709961 + ], + "logprob_reward": -0.004659746307879686, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004659746307879686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1204170435667038, + 0.31101927161216736, + 0.06699006259441376, + 0.017961984500288963, + 0.05095551535487175, + 0.06460641324520111, + 0.11379951983690262, + 0.11781962960958481, + 0.048852283507585526, + 0.08757825940847397 + ], + "credit_map": [ + 0.014172513037919998, + 0.014898371882736683, + 0.013025646097958088, + 0.015965502709150314, + 0.020925981923937798, + 0.01657700166106224, + 0.012378809042274952, + 0.012473214417696, + 0.03185097128152847, + 0.03603937476873398, + 0.03498881310224533, + 0.051014870405197144, + 0.04669836536049843, + 0.03554887697100639, + 0.03656381368637085, + 0.03831420838832855, + 0.007782805245369673, + 0.007569035980850458, + 0.006234490312635899, + 0.011705008335411549, + 0.007859400473535061, + 0.006246402394026518, + 0.012061497196555138, + 0.007531429640948772, + 0.0027864519506692886, + 0.0020200293511152267, + 0.0016117041232064366, + 0.002128581516444683, + 0.002763007301837206, + 0.0019667639862746, + 0.0028415678534656763, + 0.0018438788829371333, + 0.006286598276346922, + 0.006360982544720173, + 0.007293886039406061, + 0.005382901523262262, + 0.007519759703427553, + 0.0060736555606126785, + 0.00692229438573122, + 0.005115439649671316, + 0.006217779126018286, + 0.008108140900731087, + 0.007977607659995556, + 0.008430859073996544, + 0.010383046232163906, + 0.0071394555270671844, + 0.01061797421425581, + 0.005731554701924324, + 0.014852984808385372, + 0.014045121148228645, + 0.015183786861598492, + 0.013181318528950214, + 0.017657455056905746, + 0.013506432063877583, + 0.01172503363341093, + 0.013647385872900486, + 0.023934381082654, + 0.014952294528484344, + 0.010439408011734486, + 0.01352501567453146, + 0.017229188233613968, + 0.013959677889943123, + 0.010692372918128967, + 0.01308728288859129, + 0.005797252990305424, + 0.006281763780862093, + 0.004143259953707457, + 0.00611252011731267, + 0.008190218359231949, + 0.005534389056265354, + 0.008020269684493542, + 0.004772610031068325, + 0.02278839983046055, + 0.022925283759832382, + 0.0077513703145086765, + 0.006520479917526245, + 0.01001668255776167, + 0.005409078672528267, + 0.006267268676310778, + 0.005899694282561541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.004659746307879686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013431822881102562 + ], + "opposite_logprob_tokens": [ + -6.626343250274658 + ], + "logprob_reward": -0.0013431822881102562, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013431822881102562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1892780363559723, + 0.07134009897708893, + 0.03803062066435814, + 0.12107276916503906, + 0.07843936234712601, + 0.05301419645547867, + 0.17609138786792755, + 0.14377009868621826, + 0.0740744099020958, + 0.05488897114992142 + ], + "credit_map": [ + 0.022007785737514496, + 0.021275673061609268, + 0.019310642033815384, + 0.0291327852755785, + 0.026040762662887573, + 0.022254353389143944, + 0.030244389548897743, + 0.019011642783880234, + 0.011485843919217587, + 0.009375258348882198, + 0.010681575164198875, + 0.010663129389286041, + 0.00794542208313942, + 0.006222805939614773, + 0.007905551232397556, + 0.007060511969029903, + 0.004913804586976767, + 0.005710281431674957, + 0.004745354410260916, + 0.005607291590422392, + 0.00511588342487812, + 0.0034227243158966303, + 0.005591760855168104, + 0.002923524472862482, + 0.014079510234296322, + 0.013934570364654064, + 0.01379425823688507, + 0.018116561695933342, + 0.016858713701367378, + 0.01381905097514391, + 0.0168608445674181, + 0.013609262183308601, + 0.008573121391236782, + 0.0098666837438941, + 0.010513820685446262, + 0.00890850368887186, + 0.009110255166888237, + 0.013800006359815598, + 0.008807363919913769, + 0.00885960552841425, + 0.006131916306912899, + 0.005599589087069035, + 0.007248895708471537, + 0.006950026843696833, + 0.00623974809423089, + 0.006098926067352295, + 0.009983106516301632, + 0.004761986434459686, + 0.01890161819756031, + 0.021435841917991638, + 0.017769113183021545, + 0.021833106875419617, + 0.031612906605005264, + 0.02261742390692234, + 0.02072848007082939, + 0.021192893385887146, + 0.013660061173141003, + 0.020322374999523163, + 0.00906788744032383, + 0.016603389754891396, + 0.03381625935435295, + 0.016793910413980484, + 0.01697036437690258, + 0.01653585396707058, + 0.009251357987523079, + 0.010032866150140762, + 0.007479677442461252, + 0.008300204761326313, + 0.01605643890798092, + 0.0069212401285767555, + 0.009714043699204922, + 0.006318577099591494, + 0.010703501291573048, + 0.012263527140021324, + 0.0062661077827215195, + 0.007285033352673054, + 0.005749684292823076, + 0.004233696963638067, + 0.004359154030680656, + 0.004028265364468098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0013431822881102562 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595969632267952 + ], + "opposite_logprob_tokens": [ + -3.8955960273742676 + ], + "logprob_reward": -0.020595969632267952, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595969632267952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19411610066890717, + 0.08847782760858536, + 0.07387138903141022, + 0.02834557369351387, + 0.05171662196516991, + 0.11275056004524231, + 0.10971614718437195, + 0.18763871490955353, + 0.09969057142734528, + 0.0536765418946743 + ], + "credit_map": [ + 0.016018692404031754, + 0.02853742055594921, + 0.019604984670877457, + 0.022959740832448006, + 0.04268113151192665, + 0.016957174986600876, + 0.02952386811375618, + 0.017833085730671883, + 0.01311890222132206, + 0.01189185120165348, + 0.012888846918940544, + 0.010758083313703537, + 0.011019183322787285, + 0.010627425275743008, + 0.00959647074341774, + 0.008577069267630577, + 0.009211846627295017, + 0.012027020566165447, + 0.0059599727392196655, + 0.009561950340867043, + 0.01105392538011074, + 0.009319273754954338, + 0.00977566372603178, + 0.00696173682808876, + 0.004958901088684797, + 0.003162563545629382, + 0.0031425761990249157, + 0.0037601334042847157, + 0.003231134731322527, + 0.0040511018596589565, + 0.003289590124040842, + 0.0027495739050209522, + 0.005380807910114527, + 0.006190976593643427, + 0.00553248031064868, + 0.006922656204551458, + 0.006235065869987011, + 0.008788901381194592, + 0.007194153033196926, + 0.005471581127494574, + 0.01257714256644249, + 0.015771841630339622, + 0.012272964231669903, + 0.015557396225631237, + 0.01596122793853283, + 0.010848416946828365, + 0.020259669050574303, + 0.009501894935965538, + 0.011616168543696404, + 0.012625166215002537, + 0.010985687375068665, + 0.013377103954553604, + 0.01983104646205902, + 0.012879172340035439, + 0.014986376278102398, + 0.013415436260402203, + 0.01418905146420002, + 0.026131005957722664, + 0.010685092769563198, + 0.020946448668837547, + 0.048096828162670135, + 0.02260674722492695, + 0.022722415626049042, + 0.02226112224161625, + 0.011263038963079453, + 0.014577601104974747, + 0.006929883733391762, + 0.009013489820063114, + 0.028119385242462158, + 0.008820410817861557, + 0.011997034773230553, + 0.008969727903604507, + 0.010820961557328701, + 0.011552279815077782, + 0.005582122597843409, + 0.006739455740898848, + 0.006666822824627161, + 0.00419645756483078, + 0.0043068695813417435, + 0.0038115764036774635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.020595969632267952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a black cast-iron pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the broth inside the pot dark red with a layer of translucent red oil on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large orange-red bubbles bursting and rising in the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin slices of beef, verdant vegetable leaves, and cross-cut shiitake mushrooms visible in the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199061319231987 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199061319231987, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199061319231987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14375148713588715, + 0.06876130402088165, + 0.048456672579050064, + 0.11475321650505066, + 0.039979178458452225, + 0.029028277844190598, + 0.24213044345378876, + 0.050022728741168976, + 0.07094249129295349, + 0.1921742856502533 + ], + "credit_map": [ + 0.015831969678401947, + 0.01839018054306507, + 0.014143474400043488, + 0.02216949313879013, + 0.022816939279437065, + 0.015041621401906013, + 0.021387923508882523, + 0.013969875872135162, + 0.005646811798214912, + 0.008246448822319508, + 0.00880387146025896, + 0.011574795469641685, + 0.011641325429081917, + 0.007895693182945251, + 0.007055679801851511, + 0.007896670140326023, + 0.006178546231240034, + 0.005187381990253925, + 0.0049700019881129265, + 0.007681297138333321, + 0.00619468605145812, + 0.004527876619249582, + 0.008270428515970707, + 0.005446446128189564, + 0.018901009112596512, + 0.012767976149916649, + 0.011424703523516655, + 0.01685478165745735, + 0.013672730885446072, + 0.01151217706501484, + 0.017507445067167282, + 0.012112380005419254, + 0.005647582467645407, + 0.004267271142452955, + 0.004836100619286299, + 0.0038046909030526876, + 0.0040858713909983635, + 0.007339603267610073, + 0.005154200829565525, + 0.00484385434538126, + 0.0035751627292484045, + 0.004047058988362551, + 0.004460385534912348, + 0.0036217933520674706, + 0.003497401950880885, + 0.003504455555230379, + 0.0032009759452193975, + 0.0031210409943014383, + 0.026293691247701645, + 0.02974061295390129, + 0.02331116609275341, + 0.028538264334201813, + 0.04520631954073906, + 0.030167244374752045, + 0.028583670035004616, + 0.030289463698863983, + 0.004932841286063194, + 0.007129492238163948, + 0.0037483659107238054, + 0.004888959228992462, + 0.012135099619626999, + 0.005901174154132605, + 0.005928826052695513, + 0.005357965361326933, + 0.007212942466139793, + 0.010958089493215084, + 0.005160577595233917, + 0.008480087853968143, + 0.019405096769332886, + 0.007056598085910082, + 0.0065084174275398254, + 0.006160680204629898, + 0.039407193660736084, + 0.04395409673452377, + 0.02062942460179329, + 0.024924490600824356, + 0.021594412624835968, + 0.014682603999972343, + 0.014281029812991619, + 0.012701029889285564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.014199061319231987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_tensor([1033, 1033])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_tensor([1033, 1033])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..46cbd1cd068809ac5f98d95a6505a7e0c1e2df9a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_tensor([1033, 1033])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot closely follows a sleek fighter jet, painted matte dark gray, with sharp angles and a futuristic design, as it flies at breakneck speed across the vast desert floor. Illuminated by the golden evening sun, the rolling sand dunes cast long, dramatic shadows, their delicate, wind-eroded textures clearly visible in the slanting light. The fighter jet flies extremely low over the sharp ridges of the dunes, its fuselage mere meters above the sand. The tips of its swept wings violently slice through the air at high speed, creating two distinct, intertwining, ribbon-like vortices from the immense pressure differential. These vortices, clinging tightly to the wingtips, tumble and stretch behind the fighter, briefly outlining its trajectory before dissipating into the air distorted by the engine's heatwave, the swirling sand particles forming a fleeting dust trail. The entire picture is full of extreme sense of speed, precise control and awe-inspiring mechanical aesthetics, forming a strong visual contrast between the magnificent natural scenery in warm colors and the cold cutting-edge technology, creating a tense and magnificent atmosphere.", + "seed": 16461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_1033_seed_16461_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0212995330337435, + "vlm_logprob": -0.0212995330337435, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.0212995330337435, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026103148236870766 + ], + "opposite_logprob_tokens": [ + -3.7761032581329346 + ], + "logprob_reward": -0.026103148236870766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026103148236870766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0685306042432785, + 0.0610179677605629, + 0.07600370794534683, + 0.05702469125390053, + 0.3468337953090668, + 0.12989874184131622, + 0.12466463446617126, + 0.03470078855752945, + 0.03462221100926399, + 0.06670286506414413 + ], + "credit_map": [ + 0.00622443575412035, + 0.009370016865432262, + 0.007952642627060413, + 0.008280551061034203, + 0.010700437240302563, + 0.011082335375249386, + 0.007899109274148941, + 0.007021073251962662, + 0.004337449092417955, + 0.007019868120551109, + 0.01848761923611164, + 0.0043587833642959595, + 0.003842846257612109, + 0.005219329614192247, + 0.003677056171000004, + 0.014075015671551228, + 0.008805775083601475, + 0.009980211965739727, + 0.00930328480899334, + 0.00788009725511074, + 0.014318996109068394, + 0.009801519103348255, + 0.010230110958218575, + 0.0056837149895727634, + 0.006244756281375885, + 0.005411663558334112, + 0.006437642965465784, + 0.010952724143862724, + 0.006485336925834417, + 0.007093543186783791, + 0.009015006013214588, + 0.005384016316384077, + 0.039211541414260864, + 0.024276014417409897, + 0.023831583559513092, + 0.032896436750888824, + 0.08980032801628113, + 0.035972967743873596, + 0.021657897159457207, + 0.07918700575828552, + 0.011725891381502151, + 0.011868634261190891, + 0.014797158539295197, + 0.018545111641287804, + 0.016605114564299583, + 0.013624484650790691, + 0.02122092992067337, + 0.021511418744921684, + 0.01224316656589508, + 0.02086653932929039, + 0.012411637231707573, + 0.01653137058019638, + 0.01265490148216486, + 0.011339131742715836, + 0.025905486196279526, + 0.01271240133792162, + 0.005518451798707247, + 0.005498331505805254, + 0.003695437218993902, + 0.002836743136867881, + 0.0033855338115245104, + 0.004743055906146765, + 0.0034929844550788403, + 0.00553025072440505, + 0.005812039133161306, + 0.005402756854891777, + 0.003049359656870365, + 0.0027865173760801554, + 0.00679483637213707, + 0.004302558489143848, + 0.003105355193838477, + 0.003368786768987775, + 0.005735230632126331, + 0.007458042819052935, + 0.010249496437609196, + 0.008347541093826294, + 0.010142059996724129, + 0.008296379819512367, + 0.009517925791442394, + 0.006956188939511776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.026103148236870766 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642790675163269 + ], + "opposite_logprob_tokens": [ + -4.88264274597168 + ], + "logprob_reward": -0.007642790675163269, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642790675163269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06589917093515396, + 0.03857271745800972, + 0.18655073642730713, + 0.03248069807887077, + 0.2169724553823471, + 0.08374468982219696, + 0.04694807901978493, + 0.03318750113248825, + 0.03000470995903015, + 0.2656392753124237 + ], + "credit_map": [ + 0.006134110502898693, + 0.009853942319750786, + 0.010931247845292091, + 0.009188272058963776, + 0.006410686764866114, + 0.006637852173298597, + 0.009025110863149166, + 0.007717946078628302, + 0.003000478260219097, + 0.004717652220278978, + 0.006490321364253759, + 0.0055158971808850765, + 0.003543025115504861, + 0.006599122192710638, + 0.003980179782956839, + 0.004726035054773092, + 0.008516150526702404, + 0.006009116303175688, + 0.007556176744401455, + 0.011309939436614513, + 0.011994774453341961, + 0.06273096054792404, + 0.06819859892129898, + 0.010235000401735306, + 0.0038973866030573845, + 0.0035123974084854126, + 0.0037342095747590065, + 0.0057602436281740665, + 0.003567991079762578, + 0.00332952500320971, + 0.005456807091832161, + 0.0032221367582678795, + 0.028755702078342438, + 0.01726623810827732, + 0.0172378309071064, + 0.03325236588716507, + 0.050978075712919235, + 0.021786611527204514, + 0.021122904494404793, + 0.02657272480428219, + 0.006814045365899801, + 0.008403874933719635, + 0.010913277976214886, + 0.010345680639147758, + 0.011693168431520462, + 0.01224549487233162, + 0.012189721688628197, + 0.011139417998492718, + 0.005253395531326532, + 0.005571458488702774, + 0.005428878590464592, + 0.005986662115901709, + 0.006859797053039074, + 0.004111439920961857, + 0.009031732566654682, + 0.004704711027443409, + 0.0048933252692222595, + 0.0044150687754154205, + 0.003768520662561059, + 0.0028081873897463083, + 0.0027359144296497107, + 0.0038633972872048616, + 0.003915890119969845, + 0.006787197198718786, + 0.003843654179945588, + 0.004251715261489153, + 0.0029229032807052135, + 0.00376971997320652, + 0.005331546068191528, + 0.003303127596154809, + 0.0031868580263108015, + 0.0033951851073652506, + 0.02156510017812252, + 0.0377795435488224, + 0.03486600145697594, + 0.04319765791296959, + 0.03965066000819206, + 0.03237657994031906, + 0.029836591333150864, + 0.026367133483290672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.007642790675163269 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054947979748249054 + ], + "opposite_logprob_tokens": [ + -2.929948091506958 + ], + "logprob_reward": -0.054947979748249054, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054947979748249054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.066773422062397, + 0.06803058087825775, + 0.40649840235710144, + 0.04158226028084755, + 0.1739814728498459, + 0.04816703125834465, + 0.019364075735211372, + 0.07939586043357849, + 0.05938776955008507, + 0.03681914508342743 + ], + "credit_map": [ + 0.006354367360472679, + 0.01229066401720047, + 0.008261723443865776, + 0.006514035165309906, + 0.01118964608758688, + 0.0070295934565365314, + 0.004946775268763304, + 0.010186619125306606, + 0.005358492955565453, + 0.007354267872869968, + 0.01735357753932476, + 0.007162665482610464, + 0.004793323576450348, + 0.006339794024825096, + 0.004678593948483467, + 0.014989866875112057, + 0.021974045783281326, + 0.015956971794366837, + 0.01481960341334343, + 0.024756934493780136, + 0.02898704819381237, + 0.12712568044662476, + 0.146345853805542, + 0.02653222344815731, + 0.003750320989638567, + 0.0029862895607948303, + 0.005168725736439228, + 0.007205924019217491, + 0.005857332609593868, + 0.005245446693152189, + 0.006772490218281746, + 0.004595729522407055, + 0.01596442610025406, + 0.01654435507953167, + 0.016467677429318428, + 0.030115285888314247, + 0.03858008235692978, + 0.014478247612714767, + 0.01973767578601837, + 0.02209373004734516, + 0.004088213667273521, + 0.004889178089797497, + 0.005840334575623274, + 0.006468226667493582, + 0.005933963693678379, + 0.006522275973111391, + 0.007651484105736017, + 0.006773351691663265, + 0.0028080418705940247, + 0.0019867352675646544, + 0.002559045096859336, + 0.0020595802925527096, + 0.00259451218880713, + 0.002372347516939044, + 0.003113494487479329, + 0.0018703181995078921, + 0.011707902885973454, + 0.01028794888406992, + 0.007683899719268084, + 0.006424571853131056, + 0.006850333884358406, + 0.010131930001080036, + 0.009351127780973911, + 0.016958149150013924, + 0.00989498384296894, + 0.009265112690627575, + 0.0060659185983240604, + 0.005782837048172951, + 0.00837757159024477, + 0.007073610555380583, + 0.006510103587061167, + 0.006417636759579182, + 0.003446287242695689, + 0.004538927227258682, + 0.0052962289191782475, + 0.006519886199384928, + 0.004286173265427351, + 0.004001412075012922, + 0.004341550171375275, + 0.004388683941215277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.054947979748249054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036143234465271235 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036143234465271235, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036143234465271235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.063230961561203, + 0.07329924404621124, + 0.48468899726867676, + 0.07173468917608261, + 0.08140099793672562, + 0.0284737441688776, + 0.05145122483372688, + 0.06106233596801758, + 0.03740623965859413, + 0.047251541167497635 + ], + "credit_map": [ + 0.006161182187497616, + 0.007500697858631611, + 0.009493724443018436, + 0.007914931513369083, + 0.009741239249706268, + 0.0072342571802437305, + 0.005880242679268122, + 0.009304684586822987, + 0.004500005394220352, + 0.004283216781914234, + 0.02009395696222782, + 0.00783940963447094, + 0.006593739148229361, + 0.00933628436177969, + 0.005551671143621206, + 0.015100963413715363, + 0.022109918296337128, + 0.011893393471837044, + 0.012653796933591366, + 0.02860684134066105, + 0.029747357591986656, + 0.16680379211902618, + 0.1852227747440338, + 0.02765113115310669, + 0.007273340132087469, + 0.006248690187931061, + 0.009532562457025051, + 0.011059005744755268, + 0.011411871761083603, + 0.00959783885627985, + 0.010676686652004719, + 0.005934686399996281, + 0.010799639858305454, + 0.005958318244665861, + 0.00744392815977335, + 0.009230060502886772, + 0.016865069046616554, + 0.011625970713794231, + 0.006614771205931902, + 0.01286324206739664, + 0.002949689282104373, + 0.002089025219902396, + 0.003928337711840868, + 0.0030139104928821325, + 0.002948624547570944, + 0.005304133519530296, + 0.003488551126793027, + 0.004751470871269703, + 0.008482473902404308, + 0.006278626620769501, + 0.007190844509750605, + 0.0036581000313162804, + 0.0082045067101717, + 0.008131511509418488, + 0.006232854910194874, + 0.0032723071053624153, + 0.006501029245555401, + 0.005857856944203377, + 0.017151392996311188, + 0.0037695802748203278, + 0.0038113256450742483, + 0.004940344486385584, + 0.004326868802309036, + 0.014703941531479359, + 0.005185981746762991, + 0.0055369115434587, + 0.003243246814236045, + 0.003860432654619217, + 0.00812721811234951, + 0.0038979798555374146, + 0.0036161753814667463, + 0.003938291687518358, + 0.005336635280400515, + 0.004919607657939196, + 0.0066431001760065556, + 0.0038359418977051973, + 0.0072104912251234055, + 0.006320290267467499, + 0.0074532125145196915, + 0.005532263312488794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.0036143234465271235 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189423061907291 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189423061907291, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189423061907291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16375766694545746, + 0.06314075738191605, + 0.1285659521818161, + 0.050029486417770386, + 0.14826087653636932, + 0.04735949635505676, + 0.026777364313602448, + 0.25988200306892395, + 0.07027019560337067, + 0.04195617884397507 + ], + "credit_map": [ + 0.014200975187122822, + 0.01661381870508194, + 0.03277596831321716, + 0.024011382833123207, + 0.018526162952184677, + 0.013647280633449554, + 0.01840723305940628, + 0.02557484619319439, + 0.005166081245988607, + 0.0064635081216692924, + 0.017748665064573288, + 0.004823659081012011, + 0.003526929533109069, + 0.004862121772021055, + 0.0047037978656589985, + 0.015846000984311104, + 0.010721112601459026, + 0.010642685927450657, + 0.007491392083466053, + 0.009114578366279602, + 0.01748993992805481, + 0.026230260729789734, + 0.03716801106929779, + 0.009707971476018429, + 0.005544326733797789, + 0.004472130443900824, + 0.006477118469774723, + 0.009177463129162788, + 0.006789018865674734, + 0.006094903219491243, + 0.007248011417686939, + 0.004226513672620058, + 0.022863119840621948, + 0.013397712260484695, + 0.01640351675450802, + 0.03044883906841278, + 0.012776313349604607, + 0.022538114339113235, + 0.016201630234718323, + 0.013631630688905716, + 0.003770733019337058, + 0.004226830787956715, + 0.006981268059462309, + 0.006155064329504967, + 0.006454748101532459, + 0.008574625477194786, + 0.006463891826570034, + 0.004732332658022642, + 0.004113357979804277, + 0.0027866146992892027, + 0.0036406172439455986, + 0.002940853824838996, + 0.0031606287229806185, + 0.00439573684707284, + 0.0032136780209839344, + 0.0025258795358240604, + 0.02502584457397461, + 0.024106016382575035, + 0.06465216726064682, + 0.01240537315607071, + 0.012868051417171955, + 0.021457988768815994, + 0.022315479815006256, + 0.0770510882139206, + 0.009575637057423592, + 0.01024249941110611, + 0.005599164869636297, + 0.0063989367336034775, + 0.015414021909236908, + 0.007418452762067318, + 0.007410572841763496, + 0.008210910484194756, + 0.004588816314935684, + 0.004855259787291288, + 0.006507085170596838, + 0.004142895806580782, + 0.005198508035391569, + 0.0058771283365786076, + 0.005857208743691444, + 0.004929274320602417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.014189423061907291 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot closely follows a sleek fighter jet, painted matte dark gray, with sharp angles and a futuristic design, as it flies at breakneck speed across the vast desert floor. Illuminated by the golden evening sun, the rolling sand dunes cast long, dramatic shadows, their delicate, wind-eroded textures clearly visible in the slanting light. The fighter jet flies extremely low over the sharp ridges of the dunes, its fuselage mere meters above the sand. The tips of its swept wings violently slice through the air at high speed, creating two distinct, intertwining, ribbon-like vortices from the immense pressure differential. These vortices, clinging tightly to the wingtips, tumble and stretch behind the fighter, briefly outlining its trajectory before dissipating into the air distorted by the engine's heatwave, the swirling sand particles forming a fleeting dust trail. The entire picture is full of extreme sense of speed, precise control and awe-inspiring mechanical aesthetics, forming a strong visual contrast between the magnificent natural scenery in warm colors and the cold cutting-edge technology, creating a tense and magnificent atmosphere.", + "seed": 16462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_1033_seed_16462_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020585077442228795, + "vlm_logprob": -0.020585077442228795, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.020585077442228795, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.046674925833940506 + ], + "opposite_logprob_tokens": [ + -3.171674966812134 + ], + "logprob_reward": -0.046674925833940506, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.046674925833940506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12749245762825012, + 0.0782613530755043, + 0.0742475837469101, + 0.17444263398647308, + 0.08485133945941925, + 0.06509595364332199, + 0.1703271120786667, + 0.05854204669594765, + 0.09865878522396088, + 0.06808073073625565 + ], + "credit_map": [ + 0.01687830500304699, + 0.01893572323024273, + 0.02231685072183609, + 0.017168836668133736, + 0.018742550164461136, + 0.01031521800905466, + 0.012157863937318325, + 0.010977108031511307, + 0.0077756838873028755, + 0.007797325495630503, + 0.007137293461710215, + 0.011172791942954063, + 0.015335183590650558, + 0.012393894605338573, + 0.008816638961434364, + 0.007832538336515427, + 0.012088056653738022, + 0.005804112181067467, + 0.007644748780876398, + 0.006901240441948175, + 0.00456724688410759, + 0.005865294486284256, + 0.01107926294207573, + 0.020297618582844734, + 0.01237448863685131, + 0.017344944179058075, + 0.01780976727604866, + 0.01648583635687828, + 0.01984742097556591, + 0.038075681775808334, + 0.02435627020895481, + 0.028148228302598, + 0.009445177391171455, + 0.007966795936226845, + 0.010247280821204185, + 0.014832533895969391, + 0.01064873393625021, + 0.0107760364189744, + 0.012011837214231491, + 0.008922940120100975, + 0.0061563048511743546, + 0.009453033097088337, + 0.005883956793695688, + 0.0049973139539361, + 0.011270497925579548, + 0.010459279641509056, + 0.011059992015361786, + 0.005815570708364248, + 0.01534652803093195, + 0.019312553107738495, + 0.017358317971229553, + 0.014727490022778511, + 0.020966755226254463, + 0.034803956747055054, + 0.026508139446377754, + 0.02130337618291378, + 0.004446341190487146, + 0.005343700293451548, + 0.014308521524071693, + 0.006044111214578152, + 0.005516144447028637, + 0.008513731881976128, + 0.005763583816587925, + 0.008605913259088993, + 0.008703135885298252, + 0.01852617971599102, + 0.006968938745558262, + 0.010433418676257133, + 0.023513903841376305, + 0.009795011952519417, + 0.012034635059535503, + 0.008683566004037857, + 0.005491782445460558, + 0.00920379813760519, + 0.0043084146454930305, + 0.007811244577169418, + 0.013638434931635857, + 0.010134538635611534, + 0.00915000680834055, + 0.008342508226633072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.046674925833940506 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270870868116617 + ], + "opposite_logprob_tokens": [ + -5.255270957946777 + ], + "logprob_reward": -0.005270870868116617, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005270870868116617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10400838404893875, + 0.0754232406616211, + 0.04480135813355446, + 0.059795260429382324, + 0.0432538166642189, + 0.3357188105583191, + 0.21961526572704315, + 0.04937449097633362, + 0.03678588196635246, + 0.03122347965836525 + ], + "credit_map": [ + 0.011744560673832893, + 0.009884205646812916, + 0.02478119730949402, + 0.012475567869842052, + 0.010479495860636234, + 0.009819318540394306, + 0.011973228305578232, + 0.012850809842348099, + 0.007054409943521023, + 0.006160259712487459, + 0.011805434711277485, + 0.008411538787186146, + 0.005359013564884663, + 0.009644215926527977, + 0.01657799258828163, + 0.010410373099148273, + 0.00604950450360775, + 0.005261865444481373, + 0.006302892696112394, + 0.006337856408208609, + 0.0030500360298901796, + 0.005074864719063044, + 0.005183153320103884, + 0.007541184779256582, + 0.006361254490911961, + 0.007457769010215998, + 0.007487546186894178, + 0.004721599165350199, + 0.004636313766241074, + 0.005068013910204172, + 0.008850022219121456, + 0.015212741680443287, + 0.0052434951066970825, + 0.005318895913660526, + 0.005645109806209803, + 0.007358698174357414, + 0.006257880479097366, + 0.004731677006930113, + 0.004371108487248421, + 0.004326948896050453, + 0.032911308109760284, + 0.09150296449661255, + 0.03576773405075073, + 0.034220416098833084, + 0.046762894839048386, + 0.02453215979039669, + 0.030165815725922585, + 0.03985551372170448, + 0.017611095681786537, + 0.06633318215608597, + 0.02082870714366436, + 0.024003125727176666, + 0.01810898631811142, + 0.018537359312176704, + 0.016931762918829918, + 0.03726104646921158, + 0.005916749592870474, + 0.007185663096606731, + 0.00684410659596324, + 0.0064598917961120605, + 0.00640021963045001, + 0.00739491730928421, + 0.0036502631846815348, + 0.005522677209228277, + 0.003130940953269601, + 0.005363027099519968, + 0.0034461261238902807, + 0.004927786532789469, + 0.006539897061884403, + 0.0037496506702154875, + 0.004203267861157656, + 0.005425188224762678, + 0.003784275846555829, + 0.0042605288326740265, + 0.002508825622498989, + 0.004353003576397896, + 0.004937360994517803, + 0.0041470034047961235, + 0.004004911053925753, + 0.003227570094168186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.005270870868116617 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810051456093788 + ], + "opposite_logprob_tokens": [ + -3.288100481033325 + ], + "logprob_reward": -0.03810051456093788, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810051456093788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11184411495923996, + 0.057437796145677567, + 0.2763965427875519, + 0.060607098042964935, + 0.04622672498226166, + 0.10489976406097412, + 0.05450702831149101, + 0.14059707522392273, + 0.08954349905252457, + 0.057940300554037094 + ], + "credit_map": [ + 0.012390939518809319, + 0.014379099011421204, + 0.019503837451338768, + 0.018076416105031967, + 0.00892534852027893, + 0.012634120881557465, + 0.014261940494179726, + 0.011672420427203178, + 0.007388426922261715, + 0.005657962989062071, + 0.004130839370191097, + 0.009008937515318394, + 0.004760247655212879, + 0.008151139132678509, + 0.011712688021361828, + 0.006627553142607212, + 0.018946584314107895, + 0.007311551366001368, + 0.007694779895246029, + 0.010056138969957829, + 0.008552676998078823, + 0.013050615787506104, + 0.1003040075302124, + 0.11048019677400589, + 0.003637832822278142, + 0.0050026788376271725, + 0.013350168243050575, + 0.00494942395016551, + 0.004934083204716444, + 0.012961694970726967, + 0.004379746038466692, + 0.011391467414796352, + 0.005375104025006294, + 0.00556421373039484, + 0.005196745041757822, + 0.008759918622672558, + 0.0057703410275280476, + 0.005133288446813822, + 0.006000260356813669, + 0.004426855128258467, + 0.009869096800684929, + 0.02594253048300743, + 0.011292190290987492, + 0.007426274940371513, + 0.011218392290174961, + 0.011305679567158222, + 0.015839790925383568, + 0.012005806900560856, + 0.005235685966908932, + 0.013085969723761082, + 0.006233426742255688, + 0.0045457095839083195, + 0.005667588673532009, + 0.00703125586733222, + 0.005797898396849632, + 0.006909488700330257, + 0.010556650348007679, + 0.01962861232459545, + 0.022932488471269608, + 0.012504106387495995, + 0.024042760953307152, + 0.02008742466568947, + 0.018122542649507523, + 0.012722485698759556, + 0.007876517251133919, + 0.0125749371945858, + 0.00470132427290082, + 0.012517415918409824, + 0.013137738220393658, + 0.008739184588193893, + 0.007904937490820885, + 0.022091440856456757, + 0.006392480339854956, + 0.007773216813802719, + 0.006005631759762764, + 0.007049063686281443, + 0.008313828147947788, + 0.008922537788748741, + 0.007288515102118254, + 0.006195024121552706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.03810051456093788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253438837826252 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253438837826252, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253438837826252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13177870213985443, + 0.07244927436113358, + 0.21552647650241852, + 0.11169884353876114, + 0.03468751162290573, + 0.13923120498657227, + 0.1463327258825302, + 0.09010657668113708, + 0.03178771212697029, + 0.026400921866297722 + ], + "credit_map": [ + 0.016864219680428505, + 0.016045331954956055, + 0.019880859181284904, + 0.018127908930182457, + 0.014493975788354874, + 0.017332738265395164, + 0.014273577369749546, + 0.01476009376347065, + 0.006783636752516031, + 0.005789564922451973, + 0.00716323172673583, + 0.0076795415952801704, + 0.00870231818407774, + 0.01329160388559103, + 0.008046136237680912, + 0.014993243850767612, + 0.017733054235577583, + 0.012296737171709538, + 0.025173038244247437, + 0.023698173463344574, + 0.016562076285481453, + 0.01873178221285343, + 0.0469718873500824, + 0.054359760135412216, + 0.0059866271913051605, + 0.007734791375696659, + 0.012013961561024189, + 0.009212876670062542, + 0.01703399419784546, + 0.03694949671626091, + 0.0134956743568182, + 0.009271436370909214, + 0.003992204554378986, + 0.0038058976642787457, + 0.0048888009041547775, + 0.006451338529586792, + 0.005092601291835308, + 0.004166162572801113, + 0.0032644488383084536, + 0.0030260614585131407, + 0.013407085090875626, + 0.036800943315029144, + 0.016906455159187317, + 0.01700573042035103, + 0.015384739264845848, + 0.00958844181150198, + 0.010949725285172462, + 0.01918810047209263, + 0.009891515597701073, + 0.028013400733470917, + 0.0177987739443779, + 0.016202466562390327, + 0.013135742396116257, + 0.02025020867586136, + 0.02063695713877678, + 0.020403672009706497, + 0.008902829140424728, + 0.011785418726503849, + 0.02127135545015335, + 0.013135720044374466, + 0.00862920843064785, + 0.00642153387889266, + 0.0059010484255850315, + 0.014059470035135746, + 0.003175569698214531, + 0.005010286811739206, + 0.004992721602320671, + 0.003250627778470516, + 0.004512529354542494, + 0.004126228392124176, + 0.0032027510460466146, + 0.003517002100124955, + 0.0030192823614925146, + 0.0030580207239836454, + 0.0023138243705034256, + 0.003362273098900914, + 0.003717410145327449, + 0.004499894566833973, + 0.003311307867988944, + 0.0031189110595732927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.005253438837826252 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625637110322714 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625637110322714, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625637110322714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12950244545936584, + 0.04830317944288254, + 0.33594372868537903, + 0.053567104041576385, + 0.05066786706447601, + 0.06732012331485748, + 0.1300983875989914, + 0.11558517813682556, + 0.0430058017373085, + 0.026006177067756653 + ], + "credit_map": [ + 0.015720978379249573, + 0.01732572354376316, + 0.0187180507928133, + 0.015568316914141178, + 0.015836963430047035, + 0.01567995920777321, + 0.01552248653024435, + 0.015129967592656612, + 0.007949275895953178, + 0.003800006350502372, + 0.004244405310600996, + 0.0038324538618326187, + 0.004280664026737213, + 0.005676419008523226, + 0.0063644889742136, + 0.01215546578168869, + 0.025607120245695114, + 0.008681995794177055, + 0.0097985053434968, + 0.010034002363681793, + 0.010366491973400116, + 0.014583663083612919, + 0.12003888934850693, + 0.1368330717086792, + 0.005164698231965303, + 0.008479036390781403, + 0.004475071560591459, + 0.005494754295796156, + 0.005334600806236267, + 0.008363349363207817, + 0.00781980250030756, + 0.008435788564383984, + 0.005639118142426014, + 0.005410128738731146, + 0.006314459722489119, + 0.008336811326444149, + 0.007900550961494446, + 0.006775127723813057, + 0.006036646664142609, + 0.004255020059645176, + 0.007300175726413727, + 0.013206996954977512, + 0.007527569308876991, + 0.007387068588286638, + 0.010376645252108574, + 0.006312891375273466, + 0.008287247270345688, + 0.006921527907252312, + 0.010057750158011913, + 0.02845732681453228, + 0.014751486480236053, + 0.019609542563557625, + 0.011340314522385597, + 0.013487821444869041, + 0.015287421643733978, + 0.01710672304034233, + 0.007943755947053432, + 0.010016191750764847, + 0.036385320127010345, + 0.008351345546543598, + 0.009709486737847328, + 0.010671502910554409, + 0.008819995447993279, + 0.023687580600380898, + 0.004113072995096445, + 0.004648472182452679, + 0.005389276426285505, + 0.006776981055736542, + 0.006015084683895111, + 0.004270248580724001, + 0.004008082207292318, + 0.007784583140164614, + 0.002585682785138488, + 0.003693646052852273, + 0.0018347270088270307, + 0.0022861745674163103, + 0.004109645262360573, + 0.004766782745718956, + 0.0030376689974218607, + 0.0036918488331139088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.007625637110322714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot closely follows a sleek fighter jet, painted matte dark gray, with sharp angles and a futuristic design, as it flies at breakneck speed across the vast desert floor. Illuminated by the golden evening sun, the rolling sand dunes cast long, dramatic shadows, their delicate, wind-eroded textures clearly visible in the slanting light. The fighter jet flies extremely low over the sharp ridges of the dunes, its fuselage mere meters above the sand. The tips of its swept wings violently slice through the air at high speed, creating two distinct, intertwining, ribbon-like vortices from the immense pressure differential. These vortices, clinging tightly to the wingtips, tumble and stretch behind the fighter, briefly outlining its trajectory before dissipating into the air distorted by the engine's heatwave, the swirling sand particles forming a fleeting dust trail. The entire picture is full of extreme sense of speed, precise control and awe-inspiring mechanical aesthetics, forming a strong visual contrast between the magnificent natural scenery in warm colors and the cold cutting-edge technology, creating a tense and magnificent atmosphere.", + "seed": 16463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_1033_seed_16463_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05631269617006183, + "vlm_logprob": -0.05631269617006183, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.05631269617006183, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20592617988586426 + ], + "opposite_logprob_tokens": [ + -1.7059261798858643 + ], + "logprob_reward": -0.20592617988586426, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20592617988586426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3440016210079193, + 0.2934226095676422, + 0.03413956239819527, + 0.04761940613389015, + 0.04647913575172424, + 0.02302347496151924, + 0.069969043135643, + 0.023723402991890907, + 0.05091340094804764, + 0.06670840829610825 + ], + "credit_map": [ + 0.04434368759393692, + 0.03834328055381775, + 0.03583331033587456, + 0.027560193091630936, + 0.05801713839173317, + 0.03317902609705925, + 0.0559915229678154, + 0.05073345825076103, + 0.01668926142156124, + 0.008637615479528904, + 0.07387152314186096, + 0.016087215393781662, + 0.0279314536601305, + 0.11414969712495804, + 0.017979023978114128, + 0.018076810985803604, + 0.00399328488856554, + 0.003819236299023032, + 0.004281762521713972, + 0.003916039597243071, + 0.004146715626120567, + 0.006919744424521923, + 0.003924681805074215, + 0.00313809746876359, + 0.004071714822202921, + 0.006012818310409784, + 0.005985868163406849, + 0.003967369440943003, + 0.013778235763311386, + 0.003513842122629285, + 0.005847068503499031, + 0.004442491102963686, + 0.006047066766768694, + 0.008187745697796345, + 0.006220578216016293, + 0.007338513620197773, + 0.005476253107190132, + 0.0035237963311374187, + 0.006244877818971872, + 0.0034403030294924974, + 0.0024254131130874157, + 0.003024378325790167, + 0.0031949530821293592, + 0.0029764273203909397, + 0.0018687581177800894, + 0.0037880532909184694, + 0.0031228535808622837, + 0.002622636267915368, + 0.009553728625178337, + 0.010201864875853062, + 0.008762119337916374, + 0.014018937014043331, + 0.0076783811673521996, + 0.005805722903460264, + 0.0070428065955638885, + 0.006905478425323963, + 0.0027102180756628513, + 0.0017547309398651123, + 0.0050422679632902145, + 0.00218889769166708, + 0.003396559040993452, + 0.0034186472184956074, + 0.0026968789752572775, + 0.0025152042508125305, + 0.0074137067422270775, + 0.008171497844159603, + 0.005325171165168285, + 0.005289706867188215, + 0.006045707035809755, + 0.006083430256694555, + 0.0054765017703175545, + 0.007107682526111603, + 0.005656752735376358, + 0.003860285971313715, + 0.01938270404934883, + 0.0045534707605838776, + 0.0025941438507288694, + 0.0077449241653084755, + 0.007431536912918091, + 0.015484587289392948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.20592617988586426 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006759160198271275 + ], + "opposite_logprob_tokens": [ + -5.006759166717529 + ], + "logprob_reward": -0.006759160198271275, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006759160198271275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08100631088018417, + 0.4628339409828186, + 0.06590383499860764, + 0.05034022778272629, + 0.025253960862755775, + 0.11537470668554306, + 0.030433420091867447, + 0.04126093164086342, + 0.040253255516290665, + 0.08733948320150375 + ], + "credit_map": [ + 0.006783400662243366, + 0.009556646458804607, + 0.009233862161636353, + 0.0069318898022174835, + 0.009895357303321362, + 0.013843429274857044, + 0.012755807489156723, + 0.012005913071334362, + 0.029245147481560707, + 0.012485835701227188, + 0.1539676934480667, + 0.028178749606013298, + 0.02686731144785881, + 0.14124827086925507, + 0.03294197842478752, + 0.037898920476436615, + 0.010334942489862442, + 0.007632677908986807, + 0.008496846072375774, + 0.00410351762548089, + 0.007414772175252438, + 0.013223610818386078, + 0.007275478448718786, + 0.007421988993883133, + 0.00530563248321414, + 0.005569683853536844, + 0.006780424155294895, + 0.0032219639979302883, + 0.011692128144204617, + 0.005438425112515688, + 0.005562494974583387, + 0.0067694708704948425, + 0.003039299976080656, + 0.004046680871397257, + 0.004502088762819767, + 0.0031878547742962837, + 0.0018648178083822131, + 0.003041676478460431, + 0.002803330309689045, + 0.0027682092040777206, + 0.01729961670935154, + 0.009796424768865108, + 0.014441664330661297, + 0.013298684731125832, + 0.011775544844567776, + 0.021119005978107452, + 0.013355443254113197, + 0.014288314618170261, + 0.008504800498485565, + 0.0033098123967647552, + 0.004378932528197765, + 0.004143150988966227, + 0.0018381074769422412, + 0.002678239019587636, + 0.002345043234527111, + 0.0032353319693356752, + 0.008451159112155437, + 0.003505474189296365, + 0.005333288107067347, + 0.004121397156268358, + 0.0033612046390771866, + 0.006998913362622261, + 0.0037549221888184547, + 0.005734573118388653, + 0.004643613938242197, + 0.004445115104317665, + 0.00530525716021657, + 0.004619027487933636, + 0.005934030748903751, + 0.004793156869709492, + 0.0046376814134418964, + 0.005875367671251297, + 0.0094660809263587, + 0.006697012111544609, + 0.022374045103788376, + 0.00686277449131012, + 0.007193245459347963, + 0.00913156010210514, + 0.012125400826334953, + 0.013489355333149433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.006759160198271275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430624857544899 + ], + "opposite_logprob_tokens": [ + -3.168062448501587 + ], + "logprob_reward": -0.0430624857544899, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430624857544899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11169961839914322, + 0.5878636837005615, + 0.08594294637441635, + 0.08141161501407623, + 0.021063389256596565, + 0.028736798092722893, + 0.013403761200606823, + 0.031276438385248184, + 0.02086774818599224, + 0.017733991146087646 + ], + "credit_map": [ + 0.017280999571084976, + 0.010263165459036827, + 0.013896835967898369, + 0.007692043669521809, + 0.02020619809627533, + 0.011174553073942661, + 0.015063120052218437, + 0.01612270437180996, + 0.03306184709072113, + 0.011101597920060158, + 0.18606533110141754, + 0.03517499566078186, + 0.04508082941174507, + 0.21065464615821838, + 0.038904786109924316, + 0.027819659560918808, + 0.012381725013256073, + 0.007002110127359629, + 0.011535406112670898, + 0.0076801348477602005, + 0.010207341983914375, + 0.017299655824899673, + 0.00991431437432766, + 0.009922264143824577, + 0.006643506232649088, + 0.010587350465357304, + 0.012324226088821888, + 0.006226629018783569, + 0.022783007472753525, + 0.006390644237399101, + 0.009423903189599514, + 0.007032350637018681, + 0.0039615207351744175, + 0.002219908870756626, + 0.0043432265520095825, + 0.0019550523720681667, + 0.001749149989336729, + 0.0028949177358299494, + 0.0018309562001377344, + 0.0021086568012833595, + 0.003200249280780554, + 0.002619900507852435, + 0.003785502864047885, + 0.004201293457299471, + 0.0034636014606803656, + 0.004451642278581858, + 0.0034959015902131796, + 0.003518708050251007, + 0.00266645522788167, + 0.0020072481129318476, + 0.0015948188956826925, + 0.002367738401517272, + 0.0012963938061147928, + 0.0010139241348952055, + 0.0011279722675681114, + 0.0013292115181684494, + 0.005840157624334097, + 0.0028739881236106157, + 0.0036934767849743366, + 0.0026332137640565634, + 0.0025756710674613714, + 0.005991404876112938, + 0.0030554994009435177, + 0.004613026045262814, + 0.002484150230884552, + 0.002158442512154579, + 0.0031550440471619368, + 0.0021946937777101994, + 0.0025155304465442896, + 0.0023256647400557995, + 0.0027984739281237125, + 0.0032357482705265284, + 0.0016767007764428854, + 0.001258857431821525, + 0.0024193713907152414, + 0.0018662333022803068, + 0.001765314838849008, + 0.002717335009947419, + 0.0015937291318550706, + 0.004436451476067305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0430624857544899 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188007175922394 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.018188007175922394, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188007175922394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09439508616924286, + 0.1726432889699936, + 0.1600044220685959, + 0.17697769403457642, + 0.027318943291902542, + 0.16041983664035797, + 0.058773573487997055, + 0.038436103612184525, + 0.04336055740714073, + 0.06767048686742783 + ], + "credit_map": [ + 0.012553058564662933, + 0.011076305992901325, + 0.015155453234910965, + 0.009052191860973835, + 0.010695347562432289, + 0.0077118887566030025, + 0.015878286212682724, + 0.012272560968995094, + 0.021992841735482216, + 0.012774186208844185, + 0.021451594308018684, + 0.007205631583929062, + 0.019960157573223114, + 0.04125651717185974, + 0.011496443301439285, + 0.036505937576293945, + 0.026515323668718338, + 0.01119152456521988, + 0.023360172286629677, + 0.01348034106194973, + 0.015629876405000687, + 0.03370628505945206, + 0.017946723848581314, + 0.018174180760979652, + 0.015861468389630318, + 0.02239891141653061, + 0.027862751856446266, + 0.014022642746567726, + 0.04820113256573677, + 0.013981088064610958, + 0.020504465326666832, + 0.01414525043219328, + 0.0028080239426344633, + 0.004497148096561432, + 0.0043704030103981495, + 0.0038042620290070772, + 0.0026851750444620848, + 0.0030513557139784098, + 0.0028783422894775867, + 0.0032242347951978445, + 0.021721776574850082, + 0.013375922106206417, + 0.02144414186477661, + 0.020187824964523315, + 0.01838143728673458, + 0.026809219270944595, + 0.018862031400203705, + 0.019637495279312134, + 0.014557822607457638, + 0.006717996206134558, + 0.007798746693879366, + 0.009600814431905746, + 0.0042921872809529305, + 0.004860043991357088, + 0.005023426841944456, + 0.005922534503042698, + 0.00675056129693985, + 0.003942206036299467, + 0.004656885750591755, + 0.004418622702360153, + 0.002948351437225938, + 0.007221876177936792, + 0.0036779718939214945, + 0.004819633439183235, + 0.0035989892203360796, + 0.0051508755423128605, + 0.006655104458332062, + 0.0053017521277070045, + 0.0055238003842532635, + 0.005292768124490976, + 0.005210444331169128, + 0.006626825779676437, + 0.006563767325133085, + 0.004047799855470657, + 0.019780388101935387, + 0.004286725539714098, + 0.004312346689403057, + 0.00716229435056448, + 0.008581522852182388, + 0.012935644946992397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.018188007175922394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627647835761309 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627647835761309, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627647835761309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15306302905082703, + 0.3355069160461426, + 0.06209995970129967, + 0.10160153359174728, + 0.03968983143568039, + 0.05511292442679405, + 0.06882084161043167, + 0.03538554534316063, + 0.06950697302818298, + 0.07921239733695984 + ], + "credit_map": [ + 0.017572926357388496, + 0.020172463729977608, + 0.013838838785886765, + 0.014069457538425922, + 0.02835189737379551, + 0.01359503623098135, + 0.0223950557410717, + 0.023067351430654526, + 0.018726198002696037, + 0.007635762915015221, + 0.09115732461214066, + 0.018334897235035896, + 0.031044477596879005, + 0.13110719621181488, + 0.020617954432964325, + 0.016883086413145065, + 0.008958568796515465, + 0.006291157566010952, + 0.009423800744116306, + 0.004717376083135605, + 0.005897970870137215, + 0.00975979957729578, + 0.008955991826951504, + 0.008095297031104565, + 0.007788595277816057, + 0.01243617758154869, + 0.016022874042391777, + 0.008618716150522232, + 0.02769971638917923, + 0.007806194014847279, + 0.011842241510748863, + 0.009387023746967316, + 0.004906573798507452, + 0.0065653664059937, + 0.006127689499408007, + 0.005184904206544161, + 0.0038492733146995306, + 0.004320352803915739, + 0.005091286730021238, + 0.0036443837452679873, + 0.006798887625336647, + 0.004906084388494492, + 0.008100069127976894, + 0.0073324753902852535, + 0.004944974090903997, + 0.0071589285507798195, + 0.009041939862072468, + 0.006829558871686459, + 0.011676025576889515, + 0.009612653404474258, + 0.008481141179800034, + 0.0141370240598917, + 0.006348078604787588, + 0.004995514638721943, + 0.007129823789000511, + 0.006440582685172558, + 0.004925163462758064, + 0.002853657817468047, + 0.005455940030515194, + 0.003310148138552904, + 0.0040655676275491714, + 0.006283463444560766, + 0.003928330261260271, + 0.00456327386200428, + 0.011460437439382076, + 0.008756321854889393, + 0.00794868916273117, + 0.006676987744867802, + 0.008336580358445644, + 0.008655732497572899, + 0.007714475970715284, + 0.009957743808627129, + 0.010424107313156128, + 0.005249010398983955, + 0.019621727988123894, + 0.00617310730740428, + 0.006505600642412901, + 0.008320426568388939, + 0.011201249435544014, + 0.011717170476913452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.007627647835761309 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot closely follows a sleek fighter jet, painted matte dark gray, with sharp angles and a futuristic design, as it flies at breakneck speed across the vast desert floor. Illuminated by the golden evening sun, the rolling sand dunes cast long, dramatic shadows, their delicate, wind-eroded textures clearly visible in the slanting light. The fighter jet flies extremely low over the sharp ridges of the dunes, its fuselage mere meters above the sand. The tips of its swept wings violently slice through the air at high speed, creating two distinct, intertwining, ribbon-like vortices from the immense pressure differential. These vortices, clinging tightly to the wingtips, tumble and stretch behind the fighter, briefly outlining its trajectory before dissipating into the air distorted by the engine's heatwave, the swirling sand particles forming a fleeting dust trail. The entire picture is full of extreme sense of speed, precise control and awe-inspiring mechanical aesthetics, forming a strong visual contrast between the magnificent natural scenery in warm colors and the cold cutting-edge technology, creating a tense and magnificent atmosphere.", + "seed": 16464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_7/index_1033_seed_16464_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13196614561602474, + "vlm_logprob": -0.13196614561602474, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.13196614561602474, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5796933174133301 + ], + "opposite_logprob_tokens": [ + -0.8296933174133301 + ], + "logprob_reward": -0.5796933174133301, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5796933174133301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11985360085964203, + 0.03939821571111679, + 0.04174630716443062, + 0.11157987266778946, + 0.12638404965400696, + 0.051256150007247925, + 0.06439652293920517, + 0.03596780076622963, + 0.028520535677671432, + 0.3808968961238861 + ], + "credit_map": [ + 0.017828602343797684, + 0.01681194081902504, + 0.011196350678801537, + 0.011521872133016586, + 0.02015695348381996, + 0.015265426598489285, + 0.013753672130405903, + 0.013318795710802078, + 0.004268770571798086, + 0.006674258969724178, + 0.004723729100078344, + 0.004744143225252628, + 0.005529031623154879, + 0.004699131473898888, + 0.004901826847344637, + 0.003857326926663518, + 0.003870256245136261, + 0.005919285584241152, + 0.0039707375690341, + 0.004063895437866449, + 0.006118178367614746, + 0.004797612316906452, + 0.0041684494353830814, + 0.008837895467877388, + 0.012570750899612904, + 0.016831154003739357, + 0.008298209868371487, + 0.010315905325114727, + 0.01372553501278162, + 0.009330912493169308, + 0.017738891765475273, + 0.022768516093492508, + 0.011406979523599148, + 0.007846730761229992, + 0.03440992534160614, + 0.006405189633369446, + 0.03977441042661667, + 0.013287502340972424, + 0.0064607905223965645, + 0.006792522966861725, + 0.005883245263248682, + 0.00598500669002533, + 0.004092635586857796, + 0.008049564436078072, + 0.005261160898953676, + 0.005079443100839853, + 0.008676072582602501, + 0.008229020982980728, + 0.005179295781999826, + 0.006076365243643522, + 0.00545850396156311, + 0.006799045018851757, + 0.014744308777153492, + 0.008138569071888924, + 0.009887492284178734, + 0.008112948387861252, + 0.0055307066068053246, + 0.003082602983340621, + 0.0052837831899523735, + 0.003955688793212175, + 0.004341095685958862, + 0.004951719660311937, + 0.004414515104144812, + 0.004407691303640604, + 0.0025233463384211063, + 0.0029456093907356262, + 0.004425527527928352, + 0.0044558607041835785, + 0.0033372945617884398, + 0.003350809682160616, + 0.004197147209197283, + 0.0032849430572241545, + 0.010094691067934036, + 0.006153522524982691, + 0.009322098456323147, + 0.007132857106626034, + 0.014215600676834583, + 0.15485098958015442, + 0.012846890836954117, + 0.16628029942512512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.5796933174133301 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008665459230542183 + ], + "opposite_logprob_tokens": [ + -4.758665561676025 + ], + "logprob_reward": -0.008665459230542183, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008665459230542183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07547792047262192, + 0.0405116006731987, + 0.02456585131585598, + 0.0999305471777916, + 0.05129386857151985, + 0.04167413339018822, + 0.05335284024477005, + 0.028789272531867027, + 0.016719840466976166, + 0.5676841139793396 + ], + "credit_map": [ + 0.008952232077717781, + 0.009790683165192604, + 0.007948431186378002, + 0.008176161907613277, + 0.0077293794602155685, + 0.010631078854203224, + 0.009976302273571491, + 0.012273651547729969, + 0.004616678226739168, + 0.00657854788005352, + 0.006390138529241085, + 0.003946396056562662, + 0.004990110639482737, + 0.004606257192790508, + 0.004595414735376835, + 0.004788057412952185, + 0.0023280351888388395, + 0.0040580048225820065, + 0.003960807342082262, + 0.003063095035031438, + 0.002920351456850767, + 0.0029152855277061462, + 0.002474050736054778, + 0.002846222370862961, + 0.014567326754331589, + 0.014926246367394924, + 0.008458144031465054, + 0.010034558363258839, + 0.017508210614323616, + 0.009964640252292156, + 0.011005637235939503, + 0.01346578635275364, + 0.004600278567522764, + 0.003241018857806921, + 0.008975778706371784, + 0.003549163695424795, + 0.017673609778285027, + 0.00527453888207674, + 0.0029444124083966017, + 0.00503506837412715, + 0.005686647724360228, + 0.004085994325578213, + 0.003548668697476387, + 0.004583396017551422, + 0.0040771737694740295, + 0.004652156960219145, + 0.00552103016525507, + 0.009519065730273724, + 0.00501097459346056, + 0.004718557000160217, + 0.0056311823427677155, + 0.0053126453422009945, + 0.010999331250786781, + 0.007193524390459061, + 0.007812249008566141, + 0.006674372591078281, + 0.005715719424188137, + 0.002643402898684144, + 0.0036749886348843575, + 0.0027078904677182436, + 0.0036084752064198256, + 0.0037056435830891132, + 0.0031057558953762054, + 0.003627397818490863, + 0.0011893940391018987, + 0.0018129345262423158, + 0.002862259978428483, + 0.002497768960893154, + 0.0019857033621519804, + 0.00259256805293262, + 0.0017685265047475696, + 0.002010685158893466, + 0.015253902412950993, + 0.00923088937997818, + 0.009868803434073925, + 0.008707457222044468, + 0.017284007743000984, + 0.23340852558612823, + 0.01807352341711521, + 0.255856990814209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.008665459230542183 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866760969161987 + ], + "opposite_logprob_tokens": [ + -3.0486676692962646 + ], + "logprob_reward": -0.04866760969161987, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866760969161987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10310954600572586, + 0.07019422948360443, + 0.06633080542087555, + 0.20951540768146515, + 0.10209707170724869, + 0.08592311292886734, + 0.06596657633781433, + 0.08193235844373703, + 0.15844179689884186, + 0.05648909881711006 + ], + "credit_map": [ + 0.017283383756875992, + 0.01530429907143116, + 0.012382817454636097, + 0.011157759465277195, + 0.014972829259932041, + 0.014553687535226345, + 0.009879399091005325, + 0.007575374562293291, + 0.006627371069043875, + 0.01176503300666809, + 0.009224159643054008, + 0.007931523956358433, + 0.009011032059788704, + 0.009568111039698124, + 0.008356555365025997, + 0.007710441946983337, + 0.0067000179551541805, + 0.010357176885008812, + 0.008010963909327984, + 0.009793426841497421, + 0.008192062377929688, + 0.007418791763484478, + 0.0074661048129200935, + 0.00839226320385933, + 0.031732138246297836, + 0.027115384116768837, + 0.020424561575055122, + 0.01795966923236847, + 0.04313565418124199, + 0.024536527693271637, + 0.017551695927977562, + 0.027059761807322502, + 0.008271395228803158, + 0.007260337006300688, + 0.012867946177721024, + 0.0065391515381634235, + 0.033086903393268585, + 0.009898367337882519, + 0.010703345760703087, + 0.013469630852341652, + 0.009651130996644497, + 0.009442591108381748, + 0.007417822256684303, + 0.009999816305935383, + 0.010959386825561523, + 0.011126650497317314, + 0.013094350695610046, + 0.014231361448764801, + 0.006601753178983927, + 0.007185967173427343, + 0.006014345679432154, + 0.00639033829793334, + 0.014116713777184486, + 0.006505876313894987, + 0.010232571512460709, + 0.008919010870158672, + 0.008613639511168003, + 0.008395113050937653, + 0.007318504620343447, + 0.007514419034123421, + 0.017451263964176178, + 0.009617858566343784, + 0.012527457438409328, + 0.010494104586541653, + 0.030910078436136246, + 0.018044672906398773, + 0.011036784388124943, + 0.009895953349769115, + 0.02210620976984501, + 0.02183932065963745, + 0.02514960989356041, + 0.01945917308330536, + 0.004933974705636501, + 0.003990352153778076, + 0.004162785597145557, + 0.0033854106441140175, + 0.004965241067111492, + 0.01407871674746275, + 0.003785019740462303, + 0.01718759536743164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.04866760969161987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738202180713415 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738202180713415, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738202180713415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10882989317178726, + 0.14913298189640045, + 0.063728466629982, + 0.1115730032324791, + 0.09158571809530258, + 0.028255894780158997, + 0.03104236349463463, + 0.034410156309604645, + 0.03590276837348938, + 0.34553876519203186 + ], + "credit_map": [ + 0.014535622671246529, + 0.015159490518271923, + 0.012846109457314014, + 0.014020121656358242, + 0.013082869350910187, + 0.016309194266796112, + 0.01401517540216446, + 0.008861307986080647, + 0.01335891429334879, + 0.014932269230484962, + 0.02514549344778061, + 0.017382241785526276, + 0.014897565357387066, + 0.021229654550552368, + 0.01766064204275608, + 0.02452620305120945, + 0.005966756958514452, + 0.008471048437058926, + 0.006665471941232681, + 0.0054907528683543205, + 0.013189064338803291, + 0.007521369494497776, + 0.00709534203633666, + 0.009328658692538738, + 0.013646838255226612, + 0.016922755166888237, + 0.0058517721481621265, + 0.00924611184746027, + 0.00783679448068142, + 0.010505969636142254, + 0.020177805796265602, + 0.02738495171070099, + 0.008990033529698849, + 0.005377091001719236, + 0.028492415323853493, + 0.0050207204185426235, + 0.017680982127785683, + 0.007720177061855793, + 0.00430649658665061, + 0.013997803442180157, + 0.005404659081250429, + 0.003983575385063887, + 0.0027662296779453754, + 0.004086886532604694, + 0.0022325406316667795, + 0.002384804654866457, + 0.003583006327971816, + 0.0038141929544508457, + 0.002201332710683346, + 0.0026859520003199577, + 0.003271650755777955, + 0.003007392166182399, + 0.005775033496320248, + 0.005935030058026314, + 0.004068584181368351, + 0.004097387194633484, + 0.00974369514733553, + 0.0043135411106050014, + 0.0032217507250607014, + 0.002121923491358757, + 0.002940086182206869, + 0.0031668227165937424, + 0.0035091410391032696, + 0.005393194500356913, + 0.006957458797842264, + 0.003705244744196534, + 0.00363728916272521, + 0.002817722037434578, + 0.005251810420304537, + 0.005356042180210352, + 0.0035256361588835716, + 0.004651566967368126, + 0.009311629459261894, + 0.005107965786010027, + 0.007684910669922829, + 0.006035164464265108, + 0.011456473730504513, + 0.14154879748821259, + 0.011032870970666409, + 0.15336094796657562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.006738202180713415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek fighter jet painted matte dark gray with sharp angles and a futuristic design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and tracking the jet closely as it flies extremely low over sharp sand dune ridges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two distinct, intertwining, ribbon-like vortices visible at the wingtips, clinging tightly and tumbling behind the jet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a golden evening sun casting long shadows on rolling sand dunes with visible wind-eroded textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the jet's fuselage shown mere meters above the sand, with swirling sand particles forming a dust trail distorted by a heatwave?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066139563918114 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066139563918114, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066139563918114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12617003917694092, + 0.0744970366358757, + 0.047856055200099945, + 0.03381504490971565, + 0.09959564357995987, + 0.04535635560750961, + 0.06316126883029938, + 0.03792428597807884, + 0.05251406878232956, + 0.4191102385520935 + ], + "credit_map": [ + 0.025354770943522453, + 0.014750521630048752, + 0.01507460605353117, + 0.012020370922982693, + 0.020527418702840805, + 0.016828952357172966, + 0.011299152858555317, + 0.010314250364899635, + 0.009012749418616295, + 0.012221330776810646, + 0.00952441617846489, + 0.005499042104929686, + 0.009392227977514267, + 0.009752528741955757, + 0.01070318091660738, + 0.008391562849283218, + 0.0042223380878567696, + 0.008592478930950165, + 0.004565761424601078, + 0.005139537155628204, + 0.010604490526020527, + 0.0065936981700360775, + 0.0038950536400079727, + 0.004242696799337864, + 0.005952930543571711, + 0.005218584556132555, + 0.003228861605748534, + 0.0035440484061837196, + 0.003867201739922166, + 0.0037848877254873514, + 0.003405727446079254, + 0.004812804516404867, + 0.007998607121407986, + 0.005345211364328861, + 0.018504852429032326, + 0.0031189543660730124, + 0.04037702456116676, + 0.009330203756690025, + 0.003495576558634639, + 0.01142521295696497, + 0.00523839145898819, + 0.004516542423516512, + 0.0046738386154174805, + 0.006976681295782328, + 0.004809725563973188, + 0.0044134221971035, + 0.00606816029176116, + 0.008659595623612404, + 0.00662850821390748, + 0.00561828026548028, + 0.0076988087967038155, + 0.005813021678477526, + 0.011987032368779182, + 0.007287480402737856, + 0.009258256293833256, + 0.008869882673025131, + 0.00800977274775505, + 0.0033842416014522314, + 0.005754250101745129, + 0.0038013923913240433, + 0.0031788332853466272, + 0.004896396771073341, + 0.004039797466248274, + 0.004859598353505135, + 0.013362006284296513, + 0.005419272463768721, + 0.005975347477942705, + 0.004257652442902327, + 0.006785417906939983, + 0.00543211167678237, + 0.004845486022531986, + 0.006436775904148817, + 0.011560995131731033, + 0.006397855002433062, + 0.008499211631715298, + 0.007315822411328554, + 0.013962985016405582, + 0.1713302582502365, + 0.01309032179415226, + 0.18695275485515594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.016066139563918114 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_8/index_tensor([9333, 9333])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_8/index_tensor([9333, 9333])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4babd4feab2806451a276fdc489f7ffbc28b2b7e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_8/index_tensor([9333, 9333])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static shot, shot from an extreme close-up and slightly upward, focuses on an expressionless porcelain doll. Its smooth, cold porcelain cheeks, painted with hand-painted pink rouge, have faded slightly with age, while the corners of its mouth maintain a timeless, faint curve. It sits stiffly, clad in a yellowed, lace-trimmed Victorian dress, its carefully coiffed, though slightly dusty, blond curls cascading over its shoulders. Its gaze, through its deep, hollow blue glass eyes, is fixed motionlessly on an old rag doll across from it. The doll, made of rough, faded cotton and linen, its body sagging with age, sits limply, two black buttons serving as eyes, and a slightly crooked smile, stitched with thick red thread, is faintly visible in the shadows. A dim beam of light slants in from outside the painting, penetrating the fine dust particles floating in the air and precisely illuminating half of the porcelain doll's smooth cheek. A cold, lifeless highlight appears on its glass eye, while the rough texture of the doll absorbs most of the light, fading it into deeper shadows. The painting's somber tones, replete with stark contrasts of light and dark, create an atmosphere of unsettling stillness, frozen in time, as if a silent confrontation is eternally unfolding between two inanimate objects.", + "seed": 9615, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_8/index_9333_seed_9615_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11205842876806855, + "vlm_logprob": -0.11205842876806855, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.11205842876806855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16883431375026703 + ], + "opposite_logprob_tokens": [ + -1.9188343286514282 + ], + "logprob_reward": -0.16883431375026703, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16883431375026703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3401113748550415, + 0.1272912621498108, + 0.06941666454076767, + 0.12482555955648422, + 0.046807028353214264, + 0.03534357622265816, + 0.0333683043718338, + 0.07509903609752655, + 0.07420742511749268, + 0.07352982461452484 + ], + "credit_map": [ + 0.050991397351026535, + 0.04208371043205261, + 0.051627952605485916, + 0.048364534974098206, + 0.06848409026861191, + 0.025929493829607964, + 0.03253040090203285, + 0.020099760964512825, + 0.019847551360726357, + 0.01876818574965, + 0.020442482084035873, + 0.019281452521681786, + 0.012390481308102608, + 0.011380481533706188, + 0.012393624521791935, + 0.012786990962922573, + 0.0112466337159276, + 0.011090649291872978, + 0.01186960656195879, + 0.011260400526225567, + 0.010529694147408009, + 0.003650092985481024, + 0.005429868586361408, + 0.004339715465903282, + 0.019681280478835106, + 0.015716729685664177, + 0.020228656008839607, + 0.018784359097480774, + 0.014339629560709, + 0.01496279425919056, + 0.009497081860899925, + 0.011615031398832798, + 0.007404375821352005, + 0.009401794523000717, + 0.008947786875069141, + 0.0051996237598359585, + 0.004073609132319689, + 0.005071634892374277, + 0.0024893544614315033, + 0.00421884236857295, + 0.006377657875418663, + 0.004770469386130571, + 0.00843760184943676, + 0.0048714205622673035, + 0.0025653867051005363, + 0.0030542202293872833, + 0.0021015361417084932, + 0.0031652774196118116, + 0.006530974991619587, + 0.0038592959754168987, + 0.007807509507983923, + 0.005970305763185024, + 0.002957958960905671, + 0.0017017531208693981, + 0.002355185803025961, + 0.0021853153593838215, + 0.012388945557177067, + 0.01220835093408823, + 0.01181744784116745, + 0.010907226242125034, + 0.008067246526479721, + 0.007049567997455597, + 0.00450031366199255, + 0.008159938268363476, + 0.01152853760868311, + 0.011611941270530224, + 0.007974269799888134, + 0.010284657590091228, + 0.0054989284835755825, + 0.008777664043009281, + 0.005133932456374168, + 0.013397489674389362, + 0.012839102186262608, + 0.011245480738580227, + 0.007955479435622692, + 0.009934133850038052, + 0.0055157579481601715, + 0.008890671655535698, + 0.00575083727017045, + 0.01139836385846138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.16883431375026703 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561347484588623 + ], + "opposite_logprob_tokens": [ + -1.6006134748458862 + ], + "logprob_reward": -0.22561347484588623, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561347484588623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2943339943885803, + 0.2050081044435501, + 0.07420027256011963, + 0.06274335086345673, + 0.06341198086738586, + 0.053534992039203644, + 0.054747410118579865, + 0.08869583904743195, + 0.05074340105056763, + 0.05258075147867203 + ], + "credit_map": [ + 0.04965271055698395, + 0.038326289504766464, + 0.04601673409342766, + 0.04166736453771591, + 0.035374805331230164, + 0.035775452852249146, + 0.023089632391929626, + 0.02443098835647106, + 0.031832195818424225, + 0.027953561395406723, + 0.030529886484146118, + 0.031120872125029564, + 0.022701486945152283, + 0.023279966786503792, + 0.01743890717625618, + 0.020151209086179733, + 0.011619948782026768, + 0.009805130772292614, + 0.009323520585894585, + 0.008117197081446648, + 0.012032520025968552, + 0.0053518060594797134, + 0.007614274509251118, + 0.010335867293179035, + 0.01034176629036665, + 0.005729881580919027, + 0.008444896899163723, + 0.00958652701228857, + 0.0062566958367824554, + 0.00887064728885889, + 0.0043180896900594234, + 0.009194842539727688, + 0.011242765933275223, + 0.005362629424780607, + 0.013615661300718784, + 0.00952098984271288, + 0.006877607200294733, + 0.004608748946338892, + 0.004308741539716721, + 0.007874825038015842, + 0.009560897946357727, + 0.008005300536751747, + 0.01289100106805563, + 0.008508351631462574, + 0.003496503224596381, + 0.0037807233165949583, + 0.003039759350940585, + 0.004252446815371513, + 0.010603339411318302, + 0.008497912436723709, + 0.009829578921198845, + 0.008343388326466084, + 0.004655696917325258, + 0.003752346383407712, + 0.004748388659209013, + 0.004316750913858414, + 0.015243018977344036, + 0.014863967895507812, + 0.013405323959887028, + 0.009226945228874683, + 0.008869669400155544, + 0.010382069274783134, + 0.005306608509272337, + 0.01139822043478489, + 0.008265318349003792, + 0.007059228140860796, + 0.00846906565129757, + 0.0069800312630832195, + 0.004606971051543951, + 0.00569869065657258, + 0.004189054947346449, + 0.005475039128214121, + 0.010186304338276386, + 0.005834490992128849, + 0.007525203283876181, + 0.009677513502538204, + 0.004391667898744345, + 0.004528528545051813, + 0.004972093738615513, + 0.005464943125844002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.22561347484588623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067553711123764515 + ], + "opposite_logprob_tokens": [ + -5.006755352020264 + ], + "logprob_reward": -0.0067553711123764515, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067553711123764515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20804990828037262, + 0.13035675883293152, + 0.03574772924184799, + 0.11911378800868988, + 0.09828328341245651, + 0.11124461889266968, + 0.028471756726503372, + 0.11514037102460861, + 0.06433568894863129, + 0.08925605565309525 + ], + "credit_map": [ + 0.040126655250787735, + 0.024784043431282043, + 0.03543027490377426, + 0.03585034981369972, + 0.024286000058054924, + 0.015167400240898132, + 0.016088128089904785, + 0.016317056491971016, + 0.02520975098013878, + 0.014560883864760399, + 0.02605782262980938, + 0.017234567552804947, + 0.01365444902330637, + 0.007192919030785561, + 0.01374579407274723, + 0.01270057912915945, + 0.005429297685623169, + 0.00449877604842186, + 0.006388809997588396, + 0.006753962021321058, + 0.00350851658731699, + 0.002793287392705679, + 0.0028415133710950613, + 0.003533565439283848, + 0.019379207864403725, + 0.01519076619297266, + 0.02199229598045349, + 0.020690007135272026, + 0.011614520102739334, + 0.011882533319294453, + 0.00660941656678915, + 0.011755040846765041, + 0.01820092462003231, + 0.01956784538924694, + 0.02309895120561123, + 0.012600699439644814, + 0.007641668431460857, + 0.005926358513534069, + 0.004374085925519466, + 0.006872751750051975, + 0.023330973461270332, + 0.015598738566040993, + 0.03189484775066376, + 0.017064081504940987, + 0.006649605929851532, + 0.004691519774496555, + 0.00581063237041235, + 0.0062042283825576305, + 0.004846790339797735, + 0.004229693207889795, + 0.003923608455806971, + 0.0033990752417594194, + 0.003237223718315363, + 0.0028846748173236847, + 0.0021420824341475964, + 0.003808611072599888, + 0.020531320944428444, + 0.01877618581056595, + 0.018641555681824684, + 0.019127948209643364, + 0.009705204516649246, + 0.008957390673458576, + 0.007523398380726576, + 0.011877362616360188, + 0.010081439279019833, + 0.010320693254470825, + 0.0069323270581662655, + 0.011232495307922363, + 0.0067748818546533585, + 0.0060600596480071545, + 0.004102989565581083, + 0.008830806240439415, + 0.014925437048077583, + 0.01258875522762537, + 0.012860422022640705, + 0.024040723219513893, + 0.007467618677765131, + 0.004956906661391258, + 0.007952581159770489, + 0.004463613964617252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0067553711123764515 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003960013389587 + ], + "opposite_logprob_tokens": [ + -2.6950395107269287 + ], + "logprob_reward": -0.07003960013389587, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003960013389587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3299291431903839, + 0.1371939778327942, + 0.10250815004110336, + 0.05794009566307068, + 0.06687142699956894, + 0.042053673416376114, + 0.04492221772670746, + 0.06847219914197922, + 0.09766629338264465, + 0.05244278162717819 + ], + "credit_map": [ + 0.06396883726119995, + 0.03592502325773239, + 0.06066572293639183, + 0.053458649665117264, + 0.04763355106115341, + 0.023456532508134842, + 0.022663326933979988, + 0.022157521918416023, + 0.019795594736933708, + 0.02042657881975174, + 0.021231917664408684, + 0.018885208293795586, + 0.02237427420914173, + 0.010048426687717438, + 0.012014037929475307, + 0.012417960911989212, + 0.01745307818055153, + 0.014184534549713135, + 0.013706739991903305, + 0.01332788448780775, + 0.01511254720389843, + 0.00818630214780569, + 0.010126098059117794, + 0.010410971008241177, + 0.008739686571061611, + 0.006581271532922983, + 0.00899372436106205, + 0.009473314508795738, + 0.006730674300342798, + 0.005922408774495125, + 0.004937977995723486, + 0.006561045069247484, + 0.010838255286216736, + 0.010076062753796577, + 0.012130266055464745, + 0.008585886098444462, + 0.006552157457917929, + 0.0063993665389716625, + 0.007142501883208752, + 0.005146934650838375, + 0.0077878618612885475, + 0.003787197871133685, + 0.01007765531539917, + 0.007212968543171883, + 0.002567499876022339, + 0.003679152112454176, + 0.0035977568477392197, + 0.0033435868099331856, + 0.008035139180719852, + 0.0057604373432695866, + 0.007732144556939602, + 0.006931317038834095, + 0.00382415228523314, + 0.003562235040590167, + 0.004476266913115978, + 0.004600526299327612, + 0.010751613415777683, + 0.011723417788743973, + 0.010595504194498062, + 0.007511110045015812, + 0.007319806143641472, + 0.007676218170672655, + 0.004187866114079952, + 0.008706665597856045, + 0.01695849932730198, + 0.016136685386300087, + 0.016431812196969986, + 0.014857918955385685, + 0.00911085307598114, + 0.009021496400237083, + 0.005856531672179699, + 0.009292500093579292, + 0.008459489792585373, + 0.0060376133769750595, + 0.0056255050003528595, + 0.011507736518979073, + 0.003977341577410698, + 0.006063032895326614, + 0.00464291637763381, + 0.006129152607172728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07003960013389587 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904938399791718 + ], + "opposite_logprob_tokens": [ + -2.4640493392944336 + ], + "logprob_reward": -0.08904938399791718, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904938399791718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.505134105682373, + 0.06325086206197739, + 0.0286798644810915, + 0.05892977863550186, + 0.0752321183681488, + 0.04806651175022125, + 0.03720083087682724, + 0.04996122047305107, + 0.08253447711467743, + 0.05101018026471138 + ], + "credit_map": [ + 0.07385159283876419, + 0.06580270826816559, + 0.0668862983584404, + 0.06376516819000244, + 0.09757732599973679, + 0.05370189994573593, + 0.045400213450193405, + 0.0381489135324955, + 0.010652747005224228, + 0.00817870069295168, + 0.008930944837629795, + 0.010031107813119888, + 0.005645354278385639, + 0.007446916773915291, + 0.0048499745316803455, + 0.007515114266425371, + 0.004760532174259424, + 0.0031008257064968348, + 0.004024303052574396, + 0.003100332571193576, + 0.003119096625596285, + 0.004049159120768309, + 0.0024862713180482388, + 0.004039342515170574, + 0.009471177123486996, + 0.008481008931994438, + 0.009107262827455997, + 0.01049930788576603, + 0.005453417077660561, + 0.007005869410932064, + 0.0036167106591165066, + 0.005295026581734419, + 0.01456030085682869, + 0.010689697228372097, + 0.01558971032500267, + 0.00789590459316969, + 0.006234296131879091, + 0.0075394511222839355, + 0.005880187265574932, + 0.006842578295618296, + 0.009771463461220264, + 0.006255264859646559, + 0.01162812951952219, + 0.007148395292460918, + 0.004175246227532625, + 0.0030870770569890738, + 0.002539197215810418, + 0.003461738582700491, + 0.006766790524125099, + 0.005623991601169109, + 0.006068361457437277, + 0.00590994069352746, + 0.0037449582014232874, + 0.0034243755508214235, + 0.0021845377050340176, + 0.0034778749104589224, + 0.009185269474983215, + 0.007784734014421701, + 0.0072123464196920395, + 0.006218848284333944, + 0.005620482377707958, + 0.0055693527683615685, + 0.00351791107095778, + 0.0048522790893912315, + 0.014844776131212711, + 0.013156174682080746, + 0.01159747876226902, + 0.010014187544584274, + 0.006666058208793402, + 0.007834167219698429, + 0.004959639627486467, + 0.013461999595165253, + 0.009185987524688244, + 0.005113914608955383, + 0.006489869672805071, + 0.007056170608848333, + 0.0038623614236712456, + 0.006990787573158741, + 0.0037967911921441555, + 0.008514299988746643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.08904938399791718 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, shot from an extreme close-up and slightly upward, focuses on an expressionless porcelain doll. Its smooth, cold porcelain cheeks, painted with hand-painted pink rouge, have faded slightly with age, while the corners of its mouth maintain a timeless, faint curve. It sits stiffly, clad in a yellowed, lace-trimmed Victorian dress, its carefully coiffed, though slightly dusty, blond curls cascading over its shoulders. Its gaze, through its deep, hollow blue glass eyes, is fixed motionlessly on an old rag doll across from it. The doll, made of rough, faded cotton and linen, its body sagging with age, sits limply, two black buttons serving as eyes, and a slightly crooked smile, stitched with thick red thread, is faintly visible in the shadows. A dim beam of light slants in from outside the painting, penetrating the fine dust particles floating in the air and precisely illuminating half of the porcelain doll's smooth cheek. A cold, lifeless highlight appears on its glass eye, while the rough texture of the doll absorbs most of the light, fading it into deeper shadows. The painting's somber tones, replete with stark contrasts of light and dark, create an atmosphere of unsettling stillness, frozen in time, as if a silent confrontation is eternally unfolding between two inanimate objects.", + "seed": 9616, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_8/index_9333_seed_9616_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06157672423869372, + "vlm_logprob": -0.06157672423869372, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.06157672423869372, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012275266461074352 + ], + "opposite_logprob_tokens": [ + -5.387275218963623 + ], + "logprob_reward": -0.012275266461074352, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.012275266461074352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05481862649321556, + 0.03564189374446869, + 0.026522135362029076, + 0.017384981736540794, + 0.023827489465475082, + 0.041018418967723846, + 0.021833905950188637, + 0.6519452929496765, + 0.06490956991910934, + 0.06209768354892731 + ], + "credit_map": [ + 0.006564631126821041, + 0.005620872601866722, + 0.008753905072808266, + 0.007153158076107502, + 0.0076399496756494045, + 0.005809412337839603, + 0.005270624533295631, + 0.008006074465811253, + 0.0024427028838545084, + 0.002980972873046994, + 0.00679565966129303, + 0.008448238484561443, + 0.006541941314935684, + 0.003146945498883724, + 0.0023474134504795074, + 0.002938017249107361, + 0.001976372441276908, + 0.002120181219652295, + 0.0022866365034133196, + 0.002191658364608884, + 0.004264675080776215, + 0.005044246558099985, + 0.0024101687595248222, + 0.006228195037692785, + 0.0019917567260563374, + 0.0009185426752083004, + 0.002535308711230755, + 0.0024735620245337486, + 0.0025931652635335922, + 0.0015765008283779025, + 0.0028814866673201323, + 0.0024146584328264, + 0.002359817735850811, + 0.0016824573976919055, + 0.0034605117980390787, + 0.0030571650713682175, + 0.0032543546985834837, + 0.003124021925032139, + 0.0027727042324841022, + 0.004116457886993885, + 0.003099615452811122, + 0.0023059016093611717, + 0.005004549399018288, + 0.004111682064831257, + 0.005030811298638582, + 0.002859157044440508, + 0.00381575059145689, + 0.014790952205657959, + 0.00252908025868237, + 0.0015310785965994, + 0.0026474404148757458, + 0.00167535908985883, + 0.003258269978687167, + 0.003566900035366416, + 0.003255879506468773, + 0.0033698983024805784, + 0.038058288395404816, + 0.043092697858810425, + 0.08718620985746384, + 0.03752874955534935, + 0.24648597836494446, + 0.032265592366456985, + 0.13379620015621185, + 0.033531591296195984, + 0.006243580020964146, + 0.006785794626921415, + 0.005670224316418171, + 0.005082296207547188, + 0.013406777754426003, + 0.0066438824869692326, + 0.010854586027562618, + 0.010222426615655422, + 0.007921834476292133, + 0.005650300066918135, + 0.008845638483762741, + 0.005087205674499273, + 0.008577522821724415, + 0.00824027881026268, + 0.00814760010689497, + 0.009627308696508408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.012275266461074352 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015935182571411 + ], + "opposite_logprob_tokens": [ + -1.7015935182571411 + ], + "logprob_reward": -0.2015935182571411, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015935182571411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2428852617740631, + 0.02500210702419281, + 0.04457511752843857, + 0.042180296033620834, + 0.08009279519319534, + 0.0996379405260086, + 0.06822160631418228, + 0.21265988051891327, + 0.09012014418840408, + 0.0946248322725296 + ], + "credit_map": [ + 0.028895968571305275, + 0.019890187308192253, + 0.044725630432367325, + 0.03431003540754318, + 0.028967982158064842, + 0.032185640186071396, + 0.02544756419956684, + 0.028462260961532593, + 0.0022152389865368605, + 0.002241741167381406, + 0.005545699968934059, + 0.0036372037138789892, + 0.0029377827886492014, + 0.0028726228047162294, + 0.002787579083815217, + 0.002764240140095353, + 0.005529533606022596, + 0.0027363651897758245, + 0.007005342282354832, + 0.004079924896359444, + 0.0069657498970627785, + 0.0036822855472564697, + 0.00487107178196311, + 0.009704851545393467, + 0.006055168341845274, + 0.0037411858793348074, + 0.003836052492260933, + 0.005568766966462135, + 0.006576646585017443, + 0.00582075584679842, + 0.005251326598227024, + 0.005330396816134453, + 0.008533025160431862, + 0.006340626627206802, + 0.00781391840428114, + 0.007892034016549587, + 0.016505716368556023, + 0.009723308496177197, + 0.010022168047726154, + 0.013262000866234303, + 0.007473160047084093, + 0.0037726189475506544, + 0.023263923823833466, + 0.007318930700421333, + 0.013823594897985458, + 0.0072074634954333305, + 0.011218713596463203, + 0.025559552013874054, + 0.011720779351890087, + 0.007077936548739672, + 0.005939009599387646, + 0.005767430178821087, + 0.009629467502236366, + 0.00644690589979291, + 0.010685882531106472, + 0.01095419842749834, + 0.012963852845132351, + 0.013841225765645504, + 0.007865918800234795, + 0.011053903959691525, + 0.08143973350524902, + 0.011523355729877949, + 0.044403545558452606, + 0.02956835925579071, + 0.00960406381636858, + 0.010429824702441692, + 0.009579240344464779, + 0.0066215405240654945, + 0.01396914478391409, + 0.012346074916422367, + 0.012499641627073288, + 0.015070619061589241, + 0.011727303266525269, + 0.00901826098561287, + 0.010889931581914425, + 0.007867624051868916, + 0.018458332866430283, + 0.010766370221972466, + 0.014596151188015938, + 0.01130086462944746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.2015935182571411 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008654231205582619 + ], + "opposite_logprob_tokens": [ + -4.7586541175842285 + ], + "logprob_reward": -0.008654231205582619, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008654231205582619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22885265946388245, + 0.04172250255942345, + 0.05789710581302643, + 0.03128372132778168, + 0.025691622868180275, + 0.05964181572198868, + 0.04235195368528366, + 0.4072937071323395, + 0.05963863059878349, + 0.04562629014253616 + ], + "credit_map": [ + 0.03192591667175293, + 0.025597132742404938, + 0.031584370881319046, + 0.03416577726602554, + 0.034111279994249344, + 0.03185958042740822, + 0.019232992082834244, + 0.020375613123178482, + 0.004391249269247055, + 0.003594441572204232, + 0.0099782170727849, + 0.00762200215831399, + 0.00454235402867198, + 0.003993686754256487, + 0.002964240498840809, + 0.004636310040950775, + 0.008215729147195816, + 0.004907046444714069, + 0.0075501552782952785, + 0.009829163551330566, + 0.008858906105160713, + 0.006554886698722839, + 0.006524805445224047, + 0.0054564145393669605, + 0.0036145320627838373, + 0.002354802330955863, + 0.0036264541558921337, + 0.004515080712735653, + 0.005313329864293337, + 0.0035922438837587833, + 0.004169795196503401, + 0.004097483586519957, + 0.0026389220729470253, + 0.0021919754799455404, + 0.004620961844921112, + 0.003181776963174343, + 0.004667586646974087, + 0.0031168826390057802, + 0.0025904683861881495, + 0.0026830509305000305, + 0.005808473099023104, + 0.004004703834652901, + 0.00612478144466877, + 0.006201293785125017, + 0.006477289367467165, + 0.0034713835921138525, + 0.010272598825395107, + 0.017281293869018555, + 0.007346414960920811, + 0.00366966612637043, + 0.0077547235414385796, + 0.003892230335623026, + 0.005133774597197771, + 0.005557886324822903, + 0.004849699325859547, + 0.004147561267018318, + 0.021618103608489037, + 0.02695583738386631, + 0.03590793162584305, + 0.020741146057844162, + 0.1604282408952713, + 0.02038201503455639, + 0.08626756817102432, + 0.03499286621809006, + 0.006233480293303728, + 0.0060362000949680805, + 0.005730483680963516, + 0.005438787396997213, + 0.01408383622765541, + 0.006983591243624687, + 0.008345769718289375, + 0.006786484271287918, + 0.00622195191681385, + 0.004453030880540609, + 0.004964215215295553, + 0.0031859194859862328, + 0.007594306953251362, + 0.005528208799660206, + 0.006818740162998438, + 0.006859918124973774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.008654231205582619 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330111712217331 + ], + "opposite_logprob_tokens": [ + -3.773301124572754 + ], + "logprob_reward": -0.02330111712217331, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330111712217331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30550679564476013, + 0.08977244049310684, + 0.0360836535692215, + 0.0328204408288002, + 0.06506600230932236, + 0.05688944458961487, + 0.07911071181297302, + 0.21016283333301544, + 0.06324679404497147, + 0.061340902000665665 + ], + "credit_map": [ + 0.040408290922641754, + 0.022708231583237648, + 0.05550181120634079, + 0.040571484714746475, + 0.03690871223807335, + 0.04109640792012215, + 0.03351997956633568, + 0.034791864454746246, + 0.006529377307742834, + 0.006856714375317097, + 0.01592383161187172, + 0.020670311525464058, + 0.0170047078281641, + 0.006934814620763063, + 0.0059665655717253685, + 0.0098861213773489, + 0.004431269597262144, + 0.0027613129932433367, + 0.003464941866695881, + 0.004012420307844877, + 0.0054704369977116585, + 0.004891727119684219, + 0.004621230531483889, + 0.006430311594158411, + 0.004225276410579681, + 0.0026177626568824053, + 0.004522114526480436, + 0.004835800733417273, + 0.005024717655032873, + 0.00337636424228549, + 0.00478059658780694, + 0.0034378112759441137, + 0.006342901848256588, + 0.004607874900102615, + 0.008272876031696796, + 0.0072190784849226475, + 0.011471827514469624, + 0.008769652806222439, + 0.009122375398874283, + 0.009259414859116077, + 0.005574886221438646, + 0.003916291985660791, + 0.013167180120944977, + 0.004690066445618868, + 0.007562468759715557, + 0.0040361336432397366, + 0.011541596613824368, + 0.006400815676897764, + 0.0163359884172678, + 0.0056689428165555, + 0.00932942796498537, + 0.007492217235267162, + 0.009862069971859455, + 0.008857613429427147, + 0.00775870680809021, + 0.013805747963488102, + 0.010249018669128418, + 0.011635295115411282, + 0.012666082940995693, + 0.008218392729759216, + 0.07734822481870651, + 0.011480722576379776, + 0.04198826849460602, + 0.03657682612538338, + 0.00570607790723443, + 0.00494726886972785, + 0.005204557906836271, + 0.005474571604281664, + 0.013191635720431805, + 0.007380221504718065, + 0.010822072625160217, + 0.010520384646952152, + 0.007197139784693718, + 0.005547875538468361, + 0.008909902535378933, + 0.004603488836437464, + 0.010082519613206387, + 0.0073135364800691605, + 0.010654337704181671, + 0.007032100576907396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02330111712217331 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205948814749718 + ], + "opposite_logprob_tokens": [ + -2.8120594024658203 + ], + "logprob_reward": -0.06205948814749718, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205948814749718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16375380754470825, + 0.0801253393292427, + 0.04061049222946167, + 0.024490421637892723, + 0.03992920368909836, + 0.030519215390086174, + 0.04529161751270294, + 0.4538257122039795, + 0.06790269911289215, + 0.053551509976387024 + ], + "credit_map": [ + 0.02120322547852993, + 0.013182492926716805, + 0.031351350247859955, + 0.019564667716622353, + 0.021292846649885178, + 0.024586493149399757, + 0.012910707853734493, + 0.01966201700270176, + 0.005820436403155327, + 0.006280632223933935, + 0.014718158170580864, + 0.019776225090026855, + 0.015548201277852058, + 0.006951311137527227, + 0.004929219838231802, + 0.006101157050579786, + 0.005826084408909082, + 0.002788265934213996, + 0.00406028563156724, + 0.005471699871122837, + 0.006677981466054916, + 0.0036175306886434555, + 0.005014404188841581, + 0.007154242601245642, + 0.0032500356901437044, + 0.0017582615837454796, + 0.002311130752786994, + 0.003854029346257448, + 0.0037641292437911034, + 0.002337893471121788, + 0.003501766361296177, + 0.0037131733261048794, + 0.0032850271090865135, + 0.002291542710736394, + 0.005252052564173937, + 0.005391613580286503, + 0.006327194627374411, + 0.0060500591062009335, + 0.006863831542432308, + 0.004467884544283152, + 0.004288879688829184, + 0.002188951475545764, + 0.00518054561689496, + 0.002730122534558177, + 0.003426678478717804, + 0.003354032523930073, + 0.0034396753180772066, + 0.0059103281237185, + 0.005302814766764641, + 0.004008497577160597, + 0.0037898942828178406, + 0.003551621688529849, + 0.008259428665041924, + 0.004619842395186424, + 0.009707547724246979, + 0.006051972974091768, + 0.026790307834744453, + 0.031173452734947205, + 0.051597267389297485, + 0.026470189914107323, + 0.17482592165470123, + 0.022955631837248802, + 0.09421195834875107, + 0.025800934061408043, + 0.006922190077602863, + 0.00804040115326643, + 0.00708639295771718, + 0.006487003993242979, + 0.010275186039507389, + 0.008396181277930737, + 0.010290022939443588, + 0.010405320674180984, + 0.006739554461091757, + 0.004200782161206007, + 0.00623235572129488, + 0.0041758595034480095, + 0.009483769536018372, + 0.0065231043845415115, + 0.008807220496237278, + 0.007388866972178221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.06205948814749718 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, shot from an extreme close-up and slightly upward, focuses on an expressionless porcelain doll. Its smooth, cold porcelain cheeks, painted with hand-painted pink rouge, have faded slightly with age, while the corners of its mouth maintain a timeless, faint curve. It sits stiffly, clad in a yellowed, lace-trimmed Victorian dress, its carefully coiffed, though slightly dusty, blond curls cascading over its shoulders. Its gaze, through its deep, hollow blue glass eyes, is fixed motionlessly on an old rag doll across from it. The doll, made of rough, faded cotton and linen, its body sagging with age, sits limply, two black buttons serving as eyes, and a slightly crooked smile, stitched with thick red thread, is faintly visible in the shadows. A dim beam of light slants in from outside the painting, penetrating the fine dust particles floating in the air and precisely illuminating half of the porcelain doll's smooth cheek. A cold, lifeless highlight appears on its glass eye, while the rough texture of the doll absorbs most of the light, fading it into deeper shadows. The painting's somber tones, replete with stark contrasts of light and dark, create an atmosphere of unsettling stillness, frozen in time, as if a silent confrontation is eternally unfolding between two inanimate objects.", + "seed": 9617, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_8/index_9333_seed_9617_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04736349210143089, + "vlm_logprob": -0.04736349210143089, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.04736349210143089, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02140655741095543 + ], + "opposite_logprob_tokens": [ + -4.021406650543213 + ], + "logprob_reward": -0.02140655741095543, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02140655741095543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10001648217439651, + 0.056536734104156494, + 0.148832768201828, + 0.11917033046483994, + 0.07598155736923218, + 0.09635505080223083, + 0.17712682485580444, + 0.05228137969970703, + 0.08474476635456085, + 0.08895409107208252 + ], + "credit_map": [ + 0.0073380074463784695, + 0.008112996816635132, + 0.01233020331710577, + 0.01115365233272314, + 0.014786595478653908, + 0.0121080931276083, + 0.021702509373426437, + 0.012484420090913773, + 0.006515192799270153, + 0.007990117184817791, + 0.004271111451089382, + 0.006071724463254213, + 0.010066138580441475, + 0.006919785402715206, + 0.008264413103461266, + 0.006438250187784433, + 0.004134903196245432, + 0.008134842850267887, + 0.004650171846151352, + 0.00643601268529892, + 0.010205435566604137, + 0.05969103425741196, + 0.008382787927985191, + 0.047197580337524414, + 0.0055825794115662575, + 0.008015687577426434, + 0.005092198494821787, + 0.0055176736786961555, + 0.013643275015056133, + 0.03731995075941086, + 0.012560313567519188, + 0.03143865615129471, + 0.006458446383476257, + 0.008568880148231983, + 0.005187759175896645, + 0.0055573685094714165, + 0.010706417262554169, + 0.017374278977513313, + 0.00804146658629179, + 0.014086946845054626, + 0.005243683699518442, + 0.012105327099561691, + 0.006003017537295818, + 0.006255825981497765, + 0.02436966449022293, + 0.012707029469311237, + 0.016980985179543495, + 0.012689518742263317, + 0.016183067113161087, + 0.02783278375864029, + 0.017241893336176872, + 0.020074384286999702, + 0.021368423476815224, + 0.021622858941555023, + 0.026826832443475723, + 0.02597658894956112, + 0.006121944170445204, + 0.006300751119852066, + 0.0041296472772955894, + 0.005328191444277763, + 0.00925730261951685, + 0.0067444657906889915, + 0.007471741642802954, + 0.006927333306521177, + 0.007803359068930149, + 0.0182571429759264, + 0.009894345887005329, + 0.010927229188382626, + 0.009174371138215065, + 0.010700237937271595, + 0.00903845764696598, + 0.008949628099799156, + 0.004949782509356737, + 0.015212373808026314, + 0.0057031563483178616, + 0.01303128618746996, + 0.006744868122041225, + 0.018586894497275352, + 0.008469963446259499, + 0.01625576615333557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.02140655741095543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296205222606659 + ], + "opposite_logprob_tokens": [ + -2.237962007522583 + ], + "logprob_reward": -0.11296205222606659, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296205222606659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1710314005613327, + 0.16712501645088196, + 0.13669827580451965, + 0.05668789520859718, + 0.06492579728364944, + 0.03858255222439766, + 0.10160325467586517, + 0.09558694064617157, + 0.07723701745271683, + 0.09052189439535141 + ], + "credit_map": [ + 0.017462771385908127, + 0.01715339906513691, + 0.015482769347727299, + 0.017935730516910553, + 0.03204760700464249, + 0.022151771932840347, + 0.029397279024124146, + 0.019400063902139664, + 0.010432952083647251, + 0.022360945120453835, + 0.013351131230592728, + 0.016813043504953384, + 0.037275854498147964, + 0.024287082254886627, + 0.024308087304234505, + 0.01829592138528824, + 0.007691469043493271, + 0.019658733159303665, + 0.010269740596413612, + 0.012647716328501701, + 0.017446041107177734, + 0.03313646838068962, + 0.013810021802783012, + 0.022038089111447334, + 0.005100415553897619, + 0.007896964438259602, + 0.004647562745958567, + 0.006565905641764402, + 0.005339278839528561, + 0.011565109714865685, + 0.005875599570572376, + 0.009697063826024532, + 0.002519152360036969, + 0.005157914012670517, + 0.0027246344834566116, + 0.0027018089313060045, + 0.008256908506155014, + 0.021017616614699364, + 0.005044530611485243, + 0.01750323176383972, + 0.003511386923491955, + 0.0041653867810964584, + 0.003402670379728079, + 0.003353552892804146, + 0.007889117114245892, + 0.005220665596425533, + 0.00596378231421113, + 0.005075990688055754, + 0.01178273931145668, + 0.018455592915415764, + 0.007936922833323479, + 0.009122897870838642, + 0.015256931073963642, + 0.012327417731285095, + 0.014201145619153976, + 0.01251960825175047, + 0.010094874538481236, + 0.016776710748672485, + 0.006851137150079012, + 0.0074487184174358845, + 0.01516791619360447, + 0.011346698738634586, + 0.013925102539360523, + 0.013975790701806545, + 0.006876402068883181, + 0.012508297339081764, + 0.007996608503162861, + 0.008863849565386772, + 0.010699445381760597, + 0.011072606779634953, + 0.008726442232728004, + 0.010493368841707706, + 0.006094234995543957, + 0.016125354915857315, + 0.008929566480219364, + 0.009897167794406414, + 0.01051845122128725, + 0.017430633306503296, + 0.008243745192885399, + 0.01328273955732584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.11296205222606659 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821937784552574 + ], + "opposite_logprob_tokens": [ + -4.018219470977783 + ], + "logprob_reward": -0.01821937784552574, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821937784552574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17732585966587067, + 0.07763558626174927, + 0.19978079199790955, + 0.11818291246891022, + 0.0563017763197422, + 0.036729540675878525, + 0.07503455877304077, + 0.0966871902346611, + 0.05103304982185364, + 0.11128871887922287 + ], + "credit_map": [ + 0.02463608793914318, + 0.024982398375868797, + 0.019820036366581917, + 0.022845152765512466, + 0.0220448300242424, + 0.016593972221016884, + 0.028665907680988312, + 0.017737485468387604, + 0.006189500913023949, + 0.009119288995862007, + 0.006695046089589596, + 0.009046564809978008, + 0.01564605161547661, + 0.008919685147702694, + 0.012816235423088074, + 0.00920321885496378, + 0.010813772678375244, + 0.016204513609409332, + 0.007093494758009911, + 0.012859268113970757, + 0.013050021603703499, + 0.0719057023525238, + 0.00831140298396349, + 0.059542618691921234, + 0.007845867425203323, + 0.008164151571691036, + 0.005778271239250898, + 0.004787130281329155, + 0.008313109166920185, + 0.03773718327283859, + 0.012204150669276714, + 0.0333530530333519, + 0.004395156633108854, + 0.006723950617015362, + 0.005123692099004984, + 0.005550357047468424, + 0.006795303896069527, + 0.00897605624049902, + 0.009177222847938538, + 0.009560035541653633, + 0.003675151849165559, + 0.005975661333650351, + 0.0034552125725895166, + 0.003484732238575816, + 0.006672619841992855, + 0.005934988148510456, + 0.004253501072525978, + 0.0032776726875454187, + 0.008698288351297379, + 0.011272750794887543, + 0.0076757073402404785, + 0.005693982820957899, + 0.010653545148670673, + 0.01082760188728571, + 0.008782902732491493, + 0.011429783888161182, + 0.008469109423458576, + 0.01669614017009735, + 0.007306430023163557, + 0.007348031736910343, + 0.01453167013823986, + 0.012796946801245213, + 0.01699776202440262, + 0.012541104108095169, + 0.004094335716217756, + 0.01147075928747654, + 0.006436086259782314, + 0.006980735808610916, + 0.004735640715807676, + 0.0068565960973501205, + 0.004718379117548466, + 0.005740517284721136, + 0.009763289242982864, + 0.016787132248282433, + 0.007635481655597687, + 0.01216085534542799, + 0.009644920006394386, + 0.02299809455871582, + 0.012321772053837776, + 0.019977176561951637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.01821937784552574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201764017343521 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.014201764017343521, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201764017343521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10357807576656342, + 0.09813567996025085, + 0.3031263053417206, + 0.10159879177808762, + 0.06115071102976799, + 0.12005947530269623, + 0.07613646239042282, + 0.06991028040647507, + 0.03650515526533127, + 0.029799077659845352 + ], + "credit_map": [ + 0.013801094144582748, + 0.010578541085124016, + 0.012735804542899132, + 0.014340650290250778, + 0.014213239774107933, + 0.012786809355020523, + 0.012977277860045433, + 0.01214465033262968, + 0.006232272833585739, + 0.010241555981338024, + 0.006772216875106096, + 0.00910111516714096, + 0.021847615018486977, + 0.01429012231528759, + 0.017253873869776726, + 0.012396903708577156, + 0.007841411046683788, + 0.0093384413048625, + 0.0059381332248449326, + 0.006425913888961077, + 0.016448751091957092, + 0.13473168015480042, + 0.012436454184353352, + 0.10996551811695099, + 0.003306615399196744, + 0.004356508608907461, + 0.003623386612161994, + 0.00343981571495533, + 0.010155366733670235, + 0.038023579865694046, + 0.007730289828032255, + 0.030963227152824402, + 0.004389844834804535, + 0.005457998253405094, + 0.004109937231987715, + 0.005277630407363176, + 0.014431973919272423, + 0.008763347752392292, + 0.012175212614238262, + 0.0065447669476270676, + 0.006071837618947029, + 0.010490692220628262, + 0.00541944894939661, + 0.00867478083819151, + 0.030743911862373352, + 0.019702695310115814, + 0.020250912755727768, + 0.01870519481599331, + 0.006694222800433636, + 0.010294418781995773, + 0.006098009180277586, + 0.00793321244418621, + 0.009509979747235775, + 0.011940645053982735, + 0.010762852616608143, + 0.012903125025331974, + 0.008339461870491505, + 0.012262918055057526, + 0.006546929478645325, + 0.005411684047430754, + 0.007319248281419277, + 0.008401001803576946, + 0.01296378392726183, + 0.008665252476930618, + 0.003275451948866248, + 0.005506683606654406, + 0.004344522953033447, + 0.004545848350971937, + 0.005950898863375187, + 0.004571766592562199, + 0.004917595535516739, + 0.0033923897426575422, + 0.0015812832862138748, + 0.004802620969712734, + 0.002319039311259985, + 0.005883784499019384, + 0.004380872007459402, + 0.0038906363770365715, + 0.003752835327759385, + 0.0031880061142146587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.014201764017343521 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002770900726318 + ], + "opposite_logprob_tokens": [ + -2.6950278282165527 + ], + "logprob_reward": -0.07002770900726318, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002770900726318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15131233632564545, + 0.1162121519446373, + 0.10513708740472794, + 0.13737940788269043, + 0.039977774024009705, + 0.04537491872906685, + 0.048241470009088516, + 0.1295013278722763, + 0.05315867066383362, + 0.17370490729808807 + ], + "credit_map": [ + 0.013825436122715473, + 0.01509352307766676, + 0.012120158411562443, + 0.014126836322247982, + 0.021315021440386772, + 0.02338084951043129, + 0.029402008280158043, + 0.022048497572541237, + 0.009465336799621582, + 0.014343234710395336, + 0.009844580665230751, + 0.012512250803411007, + 0.018845725804567337, + 0.015380514785647392, + 0.023114679381251335, + 0.012705829925835133, + 0.005647333804517984, + 0.008619534783065319, + 0.004648254252970219, + 0.004288271069526672, + 0.012739666737616062, + 0.032432373613119125, + 0.00960311945527792, + 0.027158526703715324, + 0.003719000145792961, + 0.005568533670157194, + 0.005338700953871012, + 0.00431253295391798, + 0.017655499279499054, + 0.05235671252012253, + 0.00731277372688055, + 0.041115652769804, + 0.004347634967416525, + 0.004462455864995718, + 0.0034831599332392216, + 0.0038599467370659113, + 0.006525774952024221, + 0.004932788200676441, + 0.007116054650396109, + 0.005249958485364914, + 0.00438281986862421, + 0.005887816660106182, + 0.0032329545356333256, + 0.0035499664954841137, + 0.007263257633894682, + 0.005642983131110668, + 0.008865291252732277, + 0.00654982915148139, + 0.0038856472820043564, + 0.006205471232533455, + 0.004440582823008299, + 0.004170077852904797, + 0.010070965625345707, + 0.006293697282671928, + 0.006253536324948072, + 0.006921493913978338, + 0.011234776116907597, + 0.01568732224404812, + 0.01322691235691309, + 0.010418487712740898, + 0.019119467586278915, + 0.01775023713707924, + 0.020675605162978172, + 0.021388519555330276, + 0.006977476645261049, + 0.009139487519860268, + 0.005242159124463797, + 0.006326197646558285, + 0.005225888919085264, + 0.0066319904290139675, + 0.007383864838629961, + 0.006231601350009441, + 0.011497600004076958, + 0.016614172607660294, + 0.008164744824171066, + 0.011455410160124302, + 0.016232317313551903, + 0.045782074332237244, + 0.024380961433053017, + 0.03957761824131012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.07002770900726318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, shot from an extreme close-up and slightly upward, focuses on an expressionless porcelain doll. Its smooth, cold porcelain cheeks, painted with hand-painted pink rouge, have faded slightly with age, while the corners of its mouth maintain a timeless, faint curve. It sits stiffly, clad in a yellowed, lace-trimmed Victorian dress, its carefully coiffed, though slightly dusty, blond curls cascading over its shoulders. Its gaze, through its deep, hollow blue glass eyes, is fixed motionlessly on an old rag doll across from it. The doll, made of rough, faded cotton and linen, its body sagging with age, sits limply, two black buttons serving as eyes, and a slightly crooked smile, stitched with thick red thread, is faintly visible in the shadows. A dim beam of light slants in from outside the painting, penetrating the fine dust particles floating in the air and precisely illuminating half of the porcelain doll's smooth cheek. A cold, lifeless highlight appears on its glass eye, while the rough texture of the doll absorbs most of the light, fading it into deeper shadows. The painting's somber tones, replete with stark contrasts of light and dark, create an atmosphere of unsettling stillness, frozen in time, as if a silent confrontation is eternally unfolding between two inanimate objects.", + "seed": 9618, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_8/index_9333_seed_9618_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.331182661280036, + "vlm_logprob": -0.331182661280036, + "vlm_yes_no_margin": 1.7, + "vlm_token_credit": -0.331182661280036, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9008700847625732 + ], + "opposite_logprob_tokens": [ + -0.5258700847625732 + ], + "logprob_reward": -0.9008700847625732, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9008700847625732, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3568069338798523, + 0.09858156740665436, + 0.03736715763807297, + 0.027992872521281242, + 0.07475107908248901, + 0.051376014947891235, + 0.04341130331158638, + 0.029151393100619316, + 0.09632111340761185, + 0.18424056470394135 + ], + "credit_map": [ + 0.031055502593517303, + 0.02725084312260151, + 0.036559123545885086, + 0.06221634894609451, + 0.04749365895986557, + 0.046520840376615524, + 0.05668124929070473, + 0.04902935028076172, + 0.007881912402808666, + 0.007622445933520794, + 0.009652461856603622, + 0.01823338121175766, + 0.012013787403702736, + 0.013401237316429615, + 0.016502363607287407, + 0.013273978605866432, + 0.0030562058091163635, + 0.002969292690977454, + 0.003731281729415059, + 0.005217812489718199, + 0.004812794737517834, + 0.00525957765057683, + 0.0070591396652162075, + 0.00526105472818017, + 0.0025588099379092455, + 0.001971832010895014, + 0.0024363123811781406, + 0.003579753218218684, + 0.0051895263604819775, + 0.0035781078040599823, + 0.004590545780956745, + 0.004087985958904028, + 0.007648571394383907, + 0.009471885859966278, + 0.00824008695781231, + 0.010045024566352367, + 0.009989750571548939, + 0.009713396430015564, + 0.01147058978676796, + 0.00817177351564169, + 0.00508079631254077, + 0.005318212788552046, + 0.0064535075798630714, + 0.006909582763910294, + 0.01023044716566801, + 0.0063933017663657665, + 0.006068754941225052, + 0.0049214111641049385, + 0.0036212338600307703, + 0.004114960320293903, + 0.00589602580294013, + 0.007469307631254196, + 0.004775333683937788, + 0.0062358020804822445, + 0.0056380280293524265, + 0.005660609342157841, + 0.0028873314149677753, + 0.002142653102055192, + 0.002405100502073765, + 0.003641478018835187, + 0.006213515996932983, + 0.00400666194036603, + 0.004282774869352579, + 0.003571879118680954, + 0.012369156815111637, + 0.006138521246612072, + 0.008803803473711014, + 0.010523824952542782, + 0.020450372248888016, + 0.012439383193850517, + 0.014168495312333107, + 0.011427556164562702, + 0.018070584163069725, + 0.018364084884524345, + 0.017453042790293694, + 0.023186786100268364, + 0.03346776217222214, + 0.023141564801335335, + 0.028986835852265358, + 0.021569903939962387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.9008700847625732 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761767625808716 + ], + "opposite_logprob_tokens": [ + -0.8261767625808716 + ], + "logprob_reward": -0.5761767625808716, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761767625808716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3228151798248291, + 0.13569654524326324, + 0.05997796729207039, + 0.06574499607086182, + 0.08026331663131714, + 0.061027731746435165, + 0.041130512952804565, + 0.04106137529015541, + 0.03780142962932587, + 0.15448099374771118 + ], + "credit_map": [ + 0.03270241990685463, + 0.02316332422196865, + 0.026354458183050156, + 0.05120798945426941, + 0.04962804168462753, + 0.04358392953872681, + 0.052124906331300735, + 0.04405008256435394, + 0.012565599754452705, + 0.008160398341715336, + 0.012003645300865173, + 0.018792305141687393, + 0.019508853554725647, + 0.01822754554450512, + 0.025882240384817123, + 0.020555956289172173, + 0.006229879800230265, + 0.00370811834000051, + 0.0054731848649680614, + 0.007287158630788326, + 0.009799926541745663, + 0.007961551658809185, + 0.01065167784690857, + 0.008866471238434315, + 0.005930377170443535, + 0.004899611696600914, + 0.006240359041839838, + 0.008216852322220802, + 0.009046425111591816, + 0.008673182688653469, + 0.012789064086973667, + 0.009949124418199062, + 0.008274780586361885, + 0.009688714519143105, + 0.010699167847633362, + 0.010659580118954182, + 0.014476154930889606, + 0.009805851615965366, + 0.009853694587945938, + 0.006805373355746269, + 0.005400823894888163, + 0.003956390544772148, + 0.006591896526515484, + 0.006638787221163511, + 0.00897911936044693, + 0.008691380731761456, + 0.011146758683025837, + 0.009622574783861637, + 0.004935715347528458, + 0.004118555691093206, + 0.0029474650509655476, + 0.006815242115408182, + 0.006686038337647915, + 0.0055868844501674175, + 0.005678507499396801, + 0.004362103063613176, + 0.004102214239537716, + 0.0023957216180860996, + 0.004298144020140171, + 0.0041541229002177715, + 0.005262406542897224, + 0.005650334060192108, + 0.008239071816205978, + 0.006959360092878342, + 0.003544636769220233, + 0.003434574231505394, + 0.003536294912919402, + 0.0045731631107628345, + 0.006106903776526451, + 0.0053762332536280155, + 0.006394654978066683, + 0.0048349713906645775, + 0.01612597145140171, + 0.01384095661342144, + 0.013243758119642735, + 0.022147493436932564, + 0.025880033150315285, + 0.020392851904034615, + 0.02487650141119957, + 0.01797342300415039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.5761767625808716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029832107946276665 + ], + "opposite_logprob_tokens": [ + -3.529832124710083 + ], + "logprob_reward": -0.029832107946276665, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029832107946276665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39509886503219604, + 0.11326097697019577, + 0.06322578340768814, + 0.03841879218816757, + 0.05506432056427002, + 0.038104768842458725, + 0.025160888209939003, + 0.037810854613780975, + 0.09135399758815765, + 0.14250071346759796 + ], + "credit_map": [ + 0.036624565720558167, + 0.030968081206083298, + 0.03922298178076744, + 0.061782434582710266, + 0.0520976223051548, + 0.05325503274798393, + 0.06282390654087067, + 0.05832425132393837, + 0.011088746599853039, + 0.008239497430622578, + 0.010266206227242947, + 0.016194432973861694, + 0.015299052931368351, + 0.015256383456289768, + 0.019653698429465294, + 0.01726294681429863, + 0.006834087427705526, + 0.005054238718003035, + 0.006269296631217003, + 0.007973024621605873, + 0.008813120424747467, + 0.008776948787271976, + 0.010347253642976284, + 0.009157813154160976, + 0.004235411994159222, + 0.0029738598968833685, + 0.0030213382560759783, + 0.004586693365126848, + 0.005624039098620415, + 0.005132846999913454, + 0.007322929799556732, + 0.005521673243492842, + 0.006011934485286474, + 0.0049098641611635685, + 0.005942224059253931, + 0.0067243268713355064, + 0.0077321091666817665, + 0.007463203743100166, + 0.008527633734047413, + 0.007753028068691492, + 0.0038916703779250383, + 0.0027408208698034286, + 0.00424814410507679, + 0.004225808195769787, + 0.004021030850708485, + 0.005387180484831333, + 0.007552186958491802, + 0.006037924438714981, + 0.0019266445888206363, + 0.002037877682596445, + 0.0030453004874289036, + 0.004241690039634705, + 0.004119229037314653, + 0.003606450743973255, + 0.0034451792016625404, + 0.002738516079261899, + 0.00373367452993989, + 0.0028078602626919746, + 0.003368680365383625, + 0.0037301306147128344, + 0.007006035652011633, + 0.005386705975979567, + 0.006494048051536083, + 0.005283720791339874, + 0.012090768665075302, + 0.005444257985800505, + 0.0077622695825994015, + 0.00885065644979477, + 0.01786164939403534, + 0.012632607482373714, + 0.014859414659440517, + 0.011852376163005829, + 0.014691594988107681, + 0.011377747170627117, + 0.012559350579977036, + 0.015150079503655434, + 0.03072379156947136, + 0.01821930706501007, + 0.02461809292435646, + 0.015160741284489632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.029832107946276665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004182040691376 + ], + "opposite_logprob_tokens": [ + -2.6950418949127197 + ], + "logprob_reward": -0.07004182040691376, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004182040691376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43731382489204407, + 0.15461651980876923, + 0.044115085154771805, + 0.03479789197444916, + 0.019845154136419296, + 0.020459020510315895, + 0.06668153405189514, + 0.024337850511074066, + 0.04727284982800484, + 0.15056024491786957 + ], + "credit_map": [ + 0.042531583458185196, + 0.031964704394340515, + 0.0416540801525116, + 0.07299398630857468, + 0.05317779630422592, + 0.05801309645175934, + 0.07139745354652405, + 0.06558114290237427, + 0.013724918477237225, + 0.011433959938585758, + 0.015070972964167595, + 0.027571337297558784, + 0.017778649926185608, + 0.021420804783701897, + 0.026008905842900276, + 0.021606985479593277, + 0.0038269537035375834, + 0.0032748717349022627, + 0.004688425920903683, + 0.007151265628635883, + 0.005730564706027508, + 0.006108635570853949, + 0.007236082572489977, + 0.006098288577049971, + 0.0033871042542159557, + 0.0030919057317078114, + 0.004209308885037899, + 0.0050597842782735825, + 0.004163667093962431, + 0.0048026517033576965, + 0.005438055377453566, + 0.004645416513085365, + 0.0017562430584803224, + 0.0016559415962547064, + 0.002432285575196147, + 0.003272718284279108, + 0.00283806468360126, + 0.002668519737198949, + 0.002467392012476921, + 0.0027539916336536407, + 0.0018690708093345165, + 0.001739362720400095, + 0.002307059010490775, + 0.002908202586695552, + 0.002932792529463768, + 0.0026416212785989046, + 0.0028470056131482124, + 0.0032139080576598644, + 0.0070482478477060795, + 0.004312923643738031, + 0.006456795148551464, + 0.009607610292732716, + 0.009775080718100071, + 0.009227657690644264, + 0.011023792438209057, + 0.009229431860148907, + 0.002694434719160199, + 0.0015878018457442522, + 0.001846838160417974, + 0.003533755661919713, + 0.0036923119332641363, + 0.0032546345610171556, + 0.00394056923687458, + 0.0037875063717365265, + 0.005494295619428158, + 0.004075648263096809, + 0.00529552111402154, + 0.005872090347111225, + 0.007598473224788904, + 0.00643205689266324, + 0.006877786014229059, + 0.005626981612294912, + 0.015384061262011528, + 0.016843382269144058, + 0.013828611932694912, + 0.021282320842146873, + 0.02148103155195713, + 0.01907810941338539, + 0.023064563050866127, + 0.019598176702857018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07004182040691376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on an expressionless porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward looking at the doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the doll wearing a yellowed, lace-trimmed Victorian dress with dusty blond curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second rag doll made of rough cotton and linen visible sitting opposite the porcelain doll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a dim beam of light slanting in from outside that illuminates half of the porcelain doll's cheek?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899253070354462 + ], + "opposite_logprob_tokens": [ + -2.5789926052093506 + ], + "logprob_reward": -0.07899253070354462, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899253070354462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3042609393596649, + 0.1064438596367836, + 0.05254271253943443, + 0.035614874213933945, + 0.04775214567780495, + 0.034072618931531906, + 0.05598438158631325, + 0.03759846091270447, + 0.0875868871808052, + 0.23814313113689423 + ], + "credit_map": [ + 0.024077028036117554, + 0.025881526991724968, + 0.02638535015285015, + 0.055246319621801376, + 0.03581484034657478, + 0.04007963463664055, + 0.05311837047338486, + 0.04365786910057068, + 0.007122411392629147, + 0.009154899045825005, + 0.009359847754240036, + 0.016639644280076027, + 0.012098196893930435, + 0.0148744722828269, + 0.020925814285874367, + 0.016268564388155937, + 0.004869847558438778, + 0.004493944812566042, + 0.005047030281275511, + 0.006861373316496611, + 0.006888755597174168, + 0.007462694309651852, + 0.008798494935035706, + 0.008120568469166756, + 0.004574249964207411, + 0.004742830991744995, + 0.004413606133311987, + 0.004151953849941492, + 0.004923232831060886, + 0.004430476110428572, + 0.005179725121706724, + 0.0031987973488867283, + 0.005159816704690456, + 0.005507195368409157, + 0.005586537066847086, + 0.006086857058107853, + 0.008042645640671253, + 0.005975452251732349, + 0.006469834595918655, + 0.004923805128782988, + 0.002714857691898942, + 0.0027564556803554296, + 0.003729119896888733, + 0.003321073716506362, + 0.004925411660224199, + 0.004941747989505529, + 0.006398278288543224, + 0.005285671446472406, + 0.005023722071200609, + 0.004481728188693523, + 0.005039115436375141, + 0.007151065394282341, + 0.007643206510692835, + 0.007933284156024456, + 0.00939914770424366, + 0.009313108399510384, + 0.005654133856296539, + 0.0033754550386220217, + 0.004403094295412302, + 0.006942077539861202, + 0.005681134294718504, + 0.005058723501861095, + 0.0036199879832565784, + 0.0028638511430472136, + 0.012074992060661316, + 0.007609209977090359, + 0.008630689233541489, + 0.010540789924561977, + 0.013935737311840057, + 0.012658322229981422, + 0.011437423527240753, + 0.01069971639662981, + 0.025019504129886627, + 0.02757662907242775, + 0.023802265524864197, + 0.03851409628987312, + 0.038427550345659256, + 0.028647182509303093, + 0.033308226615190506, + 0.02284766547381878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07899253070354462 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_9/index_tensor([8125, 8125])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_9/index_tensor([8125, 8125])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c3c8bbdb068b9742b12290734b65fec0dd64580d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_9/index_tensor([8125, 8125])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium close-up shot, shot from a slightly overhead angle, focuses on an ancient young woman half-reclining on a sickbed. Her face is delicate yet pale, her long, black hair loosely tied, a few strands, damp with sweat, plastered to her smooth forehead. She wears a loose, light pink satin tunic, slightly open at the neck. Suddenly, a violent coughing fit seizes her frail body. Her entire body hunches in agony, her shoulders tensing. She weakly but swiftly raises her right hand and tightly covers her mouth with a pristine white handkerchief, her slender fingers whitening from the pressure. The handkerchief, pressed tightly against her lips, rises and falls slightly with her violent trembling. When the heart-wrenching coughing subsides slightly, she slowly lowers the handkerchief. Unshed tears well in her clear, almond-shaped eyes, a look filled with deep sorrow and helplessness, as she breathes wearily. The scene is illuminated by a blend of two lights: one, the cool, soft natural light filtering through the window paper, outlining her delicate features; the other, the warm, flickering glow of an oil lamp on the dark wooden bedside table, casting dancing shadows across her sickly features, creating an atmosphere of tranquility and melancholy. The entire painting exudes the delicate and elegant texture of a classical silk painting, with subdued, restrained colors and refined, graceful lines, imbued with a heartbreakingly fragile beauty.", + "seed": 21069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_9/index_8125_seed_21069_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02802952565252781, + "vlm_logprob": -0.02802952565252781, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02802952565252781, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02070561982691288 + ], + "opposite_logprob_tokens": [ + -5.395705699920654 + ], + "logprob_reward": -0.02070561982691288, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.02070561982691288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18047697842121124, + 0.08713316917419434, + 0.013998974114656448, + 0.009522441774606705, + 0.054063957184553146, + 0.025888042524456978, + 0.014062270522117615, + 0.4093315601348877, + 0.18299922347068787, + 0.02252342365682125 + ], + "credit_map": [ + 0.03327936679124832, + 0.013181758113205433, + 0.011725331656634808, + 0.013174675405025482, + 0.023904429748654366, + 0.058960773050785065, + 0.009805402718484402, + 0.016445225104689598, + 0.008489219471812248, + 0.00921681709587574, + 0.007239597849547863, + 0.03321628272533417, + 0.005382333882153034, + 0.013112069107592106, + 0.004512938670814037, + 0.00596390338614583, + 0.0014645080082118511, + 0.0013392898254096508, + 0.0010356062557548285, + 0.0029839600902050734, + 0.003605709644034505, + 0.001498244353570044, + 0.0010552277090027928, + 0.0010164270643144846, + 0.0009127483353950083, + 0.0009691687882877886, + 0.0009448083001188934, + 0.0015571963740512729, + 0.0008233081316575408, + 0.0020906964782625437, + 0.0007141249370761216, + 0.0015103889163583517, + 0.002899159211665392, + 0.020519733428955078, + 0.0038779363967478275, + 0.009585198014974594, + 0.0022717416286468506, + 0.004853497724980116, + 0.002735460875555873, + 0.007321223150938749, + 0.001630613929592073, + 0.0019221208058297634, + 0.004369413945823908, + 0.010001227259635925, + 0.0008532975916750729, + 0.004297928418964148, + 0.0014240100281313062, + 0.0013894280418753624, + 0.0015026949113234878, + 0.0021970642264932394, + 0.0017454170156270266, + 0.002109780441969633, + 0.0014708915259689093, + 0.001295774825848639, + 0.002116684801876545, + 0.00162396056111902, + 0.0177563838660717, + 0.03674767538905144, + 0.16815011203289032, + 0.11826406419277191, + 0.005758062470704317, + 0.028568850830197334, + 0.011100691743195057, + 0.02298569306731224, + 0.0072686620987951756, + 0.008240659721195698, + 0.03602217137813568, + 0.07136546820402145, + 0.008210880681872368, + 0.03820401802659035, + 0.0072303949855268, + 0.006456962786614895, + 0.002092247363179922, + 0.004558161366730928, + 0.0030497247353196144, + 0.004273657687008381, + 0.0022079653572291136, + 0.0021507723722606897, + 0.0023447666317224503, + 0.0018461250001564622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.02070561982691288 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060624770820141 + ], + "opposite_logprob_tokens": [ + -3.895606279373169 + ], + "logprob_reward": -0.02060624770820141, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060624770820141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5767877697944641, + 0.054998431354761124, + 0.0473385825753212, + 0.014112447388470173, + 0.023456167429685593, + 0.026980852708220482, + 0.05078130215406418, + 0.08559449017047882, + 0.10353746265172958, + 0.016412479802966118 + ], + "credit_map": [ + 0.12113863229751587, + 0.0393325500190258, + 0.03996706381440163, + 0.03747188672423363, + 0.07145451754331589, + 0.1878565549850464, + 0.029535017907619476, + 0.05003153532743454, + 0.00563776632770896, + 0.00577561417594552, + 0.004789575934410095, + 0.02130974270403385, + 0.004029389005154371, + 0.006604737602174282, + 0.0031499078031629324, + 0.0037016975693404675, + 0.0034163489472121, + 0.004375435411930084, + 0.0036718465853482485, + 0.01205111388117075, + 0.011631444096565247, + 0.004227503202855587, + 0.004109792411327362, + 0.003855097806081176, + 0.0013113307068124413, + 0.001638225745409727, + 0.0016721274005249143, + 0.00343118142336607, + 0.0010189495515078306, + 0.0021361245308071375, + 0.0016320159193128347, + 0.0012724912958219647, + 0.0011077886447310448, + 0.008613198064267635, + 0.0015248351264744997, + 0.004331598058342934, + 0.0012603660579770803, + 0.0021394880022853613, + 0.0013668035389855504, + 0.0031120874918997288, + 0.0016729537164792418, + 0.003548430046066642, + 0.003039015457034111, + 0.008303995244204998, + 0.0019012234406545758, + 0.005068477708846331, + 0.001612884341739118, + 0.001833869144320488, + 0.003921215422451496, + 0.007793900556862354, + 0.005150206852704287, + 0.005887977313250303, + 0.010048160329461098, + 0.009711119346320629, + 0.003859792836010456, + 0.0044089267030358315, + 0.0034042412880808115, + 0.007213128264993429, + 0.033252790570259094, + 0.02549165114760399, + 0.0023412546142935753, + 0.006943638436496258, + 0.002567899413406849, + 0.0043798815459012985, + 0.004289502743631601, + 0.005122150294482708, + 0.013823184184730053, + 0.03573356941342354, + 0.007282755803316832, + 0.02789829671382904, + 0.004372147843241692, + 0.0050158509984612465, + 0.0015050051733851433, + 0.004133233800530434, + 0.0016319017158821225, + 0.0019931804854422808, + 0.0017260779859498143, + 0.0016264010919257998, + 0.0017870523734018207, + 0.00200962508097291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.02060624770820141 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054965127259492874 + ], + "opposite_logprob_tokens": [ + -2.929965019226074 + ], + "logprob_reward": -0.054965127259492874, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054965127259492874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.632101833820343, + 0.12220586836338043, + 0.016733236610889435, + 0.06293725222349167, + 0.033512625843286514, + 0.0176369771361351, + 0.005421576555818319, + 0.02194451354444027, + 0.08000623434782028, + 0.00749979680404067 + ], + "credit_map": [ + 0.12860476970672607, + 0.04289057105779648, + 0.042662832885980606, + 0.045310020446777344, + 0.0777069702744484, + 0.2061445415019989, + 0.03376824036240578, + 0.05501391366124153, + 0.011711752973496914, + 0.012517421506345272, + 0.010168020613491535, + 0.04682714492082596, + 0.007956367917358875, + 0.017844779416918755, + 0.006637459620833397, + 0.008542930707335472, + 0.0014668569201603532, + 0.0023323968052864075, + 0.0015047321794554591, + 0.0038222544826567173, + 0.003665498225018382, + 0.0014346379321068525, + 0.0012676563346758485, + 0.0012392043136060238, + 0.005801299586892128, + 0.006029816810041666, + 0.007628096733242273, + 0.01524203922599554, + 0.005130142904818058, + 0.01146736927330494, + 0.005998014472424984, + 0.005640481133013964, + 0.0022698345128446817, + 0.011819151230156422, + 0.004095757845789194, + 0.004537698347121477, + 0.0021516112610697746, + 0.0025026549119502306, + 0.002311530290171504, + 0.003824390470981598, + 0.001178056700155139, + 0.002849734388291836, + 0.003406676696613431, + 0.003518520388752222, + 0.001271408749744296, + 0.0012085124617442489, + 0.00296325096860528, + 0.0012408173643052578, + 0.0008974330266937613, + 0.0009493024554103613, + 0.0005299110198393464, + 0.0006568875978700817, + 0.0006052099633961916, + 0.0005989083438180387, + 0.000633706571534276, + 0.0005502181011252105, + 0.001232536043971777, + 0.0026388554833829403, + 0.008229343220591545, + 0.0041732871904969215, + 0.0010134631302207708, + 0.002330295043066144, + 0.0008917829254642129, + 0.0014349493430927396, + 0.0037818429991602898, + 0.004473892506211996, + 0.01913994736969471, + 0.03229814022779465, + 0.0017782532377168536, + 0.012004757300019264, + 0.0033026868477463722, + 0.0032267190981656313, + 0.0007615189533680677, + 0.0015328378649428487, + 0.0016503611113876104, + 0.0010624557035043836, + 0.0005517700337804854, + 0.0004736997070722282, + 0.0008430590969510376, + 0.000624094798695296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.054965127259492874 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058989740908146 + ], + "opposite_logprob_tokens": [ + -3.895589828491211 + ], + "logprob_reward": -0.02058989740908146, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058989740908146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35741883516311646, + 0.015764106065034866, + 0.030393371358513832, + 0.023993968963623047, + 0.04328436776995659, + 0.024542193859815598, + 0.0265425443649292, + 0.36932140588760376, + 0.08925861865282059, + 0.0194806270301342 + ], + "credit_map": [ + 0.07751443237066269, + 0.024243833497166634, + 0.024974163621664047, + 0.019577328115701675, + 0.04359511286020279, + 0.11505461484193802, + 0.0192403681576252, + 0.033218979835510254, + 0.0014580230927094817, + 0.0021135290153324604, + 0.0012830960331484675, + 0.004750589374452829, + 0.0016824881313368678, + 0.0017836210317909718, + 0.0014331775018945336, + 0.0012595830485224724, + 0.002074542222544551, + 0.002915202407166362, + 0.0022053273860365152, + 0.007538887206465006, + 0.007741935551166534, + 0.0034755640663206577, + 0.002171854954212904, + 0.002270056866109371, + 0.00209400849416852, + 0.0028527823742479086, + 0.003062271745875478, + 0.005959761328995228, + 0.0022844597697257996, + 0.0024982052855193615, + 0.0024822871200740337, + 0.002760193543508649, + 0.002222212264314294, + 0.016062546521425247, + 0.0033489943016320467, + 0.007160105276852846, + 0.002369479974731803, + 0.003983247559517622, + 0.0022346186451613903, + 0.005903160199522972, + 0.0019375082338228822, + 0.002534126164391637, + 0.004291294142603874, + 0.008372779935598373, + 0.0013753834646195173, + 0.0028867325745522976, + 0.0015247099800035357, + 0.0016196580836549401, + 0.002943771192803979, + 0.0027352566830813885, + 0.0024499516002833843, + 0.0035473371390253305, + 0.005380833987146616, + 0.004759654402732849, + 0.0019656019285321236, + 0.0027601358015090227, + 0.016747457906603813, + 0.03453799709677696, + 0.1522555649280548, + 0.10414556413888931, + 0.005030553322285414, + 0.02506800927221775, + 0.010439998470246792, + 0.02109627239406109, + 0.003598666749894619, + 0.004465920384973288, + 0.012112618423998356, + 0.03126407414674759, + 0.006105415523052216, + 0.023733237758278847, + 0.0038912310265004635, + 0.004087452776730061, + 0.0025261021219193935, + 0.0034782516304403543, + 0.0016929918201640248, + 0.003913240507245064, + 0.001942505594342947, + 0.002890807343646884, + 0.0015774143394082785, + 0.001459314371459186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.02058989740908146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280736058950424 + ], + "opposite_logprob_tokens": [ + -3.773280620574951 + ], + "logprob_reward": -0.023280736058950424, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280736058950424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35116806626319885, + 0.12707918882369995, + 0.04157128185033798, + 0.03488605096936226, + 0.10618073493242264, + 0.04369375482201576, + 0.033088911324739456, + 0.08845680952072144, + 0.14475835859775543, + 0.02911682240664959 + ], + "credit_map": [ + 0.07444386184215546, + 0.024932067841291428, + 0.023949509486556053, + 0.020918842405080795, + 0.04362942650914192, + 0.11391835659742355, + 0.01828988641500473, + 0.031086141243577003, + 0.01089387945830822, + 0.013392479158937931, + 0.010249716229736805, + 0.04736325517296791, + 0.007916557602584362, + 0.020828036591410637, + 0.006721996236592531, + 0.009713280014693737, + 0.0031982334330677986, + 0.004643102642148733, + 0.0027474272064864635, + 0.010397219099104404, + 0.01001571211963892, + 0.0036683969665318727, + 0.0033851999323815107, + 0.0035159930121153593, + 0.0030482313595712185, + 0.003629990853369236, + 0.003635224187746644, + 0.0072615486569702625, + 0.0033906339667737484, + 0.007824135012924671, + 0.002851212862879038, + 0.0032450780272483826, + 0.005036149639636278, + 0.039837513118982315, + 0.006996056530624628, + 0.019858144223690033, + 0.004880038555711508, + 0.010081345215439796, + 0.005196618381887674, + 0.014294885098934174, + 0.00264423293992877, + 0.004145586863160133, + 0.007123275194317102, + 0.014735690318048, + 0.002564246766269207, + 0.00646462244912982, + 0.003736210521310568, + 0.0022798923309892416, + 0.0036993606481701136, + 0.0051553454250097275, + 0.0034963414072990417, + 0.004720780067145824, + 0.005149666219949722, + 0.0042548454366624355, + 0.0025592984165996313, + 0.004053275566548109, + 0.005668416619300842, + 0.009303689002990723, + 0.030065182596445084, + 0.015827704221010208, + 0.006849752273410559, + 0.012102685868740082, + 0.003248251974582672, + 0.005391141399741173, + 0.006082495674490929, + 0.007260281126946211, + 0.018684420734643936, + 0.0496557243168354, + 0.010267239063978195, + 0.03929711505770683, + 0.006491366308182478, + 0.007019717246294022, + 0.002637824509292841, + 0.006469380110502243, + 0.004206699784845114, + 0.005726840812712908, + 0.0024574282579123974, + 0.0025055392179638147, + 0.002782823285087943, + 0.0023302892223000526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.023280736058950424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up shot, shot from a slightly overhead angle, focuses on an ancient young woman half-reclining on a sickbed. Her face is delicate yet pale, her long, black hair loosely tied, a few strands, damp with sweat, plastered to her smooth forehead. She wears a loose, light pink satin tunic, slightly open at the neck. Suddenly, a violent coughing fit seizes her frail body. Her entire body hunches in agony, her shoulders tensing. She weakly but swiftly raises her right hand and tightly covers her mouth with a pristine white handkerchief, her slender fingers whitening from the pressure. The handkerchief, pressed tightly against her lips, rises and falls slightly with her violent trembling. When the heart-wrenching coughing subsides slightly, she slowly lowers the handkerchief. Unshed tears well in her clear, almond-shaped eyes, a look filled with deep sorrow and helplessness, as she breathes wearily. The scene is illuminated by a blend of two lights: one, the cool, soft natural light filtering through the window paper, outlining her delicate features; the other, the warm, flickering glow of an oil lamp on the dark wooden bedside table, casting dancing shadows across her sickly features, creating an atmosphere of tranquility and melancholy. The entire painting exudes the delicate and elegant texture of a classical silk painting, with subdued, restrained colors and refined, graceful lines, imbued with a heartbreakingly fragile beauty.", + "seed": 21070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_9/index_8125_seed_21070_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02594723366200924, + "vlm_logprob": -0.02594723366200924, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02594723366200924, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032271213829517365 + ], + "opposite_logprob_tokens": [ + -5.157271385192871 + ], + "logprob_reward": -0.032271213829517365, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.032271213829517365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15404105186462402, + 0.08137452602386475, + 0.029434718191623688, + 0.04053923115134239, + 0.03676741197705269, + 0.06246674805879593, + 0.15157467126846313, + 0.13602520525455475, + 0.11834997683763504, + 0.1894264817237854 + ], + "credit_map": [ + 0.014655394479632378, + 0.02370767667889595, + 0.014347277581691742, + 0.016582317650318146, + 0.029166094958782196, + 0.019265543669462204, + 0.02187431789934635, + 0.014442440122365952, + 0.0071418615989387035, + 0.008271628990769386, + 0.00870557688176632, + 0.013838897459208965, + 0.008576425723731518, + 0.00910355057567358, + 0.013612900860607624, + 0.012123696506023407, + 0.002341226674616337, + 0.0028865148779004812, + 0.004052703268826008, + 0.003459029132500291, + 0.004245228134095669, + 0.0032219556160271168, + 0.005128844175487757, + 0.004099217243492603, + 0.005004961043596268, + 0.0035924443509429693, + 0.005916874390095472, + 0.005071797873824835, + 0.005164861213415861, + 0.0047891768626868725, + 0.005957217887043953, + 0.005041902884840965, + 0.004224597476422787, + 0.0028036406729370356, + 0.0029992051422595978, + 0.004830309189856052, + 0.003663719864562154, + 0.00477779982611537, + 0.006720297038555145, + 0.0067478446289896965, + 0.005273988004773855, + 0.003070242702960968, + 0.003163533750921488, + 0.005858002230525017, + 0.003504580818116665, + 0.009933297522366047, + 0.012782925739884377, + 0.01888018473982811, + 0.007800344377756119, + 0.003131614299491048, + 0.007413294166326523, + 0.007386037148535252, + 0.010154367424547672, + 0.025039859116077423, + 0.03719877079129219, + 0.053450386971235275, + 0.012605699710547924, + 0.012939805164933205, + 0.010671278461813927, + 0.019683003425598145, + 0.01633748784661293, + 0.012032722122967243, + 0.033019840717315674, + 0.018735371530056, + 0.009228741750121117, + 0.016977526247501373, + 0.009188218042254448, + 0.016380878165364265, + 0.01862623356282711, + 0.01362159289419651, + 0.02080170251429081, + 0.01352508645504713, + 0.012782705947756767, + 0.025155482813715935, + 0.013493912294507027, + 0.022456640377640724, + 0.025685235857963562, + 0.02262972481548786, + 0.02803899347782135, + 0.039183784276247025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.032271213829517365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018206853419542313 + ], + "opposite_logprob_tokens": [ + -4.01820707321167 + ], + "logprob_reward": -0.018206853419542313, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018206853419542313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1713041365146637, + 0.081490159034729, + 0.04001852124929428, + 0.0417456328868866, + 0.0644499659538269, + 0.18555517494678497, + 0.0802149623632431, + 0.13475413620471954, + 0.10012666136026382, + 0.10034061968326569 + ], + "credit_map": [ + 0.015530779957771301, + 0.01510720606893301, + 0.013314740732312202, + 0.033188700675964355, + 0.019774142652750015, + 0.02016567997634411, + 0.022932833060622215, + 0.03129005804657936, + 0.006245381198823452, + 0.006689110305160284, + 0.0076453364454209805, + 0.015476328320801258, + 0.008449374698102474, + 0.008602151647210121, + 0.014012533240020275, + 0.014369938522577286, + 0.0037784173619002104, + 0.0033881014678627253, + 0.0047490899451076984, + 0.00495030777528882, + 0.005095559172332287, + 0.004216415341943502, + 0.007341718301177025, + 0.006498910021036863, + 0.00362911494448781, + 0.003904060460627079, + 0.0044877235777676105, + 0.003250862006098032, + 0.008784124627709389, + 0.005462832748889923, + 0.007128091994673014, + 0.005098822060972452, + 0.005633488297462463, + 0.0031720330007374287, + 0.004588885232806206, + 0.010966494679450989, + 0.005411652848124504, + 0.006450901739299297, + 0.012233614921569824, + 0.015992896631360054, + 0.013319605961441994, + 0.00464037898927927, + 0.006364012137055397, + 0.0138757498934865, + 0.010240335948765278, + 0.03263714909553528, + 0.04273547604680061, + 0.06174246966838837, + 0.010402203537523746, + 0.0052131712436676025, + 0.007018730044364929, + 0.012631088495254517, + 0.009638480842113495, + 0.006734446622431278, + 0.017226487398147583, + 0.011350356973707676, + 0.011085408739745617, + 0.008844486437737942, + 0.010815528221428394, + 0.021588385105133057, + 0.013924866914749146, + 0.013116513378918171, + 0.034465350210666656, + 0.02091359719634056, + 0.007428424432873726, + 0.008274594321846962, + 0.011116623878479004, + 0.015145763754844666, + 0.013712539337575436, + 0.009399479255080223, + 0.02262483909726143, + 0.012424387969076633, + 0.006140741985291243, + 0.004046374931931496, + 0.006793975830078125, + 0.006602344568818808, + 0.009305872023105621, + 0.007135598920285702, + 0.0340205617249012, + 0.02629515342414379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.018206853419542313 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810166195034981 + ], + "opposite_logprob_tokens": [ + -3.2881016731262207 + ], + "logprob_reward": -0.03810166195034981, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810166195034981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13063254952430725, + 0.03625384345650673, + 0.03655082359910011, + 0.05293409526348114, + 0.05108543485403061, + 0.06517310440540314, + 0.11326950043439865, + 0.10110634565353394, + 0.16130565106868744, + 0.2516886293888092 + ], + "credit_map": [ + 0.01361029502004385, + 0.01319038961082697, + 0.00794688519090414, + 0.024202659726142883, + 0.01467338390648365, + 0.014461713843047619, + 0.022373031824827194, + 0.02017420344054699, + 0.004292014986276627, + 0.0032610122580081224, + 0.004181907512247562, + 0.005857525393366814, + 0.004126884043216705, + 0.003981506917625666, + 0.005437665618956089, + 0.005115331616252661, + 0.0036611747927963734, + 0.005970495752990246, + 0.00371505762450397, + 0.004408345092087984, + 0.005099230445921421, + 0.00517779216170311, + 0.004342477302998304, + 0.004176254849880934, + 0.005616999696940184, + 0.004276080057024956, + 0.006662813946604729, + 0.007704244926571846, + 0.005516913719475269, + 0.005444671958684921, + 0.009051909670233727, + 0.008660468272864819, + 0.005073040258139372, + 0.0029838765040040016, + 0.0035267528146505356, + 0.00665917806327343, + 0.0035723161417990923, + 0.005873096641153097, + 0.010683552362024784, + 0.012713623233139515, + 0.005291917361319065, + 0.0034912130795419216, + 0.003635491942986846, + 0.006323480978608131, + 0.005014090333133936, + 0.010062482208013535, + 0.012961331754922867, + 0.018393101170659065, + 0.008302128873765469, + 0.008534816093742847, + 0.011478872038424015, + 0.010695608332753181, + 0.007420998997986317, + 0.01335151493549347, + 0.023154858499765396, + 0.030330715700984, + 0.009605527855455875, + 0.005885858088731766, + 0.012251335196197033, + 0.015036585740745068, + 0.010651174001395702, + 0.00786295160651207, + 0.027878135442733765, + 0.011934778653085232, + 0.016327936202287674, + 0.010530317202210426, + 0.016655389219522476, + 0.031945448368787766, + 0.016574902459979057, + 0.013169347308576107, + 0.03379257023334503, + 0.022309761494398117, + 0.015442080795764923, + 0.010233599692583084, + 0.007746782619506121, + 0.012235361151397228, + 0.01306608971208334, + 0.013436079025268555, + 0.08820287883281708, + 0.09132576733827591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.03810166195034981 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584525540471077 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.020584525540471077, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584525540471077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13029494881629944, + 0.05617937445640564, + 0.024011190980672836, + 0.03699071705341339, + 0.025929169729351997, + 0.15146473050117493, + 0.2000115066766739, + 0.2516336143016815, + 0.028725551441311836, + 0.09475915133953094 + ], + "credit_map": [ + 0.017063146457076073, + 0.018238455057144165, + 0.014147150330245495, + 0.018220597878098488, + 0.019272487610578537, + 0.01510348916053772, + 0.017708670347929, + 0.010540954768657684, + 0.005730144679546356, + 0.0066697243601083755, + 0.006494022905826569, + 0.007953147403895855, + 0.00821837130934, + 0.007005755323916674, + 0.0072898659855127335, + 0.006818345747888088, + 0.002889985917136073, + 0.0015926639316603541, + 0.0018753267358988523, + 0.004151741974055767, + 0.0021036549005657434, + 0.002244628267362714, + 0.003810266265645623, + 0.0053429254330694675, + 0.007070992141962051, + 0.0031636254861950874, + 0.003711760276928544, + 0.006023341789841652, + 0.0039408570155501366, + 0.004167168866842985, + 0.004869632422924042, + 0.004043339751660824, + 0.005166397895663977, + 0.002795568434521556, + 0.0024132304824888706, + 0.003051272826269269, + 0.002535753883421421, + 0.0026374964509159327, + 0.0034527876414358616, + 0.003876663511618972, + 0.01679176092147827, + 0.00719642685726285, + 0.012613930739462376, + 0.009334108792245388, + 0.009443636983633041, + 0.02454555407166481, + 0.029043851420283318, + 0.04249546304345131, + 0.014114288613200188, + 0.007406964898109436, + 0.00840748194605112, + 0.0095801567658782, + 0.013647799380123615, + 0.028874201700091362, + 0.04819187894463539, + 0.06978874653577805, + 0.03573187440633774, + 0.03746607154607773, + 0.02009415812790394, + 0.030644511803984642, + 0.03785062953829765, + 0.024227848276495934, + 0.03412460908293724, + 0.031493932008743286, + 0.002284344518557191, + 0.0026678498834371567, + 0.002673718612641096, + 0.005071708001196384, + 0.003997967578470707, + 0.0027002820279449224, + 0.005568705964833498, + 0.003760977415367961, + 0.006510350853204727, + 0.006675995886325836, + 0.006430808454751968, + 0.010442893020808697, + 0.00837930291891098, + 0.011095833033323288, + 0.017890959978103638, + 0.027333009988069534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.020584525540471077 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571913570165634 + ], + "opposite_logprob_tokens": [ + -3.8955719470977783 + ], + "logprob_reward": -0.020571913570165634, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571913570165634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21359270811080933, + 0.0890682190656662, + 0.05575745552778244, + 0.04129913076758385, + 0.042695995420217514, + 0.03971116617321968, + 0.10750448703765869, + 0.14789319038391113, + 0.09122804552316666, + 0.1712496131658554 + ], + "credit_map": [ + 0.02296319231390953, + 0.018150407820940018, + 0.01832648180425167, + 0.04900997504591942, + 0.021791532635688782, + 0.026534615084528923, + 0.030001524835824966, + 0.026814976707100868, + 0.008323370479047298, + 0.007751602213829756, + 0.008692783303558826, + 0.017337795346975327, + 0.009851930662989616, + 0.010030069388449192, + 0.013619069010019302, + 0.013461606577038765, + 0.0068402341566979885, + 0.006308810785412788, + 0.005839907564222813, + 0.006976970937103033, + 0.008731210604310036, + 0.006529552396386862, + 0.008391005918383598, + 0.006139766424894333, + 0.005352864507585764, + 0.0034454534761607647, + 0.004042237531393766, + 0.007606436498463154, + 0.004450223874300718, + 0.004717797040939331, + 0.006134447176009417, + 0.0055496711283922195, + 0.006503463722765446, + 0.0034689067397266626, + 0.0033587238285690546, + 0.0073293051682412624, + 0.0038237241096794605, + 0.004018730018287897, + 0.006602754816412926, + 0.007590386550873518, + 0.005644940305501223, + 0.005392559338361025, + 0.003155158832669258, + 0.0054271025583148, + 0.005114361178129911, + 0.0054673096165061, + 0.004843149799853563, + 0.004666585475206375, + 0.007584307808429003, + 0.004716155584901571, + 0.006802898831665516, + 0.007778109982609749, + 0.007663988973945379, + 0.016974909231066704, + 0.02349444292485714, + 0.03248966857790947, + 0.012588533572852612, + 0.010707455687224865, + 0.013393502682447433, + 0.023113684728741646, + 0.017972270026803017, + 0.013908441178500652, + 0.037846338003873825, + 0.01836295798420906, + 0.01003308780491352, + 0.010444051586091518, + 0.007718482054769993, + 0.013796785846352577, + 0.013781361281871796, + 0.008871487341821194, + 0.014963088557124138, + 0.01161970105022192, + 0.009963116608560085, + 0.008774129673838615, + 0.007839162833988667, + 0.010907786898314953, + 0.011172985658049583, + 0.01029552798718214, + 0.06122862175107002, + 0.05106829106807709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.020571913570165634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up shot, shot from a slightly overhead angle, focuses on an ancient young woman half-reclining on a sickbed. Her face is delicate yet pale, her long, black hair loosely tied, a few strands, damp with sweat, plastered to her smooth forehead. She wears a loose, light pink satin tunic, slightly open at the neck. Suddenly, a violent coughing fit seizes her frail body. Her entire body hunches in agony, her shoulders tensing. She weakly but swiftly raises her right hand and tightly covers her mouth with a pristine white handkerchief, her slender fingers whitening from the pressure. The handkerchief, pressed tightly against her lips, rises and falls slightly with her violent trembling. When the heart-wrenching coughing subsides slightly, she slowly lowers the handkerchief. Unshed tears well in her clear, almond-shaped eyes, a look filled with deep sorrow and helplessness, as she breathes wearily. The scene is illuminated by a blend of two lights: one, the cool, soft natural light filtering through the window paper, outlining her delicate features; the other, the warm, flickering glow of an oil lamp on the dark wooden bedside table, casting dancing shadows across her sickly features, creating an atmosphere of tranquility and melancholy. The entire painting exudes the delicate and elegant texture of a classical silk painting, with subdued, restrained colors and refined, graceful lines, imbued with a heartbreakingly fragile beauty.", + "seed": 21071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_9/index_8125_seed_21071_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03753373995423317, + "vlm_logprob": -0.03753373995423317, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.03753373995423317, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038432322442531586 + ], + "opposite_logprob_tokens": [ + -4.7884321212768555 + ], + "logprob_reward": -0.038432322442531586, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.038432322442531586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1754976212978363, + 0.07368510961532593, + 0.1074971929192543, + 0.08857077360153198, + 0.0801498293876648, + 0.09836190938949585, + 0.10288778692483902, + 0.07777440547943115, + 0.10081032663583755, + 0.0947650820016861 + ], + "credit_map": [ + 0.016594944521784782, + 0.01148228906095028, + 0.01386579591780901, + 0.02934129536151886, + 0.020451931282877922, + 0.015323400497436523, + 0.03909597545862198, + 0.02934197708964348, + 0.01085249986499548, + 0.006898314226418734, + 0.007049208506941795, + 0.014571047388017178, + 0.005228791385889053, + 0.004494689870625734, + 0.012377175502479076, + 0.012213381938636303, + 0.017760707065463066, + 0.01799255982041359, + 0.011377226561307907, + 0.013354666531085968, + 0.008634988218545914, + 0.008639904670417309, + 0.016159018501639366, + 0.013578120619058609, + 0.010741944424808025, + 0.007250546012073755, + 0.008551361039280891, + 0.017308689653873444, + 0.011660825461149216, + 0.006520620081573725, + 0.015328719280660152, + 0.011208068579435349, + 0.006101762875914574, + 0.00404579471796751, + 0.008559908717870712, + 0.01892637461423874, + 0.004967300221323967, + 0.005756861064583063, + 0.017657333984971046, + 0.014134502038359642, + 0.013070463202893734, + 0.013926700688898563, + 0.011479301378130913, + 0.019580792635679245, + 0.007419347297400236, + 0.0055360859259963036, + 0.013852668926119804, + 0.013496543280780315, + 0.013348999433219433, + 0.009114595130085945, + 0.009005694650113583, + 0.016988730058073997, + 0.006897979415953159, + 0.006780358497053385, + 0.026513122022151947, + 0.014238305389881134, + 0.012282307259738445, + 0.00733577087521553, + 0.007406824734061956, + 0.012224440462887287, + 0.0052804783917963505, + 0.0059187826700508595, + 0.01709405891597271, + 0.01023174449801445, + 0.01478869654238224, + 0.010802090167999268, + 0.008320649154484272, + 0.013699125498533249, + 0.00835275836288929, + 0.007814668118953705, + 0.023459868505597115, + 0.013572469353675842, + 0.00950448215007782, + 0.015572513453662395, + 0.007132468279451132, + 0.013832629658281803, + 0.00910657923668623, + 0.012200023047626019, + 0.015993403270840645, + 0.011422980576753616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.038432322442531586 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029831530526280403 + ], + "opposite_logprob_tokens": [ + -3.529831647872925 + ], + "logprob_reward": -0.029831530526280403, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029831530526280403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13344594836235046, + 0.1040060818195343, + 0.13040810823440552, + 0.07604007422924042, + 0.1979409158229828, + 0.10586351156234741, + 0.08136797696352005, + 0.09074858576059341, + 0.05096631869673729, + 0.029212461784482002 + ], + "credit_map": [ + 0.012400954030454159, + 0.008806319907307625, + 0.013430042192339897, + 0.026645241305232048, + 0.010385476984083652, + 0.009069873951375484, + 0.029953327029943466, + 0.022754712030291557, + 0.010874977335333824, + 0.006904326844960451, + 0.010648696683347225, + 0.023875655606389046, + 0.00793884415179491, + 0.005965108517557383, + 0.021375929936766624, + 0.01642254739999771, + 0.011117581278085709, + 0.008050301112234592, + 0.01020036730915308, + 0.026028204709291458, + 0.0074579776264727116, + 0.00921104196459055, + 0.037321142852306366, + 0.021021489053964615, + 0.007547570392489433, + 0.006257642526179552, + 0.009251090697944164, + 0.017681671306490898, + 0.008004679344594479, + 0.0052553568966686726, + 0.011028122156858444, + 0.011013942770659924, + 0.020122315734624863, + 0.019999846816062927, + 0.02165761962532997, + 0.023866156116127968, + 0.02838234417140484, + 0.02167733572423458, + 0.03601151704788208, + 0.026223773136734962, + 0.0127397570759058, + 0.009743757545948029, + 0.011852780357003212, + 0.017546722665429115, + 0.008367872796952724, + 0.0063231587409973145, + 0.02395753189921379, + 0.015331937931478024, + 0.005117964465171099, + 0.004403076134622097, + 0.006180805154144764, + 0.013010028749704361, + 0.004526169504970312, + 0.0061409915797412395, + 0.027371615171432495, + 0.014617319218814373, + 0.011120068840682507, + 0.00945617351680994, + 0.008129674941301346, + 0.019820643588900566, + 0.0058454424142837524, + 0.005392777267843485, + 0.018415654078125954, + 0.012568152509629726, + 0.005369385704398155, + 0.008241386152803898, + 0.004643612075597048, + 0.0070913005620241165, + 0.007462238427251577, + 0.004997655749320984, + 0.007656709756702185, + 0.005504028871655464, + 0.0051282672211527824, + 0.0036391038447618484, + 0.002377673052251339, + 0.004514685366302729, + 0.0019663588609546423, + 0.0033241696655750275, + 0.004803900606930256, + 0.003458302700892091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.029831530526280403 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495542660355568 + ], + "opposite_logprob_tokens": [ + -2.92995548248291 + ], + "logprob_reward": -0.05495542660355568, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495542660355568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24851463735103607, + 0.08654626458883286, + 0.1112842708826065, + 0.10645407438278198, + 0.16932600736618042, + 0.06124996766448021, + 0.031857799738645554, + 0.09352223575115204, + 0.06734559684991837, + 0.023899130523204803 + ], + "credit_map": [ + 0.023263445124030113, + 0.018430208787322044, + 0.01743072271347046, + 0.03707956150174141, + 0.03796803951263428, + 0.019633824005723, + 0.05466652289032936, + 0.040042340755462646, + 0.008817896246910095, + 0.007905097678303719, + 0.008209643885493279, + 0.012304681353271008, + 0.0068377479910850525, + 0.008854644373059273, + 0.021018002182245255, + 0.012598556466400623, + 0.011041704565286636, + 0.007338453084230423, + 0.008511528372764587, + 0.02032860368490219, + 0.0075501855462789536, + 0.011013241484761238, + 0.027698297053575516, + 0.017802270129323006, + 0.008914656937122345, + 0.004740249831229448, + 0.012451241724193096, + 0.02746620774269104, + 0.00624428316950798, + 0.006802321411669254, + 0.021423107013106346, + 0.018412023782730103, + 0.020239967852830887, + 0.018741631880402565, + 0.019063374027609825, + 0.027540042996406555, + 0.026792751625180244, + 0.01216772012412548, + 0.02443455532193184, + 0.020345985889434814, + 0.006445949897170067, + 0.0066239493899047375, + 0.007092901971191168, + 0.011265912093222141, + 0.007310180924832821, + 0.004256713204085827, + 0.010020687244832516, + 0.008233681321144104, + 0.0040099346078932285, + 0.0026698033325374126, + 0.004337858408689499, + 0.004955299664288759, + 0.0022371457889676094, + 0.00316444574855268, + 0.0060163987800478935, + 0.004466916900128126, + 0.008947386406362057, + 0.007740229368209839, + 0.0078069670125842094, + 0.014575903303921223, + 0.0072737811133265495, + 0.007519765757024288, + 0.026344768702983856, + 0.013313435949385166, + 0.005808404181152582, + 0.005271285306662321, + 0.005313120316714048, + 0.010491413995623589, + 0.0046010869555175304, + 0.005548554938286543, + 0.020148906856775284, + 0.010162828490138054, + 0.0035633398219943047, + 0.002886429661884904, + 0.0021653533913195133, + 0.0030964117031544447, + 0.0032619426492601633, + 0.0015941259916871786, + 0.004465782083570957, + 0.002865745220333338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.05495542660355568 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810705617070198 + ], + "opposite_logprob_tokens": [ + -3.28810715675354 + ], + "logprob_reward": -0.03810705617070198, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810705617070198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17425814270973206, + 0.09151958674192429, + 0.10594898462295532, + 0.10854879021644592, + 0.14627236127853394, + 0.0400146059691906, + 0.07110666483640671, + 0.0662173330783844, + 0.09490282088518143, + 0.10121071338653564 + ], + "credit_map": [ + 0.025792844593524933, + 0.014485323801636696, + 0.022573266178369522, + 0.027994347736239433, + 0.024123581126332283, + 0.010953201912343502, + 0.02558021992444992, + 0.02275536023080349, + 0.013978051953017712, + 0.008197867311537266, + 0.008985446766018867, + 0.012039467692375183, + 0.011414328590035439, + 0.007793241646140814, + 0.01628100872039795, + 0.012830173596739769, + 0.01895836554467678, + 0.012946227565407753, + 0.012971174903213978, + 0.016820430755615234, + 0.008666097186505795, + 0.006043630186468363, + 0.015105028636753559, + 0.014438029378652573, + 0.010251675732433796, + 0.005295331124216318, + 0.012193955481052399, + 0.02758946642279625, + 0.00629529170691967, + 0.006525930482894182, + 0.021673141047358513, + 0.018723998218774796, + 0.01787565089762211, + 0.017732374370098114, + 0.017236672341823578, + 0.017657706514000893, + 0.023081598803400993, + 0.014829014427959919, + 0.021425984799861908, + 0.016433365643024445, + 0.004013646859675646, + 0.0039963251911103725, + 0.004748732782900333, + 0.00653764046728611, + 0.003373062936589122, + 0.0037505594082176685, + 0.007464938331395388, + 0.0061296988278627396, + 0.012308974750339985, + 0.008851154707372189, + 0.008645706810057163, + 0.010830540210008621, + 0.005673666950315237, + 0.005712379701435566, + 0.010562055744230747, + 0.008522185496985912, + 0.007599587086588144, + 0.005718736443668604, + 0.0056678722612559795, + 0.010052644647657871, + 0.00427600322291255, + 0.005713724531233311, + 0.01786976121366024, + 0.009319006465375423, + 0.007034446112811565, + 0.009791849181056023, + 0.005783033557236195, + 0.01435727160423994, + 0.009919431060552597, + 0.005388119723647833, + 0.02875273860991001, + 0.013875934295356274, + 0.020971860736608505, + 0.012296161614358425, + 0.013309483416378498, + 0.014301765710115433, + 0.005917547270655632, + 0.007980020716786385, + 0.01458802167326212, + 0.01184585876762867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.03810705617070198 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263423640280962 + ], + "opposite_logprob_tokens": [ + -3.6513423919677734 + ], + "logprob_reward": -0.0263423640280962, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263423640280962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24124407768249512, + 0.09672217816114426, + 0.10439636558294296, + 0.047877728939056396, + 0.12409929186105728, + 0.09030521661043167, + 0.0955454558134079, + 0.10673937201499939, + 0.053172897547483444, + 0.039897311478853226 + ], + "credit_map": [ + 0.02778264507651329, + 0.015872841700911522, + 0.02107098139822483, + 0.04797756299376488, + 0.04132981598377228, + 0.022554071620106697, + 0.03293387219309807, + 0.03172234073281288, + 0.01050108578056097, + 0.008925283327698708, + 0.009787001647055149, + 0.018297968432307243, + 0.012021800503134727, + 0.008008225820958614, + 0.015357925556600094, + 0.013822907581925392, + 0.015417976304888725, + 0.014674142934381962, + 0.010394022800028324, + 0.016748864203691483, + 0.0069168685004115105, + 0.006265733391046524, + 0.018920645117759705, + 0.015058132819831371, + 0.007469301111996174, + 0.006032598204910755, + 0.0055830273777246475, + 0.007585694082081318, + 0.006410347763448954, + 0.0035677822306752205, + 0.005945773795247078, + 0.005283214617520571, + 0.015216133557260036, + 0.012042087502777576, + 0.014326660893857479, + 0.019938986748456955, + 0.016745608299970627, + 0.009018480777740479, + 0.02069457620382309, + 0.016116788610816002, + 0.012370300479233265, + 0.011867227032780647, + 0.010001207701861858, + 0.011624027974903584, + 0.005167708266526461, + 0.007077851332724094, + 0.019364675506949425, + 0.012832238338887691, + 0.010885951109230518, + 0.008157323114573956, + 0.009702966548502445, + 0.012746053747832775, + 0.006076558027416468, + 0.008614442311227322, + 0.0246003158390522, + 0.014761866070330143, + 0.010024637915194035, + 0.010537749156355858, + 0.008739444427192211, + 0.01821935921907425, + 0.008012151345610619, + 0.009364205412566662, + 0.027358680963516235, + 0.014483173377811909, + 0.007939325645565987, + 0.005403915420174599, + 0.003486134810373187, + 0.011300954967737198, + 0.004025239963084459, + 0.004603156819939613, + 0.009510883130133152, + 0.006903296336531639, + 0.005774165969341993, + 0.003824472427368164, + 0.004933704622089863, + 0.00738733634352684, + 0.0033692389260977507, + 0.0032161849085241556, + 0.006150897592306137, + 0.005241319537162781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0263423640280962 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up shot, shot from a slightly overhead angle, focuses on an ancient young woman half-reclining on a sickbed. Her face is delicate yet pale, her long, black hair loosely tied, a few strands, damp with sweat, plastered to her smooth forehead. She wears a loose, light pink satin tunic, slightly open at the neck. Suddenly, a violent coughing fit seizes her frail body. Her entire body hunches in agony, her shoulders tensing. She weakly but swiftly raises her right hand and tightly covers her mouth with a pristine white handkerchief, her slender fingers whitening from the pressure. The handkerchief, pressed tightly against her lips, rises and falls slightly with her violent trembling. When the heart-wrenching coughing subsides slightly, she slowly lowers the handkerchief. Unshed tears well in her clear, almond-shaped eyes, a look filled with deep sorrow and helplessness, as she breathes wearily. The scene is illuminated by a blend of two lights: one, the cool, soft natural light filtering through the window paper, outlining her delicate features; the other, the warm, flickering glow of an oil lamp on the dark wooden bedside table, casting dancing shadows across her sickly features, creating an atmosphere of tranquility and melancholy. The entire painting exudes the delicate and elegant texture of a classical silk painting, with subdued, restrained colors and refined, graceful lines, imbued with a heartbreakingly fragile beauty.", + "seed": 21072, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000089/group_9/index_8125_seed_21072_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023563284799456595, + "vlm_logprob": -0.023563284799456595, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.023563284799456595, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02071157470345497 + ], + "opposite_logprob_tokens": [ + -5.395711421966553 + ], + "logprob_reward": -0.02071157470345497, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.02071157470345497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4544960856437683, + 0.06550699472427368, + 0.06525023281574249, + 0.0162515789270401, + 0.16010567545890808, + 0.12348388135433197, + 0.031883224844932556, + 0.036190591752529144, + 0.023757241666316986, + 0.023074541240930557 + ], + "credit_map": [ + 0.0377090647816658, + 0.020286494866013527, + 0.09821279346942902, + 0.03240230679512024, + 0.01766907423734665, + 0.03562448173761368, + 0.050961386412382126, + 0.16163048148155212, + 0.006837265100330114, + 0.009764148853719234, + 0.011494340375065804, + 0.005142662674188614, + 0.0066949413157999516, + 0.005385986994951963, + 0.0057347179390490055, + 0.014452928677201271, + 0.00791811104863882, + 0.014530960470438004, + 0.006459173746407032, + 0.006137196905910969, + 0.009092732332646847, + 0.0072374832816421986, + 0.007413787767291069, + 0.006460784003138542, + 0.0019285596208646894, + 0.001731163589283824, + 0.002260844688862562, + 0.0013565297704190016, + 0.0025207323487848043, + 0.002392786554992199, + 0.002443106146529317, + 0.0016178570222109556, + 0.01248369924724102, + 0.004673874005675316, + 0.041942015290260315, + 0.005505736917257309, + 0.010087189264595509, + 0.07042278349399567, + 0.008110021241009235, + 0.006880365312099457, + 0.009840345941483974, + 0.007688083220273256, + 0.026073448359966278, + 0.00563886109739542, + 0.009925952181220055, + 0.007240684702992439, + 0.044916607439517975, + 0.012159899808466434, + 0.005079495720565319, + 0.00302594224922359, + 0.006337438710033894, + 0.0016670823097229004, + 0.00288849207572639, + 0.0030170718673616648, + 0.006657895632088184, + 0.0032098081428557634, + 0.003935669083148241, + 0.006033027078956366, + 0.006514066830277443, + 0.00224613631144166, + 0.00435128016397357, + 0.0043868194334208965, + 0.0045121824368834496, + 0.004211412742733955, + 0.002884046407416463, + 0.00384645932354033, + 0.003845429979264736, + 0.0016839290037751198, + 0.0030243624933063984, + 0.003077959641814232, + 0.0030863333959132433, + 0.0023087188601493835, + 0.0027111710514873266, + 0.004090025555342436, + 0.0029390098061412573, + 0.001496048178523779, + 0.0037609117571264505, + 0.003211519680917263, + 0.003018702380359173, + 0.0018471548100933433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.02071157470345497 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008657541126012802 + ], + "opposite_logprob_tokens": [ + -4.758657455444336 + ], + "logprob_reward": -0.008657541126012802, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008657541126012802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1352638304233551, + 0.1092400774359703, + 0.06119104474782944, + 0.017912430688738823, + 0.19498446583747864, + 0.33741074800491333, + 0.02205435000360012, + 0.055366553366184235, + 0.0316842682659626, + 0.034892208874225616 + ], + "credit_map": [ + 0.010460861027240753, + 0.011147628538310528, + 0.018707288429141045, + 0.008762224577367306, + 0.009030421264469624, + 0.019862735643982887, + 0.01556765753775835, + 0.041725024580955505, + 0.010106132365763187, + 0.007148663979023695, + 0.023332461714744568, + 0.007027864921838045, + 0.007890376262366772, + 0.008427773602306843, + 0.011500973254442215, + 0.033805835992097855, + 0.006563200149685144, + 0.013006164692342281, + 0.007045211736112833, + 0.006746323779225349, + 0.009865152649581432, + 0.005288267508149147, + 0.006533314473927021, + 0.006143415812402964, + 0.0016092545120045543, + 0.003050884697586298, + 0.001940294518135488, + 0.001404763082973659, + 0.0028051140252500772, + 0.0026938498485833406, + 0.002515735337510705, + 0.0018925364129245281, + 0.015264447778463364, + 0.005905927158892155, + 0.051952384412288666, + 0.0066949171014130116, + 0.013200022280216217, + 0.08484269678592682, + 0.008480059914290905, + 0.008644006215035915, + 0.02636057138442993, + 0.020106660202145576, + 0.06412231177091599, + 0.014722769148647785, + 0.026088446378707886, + 0.027514057233929634, + 0.12485606968402863, + 0.033639874309301376, + 0.004233629908412695, + 0.002862227149307728, + 0.003333739936351776, + 0.0016728545306250453, + 0.0025531949941068888, + 0.0022481963969767094, + 0.0027947567868977785, + 0.002355752279981971, + 0.007450102362781763, + 0.007407135330140591, + 0.010173858143389225, + 0.003845298895612359, + 0.007884752936661243, + 0.006553820800036192, + 0.006590680219233036, + 0.005460907705128193, + 0.003153269411996007, + 0.006916136480867863, + 0.00293936301022768, + 0.0020106059964746237, + 0.004044022411108017, + 0.004777542781084776, + 0.002864729380235076, + 0.0049786027520895, + 0.003709295764565468, + 0.0076178996823728085, + 0.0058342935517430305, + 0.0024343403056263924, + 0.00469962926581502, + 0.004984019324183464, + 0.0030722699593752623, + 0.0025404621846973896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.008657541126012802 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002348452806473 + ], + "opposite_logprob_tokens": [ + -2.695023536682129 + ], + "logprob_reward": -0.07002348452806473, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002348452806473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.342501163482666, + 0.07802187651395798, + 0.04789268225431442, + 0.05907990410923958, + 0.1509079486131668, + 0.17212891578674316, + 0.03401030972599983, + 0.05345657467842102, + 0.033643290400505066, + 0.028357410803437233 + ], + "credit_map": [ + 0.028972912579774857, + 0.02081434428691864, + 0.07285591959953308, + 0.02523624897003174, + 0.01995353400707245, + 0.03308945894241333, + 0.034909624606370926, + 0.10666908323764801, + 0.007583871949464083, + 0.015391852706670761, + 0.012818762101233006, + 0.007609531283378601, + 0.009445649571716785, + 0.009912511333823204, + 0.007841164246201515, + 0.007418527267873287, + 0.004205077420920134, + 0.010166964493691921, + 0.00490593584254384, + 0.004224387928843498, + 0.0068797157146036625, + 0.006217591464519501, + 0.004292154684662819, + 0.0070008509792387486, + 0.0042119743302464485, + 0.013192168436944485, + 0.007389964070171118, + 0.00503893056884408, + 0.011682795360684395, + 0.006113188341259956, + 0.005595908500254154, + 0.0058549693785607815, + 0.012667716480791569, + 0.004461536183953285, + 0.041853003203868866, + 0.005171454045921564, + 0.00958620198071003, + 0.06489969789981842, + 0.006170656066387892, + 0.0060976771637797356, + 0.01277219783514738, + 0.008327086456120014, + 0.026213739067316055, + 0.006065789144486189, + 0.012259098701179028, + 0.024665461853146553, + 0.06487517058849335, + 0.016950372606515884, + 0.003948655910789967, + 0.005022726953029633, + 0.0054025957360863686, + 0.002783543197438121, + 0.004487654194235802, + 0.006782955024391413, + 0.003459424478933215, + 0.002122750971466303, + 0.005794305354356766, + 0.011287730187177658, + 0.0059037054888904095, + 0.0040109120309352875, + 0.007833307608962059, + 0.00516345351934433, + 0.006114252842962742, + 0.007348902523517609, + 0.0047539519146084785, + 0.005979970563203096, + 0.004777096211910248, + 0.0021816189400851727, + 0.0051140375435352325, + 0.0037696443032473326, + 0.0037748597096651793, + 0.003292107954621315, + 0.0038565436843782663, + 0.005317486822605133, + 0.003044233890250325, + 0.002142594661563635, + 0.005024699494242668, + 0.0038602089043706656, + 0.0028591337613761425, + 0.0022525088861584663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.07002348452806473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645131252706051 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008645131252706051, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645131252706051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.256208598613739, + 0.01978183165192604, + 0.06245110183954239, + 0.019356412813067436, + 0.13904878497123718, + 0.1364450454711914, + 0.14134624600410461, + 0.17111270129680634, + 0.042515795677900314, + 0.011733464896678925 + ], + "credit_map": [ + 0.020236339420080185, + 0.015349596738815308, + 0.05241528898477554, + 0.019890084862709045, + 0.014160092920064926, + 0.0218178853392601, + 0.027888953685760498, + 0.08445034921169281, + 0.0019509458215907216, + 0.0038686515763401985, + 0.0028190959710627794, + 0.0018929608631879091, + 0.0025451246183365583, + 0.0024717648047953844, + 0.0021243358496576548, + 0.002108953660354018, + 0.007057019043713808, + 0.011699067428708076, + 0.007296639960259199, + 0.00633458886295557, + 0.008558000437915325, + 0.007566447835415602, + 0.006637001410126686, + 0.007302336394786835, + 0.0019747538026422262, + 0.0032935410272330046, + 0.002572528552263975, + 0.001868633204139769, + 0.003050797153264284, + 0.0030385165009647608, + 0.0018371287733316422, + 0.0017205142648890615, + 0.011441411450505257, + 0.0037612204905599356, + 0.03605683147907257, + 0.004771144594997168, + 0.00877988338470459, + 0.061163872480392456, + 0.007176603190600872, + 0.005897809285670519, + 0.009619705379009247, + 0.006265263073146343, + 0.01481341291218996, + 0.004595090635120869, + 0.009727113880217075, + 0.028543943539261818, + 0.050196755677461624, + 0.012683757580816746, + 0.014149896800518036, + 0.02055140770971775, + 0.020479736849665642, + 0.009158158674836159, + 0.016952358186244965, + 0.013580050319433212, + 0.034425899386405945, + 0.012048738077282906, + 0.017687689512968063, + 0.01921052299439907, + 0.030014153569936752, + 0.010208051651716232, + 0.024879736825823784, + 0.034686941653490067, + 0.01856842264533043, + 0.015857195481657982, + 0.0036347282584756613, + 0.004699279088526964, + 0.007427436299622059, + 0.0023211075458675623, + 0.004570428282022476, + 0.0048386151902377605, + 0.004448967520147562, + 0.010575233958661556, + 0.0012156893499195576, + 0.0021854452788829803, + 0.0025272523052990437, + 0.0007125771371647716, + 0.001457013888284564, + 0.0017537754029035568, + 0.0009805584559217095, + 0.0009011532529257238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.008645131252706051 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an ancient young woman half-reclining on a sickbed wearing a loose light pink satin tunic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman perform a violent coughing fit where she hunches, raises her right hand to cover her mouth with a white handkerchief, and then lowers it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by cool natural light from a window and a warm flickering oil lamp on a dark wooden bedside table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit the texture and style of a classical silk painting with subdued colors and refined lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977869238704443 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.00977869238704443, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977869238704443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15902143716812134, + 0.0702688917517662, + 0.053488217294216156, + 0.027647715061903, + 0.05027981474995613, + 0.41295960545539856, + 0.04630565643310547, + 0.08161967992782593, + 0.07553843408823013, + 0.022870544344186783 + ], + "credit_map": [ + 0.012982838787138462, + 0.01370218489319086, + 0.026005953550338745, + 0.012111566960811615, + 0.010186287574470043, + 0.015594340860843658, + 0.016484906896948814, + 0.05195334926247597, + 0.005439751781523228, + 0.010632742196321487, + 0.00850716046988964, + 0.006269670557230711, + 0.008469106629490852, + 0.006315330043435097, + 0.008449021726846695, + 0.01618611253798008, + 0.00458554457873106, + 0.010897597298026085, + 0.006002670619636774, + 0.004472195170819759, + 0.010516989976167679, + 0.007040964439511299, + 0.0046321130357682705, + 0.005340140778571367, + 0.002301247324794531, + 0.0049532935954630375, + 0.0034522314090281725, + 0.0022088198456913233, + 0.004533137194812298, + 0.0036602874752134085, + 0.003618472022935748, + 0.0029202294535934925, + 0.004632576368749142, + 0.006212763022631407, + 0.00932285562157631, + 0.0023941444233059883, + 0.0064936731941998005, + 0.012635278515517712, + 0.004577629268169403, + 0.004010898992419243, + 0.029961060732603073, + 0.02410344034433365, + 0.0634743720293045, + 0.016274960711598396, + 0.031034786254167557, + 0.05785391852259636, + 0.15101733803749084, + 0.03923971578478813, + 0.006196242291480303, + 0.007332897745072842, + 0.00730692595243454, + 0.003312397748231888, + 0.0068704127334058285, + 0.0046068779192864895, + 0.005426978692412376, + 0.005252928007394075, + 0.0070854676887393, + 0.015883183106780052, + 0.011317899450659752, + 0.004568338394165039, + 0.011987926438450813, + 0.007634232752025127, + 0.008204417303204536, + 0.014938217587769032, + 0.007107544224709272, + 0.010763302445411682, + 0.012068798765540123, + 0.005092711187899113, + 0.007971025072038174, + 0.008665579371154308, + 0.007604900747537613, + 0.016264572739601135, + 0.0026513757184147835, + 0.0054751234129071236, + 0.00275669083930552, + 0.0017557785613462329, + 0.0032835379242897034, + 0.0033573093824088573, + 0.0019314070232212543, + 0.0016593228792771697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.00977869238704443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_0/index_tensor([9434, 9434])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_0/index_tensor([9434, 9434])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1f730298758dba35fe7531c756370557a18c8087 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_0/index_tensor([9434, 9434])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from an extremely low angle, the camera slowly tilts upward, following the subject's movements. On a damp, slightly pebbled ground bathed in cool moonlight, an elegant black cat crouches compactly. Its glossy black fur, like smooth silk, reflects a faint silver sheen in the dim light, outlining the muscular lines beneath its back and shoulder blades. Initially, its head is slightly lowered, then, with a steady and deliberate rhythm, it slowly raises its head, finally fixing its gaze on the night sky. In this movement, its jewel-like emerald green eyes are fully revealed, their clear pupils clearly reflecting the bright full moon, as if absorbing the entire night sky's brilliance. Its gaze is focused and serene. The entire scene, a stark contrast between the deep black and the brilliant silver and emerald green, creates an atmosphere of mystery, tranquility, and spirituality.", + "seed": 29584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_0/index_9434_seed_29584_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012728845467790961, + "vlm_logprob": -0.012728845467790961, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.012728845467790961, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0286283940076828 + ], + "opposite_logprob_tokens": [ + -4.403628349304199 + ], + "logprob_reward": -0.0286283940076828, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0286283940076828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15269562602043152, + 0.06357129663228989, + 0.05929455906152725, + 0.030624819919466972, + 0.1094968393445015, + 0.10880641639232635, + 0.2800242602825165, + 0.025010384619235992, + 0.04275759309530258, + 0.12771813571453094 + ], + "credit_map": [ + 0.017772773280739784, + 0.019341839477419853, + 0.017173079773783684, + 0.017905473709106445, + 0.02163797989487648, + 0.01647529937326908, + 0.02331569604575634, + 0.019073480740189552, + 0.008793175220489502, + 0.0035156086087226868, + 0.005972500424832106, + 0.003877688432112336, + 0.015037024393677711, + 0.006467674858868122, + 0.006619044113904238, + 0.013288584537804127, + 0.010309597477316856, + 0.0048230779357254505, + 0.003666400443762541, + 0.004480286035686731, + 0.004974556155502796, + 0.008874544873833656, + 0.005440821871161461, + 0.01672527752816677, + 0.004194109234958887, + 0.003649895777925849, + 0.004785899538546801, + 0.0027052101213485003, + 0.0031387396156787872, + 0.0044852569699287415, + 0.0017155737150460482, + 0.005950133316218853, + 0.014087362214922905, + 0.011378257535398006, + 0.01720523089170456, + 0.010491193272173405, + 0.017750106751918793, + 0.01196887344121933, + 0.013470279984176159, + 0.013145534321665764, + 0.007997974753379822, + 0.00972280465066433, + 0.008923033252358437, + 0.018937988206744194, + 0.014199487864971161, + 0.02076779492199421, + 0.011412417516112328, + 0.01684490777552128, + 0.06876864284276962, + 0.010077658109366894, + 0.015152049250900745, + 0.02646331861615181, + 0.014021297916769981, + 0.018570221960544586, + 0.022307327017188072, + 0.10466375201940536, + 0.003824211424216628, + 0.002184111624956131, + 0.0024490985088050365, + 0.0022057814057916403, + 0.002037061844021082, + 0.0023470751475542784, + 0.006182756274938583, + 0.0037802865263074636, + 0.006051892414689064, + 0.0036166890058666468, + 0.004443800076842308, + 0.003591295098885894, + 0.0050708334892988205, + 0.006948920898139477, + 0.007544033229351044, + 0.005490127019584179, + 0.039151452481746674, + 0.004482550546526909, + 0.003669228171929717, + 0.006039626896381378, + 0.0036567167844623327, + 0.04434943199157715, + 0.012511459179222584, + 0.013857663609087467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.0286283940076828 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637290023267269 + ], + "opposite_logprob_tokens": [ + -3.6513729095458984 + ], + "logprob_reward": -0.02637290023267269, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637290023267269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12393695116043091, + 0.038192667067050934, + 0.03193815425038338, + 0.019614052027463913, + 0.10007402300834656, + 0.13637875020503998, + 0.4241926670074463, + 0.02929677814245224, + 0.039803989231586456, + 0.05657191574573517 + ], + "credit_map": [ + 0.015660937875509262, + 0.014534355141222477, + 0.013354665599763393, + 0.013421066105365753, + 0.017237311229109764, + 0.014895671978592873, + 0.018785610795021057, + 0.01604732684791088, + 0.0051835766062140465, + 0.0026583641301840544, + 0.004111099056899548, + 0.003255577525123954, + 0.006926253437995911, + 0.0047095417976379395, + 0.005062088370323181, + 0.006286165211349726, + 0.004598912317305803, + 0.0035861413925886154, + 0.0027916855178773403, + 0.0024298515636473894, + 0.003143543843179941, + 0.004704831168055534, + 0.004786405712366104, + 0.005896782502532005, + 0.0022282800637185574, + 0.002183453179895878, + 0.001936939894221723, + 0.0013889005640521646, + 0.0037323639262467623, + 0.004296587780117989, + 0.0013411048566922545, + 0.0025064232759177685, + 0.009011522866785526, + 0.011252209544181824, + 0.015421238727867603, + 0.009855255484580994, + 0.019271135330200195, + 0.009781470522284508, + 0.014260382391512394, + 0.011220811866223812, + 0.009947504848241806, + 0.010838359594345093, + 0.010000539012253284, + 0.025636032223701477, + 0.016675978899002075, + 0.02675228752195835, + 0.013457185588777065, + 0.023070862516760826, + 0.10849156230688095, + 0.015222961083054543, + 0.02165095880627632, + 0.04041821137070656, + 0.013604518957436085, + 0.027498336508870125, + 0.029079994186758995, + 0.16822615265846252, + 0.00423736497759819, + 0.0022419635206460953, + 0.0026461046654731035, + 0.0026161163114011288, + 0.0033920903224498034, + 0.0034199184738099575, + 0.00706493342295289, + 0.003678287146613002, + 0.004860337357968092, + 0.002336532110348344, + 0.002679604571312666, + 0.007711690850555897, + 0.0023249881342053413, + 0.004615185782313347, + 0.010530889965593815, + 0.004744761623442173, + 0.016672812402248383, + 0.001979232532903552, + 0.002177588641643524, + 0.003282505553215742, + 0.0020036413334310055, + 0.018804477527737617, + 0.005978865548968315, + 0.005672795698046684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.02637290023267269 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003622637828812003 + ], + "opposite_logprob_tokens": [ + -5.628622531890869 + ], + "logprob_reward": -0.003622637828812003, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003622637828812003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17405347526073456, + 0.03336123377084732, + 0.03464432805776596, + 0.02795937843620777, + 0.045360833406448364, + 0.22927477955818176, + 0.28561240434646606, + 0.04254785552620888, + 0.04036092758178711, + 0.08682478219270706 + ], + "credit_map": [ + 0.017321737483143806, + 0.014247067272663116, + 0.012374921701848507, + 0.013044901192188263, + 0.027019688859581947, + 0.024850554764270782, + 0.0334760956466198, + 0.03171849250793457, + 0.006235246080905199, + 0.0019459602190181613, + 0.0022066638339310884, + 0.0021734980400651693, + 0.009426292963325977, + 0.003459977451711893, + 0.002476692898198962, + 0.0054368977434933186, + 0.0062399134039878845, + 0.002841860055923462, + 0.0025499567855149508, + 0.0020337849855422974, + 0.0033850253093987703, + 0.004534045699983835, + 0.003323443466797471, + 0.009736292995512486, + 0.0034566796384751797, + 0.0036701681092381477, + 0.00302023789845407, + 0.002408097730949521, + 0.003441658802330494, + 0.006192550994455814, + 0.002359127625823021, + 0.003410854609683156, + 0.003967883065342903, + 0.005270231980830431, + 0.006851803511381149, + 0.00522986613214016, + 0.009324988350272179, + 0.004204184748232365, + 0.0048058214597404, + 0.005706049967557192, + 0.016523858532309532, + 0.019599545747041702, + 0.018587591126561165, + 0.04186936840415001, + 0.026588045060634613, + 0.040886230766773224, + 0.02348138391971588, + 0.041738737374544144, + 0.06494756788015366, + 0.013258029706776142, + 0.0118948332965374, + 0.026145799085497856, + 0.009822951629757881, + 0.020055223256349564, + 0.01951317861676216, + 0.11997479945421219, + 0.0059955730102956295, + 0.003462513443082571, + 0.0021643717773258686, + 0.004765413701534271, + 0.0026708347722887993, + 0.003997206222265959, + 0.01363827008754015, + 0.005853667389601469, + 0.008017872460186481, + 0.002263805130496621, + 0.002129943110048771, + 0.003299832111224532, + 0.0031446446664631367, + 0.007723894435912371, + 0.00803255196660757, + 0.00574837950989604, + 0.01934133656322956, + 0.006681170780211687, + 0.005106574855744839, + 0.006584085989743471, + 0.013590452261269093, + 0.019688721746206284, + 0.006251374259591103, + 0.009581055492162704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.003622637828812003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002199494047090411 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.002199494047090411, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002199494047090411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04998655244708061, + 0.08047834783792496, + 0.01914292201399803, + 0.05630224943161011, + 0.055363234132528305, + 0.1393495351076126, + 0.1641237735748291, + 0.03148927539587021, + 0.040172792971134186, + 0.3635912835597992 + ], + "credit_map": [ + 0.005985264666378498, + 0.007491820026189089, + 0.0035653209779411554, + 0.005747729446738958, + 0.00876440666615963, + 0.005222972482442856, + 0.005676321219652891, + 0.007532714866101742, + 0.01637781597673893, + 0.004218636080622673, + 0.004310508258640766, + 0.0033859051764011383, + 0.028352683410048485, + 0.005058093927800655, + 0.004920910578221083, + 0.013853795826435089, + 0.0029264239128679037, + 0.0020628836937248707, + 0.0015354042407125235, + 0.001587062724865973, + 0.0021070935763418674, + 0.0031549851410090923, + 0.002625448862090707, + 0.003143619978800416, + 0.007188248448073864, + 0.008224268443882465, + 0.006924906279891729, + 0.0051438710652291775, + 0.006718593183904886, + 0.01229746825993061, + 0.004199708811938763, + 0.005605184473097324, + 0.007868913002312183, + 0.00812943559139967, + 0.009294010698795319, + 0.0058218324556946754, + 0.005753750912845135, + 0.004960328806191683, + 0.005728662945330143, + 0.007806298788636923, + 0.009902521036565304, + 0.011901774443686008, + 0.011592429131269455, + 0.024020446464419365, + 0.017799893394112587, + 0.025279229506850243, + 0.014721355400979519, + 0.02413187548518181, + 0.0413338765501976, + 0.006283453665673733, + 0.010032640770077705, + 0.015562785789370537, + 0.005626800935715437, + 0.009931241162121296, + 0.011211737059056759, + 0.06414124369621277, + 0.004598366562277079, + 0.0020149117335677147, + 0.0035793946590274572, + 0.0028940720949321985, + 0.0032847432885318995, + 0.0034235890489071608, + 0.005049646832048893, + 0.006644552107900381, + 0.0069217015989124775, + 0.0028949896804988384, + 0.003771393559873104, + 0.00315744592808187, + 0.0024221083149313927, + 0.004464288242161274, + 0.007469210773706436, + 0.009071655571460724, + 0.063575878739357, + 0.02957652322947979, + 0.026266921311616898, + 0.020789407193660736, + 0.06473924964666367, + 0.07885532081127167, + 0.02838086150586605, + 0.051407117396593094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.002199494047090411 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028208012226969004 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028208012226969004, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028208012226969004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07092022895812988, + 0.053399618715047836, + 0.03708226978778839, + 0.03655735030770302, + 0.03259066119790077, + 0.17610841989517212, + 0.3918340504169464, + 0.03467497602105141, + 0.023462502285838127, + 0.1433698683977127 + ], + "credit_map": [ + 0.008439439348876476, + 0.011435958556830883, + 0.005934328306466341, + 0.006608930882066488, + 0.010666927322745323, + 0.007145915646106005, + 0.009106609039008617, + 0.011582120321691036, + 0.012283549644052982, + 0.0030752476304769516, + 0.002881309948861599, + 0.002529048128053546, + 0.017524899914860725, + 0.004801299888640642, + 0.004025876522064209, + 0.0062783886678516865, + 0.006599909625947475, + 0.002961215563118458, + 0.0026369851548224688, + 0.0034080022014677525, + 0.003431540448218584, + 0.0038857527542859316, + 0.003619666211307049, + 0.010539194568991661, + 0.0062175896018743515, + 0.006463206373155117, + 0.00350241013802588, + 0.003688662778586149, + 0.004337116610258818, + 0.005851288326084614, + 0.002815286861732602, + 0.0036817891523241997, + 0.002522381255403161, + 0.003969548735767603, + 0.0047062598168849945, + 0.003574890550225973, + 0.005196967627853155, + 0.0035789478570222855, + 0.004477668087929487, + 0.004563997499644756, + 0.012507280334830284, + 0.015736503526568413, + 0.014686085283756256, + 0.030850157141685486, + 0.021638276055455208, + 0.0297007467597723, + 0.019904281944036484, + 0.03108508139848709, + 0.0993894636631012, + 0.014210367575287819, + 0.018476679921150208, + 0.037397611886262894, + 0.012134335935115814, + 0.025625528767704964, + 0.02648780308663845, + 0.15811225771903992, + 0.005081736482679844, + 0.0024528412614017725, + 0.0038485443219542503, + 0.0029579787515103817, + 0.0040137567557394505, + 0.0034735172521322966, + 0.007198130711913109, + 0.005648469086736441, + 0.004929502960294485, + 0.001400061883032322, + 0.002232614438980818, + 0.0015633784933015704, + 0.0013926672982051969, + 0.0044370535761117935, + 0.002596144797280431, + 0.0049110776744782925, + 0.038158148527145386, + 0.007786115165799856, + 0.007008977234363556, + 0.004590488970279694, + 0.015123189426958561, + 0.04562298581004143, + 0.006760928314179182, + 0.018319021910429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.0028208012226969004 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, the camera slowly tilts upward, following the subject's movements. On a damp, slightly pebbled ground bathed in cool moonlight, an elegant black cat crouches compactly. Its glossy black fur, like smooth silk, reflects a faint silver sheen in the dim light, outlining the muscular lines beneath its back and shoulder blades. Initially, its head is slightly lowered, then, with a steady and deliberate rhythm, it slowly raises its head, finally fixing its gaze on the night sky. In this movement, its jewel-like emerald green eyes are fully revealed, their clear pupils clearly reflecting the bright full moon, as if absorbing the entire night sky's brilliance. Its gaze is focused and serene. The entire scene, a stark contrast between the deep black and the brilliant silver and emerald green, creates an atmosphere of mystery, tranquility, and spirituality.", + "seed": 29585, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_0/index_9434_seed_29585_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01965942676179111, + "vlm_logprob": -0.01965942676179111, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.01965942676179111, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056009650230407715 + ], + "opposite_logprob_tokens": [ + -3.5560097694396973 + ], + "logprob_reward": -0.056009650230407715, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.056009650230407715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2468191236257553, + 0.09720884263515472, + 0.11201281100511551, + 0.09465120732784271, + 0.13972441852092743, + 0.07210036367177963, + 0.05432392656803131, + 0.049816396087408066, + 0.07576794177293777, + 0.057575006037950516 + ], + "credit_map": [ + 0.014548267237842083, + 0.0326240174472332, + 0.03439135104417801, + 0.026616426184773445, + 0.03869527950882912, + 0.024321921169757843, + 0.054578863084316254, + 0.02104296162724495, + 0.010709178633987904, + 0.01626604236662388, + 0.011192378588020802, + 0.009793957695364952, + 0.018369875848293304, + 0.009825182147324085, + 0.01214582659304142, + 0.008906390517950058, + 0.00765214953571558, + 0.017120711505413055, + 0.011848418042063713, + 0.013828694820404053, + 0.01914081908762455, + 0.01342547032982111, + 0.013905515894293785, + 0.015091028064489365, + 0.010515422560274601, + 0.013623797334730625, + 0.007804366294294596, + 0.011852744966745377, + 0.011948281899094582, + 0.008262557908892632, + 0.01576473005115986, + 0.014879296533763409, + 0.005841491278260946, + 0.047648195177316666, + 0.004365609958767891, + 0.02280540205538273, + 0.030744697898626328, + 0.0055311620235443115, + 0.014353357255458832, + 0.008434486575424671, + 0.004354436416178942, + 0.007499769330024719, + 0.0038338915910571814, + 0.021070869639515877, + 0.01066303625702858, + 0.007046801503747702, + 0.013012748211622238, + 0.004618805833160877, + 0.0019243969582021236, + 0.006925595458596945, + 0.0029156070668250322, + 0.009286158718168736, + 0.010350249707698822, + 0.009098248556256294, + 0.007390805520117283, + 0.006432858761399984, + 0.0029006123077124357, + 0.005192796234041452, + 0.004963011480867863, + 0.006269033532589674, + 0.009042590856552124, + 0.008618079125881195, + 0.007511468138545752, + 0.005318799521774054, + 0.0033024309668689966, + 0.006831214297562838, + 0.0051550958305597305, + 0.011526267044246197, + 0.015071186237037182, + 0.009051944129168987, + 0.014204767532646656, + 0.01062503270804882, + 0.005244254134595394, + 0.006336349993944168, + 0.006524394266307354, + 0.008317343890666962, + 0.006837361957877874, + 0.005798042751848698, + 0.009095659479498863, + 0.009421592578291893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.056009650230407715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026382071897387505 + ], + "opposite_logprob_tokens": [ + -3.6513819694519043 + ], + "logprob_reward": -0.026382071897387505, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026382071897387505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23767642676830292, + 0.1027875766158104, + 0.225160151720047, + 0.09414997696876526, + 0.06155804172158241, + 0.03301399201154709, + 0.037637967616319656, + 0.04410654678940773, + 0.1232842430472374, + 0.04062514379620552 + ], + "credit_map": [ + 0.017680678516626358, + 0.0282925795763731, + 0.04287256672978401, + 0.023559166118502617, + 0.03321078047156334, + 0.027620593085885048, + 0.04409077391028404, + 0.020349280908703804, + 0.010158533230423927, + 0.012353072874248028, + 0.013904998078942299, + 0.01021141279488802, + 0.017584271728992462, + 0.014979381114244461, + 0.011492353864014149, + 0.01210353896021843, + 0.0155504010617733, + 0.03507554531097412, + 0.022130796685814857, + 0.027160467579960823, + 0.03792387247085571, + 0.026113402098417282, + 0.0295375045388937, + 0.03166816383600235, + 0.005464385729283094, + 0.01968754455447197, + 0.005771559663116932, + 0.00897939596325159, + 0.017276162281632423, + 0.007299124263226986, + 0.018856491893529892, + 0.010815305635333061, + 0.0054716141894459724, + 0.015335108153522015, + 0.0044061122462153435, + 0.009433656930923462, + 0.011228833347558975, + 0.003462250577285886, + 0.007253176998347044, + 0.004967283457517624, + 0.0022522001527249813, + 0.0036223260685801506, + 0.0017799997003749013, + 0.008611681871116161, + 0.005270382389426231, + 0.0020489529706537724, + 0.007396077737212181, + 0.0020323668140918016, + 0.0018301079981029034, + 0.0032188864424824715, + 0.002077233511954546, + 0.010335131548345089, + 0.005981775466352701, + 0.0028227732982486486, + 0.008532767184078693, + 0.002839290304109454, + 0.0023139063268899918, + 0.004950891248881817, + 0.0018967832438647747, + 0.007461437955498695, + 0.008654254488646984, + 0.008147279731929302, + 0.007032496854662895, + 0.003649495542049408, + 0.007704656105488539, + 0.01224503293633461, + 0.008329395204782486, + 0.01814272627234459, + 0.029624411836266518, + 0.01342235878109932, + 0.02261730283498764, + 0.011198352091014385, + 0.005172755103558302, + 0.0034635348711162806, + 0.005000704433768988, + 0.0059854742139577866, + 0.0043155234307050705, + 0.003790759015828371, + 0.007896777242422104, + 0.004999611992388964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.026382071897387505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794275276362896 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.009794275276362896, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794275276362896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0934690460562706, + 0.0771474689245224, + 0.09036257117986679, + 0.09673294425010681, + 0.13119177520275116, + 0.2870844006538391, + 0.05815732851624489, + 0.046109624207019806, + 0.07710018008947372, + 0.04264473170042038 + ], + "credit_map": [ + 0.006606847047805786, + 0.011464673094451427, + 0.016794828698039055, + 0.011741241440176964, + 0.012233102694153786, + 0.010547198355197906, + 0.01572479121387005, + 0.00835635419934988, + 0.007943838834762573, + 0.010169456712901592, + 0.00994484219700098, + 0.009467050433158875, + 0.010965308174490929, + 0.010233093053102493, + 0.009927748702466488, + 0.008496127091348171, + 0.005790407303720713, + 0.02182106301188469, + 0.0060544246807694435, + 0.007695202250033617, + 0.019339608028531075, + 0.008540344424545765, + 0.009768224321305752, + 0.011353292502462864, + 0.010709747672080994, + 0.021520694717764854, + 0.006078728009015322, + 0.00939988438040018, + 0.0184769406914711, + 0.007447625510394573, + 0.012941456399857998, + 0.010157857090234756, + 0.00947600044310093, + 0.04235365614295006, + 0.005581096280366182, + 0.019071713089942932, + 0.02632473222911358, + 0.008632875047624111, + 0.011169839650392532, + 0.008581847883760929, + 0.017231067642569542, + 0.035787977278232574, + 0.028994616121053696, + 0.0695103332400322, + 0.03983825445175171, + 0.02821248769760132, + 0.045794527977705, + 0.021715126931667328, + 0.0035820899065583944, + 0.006865488365292549, + 0.0041837748140096664, + 0.013998163864016533, + 0.008164780214428902, + 0.005068507976830006, + 0.011951153166592121, + 0.0043433657847344875, + 0.004444443620741367, + 0.0052414545789361, + 0.0032089385204017162, + 0.005780417937785387, + 0.008916286751627922, + 0.007876412943005562, + 0.007018324453383684, + 0.003623344236984849, + 0.004078395199030638, + 0.01202621590346098, + 0.0050721969455480576, + 0.010661708191037178, + 0.011353151872754097, + 0.011460402980446815, + 0.015193790197372437, + 0.007254305761307478, + 0.006190349347889423, + 0.0041199070401489735, + 0.0064743682742118835, + 0.006233744323253632, + 0.0037651469465345144, + 0.004264220129698515, + 0.006686423905193806, + 0.004910570569336414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.009794275276362896 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619193332269788 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.003619193332269788, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619193332269788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0901503935456276, + 0.059702806174755096, + 0.13607069849967957, + 0.06615309417247772, + 0.1519511491060257, + 0.06478007137775421, + 0.05853680148720741, + 0.1666441410779953, + 0.08207341283559799, + 0.1239374503493309 + ], + "credit_map": [ + 0.00999962817877531, + 0.009373173117637634, + 0.01244524959474802, + 0.0070608980022370815, + 0.013038051314651966, + 0.01181124523282051, + 0.014890823513269424, + 0.011531315743923187, + 0.003961669746786356, + 0.01023099571466446, + 0.006662913132458925, + 0.005052767228335142, + 0.013444339856505394, + 0.008172141388058662, + 0.006947589572519064, + 0.0052303834818303585, + 0.008980664424598217, + 0.029862171038985252, + 0.010509098879992962, + 0.01427219808101654, + 0.026968887075781822, + 0.011778193525969982, + 0.016465874388813972, + 0.0172335933893919, + 0.006590003613382578, + 0.012308604083955288, + 0.0046578156761825085, + 0.007763539906591177, + 0.007656649686396122, + 0.005919821560382843, + 0.009347858838737011, + 0.011908800341188908, + 0.012252790853381157, + 0.03976582735776901, + 0.010764272883534431, + 0.026376044377684593, + 0.028310153633356094, + 0.008192222565412521, + 0.016958480700850487, + 0.009331348352134228, + 0.004307959228754044, + 0.007198697421699762, + 0.004501401912420988, + 0.014920382760465145, + 0.010857070796191692, + 0.0070997243747115135, + 0.010765613056719303, + 0.005129221361130476, + 0.004904524888843298, + 0.007771489210426807, + 0.006651374977082014, + 0.007897845469415188, + 0.011248859576880932, + 0.007485206238925457, + 0.007558897137641907, + 0.005018598400056362, + 0.012126389890909195, + 0.018886176869273186, + 0.010911760851740837, + 0.028229817748069763, + 0.029223168268799782, + 0.019540879875421524, + 0.02549281343817711, + 0.02223312109708786, + 0.003688151715323329, + 0.008599979802966118, + 0.00717137660831213, + 0.011385639198124409, + 0.011380678042769432, + 0.013035724870860577, + 0.01697477698326111, + 0.009837072342634201, + 0.020862018689513206, + 0.012803487479686737, + 0.014623786322772503, + 0.012759611941874027, + 0.014423776417970657, + 0.012675370089709759, + 0.01906733028590679, + 0.016722064465284348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.003619193332269788 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002491943072527647 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.002491943072527647, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002491943072527647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15153726935386658, + 0.09146237373352051, + 0.07972495257854462, + 0.0826263353228569, + 0.12271296977996826, + 0.16044963896274567, + 0.07758311927318573, + 0.08057697862386703, + 0.07309233397245407, + 0.08023407310247421 + ], + "credit_map": [ + 0.01765240915119648, + 0.017286809161305428, + 0.027660798281431198, + 0.017884718254208565, + 0.018829647451639175, + 0.018217453733086586, + 0.02079135924577713, + 0.013214058242738247, + 0.005224127322435379, + 0.014133807271718979, + 0.007398630492389202, + 0.011159650981426239, + 0.01825817860662937, + 0.01478575449436903, + 0.010966690257191658, + 0.009535539895296097, + 0.006269446574151516, + 0.013227864168584347, + 0.0081788320094347, + 0.008991983719170094, + 0.014055371284484863, + 0.00940665602684021, + 0.009599406272172928, + 0.009995394386351109, + 0.010345117188990116, + 0.014076187275350094, + 0.00611380347982049, + 0.008357658982276917, + 0.011683270335197449, + 0.00775808235630393, + 0.010991329327225685, + 0.013300878927111626, + 0.007629982195794582, + 0.025411969050765038, + 0.007602063473314047, + 0.020924678072333336, + 0.01993751712143421, + 0.016934216022491455, + 0.013752033933997154, + 0.010520512238144875, + 0.00813101977109909, + 0.015581715852022171, + 0.00894960481673479, + 0.04560449346899986, + 0.02883991412818432, + 0.017362236976623535, + 0.025000987574458122, + 0.010979671031236649, + 0.004136761650443077, + 0.010819909162819386, + 0.005098662804812193, + 0.016003558412194252, + 0.014421404339373112, + 0.007257778663188219, + 0.01414258498698473, + 0.005702461581677198, + 0.0058061606250703335, + 0.009453162550926208, + 0.00597025454044342, + 0.010755130089819431, + 0.014408262446522713, + 0.012741493992507458, + 0.011873948387801647, + 0.009568563662469387, + 0.004978583659976721, + 0.007356996648013592, + 0.0073492759838700294, + 0.010844818316400051, + 0.013442116789519787, + 0.008147278800606728, + 0.01347285881638527, + 0.007500401232391596, + 0.01138798426836729, + 0.007661353796720505, + 0.00944104976952076, + 0.011978711001574993, + 0.007796754594892263, + 0.006913376040756702, + 0.0120529904961586, + 0.013001851737499237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.002491943072527647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, the camera slowly tilts upward, following the subject's movements. On a damp, slightly pebbled ground bathed in cool moonlight, an elegant black cat crouches compactly. Its glossy black fur, like smooth silk, reflects a faint silver sheen in the dim light, outlining the muscular lines beneath its back and shoulder blades. Initially, its head is slightly lowered, then, with a steady and deliberate rhythm, it slowly raises its head, finally fixing its gaze on the night sky. In this movement, its jewel-like emerald green eyes are fully revealed, their clear pupils clearly reflecting the bright full moon, as if absorbing the entire night sky's brilliance. Its gaze is focused and serene. The entire scene, a stark contrast between the deep black and the brilliant silver and emerald green, creates an atmosphere of mystery, tranquility, and spirituality.", + "seed": 29586, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_0/index_9434_seed_29586_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03619470959529281, + "vlm_logprob": -0.03619470959529281, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.03619470959529281, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06109674647450447 + ], + "opposite_logprob_tokens": [ + -3.186096668243408 + ], + "logprob_reward": -0.06109674647450447, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06109674647450447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06078660860657692, + 0.5917262434959412, + 0.03531354293227196, + 0.030951127409934998, + 0.05047453194856644, + 0.09397359937429428, + 0.07633716613054276, + 0.029578933492302895, + 0.01673741824924946, + 0.014120753854513168 + ], + "credit_map": [ + 0.008801509626209736, + 0.007634441368281841, + 0.004733366891741753, + 0.004856166895478964, + 0.008206972852349281, + 0.008715411648154259, + 0.012730604037642479, + 0.005108144134283066, + 0.02354811131954193, + 0.24152718484401703, + 0.01433870941400528, + 0.016027189791202545, + 0.03685736283659935, + 0.012875179760158062, + 0.2320771962404251, + 0.014475364238023758, + 0.004373941570520401, + 0.00408156868070364, + 0.007087918929755688, + 0.003658932400867343, + 0.005355258472263813, + 0.003038175869733095, + 0.004030618350952864, + 0.0036871335469186306, + 0.0032841279171407223, + 0.005494690500199795, + 0.004186606965959072, + 0.002915730467066169, + 0.004518586210906506, + 0.003006373066455126, + 0.004831691272556782, + 0.002713326131924987, + 0.0035108027514070272, + 0.007374900393188, + 0.005650647915899754, + 0.005349087528884411, + 0.009047216735780239, + 0.007142155896872282, + 0.007270517759025097, + 0.005129208788275719, + 0.009683484211564064, + 0.01602952554821968, + 0.014601373113691807, + 0.008975721895694733, + 0.010007311590015888, + 0.009631636552512646, + 0.01143322791904211, + 0.013611328788101673, + 0.011268086731433868, + 0.018214210867881775, + 0.006062373053282499, + 0.00998402014374733, + 0.0072681182064116, + 0.005117036867886782, + 0.010304522700607777, + 0.008118807338178158, + 0.002352821873500943, + 0.004835368134081364, + 0.0031575518660247326, + 0.0032598513644188643, + 0.0017968551255762577, + 0.00350520689971745, + 0.006608078256249428, + 0.00406320346519351, + 0.0016247185412794352, + 0.0020835869945585728, + 0.0019411054672673345, + 0.002554995007812977, + 0.0025272490456700325, + 0.0017798648914322257, + 0.0020759934559464455, + 0.0021499055437743664, + 0.0011528024915605783, + 0.001454485347494483, + 0.001469134702347219, + 0.0027698904741555452, + 0.0018397121457383037, + 0.0021714670583605766, + 0.0017943276325240731, + 0.0014689357485622168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.06109674647450447 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003585234284401 + ], + "opposite_logprob_tokens": [ + -2.350358486175537 + ], + "logprob_reward": -0.1003585234284401, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003585234284401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06578067690134048, + 0.34277939796447754, + 0.04273826628923416, + 0.04012719541788101, + 0.08916736394166946, + 0.08883678168058395, + 0.0849454328417778, + 0.07090996950864792, + 0.09010647237300873, + 0.08460843563079834 + ], + "credit_map": [ + 0.01016585435718298, + 0.008452548645436764, + 0.004922524560242891, + 0.004775420296937227, + 0.012472721748054028, + 0.00863339751958847, + 0.011914072558283806, + 0.0044441381469368935, + 0.014926583506166935, + 0.13629506528377533, + 0.010060473345220089, + 0.009876511991024017, + 0.02124077081680298, + 0.008353294804692268, + 0.13301877677440643, + 0.009007919579744339, + 0.006924774497747421, + 0.00562801631167531, + 0.006406821776181459, + 0.0043307701125741005, + 0.006426795851439238, + 0.003938564099371433, + 0.00528474897146225, + 0.003797771641984582, + 0.0035874787718057632, + 0.007062193006277084, + 0.00550002371892333, + 0.004644184373319149, + 0.005376284476369619, + 0.0036450026091188192, + 0.006329112686216831, + 0.0039829169400036335, + 0.005923937540501356, + 0.012845098972320557, + 0.010671635158360004, + 0.009114849381148815, + 0.01587710902094841, + 0.012918593361973763, + 0.012502250261604786, + 0.009313886985182762, + 0.007591992150992155, + 0.0164407417178154, + 0.015406719408929348, + 0.009367226622998714, + 0.008192556910216808, + 0.007467464078217745, + 0.012666299007833004, + 0.01170377992093563, + 0.010560151189565659, + 0.01945238932967186, + 0.007363962475210428, + 0.008348105475306511, + 0.007541092578321695, + 0.006205435376614332, + 0.012900506146252155, + 0.012573791667819023, + 0.005581414792686701, + 0.014633717015385628, + 0.005689352750778198, + 0.009989527985453606, + 0.009338797070086002, + 0.005077797453850508, + 0.010611922480165958, + 0.009987437166273594, + 0.008866635151207447, + 0.011862458661198616, + 0.008650390431284904, + 0.018249141052365303, + 0.013520900160074234, + 0.008757955394685268, + 0.01082819513976574, + 0.009370791725814342, + 0.005436155013740063, + 0.00893027987331152, + 0.006535632070153952, + 0.014087453484535217, + 0.02064080536365509, + 0.009440016932785511, + 0.011216728016734123, + 0.008321364410221577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.1003585234284401 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204702340066433 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204702340066433, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204702340066433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11198703944683075, + 0.3718720078468323, + 0.04683289676904678, + 0.03834031522274017, + 0.08846357464790344, + 0.09634406864643097, + 0.09115058928728104, + 0.03636879101395607, + 0.09935405850410461, + 0.019286690279841423 + ], + "credit_map": [ + 0.017471082508563995, + 0.01627897284924984, + 0.009910056367516518, + 0.010983286425471306, + 0.016982823610305786, + 0.012488653883337975, + 0.015115335583686829, + 0.012756824493408203, + 0.017439192160964012, + 0.14113271236419678, + 0.015495400875806808, + 0.01479765772819519, + 0.024859977886080742, + 0.012786583974957466, + 0.13271653652191162, + 0.012643941678106785, + 0.004501217044889927, + 0.005504891742020845, + 0.008809177204966545, + 0.004489492624998093, + 0.008926015347242355, + 0.004216286353766918, + 0.005879241973161697, + 0.0045065730810165405, + 0.004492779728025198, + 0.008619888685643673, + 0.004431175533682108, + 0.005182235036045313, + 0.004566347226500511, + 0.0027411195915192366, + 0.0044917212799191475, + 0.0038150486070662737, + 0.0071447161026299, + 0.013672544620931149, + 0.009830423630774021, + 0.008807490579783916, + 0.015210265293717384, + 0.012979048304259777, + 0.012136027216911316, + 0.008683057501912117, + 0.010390639305114746, + 0.015189750120043755, + 0.015396618284285069, + 0.009141601622104645, + 0.011484142392873764, + 0.009205558337271214, + 0.011627371422946453, + 0.013908385299146175, + 0.010380898602306843, + 0.025526123121380806, + 0.006674893666058779, + 0.011527720838785172, + 0.00978176947683096, + 0.005679943598806858, + 0.011699032038450241, + 0.009880210272967815, + 0.0026450601872056723, + 0.008129910565912724, + 0.0028530587442219257, + 0.005383871030062437, + 0.00423415144905448, + 0.0030959544237703085, + 0.0048644463531672955, + 0.005162336397916079, + 0.0075603267177939415, + 0.01461109146475792, + 0.00600979570299387, + 0.028866546228528023, + 0.012080653570592403, + 0.007803960237652063, + 0.014112631790339947, + 0.008309056982398033, + 0.0025819456204771996, + 0.0019697104580700397, + 0.0024256655015051365, + 0.00336451712064445, + 0.0019471151754260063, + 0.0035269190557301044, + 0.0021290883887559175, + 0.0013417290756478906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.014204702340066433 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028224652633070946 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.0028224652633070946, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028224652633070946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05052310600876808, + 0.2523993253707886, + 0.021044138818979263, + 0.018535064533352852, + 0.04302327334880829, + 0.07158610224723816, + 0.11372145265340805, + 0.03161104395985603, + 0.35499659180641174, + 0.042559895664453506 + ], + "credit_map": [ + 0.00727443303912878, + 0.0061041973531246185, + 0.005331488326191902, + 0.005116303917020559, + 0.007655584719032049, + 0.005976053886115551, + 0.008445077575743198, + 0.004619967192411423, + 0.010266410186886787, + 0.09767287224531174, + 0.009899219498038292, + 0.007684001699090004, + 0.016618337482213974, + 0.006854868493974209, + 0.09574779868125916, + 0.007655823137611151, + 0.0023551913909614086, + 0.002704533515498042, + 0.00250779720954597, + 0.002238684566691518, + 0.002907753689214587, + 0.002618941944092512, + 0.002930559916421771, + 0.0027806758880615234, + 0.0019625320564955473, + 0.0024483276065438986, + 0.003456246107816696, + 0.002751939930021763, + 0.002153231529518962, + 0.0016520376084372401, + 0.0022071024868637323, + 0.0019036465091630816, + 0.0027675298042595387, + 0.005645175464451313, + 0.004394365940243006, + 0.004063768312335014, + 0.008498650044202805, + 0.006575137842446566, + 0.0065153841860592365, + 0.004563263151794672, + 0.006534365471452475, + 0.012207860127091408, + 0.011101286858320236, + 0.0069394176825881, + 0.009636002592742443, + 0.007000844459980726, + 0.008248327299952507, + 0.009917995892465115, + 0.020374072715640068, + 0.027021553367376328, + 0.00919738132506609, + 0.010067658498883247, + 0.010042690671980381, + 0.006657909601926804, + 0.016769783571362495, + 0.013590396381914616, + 0.0024393920321017504, + 0.006164431571960449, + 0.00231278408318758, + 0.004246237222105265, + 0.00250800303183496, + 0.0032190328929573298, + 0.006051781587302685, + 0.0046693808399140835, + 0.038191478699445724, + 0.04958334565162659, + 0.02462293952703476, + 0.10178025811910629, + 0.03442608565092087, + 0.03313601762056351, + 0.047200679779052734, + 0.02605578862130642, + 0.005418411456048489, + 0.005140152294188738, + 0.0035395375452935696, + 0.008317669853568077, + 0.006159722339361906, + 0.004809695761650801, + 0.0056390706449747086, + 0.0035356362350285053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.0028224652633070946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024911104701459408 + ], + "opposite_logprob_tokens": [ + -6.002490997314453 + ], + "logprob_reward": -0.0024911104701459408, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024911104701459408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07949738204479218, + 0.24927233159542084, + 0.04833746701478958, + 0.01497361809015274, + 0.04797882214188576, + 0.08295056223869324, + 0.09363758563995361, + 0.07106456905603409, + 0.20762641727924347, + 0.1046612560749054 + ], + "credit_map": [ + 0.011464294977486134, + 0.008099358528852463, + 0.008219190873205662, + 0.00880358461290598, + 0.009664655663073063, + 0.010989625006914139, + 0.01284794695675373, + 0.009408727288246155, + 0.01024894043803215, + 0.09713948518037796, + 0.00841650553047657, + 0.007176386192440987, + 0.016251908615231514, + 0.005905495025217533, + 0.09534095972776413, + 0.008792652748525143, + 0.0072347912937402725, + 0.005988415330648422, + 0.004577904473990202, + 0.005535770673304796, + 0.007935620844364166, + 0.005708916112780571, + 0.005714496597647667, + 0.005641550291329622, + 0.0022562795784324408, + 0.0022407423239201307, + 0.0026133714709430933, + 0.0017939548706635833, + 0.0017042160034179688, + 0.0013167794095352292, + 0.001713592209853232, + 0.001334681990556419, + 0.004381482023745775, + 0.007196494378149509, + 0.005875066854059696, + 0.0033330463338643312, + 0.007408750243484974, + 0.007345781195908785, + 0.006777907256036997, + 0.00566028943285346, + 0.011168320663273335, + 0.01385016180574894, + 0.011976053938269615, + 0.007616384886205196, + 0.008832933381199837, + 0.0068654706701636314, + 0.010476373136043549, + 0.012164857238531113, + 0.015099061653017998, + 0.022952962666749954, + 0.007579150144010782, + 0.008845174685120583, + 0.008817553520202637, + 0.004897301085293293, + 0.013885539025068283, + 0.011560851708054543, + 0.005849762819707394, + 0.01514084730297327, + 0.005406377371400595, + 0.009379460476338863, + 0.007463811431080103, + 0.0045154220424592495, + 0.013791147619485855, + 0.009517744183540344, + 0.023033972829580307, + 0.02797367051243782, + 0.013794121332466602, + 0.059226009994745255, + 0.019446197897195816, + 0.018194640055298805, + 0.028678663074970245, + 0.017279138788580894, + 0.006829578895121813, + 0.012089135125279427, + 0.008007999509572983, + 0.01643953286111355, + 0.029450885951519012, + 0.010643992573022842, + 0.013650394976139069, + 0.007549743168056011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.0024911104701459408 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from an extremely low angle, the camera slowly tilts upward, following the subject's movements. On a damp, slightly pebbled ground bathed in cool moonlight, an elegant black cat crouches compactly. Its glossy black fur, like smooth silk, reflects a faint silver sheen in the dim light, outlining the muscular lines beneath its back and shoulder blades. Initially, its head is slightly lowered, then, with a steady and deliberate rhythm, it slowly raises its head, finally fixing its gaze on the night sky. In this movement, its jewel-like emerald green eyes are fully revealed, their clear pupils clearly reflecting the bright full moon, as if absorbing the entire night sky's brilliance. Its gaze is focused and serene. The entire scene, a stark contrast between the deep black and the brilliant silver and emerald green, creates an atmosphere of mystery, tranquility, and spirituality.", + "seed": 29587, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_0/index_9434_seed_29587_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01664197649806738, + "vlm_logprob": -0.01664197649806738, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.01664197649806738, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03933002054691315 + ], + "opposite_logprob_tokens": [ + -3.789330005645752 + ], + "logprob_reward": -0.03933002054691315, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03933002054691315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19846679270267487, + 0.06511261314153671, + 0.1399204283952713, + 0.11454084515571594, + 0.11224161088466644, + 0.07736214995384216, + 0.08732905983924866, + 0.10946083813905716, + 0.08251351118087769, + 0.01305212453007698 + ], + "credit_map": [ + 0.021357547491788864, + 0.04868170619010925, + 0.024888332933187485, + 0.014763524755835533, + 0.04177651181817055, + 0.012629976496100426, + 0.014738871715962887, + 0.019630316644906998, + 0.006792630068957806, + 0.010547786019742489, + 0.00898320134729147, + 0.012001165188848972, + 0.009679657407104969, + 0.004233946092426777, + 0.00770910969004035, + 0.005165119655430317, + 0.009308261796832085, + 0.03557581454515457, + 0.018060928210616112, + 0.0115541722625494, + 0.01731349527835846, + 0.011180459521710873, + 0.02513611689209938, + 0.011791180819272995, + 0.009662657976150513, + 0.02171202190220356, + 0.01611001044511795, + 0.016258882358670235, + 0.0066856262274086475, + 0.009916246868669987, + 0.02593352273106575, + 0.008261867798864841, + 0.007748132571578026, + 0.026337601244449615, + 0.006588684860616922, + 0.018787236884236336, + 0.005600558128207922, + 0.014681929722428322, + 0.01024096179753542, + 0.022256510332226753, + 0.012688479386270046, + 0.008559256792068481, + 0.010096929036080837, + 0.010779470205307007, + 0.007484497502446175, + 0.01557901967316866, + 0.00449718302115798, + 0.007677313406020403, + 0.009076926857233047, + 0.0215835589915514, + 0.00892205722630024, + 0.012462795712053776, + 0.008330967277288437, + 0.007771726232022047, + 0.011123632080852985, + 0.008057404309511185, + 0.015274625271558762, + 0.028264474123716354, + 0.017351891845464706, + 0.012896275147795677, + 0.013902985490858555, + 0.0071201627142727375, + 0.007333815097808838, + 0.007316608913242817, + 0.008002219721674919, + 0.027639375999569893, + 0.014012979343533516, + 0.009188474155962467, + 0.011134748347103596, + 0.0038465645629912615, + 0.004478737711906433, + 0.004210412502288818, + 0.0018533837283030152, + 0.0020816910546272993, + 0.0022136650513857603, + 0.0017178537091240287, + 0.0016553503228351474, + 0.0016924859955906868, + 0.0008710387628525496, + 0.0009666560217738152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.03933002054691315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373829692602158 + ], + "opposite_logprob_tokens": [ + -3.651373863220215 + ], + "logprob_reward": -0.026373829692602158, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373829692602158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13352368772029877, + 0.08734700828790665, + 0.13927854597568512, + 0.14594897627830505, + 0.0683123916387558, + 0.0870259702205658, + 0.1268433928489685, + 0.12838274240493774, + 0.05524475872516632, + 0.028092585504055023 + ], + "credit_map": [ + 0.015958435833454132, + 0.02034064754843712, + 0.019209904596209526, + 0.01690746285021305, + 0.031890932470560074, + 0.00882756244391203, + 0.010870445519685745, + 0.00951829832047224, + 0.0068442681804299355, + 0.013548365794122219, + 0.011409351602196693, + 0.014681474305689335, + 0.015484820120036602, + 0.00746281398460269, + 0.010517020709812641, + 0.0073988912627100945, + 0.010267829522490501, + 0.04032392427325249, + 0.01723015122115612, + 0.01603231579065323, + 0.014257620088756084, + 0.005866216961294413, + 0.02664356492459774, + 0.008656935766339302, + 0.00852636806666851, + 0.04083879664540291, + 0.009934330359101295, + 0.052179720252752304, + 0.007070515770465136, + 0.006034549325704575, + 0.01403701864182949, + 0.007327681872993708, + 0.005157999228686094, + 0.010916173458099365, + 0.005925052333623171, + 0.011179011315107346, + 0.004549643490463495, + 0.006214491091668606, + 0.007526840548962355, + 0.01684318482875824, + 0.010506747290492058, + 0.009858090430498123, + 0.011068112216889858, + 0.011390099301934242, + 0.006609148345887661, + 0.02167433127760887, + 0.0044799260795116425, + 0.011439514346420765, + 0.015424070879817009, + 0.03454294055700302, + 0.011652571149170399, + 0.018114367499947548, + 0.012487219646573067, + 0.010493413545191288, + 0.013471047393977642, + 0.01065775379538536, + 0.01717829331755638, + 0.03813342750072479, + 0.019771838560700417, + 0.014967442490160465, + 0.013473782688379288, + 0.010906309820711613, + 0.006661836989223957, + 0.007289806380867958, + 0.00452717998996377, + 0.016336537897586823, + 0.009097717702388763, + 0.006191367283463478, + 0.009989391081035137, + 0.0029985224828124046, + 0.003142641158774495, + 0.0029613994993269444, + 0.0032042101956903934, + 0.006251673214137554, + 0.004930596332997084, + 0.002520959824323654, + 0.0036547428462654352, + 0.0031120844651013613, + 0.0019556598272174597, + 0.0024626569356769323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.026373829692602158 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548773549497128 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012548773549497128, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548773549497128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4294668138027191, + 0.03761995956301689, + 0.10665903240442276, + 0.11307647824287415, + 0.05657902732491493, + 0.05412030220031738, + 0.07179973274469376, + 0.06987424194812775, + 0.04059305042028427, + 0.020211346447467804 + ], + "credit_map": [ + 0.05234452709555626, + 0.07378051429986954, + 0.07907986640930176, + 0.039047956466674805, + 0.06160649284720421, + 0.029697690159082413, + 0.03627794235944748, + 0.057631831616163254, + 0.003321901662275195, + 0.006310718134045601, + 0.005308186169713736, + 0.004625923465937376, + 0.007035269867628813, + 0.003349462291225791, + 0.004301838111132383, + 0.0033666626550257206, + 0.010623413138091564, + 0.029108304530382156, + 0.015471560880541801, + 0.011976906098425388, + 0.008723522536456585, + 0.004518710542470217, + 0.019003331661224365, + 0.007233279291540384, + 0.011024841107428074, + 0.021623097360134125, + 0.014638395048677921, + 0.022789379581809044, + 0.005324496887624264, + 0.00995214469730854, + 0.021106228232383728, + 0.006617904175072908, + 0.004648329224437475, + 0.017606548964977264, + 0.00280336057767272, + 0.010682225227355957, + 0.003268943168222904, + 0.00861893966794014, + 0.005419068969786167, + 0.003531609196215868, + 0.006496678106486797, + 0.010277732275426388, + 0.005557576194405556, + 0.013663559220731258, + 0.0033128883223980665, + 0.008006357587873936, + 0.0024381172843277454, + 0.004367392044514418, + 0.007103439886122942, + 0.01580231450498104, + 0.0088217593729496, + 0.009915717877447605, + 0.006531118880957365, + 0.00749603845179081, + 0.009796859696507454, + 0.006332488730549812, + 0.009740522131323814, + 0.01630621775984764, + 0.011815284378826618, + 0.007594097871333361, + 0.0074593001045286655, + 0.006944813299924135, + 0.004532718099653721, + 0.005481293890625238, + 0.004912352655082941, + 0.01199740543961525, + 0.006854054518043995, + 0.0034701975528150797, + 0.005336576607078314, + 0.003505285596475005, + 0.0017932463670149446, + 0.002723932731896639, + 0.0020690090022981167, + 0.006121628452092409, + 0.00341408746317029, + 0.0019975865725427866, + 0.0027331113815307617, + 0.0014400543877854943, + 0.001118777203373611, + 0.0013170924503356218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.012548773549497128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013390155509114265 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013390155509114265, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013390155509114265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14842477440834045, + 0.06264267861843109, + 0.20584653317928314, + 0.0949995145201683, + 0.11931398510932922, + 0.13568015396595, + 0.05925053358078003, + 0.07030930370092392, + 0.06816685199737549, + 0.03536568209528923 + ], + "credit_map": [ + 0.01886986568570137, + 0.024959437549114227, + 0.027463339269161224, + 0.015238314867019653, + 0.028714023530483246, + 0.010796519927680492, + 0.009244254790246487, + 0.01313900388777256, + 0.00740011315792799, + 0.010530506260693073, + 0.006430434063076973, + 0.012361794710159302, + 0.01017716247588396, + 0.0035152144264429808, + 0.007790896575897932, + 0.004436554852873087, + 0.020013265311717987, + 0.05286256596446037, + 0.03012060932815075, + 0.023620299994945526, + 0.020134614780545235, + 0.007646805141121149, + 0.035283397883176804, + 0.01616496406495571, + 0.006445486098527908, + 0.020464958623051643, + 0.014135032892227173, + 0.023181702941656113, + 0.00479685515165329, + 0.0066458918154239655, + 0.013883387669920921, + 0.00544620119035244, + 0.008914308622479439, + 0.03155461326241493, + 0.008989011868834496, + 0.021214038133621216, + 0.006135196890681982, + 0.01606021635234356, + 0.012293940410017967, + 0.014152657240629196, + 0.015524115413427353, + 0.016202352941036224, + 0.013370092958211899, + 0.020255668088793755, + 0.011030560359358788, + 0.03368290513753891, + 0.008230706676840782, + 0.017383748665452003, + 0.006617643404752016, + 0.012605638243258, + 0.00504064466804266, + 0.006759822368621826, + 0.006226019933819771, + 0.00960680190473795, + 0.005784622393548489, + 0.006609336938709021, + 0.00712165329605341, + 0.016621755436062813, + 0.010745557025074959, + 0.00690496526658535, + 0.01002210658043623, + 0.00670177536085248, + 0.005832699127495289, + 0.006358793471008539, + 0.006438816897571087, + 0.020059991627931595, + 0.010839138180017471, + 0.006760661024600267, + 0.009809219278395176, + 0.005201191641390324, + 0.0038440399803221226, + 0.005213792901486158, + 0.005038531497120857, + 0.006713146809488535, + 0.005035627633333206, + 0.005593384150415659, + 0.003424736438319087, + 0.003474961966276169, + 0.0028176922351121902, + 0.0032676004339009523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0013390155509114265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up filmed from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward following the subject's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject an elegant black cat crouching on damp, slightly pebbled ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat slowly raise its head to fix its gaze on the night sky, revealing emerald green eyes reflecting a full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool moonlight with a color contrast of deep black, silver, and emerald green?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036182431504130363 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036182431504130363, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036182431504130363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2540596127510071, + 0.07516701519489288, + 0.09207227826118469, + 0.1448805332183838, + 0.1460508555173874, + 0.07463019341230392, + 0.06791697442531586, + 0.07151457667350769, + 0.04073932766914368, + 0.032968588173389435 + ], + "credit_map": [ + 0.03281581029295921, + 0.04771095886826515, + 0.0424208827316761, + 0.02326575666666031, + 0.04855091869831085, + 0.01671087183058262, + 0.01630300283432007, + 0.026281412690877914, + 0.00794056337326765, + 0.012493794783949852, + 0.010826761834323406, + 0.011859712190926075, + 0.011907069012522697, + 0.005131673067808151, + 0.008813120424747467, + 0.006194320973008871, + 0.009886844083666801, + 0.016058389097452164, + 0.015091636218130589, + 0.010847564786672592, + 0.00944902841001749, + 0.006358359940350056, + 0.014060834422707558, + 0.010319620370864868, + 0.010613192804157734, + 0.031099196523427963, + 0.020490946248173714, + 0.03839895874261856, + 0.0064018950797617435, + 0.011947784572839737, + 0.01952570490539074, + 0.0064028664492070675, + 0.011198300868272781, + 0.04215318709611893, + 0.011887701228260994, + 0.025865236297249794, + 0.00762641616165638, + 0.020584186539053917, + 0.013148295693099499, + 0.013587526977062225, + 0.010277669876813889, + 0.009138064458966255, + 0.00780708622187376, + 0.01172136701643467, + 0.007163493428379297, + 0.013818946667015553, + 0.005845886189490557, + 0.008857685141265392, + 0.0071672131307423115, + 0.021199490875005722, + 0.006053958088159561, + 0.009295339696109295, + 0.007223055232316256, + 0.0052757468074560165, + 0.006866678595542908, + 0.004835502244532108, + 0.010142521001398563, + 0.01627313904464245, + 0.009084975346922874, + 0.009007623419165611, + 0.008134471252560616, + 0.006680516060441732, + 0.004514759872108698, + 0.007676573470234871, + 0.005411461926996708, + 0.008528810925781727, + 0.0064470828510820866, + 0.0031475231517106295, + 0.0041017900221049786, + 0.006025352049618959, + 0.002047138288617134, + 0.005030171480029821, + 0.00397768709808588, + 0.009357865899801254, + 0.005988030694425106, + 0.0034090655390173197, + 0.003889569081366062, + 0.0023409200366586447, + 0.0016472729621455073, + 0.0023581755813211203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0036182431504130363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_1/index_tensor([3254, 3254])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_1/index_tensor([3254, 3254])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..07de5455257b3cce9a98fcf4c22985b229ff2a52 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_1/index_tensor([3254, 3254])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a surreal scene crafted in the style of stop-motion clay animation, a slow and deliberate wide-angle shot pans leftward, slightly tilted upward, revealing a vast, dreamlike desert landscape. A massive, pale yellow clay moon hangs motionless in a deep violet sky, casting a soft, ethereal glow over a landscape composed of smooth ochre clay. Scattered throughout, several bizarre clay cacti, their surfaces bearing visible hand-kneading fingerprints and subtle tool marks, gently sway back and forth in an unnatural, breathing rhythm. These cacti serve as their own light source, their colors gently pulsing between deep magenta, electric blue, and soft lavender, casting iridescent shadows on the ground. A small, rough-surfaced camel, made of sandy clay, trudges into view from the right side of the frame. Its four legs laboriously move in a staggered, slightly jerky sequence characteristic of stop-motion animation, its body sinking slightly with each step, leaving a series of distinct, deep indentations in the soft, clay-like sand. Its head lowered slightly, as if enduring the fatigue of a long journey, the camel steadily and calmly traverses this strange and silent world. The entire scene is imbued with a surreal stillness and a touch of solitary poetry, creating a dreamlike atmosphere that is both eerie and tranquil.", + "seed": 14816, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_1/index_3254_seed_14816_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07063376568257809, + "vlm_logprob": -0.07063376568257809, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.07063376568257809, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04401193931698799 + ], + "opposite_logprob_tokens": [ + -3.2940118312835693 + ], + "logprob_reward": -0.04401193931698799, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04401193931698799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15421748161315918, + 0.10820455104112625, + 0.04783613607287407, + 0.06939230114221573, + 0.044547051191329956, + 0.1519884467124939, + 0.07792088389396667, + 0.16551029682159424, + 0.1174575611948967, + 0.06292525678873062 + ], + "credit_map": [ + 0.029878973960876465, + 0.028637249022722244, + 0.012513305060565472, + 0.010611484758555889, + 0.030018575489521027, + 0.01453548762947321, + 0.011920981109142303, + 0.016101447865366936, + 0.011764281429350376, + 0.013565042056143284, + 0.01485480833798647, + 0.007544531021267176, + 0.013651409186422825, + 0.01686563715338707, + 0.008640307933092117, + 0.02131853997707367, + 0.006617676466703415, + 0.00559175293892622, + 0.00564446859061718, + 0.00535038486123085, + 0.006528508383780718, + 0.008072515949606895, + 0.003968134522438049, + 0.006062701810151339, + 0.012160558253526688, + 0.009724749252200127, + 0.005601169541478157, + 0.006344255059957504, + 0.013051445595920086, + 0.00765501894056797, + 0.005131931509822607, + 0.0097231799736619, + 0.005163883790373802, + 0.0053338827565312386, + 0.003591290907934308, + 0.003881176235154271, + 0.003815674688667059, + 0.007915589027106762, + 0.004531822632998228, + 0.010313737206161022, + 0.022455081343650818, + 0.01859372667968273, + 0.020189162343740463, + 0.019013386219739914, + 0.017003338783979416, + 0.029658358544111252, + 0.010228361934423447, + 0.01484704203903675, + 0.012065739370882511, + 0.00957030151039362, + 0.007094805594533682, + 0.007561530917882919, + 0.007890299893915653, + 0.012793116271495819, + 0.007856565527617931, + 0.013088527135550976, + 0.031996071338653564, + 0.03939638286828995, + 0.016844075173139572, + 0.016275465488433838, + 0.014473721385002136, + 0.019196676090359688, + 0.012487522326409817, + 0.014840393327176571, + 0.016056274995207787, + 0.02047893963754177, + 0.008135543204843998, + 0.006836130749434233, + 0.009967333637177944, + 0.03726577386260033, + 0.009549190290272236, + 0.009168379008769989, + 0.009700009599328041, + 0.00537565303966403, + 0.0059134685434401035, + 0.008489865809679031, + 0.006906337104737759, + 0.008565962314605713, + 0.009229792281985283, + 0.008744177408516407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.04401193931698799 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902226597070694 + ], + "opposite_logprob_tokens": [ + -2.579022169113159 + ], + "logprob_reward": -0.07902226597070694, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902226597070694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10679764300584793, + 0.03344348818063736, + 0.046832166612148285, + 0.027175728231668472, + 0.04788186773657799, + 0.11161282658576965, + 0.12460942566394806, + 0.1792955845594406, + 0.30831241607666016, + 0.014038825407624245 + ], + "credit_map": [ + 0.025434786453843117, + 0.030854059383273125, + 0.007194013334810734, + 0.005373718682676554, + 0.013772326521575451, + 0.007093763444572687, + 0.008502966724336147, + 0.008572009392082691, + 0.005565991159528494, + 0.005141960456967354, + 0.002903459593653679, + 0.00195715157315135, + 0.004844123497605324, + 0.00602485379204154, + 0.002159849740564823, + 0.004846096504479647, + 0.004475173074752092, + 0.004935071337968111, + 0.006384465843439102, + 0.0039551458321511745, + 0.00632680905982852, + 0.0068105366080999374, + 0.004288921132683754, + 0.009656042791903019, + 0.004363550338894129, + 0.004116338677704334, + 0.0022808020003139973, + 0.0022054805886000395, + 0.004362683743238449, + 0.0044738962315022945, + 0.0018069820944219828, + 0.0035659947898238897, + 0.008947494439780712, + 0.00919841043651104, + 0.004655324388295412, + 0.004544246010482311, + 0.006720455829054117, + 0.005620642099529505, + 0.0036919517442584038, + 0.004503339994698763, + 0.020662089809775352, + 0.017405543476343155, + 0.022038282826542854, + 0.012571017257869244, + 0.01090906374156475, + 0.009799757041037083, + 0.007942278869450092, + 0.010284798219799995, + 0.03165742754936218, + 0.017897875979542732, + 0.0206864345818758, + 0.011276176199316978, + 0.011885439045727253, + 0.010973123833537102, + 0.009611512534320354, + 0.010621435008943081, + 0.03464041277766228, + 0.036004893481731415, + 0.015144468285143375, + 0.020346418023109436, + 0.017498617991805077, + 0.019616495817899704, + 0.014013559557497501, + 0.02203071489930153, + 0.047609467059373856, + 0.045163288712501526, + 0.016685742884874344, + 0.02606394700706005, + 0.032625045627355576, + 0.08488933742046356, + 0.0254821814596653, + 0.02979341708123684, + 0.0023317718878388405, + 0.0018520740559324622, + 0.0018780323443934321, + 0.0017934286734089255, + 0.0017185361357405782, + 0.001694204518571496, + 0.001398699707351625, + 0.0013720786664634943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.07902226597070694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015156697481870651 + ], + "opposite_logprob_tokens": [ + -6.501515865325928 + ], + "logprob_reward": -0.0015156697481870651, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015156697481870651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17348285019397736, + 0.0844741091132164, + 0.04310745745897293, + 0.03923853486776352, + 0.0808427706360817, + 0.12704463303089142, + 0.12627892196178436, + 0.19357138872146606, + 0.11427955329418182, + 0.017679816111922264 + ], + "credit_map": [ + 0.029336974024772644, + 0.03639904409646988, + 0.012948508374392986, + 0.009921898134052753, + 0.046679869294166565, + 0.012582085095345974, + 0.009196106344461441, + 0.016418365761637688, + 0.01135231088846922, + 0.011997858993709087, + 0.00831313244998455, + 0.006667038891464472, + 0.013077535666525364, + 0.00848359800875187, + 0.00665636733174324, + 0.017926258966326714, + 0.004906000103801489, + 0.005339684896171093, + 0.00545576773583889, + 0.004028290510177612, + 0.007389760110527277, + 0.005194069817662239, + 0.004309786017984152, + 0.006484095938503742, + 0.004943716339766979, + 0.004917245823889971, + 0.003363364841789007, + 0.0036847435403615236, + 0.0070123313926160336, + 0.006662675179541111, + 0.0027290410362184048, + 0.0059254104271531105, + 0.010790470987558365, + 0.011397399939596653, + 0.0066361455246806145, + 0.00573075283318758, + 0.009341780096292496, + 0.01316901482641697, + 0.007372384890913963, + 0.01640481688082218, + 0.02964700385928154, + 0.018018096685409546, + 0.018986163660883904, + 0.011949910782277584, + 0.013150683604180813, + 0.016467928886413574, + 0.008255122229456902, + 0.01056972611695528, + 0.026388835161924362, + 0.018102120608091354, + 0.015815820544958115, + 0.015279511921107769, + 0.015186450444161892, + 0.011157391592860222, + 0.010981937870383263, + 0.013366853818297386, + 0.040632642805576324, + 0.039435796439647675, + 0.022008584812283516, + 0.021075144410133362, + 0.023495813831686974, + 0.016054686158895493, + 0.012205922044813633, + 0.01866278424859047, + 0.010310492478311062, + 0.03389774635434151, + 0.0038559006061404943, + 0.005393617786467075, + 0.007032249588519335, + 0.03937366232275963, + 0.007979398593306541, + 0.006436482071876526, + 0.0027700550854206085, + 0.002587685827165842, + 0.0014262470649555326, + 0.0014899962116032839, + 0.0018437529215589166, + 0.003476796904578805, + 0.0022383672185242176, + 0.0018469140632078052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.0015156697481870651 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17997200787067413 + ], + "opposite_logprob_tokens": [ + -1.8049720525741577 + ], + "logprob_reward": -0.17997200787067413, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17997200787067413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13476692140102386, + 0.06122959405183792, + 0.03507500886917114, + 0.03672831878066063, + 0.03408040106296539, + 0.1106373518705368, + 0.09959821403026581, + 0.2188599407672882, + 0.2288549244403839, + 0.04016930237412453 + ], + "credit_map": [ + 0.020926544442772865, + 0.026319459080696106, + 0.01123241987079382, + 0.008776155300438404, + 0.027283426374197006, + 0.01278716791421175, + 0.010539607144892216, + 0.016902148723602295, + 0.006693743169307709, + 0.007841806858778, + 0.007781540509313345, + 0.004320645239204168, + 0.006728772073984146, + 0.00905337929725647, + 0.004095250740647316, + 0.014714454300701618, + 0.004971872083842754, + 0.004050871357321739, + 0.004158133175224066, + 0.004335036966949701, + 0.005209874827414751, + 0.004520830232650042, + 0.003215963486582041, + 0.004612427204847336, + 0.008055123500525951, + 0.005211811512708664, + 0.003917862195521593, + 0.0031895742285996675, + 0.005044375546276569, + 0.004961048252880573, + 0.002434719819575548, + 0.00391380162909627, + 0.0044446540996432304, + 0.004714071750640869, + 0.0028538471087813377, + 0.0035283772740513086, + 0.004747532308101654, + 0.00404199305921793, + 0.004487259313464165, + 0.005262665916234255, + 0.02115466631948948, + 0.014834052883088589, + 0.021399514749646187, + 0.013103954493999481, + 0.013009203597903252, + 0.009755984880030155, + 0.007853280752897263, + 0.009526695124804974, + 0.023347843438386917, + 0.013952953740954399, + 0.016382910311222076, + 0.010823297314345837, + 0.010665719397366047, + 0.007603581063449383, + 0.008798358030617237, + 0.008023550733923912, + 0.045867156237363815, + 0.046205952763557434, + 0.027095003053545952, + 0.019854940474033356, + 0.021253269165754318, + 0.019980939105153084, + 0.015701837837696075, + 0.02290085330605507, + 0.017518192529678345, + 0.06950408220291138, + 0.005144116934388876, + 0.008777672424912453, + 0.011224323883652687, + 0.09336981922388077, + 0.010411472991108894, + 0.012905222363770008, + 0.007476392667740583, + 0.00415540998801589, + 0.004766657017171383, + 0.003198168706148863, + 0.004462081473320723, + 0.007488193456083536, + 0.0040305620059370995, + 0.004591834731400013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.17997200787067413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048646945506334305 + ], + "opposite_logprob_tokens": [ + -3.048646926879883 + ], + "logprob_reward": -0.048646945506334305, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048646945506334305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05310387164354324, + 0.09091116487979889, + 0.04332265630364418, + 0.08343983441591263, + 0.030619263648986816, + 0.08444087207317352, + 0.0816139206290245, + 0.09581457823514938, + 0.39660248160362244, + 0.040131378918886185 + ], + "credit_map": [ + 0.006876267958432436, + 0.005901635158807039, + 0.0061249276623129845, + 0.003923848271369934, + 0.00769140524789691, + 0.004231942817568779, + 0.006517560221254826, + 0.0118362782523036, + 0.010171188972890377, + 0.01014569029211998, + 0.008950420655310154, + 0.005842296406626701, + 0.010784417390823364, + 0.011205313727259636, + 0.008490459993481636, + 0.025321373715996742, + 0.004438455682247877, + 0.004703113809227943, + 0.004616914317011833, + 0.003680045483633876, + 0.006916347425431013, + 0.0075629716739058495, + 0.003683191491290927, + 0.007721613626927137, + 0.010469804517924786, + 0.009857201017439365, + 0.005975055042654276, + 0.0075651961378753185, + 0.009066648781299591, + 0.01788618043065071, + 0.006824966985732317, + 0.015794772654771805, + 0.005969711113721132, + 0.005916356109082699, + 0.002721362514421344, + 0.002083276631310582, + 0.005418750457465649, + 0.003565857419744134, + 0.0020829709246754646, + 0.002860975917428732, + 0.009592783637344837, + 0.010027289390563965, + 0.01920323073863983, + 0.009874866344034672, + 0.007836002856492996, + 0.014585889875888824, + 0.005275735631585121, + 0.008045067079365253, + 0.018948713317513466, + 0.011593461036682129, + 0.0107268700376153, + 0.0067675309255719185, + 0.006831696722656488, + 0.01183801144361496, + 0.006750586908310652, + 0.008157049305737019, + 0.014879170805215836, + 0.021178022027015686, + 0.008918763138353825, + 0.013162053190171719, + 0.009106963872909546, + 0.01104878168553114, + 0.008041612803936005, + 0.009479206055402756, + 0.031784262508153915, + 0.1113843247294426, + 0.009350663982331753, + 0.018643693998456, + 0.02441776916384697, + 0.15320976078510284, + 0.023660369217395782, + 0.024151578545570374, + 0.0056740050204098225, + 0.004304904490709305, + 0.0027776765637099743, + 0.0028249279130250216, + 0.0045056878589093685, + 0.009596194140613079, + 0.005111088976264, + 0.005336891394108534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.048646945506334305 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal scene crafted in the style of stop-motion clay animation, a slow and deliberate wide-angle shot pans leftward, slightly tilted upward, revealing a vast, dreamlike desert landscape. A massive, pale yellow clay moon hangs motionless in a deep violet sky, casting a soft, ethereal glow over a landscape composed of smooth ochre clay. Scattered throughout, several bizarre clay cacti, their surfaces bearing visible hand-kneading fingerprints and subtle tool marks, gently sway back and forth in an unnatural, breathing rhythm. These cacti serve as their own light source, their colors gently pulsing between deep magenta, electric blue, and soft lavender, casting iridescent shadows on the ground. A small, rough-surfaced camel, made of sandy clay, trudges into view from the right side of the frame. Its four legs laboriously move in a staggered, slightly jerky sequence characteristic of stop-motion animation, its body sinking slightly with each step, leaving a series of distinct, deep indentations in the soft, clay-like sand. Its head lowered slightly, as if enduring the fatigue of a long journey, the camel steadily and calmly traverses this strange and silent world. The entire scene is imbued with a surreal stillness and a touch of solitary poetry, creating a dreamlike atmosphere that is both eerie and tranquil.", + "seed": 14817, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_1/index_3254_seed_14817_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21572170932777227, + "vlm_logprob": -0.21572170932777227, + "vlm_yes_no_margin": 2.2749999880790712, + "vlm_token_credit": -0.21572170932777227, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23349155485630035 + ], + "opposite_logprob_tokens": [ + -1.6084915399551392 + ], + "logprob_reward": -0.23349155485630035, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23349155485630035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37714242935180664, + 0.050469204783439636, + 0.06123431771993637, + 0.025330690667033195, + 0.0787462368607521, + 0.16199727356433868, + 0.1572890281677246, + 0.030142098665237427, + 0.025363942608237267, + 0.032284755259752274 + ], + "credit_map": [ + 0.046641696244478226, + 0.018009599298238754, + 0.027613883838057518, + 0.020690349861979485, + 0.09761200100183487, + 0.0964367538690567, + 0.02853226289153099, + 0.04160588979721069, + 0.00858744140714407, + 0.0030804884154349566, + 0.007263461127877235, + 0.005009131506085396, + 0.006430499255657196, + 0.006863549817353487, + 0.00616149976849556, + 0.007073134649544954, + 0.00691751204431057, + 0.005902345757931471, + 0.007209551054984331, + 0.005417096894234419, + 0.006860553286969662, + 0.009065507911145687, + 0.006200617644935846, + 0.013661129400134087, + 0.0030601811595261097, + 0.0017572448123246431, + 0.0028587044216692448, + 0.0023794977460056543, + 0.0030176264699548483, + 0.0034273499622941017, + 0.002274986356496811, + 0.006555098108947277, + 0.007819639518857002, + 0.011008268222212791, + 0.011575737968087196, + 0.010565640404820442, + 0.00852823257446289, + 0.011570749804377556, + 0.007845965214073658, + 0.009832009673118591, + 0.024229571223258972, + 0.02055450901389122, + 0.020122583955526352, + 0.017998095601797104, + 0.021519146859645844, + 0.02153354324400425, + 0.015666743740439415, + 0.02037307620048523, + 0.016745513305068016, + 0.015601003542542458, + 0.018434660509228706, + 0.018579844385385513, + 0.014141266234219074, + 0.017877191305160522, + 0.023708689957857132, + 0.032200854271650314, + 0.003971078433096409, + 0.0023387998808175325, + 0.004003896843641996, + 0.002818587003275752, + 0.003782059298828244, + 0.004375325050204992, + 0.0033216201700270176, + 0.005530731752514839, + 0.0032500626984983683, + 0.0029054470360279083, + 0.0037231699097901583, + 0.0036449774634093046, + 0.002516878303140402, + 0.0025506275705993176, + 0.003058993723243475, + 0.0037137847393751144, + 0.0033734655007719994, + 0.0025598444044589996, + 0.003545616753399372, + 0.005763969849795103, + 0.0036324451211839914, + 0.003625475103035569, + 0.004703517537564039, + 0.0050804209895431995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.23349155485630035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.313483327627182 + ], + "opposite_logprob_tokens": [ + -1.3134833574295044 + ], + "logprob_reward": -0.313483327627182, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.313483327627182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22124825417995453, + 0.10360344499349594, + 0.0978032574057579, + 0.07128236442804337, + 0.165772944688797, + 0.08070949465036392, + 0.05005456507205963, + 0.060156356543302536, + 0.10479719936847687, + 0.0445721372961998 + ], + "credit_map": [ + 0.03069978766143322, + 0.009091680869460106, + 0.021780801936984062, + 0.01487930677831173, + 0.047029655426740646, + 0.04784108325839043, + 0.015749024227261543, + 0.03417690470814705, + 0.012183452025055885, + 0.006628931034356356, + 0.01779114454984665, + 0.01583821326494217, + 0.008794596418738365, + 0.008231722749769688, + 0.013662409037351608, + 0.020472979173064232, + 0.008078495971858501, + 0.006275296676903963, + 0.013341688551008701, + 0.010917367413640022, + 0.012337892316281796, + 0.009237341582775116, + 0.015112756751477718, + 0.022502418607473373, + 0.008361846208572388, + 0.005453518126159906, + 0.009781157597899437, + 0.007348737679421902, + 0.009305177256464958, + 0.00689029973000288, + 0.006062494590878487, + 0.018079129979014397, + 0.021845577284693718, + 0.012796301394701004, + 0.012990930117666721, + 0.012158402241766453, + 0.028944451361894608, + 0.028556037694215775, + 0.0208421740680933, + 0.027639076113700867, + 0.015185656026005745, + 0.009479020722210407, + 0.006152964197099209, + 0.005712010432034731, + 0.01718781143426895, + 0.011233930476009846, + 0.007540235295891762, + 0.008217871189117432, + 0.00584667781367898, + 0.006459538359194994, + 0.006286332383751869, + 0.006229051388800144, + 0.00484064407646656, + 0.00546942139044404, + 0.006316155195236206, + 0.008606743067502975, + 0.006209348328411579, + 0.004937032703310251, + 0.006076375488191843, + 0.006072799675166607, + 0.0062385606579482555, + 0.008381478488445282, + 0.010958625003695488, + 0.011282140389084816, + 0.013755591586232185, + 0.006440632976591587, + 0.011850301176309586, + 0.010223782621324062, + 0.016668861731886864, + 0.008099328726530075, + 0.023665759712457657, + 0.0140929464250803, + 0.005669733975082636, + 0.003926069010049105, + 0.004650828894227743, + 0.006212241481989622, + 0.00658826157450676, + 0.00622933404520154, + 0.005412442144006491, + 0.00588322663679719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.313483327627182 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017149040941148996 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017149040941148996, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017149040941148996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1970534324645996, + 0.12205660343170166, + 0.22075557708740234, + 0.1217188760638237, + 0.023295830935239792, + 0.09544582664966583, + 0.038733821362257004, + 0.039526112377643585, + 0.07626110315322876, + 0.06515280902385712 + ], + "credit_map": [ + 0.030665110796689987, + 0.022869639098644257, + 0.01721530221402645, + 0.02180851250886917, + 0.03240825980901718, + 0.036618784070014954, + 0.01467194501310587, + 0.02079588547348976, + 0.013359549455344677, + 0.012381196953356266, + 0.01889340952038765, + 0.012284992262721062, + 0.0194503515958786, + 0.013960909098386765, + 0.016429508104920387, + 0.015296680852770805, + 0.03241954743862152, + 0.028518537059426308, + 0.02622518502175808, + 0.01915915496647358, + 0.0368376150727272, + 0.026630539447069168, + 0.01642068848013878, + 0.03454430028796196, + 0.02103469707071781, + 0.010985099710524082, + 0.011912564747035503, + 0.014218480326235294, + 0.02245609275996685, + 0.018175410106778145, + 0.008215557783842087, + 0.0147209782153368, + 0.002653537318110466, + 0.004551087040454149, + 0.0026461859233677387, + 0.0022227962035685778, + 0.0030705630779266357, + 0.002911592833697796, + 0.0016652338672429323, + 0.003574834205210209, + 0.017390694469213486, + 0.01082249078899622, + 0.009075349196791649, + 0.008789101615548134, + 0.019003793597221375, + 0.013378571718931198, + 0.006986350752413273, + 0.009999476373195648, + 0.005279566161334515, + 0.0035189660266041756, + 0.0035417526960372925, + 0.004728971980512142, + 0.006537517532706261, + 0.005096821580082178, + 0.003974548541009426, + 0.006055680103600025, + 0.005275999661535025, + 0.004460844211280346, + 0.004452483728528023, + 0.00521255936473608, + 0.004745809361338615, + 0.004914681892842054, + 0.006180876865983009, + 0.004282859619706869, + 0.011305471882224083, + 0.011218844912946224, + 0.00622215773910284, + 0.007669576443731785, + 0.014645940624177456, + 0.0047073401510715485, + 0.012719610705971718, + 0.007772155571728945, + 0.006941257975995541, + 0.007625381462275982, + 0.007005337160080671, + 0.010208956897258759, + 0.004964764229953289, + 0.00934371817857027, + 0.006116317585110664, + 0.012947076000273228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0017149040941148996 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871176838874817 + ], + "opposite_logprob_tokens": [ + -1.137117624282837 + ], + "logprob_reward": -0.3871176838874817, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.3871176838874817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37217187881469727, + 0.053899988532066345, + 0.1600823700428009, + 0.05973216891288757, + 0.0560491606593132, + 0.06037938594818115, + 0.06248435750603676, + 0.04273347556591034, + 0.06595086306333542, + 0.06651636213064194 + ], + "credit_map": [ + 0.04144739732146263, + 0.018321305513381958, + 0.029489923268556595, + 0.02940865233540535, + 0.07237326353788376, + 0.08459345996379852, + 0.03932259604334831, + 0.05721527710556984, + 0.005641221534460783, + 0.0055838474072515965, + 0.006753799971193075, + 0.006636260077357292, + 0.0055807000026106834, + 0.006631898228079081, + 0.010523292236030102, + 0.006548967212438583, + 0.016396058723330498, + 0.014202832244336605, + 0.02315993793308735, + 0.014397009275853634, + 0.024217618629336357, + 0.025821421295404434, + 0.019276434555649757, + 0.022611059248447418, + 0.010278948582708836, + 0.003441083477810025, + 0.004262356553226709, + 0.005503121763467789, + 0.013789444230496883, + 0.008441956713795662, + 0.004134479444473982, + 0.00988078210502863, + 0.008551937527954578, + 0.006616337690502405, + 0.006541132461279631, + 0.0052277036011219025, + 0.009800806641578674, + 0.006989188492298126, + 0.005159203894436359, + 0.007162847556173801, + 0.008160993456840515, + 0.008166034705936909, + 0.006629951298236847, + 0.008876657113432884, + 0.007161823101341724, + 0.008258691988885403, + 0.006394957657903433, + 0.0067302812822163105, + 0.007209392264485359, + 0.007740227039903402, + 0.009850887581706047, + 0.005925345234572887, + 0.005491201765835285, + 0.008008532226085663, + 0.007281447295099497, + 0.010977324098348618, + 0.005039317067712545, + 0.005282261408865452, + 0.005101402755826712, + 0.0037992624565958977, + 0.0043542152270674706, + 0.006423022598028183, + 0.006968441419303417, + 0.005765549372881651, + 0.009840928949415684, + 0.006787370890378952, + 0.00793392676860094, + 0.007252158597111702, + 0.006371004041284323, + 0.009409921243786812, + 0.0074143786914646626, + 0.010941174812614918, + 0.007404008414596319, + 0.0059578996151685715, + 0.00669587729498744, + 0.009566148743033409, + 0.009944532066583633, + 0.0073853749781847, + 0.009294167160987854, + 0.010268354788422585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.3871176838874817 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280107617378235 + ], + "opposite_logprob_tokens": [ + -2.01780104637146 + ], + "logprob_reward": -0.14280107617378235, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280107617378235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20370951294898987, + 0.06157940998673439, + 0.10970443487167358, + 0.05729558318853378, + 0.25307103991508484, + 0.027710789814591408, + 0.02918994054198265, + 0.1327856332063675, + 0.05693645402789116, + 0.06801711767911911 + ], + "credit_map": [ + 0.0252725500613451, + 0.00899440236389637, + 0.014879920519888401, + 0.010408958420157433, + 0.05497608333826065, + 0.05204126611351967, + 0.01682722009718418, + 0.020309120416641235, + 0.0067124138586223125, + 0.004965684376657009, + 0.008293865248560905, + 0.012530065141618252, + 0.006007979158312082, + 0.0066932509653270245, + 0.00709390128031373, + 0.009282255545258522, + 0.01341947726905346, + 0.006648679729551077, + 0.017111746594309807, + 0.009835510514676571, + 0.013326747342944145, + 0.012726371176540852, + 0.02189197577536106, + 0.01474393717944622, + 0.00637427531182766, + 0.005265700165182352, + 0.008234155364334583, + 0.006803033407777548, + 0.008015316911041737, + 0.006138939410448074, + 0.006485987454652786, + 0.009978183545172215, + 0.024855324998497963, + 0.024341469630599022, + 0.03759092465043068, + 0.035966407507658005, + 0.035182204097509384, + 0.03433018922805786, + 0.019003812223672867, + 0.04180074855685234, + 0.003901991294696927, + 0.0034782872535288334, + 0.0032607433386147022, + 0.003023112891241908, + 0.003565208986401558, + 0.0035805334337055683, + 0.0030228185933083296, + 0.003878096817061305, + 0.0034714119974523783, + 0.0031121978536248207, + 0.003878173651173711, + 0.0039505381137132645, + 0.002600046107545495, + 0.0023449792061001062, + 0.005537588149309158, + 0.00429501011967659, + 0.01564299501478672, + 0.010770220309495926, + 0.013117065653204918, + 0.012547828257083893, + 0.01755174621939659, + 0.021753134205937386, + 0.018245849758386612, + 0.023156797513365746, + 0.008496427908539772, + 0.004428141284734011, + 0.004589204676449299, + 0.005488645285367966, + 0.011336103081703186, + 0.005778797902166843, + 0.009933163411915302, + 0.006885976996272802, + 0.008127347566187382, + 0.005122989881783724, + 0.00790572538971901, + 0.008039933629333973, + 0.007511709816753864, + 0.010455649346113205, + 0.009175999090075493, + 0.011677773669362068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.14280107617378235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal scene crafted in the style of stop-motion clay animation, a slow and deliberate wide-angle shot pans leftward, slightly tilted upward, revealing a vast, dreamlike desert landscape. A massive, pale yellow clay moon hangs motionless in a deep violet sky, casting a soft, ethereal glow over a landscape composed of smooth ochre clay. Scattered throughout, several bizarre clay cacti, their surfaces bearing visible hand-kneading fingerprints and subtle tool marks, gently sway back and forth in an unnatural, breathing rhythm. These cacti serve as their own light source, their colors gently pulsing between deep magenta, electric blue, and soft lavender, casting iridescent shadows on the ground. A small, rough-surfaced camel, made of sandy clay, trudges into view from the right side of the frame. Its four legs laboriously move in a staggered, slightly jerky sequence characteristic of stop-motion animation, its body sinking slightly with each step, leaving a series of distinct, deep indentations in the soft, clay-like sand. Its head lowered slightly, as if enduring the fatigue of a long journey, the camel steadily and calmly traverses this strange and silent world. The entire scene is imbued with a surreal stillness and a touch of solitary poetry, creating a dreamlike atmosphere that is both eerie and tranquil.", + "seed": 14818, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_1/index_3254_seed_14818_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13038015404017642, + "vlm_logprob": -0.13038015404017642, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.13038015404017642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.057164546102285385 + ], + "opposite_logprob_tokens": [ + -3.057164430618286 + ], + "logprob_reward": -0.057164546102285385, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.057164546102285385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17815348505973816, + 0.031362660229206085, + 0.020466141402721405, + 0.026660306379199028, + 0.04577424004673958, + 0.18185560405254364, + 0.13838040828704834, + 0.03888387978076935, + 0.05731987953186035, + 0.2811433970928192 + ], + "credit_map": [ + 0.0185691025108099, + 0.02239951863884926, + 0.015950355678796768, + 0.01605992391705513, + 0.027538364753127098, + 0.022850574925541878, + 0.03491640090942383, + 0.0198692437261343, + 0.003603476332500577, + 0.0037239915691316128, + 0.0037430135998874903, + 0.0029459046199917793, + 0.004276512190699577, + 0.004380309022963047, + 0.0035631186328828335, + 0.005126334726810455, + 0.002901999279856682, + 0.003305712016299367, + 0.002301560714840889, + 0.0025001922622323036, + 0.0027027784381061792, + 0.002068938221782446, + 0.0018830663757398725, + 0.0028018930461257696, + 0.00261873216368258, + 0.00356631213799119, + 0.0035665747709572315, + 0.0028094053268432617, + 0.0036584383342415094, + 0.002719433046877384, + 0.00230625132098794, + 0.005415159743279219, + 0.006581979803740978, + 0.0035274145193398, + 0.010239146649837494, + 0.002616569632664323, + 0.0043810573406517506, + 0.008665948174893856, + 0.003336557187139988, + 0.0064255669713020325, + 0.0344255305826664, + 0.015883667394518852, + 0.03205600380897522, + 0.013560503721237183, + 0.03164894878864288, + 0.016237841919064522, + 0.014470169320702553, + 0.023572934791445732, + 0.02484774962067604, + 0.018745599314570427, + 0.0380968376994133, + 0.010557190515100956, + 0.009113634005188942, + 0.011220297776162624, + 0.008501383475959301, + 0.01729772239923477, + 0.0048535908572375774, + 0.005628332961350679, + 0.005049285478889942, + 0.003975654952228069, + 0.004475334193557501, + 0.004174014553427696, + 0.005017132498323917, + 0.005710532423108816, + 0.009501909837126732, + 0.007411920931190252, + 0.010856705717742443, + 0.0060591623187065125, + 0.005259414203464985, + 0.007083909120410681, + 0.005658003501594067, + 0.005488854832947254, + 0.07167156040668488, + 0.01537473127245903, + 0.1122690811753273, + 0.014782613143324852, + 0.017027314752340317, + 0.016604792326688766, + 0.01845874823629856, + 0.014954566024243832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.057164546102285385 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161154866218567 + ], + "opposite_logprob_tokens": [ + -1.7016115188598633 + ], + "logprob_reward": -0.20161154866218567, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161154866218567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052757345139980316, + 0.032270681113004684, + 0.04923848807811737, + 0.04682875797152519, + 0.2195344865322113, + 0.25509050488471985, + 0.13617254793643951, + 0.03190348669886589, + 0.033907629549503326, + 0.14229603111743927 + ], + "credit_map": [ + 0.006602316163480282, + 0.012548882514238358, + 0.004638017620891333, + 0.0043756416998803616, + 0.007338396273553371, + 0.0069874548353254795, + 0.004476513247936964, + 0.0057901302352547646, + 0.0037648642901331186, + 0.005909419618546963, + 0.0032013803720474243, + 0.002776217879727483, + 0.004485394340008497, + 0.004623778630048037, + 0.0026738252490758896, + 0.00483580632135272, + 0.007678113412111998, + 0.005984233692288399, + 0.014823127537965775, + 0.0035618722904473543, + 0.004254654515534639, + 0.005215466022491455, + 0.004321202170103788, + 0.003399819601327181, + 0.007883752696216106, + 0.005558960139751434, + 0.005593964364379644, + 0.005199559964239597, + 0.004208890255540609, + 0.0061922837048769, + 0.003980782348662615, + 0.008210568688809872, + 0.05582847818732262, + 0.011220532469451427, + 0.02279258705675602, + 0.011863046325743198, + 0.017958732321858406, + 0.06189906224608421, + 0.011279015801846981, + 0.026693059131503105, + 0.04463954269886017, + 0.020822975784540176, + 0.03379548713564873, + 0.023976827040314674, + 0.03585795685648918, + 0.02880711667239666, + 0.020168934017419815, + 0.04702167958021164, + 0.01946665719151497, + 0.012106601148843765, + 0.04871981218457222, + 0.012134293094277382, + 0.00834195502102375, + 0.01552046462893486, + 0.009924215264618397, + 0.009958576411008835, + 0.004205630160868168, + 0.0034904314670711756, + 0.006604333873838186, + 0.005737334955483675, + 0.0019884207285940647, + 0.0023749254178255796, + 0.004425432998687029, + 0.003076978726312518, + 0.006525017321109772, + 0.0046260240487754345, + 0.00861254520714283, + 0.002890364732593298, + 0.0035191811621189117, + 0.002143983030691743, + 0.0032653200905770063, + 0.002325196750462055, + 0.03694837912917137, + 0.007677755784243345, + 0.057117581367492676, + 0.007822292856872082, + 0.0086338110268116, + 0.00819393154233694, + 0.007747299503535032, + 0.008154993876814842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.20161154866218567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006370898918248713 + ], + "opposite_logprob_tokens": [ + -7.375637054443359 + ], + "logprob_reward": -0.0006370898918248713, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006370898918248713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08106083422899246, + 0.06615567207336426, + 0.045332252979278564, + 0.04484276846051216, + 0.052617140114307404, + 0.07159975916147232, + 0.1265176236629486, + 0.05761710926890373, + 0.06593329459428787, + 0.3883235454559326 + ], + "credit_map": [ + 0.01065223291516304, + 0.01784135401248932, + 0.006905366200953722, + 0.006541645620018244, + 0.010477139614522457, + 0.01109100878238678, + 0.006736976094543934, + 0.010815110988914967, + 0.007283967919647694, + 0.011952443979680538, + 0.005299814976751804, + 0.004693014547228813, + 0.01573583483695984, + 0.007243383210152388, + 0.00611870177090168, + 0.007828515954315662, + 0.006797917187213898, + 0.007052472792565823, + 0.008551533333957195, + 0.003846072591841221, + 0.0036524871829897165, + 0.005431939382106066, + 0.004199816379696131, + 0.0058000111021101475, + 0.006331995595246553, + 0.006358328741043806, + 0.009595842100679874, + 0.0049898503348231316, + 0.003599186660721898, + 0.0040759919211268425, + 0.004106371197849512, + 0.0057852016761898994, + 0.004969295114278793, + 0.0033582996111363173, + 0.00511506712064147, + 0.0033235817681998014, + 0.004108656197786331, + 0.014571578241884708, + 0.0038345407228916883, + 0.013336123898625374, + 0.012211527675390244, + 0.013532034121453762, + 0.009349529631435871, + 0.005961643066257238, + 0.00962461531162262, + 0.0073175979778170586, + 0.005054583307355642, + 0.008548218756914139, + 0.023378051817417145, + 0.008362365886569023, + 0.04801112040877342, + 0.011037576012313366, + 0.006007780786603689, + 0.011096919886767864, + 0.007909526117146015, + 0.010714287869632244, + 0.005785734858363867, + 0.008223014883697033, + 0.005994941107928753, + 0.011277684941887856, + 0.006690443027764559, + 0.005763464607298374, + 0.007549431174993515, + 0.006332393269985914, + 0.008039173670113087, + 0.008607874624431133, + 0.009847993031144142, + 0.007060914300382137, + 0.010458119213581085, + 0.004809685051441193, + 0.01036123838275671, + 0.006748298183083534, + 0.1109800934791565, + 0.018095649778842926, + 0.1649913191795349, + 0.020323898643255234, + 0.018510719761252403, + 0.015439563430845737, + 0.021044770255684853, + 0.018937531858682632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0006370898918248713 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31352028250694275 + ], + "opposite_logprob_tokens": [ + -1.3135203123092651 + ], + "logprob_reward": -0.31352028250694275, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31352028250694275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1189134269952774, + 0.03159940615296364, + 0.021778080612421036, + 0.03990699350833893, + 0.09738789498806, + 0.09853490442037582, + 0.10883087664842606, + 0.024020526558160782, + 0.0407535657286644, + 0.41827425360679626 + ], + "credit_map": [ + 0.01273308228701353, + 0.015108486637473106, + 0.013313855975866318, + 0.011029926128685474, + 0.01865079626441002, + 0.015292001888155937, + 0.019151195883750916, + 0.0136340893805027, + 0.004061372950673103, + 0.0052568381652235985, + 0.003939504269510508, + 0.0028796803671866655, + 0.0038227441255003214, + 0.0031799450516700745, + 0.003938394598662853, + 0.004520931281149387, + 0.0038390520494431257, + 0.0024607894010841846, + 0.00469037564471364, + 0.0020315530709922314, + 0.002511924598366022, + 0.0022049786057323217, + 0.0018468908965587616, + 0.0021925189066678286, + 0.006463363766670227, + 0.00524790957570076, + 0.009633170440793037, + 0.003811908420175314, + 0.0037495922297239304, + 0.0030532691162079573, + 0.0030786481220275164, + 0.0048691341653466225, + 0.021895017474889755, + 0.004317661747336388, + 0.012261059135198593, + 0.004874556791037321, + 0.0063779610209167, + 0.0261672530323267, + 0.004679132252931595, + 0.016815263777971268, + 0.00930430181324482, + 0.009941493161022663, + 0.00997895561158657, + 0.012273014523088932, + 0.02186467871069908, + 0.012390866875648499, + 0.010323500260710716, + 0.01245810091495514, + 0.0135049968957901, + 0.014955814927816391, + 0.03554856777191162, + 0.009447522461414337, + 0.008058586157858372, + 0.011733882129192352, + 0.0077937887981534, + 0.007787721697241068, + 0.003251264803111553, + 0.0027218300383538008, + 0.002977998461574316, + 0.0026356771122664213, + 0.002987948013469577, + 0.0023625760804861784, + 0.0032502147369086742, + 0.0038330203387886286, + 0.005380412098020315, + 0.0060014137998223305, + 0.007029225118458271, + 0.004853406455367804, + 0.0038933833129704, + 0.003494031261652708, + 0.004838577937334776, + 0.005263120401650667, + 0.11290634423494339, + 0.020952655002474785, + 0.17315155267715454, + 0.021816646680235863, + 0.023444917052984238, + 0.020121991634368896, + 0.025267213582992554, + 0.02061295509338379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.31352028250694275 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896730303764343 + ], + "opposite_logprob_tokens": [ + -2.578967332839966 + ], + "logprob_reward": -0.07896730303764343, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896730303764343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050014350563287735, + 0.03507016599178314, + 0.00832983385771513, + 0.021093489602208138, + 0.021332023665308952, + 0.037941448390483856, + 0.2109779268503189, + 0.04750203341245651, + 0.07218143343925476, + 0.4955572485923767 + ], + "credit_map": [ + 0.005733355414122343, + 0.007286782842129469, + 0.0045633031986653805, + 0.0044287508353590965, + 0.005469885189086199, + 0.006019584368914366, + 0.011265245266258717, + 0.005247441586107016, + 0.004454460926353931, + 0.0023410331923514605, + 0.004842555150389671, + 0.0020858708303421736, + 0.00671762228012085, + 0.0058396682143211365, + 0.0034095149021595716, + 0.005379439331591129, + 0.0011713140411302447, + 0.0008135823300108314, + 0.0012345563154667616, + 0.0007447340176440775, + 0.001371961203403771, + 0.0010250990744680166, + 0.0008448522421531379, + 0.0011237353319302201, + 0.003603438613936305, + 0.0020606268662959337, + 0.0018860666314139962, + 0.0014127050526440144, + 0.0022590290755033493, + 0.002858873223885894, + 0.0017877909122034907, + 0.005224959924817085, + 0.005320909433066845, + 0.0013731918297708035, + 0.0025398628786206245, + 0.0011593266390264034, + 0.0018011126667261124, + 0.004384562373161316, + 0.0013494889717549086, + 0.003403568407520652, + 0.007402221672236919, + 0.0024617239832878113, + 0.006931293290108442, + 0.0026459991931915283, + 0.005947106517851353, + 0.004216479603201151, + 0.003213514806702733, + 0.005123107228428125, + 0.03540823981165886, + 0.011052489280700684, + 0.08090970665216446, + 0.01941170170903206, + 0.011146439239382744, + 0.023025179281830788, + 0.013924935832619667, + 0.016099240630865097, + 0.006550105754286051, + 0.006589662283658981, + 0.004388355650007725, + 0.004963702522218227, + 0.006205967627465725, + 0.0067143989726901054, + 0.006292546633630991, + 0.0057972935028374195, + 0.00809889379888773, + 0.00970499124377966, + 0.007979768328368664, + 0.008105390705168247, + 0.011372932232916355, + 0.005507937166839838, + 0.013273989781737328, + 0.008137528784573078, + 0.13414433598518372, + 0.02457422763109207, + 0.2045707255601883, + 0.026849647983908653, + 0.026523657143115997, + 0.024913936853408813, + 0.029436323791742325, + 0.024544375017285347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.07896730303764343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal scene crafted in the style of stop-motion clay animation, a slow and deliberate wide-angle shot pans leftward, slightly tilted upward, revealing a vast, dreamlike desert landscape. A massive, pale yellow clay moon hangs motionless in a deep violet sky, casting a soft, ethereal glow over a landscape composed of smooth ochre clay. Scattered throughout, several bizarre clay cacti, their surfaces bearing visible hand-kneading fingerprints and subtle tool marks, gently sway back and forth in an unnatural, breathing rhythm. These cacti serve as their own light source, their colors gently pulsing between deep magenta, electric blue, and soft lavender, casting iridescent shadows on the ground. A small, rough-surfaced camel, made of sandy clay, trudges into view from the right side of the frame. Its four legs laboriously move in a staggered, slightly jerky sequence characteristic of stop-motion animation, its body sinking slightly with each step, leaving a series of distinct, deep indentations in the soft, clay-like sand. Its head lowered slightly, as if enduring the fatigue of a long journey, the camel steadily and calmly traverses this strange and silent world. The entire scene is imbued with a surreal stillness and a touch of solitary poetry, creating a dreamlike atmosphere that is both eerie and tranquil.", + "seed": 14819, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_1/index_3254_seed_14819_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07239515604451299, + "vlm_logprob": -0.07239515604451299, + "vlm_yes_no_margin": 3.324999976158142, + "vlm_token_credit": -0.07239515604451299, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02524208091199398 + ], + "opposite_logprob_tokens": [ + -3.9002420902252197 + ], + "logprob_reward": -0.02524208091199398, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02524208091199398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1392488181591034, + 0.08451516181230545, + 0.08291472494602203, + 0.03428631275892258, + 0.4203197956085205, + 0.03687649965286255, + 0.038697708398103714, + 0.0693741887807846, + 0.031086938455700874, + 0.06267982721328735 + ], + "credit_map": [ + 0.013030900619924068, + 0.013037985190749168, + 0.01361765619367361, + 0.0133096594363451, + 0.03705078363418579, + 0.0209661778062582, + 0.015411609783768654, + 0.012824047356843948, + 0.006027232855558395, + 0.009330876171588898, + 0.013805292546749115, + 0.004715089686214924, + 0.009723338298499584, + 0.01566249318420887, + 0.01790783926844597, + 0.00734299561008811, + 0.009113132022321224, + 0.013628708198666573, + 0.010072329081594944, + 0.008926372975111008, + 0.008368898183107376, + 0.0085251834243536, + 0.012912916950881481, + 0.011367185972630978, + 0.0033341487869620323, + 0.009006239473819733, + 0.0027590906247496605, + 0.0047326176427304745, + 0.00456991558894515, + 0.0036272972356528044, + 0.0027273320592939854, + 0.0035296704154461622, + 0.04454157128930092, + 0.13230781257152557, + 0.03474847599864006, + 0.1309979408979416, + 0.019154703244566917, + 0.008940545842051506, + 0.020994972437620163, + 0.02863379567861557, + 0.00464894063770771, + 0.006059674546122551, + 0.004215823020786047, + 0.006054932717233896, + 0.0024186852388083935, + 0.0037591983564198017, + 0.004742637276649475, + 0.0049766055308282375, + 0.0035762665793299675, + 0.005070565268397331, + 0.003714467165991664, + 0.00396843021735549, + 0.00504241744056344, + 0.004086876288056374, + 0.005389525555074215, + 0.007849159650504589, + 0.009045475162565708, + 0.0057289088144898415, + 0.010626665316522121, + 0.013538738712668419, + 0.008617224171757698, + 0.0052263131365180016, + 0.00991901010274887, + 0.006671855691820383, + 0.003276386298239231, + 0.00196850742213428, + 0.0025517710018903017, + 0.004915245808660984, + 0.0049809180200099945, + 0.0019455964211374521, + 0.008541340939700603, + 0.002907171845436096, + 0.011020654812455177, + 0.0052344175055623055, + 0.004974284674972296, + 0.007334370631724596, + 0.006454162299633026, + 0.005973041523247957, + 0.014918043278157711, + 0.006770854815840721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.02524208091199398 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297333985567093 + ], + "opposite_logprob_tokens": [ + -2.23797345161438 + ], + "logprob_reward": -0.11297333985567093, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297333985567093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02781509794294834, + 0.08047361671924591, + 0.014713332988321781, + 0.034424733370542526, + 0.6845383644104004, + 0.040293797850608826, + 0.03497609123587608, + 0.06144111603498459, + 0.01091005653142929, + 0.010413752868771553 + ], + "credit_map": [ + 0.0028892955742776394, + 0.007104157470166683, + 0.0037494059652090073, + 0.002141111297532916, + 0.004639679566025734, + 0.0025256210938096046, + 0.002617484424263239, + 0.0021483427844941616, + 0.004877200350165367, + 0.004433053079992533, + 0.015209496952593327, + 0.00414222152903676, + 0.008307511918246746, + 0.011915205046534538, + 0.0187485720962286, + 0.012840360403060913, + 0.001659665140323341, + 0.002470848150551319, + 0.0016748508205637336, + 0.0016585616394877434, + 0.0014679392334073782, + 0.0019545925315469503, + 0.001844481099396944, + 0.00198239553719759, + 0.003200381528586149, + 0.007395160850137472, + 0.002406541956588626, + 0.0032586222514510155, + 0.0037587580736726522, + 0.0054063391871750355, + 0.0034128257539123297, + 0.005586105398833752, + 0.05151018127799034, + 0.19210532307624817, + 0.06120472401380539, + 0.24725031852722168, + 0.022928539663553238, + 0.013808798976242542, + 0.03506698086857796, + 0.060663484036922455, + 0.0042017786763608456, + 0.00489032082259655, + 0.004971873946487904, + 0.012501491233706474, + 0.0021104204934090376, + 0.003369272919371724, + 0.004332940559834242, + 0.003915699664503336, + 0.00580425513908267, + 0.007422695867717266, + 0.003337927395477891, + 0.005436066538095474, + 0.003618770046159625, + 0.0019371365197002888, + 0.0038477841299027205, + 0.003571456065401435, + 0.007085861172527075, + 0.004312943201512098, + 0.0053884186781942844, + 0.01184277143329382, + 0.010726138018071651, + 0.005805824883282185, + 0.010994290001690388, + 0.005284874700009823, + 0.001602653763256967, + 0.0006600739434361458, + 0.0009670202271081507, + 0.0017283445922657847, + 0.0010916226310655475, + 0.0006346157169900835, + 0.0031222510151565075, + 0.0011034754570573568, + 0.001544311409816146, + 0.0008736391318961978, + 0.0007883724756538868, + 0.0009727997821755707, + 0.0012628204422071576, + 0.0010950472205877304, + 0.0023044089321047068, + 0.0015723543474450707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.11297333985567093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019407500512897968 + ], + "opposite_logprob_tokens": [ + -6.251940727233887 + ], + "logprob_reward": -0.0019407500512897968, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019407500512897968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1631777137517929, + 0.36612755060195923, + 0.06431850790977478, + 0.11387936770915985, + 0.11175614595413208, + 0.06498847901821136, + 0.020921768620610237, + 0.026938170194625854, + 0.01601984351873398, + 0.05187245458364487 + ], + "credit_map": [ + 0.015330196358263493, + 0.029679592698812485, + 0.017479777336120605, + 0.012577464804053307, + 0.03971782326698303, + 0.01911240816116333, + 0.017619458958506584, + 0.011660999618470669, + 0.02175007574260235, + 0.044557124376297, + 0.08766894042491913, + 0.02759185992181301, + 0.019739368930459023, + 0.036739036440849304, + 0.10320163518190384, + 0.02487947978079319, + 0.005890873726457357, + 0.009552424773573875, + 0.008178167976439, + 0.007014766335487366, + 0.005644777324050665, + 0.008027670904994011, + 0.008761968463659286, + 0.011247865855693817, + 0.016604771837592125, + 0.04342774674296379, + 0.00639706663787365, + 0.015388213098049164, + 0.014393532648682594, + 0.005927028600126505, + 0.00468755979090929, + 0.0070534576661884785, + 0.021001510322093964, + 0.04149745777249336, + 0.00625335518270731, + 0.013681325130164623, + 0.008439300581812859, + 0.007288464345037937, + 0.006423673592507839, + 0.007171051111072302, + 0.009016809053719044, + 0.011529449373483658, + 0.0064668599516153336, + 0.015650752931833267, + 0.0038512609899044037, + 0.004378852434456348, + 0.00563135975971818, + 0.008463136851787567, + 0.002281083259731531, + 0.0026550451293587685, + 0.002132657216861844, + 0.002673932584002614, + 0.0017892020987346768, + 0.0018917135894298553, + 0.0026932493783533573, + 0.004804885480552912, + 0.0020581132266670465, + 0.0016447145026177168, + 0.003630704013630748, + 0.004163051024079323, + 0.003942460287362337, + 0.0020974900107830763, + 0.0064830961637198925, + 0.0029185405001044273, + 0.001270862529054284, + 0.0009711830643936992, + 0.0015401104465126991, + 0.002691862406209111, + 0.0028460705652832985, + 0.001262764330022037, + 0.0038417330943048, + 0.0015952567337080836, + 0.006381937302649021, + 0.005316201597452164, + 0.0061662946827709675, + 0.005682299844920635, + 0.008141572587192059, + 0.0063744597136974335, + 0.006790975574404001, + 0.007018715143203735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0019407500512897968 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.142861008644104 + ], + "opposite_logprob_tokens": [ + -2.0178608894348145 + ], + "logprob_reward": -0.142861008644104, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.142861008644104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15145528316497803, + 0.14821307361125946, + 0.02676694095134735, + 0.06485938280820847, + 0.4202524721622467, + 0.03010551445186138, + 0.043974485248327255, + 0.03546195849776268, + 0.011400457471609116, + 0.0675104483962059 + ], + "credit_map": [ + 0.014046507887542248, + 0.021986043080687523, + 0.015594585798680782, + 0.02088085375726223, + 0.03243085369467735, + 0.01943504996597767, + 0.012378172017633915, + 0.014703220687806606, + 0.008471428416669369, + 0.02176460064947605, + 0.03265424072742462, + 0.013034644536674023, + 0.007868540473282337, + 0.015605940483510494, + 0.036992765963077545, + 0.0118208983913064, + 0.003526128362864256, + 0.004431300330907106, + 0.004015851300209761, + 0.003361854702234268, + 0.002673868555575609, + 0.0026055395137518644, + 0.0031016685534268618, + 0.003050729399546981, + 0.008348485454916954, + 0.02396468259394169, + 0.0034119144547730684, + 0.010271333158016205, + 0.007175495382398367, + 0.003494761884212494, + 0.003323183162137866, + 0.004869528580456972, + 0.03073911741375923, + 0.11645961552858353, + 0.03777754679322243, + 0.1546306163072586, + 0.01392220240086317, + 0.007401200011372566, + 0.0209655724465847, + 0.03835662826895714, + 0.0029693192336708307, + 0.0020223772153258324, + 0.00400184653699398, + 0.006865406408905983, + 0.002011895412579179, + 0.003294158959761262, + 0.00394828338176012, + 0.00499222707003355, + 0.005206898786127567, + 0.00547116668894887, + 0.0036026285961270332, + 0.005072219762951136, + 0.007099303882569075, + 0.0054961685091257095, + 0.007413655053824186, + 0.004612443968653679, + 0.003993422258645296, + 0.00277522811666131, + 0.0028955943416804075, + 0.0052271350286901, + 0.006351014133542776, + 0.004467304330319166, + 0.006753688212484121, + 0.0029985676519572735, + 0.00157461641356349, + 0.0007921470678411424, + 0.0008798488997854292, + 0.002296874299645424, + 0.00136627780739218, + 0.001111217774450779, + 0.0021307480055838823, + 0.001248727086931467, + 0.0093626007437706, + 0.005511207971721888, + 0.00562990503385663, + 0.007989833131432533, + 0.009144765324890614, + 0.006116201169788837, + 0.015086542814970016, + 0.008669386617839336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.142861008644104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style a stop-motion clay animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan leftward with a slight upward tilt in a wide-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive pale yellow clay moon visible in a deep violet sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the clay cacti display visible hand-kneading fingerprints and tool marks on their surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a small rough-surfaced sandy clay camel trudge into the frame from the right side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895860075950623 + ], + "opposite_logprob_tokens": [ + -2.578958511352539 + ], + "logprob_reward": -0.07895860075950623, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895860075950623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10084013640880585, + 0.12884053587913513, + 0.021080531179904938, + 0.049960315227508545, + 0.4599507749080658, + 0.03587592765688896, + 0.038485463708639145, + 0.043176230043172836, + 0.10205404460430145, + 0.019736047834157944 + ], + "credit_map": [ + 0.009046019054949284, + 0.015799878165125847, + 0.010046407580375671, + 0.008028565905988216, + 0.03305615112185478, + 0.007535747718065977, + 0.009506717324256897, + 0.007820657454431057, + 0.008023528382182121, + 0.009637265466153622, + 0.019317317754030228, + 0.0072435676120221615, + 0.013669699430465698, + 0.02227979525923729, + 0.02570289559662342, + 0.022966474294662476, + 0.0020755676086992025, + 0.002969134133309126, + 0.002740908181294799, + 0.0029035250190645456, + 0.0022314568050205708, + 0.0022095346357673407, + 0.00348785356618464, + 0.002462550764903426, + 0.002948877401649952, + 0.005176437087357044, + 0.0043505290523171425, + 0.005857400596141815, + 0.005664620082825422, + 0.008801289834082127, + 0.007178537081927061, + 0.009982625022530556, + 0.046010371297597885, + 0.14355076849460602, + 0.03680843114852905, + 0.15362422168254852, + 0.017620712518692017, + 0.006854035425931215, + 0.0207073874771595, + 0.034774843603372574, + 0.002951559843495488, + 0.0030424022115767, + 0.005114572588354349, + 0.006091210059821606, + 0.0035624082665890455, + 0.005214973818510771, + 0.004711543209850788, + 0.005187259055674076, + 0.006032872013747692, + 0.00652748066931963, + 0.0034079020842909813, + 0.004641910549253225, + 0.0038930196315050125, + 0.0030982934404164553, + 0.006401010788977146, + 0.004482975695282221, + 0.003694609971717, + 0.0031728700269013643, + 0.0036075222305953503, + 0.006174159701913595, + 0.00739252520725131, + 0.0038392182905226946, + 0.009858128614723682, + 0.0054371957667171955, + 0.007691128645092249, + 0.006101617123931646, + 0.007844609208405018, + 0.0175034012645483, + 0.01602151244878769, + 0.005718862172216177, + 0.032249439507722855, + 0.008923472836613655, + 0.0033363099209964275, + 0.0011869075242429972, + 0.002058178186416626, + 0.002261939225718379, + 0.0017106950981542468, + 0.0012878271518275142, + 0.005705089308321476, + 0.002189100021496415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.07895860075950623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_10/index_tensor([5036, 5036])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_10/index_tensor([5036, 5036])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..85b3c34f636fb1faa9f9e51330b607d607cb6a8a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_10/index_tensor([5036, 5036])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the refined style of Pixar cartoons, the camera offers a steady, majestic crane shot, beginning with a low-angle close-up of the joyful splashing at the base of a vibrantly colored fountain. This whimsical fountain, shaped like a giant, smiling sunflower with glossy, saturated yellow petals, shoots crystal-clear water high into the air, where it then cascades down in a series of shimmering, sun-drenched droplets. As the camera steadily ascends, the plump, cheerful fountain is revealed, then the perspective widens to reveal a sprawling, whimsical water park beneath a brilliant, cloudless blue sky. The scene is a vibrant tapestry of color and movement: giant, intertwining water slides, shimmering in shades of candy red, electric blue, and lime green, twist and loop around each other, their rounded, polished surfaces reflecting the sunlight. Below, a variety of organically shaped, shimmering cyan pools are filled with stylized cartoon characters splashing and playing. As the camera peaks into a breathtaking high-angle wide shot, a magnificent, vibrant rainbow, born from the fine mist hanging in the air, arcs perfectly across the scene, its unique, saturated band of color stretching from one end of the park to the other, framing the winding slides and sparkling pools in a magical glow and creating a breathtaking atmosphere of pure, vibrant joy and the wonder of summer.", + "seed": 24381, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_10/index_5036_seed_24381_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03134244754910469, + "vlm_logprob": -0.03134244754910469, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03134244754910469, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013270976953208447 + ], + "opposite_logprob_tokens": [ + -4.388270854949951 + ], + "logprob_reward": -0.013270976953208447, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013270976953208447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10514696687459946, + 0.04077679663896561, + 0.049009162932634354, + 0.3119964897632599, + 0.11029183119535446, + 0.04878488928079605, + 0.12747792899608612, + 0.11724653095006943, + 0.0419003926217556, + 0.04736893251538277 + ], + "credit_map": [ + 0.008470160886645317, + 0.012405529618263245, + 0.008583948016166687, + 0.011137382127344608, + 0.012289776466786861, + 0.012407692149281502, + 0.01803322322666645, + 0.021819254383444786, + 0.004328608978539705, + 0.006033003330230713, + 0.004385577514767647, + 0.007081389892846346, + 0.004732360597699881, + 0.005120808724313974, + 0.0046448237262666225, + 0.0044502210803329945, + 0.006751051172614098, + 0.0064576854929327965, + 0.00464943703263998, + 0.006104026455432177, + 0.006499821320176125, + 0.006294558756053448, + 0.005163917783647776, + 0.007088663522154093, + 0.04319872707128525, + 0.11121288686990738, + 0.018352221697568893, + 0.029447345063090324, + 0.03778519481420517, + 0.029221242293715477, + 0.02435934916138649, + 0.018419509753584862, + 0.013258752413094044, + 0.021459588780999184, + 0.007931808941066265, + 0.012120340950787067, + 0.026059124618768692, + 0.00758814811706543, + 0.009728522971272469, + 0.012145538814365864, + 0.004624661989510059, + 0.010605651885271072, + 0.003489329246804118, + 0.005277401302009821, + 0.01005184929817915, + 0.006149784196168184, + 0.004364755470305681, + 0.004221460781991482, + 0.009988264180719852, + 0.016487225890159607, + 0.019589738920331, + 0.01493253093212843, + 0.018228653818368912, + 0.01872101053595543, + 0.014756588265299797, + 0.014773910865187645, + 0.006811376195400953, + 0.011851555667817593, + 0.009060917422175407, + 0.014914150349795818, + 0.015353032387793064, + 0.01556478999555111, + 0.01624962128698826, + 0.02744108997285366, + 0.0034642296377569437, + 0.0047679501585662365, + 0.005689850077033043, + 0.004340194631367922, + 0.005188527517020702, + 0.005056467372924089, + 0.008411170914769173, + 0.004982001148164272, + 0.0037371942307800055, + 0.005607676226645708, + 0.008996869437396526, + 0.0070481873117387295, + 0.004918083548545837, + 0.0027215357404202223, + 0.009037179872393608, + 0.005302205681800842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.013270976953208447 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059619685634970665 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.0059619685634970665, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059619685634970665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18886950612068176, + 0.07982191443443298, + 0.08187174052000046, + 0.24431203305721283, + 0.08478506654500961, + 0.07319740951061249, + 0.04011239856481552, + 0.07530826330184937, + 0.058023735880851746, + 0.07369798421859741 + ], + "credit_map": [ + 0.01563415862619877, + 0.01821695640683174, + 0.015755100175738335, + 0.01898408681154251, + 0.018424177542328835, + 0.021677441895008087, + 0.03906645625829697, + 0.041111093014478683, + 0.00931274052709341, + 0.009267639368772507, + 0.008158735930919647, + 0.011422696523368359, + 0.008515986613929272, + 0.008373701013624668, + 0.015415272675454617, + 0.009355138055980206, + 0.007862454280257225, + 0.01287388801574707, + 0.011697815731167793, + 0.014241146855056286, + 0.00876368023455143, + 0.010935661382973194, + 0.009333324618637562, + 0.006163759622722864, + 0.033499959856271744, + 0.08608142286539078, + 0.013818222098052502, + 0.022487858310341835, + 0.027712779119610786, + 0.02350449562072754, + 0.020843131467700005, + 0.016364142298698425, + 0.00900438241660595, + 0.014566739089787006, + 0.005774116143584251, + 0.01072616782039404, + 0.019116969779133797, + 0.00608087470754981, + 0.01085820235311985, + 0.008657609112560749, + 0.007381170056760311, + 0.01653354987502098, + 0.00870551448315382, + 0.007784334011375904, + 0.009753596968948841, + 0.008023870177567005, + 0.005862864200025797, + 0.009152501821517944, + 0.004547925665974617, + 0.008607480674982071, + 0.0025502091739326715, + 0.00468760821968317, + 0.006625018548220396, + 0.004477484151721001, + 0.005051904823631048, + 0.0035647633485496044, + 0.0038728041108697653, + 0.012286268174648285, + 0.007620635908097029, + 0.007872527465224266, + 0.009444608353078365, + 0.00845797173678875, + 0.01179674081504345, + 0.01395670510828495, + 0.004348840098828077, + 0.006204594857990742, + 0.006690084934234619, + 0.005424264818429947, + 0.00815449096262455, + 0.009478938765823841, + 0.011306301690638065, + 0.006416213698685169, + 0.005480668041855097, + 0.009187965653836727, + 0.009946904145181179, + 0.009433646686375141, + 0.007444681134074926, + 0.004422925412654877, + 0.01821112632751465, + 0.009570064023137093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0059619685634970665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494876950979233 + ], + "opposite_logprob_tokens": [ + -2.9299488067626953 + ], + "logprob_reward": -0.05494876950979233, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494876950979233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2212938815355301, + 0.08583246171474457, + 0.1580495834350586, + 0.07432103157043457, + 0.08701012283563614, + 0.11186031997203827, + 0.14330367743968964, + 0.049718327820301056, + 0.02925303764641285, + 0.03935754671692848 + ], + "credit_map": [ + 0.025129836052656174, + 0.030183563008904457, + 0.027552176266908646, + 0.02362871542572975, + 0.019334107637405396, + 0.022506091743707657, + 0.04033467546105385, + 0.03262472152709961, + 0.013056016527116299, + 0.0114042479544878, + 0.009046602062880993, + 0.011137913912534714, + 0.0102243572473526, + 0.008443105965852737, + 0.010727732442319393, + 0.011792482808232307, + 0.025248251855373383, + 0.023786041885614395, + 0.024754049256443977, + 0.022459495812654495, + 0.017435891553759575, + 0.016685379669070244, + 0.01416194811463356, + 0.013518516905605793, + 0.008968044072389603, + 0.02107846736907959, + 0.0059823766350746155, + 0.008216094225645065, + 0.010006706230342388, + 0.008465743623673916, + 0.005481358151882887, + 0.006122243590652943, + 0.010326970368623734, + 0.013337386772036552, + 0.007174276746809483, + 0.011226239614188671, + 0.017900271341204643, + 0.0072685726918280125, + 0.00962992012500763, + 0.010146488435566425, + 0.011990192346274853, + 0.027785154059529305, + 0.012119309976696968, + 0.013894177041947842, + 0.015952542424201965, + 0.008910576812922955, + 0.012378714047372341, + 0.008829646743834019, + 0.01647098921239376, + 0.023416204378008842, + 0.0177919864654541, + 0.013838360086083412, + 0.01941605657339096, + 0.019104303792119026, + 0.018826816231012344, + 0.014438952319324017, + 0.00327887712046504, + 0.0071723745204508305, + 0.004648908041417599, + 0.006043161265552044, + 0.007102256640791893, + 0.005754274316132069, + 0.00887211598455906, + 0.006846360396593809, + 0.0024942238815128803, + 0.003875478869304061, + 0.0038335451390594244, + 0.0023677386343479156, + 0.004708650056272745, + 0.003989186603575945, + 0.004737406503409147, + 0.0032468088902533054, + 0.00420448137447238, + 0.0028462279587984085, + 0.006216154433786869, + 0.006249322555959225, + 0.002809729892760515, + 0.002804381540045142, + 0.008985978551208973, + 0.005241269711405039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.05494876950979233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533823028206825 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533823028206825, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533823028206825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.114521324634552, + 0.05630600079894066, + 0.06059694662690163, + 0.10379588603973389, + 0.03247198462486267, + 0.05984624847769737, + 0.10171788930892944, + 0.08603767305612564, + 0.11303265392780304, + 0.27167338132858276 + ], + "credit_map": [ + 0.01127612218260765, + 0.010930796153843403, + 0.010879883542656898, + 0.009840408340096474, + 0.009903828613460064, + 0.01337885670363903, + 0.025125352665781975, + 0.023186080157756805, + 0.004882920067757368, + 0.005702786147594452, + 0.007246873341500759, + 0.01075790449976921, + 0.008420523256063461, + 0.007569071836769581, + 0.006104699335992336, + 0.005621220450848341, + 0.006100676488131285, + 0.00837293453514576, + 0.005585670471191406, + 0.009712640196084976, + 0.009259640239179134, + 0.008824706077575684, + 0.007230829447507858, + 0.005509847309440374, + 0.014701148495078087, + 0.034933239221572876, + 0.006381328217685223, + 0.00960832554847002, + 0.01085480023175478, + 0.008761945180594921, + 0.011482283473014832, + 0.007072816602885723, + 0.0039780051447451115, + 0.00711709912866354, + 0.003440453903749585, + 0.004423741716891527, + 0.00551613699644804, + 0.0022510283160954714, + 0.0037351723294705153, + 0.0020103463903069496, + 0.006437309551984072, + 0.013085843995213509, + 0.006951502989977598, + 0.006000331602990627, + 0.008368423208594322, + 0.009188207797706127, + 0.0046903714537620544, + 0.005124255083501339, + 0.01181032881140709, + 0.01734277233481407, + 0.00998743437230587, + 0.012180901132524014, + 0.015027696266770363, + 0.00957825593650341, + 0.01638411544263363, + 0.009406383149325848, + 0.008349470794200897, + 0.010096224956214428, + 0.0067921169102191925, + 0.010135596618056297, + 0.015482751652598381, + 0.00835467129945755, + 0.01521284319460392, + 0.011613995768129826, + 0.009300134144723415, + 0.009428421966731548, + 0.013721263967454433, + 0.010240375064313412, + 0.013881877064704895, + 0.01239654328674078, + 0.029202312231063843, + 0.014861728064715862, + 0.02500416338443756, + 0.024276984855532646, + 0.045092351734638214, + 0.0321093425154686, + 0.024532536044716835, + 0.02053760550916195, + 0.05283655226230621, + 0.047283854335546494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.012533823028206825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999669969081879 + ], + "opposite_logprob_tokens": [ + -2.6949965953826904 + ], + "logprob_reward": -0.06999669969081879, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999669969081879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1367921680212021, + 0.058066219091415405, + 0.05538509786128998, + 0.17791059613227844, + 0.08694760501384735, + 0.0994938462972641, + 0.1296369582414627, + 0.07646095007658005, + 0.06393083930015564, + 0.11537571996450424 + ], + "credit_map": [ + 0.0136825330555439, + 0.015656907111406326, + 0.015607033856213093, + 0.01571046933531761, + 0.01563935913145542, + 0.015804914757609367, + 0.023186413571238518, + 0.021504545584321022, + 0.007203494664281607, + 0.007136660628020763, + 0.007504855748265982, + 0.0075909728184342384, + 0.005726680625230074, + 0.006470296997576952, + 0.00802386924624443, + 0.008409393951296806, + 0.006766756530851126, + 0.006318299099802971, + 0.006493723019957542, + 0.0072711799293756485, + 0.005493637640029192, + 0.007850956171751022, + 0.005611802916973829, + 0.009578744880855083, + 0.020135724917054176, + 0.054940365254879, + 0.013850273564457893, + 0.01759464293718338, + 0.027793822810053825, + 0.014060229994356632, + 0.012166154570877552, + 0.017369389533996582, + 0.008961096405982971, + 0.015240035019814968, + 0.007829711772501469, + 0.011074324138462543, + 0.016510363668203354, + 0.006906048860400915, + 0.009794939309358597, + 0.010631095618009567, + 0.009664056822657585, + 0.014854292385280132, + 0.012710778042674065, + 0.010299955494701862, + 0.017269808799028397, + 0.014139180071651936, + 0.009396539069712162, + 0.011159240268170834, + 0.011834979988634586, + 0.023553218692541122, + 0.010382554493844509, + 0.020464245229959488, + 0.024244647473096848, + 0.011356168426573277, + 0.01860845647752285, + 0.009192701429128647, + 0.006991637405008078, + 0.01017241645604372, + 0.008126267232000828, + 0.009688174352049828, + 0.010002879425883293, + 0.0091550936922431, + 0.012663688510656357, + 0.00966079905629158, + 0.0052506388165056705, + 0.007849840447306633, + 0.008136230520904064, + 0.005715993698686361, + 0.011140097863972187, + 0.006890373304486275, + 0.01065840944647789, + 0.008289261721074581, + 0.010775134898722172, + 0.01067327056080103, + 0.021066220477223396, + 0.016000034287571907, + 0.00989292562007904, + 0.008457669988274574, + 0.02100662887096405, + 0.017503853887319565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.06999669969081879 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the refined style of Pixar cartoons, the camera offers a steady, majestic crane shot, beginning with a low-angle close-up of the joyful splashing at the base of a vibrantly colored fountain. This whimsical fountain, shaped like a giant, smiling sunflower with glossy, saturated yellow petals, shoots crystal-clear water high into the air, where it then cascades down in a series of shimmering, sun-drenched droplets. As the camera steadily ascends, the plump, cheerful fountain is revealed, then the perspective widens to reveal a sprawling, whimsical water park beneath a brilliant, cloudless blue sky. The scene is a vibrant tapestry of color and movement: giant, intertwining water slides, shimmering in shades of candy red, electric blue, and lime green, twist and loop around each other, their rounded, polished surfaces reflecting the sunlight. Below, a variety of organically shaped, shimmering cyan pools are filled with stylized cartoon characters splashing and playing. As the camera peaks into a breathtaking high-angle wide shot, a magnificent, vibrant rainbow, born from the fine mist hanging in the air, arcs perfectly across the scene, its unique, saturated band of color stretching from one end of the park to the other, framing the winding slides and sparkling pools in a magical glow and creating a breathtaking atmosphere of pure, vibrant joy and the wonder of summer.", + "seed": 24382, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_10/index_5036_seed_24382_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.009396763285622, + "vlm_logprob": -0.009396763285622, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.009396763285622, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004982080310583115 + ], + "opposite_logprob_tokens": [ + -5.379981994628906 + ], + "logprob_reward": -0.004982080310583115, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004982080310583115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20690886676311493, + 0.10968675464391708, + 0.05963043123483658, + 0.11573908478021622, + 0.028272313997149467, + 0.10537490248680115, + 0.1370706707239151, + 0.05105288326740265, + 0.07837823778390884, + 0.10788589715957642 + ], + "credit_map": [ + 0.02712133154273033, + 0.022026492282748222, + 0.03796207159757614, + 0.01597084477543831, + 0.03029192052781582, + 0.029374433681368828, + 0.025663621723651886, + 0.018498139455914497, + 0.010256174951791763, + 0.014903198927640915, + 0.01171275693923235, + 0.011611072346568108, + 0.025059057399630547, + 0.009865924715995789, + 0.012735805474221706, + 0.01354276668280363, + 0.004239524714648724, + 0.007196499966084957, + 0.008751893416047096, + 0.010956685990095139, + 0.01109944935888052, + 0.003727226983755827, + 0.00872576329857111, + 0.0049333879724144936, + 0.0174794252961874, + 0.02113230898976326, + 0.011577162891626358, + 0.013420987874269485, + 0.01327429711818695, + 0.013840759173035622, + 0.013909614644944668, + 0.0111045241355896, + 0.0036502506118267775, + 0.0026333159767091274, + 0.0040541780181229115, + 0.002976212417706847, + 0.00333396065980196, + 0.0031754791270941496, + 0.003296182258054614, + 0.005152735393494368, + 0.01088319718837738, + 0.03548440337181091, + 0.023608416318893433, + 0.006515198387205601, + 0.011576036922633648, + 0.003249840810894966, + 0.0062606967985630035, + 0.007797109428793192, + 0.01883523538708687, + 0.009853733703494072, + 0.017907757312059402, + 0.016537100076675415, + 0.0166111271828413, + 0.02173887938261032, + 0.014728485606610775, + 0.02085835300385952, + 0.006528580095618963, + 0.006067621521651745, + 0.006547761149704456, + 0.005662898067384958, + 0.009479888714849949, + 0.005368100944906473, + 0.006505303084850311, + 0.004892727360129356, + 0.00997542217373848, + 0.009866525419056416, + 0.008451094850897789, + 0.0069536129012703896, + 0.017234792932868004, + 0.008074034005403519, + 0.008145784959197044, + 0.00967696774750948, + 0.02068493887782097, + 0.008666112087666988, + 0.010058699175715446, + 0.012599644251167774, + 0.01871938817203045, + 0.011600018478929996, + 0.015429298393428326, + 0.010127806104719639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.004982080310583115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036235882434993982 + ], + "opposite_logprob_tokens": [ + -5.6286234855651855 + ], + "logprob_reward": -0.0036235882434993982, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036235882434993982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16136501729488373, + 0.04552968591451645, + 0.07140651345252991, + 0.07696881890296936, + 0.09104140102863312, + 0.1415814310312271, + 0.17668414115905762, + 0.07068253308534622, + 0.09601303935050964, + 0.06872741878032684 + ], + "credit_map": [ + 0.029928160831332207, + 0.018900195136666298, + 0.0190237108618021, + 0.02251867763698101, + 0.02548731118440628, + 0.015354850329458714, + 0.018180865794420242, + 0.011971259489655495, + 0.005419139284640551, + 0.006005409173667431, + 0.004398319870233536, + 0.007302049547433853, + 0.008281339891254902, + 0.003969389479607344, + 0.006128792185336351, + 0.004025246016681194, + 0.007772970013320446, + 0.007696510758250952, + 0.014567825011909008, + 0.007353451102972031, + 0.009741408750414848, + 0.00820557028055191, + 0.006458749528974295, + 0.009610028937458992, + 0.011626015417277813, + 0.00986248254776001, + 0.01419707853347063, + 0.006335940212011337, + 0.0077821239829063416, + 0.010859214700758457, + 0.0072385468520224094, + 0.009067418053746223, + 0.007614245638251305, + 0.00841444730758667, + 0.012996150180697441, + 0.00850613322108984, + 0.015535027720034122, + 0.007039644755423069, + 0.007322594989091158, + 0.023613156750798225, + 0.009418702684342861, + 0.02562945894896984, + 0.028628122061491013, + 0.019186094403266907, + 0.020051850005984306, + 0.004780254326760769, + 0.02293647639453411, + 0.01095047127455473, + 0.01914423331618309, + 0.013872778974473476, + 0.024478795006871223, + 0.03478652611374855, + 0.017050422728061676, + 0.02085796184837818, + 0.023531079292297363, + 0.022962339222431183, + 0.011327371932566166, + 0.009794019162654877, + 0.010061788372695446, + 0.006376200821250677, + 0.009888950735330582, + 0.0072538279928267, + 0.007915211841464043, + 0.00806516408920288, + 0.022476201876997948, + 0.01276868674904108, + 0.016994666308164597, + 0.010900569148361683, + 0.008082878775894642, + 0.009240307845175266, + 0.007116950117051601, + 0.008432776667177677, + 0.011782760731875896, + 0.008786335587501526, + 0.011918793432414532, + 0.008101600222289562, + 0.007589375134557486, + 0.00891308393329382, + 0.006303602363914251, + 0.005331864580512047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0036235882434993982 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107768900692463 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107768900692463, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107768900692463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1038137748837471, + 0.03110620565712452, + 0.034213777631521225, + 0.05454121530056, + 0.06756095588207245, + 0.3152998387813568, + 0.19652511179447174, + 0.03810328617691994, + 0.08968467265367508, + 0.0691511482000351 + ], + "credit_map": [ + 0.021492887288331985, + 0.01092827133834362, + 0.015642531216144562, + 0.008856900967657566, + 0.01412943471223116, + 0.01062543410807848, + 0.0143973957747221, + 0.00774091761559248, + 0.0035080271773040295, + 0.0036728456616401672, + 0.0032211849465966225, + 0.004861477762460709, + 0.005535640753805637, + 0.0033637576270848513, + 0.004504001699388027, + 0.002439269097521901, + 0.0031095854938030243, + 0.0042168027721345425, + 0.006090778391808271, + 0.0036117504350841045, + 0.006148484535515308, + 0.0032381447963416576, + 0.003658191068097949, + 0.004140041768550873, + 0.00647418387234211, + 0.0066235908307135105, + 0.010993406176567078, + 0.005879322532564402, + 0.006534950342029333, + 0.006273537874221802, + 0.0049948133528232574, + 0.006767410319298506, + 0.006753185763955116, + 0.006607783026993275, + 0.0095441285520792, + 0.0049991607666015625, + 0.012348951771855354, + 0.005470745265483856, + 0.0052499594166874886, + 0.016587041318416595, + 0.02987760119140148, + 0.10555096715688705, + 0.08272958546876907, + 0.017561662942171097, + 0.03399483114480972, + 0.009506162256002426, + 0.012453108094632626, + 0.023625923320651054, + 0.0189671628177166, + 0.019497783854603767, + 0.023643435910344124, + 0.0457223579287529, + 0.015523720532655716, + 0.02187604270875454, + 0.030012039467692375, + 0.021282562986016273, + 0.005905694328248501, + 0.004864140413701534, + 0.005044484045356512, + 0.0052702706307172775, + 0.004355766344815493, + 0.004176792223006487, + 0.004339986015111208, + 0.004146149381995201, + 0.01561844814568758, + 0.018995173275470734, + 0.008458029478788376, + 0.007294944021850824, + 0.015211550518870354, + 0.007446515839546919, + 0.007678702939301729, + 0.008981313556432724, + 0.011464182287454605, + 0.007639559917151928, + 0.009786977432668209, + 0.007507874630391598, + 0.010511217638850212, + 0.007146137300878763, + 0.00699538504704833, + 0.008099813014268875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.01107768900692463 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067322817631065845 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.0067322817631065845, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067322817631065845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07636724412441254, + 0.036060672253370285, + 0.06379400938749313, + 0.05563817173242569, + 0.028862321749329567, + 0.1178460568189621, + 0.07775598764419556, + 0.16758528351783752, + 0.1715134084224701, + 0.20457684993743896 + ], + "credit_map": [ + 0.012294414453208447, + 0.007406814023852348, + 0.01066205371171236, + 0.006633438635617495, + 0.009824142791330814, + 0.012408494018018246, + 0.011077282950282097, + 0.006060598883777857, + 0.004185903817415237, + 0.00453035207465291, + 0.004063976462930441, + 0.004205609206110239, + 0.008745137602090836, + 0.0028748775366693735, + 0.0038710436783730984, + 0.0035837700124830008, + 0.007013767026364803, + 0.005579877644777298, + 0.017054958269000053, + 0.0065601482056081295, + 0.006546800956130028, + 0.004969038534909487, + 0.005847862455993891, + 0.01022155862301588, + 0.006958676036447287, + 0.0069763073697686195, + 0.010051023215055466, + 0.006186800543218851, + 0.005928133148699999, + 0.005365646909922361, + 0.007972555235028267, + 0.006199033465236425, + 0.0033241130877286196, + 0.002493076492100954, + 0.0046990481205284595, + 0.0035239378921687603, + 0.00324633507989347, + 0.0035508854780346155, + 0.003402125556021929, + 0.004622799344360828, + 0.01076478511095047, + 0.037114016711711884, + 0.0303511843085289, + 0.007907130755484104, + 0.012608439661562443, + 0.00399999413639307, + 0.006664457265287638, + 0.008436045609414577, + 0.008678755722939968, + 0.006222344469279051, + 0.008085625246167183, + 0.013302130624651909, + 0.008589645847678185, + 0.009336556307971478, + 0.010598022490739822, + 0.012942906469106674, + 0.022769082337617874, + 0.01937880553305149, + 0.014640618115663528, + 0.01936349645256996, + 0.033209428191185, + 0.018776683136820793, + 0.021932998672127724, + 0.017514163628220558, + 0.04570823162794113, + 0.02467106282711029, + 0.024833733215928078, + 0.015688875690102577, + 0.0184123944491148, + 0.01886039413511753, + 0.009789343923330307, + 0.013549359515309334, + 0.029773108661174774, + 0.024339886382222176, + 0.03418540209531784, + 0.030071431770920753, + 0.01603240892291069, + 0.021711142733693123, + 0.02569783851504326, + 0.0227656289935112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0067322817631065845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020568177103996277 + ], + "opposite_logprob_tokens": [ + -3.8955681324005127 + ], + "logprob_reward": -0.020568177103996277, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020568177103996277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2191341668367386, + 0.12557779252529144, + 0.05911071598529816, + 0.05106145143508911, + 0.031317971646785736, + 0.04473203048110008, + 0.12058892101049423, + 0.1189049631357193, + 0.11715640127658844, + 0.11241570860147476 + ], + "credit_map": [ + 0.034651245921850204, + 0.02596867084503174, + 0.02405071258544922, + 0.025157487019896507, + 0.04130471870303154, + 0.025305375456809998, + 0.02422870323061943, + 0.018467223271727562, + 0.01191441435366869, + 0.017428532242774963, + 0.012541331350803375, + 0.012009922415018082, + 0.03451695293188095, + 0.010232307016849518, + 0.013461027294397354, + 0.01347329281270504, + 0.005712847225368023, + 0.0073271580040454865, + 0.01321894396096468, + 0.006275970954447985, + 0.008364276960492134, + 0.005185066256672144, + 0.005062740761786699, + 0.007963703013956547, + 0.0059542106464505196, + 0.008717714808881283, + 0.00790298730134964, + 0.005689759273082018, + 0.005907407496124506, + 0.0037690415047109127, + 0.007634674198925495, + 0.005485652480274439, + 0.0031906988006085157, + 0.0025318744592368603, + 0.0054317982867360115, + 0.004684765357524157, + 0.005026084370911121, + 0.0035856410395354033, + 0.0032384595833718777, + 0.0036286446265876293, + 0.004467505030333996, + 0.012664388865232468, + 0.0065374672412872314, + 0.005108391400426626, + 0.0050038439221680164, + 0.002397395670413971, + 0.00535142095759511, + 0.0032016122713685036, + 0.01750991865992546, + 0.011235738173127174, + 0.016078466549515724, + 0.019965726882219315, + 0.015224926173686981, + 0.014546852558851242, + 0.013202962465584278, + 0.01282432395964861, + 0.01967909373342991, + 0.019248146563768387, + 0.013351373374462128, + 0.014666256494820118, + 0.012495533563196659, + 0.013011696748435497, + 0.013189668767154217, + 0.013263172470033169, + 0.01957353763282299, + 0.014878123067319393, + 0.017940189689397812, + 0.013608250766992569, + 0.013795599341392517, + 0.014658521860837936, + 0.010363047942519188, + 0.01233911793678999, + 0.018551049754023552, + 0.013215715065598488, + 0.014122202061116695, + 0.01696278341114521, + 0.01327368151396513, + 0.010102363303303719, + 0.014748146757483482, + 0.011439756490290165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.020568177103996277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the refined style of Pixar cartoons, the camera offers a steady, majestic crane shot, beginning with a low-angle close-up of the joyful splashing at the base of a vibrantly colored fountain. This whimsical fountain, shaped like a giant, smiling sunflower with glossy, saturated yellow petals, shoots crystal-clear water high into the air, where it then cascades down in a series of shimmering, sun-drenched droplets. As the camera steadily ascends, the plump, cheerful fountain is revealed, then the perspective widens to reveal a sprawling, whimsical water park beneath a brilliant, cloudless blue sky. The scene is a vibrant tapestry of color and movement: giant, intertwining water slides, shimmering in shades of candy red, electric blue, and lime green, twist and loop around each other, their rounded, polished surfaces reflecting the sunlight. Below, a variety of organically shaped, shimmering cyan pools are filled with stylized cartoon characters splashing and playing. As the camera peaks into a breathtaking high-angle wide shot, a magnificent, vibrant rainbow, born from the fine mist hanging in the air, arcs perfectly across the scene, its unique, saturated band of color stretching from one end of the park to the other, framing the winding slides and sparkling pools in a magical glow and creating a breathtaking atmosphere of pure, vibrant joy and the wonder of summer.", + "seed": 24383, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_10/index_5036_seed_24383_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01933893719688058, + "vlm_logprob": -0.01933893719688058, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.01933893719688058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007130649872124195 + ], + "opposite_logprob_tokens": [ + -5.0071306228637695 + ], + "logprob_reward": -0.007130649872124195, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.007130649872124195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5115938782691956, + 0.057202186435461044, + 0.061079055070877075, + 0.044453468173742294, + 0.12712541222572327, + 0.044045478105545044, + 0.03339747339487076, + 0.03672849014401436, + 0.026409506797790527, + 0.05796511471271515 + ], + "credit_map": [ + 0.09805817902088165, + 0.08369345217943192, + 0.029625488445162773, + 0.14408937096595764, + 0.06061062961816788, + 0.036117348819971085, + 0.03457963466644287, + 0.024819787591695786, + 0.007685910910367966, + 0.008301902562379837, + 0.0040081036277115345, + 0.00536012090742588, + 0.008449241518974304, + 0.010909640230238438, + 0.005100570619106293, + 0.007386698387563229, + 0.007841631770133972, + 0.00942913070321083, + 0.003594897920265794, + 0.005773282144218683, + 0.006608978845179081, + 0.006070435978472233, + 0.012653292156755924, + 0.009107406251132488, + 0.005765235982835293, + 0.0064348503947257996, + 0.00347030907869339, + 0.003994537517428398, + 0.005427563562989235, + 0.007594839669764042, + 0.007177100516855717, + 0.004589029587805271, + 0.012507057748734951, + 0.017431646585464478, + 0.022149676457047462, + 0.010717453435063362, + 0.02363813854753971, + 0.013790866360068321, + 0.013940496370196342, + 0.012950070202350616, + 0.00638930220156908, + 0.008893989026546478, + 0.004767526872456074, + 0.002602409338578582, + 0.005586621351540089, + 0.004972145892679691, + 0.0065411352552473545, + 0.004292345605790615, + 0.0045374915935099125, + 0.004393486771732569, + 0.004857523366808891, + 0.005444606300443411, + 0.004273482598364353, + 0.0025846578646451235, + 0.004017637111246586, + 0.0032885868567973375, + 0.003836429212242365, + 0.005027200561016798, + 0.005639268551021814, + 0.006941193249076605, + 0.0036230445839464664, + 0.00264060590416193, + 0.006683298386633396, + 0.0023374471347779036, + 0.0032621161080896854, + 0.003442540764808655, + 0.004434525500983, + 0.004578765016049147, + 0.003158899489790201, + 0.0018680153880268335, + 0.003519803751260042, + 0.0021448389161378145, + 0.011322925798594952, + 0.012646608985960484, + 0.007934077642858028, + 0.009106176905333996, + 0.004336627200245857, + 0.0024476395919919014, + 0.006405474618077278, + 0.0037655846681445837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.007130649872124195 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004103098995983601 + ], + "opposite_logprob_tokens": [ + -5.504103183746338 + ], + "logprob_reward": -0.004103098995983601, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004103098995983601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5925677418708801, + 0.13205291330814362, + 0.029732752591371536, + 0.03827431797981262, + 0.04740853235125542, + 0.032046832144260406, + 0.052628591656684875, + 0.02103595621883869, + 0.025886638090014458, + 0.028365664184093475 + ], + "credit_map": [ + 0.129977285861969, + 0.0801621749997139, + 0.03626047074794769, + 0.20527926087379456, + 0.04775844141840935, + 0.03120335191488266, + 0.035741087049245834, + 0.026185741648077965, + 0.01142931543290615, + 0.014868714846670628, + 0.008301680907607079, + 0.011166203767061234, + 0.016418784856796265, + 0.03425232693552971, + 0.017644234001636505, + 0.017971651628613472, + 0.0046808659099042416, + 0.0037443882320076227, + 0.003550595371052623, + 0.003000565804541111, + 0.002842699410393834, + 0.004498488735407591, + 0.002237294800579548, + 0.0051778568886220455, + 0.005797840654850006, + 0.00510451290756464, + 0.003349172417074442, + 0.004249044694006443, + 0.004532867111265659, + 0.00481061777099967, + 0.0062459250912070274, + 0.004184338264167309, + 0.004983709193766117, + 0.01537646260112524, + 0.005369612015783787, + 0.0030614412389695644, + 0.004435171373188496, + 0.005910263862460852, + 0.003041474847123027, + 0.005230403505265713, + 0.004791954066604376, + 0.004553257953375578, + 0.004278341308236122, + 0.0033678358886390924, + 0.005084912292659283, + 0.0029873945750296116, + 0.003217258956283331, + 0.0037658773362636566, + 0.007604572921991348, + 0.01084955781698227, + 0.007974321022629738, + 0.0057542831636965275, + 0.004916290752589703, + 0.005202353000640869, + 0.005149912089109421, + 0.005177301354706287, + 0.0023125684820115566, + 0.0028151809237897396, + 0.003345297649502754, + 0.0025904374197125435, + 0.0027934142854064703, + 0.00240746489726007, + 0.003052119631320238, + 0.0017194758402183652, + 0.004202771000564098, + 0.003602794837206602, + 0.004280940629541874, + 0.0034792039077728987, + 0.0027110970113426447, + 0.00221922155469656, + 0.0030585909262299538, + 0.002332021715119481, + 0.007093253545463085, + 0.007139201276004314, + 0.0035171108320355415, + 0.0029690610244870186, + 0.001636532717384398, + 0.0015366229927167296, + 0.00302889384329319, + 0.0014449916779994965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.004103098995983601 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620361864566803 + ], + "opposite_logprob_tokens": [ + -2.8120362758636475 + ], + "logprob_reward": -0.0620361864566803, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620361864566803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48509490489959717, + 0.11316050589084625, + 0.1027417927980423, + 0.06406894326210022, + 0.04299067705869675, + 0.036655981093645096, + 0.038456182926893234, + 0.04865969344973564, + 0.029984349384903908, + 0.038186922669410706 + ], + "credit_map": [ + 0.10537681728601456, + 0.06559038162231445, + 0.029504509642720222, + 0.1712353378534317, + 0.037330418825149536, + 0.026730872690677643, + 0.026943707838654518, + 0.02238287217915058, + 0.013815534301102161, + 0.012598138302564621, + 0.011417869478464127, + 0.009602806530892849, + 0.014859771355986595, + 0.023944532498717308, + 0.012364470399916172, + 0.014557387679815292, + 0.010191543959081173, + 0.010486015118658543, + 0.005027388222515583, + 0.011077343486249447, + 0.010056301020085812, + 0.013402770273387432, + 0.029504260048270226, + 0.012996173463761806, + 0.009338146075606346, + 0.01061565987765789, + 0.0063114590011537075, + 0.007726776413619518, + 0.011983092874288559, + 0.0070823426358401775, + 0.005624999757856131, + 0.005386460572481155, + 0.003799698781222105, + 0.011671970598399639, + 0.0033288379199802876, + 0.002871089382097125, + 0.006958764977753162, + 0.004167743027210236, + 0.00460227532312274, + 0.005590297747403383, + 0.004733310081064701, + 0.005209004040807486, + 0.005026916041970253, + 0.004456306342035532, + 0.004308150615543127, + 0.004389671143144369, + 0.005152423866093159, + 0.0033801994286477566, + 0.005669041536748409, + 0.007648404221981764, + 0.005916445981711149, + 0.004497156944125891, + 0.004239950329065323, + 0.0037067693192511797, + 0.0031016930006444454, + 0.003676720429211855, + 0.005027472041547298, + 0.007180313114076853, + 0.007706334814429283, + 0.009268833324313164, + 0.004145230166614056, + 0.0032712905667722225, + 0.006555770989507437, + 0.0055044484324753284, + 0.004420184530317783, + 0.004342870321124792, + 0.00475994311273098, + 0.005288010463118553, + 0.0023863548412919044, + 0.0021624183282256126, + 0.003635072847828269, + 0.002989494940266013, + 0.009446375072002411, + 0.010128908790647984, + 0.005292018409818411, + 0.0040646931156516075, + 0.002517508342862129, + 0.001832474721595645, + 0.003328202059492469, + 0.0015767450677230954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0620361864566803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247865337878466 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005247865337878466, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247865337878466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2853166162967682, + 0.09935177117586136, + 0.04034729674458504, + 0.04112698882818222, + 0.06998283416032791, + 0.04291272163391113, + 0.127083882689476, + 0.10617717355489731, + 0.11001922190189362, + 0.07768146693706512 + ], + "credit_map": [ + 0.06025324761867523, + 0.0410657599568367, + 0.019940610975027084, + 0.09377218782901764, + 0.01941763609647751, + 0.021054837852716446, + 0.015810906887054443, + 0.014001410454511642, + 0.015750925987958908, + 0.011474372819066048, + 0.005389946047216654, + 0.009588372893631458, + 0.007675996981561184, + 0.02655491977930069, + 0.01015577744692564, + 0.012761464342474937, + 0.005253999028354883, + 0.007068085949867964, + 0.0031029670499265194, + 0.0040670582093298435, + 0.003950228914618492, + 0.003681205678731203, + 0.00847479235380888, + 0.004748957697302103, + 0.005751985590904951, + 0.005744741763919592, + 0.0036672691348940134, + 0.00455475365743041, + 0.0063227517530322075, + 0.005907642655074596, + 0.0039465418085455894, + 0.005231298040598631, + 0.007295539602637291, + 0.015207648277282715, + 0.005743491929024458, + 0.005578346084803343, + 0.01022619940340519, + 0.010607197880744934, + 0.008041705004870892, + 0.00728270597755909, + 0.005621215794235468, + 0.006392327137291431, + 0.003725319867953658, + 0.0035753410775214434, + 0.006504922639578581, + 0.007142298389226198, + 0.00564974220469594, + 0.004301551729440689, + 0.016301125288009644, + 0.015935035422444344, + 0.014025356620550156, + 0.016434509307146072, + 0.02107696793973446, + 0.017684025689959526, + 0.018075954169034958, + 0.007550896145403385, + 0.016188452020287514, + 0.015139940194785595, + 0.015059788711369038, + 0.013594492338597775, + 0.01368704903870821, + 0.009706984274089336, + 0.015570263378322124, + 0.007230211049318314, + 0.012639368884265423, + 0.014330318197607994, + 0.012883244082331657, + 0.015250272117555141, + 0.013841377571225166, + 0.010620695538818836, + 0.015232260338962078, + 0.015221686102449894, + 0.0130210742354393, + 0.012528883293271065, + 0.009519963525235653, + 0.009724617004394531, + 0.011313456110656261, + 0.006240246817469597, + 0.008813966996967793, + 0.006519260350614786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.005247865337878466 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176885321736336 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.018176885321736336, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176885321736336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5353350639343262, + 0.08309067785739899, + 0.04632257670164108, + 0.03209627419710159, + 0.05796452984213829, + 0.03854559734463692, + 0.08625257015228271, + 0.05257413908839226, + 0.02813020534813404, + 0.03968837484717369 + ], + "credit_map": [ + 0.11488605290651321, + 0.06629562377929688, + 0.03566544130444527, + 0.17840254306793213, + 0.04486990347504616, + 0.03942692652344704, + 0.029319224879145622, + 0.026469295844435692, + 0.015520395711064339, + 0.009770437143743038, + 0.0065397461876273155, + 0.008343257009983063, + 0.010183349251747131, + 0.01269044540822506, + 0.006876238156110048, + 0.013166808523237705, + 0.004913428332656622, + 0.0071263364516198635, + 0.0026489929296076298, + 0.004139844793826342, + 0.004709405824542046, + 0.0062875086441636086, + 0.010682645253837109, + 0.005814414471387863, + 0.003704850561916828, + 0.004700548946857452, + 0.003573397174477577, + 0.0040754228830337524, + 0.005017462186515331, + 0.004128519911319017, + 0.0037308880127966404, + 0.003165183588862419, + 0.0064391521736979485, + 0.015016842633485794, + 0.006205486133694649, + 0.004831351805478334, + 0.008000161498785019, + 0.006243781186640263, + 0.004377438686788082, + 0.0068503147922456264, + 0.005832953844219446, + 0.00826125219464302, + 0.003598067443817854, + 0.003369607264176011, + 0.0036022698041051626, + 0.006610518787056208, + 0.004210339393466711, + 0.0030605881474912167, + 0.011228983290493488, + 0.010664661414921284, + 0.013802888803184032, + 0.015694240108132362, + 0.008254235610365868, + 0.007367787417024374, + 0.011395039036870003, + 0.007844733074307442, + 0.0062827798537909985, + 0.006926835980266333, + 0.006312213372439146, + 0.011881629936397076, + 0.0050382125191390514, + 0.00317859067581594, + 0.00842265598475933, + 0.004531220532953739, + 0.004203476011753082, + 0.0039919898845255375, + 0.004272150807082653, + 0.004862841218709946, + 0.0028078043833374977, + 0.0026534933131188154, + 0.002729500876739621, + 0.002608948154374957, + 0.009205618873238564, + 0.010809089057147503, + 0.0033774327021092176, + 0.004521041177213192, + 0.003853413974866271, + 0.0022002633195370436, + 0.0033999013248831034, + 0.002321612788364291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.018176885321736336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the refined style of Pixar cartoons, the camera offers a steady, majestic crane shot, beginning with a low-angle close-up of the joyful splashing at the base of a vibrantly colored fountain. This whimsical fountain, shaped like a giant, smiling sunflower with glossy, saturated yellow petals, shoots crystal-clear water high into the air, where it then cascades down in a series of shimmering, sun-drenched droplets. As the camera steadily ascends, the plump, cheerful fountain is revealed, then the perspective widens to reveal a sprawling, whimsical water park beneath a brilliant, cloudless blue sky. The scene is a vibrant tapestry of color and movement: giant, intertwining water slides, shimmering in shades of candy red, electric blue, and lime green, twist and loop around each other, their rounded, polished surfaces reflecting the sunlight. Below, a variety of organically shaped, shimmering cyan pools are filled with stylized cartoon characters splashing and playing. As the camera peaks into a breathtaking high-angle wide shot, a magnificent, vibrant rainbow, born from the fine mist hanging in the air, arcs perfectly across the scene, its unique, saturated band of color stretching from one end of the park to the other, framing the winding slides and sparkling pools in a magical glow and creating a breathtaking atmosphere of pure, vibrant joy and the wonder of summer.", + "seed": 24384, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_10/index_5036_seed_24384_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012389035755768418, + "vlm_logprob": -0.012389035755768418, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.012389035755768418, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001815577270463109 + ], + "opposite_logprob_tokens": [ + -6.5018157958984375 + ], + "logprob_reward": -0.001815577270463109, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001815577270463109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11325056850910187, + 0.18155314028263092, + 0.06788314133882523, + 0.13430878520011902, + 0.032458268105983734, + 0.05359039828181267, + 0.3196537494659424, + 0.05973129719495773, + 0.022386105731129646, + 0.01518449280411005 + ], + "credit_map": [ + 0.01377738919109106, + 0.020542550832033157, + 0.011441362090408802, + 0.010910599492490292, + 0.022497307509183884, + 0.009956084191799164, + 0.012438276782631874, + 0.01168700773268938, + 0.017541084438562393, + 0.021813256666064262, + 0.014225918799638748, + 0.018458710983395576, + 0.04502611607313156, + 0.02271779626607895, + 0.026513313874602318, + 0.015256964601576328, + 0.009731194004416466, + 0.0091687748208642, + 0.006047914270311594, + 0.008105375804007053, + 0.011804061010479927, + 0.010193550027906895, + 0.005485898815095425, + 0.007346376311033964, + 0.01205898355692625, + 0.010355198755860329, + 0.007187406066805124, + 0.00718388007953763, + 0.04565906524658203, + 0.038246527314186096, + 0.00637054443359375, + 0.0072472007013857365, + 0.0027724800165742636, + 0.003805834800004959, + 0.0023415498435497284, + 0.006039017345756292, + 0.004174395464360714, + 0.003904468147084117, + 0.004350265488028526, + 0.005070262122899294, + 0.005155864171683788, + 0.005228784400969744, + 0.006410632282495499, + 0.006421631667762995, + 0.006018056999891996, + 0.006965967360883951, + 0.009259235113859177, + 0.008130230009555817, + 0.028832027688622475, + 0.03648459166288376, + 0.02831815928220749, + 0.050774767994880676, + 0.05671662837266922, + 0.04746970534324646, + 0.046371106058359146, + 0.02468680590391159, + 0.01722007803618908, + 0.003196325618773699, + 0.003340351628139615, + 0.01041482388973236, + 0.0053480640053749084, + 0.014423515647649765, + 0.003928655292838812, + 0.0018594915745779872, + 0.0033239806070923805, + 0.0030677125323563814, + 0.002126268343999982, + 0.0018651358550414443, + 0.00311701325699687, + 0.0042989980429410934, + 0.0025606879498809576, + 0.002026310423389077, + 0.00279400497674942, + 0.0008766963728703558, + 0.0017853628378361464, + 0.0016956875333562493, + 0.0016770386137068272, + 0.0024827171582728624, + 0.0023621811997145414, + 0.0015108059160411358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.001815577270463109 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645151551812887 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645151551812887, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645151551812887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06333400309085846, + 0.04952700808644295, + 0.04872330650687218, + 0.17353947460651398, + 0.04976411908864975, + 0.27331316471099854, + 0.1600543111562729, + 0.08028776198625565, + 0.08207853138446808, + 0.019378354772925377 + ], + "credit_map": [ + 0.007968106307089329, + 0.006895026657730341, + 0.007724628318101168, + 0.006946010980755091, + 0.00829159002751112, + 0.005626052618026733, + 0.01023429911583662, + 0.009648291394114494, + 0.004941392224282026, + 0.004230839665979147, + 0.009267508052289486, + 0.003701013047248125, + 0.00782399345189333, + 0.007539218757301569, + 0.004321285989135504, + 0.007701759226620197, + 0.00955792237073183, + 0.006007794290781021, + 0.008702780120074749, + 0.004121940582990646, + 0.004891849122941494, + 0.006646026857197285, + 0.003231189912185073, + 0.005563806742429733, + 0.009698526933789253, + 0.011661138385534286, + 0.005051969550549984, + 0.006746693514287472, + 0.06770651787519455, + 0.05794374644756317, + 0.0066620525903999805, + 0.008068816736340523, + 0.005635669454932213, + 0.0055727269500494, + 0.004369394853711128, + 0.011215762235224247, + 0.0054092006757855415, + 0.006355958990752697, + 0.005961286835372448, + 0.005244118627160788, + 0.037814535200595856, + 0.023013928905129433, + 0.03678159415721893, + 0.03481842577457428, + 0.03254978731274605, + 0.04133947193622589, + 0.030693294480443, + 0.0363021194934845, + 0.01629139669239521, + 0.01774793118238449, + 0.012681474909186363, + 0.016186757013201714, + 0.03511352837085724, + 0.03929727151989937, + 0.012823095545172691, + 0.009912860579788685, + 0.009896024130284786, + 0.00517638586461544, + 0.004911149386316538, + 0.025094635784626007, + 0.009462718851864338, + 0.01323652546852827, + 0.007753459271043539, + 0.004756861831992865, + 0.011094222776591778, + 0.012586268596351147, + 0.007069142069667578, + 0.010362576693296432, + 0.010413793846964836, + 0.011127312667667866, + 0.013047249987721443, + 0.006377964746206999, + 0.0037081982009112835, + 0.0020887593273073435, + 0.0014317199820652604, + 0.00181977404281497, + 0.0029565345030277967, + 0.0034799238201230764, + 0.0022593303583562374, + 0.0016341134905815125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.004645151551812887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06638439744710922, + 0.03900228068232536, + 0.03964558243751526, + 0.34023192524909973, + 0.053269416093826294, + 0.1844417303800583, + 0.1413816660642624, + 0.06635028123855591, + 0.05086343362927437, + 0.018429316580295563 + ], + "credit_map": [ + 0.008841083385050297, + 0.005378479138016701, + 0.010472806170582771, + 0.005577742122113705, + 0.006841525435447693, + 0.007370227016508579, + 0.008332287892699242, + 0.01357023511081934, + 0.004781766328960657, + 0.00413397466763854, + 0.00670453580096364, + 0.003381373593583703, + 0.005784942302852869, + 0.0054133301600813866, + 0.0032990870531648397, + 0.005503269378095865, + 0.004965286236256361, + 0.006864546798169613, + 0.0041898274794220924, + 0.004263055510818958, + 0.00654806662350893, + 0.0061337449587881565, + 0.003418263979256153, + 0.0032627906184643507, + 0.01727982796728611, + 0.022782493382692337, + 0.008466683328151703, + 0.009591495618224144, + 0.1386709213256836, + 0.11924140900373459, + 0.01140343863517046, + 0.012795638293027878, + 0.004212504252791405, + 0.0062958900816738605, + 0.00363217294216156, + 0.018156347796320915, + 0.005011606961488724, + 0.0029671553056687117, + 0.00818759948015213, + 0.004806137643754482, + 0.023741720244288445, + 0.01681586727499962, + 0.024592597037553787, + 0.024367721751332283, + 0.026248810812830925, + 0.025295766070485115, + 0.022509101778268814, + 0.020870143547654152, + 0.014703580178320408, + 0.009573662653565407, + 0.015192022547125816, + 0.014611874707043171, + 0.027320802211761475, + 0.03147958964109421, + 0.014964363537728786, + 0.013535763137042522, + 0.00839011836796999, + 0.0033252022694796324, + 0.0033428657334297895, + 0.01900041662156582, + 0.007047944236546755, + 0.015512851998209953, + 0.006228946149349213, + 0.003501931671053171, + 0.008262746967375278, + 0.005997524596750736, + 0.004249048884958029, + 0.005892727058380842, + 0.006610570475459099, + 0.007395764347165823, + 0.008935390040278435, + 0.0035196582321077585, + 0.003244817489758134, + 0.002319475868716836, + 0.001635703956708312, + 0.0015795322833582759, + 0.0025688852183520794, + 0.003247020300477743, + 0.0022877342998981476, + 0.001546145067550242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.00978010892868042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038082726299762726 + ], + "opposite_logprob_tokens": [ + -3.2880828380584717 + ], + "logprob_reward": -0.038082726299762726, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038082726299762726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05287902429699898, + 0.026144491508603096, + 0.021727602928876877, + 0.04690369963645935, + 0.040097031742334366, + 0.08632395416498184, + 0.29404786229133606, + 0.3759811222553253, + 0.039049334824085236, + 0.016845857724547386 + ], + "credit_map": [ + 0.006698213052004576, + 0.003943413496017456, + 0.007469111122190952, + 0.005785276647657156, + 0.005603325553238392, + 0.004876317922025919, + 0.011039508506655693, + 0.007463853806257248, + 0.005665008444339037, + 0.0023555115330964327, + 0.0038908785209059715, + 0.0016502657672390342, + 0.0029257037676870823, + 0.004230482038110495, + 0.0019734203815460205, + 0.0034532230347394943, + 0.0021691168658435345, + 0.003296102397143841, + 0.0024521020241081715, + 0.002937469631433487, + 0.002415187656879425, + 0.0030185282230377197, + 0.0025641866959631443, + 0.0028749092016369104, + 0.0032994248904287815, + 0.0032893819734454155, + 0.0018571536056697369, + 0.001629012986086309, + 0.01764894463121891, + 0.014598006382584572, + 0.0016651753103360534, + 0.0029165945015847683, + 0.004807407036423683, + 0.00525254663079977, + 0.004043517168611288, + 0.008133294060826302, + 0.0038343295454978943, + 0.004045694135129452, + 0.005518520716577768, + 0.004461720585823059, + 0.01215633749961853, + 0.008881757967174053, + 0.012222625315189362, + 0.009926248341798782, + 0.01061234064400196, + 0.012384110130369663, + 0.009351192973554134, + 0.01078934408724308, + 0.030435243621468544, + 0.027999557554721832, + 0.033191852271556854, + 0.04791279509663582, + 0.04268995299935341, + 0.03163450211286545, + 0.04725717380642891, + 0.03292679041624069, + 0.10867679864168167, + 0.01772652380168438, + 0.016344571486115456, + 0.058762889355421066, + 0.031201761215925217, + 0.10282297432422638, + 0.024080391973257065, + 0.016365237534046173, + 0.005965786054730415, + 0.0053662327118217945, + 0.0037922197952866554, + 0.004889131989330053, + 0.006362175568938255, + 0.00390378525480628, + 0.00577145628631115, + 0.0029985455330461264, + 0.004539711866527796, + 0.0016357688000425696, + 0.0009082389296963811, + 0.0009078450966626406, + 0.0018388662720099092, + 0.004870748147368431, + 0.0012370935874059796, + 0.0009075841517187655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.038082726299762726 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the fountain shaped like a giant, smiling sunflower with glossy yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there giant water slides in shades of candy red, electric blue, and lime green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there organically shaped cyan pools filled with stylized cartoon characters splashing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vibrant rainbow arc across the scene from one end of the park to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle close-up and ascend to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762161472812295 + ], + "opposite_logprob_tokens": [ + -4.882621765136719 + ], + "logprob_reward": -0.00762161472812295, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762161472812295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09562581032514572, + 0.02643863670527935, + 0.04353758692741394, + 0.6317229270935059, + 0.018700623884797096, + 0.025300677865743637, + 0.10092996805906296, + 0.04418526589870453, + 0.007981846109032631, + 0.0055766841396689415 + ], + "credit_map": [ + 0.01347313355654478, + 0.009059585630893707, + 0.014909022487699986, + 0.009039277210831642, + 0.009847719222307205, + 0.011485178954899311, + 0.013852180913090706, + 0.013959728181362152, + 0.003245686646550894, + 0.0032518862280994654, + 0.004573690239340067, + 0.002564367139711976, + 0.0036647289525717497, + 0.003483439330011606, + 0.0026420573703944683, + 0.003012785455211997, + 0.0032937536016106606, + 0.010052211582660675, + 0.00378590845502913, + 0.0035456062760204077, + 0.009340150281786919, + 0.007945405319333076, + 0.0029313024133443832, + 0.0026432545855641365, + 0.03183191642165184, + 0.04062959551811218, + 0.015916919335722923, + 0.013333993032574654, + 0.2657793462276459, + 0.22437970340251923, + 0.017601074650883675, + 0.02225041203200817, + 0.0018139333697035909, + 0.0018260026117786765, + 0.001583315315656364, + 0.003466720459982753, + 0.0017644722247496247, + 0.0019260499393567443, + 0.0035771294496953487, + 0.0027430022601038218, + 0.0030696741305291653, + 0.0024111312814056873, + 0.0035335454158484936, + 0.002956925891339779, + 0.002841019304469228, + 0.0040111243724823, + 0.003055969951674342, + 0.003421291010454297, + 0.010940167121589184, + 0.008900615386664867, + 0.01139463484287262, + 0.009840511716902256, + 0.019984539598226547, + 0.019834592938423157, + 0.00991727039217949, + 0.010117645375430584, + 0.01359779667109251, + 0.0019479572074487805, + 0.0020620375871658325, + 0.0017618011916056275, + 0.0038088986184448004, + 0.01798979565501213, + 0.001559181953780353, + 0.0014578017871826887, + 0.001213911920785904, + 0.0009436435648240149, + 0.0006847631302662194, + 0.0005774110904894769, + 0.0010067424736917019, + 0.0014445041306316853, + 0.001018806011416018, + 0.0010920652421191335, + 0.0007712168735451996, + 0.0005563327576965094, + 0.00042632254189811647, + 0.0005633049295283854, + 0.0007250865455716848, + 0.0007893113070167601, + 0.0011335769668221474, + 0.0006115327123552561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.00762161472812295 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_11/index_tensor([138, 138])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_11/index_tensor([138, 138])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b10dcfa8b250d3de4a58f805051eb11461faf2b2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_11/index_tensor([138, 138])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An immersive first-person point-of-view (POV) shot, the camera steadily advances at a slow, wistful pace through a narrow, damp alleyway imbued with a cyberpunk aesthetic. Rainwater, reflecting off multicolored neon lights, pools on the uneven concrete floor. Towering building walls flank the ground, a tangle of wires, rusting pipes, and torn holographic advertisements in foreign languages. The camera's focus finally settles on a simple barbecue stall deep in the alley. A thick steak, etched with a grid of caramelized grill marks, sizzles enticingly on the scorching metal grill, its rich fat melting and boiling under the intense heat. As the camera approaches, a strange phenomenon occurs: rising from the steak's charred crust is no longer ordinary steam, but rapidly condenses and transforms into a dazzling, upward-floating stream of holographic data composed of binary code, glowing circuit diagrams, and shimmering geometric symbols. This stream of digital steam, shimmering in neon blues, purples, and greens, appears both ethereal and alluring against the warm orange glow of the grill and the cool ambient light of the alley. The entire process is imbued with a possessive intensity, as if the diner's eyes are poised to devour this digital delicacy, a fusion of primitive cooking and futuristic technology, creating a psychedelic atmosphere of seeking sensory immersion in a decadent future.", + "seed": 6171, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_11/index_138_seed_6171_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01987304324284196, + "vlm_logprob": -0.01987304324284196, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.01987304324284196, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0817657932639122 + ], + "opposite_logprob_tokens": [ + -2.58176589012146 + ], + "logprob_reward": -0.0817657932639122, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0817657932639122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05234812572598457, + 0.03322218358516693, + 0.05496541038155556, + 0.22265343368053436, + 0.09076478332281113, + 0.08421964943408966, + 0.044759273529052734, + 0.03287065774202347, + 0.040151968598365784, + 0.3440445065498352 + ], + "credit_map": [ + 0.00805211067199707, + 0.0047569358721375465, + 0.006799883209168911, + 0.00721546309068799, + 0.007123631425201893, + 0.0052884346805512905, + 0.0059426636435091496, + 0.007169004529714584, + 0.002532375743612647, + 0.0029927254654467106, + 0.0049755326472222805, + 0.00588172348216176, + 0.003350840648636222, + 0.0051961797289550304, + 0.004175281617790461, + 0.0041175223886966705, + 0.004244299605488777, + 0.006663407199084759, + 0.005443872883915901, + 0.006305369548499584, + 0.006872252561151981, + 0.010112433694303036, + 0.007839477621018887, + 0.007484300062060356, + 0.01299783494323492, + 0.018099885433912277, + 0.01480049453675747, + 0.024059468880295753, + 0.030717119574546814, + 0.04787679761648178, + 0.034218091517686844, + 0.03988373652100563, + 0.006002605427056551, + 0.009878127835690975, + 0.006020242348313332, + 0.016447201371192932, + 0.020702453330159187, + 0.014032573439180851, + 0.00878544058650732, + 0.00889613851904869, + 0.006183545105159283, + 0.00698558334261179, + 0.008805755525827408, + 0.01126958429813385, + 0.01307240966707468, + 0.011859040707349777, + 0.013304718770086765, + 0.012739002704620361, + 0.006656239740550518, + 0.0035733471158891916, + 0.005638482514768839, + 0.005521234590560198, + 0.004608108662068844, + 0.007299806922674179, + 0.004730599001049995, + 0.0067314570769667625, + 0.004987647756934166, + 0.0044245426543056965, + 0.004448356106877327, + 0.0021655093878507614, + 0.005094642285257578, + 0.006700930651277304, + 0.0023262654431164265, + 0.0027227611280977726, + 0.007699454668909311, + 0.005276553798466921, + 0.0033116121776401997, + 0.0022788182832300663, + 0.008828829973936081, + 0.006714977324008942, + 0.002944527193903923, + 0.0030971975065767765, + 0.13861829042434692, + 0.013384976424276829, + 0.007281459867954254, + 0.005606368649750948, + 0.016421152278780937, + 0.14018458127975464, + 0.008068024180829525, + 0.01447964459657669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.0817657932639122 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015255491016432643 + ], + "opposite_logprob_tokens": [ + -6.501525402069092 + ], + "logprob_reward": -0.0015255491016432643, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015255491016432643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0640021488070488, + 0.020029809325933456, + 0.04796567186713219, + 0.014952040277421474, + 0.1255565881729126, + 0.036670442670583725, + 0.025136249139904976, + 0.013499081134796143, + 0.02609141357243061, + 0.6260964870452881 + ], + "credit_map": [ + 0.012056319043040276, + 0.005708209704607725, + 0.009188689291477203, + 0.008938804268836975, + 0.005516411270946264, + 0.006907834205776453, + 0.007452780846506357, + 0.008233105763792992, + 0.0016400442691519856, + 0.0023876773193478584, + 0.0021834918297827244, + 0.0023562766145914793, + 0.0019016617443412542, + 0.004193724598735571, + 0.0027238770853728056, + 0.0026430576108396053, + 0.0053245616145431995, + 0.004679806064814329, + 0.01028159074485302, + 0.005678086541593075, + 0.005022197961807251, + 0.006008508615195751, + 0.004701031371951103, + 0.0062698908150196075, + 0.0009989675600081682, + 0.0013949241256341338, + 0.0016277465038001537, + 0.0016532732406631112, + 0.0027502556331455708, + 0.0023177186958491802, + 0.0019426440121605992, + 0.002266512019559741, + 0.006134863942861557, + 0.008988526649773121, + 0.006316402927041054, + 0.02017492987215519, + 0.03629922494292259, + 0.027538880705833435, + 0.012100446037948132, + 0.00800332147628069, + 0.002541205147281289, + 0.0031242549885064363, + 0.0037638330832123756, + 0.004745737183839083, + 0.005017276853322983, + 0.005367093253880739, + 0.004403726663440466, + 0.007707315497100353, + 0.0030717861372977495, + 0.0023848963901400566, + 0.003874173853546381, + 0.0025796531699597836, + 0.003192988457158208, + 0.004194288514554501, + 0.002589847194030881, + 0.0032486170530319214, + 0.0017969007603824139, + 0.001289887586608529, + 0.0018567348597571254, + 0.0012416276149451733, + 0.002122489269822836, + 0.0018761161481961608, + 0.0010184553684666753, + 0.002296870807185769, + 0.004552172031253576, + 0.0034104809165000916, + 0.002498982008546591, + 0.002192975487560034, + 0.005580699071288109, + 0.003574205096811056, + 0.002153503242880106, + 0.002128398045897484, + 0.2531471848487854, + 0.023216646164655685, + 0.012608607299625874, + 0.009621039032936096, + 0.029766740277409554, + 0.25664976239204407, + 0.013923441991209984, + 0.02716311253607273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.0015255491016432643 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740096490830183 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006740096490830183, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740096490830183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021739810705184937, + 0.011232919059693813, + 0.02987874485552311, + 0.029708750545978546, + 0.24467216432094574, + 0.027320455759763718, + 0.07812192291021347, + 0.02996179647743702, + 0.015974244102835655, + 0.5113892555236816 + ], + "credit_map": [ + 0.004092989023774862, + 0.0020798558834940195, + 0.0038277076091617346, + 0.0030134764965623617, + 0.0019864807836711407, + 0.0019229826284572482, + 0.002206304809078574, + 0.002610014285892248, + 0.0008589399512857199, + 0.0009073158726096153, + 0.001324716373346746, + 0.0015505290357396007, + 0.001280111144296825, + 0.0019604277331382036, + 0.0016798835713416338, + 0.0016709960764274001, + 0.0018493343377485871, + 0.002492232946678996, + 0.0036673303693532944, + 0.0030950671061873436, + 0.0025691778864711523, + 0.007536949589848518, + 0.003032392356544733, + 0.005636259913444519, + 0.0019267461029812694, + 0.002687429543584585, + 0.0027690143324434757, + 0.0026008598506450653, + 0.005565331783145666, + 0.0055025797337293625, + 0.0037403900641947985, + 0.004916399251669645, + 0.01704542338848114, + 0.024455653503537178, + 0.018010592088103294, + 0.048885077238082886, + 0.054089393466711044, + 0.03830013796687126, + 0.021780529990792274, + 0.02210536226630211, + 0.002417038194835186, + 0.0027326359413564205, + 0.003040821524336934, + 0.0034541080240160227, + 0.0036331219598650932, + 0.003520489204674959, + 0.0031339717097580433, + 0.005388268269598484, + 0.010077692568302155, + 0.009031461551785469, + 0.011252081952989101, + 0.008448008447885513, + 0.010281279683113098, + 0.012918938882648945, + 0.006788079626858234, + 0.009324382990598679, + 0.0034816665574908257, + 0.0027179629541933537, + 0.003486382309347391, + 0.0038265844341367483, + 0.003860136494040489, + 0.004797591827809811, + 0.003549290122464299, + 0.0042421817779541016, + 0.002884872956201434, + 0.001943093491718173, + 0.001961569068953395, + 0.001728635048493743, + 0.0019538954365998507, + 0.0019471788546070457, + 0.0016144794644787908, + 0.0019405202474445105, + 0.20697687566280365, + 0.01896623894572258, + 0.010356116108596325, + 0.007711638230830431, + 0.023925572633743286, + 0.209927037358284, + 0.011370447464287281, + 0.0221552737057209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.006740096490830183 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622324395924807 + ], + "opposite_logprob_tokens": [ + -4.882622241973877 + ], + "logprob_reward": -0.007622324395924807, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622324395924807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038025759160518646, + 0.03388708829879761, + 0.04426393657922745, + 0.024293988943099976, + 0.06609658896923065, + 0.022529272362589836, + 0.034335218369960785, + 0.07889038324356079, + 0.05098617076873779, + 0.6066915988922119 + ], + "credit_map": [ + 0.0057050385512411594, + 0.002928918693214655, + 0.005670865532010794, + 0.005308825056999922, + 0.005498038604855537, + 0.004445916973054409, + 0.0036705813836306334, + 0.00479757459834218, + 0.0036793008912354708, + 0.0037258737720549107, + 0.004167551174759865, + 0.004479854367673397, + 0.003361878916621208, + 0.006183384917676449, + 0.004334229975938797, + 0.003955014981329441, + 0.004672365263104439, + 0.0033346957061439753, + 0.011509755626320839, + 0.005278424359858036, + 0.0031738560646772385, + 0.006546486169099808, + 0.003616292728111148, + 0.006132063455879688, + 0.0015610557748004794, + 0.0021784231066703796, + 0.0022974845487624407, + 0.0030202767811715603, + 0.00399783393368125, + 0.004005535505712032, + 0.00404546782374382, + 0.003187913214787841, + 0.005071389488875866, + 0.00947975367307663, + 0.005703875329345465, + 0.014647412113845348, + 0.009202937595546246, + 0.007120310794562101, + 0.00830922368913889, + 0.006561686750501394, + 0.0023830856662243605, + 0.0020986218005418777, + 0.0027754423208534718, + 0.0030260905623435974, + 0.004377264995127916, + 0.0023456444032490253, + 0.0017913705669343472, + 0.0037317543756216764, + 0.0033219384495168924, + 0.002461286960169673, + 0.003937208093702793, + 0.003399551846086979, + 0.004942415282130241, + 0.006734903901815414, + 0.004003916401416063, + 0.005534000229090452, + 0.012312026694417, + 0.010170563124120235, + 0.009972630068659782, + 0.004033504519611597, + 0.013223725371062756, + 0.01611904986202717, + 0.003381163813173771, + 0.009677733294665813, + 0.009243277832865715, + 0.007325873244553804, + 0.003769773757085204, + 0.0028150600846856833, + 0.011764956638216972, + 0.00827854499220848, + 0.0038413465954363346, + 0.003947339951992035, + 0.2452261596918106, + 0.023499390110373497, + 0.012854857370257378, + 0.009332568384706974, + 0.028451256453990936, + 0.24785618484020233, + 0.013790328055620193, + 0.025680894032120705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.007622324395924807 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017114529618993402 + ], + "opposite_logprob_tokens": [ + -6.376711368560791 + ], + "logprob_reward": -0.0017114529618993402, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017114529618993402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05445774272084236, + 0.019769858568906784, + 0.08563070744276047, + 0.025519248098134995, + 0.044098637998104095, + 0.028825124725699425, + 0.06106971949338913, + 0.061801258474588394, + 0.049881722778081894, + 0.5689459443092346 + ], + "credit_map": [ + 0.012237493880093098, + 0.0041422671638429165, + 0.01065328624099493, + 0.0062219067476689816, + 0.004495526198297739, + 0.00492431316524744, + 0.0052082170732319355, + 0.00657473923638463, + 0.001503869192674756, + 0.002307963091880083, + 0.0021670693531632423, + 0.0023411524016410112, + 0.0015044288011267781, + 0.0038531068712472916, + 0.0036870832554996014, + 0.002405187813565135, + 0.004710354842245579, + 0.003955853637307882, + 0.007877937518060207, + 0.006119228899478912, + 0.005108178593218327, + 0.02876991033554077, + 0.006921450607478619, + 0.022167807444930077, + 0.0016543855890631676, + 0.002201193943619728, + 0.0020108125172555447, + 0.0026548674795776606, + 0.004835125990211964, + 0.0029869244899600744, + 0.005079976748675108, + 0.004095962271094322, + 0.002883546520024538, + 0.003957610111683607, + 0.003683428280055523, + 0.00654225517064333, + 0.010646977461874485, + 0.008994597010314465, + 0.004103579558432102, + 0.003286649938672781, + 0.0029754010029137135, + 0.0033311154693365097, + 0.004270645324140787, + 0.0036876583471894264, + 0.005539115518331528, + 0.0036782841198146343, + 0.0022134436294436455, + 0.0031294659711420536, + 0.008342786692082882, + 0.00875072181224823, + 0.009581071324646473, + 0.005891697946935892, + 0.008994902484118938, + 0.008119410835206509, + 0.004312570206820965, + 0.007076560985296965, + 0.009054189547896385, + 0.0065957289189100266, + 0.0072675952687859535, + 0.004960080608725548, + 0.010076411999762058, + 0.012232647277414799, + 0.0034635260235518217, + 0.008151080459356308, + 0.007988166995346546, + 0.006807877216488123, + 0.005016534123569727, + 0.006365063134580851, + 0.00782192125916481, + 0.006816728040575981, + 0.004118467681109905, + 0.004946969449520111, + 0.2273649126291275, + 0.022671181708574295, + 0.012898415327072144, + 0.009496958926320076, + 0.02772732451558113, + 0.2304912954568863, + 0.013405094854533672, + 0.02489076368510723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.0017114529618993402 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive first-person point-of-view (POV) shot, the camera steadily advances at a slow, wistful pace through a narrow, damp alleyway imbued with a cyberpunk aesthetic. Rainwater, reflecting off multicolored neon lights, pools on the uneven concrete floor. Towering building walls flank the ground, a tangle of wires, rusting pipes, and torn holographic advertisements in foreign languages. The camera's focus finally settles on a simple barbecue stall deep in the alley. A thick steak, etched with a grid of caramelized grill marks, sizzles enticingly on the scorching metal grill, its rich fat melting and boiling under the intense heat. As the camera approaches, a strange phenomenon occurs: rising from the steak's charred crust is no longer ordinary steam, but rapidly condenses and transforms into a dazzling, upward-floating stream of holographic data composed of binary code, glowing circuit diagrams, and shimmering geometric symbols. This stream of digital steam, shimmering in neon blues, purples, and greens, appears both ethereal and alluring against the warm orange glow of the grill and the cool ambient light of the alley. The entire process is imbued with a possessive intensity, as if the diner's eyes are poised to devour this digital delicacy, a fusion of primitive cooking and futuristic technology, creating a psychedelic atmosphere of seeking sensory immersion in a decadent future.", + "seed": 6172, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_11/index_138_seed_6172_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014185862732119858, + "vlm_logprob": -0.014185862732119858, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.014185862732119858, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05616845190525055 + ], + "opposite_logprob_tokens": [ + -3.056168556213379 + ], + "logprob_reward": -0.05616845190525055, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05616845190525055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2506362199783325, + 0.14354307949543, + 0.06443944573402405, + 0.10958616435527802, + 0.08008376508951187, + 0.07251568138599396, + 0.08803659677505493, + 0.05705207213759422, + 0.07154789566993713, + 0.06255900114774704 + ], + "credit_map": [ + 0.026644214987754822, + 0.01941685378551483, + 0.04788623005151749, + 0.0338410846889019, + 0.01968677155673504, + 0.025293517857789993, + 0.02176869660615921, + 0.05609888955950737, + 0.022004801779985428, + 0.008619622327387333, + 0.029393918812274933, + 0.021254228428006172, + 0.010019286535680294, + 0.014948041178286076, + 0.011686424724757671, + 0.025616765022277832, + 0.00614283699542284, + 0.0047859749756753445, + 0.011442710645496845, + 0.007299500983208418, + 0.004884663969278336, + 0.010887108743190765, + 0.009280175901949406, + 0.009716477245092392, + 0.014229887165129185, + 0.012885567732155323, + 0.01544101256877184, + 0.014839069917798042, + 0.015995342284440994, + 0.013214487582445145, + 0.01193738542497158, + 0.011043417267501354, + 0.010692463256418705, + 0.009063483215868473, + 0.01452366728335619, + 0.01096553634852171, + 0.007063596975058317, + 0.009732311591506004, + 0.009191964752972126, + 0.008850748650729656, + 0.008685932494699955, + 0.010781927034258842, + 0.01308200228959322, + 0.00714909890666604, + 0.006378838326781988, + 0.01156570389866829, + 0.007080497685819864, + 0.00779168214648962, + 0.01071952749043703, + 0.006724813021719456, + 0.010059216059744358, + 0.008761062286794186, + 0.010538310743868351, + 0.01865115575492382, + 0.013357051648199558, + 0.009225460700690746, + 0.005936901550740004, + 0.006550731603056192, + 0.008336295373737812, + 0.0069112046621739864, + 0.007615251932293177, + 0.010925059206783772, + 0.006516436580568552, + 0.00426019262522459, + 0.00722801499068737, + 0.008073992095887661, + 0.012055317871272564, + 0.004346644040197134, + 0.01363775972276926, + 0.019221685826778412, + 0.004201394505798817, + 0.002783088944852352, + 0.007076734211295843, + 0.005470734555274248, + 0.011869304813444614, + 0.004968235269188881, + 0.006465574260801077, + 0.017830390483140945, + 0.004559269640594721, + 0.0043187630362808704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.05616845190525055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017347777029499412 + ], + "opposite_logprob_tokens": [ + -6.376734733581543 + ], + "logprob_reward": -0.0017347777029499412, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017347777029499412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09786292165517807, + 0.4042603075504303, + 0.11993308365345001, + 0.07363896071910858, + 0.056916769593954086, + 0.04598686099052429, + 0.055989135056734085, + 0.026237791404128075, + 0.04661918804049492, + 0.07255499064922333 + ], + "credit_map": [ + 0.014127871952950954, + 0.007571985945105553, + 0.02045033685863018, + 0.009350226260721684, + 0.007670914754271507, + 0.008880581706762314, + 0.006963802967220545, + 0.02284719981253147, + 0.03908003494143486, + 0.017059605568647385, + 0.1254592388868332, + 0.03590322658419609, + 0.02138008549809456, + 0.036257460713386536, + 0.021461181342601776, + 0.1076594740152359, + 0.010272289626300335, + 0.011229762807488441, + 0.02078298293054104, + 0.014242158271372318, + 0.009210336953401566, + 0.016571098938584328, + 0.018568819388747215, + 0.01905563473701477, + 0.008821762166917324, + 0.007098831702023745, + 0.0073114121332764626, + 0.008843904361128807, + 0.01071606483310461, + 0.008473386988043785, + 0.012179018929600716, + 0.010194585658609867, + 0.008025250397622585, + 0.01035765279084444, + 0.008027667179703712, + 0.007975267246365547, + 0.005679817404597998, + 0.0071376110427081585, + 0.004229128826409578, + 0.005484380759298801, + 0.006294294726103544, + 0.00544927641749382, + 0.006951197516173124, + 0.004550958052277565, + 0.004151432309299707, + 0.006147591397166252, + 0.0066058579832315445, + 0.005836255382746458, + 0.006798500195145607, + 0.004691001959145069, + 0.008380616083741188, + 0.00586796086281538, + 0.004890590440481901, + 0.012788680382072926, + 0.00718322629109025, + 0.0053885625675320625, + 0.003502767765894532, + 0.002642564009875059, + 0.004016653634607792, + 0.003152132499963045, + 0.0033292316365987062, + 0.00517240259796381, + 0.002361925784498453, + 0.002060117432847619, + 0.006010053213685751, + 0.004030040930956602, + 0.007272858172655106, + 0.004048159811645746, + 0.007388599682599306, + 0.010345758870244026, + 0.0053367153741419315, + 0.002187008736655116, + 0.010961911641061306, + 0.004028779454529285, + 0.01383897103369236, + 0.003238464007154107, + 0.006456057541072369, + 0.027928801253437996, + 0.0035917761269956827, + 0.0025102400686591864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0017347777029499412 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098350182175636 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098350182175636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098350182175636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20615793764591217, + 0.09306426346302032, + 0.052829187363386154, + 0.040042925626039505, + 0.09875394403934479, + 0.0770869106054306, + 0.08946578949689865, + 0.16149917244911194, + 0.11696469783782959, + 0.06413517892360687 + ], + "credit_map": [ + 0.029894337058067322, + 0.014960972592234612, + 0.03445703536272049, + 0.02609500102698803, + 0.013889824971556664, + 0.02803085371851921, + 0.021327434107661247, + 0.037502482533454895, + 0.009650591760873795, + 0.005174641963094473, + 0.025217654183506966, + 0.007949478924274445, + 0.0063530514016747475, + 0.00787983275949955, + 0.008817747235298157, + 0.022021271288394928, + 0.005328722298145294, + 0.005332429427653551, + 0.009326952509582043, + 0.006266376003623009, + 0.00428305147215724, + 0.008006989024579525, + 0.004705042112618685, + 0.00957962404936552, + 0.003703794442117214, + 0.004119545221328735, + 0.005741051863878965, + 0.0045679183676838875, + 0.0072778016328811646, + 0.0045004952698946, + 0.004391900729387999, + 0.005740420892834663, + 0.01343945786356926, + 0.010108000598847866, + 0.007848352193832397, + 0.014003545045852661, + 0.010209744796156883, + 0.01281098835170269, + 0.017899250611662865, + 0.012434602715075016, + 0.0077901799231767654, + 0.010622181929647923, + 0.01698010042309761, + 0.0073488918133080006, + 0.005675782915204763, + 0.009981570765376091, + 0.011512810364365578, + 0.007175392471253872, + 0.010443159379065037, + 0.008901357650756836, + 0.010276240296661854, + 0.011044878512620926, + 0.00877202395349741, + 0.016247622668743134, + 0.015702830627560616, + 0.008077673614025116, + 0.019371144473552704, + 0.012911133468151093, + 0.023142434656620026, + 0.016001813113689423, + 0.028938790783286095, + 0.026003992184996605, + 0.020722294226288795, + 0.014407570473849773, + 0.016080720350146294, + 0.009483558125793934, + 0.01775640994310379, + 0.014459570869803429, + 0.014074718579649925, + 0.02024545893073082, + 0.015597553923726082, + 0.00926670990884304, + 0.009368747472763062, + 0.005652495659887791, + 0.008872834965586662, + 0.006544801872223616, + 0.004905955400317907, + 0.015378369018435478, + 0.008352858945727348, + 0.005059113726019859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.004098350182175636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731808185577393 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006731808185577393, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731808185577393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1818038374185562, + 0.14853063225746155, + 0.04253841191530228, + 0.09255286306142807, + 0.11832824349403381, + 0.0678568035364151, + 0.11331097781658173, + 0.05096925422549248, + 0.13932225108146667, + 0.044786740094423294 + ], + "credit_map": [ + 0.025517689064145088, + 0.012722035869956017, + 0.033899370580911636, + 0.017642104998230934, + 0.012917744927108288, + 0.02419985644519329, + 0.010599072091281414, + 0.0443059466779232, + 0.011883996427059174, + 0.00929662398993969, + 0.03971090540289879, + 0.02393469773232937, + 0.006560569163411856, + 0.016075341030955315, + 0.006906501017510891, + 0.03416198492050171, + 0.0034947292879223824, + 0.003928438760340214, + 0.006177832838147879, + 0.005045871250331402, + 0.004768020939081907, + 0.00925921369343996, + 0.004062994848936796, + 0.005801305174827576, + 0.010496867820620537, + 0.010038118809461594, + 0.013758515939116478, + 0.011757872067391872, + 0.01437870692461729, + 0.009792131371796131, + 0.011647858656942844, + 0.010682792402803898, + 0.02002316527068615, + 0.012498456984758377, + 0.010892853140830994, + 0.022467972710728645, + 0.011016159318387508, + 0.015933465212583542, + 0.013629309833049774, + 0.011866851709783077, + 0.008408538065850735, + 0.012677337974309921, + 0.010081715881824493, + 0.008379667066037655, + 0.006625347770750523, + 0.009361865930259228, + 0.006836532149463892, + 0.0054857986979186535, + 0.015297501347959042, + 0.009636366739869118, + 0.017291264608502388, + 0.011772914789617062, + 0.00851006805896759, + 0.02816038392484188, + 0.013428750447928905, + 0.009213718585669994, + 0.006517586298286915, + 0.00464739091694355, + 0.006941119208931923, + 0.005857988726347685, + 0.0062432908453047276, + 0.009393255226314068, + 0.006950548849999905, + 0.004418072756379843, + 0.01378441322594881, + 0.011663681827485561, + 0.027970559895038605, + 0.007363611366599798, + 0.025659266859292984, + 0.041296254843473434, + 0.007178639993071556, + 0.004405821207910776, + 0.006647035945206881, + 0.0038922138046473265, + 0.006928251124918461, + 0.0027013435028493404, + 0.006240128539502621, + 0.012531842105090618, + 0.0035237951669842005, + 0.0023221266455948353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.006731808185577393 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195925684645772 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002195925684645772, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195925684645772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2420857697725296, + 0.07090779393911362, + 0.0272075105458498, + 0.043446194380521774, + 0.05549849942326546, + 0.10638779401779175, + 0.055344052612781525, + 0.04382706806063652, + 0.11688431352376938, + 0.23841096460819244 + ], + "credit_map": [ + 0.033963076770305634, + 0.01615302823483944, + 0.054846085608005524, + 0.019459731876850128, + 0.011476404033601284, + 0.03350011631846428, + 0.010439138859510422, + 0.06224820017814636, + 0.004793427884578705, + 0.0023421079386025667, + 0.02264028787612915, + 0.004926633555442095, + 0.004530838690698147, + 0.006688354536890984, + 0.004273833706974983, + 0.020712312310934067, + 0.0024314699694514275, + 0.0023632491938769817, + 0.004560776986181736, + 0.003828715765848756, + 0.001896057277917862, + 0.0051783351227641106, + 0.002925626467913389, + 0.004023279063403606, + 0.005900026764720678, + 0.005547155626118183, + 0.007512383628636599, + 0.006820288021117449, + 0.004855697974562645, + 0.0050178831443190575, + 0.0038838437758386135, + 0.003908912185579538, + 0.009921299293637276, + 0.005132970400154591, + 0.007116199471056461, + 0.008531615138053894, + 0.004691334441304207, + 0.006507994141429663, + 0.008105951361358166, + 0.005491137970238924, + 0.011920751072466373, + 0.016490576788783073, + 0.023053709417581558, + 0.011080684140324593, + 0.007301709149032831, + 0.01541359443217516, + 0.014549885876476765, + 0.006576884537935257, + 0.00795985572040081, + 0.006072117947041988, + 0.0068686832673847675, + 0.004940688144415617, + 0.008389911614358425, + 0.011532718315720558, + 0.005878935102373362, + 0.003701143665239215, + 0.004689785186201334, + 0.003963883966207504, + 0.006300368811935186, + 0.005265332758426666, + 0.0052578565664589405, + 0.007786870934069157, + 0.0061982194893062115, + 0.004364745691418648, + 0.012861760333180428, + 0.008815277367830276, + 0.022292930632829666, + 0.004426988773047924, + 0.021180354058742523, + 0.038234367966651917, + 0.005348381586372852, + 0.003724257694557309, + 0.03259127959609032, + 0.013394984416663647, + 0.04391172528266907, + 0.011513961479067802, + 0.02821262925863266, + 0.08716423064470291, + 0.013854701071977615, + 0.007767470553517342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.002195925684645772 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive first-person point-of-view (POV) shot, the camera steadily advances at a slow, wistful pace through a narrow, damp alleyway imbued with a cyberpunk aesthetic. Rainwater, reflecting off multicolored neon lights, pools on the uneven concrete floor. Towering building walls flank the ground, a tangle of wires, rusting pipes, and torn holographic advertisements in foreign languages. The camera's focus finally settles on a simple barbecue stall deep in the alley. A thick steak, etched with a grid of caramelized grill marks, sizzles enticingly on the scorching metal grill, its rich fat melting and boiling under the intense heat. As the camera approaches, a strange phenomenon occurs: rising from the steak's charred crust is no longer ordinary steam, but rapidly condenses and transforms into a dazzling, upward-floating stream of holographic data composed of binary code, glowing circuit diagrams, and shimmering geometric symbols. This stream of digital steam, shimmering in neon blues, purples, and greens, appears both ethereal and alluring against the warm orange glow of the grill and the cool ambient light of the alley. The entire process is imbued with a possessive intensity, as if the diner's eyes are poised to devour this digital delicacy, a fusion of primitive cooking and futuristic technology, creating a psychedelic atmosphere of seeking sensory immersion in a decadent future.", + "seed": 6173, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_11/index_138_seed_6173_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02568201133981347, + "vlm_logprob": -0.02568201133981347, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.02568201133981347, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09975051879882812 + ], + "opposite_logprob_tokens": [ + -2.474750518798828 + ], + "logprob_reward": -0.09975051879882812, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09975051879882812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17991574108600616, + 0.08737535774707794, + 0.05475624278187752, + 0.08979672193527222, + 0.0671522468328476, + 0.172418475151062, + 0.04326187074184418, + 0.21000319719314575, + 0.04093598946928978, + 0.05438413470983505 + ], + "credit_map": [ + 0.021541381254792213, + 0.014296060428023338, + 0.03491911292076111, + 0.016909990459680557, + 0.017388850450515747, + 0.030432866886258125, + 0.017377939075231552, + 0.027049535885453224, + 0.006270326673984528, + 0.008385691791772842, + 0.013972262851893902, + 0.010225459933280945, + 0.006695571355521679, + 0.015519077889621258, + 0.013519263826310635, + 0.012787695974111557, + 0.007417482323944569, + 0.004747174214571714, + 0.009507972747087479, + 0.007379693444818258, + 0.00414402037858963, + 0.005889229476451874, + 0.0046878657303750515, + 0.010982802137732506, + 0.01658676192164421, + 0.006543824914842844, + 0.016280435025691986, + 0.015604431740939617, + 0.005557759664952755, + 0.005699442699551582, + 0.015721848234534264, + 0.007802210748195648, + 0.00828910619020462, + 0.004994645714759827, + 0.010216186754405499, + 0.013430614024400711, + 0.007251869887113571, + 0.0056370701640844345, + 0.00818449817597866, + 0.0091482512652874, + 0.01762549951672554, + 0.013334687799215317, + 0.019711708649992943, + 0.030285384505987167, + 0.017455285415053368, + 0.014170683920383453, + 0.039814241230487823, + 0.020020971074700356, + 0.006509257014840841, + 0.00735029811039567, + 0.004686487838625908, + 0.004923438187688589, + 0.004385979380458593, + 0.0048804813995957375, + 0.005466359667479992, + 0.00505956681445241, + 0.027048151940107346, + 0.023819273337721825, + 0.018426213413476944, + 0.014441880397498608, + 0.056164439767599106, + 0.03807342052459717, + 0.016971465200185776, + 0.01505834236741066, + 0.00622588349506259, + 0.004836459644138813, + 0.006078251171857119, + 0.005583636462688446, + 0.00547364866361022, + 0.00498307915404439, + 0.0034255376085639, + 0.004329491872340441, + 0.008487265557050705, + 0.008568014949560165, + 0.0065547749400138855, + 0.006462522782385349, + 0.004716460593044758, + 0.007146838121116161, + 0.006670382339507341, + 0.005777870770543814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.09975051879882812 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008659549988806248 + ], + "opposite_logprob_tokens": [ + -4.758659362792969 + ], + "logprob_reward": -0.008659549988806248, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008659549988806248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38977333903312683, + 0.11029095947742462, + 0.08372453600168228, + 0.08707500249147415, + 0.0635313093662262, + 0.0397491492331028, + 0.03177304565906525, + 0.12400885671377182, + 0.037295710295438766, + 0.032778143882751465 + ], + "credit_map": [ + 0.056557390838861465, + 0.037213653326034546, + 0.08437647670507431, + 0.039214037358760834, + 0.028018640354275703, + 0.046643536537885666, + 0.04261522740125656, + 0.055134356021881104, + 0.008874865248799324, + 0.008895430713891983, + 0.017405834048986435, + 0.01463982556015253, + 0.008548295125365257, + 0.021211043000221252, + 0.01412957813590765, + 0.01658608950674534, + 0.01235915906727314, + 0.007122179958969355, + 0.012917484156787395, + 0.009497914463281631, + 0.005804089829325676, + 0.016516845673322678, + 0.006256114225834608, + 0.013250750489532948, + 0.013222305104136467, + 0.007569576613605022, + 0.014436564408242702, + 0.01361285150051117, + 0.005440115462988615, + 0.008748416788876057, + 0.01502637006342411, + 0.00901880394667387, + 0.005698165390640497, + 0.004172339104115963, + 0.010739952325820923, + 0.011244084686040878, + 0.005405839998275042, + 0.008948772214353085, + 0.007974215783178806, + 0.009347937069833279, + 0.00445836316794157, + 0.004230712074786425, + 0.0047785257920622826, + 0.008317243307828903, + 0.003948379773646593, + 0.004047675058245659, + 0.0055316430516541, + 0.004436609335243702, + 0.004303399007767439, + 0.0030168984085321426, + 0.004741752985864878, + 0.004123798571527004, + 0.002296507591381669, + 0.002800124231725931, + 0.0046045370399951935, + 0.005886028986424208, + 0.01633038930594921, + 0.012797973118722439, + 0.011186627671122551, + 0.013313598930835724, + 0.027119826525449753, + 0.01691887155175209, + 0.015692763030529022, + 0.010648800991475582, + 0.005563339218497276, + 0.0051356288604438305, + 0.00443347217515111, + 0.0032083664555102587, + 0.0064678858034312725, + 0.005786340218037367, + 0.0030990345403552055, + 0.00360164069570601, + 0.004982390906661749, + 0.005618355702608824, + 0.00396693404763937, + 0.00324775162152946, + 0.0038867658004164696, + 0.005346336867660284, + 0.0028156957123428583, + 0.0029139122925698757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.008659549988806248 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073797941207886 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073797941207886, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073797941207886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10238045454025269, + 0.06465679407119751, + 0.043333977460861206, + 0.1087847501039505, + 0.06105351075530052, + 0.12033938616514206, + 0.2239128202199936, + 0.13716071844100952, + 0.10757644474506378, + 0.030801130458712578 + ], + "credit_map": [ + 0.010478558018803596, + 0.00969784241169691, + 0.014925581403076649, + 0.014129403047263622, + 0.008233118802309036, + 0.01616295799612999, + 0.012720536440610886, + 0.016032448038458824, + 0.00458909384906292, + 0.007005419582128525, + 0.010016167536377907, + 0.009364171884953976, + 0.004978743847459555, + 0.011084377765655518, + 0.007931829430162907, + 0.009686989709734917, + 0.006538500543683767, + 0.0036705005913972855, + 0.007285516243427992, + 0.005620048847049475, + 0.003138158703222871, + 0.006480351090431213, + 0.004219099413603544, + 0.006381801329553127, + 0.015147072263062, + 0.009808412753045559, + 0.019356783479452133, + 0.01778363063931465, + 0.009484157897531986, + 0.00786061305552721, + 0.019112778827548027, + 0.01023130677640438, + 0.005707952659577131, + 0.0036966814659535885, + 0.009996180422604084, + 0.013399354182183743, + 0.005629571154713631, + 0.005952664650976658, + 0.009006877429783344, + 0.007664224598556757, + 0.016463864594697952, + 0.00918042566627264, + 0.01518702320754528, + 0.022851543501019478, + 0.011705665849149227, + 0.010176055133342743, + 0.021451588720083237, + 0.01332321111112833, + 0.03630850836634636, + 0.022361215204000473, + 0.03303035721182823, + 0.03055732138454914, + 0.022639518603682518, + 0.01765626296401024, + 0.030768167227506638, + 0.030591465532779694, + 0.015839945524930954, + 0.013659742660820484, + 0.01130410935729742, + 0.014228574931621552, + 0.026971489191055298, + 0.01989903301000595, + 0.019555198028683662, + 0.01570262759923935, + 0.017716508358716965, + 0.015061008743941784, + 0.013651395216584206, + 0.014706968329846859, + 0.01449841819703579, + 0.012602333910763264, + 0.010350261814892292, + 0.008989543654024601, + 0.004933049902319908, + 0.004369593691080809, + 0.003119841916486621, + 0.0029649296775460243, + 0.004384804982692003, + 0.004533343017101288, + 0.00339575600810349, + 0.003099810564890504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.011073797941207886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731215864419937 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006731215864419937, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731215864419937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07696652412414551, + 0.04509071633219719, + 0.033280838280916214, + 0.04252317547798157, + 0.11146578937768936, + 0.12391413003206253, + 0.09547919780015945, + 0.09161721169948578, + 0.32407623529434204, + 0.05558614432811737 + ], + "credit_map": [ + 0.010542686097323895, + 0.0057900515384972095, + 0.012475396506488323, + 0.01193653978407383, + 0.005828860215842724, + 0.008405817672610283, + 0.010783792473375797, + 0.011203384958207607, + 0.003557418240234256, + 0.003941856324672699, + 0.007280092220753431, + 0.007223520893603563, + 0.003889450104907155, + 0.008617270737886429, + 0.004086977802217007, + 0.006494125351309776, + 0.0034308198373764753, + 0.003673228435218334, + 0.00413548294454813, + 0.004975450225174427, + 0.003325794357806444, + 0.005292077548801899, + 0.004350412171334028, + 0.004097571596503258, + 0.004640253260731697, + 0.002961238846182823, + 0.006435162387788296, + 0.006913462653756142, + 0.0032189132180064917, + 0.005367075093090534, + 0.008127778768539429, + 0.004859293345361948, + 0.00829964317381382, + 0.009451881051063538, + 0.0190154779702425, + 0.022669920697808266, + 0.015277220867574215, + 0.0063444883562624454, + 0.018392520025372505, + 0.012014642357826233, + 0.012043154798448086, + 0.011179995723068714, + 0.013912037014961243, + 0.024172907695174217, + 0.011272275820374489, + 0.010076310485601425, + 0.027675556018948555, + 0.013581887818872929, + 0.015241021290421486, + 0.011594215407967567, + 0.0133574353531003, + 0.012788052670657635, + 0.010432098060846329, + 0.009700781665742397, + 0.01229425985366106, + 0.010071338154375553, + 0.010766078718006611, + 0.0100516676902771, + 0.008914753794670105, + 0.012629985809326172, + 0.008586038835346699, + 0.012168563902378082, + 0.01618201471865177, + 0.012318112887442112, + 0.04680853709578514, + 0.06744071841239929, + 0.02968405745923519, + 0.023200009018182755, + 0.05094793438911438, + 0.052702177315950394, + 0.02794330008327961, + 0.025349516421556473, + 0.008423211053013802, + 0.010125856846570969, + 0.005636473186314106, + 0.006107059307396412, + 0.0056943027302622795, + 0.007664747070521116, + 0.006304644979536533, + 0.005629851017147303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.006731215864419937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021949741058051586 + ], + "opposite_logprob_tokens": [ + -6.127194881439209 + ], + "logprob_reward": -0.0021949741058051586, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021949741058051586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17162872850894928, + 0.042536962777376175, + 0.04872570559382439, + 0.0628424808382988, + 0.11470954865217209, + 0.06099550798535347, + 0.08671963214874268, + 0.07225106656551361, + 0.24862778186798096, + 0.09096262603998184 + ], + "credit_map": [ + 0.03132200241088867, + 0.01980486512184143, + 0.02560310624539852, + 0.012679734267294407, + 0.015918800607323647, + 0.026607871055603027, + 0.014368548057973385, + 0.0253237783908844, + 0.004570446442812681, + 0.003887202125042677, + 0.008280070498585701, + 0.007634537760168314, + 0.0031142018269747496, + 0.005495613440871239, + 0.003793712705373764, + 0.00576117355376482, + 0.006954120472073555, + 0.004177629482001066, + 0.007385655771940947, + 0.005456957034766674, + 0.004201960284262896, + 0.009996658191084862, + 0.0034207922872155905, + 0.007131925318390131, + 0.00874748919159174, + 0.004724821541458368, + 0.011072113178670406, + 0.011770877987146378, + 0.005350478924810886, + 0.0048779891803860664, + 0.01081302110105753, + 0.005485686473548412, + 0.011141533963382244, + 0.00885045062750578, + 0.017891516909003258, + 0.025352712720632553, + 0.015123819001019001, + 0.009218377061188221, + 0.014374070800840855, + 0.012757052667438984, + 0.006647855043411255, + 0.00696958601474762, + 0.006979138124734163, + 0.011807690374553204, + 0.006874862127006054, + 0.006314113270491362, + 0.008954650722444057, + 0.006447602063417435, + 0.013757651671767235, + 0.012959226965904236, + 0.011794736608862877, + 0.009314962662756443, + 0.010144227184355259, + 0.012423854321241379, + 0.00884669553488493, + 0.0074782660230994225, + 0.010225838981568813, + 0.00996149517595768, + 0.006255501881241798, + 0.008192365057766438, + 0.007525582332164049, + 0.013456305488944054, + 0.008485709317028522, + 0.008148265071213245, + 0.03513335809111595, + 0.04971030727028847, + 0.020912112668156624, + 0.017902834340929985, + 0.04180208221077919, + 0.04543937370181084, + 0.020555879920721054, + 0.017171813175082207, + 0.014900377951562405, + 0.008829106576740742, + 0.008755495771765709, + 0.009185108356177807, + 0.013974986039102077, + 0.016024449840188026, + 0.010376476682722569, + 0.008916621096432209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0021949741058051586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive first-person point-of-view (POV) shot, the camera steadily advances at a slow, wistful pace through a narrow, damp alleyway imbued with a cyberpunk aesthetic. Rainwater, reflecting off multicolored neon lights, pools on the uneven concrete floor. Towering building walls flank the ground, a tangle of wires, rusting pipes, and torn holographic advertisements in foreign languages. The camera's focus finally settles on a simple barbecue stall deep in the alley. A thick steak, etched with a grid of caramelized grill marks, sizzles enticingly on the scorching metal grill, its rich fat melting and boiling under the intense heat. As the camera approaches, a strange phenomenon occurs: rising from the steak's charred crust is no longer ordinary steam, but rapidly condenses and transforms into a dazzling, upward-floating stream of holographic data composed of binary code, glowing circuit diagrams, and shimmering geometric symbols. This stream of digital steam, shimmering in neon blues, purples, and greens, appears both ethereal and alluring against the warm orange glow of the grill and the cool ambient light of the alley. The entire process is imbued with a possessive intensity, as if the diner's eyes are poised to devour this digital delicacy, a fusion of primitive cooking and futuristic technology, creating a psychedelic atmosphere of seeking sensory immersion in a decadent future.", + "seed": 6174, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_11/index_138_seed_6174_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03815138144418597, + "vlm_logprob": -0.03815138144418597, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.03815138144418597, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16959162056446075 + ], + "opposite_logprob_tokens": [ + -1.9195916652679443 + ], + "logprob_reward": -0.16959162056446075, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16959162056446075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14734090864658356, + 0.12369868159294128, + 0.0379304438829422, + 0.12021046876907349, + 0.12777312099933624, + 0.050328660756349564, + 0.15605591237545013, + 0.041792064905166626, + 0.12264098972082138, + 0.07222869992256165 + ], + "credit_map": [ + 0.01332702487707138, + 0.010073496960103512, + 0.01615600474178791, + 0.019587146118283272, + 0.023540688678622246, + 0.021063385531306267, + 0.01634552702307701, + 0.02724762260913849, + 0.011809185147285461, + 0.00860829558223486, + 0.01654762588441372, + 0.01608174480497837, + 0.013832594268023968, + 0.024593612179160118, + 0.01135594304651022, + 0.02086969092488289, + 0.004488009959459305, + 0.005521958228200674, + 0.003675110638141632, + 0.005390985868871212, + 0.004470711573958397, + 0.0052184658125042915, + 0.004221088718622923, + 0.004944114480167627, + 0.00880992691963911, + 0.008575606159865856, + 0.010187544859945774, + 0.012527598068118095, + 0.031584229320287704, + 0.016610847786068916, + 0.016150228679180145, + 0.015764478594064713, + 0.008283144794404507, + 0.012968805618584156, + 0.014313160441815853, + 0.013560847379267216, + 0.01363871805369854, + 0.022588608786463737, + 0.011120091192424297, + 0.03129974752664566, + 0.003848415333777666, + 0.004792702384293079, + 0.007888725027441978, + 0.009600009769201279, + 0.0061162798665463924, + 0.0077860611490905285, + 0.004605922382324934, + 0.0056905425153672695, + 0.012945955619215965, + 0.02080640196800232, + 0.016302788630127907, + 0.019997064024209976, + 0.010250946506857872, + 0.02812781184911728, + 0.013478903099894524, + 0.0341460295021534, + 0.004780496936291456, + 0.002963803941383958, + 0.0051087625324726105, + 0.004011362325400114, + 0.004847075790166855, + 0.0068089887499809265, + 0.0045783608220517635, + 0.008693214505910873, + 0.009181358851492405, + 0.005764090456068516, + 0.04409291222691536, + 0.014997092075645924, + 0.006013673264533281, + 0.006156107876449823, + 0.010918835178017616, + 0.025516925379633904, + 0.007856211625039577, + 0.007113605737686157, + 0.008225514553487301, + 0.008693338371813297, + 0.00876733846962452, + 0.008398600853979588, + 0.01283961534500122, + 0.010334469377994537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.16959162056446075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009810092858970165 + ], + "opposite_logprob_tokens": [ + -4.634809970855713 + ], + "logprob_reward": -0.009810092858970165, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009810092858970165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41132819652557373, + 0.13785074651241302, + 0.036895621567964554, + 0.04142969101667404, + 0.05346771702170372, + 0.029905477538704872, + 0.1075611412525177, + 0.022485801950097084, + 0.052804797887802124, + 0.10627076774835587 + ], + "credit_map": [ + 0.039972905069589615, + 0.03464552387595177, + 0.03597325086593628, + 0.04064292088150978, + 0.07130412757396698, + 0.07045930624008179, + 0.0494627021253109, + 0.06886748224496841, + 0.010438064113259315, + 0.009690511040389538, + 0.013360169716179371, + 0.022932574152946472, + 0.01677805557847023, + 0.022776909172534943, + 0.018955189734697342, + 0.022919289767742157, + 0.002508522244170308, + 0.002642606617882848, + 0.005121934227645397, + 0.005010772496461868, + 0.005837860517203808, + 0.006818369030952454, + 0.003902917494997382, + 0.005052641034126282, + 0.00419960729777813, + 0.007006110157817602, + 0.0048653301782906055, + 0.005640211049467325, + 0.004838529042899609, + 0.006146823987364769, + 0.004945273511111736, + 0.0037878125440329313, + 0.0045164842158555984, + 0.008621856570243835, + 0.010949291288852692, + 0.007002216763794422, + 0.004674848634749651, + 0.007591210771352053, + 0.004344039596617222, + 0.005767774302512407, + 0.0026561927516013384, + 0.003593015717342496, + 0.003729584626853466, + 0.0036448680330067873, + 0.004100713413208723, + 0.004441110882908106, + 0.00347428978420794, + 0.004265706054866314, + 0.014498024247586727, + 0.013519879430532455, + 0.010943328030407429, + 0.013550590723752975, + 0.01218994427472353, + 0.01726996898651123, + 0.012151249684393406, + 0.01343816053122282, + 0.0026395157910883427, + 0.001970130717381835, + 0.0028800468426197767, + 0.002171108266338706, + 0.0020383826922625303, + 0.0038533853366971016, + 0.002604734618216753, + 0.004328499082475901, + 0.004616674035787582, + 0.0022008244413882494, + 0.019282948225736618, + 0.0063452403992414474, + 0.0023820865899324417, + 0.002983772661536932, + 0.00474817119538784, + 0.010245089419186115, + 0.011123564094305038, + 0.010404180735349655, + 0.022099433466792107, + 0.01206898596137762, + 0.010943231172859669, + 0.011709822341799736, + 0.011578945443034172, + 0.01634262129664421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.009810092858970165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036185993812978268 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036185993812978268, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036185993812978268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14042770862579346, + 0.052808407694101334, + 0.05298371985554695, + 0.07594313472509384, + 0.15809762477874756, + 0.05830259248614311, + 0.28901052474975586, + 0.0565277524292469, + 0.04752769693732262, + 0.06837080419063568 + ], + "credit_map": [ + 0.013196401298046112, + 0.009972263127565384, + 0.010962294414639473, + 0.01874244026839733, + 0.01739298552274704, + 0.026300672441720963, + 0.015636412426829338, + 0.028224235400557518, + 0.00401686318218708, + 0.003420019755139947, + 0.0038451235741376877, + 0.006184868514537811, + 0.007824111729860306, + 0.009680110961198807, + 0.00877918116748333, + 0.00905812717974186, + 0.0035973438061773777, + 0.004446578677743673, + 0.003821656806394458, + 0.006314033642411232, + 0.009182814508676529, + 0.010917105711996555, + 0.006559453904628754, + 0.008144734427332878, + 0.007168481592088938, + 0.00798026379197836, + 0.009989399462938309, + 0.010721774771809578, + 0.010821178555488586, + 0.009921543300151825, + 0.011418870650231838, + 0.00792162399739027, + 0.014789492823183537, + 0.019932497292757034, + 0.02982959896326065, + 0.017463410273194313, + 0.016813460737466812, + 0.02278723381459713, + 0.010341158136725426, + 0.026140769943594933, + 0.006126911845058203, + 0.008748773485422134, + 0.00829935260117054, + 0.007208100985735655, + 0.0076529341749846935, + 0.00820537842810154, + 0.005243772175163031, + 0.006817372050136328, + 0.03461616486310959, + 0.05323254317045212, + 0.0236927829682827, + 0.03930594027042389, + 0.028180984780192375, + 0.04517605900764465, + 0.025346137583255768, + 0.039459895342588425, + 0.005631411448121071, + 0.005476113874465227, + 0.00550070172175765, + 0.004627682268619537, + 0.005000937730073929, + 0.01162893045693636, + 0.005522055551409721, + 0.01313991378992796, + 0.00434734346345067, + 0.0036254532169550657, + 0.014733143150806427, + 0.005620110780000687, + 0.003521089442074299, + 0.0030142611358314753, + 0.005013782065361738, + 0.007652511820197105, + 0.005996860098093748, + 0.007813005708158016, + 0.010454731062054634, + 0.009918098337948322, + 0.007935179397463799, + 0.006788627710193396, + 0.010617374442517757, + 0.008846928365528584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.0036185993812978268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249170120805502 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005249170120805502, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249170120805502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08277294039726257, + 0.09970522671937943, + 0.03390028700232506, + 0.05560431256890297, + 0.08587507903575897, + 0.05135536193847656, + 0.08174151182174683, + 0.21692466735839844, + 0.18791291117668152, + 0.10420773923397064 + ], + "credit_map": [ + 0.006410718429833651, + 0.006179718766361475, + 0.008910315111279488, + 0.015063067898154259, + 0.008532015606760979, + 0.011865285225212574, + 0.009905685670673847, + 0.01590612716972828, + 0.007120337802916765, + 0.008170797489583492, + 0.010213781148195267, + 0.016000114381313324, + 0.009743536822497845, + 0.017805788666009903, + 0.013879578560590744, + 0.01677127368748188, + 0.002559330780059099, + 0.0026996340602636337, + 0.003697263076901436, + 0.004223324824124575, + 0.005814290139824152, + 0.006436726078391075, + 0.003534279763698578, + 0.004935439210385084, + 0.0036847407463937998, + 0.0073141902685165405, + 0.004318549297749996, + 0.005406778771430254, + 0.011936862953007221, + 0.008101725950837135, + 0.00693821907043457, + 0.007903236895799637, + 0.008181392215192318, + 0.011238002218306065, + 0.015792211517691612, + 0.009502576664090157, + 0.011660505086183548, + 0.012674618512392044, + 0.006790815852582455, + 0.010034948587417603, + 0.006492273882031441, + 0.004940554033964872, + 0.008068472146987915, + 0.005884617567062378, + 0.004257330670952797, + 0.008219383656978607, + 0.005503954831510782, + 0.007988771423697472, + 0.011569079011678696, + 0.01579112373292446, + 0.007069635670632124, + 0.011221373453736305, + 0.007229267619550228, + 0.01232103817164898, + 0.0074416399002075195, + 0.009098352864384651, + 0.016589784994721413, + 0.016654429957270622, + 0.020280679687857628, + 0.021389959380030632, + 0.017109394073486328, + 0.04736483469605446, + 0.02355620637536049, + 0.053979359567165375, + 0.013900449499487877, + 0.00849974900484085, + 0.06644992530345917, + 0.02174166962504387, + 0.009081410244107246, + 0.01253858394920826, + 0.016377389430999756, + 0.039323702454566956, + 0.01369904913008213, + 0.009286027401685715, + 0.014637810178101063, + 0.014370927587151527, + 0.013318520970642567, + 0.013938622549176216, + 0.013791196048259735, + 0.011165580712258816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.005249170120805502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person point-of-view (POV) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily advance through a narrow, damp alleyway with cyberpunk aesthetics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a barbecue stall visible deep in the alley with a thick steak sizzling on a metal grill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does steam rising from the steak transform into a stream of holographic data composed of binary code and glowing symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the holographic data stream shimmering in neon blues, purples, and greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024874242953956127 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024874242953956127, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024874242953956127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25593677163124084, + 0.0807117149233818, + 0.030253024771809578, + 0.07554039359092712, + 0.0627003163099289, + 0.04599711298942566, + 0.11616623401641846, + 0.07444059103727341, + 0.0557534359395504, + 0.2025003880262375 + ], + "credit_map": [ + 0.014420081861317158, + 0.019491208717226982, + 0.02014497108757496, + 0.030422383919358253, + 0.042636267840862274, + 0.04826105386018753, + 0.028674291446805, + 0.05188652500510216, + 0.004440251272171736, + 0.006628596689552069, + 0.006499485578387976, + 0.011791945435106754, + 0.00982289295643568, + 0.0164302047342062, + 0.011281104758381844, + 0.013817232102155685, + 0.002239258959889412, + 0.0041993893682956696, + 0.001900061615742743, + 0.0023708553053438663, + 0.0038811967242509127, + 0.00777674512937665, + 0.003118546912446618, + 0.00476697226986289, + 0.003392576240003109, + 0.005981398746371269, + 0.004801117815077305, + 0.00648629292845726, + 0.02250327356159687, + 0.011991473846137524, + 0.007635400630533695, + 0.012748868204653263, + 0.007852302864193916, + 0.010666051879525185, + 0.009213563986122608, + 0.008657560683786869, + 0.0058344281278550625, + 0.010746785439550877, + 0.0036709539126604795, + 0.00605867151170969, + 0.005245961714535952, + 0.0074506402015686035, + 0.005321252625435591, + 0.007053188979625702, + 0.006037886720150709, + 0.0065260897390544415, + 0.003736978629603982, + 0.0046251192688941956, + 0.015557777136564255, + 0.02089243195950985, + 0.011203179135918617, + 0.013954311609268188, + 0.01391419768333435, + 0.018001792952418327, + 0.010584983974695206, + 0.012057565152645111, + 0.006742697674781084, + 0.006092352326959372, + 0.011118761263787746, + 0.007217563223093748, + 0.005735166370868683, + 0.012143428437411785, + 0.0074348836205899715, + 0.01795574277639389, + 0.005965117830783129, + 0.008060208521783352, + 0.009090273641049862, + 0.006426999811083078, + 0.006206557620316744, + 0.005966235883533955, + 0.006154244765639305, + 0.007883798331022263, + 0.02163836918771267, + 0.02096916362643242, + 0.040249988436698914, + 0.025418633595108986, + 0.023657387122511864, + 0.025456972420215607, + 0.018712567165493965, + 0.02639732137322426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.0024874242953956127 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_12/index_tensor([2476, 2476])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_12/index_tensor([2476, 2476])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5eef5ed57c9e12341248172ae9550c6409b9b8e1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_12/index_tensor([2476, 2476])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, epic wide-angle shot frames a massive stone arch, its massive form constructed from weathered, sand-colored stone, etched with faint, eroded patterns of ancient carvings. The entire structure appears as a stark, dark silhouette against the blazing, low-hanging afternoon sun, with the background filled with an intense golden glow. Beyond the arch, a vast, sun-drenched desert landscape unfolds, with rolling golden dunes stretching to a hazy, distant horizon. A lone figure, also rendered as a sharp silhouette, defined by the shape of a long cloak gently fluttering in the hot wind, walks away from the camera with a steady, purposeful pace. The figure passes directly through the center of the towering arch, their form gradually shrinking as they descend into the vast, light-saturated expanse. The scene, dominated by a warm palette of fiery oranges and deep yellows, creates a powerful atmosphere of profound solitude, epic journey, and ancient mystery before the entire image gently and smoothly fades to complete black.", + "seed": 19988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_12/index_2476_seed_19988_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14596886839717627, + "vlm_logprob": -0.14596886839717627, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.14596886839717627, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0427398607134819 + ], + "opposite_logprob_tokens": [ + -3.2927398681640625 + ], + "logprob_reward": -0.0427398607134819, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0427398607134819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09160815179347992, + 0.07253783941268921, + 0.026482203975319862, + 0.14993028342723846, + 0.07122731953859329, + 0.0262684915214777, + 0.16138777136802673, + 0.07832630723714828, + 0.08540146052837372, + 0.23683017492294312 + ], + "credit_map": [ + 0.013428892940282822, + 0.008093839511275291, + 0.01035712193697691, + 0.007230258081108332, + 0.011501908302307129, + 0.018025748431682587, + 0.008715533651411533, + 0.0142548568546772, + 0.012635287828743458, + 0.007332424633204937, + 0.007236505392938852, + 0.007037820760160685, + 0.005678774788975716, + 0.014419292099773884, + 0.00790735986083746, + 0.010290375910699368, + 0.0034192348830401897, + 0.0032293268013745546, + 0.002973458031192422, + 0.0018517837161198258, + 0.001978930551558733, + 0.007309865206480026, + 0.002172847045585513, + 0.003546759020537138, + 0.015569956973195076, + 0.012818819843232632, + 0.018807366490364075, + 0.013899587094783783, + 0.012948580086231232, + 0.0398932546377182, + 0.01254114881157875, + 0.023451579734683037, + 0.012486674822866917, + 0.006300436798483133, + 0.009519609622657299, + 0.0060599432326853275, + 0.006195394787937403, + 0.013609711080789566, + 0.006962918676435947, + 0.010092626325786114, + 0.004002267960458994, + 0.00205018138512969, + 0.003330956446006894, + 0.002923257416114211, + 0.0025516238529235125, + 0.004623947665095329, + 0.0031481818296015263, + 0.003638074267655611, + 0.027420472353696823, + 0.01521369069814682, + 0.02031783200800419, + 0.014838541857898235, + 0.01731068268418312, + 0.02955769933760166, + 0.016493283212184906, + 0.020235568284988403, + 0.008153808303177357, + 0.004068708047270775, + 0.009715380147099495, + 0.011011138558387756, + 0.0067922682501375675, + 0.00807451643049717, + 0.018124239519238472, + 0.012386242859065533, + 0.009102532640099525, + 0.009719453752040863, + 0.009045097045600414, + 0.010502302087843418, + 0.010266225785017014, + 0.010698121972382069, + 0.013904079794883728, + 0.012163650244474411, + 0.015982910990715027, + 0.02333671972155571, + 0.024940429255366325, + 0.03116626664996147, + 0.024616500362753868, + 0.03208019956946373, + 0.04891468212008476, + 0.03579246252775192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.0427398607134819 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25214454531669617 + ], + "opposite_logprob_tokens": [ + -1.5021445751190186 + ], + "logprob_reward": -0.25214454531669617, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25214454531669617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11473463475704193, + 0.07523830235004425, + 0.050615571439266205, + 0.2046571969985962, + 0.08490881323814392, + 0.032937273383140564, + 0.24481366574764252, + 0.04208752140402794, + 0.02315264753997326, + 0.1268543303012848 + ], + "credit_map": [ + 0.014927146956324577, + 0.010200563818216324, + 0.012902752496302128, + 0.009975709952414036, + 0.012213710695505142, + 0.025054151192307472, + 0.011372038163244724, + 0.01808856800198555, + 0.01394517719745636, + 0.00635968241840601, + 0.00837200041860342, + 0.007877049036324024, + 0.005781722720712423, + 0.012633359991014004, + 0.009407898411154747, + 0.010861421003937721, + 0.006665737833827734, + 0.0038539250381290913, + 0.005992643069475889, + 0.005187800619751215, + 0.0058480012230575085, + 0.012048361822962761, + 0.004475132096558809, + 0.006543975789099932, + 0.026623640209436417, + 0.017306871712207794, + 0.023896431550383568, + 0.015965629369020462, + 0.015942681580781937, + 0.05907639488577843, + 0.015159317292273045, + 0.03068624809384346, + 0.015157810412347317, + 0.00722298352047801, + 0.0106875691562891, + 0.008254330605268478, + 0.008265772834420204, + 0.01475930493324995, + 0.008515998721122742, + 0.012045050039887428, + 0.005104981828480959, + 0.0027736749034374952, + 0.004841409623622894, + 0.003298436990007758, + 0.0029122703708708286, + 0.006547464523464441, + 0.002926967805251479, + 0.004532070364803076, + 0.04122399166226387, + 0.023740943521261215, + 0.030119547620415688, + 0.023977143689990044, + 0.028019241988658905, + 0.041823405772447586, + 0.02517332322895527, + 0.030736088752746582, + 0.005825833883136511, + 0.0032693699467927217, + 0.004888220224529505, + 0.005383079871535301, + 0.003881680080667138, + 0.004621980246156454, + 0.00853198952972889, + 0.005685369484126568, + 0.002479258691892028, + 0.0014341636560857296, + 0.0026619452983140945, + 0.002264157636091113, + 0.0024011998903006315, + 0.004974175710231066, + 0.004209503531455994, + 0.002728243125602603, + 0.00787881575524807, + 0.01167831290513277, + 0.012772487476468086, + 0.014798145741224289, + 0.011785803362727165, + 0.018233565613627434, + 0.02467549778521061, + 0.025031715631484985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.25214454531669617 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811084106564522 + ], + "opposite_logprob_tokens": [ + -3.2881107330322266 + ], + "logprob_reward": -0.03811084106564522, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811084106564522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033494625240564346, + 0.01884486712515354, + 0.023137956857681274, + 0.027162795886397362, + 0.018739452585577965, + 0.009368105791509151, + 0.009040064178407192, + 0.03425706923007965, + 0.16675344109535217, + 0.6592016220092773 + ], + "credit_map": [ + 0.0047224354930222034, + 0.003515549935400486, + 0.0039399368688464165, + 0.0033650582190603018, + 0.00333614950068295, + 0.0067974417470395565, + 0.002756575122475624, + 0.0050614806823432446, + 0.00253009214065969, + 0.0017001229571178555, + 0.001930959988385439, + 0.0025443637277930975, + 0.0018743908731266856, + 0.0020966867450624704, + 0.0032353061251342297, + 0.0029329454991966486, + 0.002126551466062665, + 0.002056584693491459, + 0.0023474502377212048, + 0.0025006828363984823, + 0.0022233964409679174, + 0.004997518379241228, + 0.0028870333917438984, + 0.0039987387135624886, + 0.0041657742112874985, + 0.0023910386953502893, + 0.0027952527161687613, + 0.002929889364168048, + 0.0030580610036849976, + 0.006201897282153368, + 0.002072179690003395, + 0.003548703622072935, + 0.0028640925884246826, + 0.0017144730081781745, + 0.0024846694432199, + 0.0019901669584214687, + 0.0021071091759949923, + 0.0027031395584344864, + 0.002076973207294941, + 0.0027988292276859283, + 0.001449850620701909, + 0.0008137916447594762, + 0.001244267332367599, + 0.0012800416443496943, + 0.0008575284155085683, + 0.0012158437166363, + 0.0010861002374440432, + 0.0014206821797415614, + 0.001423080451786518, + 0.0009311052272096276, + 0.0012301658280193806, + 0.0009095814311876893, + 0.0010045220842584968, + 0.0016519796336069703, + 0.0007108395802788436, + 0.0011787903495132923, + 0.002199533162638545, + 0.00273509556427598, + 0.0046884589828550816, + 0.0041127391159534454, + 0.0039949011988937855, + 0.003372220555320382, + 0.0071419537998735905, + 0.006012164521962404, + 0.016462285071611404, + 0.01917274296283722, + 0.017065685242414474, + 0.020520659163594246, + 0.020576998591423035, + 0.022862259298563004, + 0.024936025962233543, + 0.02515677735209465, + 0.04523894563317299, + 0.06379865854978561, + 0.06951476633548737, + 0.08627493679523468, + 0.06770474463701248, + 0.09264589846134186, + 0.13733209669589996, + 0.0966915637254715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.03811084106564522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783296845853329 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783296845853329, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783296845853329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07721678167581558, + 0.05239022523164749, + 0.0574478954076767, + 0.060833707451820374, + 0.026002444326877594, + 0.02386067993938923, + 0.039054106920957565, + 0.10470230132341385, + 0.2651418447494507, + 0.2933500111103058 + ], + "credit_map": [ + 0.011462973430752754, + 0.008295265026390553, + 0.00955390278249979, + 0.006510698702186346, + 0.009062808938324451, + 0.011814028024673462, + 0.008095731027424335, + 0.012421383522450924, + 0.007421253714710474, + 0.006198502611368895, + 0.00577908381819725, + 0.006167108658701181, + 0.006618749815970659, + 0.007886302657425404, + 0.00511745223775506, + 0.007201770786195993, + 0.005500743631273508, + 0.007259204052388668, + 0.005462197586894035, + 0.005309486761689186, + 0.008320759981870651, + 0.010610207915306091, + 0.006628045812249184, + 0.008357251062989235, + 0.008231171406805515, + 0.005458814091980457, + 0.006149549037218094, + 0.007857728749513626, + 0.008085163310170174, + 0.009829233400523663, + 0.00635745283216238, + 0.00886459555476904, + 0.0032203327864408493, + 0.002755270339548588, + 0.0034515687730163336, + 0.0030714133754372597, + 0.0028654320631176233, + 0.003987236879765987, + 0.0028455150313675404, + 0.0038056757766753435, + 0.003192817559465766, + 0.0022805731277912855, + 0.0032150510232895613, + 0.0028335966635495424, + 0.002057628706097603, + 0.004392008762806654, + 0.0026123884599655867, + 0.003276616334915161, + 0.0036779334768652916, + 0.003841278376057744, + 0.005319890100508928, + 0.004699693527072668, + 0.004378697834908962, + 0.0074018388986587524, + 0.00405713077634573, + 0.00567764462903142, + 0.011017696000635624, + 0.006458688527345657, + 0.0122687304392457, + 0.015268788672983646, + 0.008090897463262081, + 0.012068355455994606, + 0.023982279002666473, + 0.01554686389863491, + 0.02515542507171631, + 0.029549647122621536, + 0.026790058240294456, + 0.033218443393707275, + 0.0326700434088707, + 0.035555399954319, + 0.04126766324043274, + 0.040935155004262924, + 0.02105168253183365, + 0.027925852686166763, + 0.029372885823249817, + 0.03604952245950699, + 0.028432972729206085, + 0.0523250550031662, + 0.05378368869423866, + 0.04440835863351822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.009783296845853329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870657980442047 + ], + "opposite_logprob_tokens": [ + -1.1370657682418823 + ], + "logprob_reward": -0.3870657980442047, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870657980442047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19807451963424683, + 0.14785560965538025, + 0.06165670230984688, + 0.07727887481451035, + 0.08246899396181107, + 0.059581954032182693, + 0.029623614624142647, + 0.1287022978067398, + 0.06649472564458847, + 0.14826266467571259 + ], + "credit_map": [ + 0.022373812273144722, + 0.024759050458669662, + 0.021511586382985115, + 0.02050287276506424, + 0.02257758006453514, + 0.04089251905679703, + 0.018694529309868813, + 0.026762573048472404, + 0.015815522521734238, + 0.015439065173268318, + 0.015306523069739342, + 0.016023024916648865, + 0.011590473353862762, + 0.02843359299004078, + 0.020000101998448372, + 0.025247303768992424, + 0.006352785974740982, + 0.005732009653002024, + 0.006805662997066975, + 0.0051230196841061115, + 0.005747696850448847, + 0.017221318557858467, + 0.005689162295311689, + 0.008985050022602081, + 0.015249793417751789, + 0.007617587689310312, + 0.007850674912333488, + 0.008504468947649002, + 0.008825574070215225, + 0.01039426401257515, + 0.007588617503643036, + 0.01124789658933878, + 0.012944010086357594, + 0.00729424599558115, + 0.00928141362965107, + 0.00998355820775032, + 0.007894588634371758, + 0.012037398293614388, + 0.011228826828300953, + 0.011804950423538685, + 0.005839490797370672, + 0.004912881646305323, + 0.006867904216051102, + 0.009264388121664524, + 0.004652139265090227, + 0.00689656799659133, + 0.01255810633301735, + 0.008590473793447018, + 0.0031017325818538666, + 0.0024372488260269165, + 0.004364499356597662, + 0.003528995905071497, + 0.003309789812192321, + 0.005973633844405413, + 0.0027695870958268642, + 0.004138126503676176, + 0.009974710643291473, + 0.008414585143327713, + 0.017295807600021362, + 0.0170616302639246, + 0.012082758359611034, + 0.012182018719613552, + 0.029123974964022636, + 0.022566812112927437, + 0.005671233404427767, + 0.006086856592446566, + 0.007345200981944799, + 0.00775757385417819, + 0.007126298267394304, + 0.009831010363996029, + 0.01197047159075737, + 0.010706081055104733, + 0.009791279211640358, + 0.014949188567698002, + 0.0164121575653553, + 0.019571570679545403, + 0.014125616289675236, + 0.024080028757452965, + 0.027573546394705772, + 0.0217592790722847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.3870657980442047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, epic wide-angle shot frames a massive stone arch, its massive form constructed from weathered, sand-colored stone, etched with faint, eroded patterns of ancient carvings. The entire structure appears as a stark, dark silhouette against the blazing, low-hanging afternoon sun, with the background filled with an intense golden glow. Beyond the arch, a vast, sun-drenched desert landscape unfolds, with rolling golden dunes stretching to a hazy, distant horizon. A lone figure, also rendered as a sharp silhouette, defined by the shape of a long cloak gently fluttering in the hot wind, walks away from the camera with a steady, purposeful pace. The figure passes directly through the center of the towering arch, their form gradually shrinking as they descend into the vast, light-saturated expanse. The scene, dominated by a warm palette of fiery oranges and deep yellows, creates a powerful atmosphere of profound solitude, epic journey, and ancient mystery before the entire image gently and smoothly fades to complete black.", + "seed": 19989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_12/index_2476_seed_19989_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0477015377022326, + "vlm_logprob": -0.0477015377022326, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.0477015377022326, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021861055865883827 + ], + "opposite_logprob_tokens": [ + -4.2718610763549805 + ], + "logprob_reward": -0.021861055865883827, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.021861055865883827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0316707007586956, + 0.017294952645897865, + 0.013928315602242947, + 0.009442348033189774, + 0.00885995663702488, + 0.11774007230997086, + 0.2074122577905655, + 0.2715342342853546, + 0.098112091422081, + 0.2240051031112671 + ], + "credit_map": [ + 0.005437955260276794, + 0.0032379846088588238, + 0.0029588197357952595, + 0.006750407163053751, + 0.003122086636722088, + 0.002486178185790777, + 0.0034435796551406384, + 0.004233688581734896, + 0.0032111501786857843, + 0.001709133037365973, + 0.002049950882792473, + 0.0023915276397019625, + 0.0024309479631483555, + 0.0012051318772137165, + 0.002446936210617423, + 0.001850175322033465, + 0.002363664796575904, + 0.0019642983097583055, + 0.0021512655075639486, + 0.0017204697942361236, + 0.0012432822259142995, + 0.0016918707406148314, + 0.0012482557212933898, + 0.0015452089719474316, + 0.0018103726906701922, + 0.000829878612421453, + 0.0013507022522389889, + 0.001621201285161078, + 0.0011234889971092343, + 0.0007937105256132782, + 0.0008709391695447266, + 0.0010420550825074315, + 0.0012712521711364388, + 0.0007638842798769474, + 0.0009472858509980142, + 0.0018042402807623148, + 0.0011557386023923755, + 0.0012297326466068625, + 0.0007263711886480451, + 0.000961451674811542, + 0.023872781544923782, + 0.015617250464856625, + 0.009033530950546265, + 0.01099073514342308, + 0.020708847790956497, + 0.009908417239785194, + 0.01846587471663952, + 0.009142628870904446, + 0.0040824199095368385, + 0.004790952429175377, + 0.0022969755809754133, + 0.0034415251575410366, + 0.09305310994386673, + 0.0037938260938972235, + 0.09224274009466171, + 0.00371069204993546, + 0.009835252538323402, + 0.08825980126857758, + 0.009840586222708225, + 0.07798699289560318, + 0.04217040166258812, + 0.006587080657482147, + 0.03173300251364708, + 0.005121110938489437, + 0.009301678277552128, + 0.011491194367408752, + 0.020871127024292946, + 0.015095850452780724, + 0.016060171648859978, + 0.008396884426474571, + 0.010587538592517376, + 0.006307643372565508, + 0.02148343250155449, + 0.021421553567051888, + 0.02057921141386032, + 0.01907208003103733, + 0.055279769003391266, + 0.025740481913089752, + 0.021386047825217247, + 0.03904252499341965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.021861055865883827 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994432151317596 + ], + "opposite_logprob_tokens": [ + -1.8049442768096924 + ], + "logprob_reward": -0.17994432151317596, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994432151317596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10159336775541306, + 0.0355742983520031, + 0.03481539711356163, + 0.016721511259675026, + 0.0103662870824337, + 0.11123322695493698, + 0.1320498287677765, + 0.09312532097101212, + 0.1665448099374771, + 0.29797589778900146 + ], + "credit_map": [ + 0.02042405866086483, + 0.01401778869330883, + 0.007701800670474768, + 0.017571834847331047, + 0.010072855278849602, + 0.0068698786199092865, + 0.010176241397857666, + 0.014758911915123463, + 0.008163206279277802, + 0.0042882454581558704, + 0.004597566090524197, + 0.004944545682519674, + 0.0026688207872211933, + 0.003047599922865629, + 0.0037034559063613415, + 0.0041608563624322414, + 0.006054442375898361, + 0.0050946190021932125, + 0.005635218229144812, + 0.004588843323290348, + 0.0025610486045479774, + 0.0035771424882113934, + 0.00366938766092062, + 0.0036346970591694117, + 0.0023264179471880198, + 0.0010753186652436852, + 0.0011612663511186838, + 0.0019039255566895008, + 0.004515463951975107, + 0.001076531712897122, + 0.002929654438048601, + 0.001732931938022375, + 0.0015618109609931707, + 0.0007780952728353441, + 0.001319150673225522, + 0.0015859309351071715, + 0.001798564800992608, + 0.001173454336822033, + 0.0011441413080319762, + 0.0010051388526335359, + 0.01551622711122036, + 0.013171104714274406, + 0.01068943366408348, + 0.015371306799352169, + 0.017095638439059258, + 0.00977263879030943, + 0.017032882198691368, + 0.012584005482494831, + 0.010388405993580818, + 0.018784895539283752, + 0.005518273916095495, + 0.008294938132166862, + 0.03368345648050308, + 0.009461586363613605, + 0.03330542519688606, + 0.012612841092050076, + 0.005230945535004139, + 0.015168179757893085, + 0.007073862478137016, + 0.02761460281908512, + 0.009414641186594963, + 0.004133197944611311, + 0.021641051396727562, + 0.0028488400857895613, + 0.01252670492976904, + 0.021159600466489792, + 0.052641287446022034, + 0.02269863337278366, + 0.023940103128552437, + 0.009807535447180271, + 0.01808220148086548, + 0.005688738543540239, + 0.029104378074407578, + 0.030344119295477867, + 0.03153840824961662, + 0.026897156611084938, + 0.07038156688213348, + 0.03147701174020767, + 0.02965431660413742, + 0.04857892543077469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.17994432151317596 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789789095520973 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.009789789095520973, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789789095520973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12793214619159698, + 0.04065941646695137, + 0.02466319501399994, + 0.014056388288736343, + 0.019635915756225586, + 0.048789843916893005, + 0.2872239649295807, + 0.20222651958465576, + 0.101070836186409, + 0.1337418407201767 + ], + "credit_map": [ + 0.02139716036617756, + 0.01470339298248291, + 0.013561833649873734, + 0.02348492480814457, + 0.01199994795024395, + 0.007754302583634853, + 0.019461210817098618, + 0.015569351613521576, + 0.006679980084300041, + 0.004796131514012814, + 0.0042050424963235855, + 0.00595488864928484, + 0.005035401321947575, + 0.0032094584312289953, + 0.005836828611791134, + 0.0049416846595704556, + 0.003519402351230383, + 0.002730396343395114, + 0.0027092837262898684, + 0.0033058219123631716, + 0.0027814405038952827, + 0.0023276323918253183, + 0.0044790576212108135, + 0.0028101576026529074, + 0.0017729237442836165, + 0.0008830608567222953, + 0.0012603981886059046, + 0.001886200625449419, + 0.003945120144635439, + 0.0006684199906885624, + 0.002181627321988344, + 0.0014586355537176132, + 0.0031174796167761087, + 0.0014381039654836059, + 0.0031864980701357126, + 0.002926908666267991, + 0.0022537356708198786, + 0.0018134746933355927, + 0.002987863263115287, + 0.0019118492491543293, + 0.008162930607795715, + 0.004623423330485821, + 0.00650058826431632, + 0.005926608107984066, + 0.007067241705954075, + 0.005125865805894136, + 0.007131693884730339, + 0.004251485224813223, + 0.010884270071983337, + 0.016239633783698082, + 0.0055462634190917015, + 0.00869976170361042, + 0.1083826869726181, + 0.015716983005404472, + 0.10571807622909546, + 0.016036273911595345, + 0.012610813602805138, + 0.04191840812563896, + 0.013731247745454311, + 0.05911222845315933, + 0.0269189290702343, + 0.009486998431384563, + 0.02853827364742756, + 0.009909594431519508, + 0.007028809282928705, + 0.009482921101152897, + 0.01538590993732214, + 0.017466343939304352, + 0.01205845084041357, + 0.009137548506259918, + 0.023735523223876953, + 0.006775314919650555, + 0.011837773025035858, + 0.012859469279646873, + 0.01372099295258522, + 0.013747291639447212, + 0.03363320976495743, + 0.014086103066802025, + 0.013328365050256252, + 0.020528636872768402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.009789789095520973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329261600971222 + ], + "opposite_logprob_tokens": [ + -3.7732925415039062 + ], + "logprob_reward": -0.02329261600971222, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329261600971222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05437672883272171, + 0.03618120774626732, + 0.04499804228544235, + 0.016037752851843834, + 0.034411683678627014, + 0.11872308701276779, + 0.10576551407575607, + 0.19991396367549896, + 0.3060230314731598, + 0.0835690125823021 + ], + "credit_map": [ + 0.006083509419113398, + 0.006307948846369982, + 0.006774671841412783, + 0.012008042074739933, + 0.004874037113040686, + 0.004419930279254913, + 0.00660034641623497, + 0.007308242376893759, + 0.005914219189435244, + 0.0033135944977402687, + 0.003478515660390258, + 0.005365286022424698, + 0.0055136894807219505, + 0.002237359993159771, + 0.006134773604571819, + 0.0042237695306539536, + 0.008326132781803608, + 0.005775209050625563, + 0.0058149010874331, + 0.006196598056703806, + 0.005096292123198509, + 0.0033940994180738926, + 0.005461628083139658, + 0.004933181218802929, + 0.0020693503320217133, + 0.0009086350328288972, + 0.0011919134994968772, + 0.002739368937909603, + 0.0038558742962777615, + 0.0012964671477675438, + 0.0016483200015500188, + 0.002327823778614402, + 0.005895482376217842, + 0.0026467591524124146, + 0.006904012057930231, + 0.005945069249719381, + 0.004235826898366213, + 0.0023719784803688526, + 0.003108152886852622, + 0.003304404206573963, + 0.01532373670488596, + 0.01097937487065792, + 0.01487252488732338, + 0.012508162297308445, + 0.019682640209794044, + 0.014390256255865097, + 0.020274145528674126, + 0.01069224439561367, + 0.015843521803617477, + 0.015924733132123947, + 0.010706698521971703, + 0.009197387844324112, + 0.018274132162332535, + 0.010932499542832375, + 0.01612234301865101, + 0.008764202706515789, + 0.024071423336863518, + 0.040945567190647125, + 0.01894446834921837, + 0.03039688616991043, + 0.026214804500341415, + 0.017804916948080063, + 0.030233772471547127, + 0.011302117258310318, + 0.020967034623026848, + 0.03339756652712822, + 0.04007182642817497, + 0.055270906537771225, + 0.0398721843957901, + 0.025826895609498024, + 0.0749698355793953, + 0.015646759420633316, + 0.005305878352373838, + 0.007170557044446468, + 0.022255271673202515, + 0.01134725846350193, + 0.01524780411273241, + 0.006441222038120031, + 0.010393975302577019, + 0.005407051648944616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.02329261600971222 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036199060268700123 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.0036199060268700123, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036199060268700123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012675803154706955, + 0.006107529159635305, + 0.0021812983322888613, + 0.003138968488201499, + 0.0034347150940448046, + 0.021259548142552376, + 0.10564513504505157, + 0.21031124889850616, + 0.18342824280261993, + 0.45181748270988464 + ], + "credit_map": [ + 0.0020378069020807743, + 0.0009780903346836567, + 0.0014050848549231887, + 0.0016794989351183176, + 0.0016578969080001116, + 0.001276580966077745, + 0.00201106327585876, + 0.001629781094379723, + 0.0007171063334681094, + 0.0004952865419909358, + 0.0005908085149712861, + 0.0009420626447536051, + 0.0009113571140915155, + 0.0005682333139702678, + 0.0011555772507563233, + 0.0007270975038409233, + 0.0004004405636806041, + 0.000244379771174863, + 0.0002999069693032652, + 0.0002776823821477592, + 0.00022319545678328723, + 0.00021452311193570495, + 0.0002962955040857196, + 0.0002248745149699971, + 0.0005218901787884533, + 0.0002492011117283255, + 0.00043197599006816745, + 0.0005714220460504293, + 0.0004551948222797364, + 0.0002553461818024516, + 0.0002523746225051582, + 0.0004015635640826076, + 0.000550312630366534, + 0.0002737409668043256, + 0.0004478651680983603, + 0.00042457605013623834, + 0.000611824681982398, + 0.00040580806671641767, + 0.00043203873792663217, + 0.00028854887932538986, + 0.005165416747331619, + 0.002678546356037259, + 0.002010165713727474, + 0.001696442486718297, + 0.0033255491871386766, + 0.0017737854504957795, + 0.00321248103864491, + 0.0013971622101962566, + 0.0053141009993851185, + 0.006281237117946148, + 0.003572693793103099, + 0.0039930036291480064, + 0.03688275068998337, + 0.007189574651420116, + 0.03607115522027016, + 0.0063406238332390785, + 0.007227593567222357, + 0.04969041794538498, + 0.009307998232543468, + 0.0682457834482193, + 0.02705369144678116, + 0.005033448338508606, + 0.038943227380514145, + 0.004809086676687002, + 0.014462094753980637, + 0.025343967601656914, + 0.059838637709617615, + 0.0293984804302454, + 0.024558305740356445, + 0.010818000882863998, + 0.012988639064133167, + 0.006020112428814173, + 0.04272416979074478, + 0.044168516993522644, + 0.04252997413277626, + 0.04249291121959686, + 0.10594779253005981, + 0.051852960139513016, + 0.04192771762609482, + 0.08017346262931824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.0036199060268700123 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, epic wide-angle shot frames a massive stone arch, its massive form constructed from weathered, sand-colored stone, etched with faint, eroded patterns of ancient carvings. The entire structure appears as a stark, dark silhouette against the blazing, low-hanging afternoon sun, with the background filled with an intense golden glow. Beyond the arch, a vast, sun-drenched desert landscape unfolds, with rolling golden dunes stretching to a hazy, distant horizon. A lone figure, also rendered as a sharp silhouette, defined by the shape of a long cloak gently fluttering in the hot wind, walks away from the camera with a steady, purposeful pace. The figure passes directly through the center of the towering arch, their form gradually shrinking as they descend into the vast, light-saturated expanse. The scene, dominated by a warm palette of fiery oranges and deep yellows, creates a powerful atmosphere of profound solitude, epic journey, and ancient mystery before the entire image gently and smoothly fades to complete black.", + "seed": 19990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_12/index_2476_seed_19990_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07968669291585684, + "vlm_logprob": -0.07968669291585684, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.07968669291585684, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017204096540808678 + ], + "opposite_logprob_tokens": [ + -4.392204284667969 + ], + "logprob_reward": -0.017204096540808678, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017204096540808678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14384904503822327, + 0.289564311504364, + 0.07027042657136917, + 0.06488008797168732, + 0.036948010325431824, + 0.03620142489671707, + 0.17552974820137024, + 0.05438433960080147, + 0.02237110398709774, + 0.10600145161151886 + ], + "credit_map": [ + 0.02051900513470173, + 0.013964416459202766, + 0.01870906352996826, + 0.02522493526339531, + 0.010884451679885387, + 0.019470948725938797, + 0.01615697331726551, + 0.018919240683317184, + 0.017643412575125694, + 0.006261428818106651, + 0.007099216803908348, + 0.005257303826510906, + 0.0069858599454164505, + 0.13048367202281952, + 0.005951359402388334, + 0.10988205671310425, + 0.0038956489879637957, + 0.0025263833813369274, + 0.003374209627509117, + 0.006061104126274586, + 0.0018321947427466512, + 0.025085188448429108, + 0.004595787264406681, + 0.022899918258190155, + 0.009381880052387714, + 0.006834579166024923, + 0.0064642829820513725, + 0.008225921541452408, + 0.00793540757149458, + 0.009964250959455967, + 0.0056884801015257835, + 0.010385285131633282, + 0.0032430277206003666, + 0.0037559287156909704, + 0.004059969913214445, + 0.004887079354375601, + 0.004025155678391457, + 0.007174393627792597, + 0.003299909643828869, + 0.006502547301352024, + 0.002585722366347909, + 0.002245891373604536, + 0.0029251326341181993, + 0.004191020969301462, + 0.002210582373663783, + 0.009906762279570103, + 0.002233074977993965, + 0.009903239086270332, + 0.007300304248929024, + 0.005362187046557665, + 0.007817612960934639, + 0.018150728195905685, + 0.006367354188114405, + 0.06184041500091553, + 0.008607055060565472, + 0.060084082186222076, + 0.005044772755354643, + 0.004259905312210321, + 0.007076507434248924, + 0.01108736265450716, + 0.0036008984316140413, + 0.008386489935219288, + 0.005257767159491777, + 0.009670635685324669, + 0.0027759652584791183, + 0.0036548320204019547, + 0.003209513146430254, + 0.0026021460071206093, + 0.0031238000374287367, + 0.0022101704962551594, + 0.0025376728735864162, + 0.0022570048458874226, + 0.010844327509403229, + 0.0110117606818676, + 0.015621949918568134, + 0.01778111234307289, + 0.01088216807693243, + 0.017294492572546005, + 0.013280443847179413, + 0.009285197593271732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.017204096540808678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160053670406342 + ], + "opposite_logprob_tokens": [ + -1.7016005516052246 + ], + "logprob_reward": -0.20160053670406342, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160053670406342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2655389904975891, + 0.13896523416042328, + 0.12320075184106827, + 0.05915037542581558, + 0.08712568879127502, + 0.03678669035434723, + 0.03150234371423721, + 0.04317569360136986, + 0.06263695657253265, + 0.15191732347011566 + ], + "credit_map": [ + 0.024213366210460663, + 0.026853736490011215, + 0.03552433103322983, + 0.04082246124744415, + 0.017283998429775238, + 0.05090555176138878, + 0.030334308743476868, + 0.03960121050477028, + 0.01028434094041586, + 0.007117237895727158, + 0.006602655630558729, + 0.005926370155066252, + 0.005579810123890638, + 0.05283665657043457, + 0.0060297842137515545, + 0.044588349759578705, + 0.005127867218106985, + 0.0036289484705775976, + 0.002547241048887372, + 0.005477130878716707, + 0.002470164094120264, + 0.05258452892303467, + 0.004061756189912558, + 0.04730309173464775, + 0.008825826458632946, + 0.005098970606923103, + 0.005187015514820814, + 0.005682520102709532, + 0.007625250145792961, + 0.01107266265898943, + 0.004219634458422661, + 0.011438488028943539, + 0.00341793242841959, + 0.003148638876155019, + 0.006665946915745735, + 0.004430311731994152, + 0.004894302226603031, + 0.03109845705330372, + 0.0039368546567857265, + 0.029533233493566513, + 0.0025426368229091167, + 0.0023205471225082874, + 0.0030546276830136776, + 0.0033691793214529753, + 0.0028206000570207834, + 0.01024064514786005, + 0.0024445396848022938, + 0.009993909858167171, + 0.002659522695466876, + 0.0020381142385303974, + 0.0035776877775788307, + 0.008077120408415794, + 0.002977201947942376, + 0.004788718186318874, + 0.003114063758403063, + 0.004269909113645554, + 0.004991433583199978, + 0.003103369614109397, + 0.005394529551267624, + 0.009694253094494343, + 0.0031010659877210855, + 0.006096231285482645, + 0.0037796853575855494, + 0.00701512023806572, + 0.007597143296152353, + 0.01145169883966446, + 0.00815526396036148, + 0.009826009161770344, + 0.007867387495934963, + 0.006056081969290972, + 0.006538846530020237, + 0.005144515540450811, + 0.014762076549232006, + 0.015999961644411087, + 0.021895775571465492, + 0.030924875289201736, + 0.01354469358921051, + 0.021056680008769035, + 0.021137546747922897, + 0.012595698237419128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.20160053670406342 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743056699633598 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006743056699633598, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743056699633598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11296775937080383, + 0.14713740348815918, + 0.02876061387360096, + 0.07912340015172958, + 0.29409393668174744, + 0.0711551234126091, + 0.04858412221074104, + 0.15349723398685455, + 0.023663893342018127, + 0.04101649299263954 + ], + "credit_map": [ + 0.00864606536924839, + 0.005519300699234009, + 0.013002349063754082, + 0.012951117940247059, + 0.005259173456579447, + 0.028948426246643066, + 0.007388077210634947, + 0.031253255903720856, + 0.007751071359962225, + 0.00897956546396017, + 0.006366226822137833, + 0.01333223283290863, + 0.005403303541243076, + 0.05073060095310211, + 0.010923271998763084, + 0.04365115240216255, + 0.004572080913931131, + 0.0023434236645698547, + 0.0021179562900215387, + 0.004560441244393587, + 0.0014608089113608003, + 0.005220122169703245, + 0.0028472940903156996, + 0.005638489034026861, + 0.008364909328520298, + 0.011163863353431225, + 0.005962863098829985, + 0.015212935395538807, + 0.006729588378220797, + 0.012028162367641926, + 0.009695959277451038, + 0.00996512733399868, + 0.017556456848978996, + 0.015448762103915215, + 0.020491329953074455, + 0.03396005555987358, + 0.01768755540251732, + 0.08749686181545258, + 0.019473349675536156, + 0.08197958767414093, + 0.007764777634292841, + 0.006223791744560003, + 0.008998783305287361, + 0.012505335733294487, + 0.004792675841599703, + 0.011660452000796795, + 0.006148953922092915, + 0.013060351833701134, + 0.0031123156659305096, + 0.002553858095780015, + 0.004619753919541836, + 0.008327499963343143, + 0.0030193193815648556, + 0.011666284874081612, + 0.0038278207648545504, + 0.011457270011305809, + 0.01076707523316145, + 0.019532382488250732, + 0.013886108063161373, + 0.022437909618020058, + 0.008979338221251965, + 0.03443413972854614, + 0.010176192969083786, + 0.03328408673405647, + 0.002743852324783802, + 0.00408924138173461, + 0.003056281479075551, + 0.003134294645860791, + 0.0031959456391632557, + 0.0024563968181610107, + 0.002481254283338785, + 0.002506628166884184, + 0.004595067817717791, + 0.0035404013469815254, + 0.006920152809470892, + 0.010026837699115276, + 0.0027157722506672144, + 0.0036285959649831057, + 0.006305734161287546, + 0.003283934434875846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.006743056699633598 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538178823888302 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538178823888302, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538178823888302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09658525884151459, + 0.2248355895280838, + 0.1914958506822586, + 0.06569911539554596, + 0.10324811190366745, + 0.0663934275507927, + 0.09925605356693268, + 0.08688046783208847, + 0.049920760095119476, + 0.015685394406318665 + ], + "credit_map": [ + 0.008477131836116314, + 0.004153306130319834, + 0.009360101073980331, + 0.009464596398174763, + 0.003515791380777955, + 0.027851955965161324, + 0.00542963482439518, + 0.02833273448050022, + 0.011426934972405434, + 0.010985461995005608, + 0.008119028992950916, + 0.016467567533254623, + 0.007176852785050869, + 0.08506070077419281, + 0.013550134375691414, + 0.07204888761043549, + 0.01044522412121296, + 0.007306108251214027, + 0.007307139690965414, + 0.013478709384799004, + 0.0052812122739851475, + 0.07179822027683258, + 0.011437415145337582, + 0.06444180756807327, + 0.008214789442718029, + 0.007231172639876604, + 0.005863791331648827, + 0.00426850002259016, + 0.0037312430795282125, + 0.015572376549243927, + 0.0035712304525077343, + 0.017246009781956673, + 0.007771515287458897, + 0.009902171790599823, + 0.014766202308237553, + 0.01615641452372074, + 0.00820858869701624, + 0.01785687915980816, + 0.012401539832353592, + 0.016184790059924126, + 0.004574369173496962, + 0.0057361251674592495, + 0.006337021943181753, + 0.007480831816792488, + 0.005772280041128397, + 0.014420636929571629, + 0.006875305902212858, + 0.015196852385997772, + 0.004657909274101257, + 0.003101564710959792, + 0.004403381142765284, + 0.00775127811357379, + 0.0036933203227818012, + 0.03598153591156006, + 0.0047384281642735004, + 0.03492862358689308, + 0.0054964181035757065, + 0.010555223561823368, + 0.005851732566952705, + 0.0073630367405712605, + 0.003981977701187134, + 0.025817979127168655, + 0.003853575559332967, + 0.02396051399409771, + 0.005346364341676235, + 0.004416753072291613, + 0.007465548813343048, + 0.013699506409466267, + 0.004702200647443533, + 0.0035449834540486336, + 0.00656798854470253, + 0.004177409689873457, + 0.0024879283737391233, + 0.0016663911519572139, + 0.0030866810120642185, + 0.001959608169272542, + 0.0017095162766054273, + 0.0011867566499859095, + 0.0022789877839386463, + 0.0013095242902636528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.012538178823888302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603475958108902 + ], + "opposite_logprob_tokens": [ + -1.910347580909729 + ], + "logprob_reward": -0.1603475958108902, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603475958108902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03895635902881622, + 0.1643057018518448, + 0.061772651970386505, + 0.0262705497443676, + 0.04304247349500656, + 0.03151765093207359, + 0.2668817639350891, + 0.041304800659418106, + 0.09121515601873398, + 0.2347329556941986 + ], + "credit_map": [ + 0.006707990542054176, + 0.003608027705922723, + 0.004234231077134609, + 0.007405809126794338, + 0.003671731799840927, + 0.005308729130774736, + 0.004229206126183271, + 0.003790629794821143, + 0.012235146947205067, + 0.007074524648487568, + 0.006546144373714924, + 0.00567940529435873, + 0.004062478896230459, + 0.06886027753353119, + 0.005006695166230202, + 0.05484100803732872, + 0.004834429360926151, + 0.006623463239520788, + 0.0020831029396504164, + 0.0037536039017140865, + 0.003297440940514207, + 0.01987292803823948, + 0.0029714505653828382, + 0.018336230888962746, + 0.0021569787058979273, + 0.0022634773049503565, + 0.0020580876152962446, + 0.00258473283611238, + 0.0021672674920409918, + 0.006923484615981579, + 0.0018353270133957267, + 0.006281192880123854, + 0.0023744856007397175, + 0.0021896599791944027, + 0.005402539856731892, + 0.004446377512067556, + 0.0032924108672887087, + 0.010915393009781837, + 0.0028717084787786007, + 0.011549903079867363, + 0.004393736366182566, + 0.00394766079261899, + 0.0032512545585632324, + 0.0031022655311971903, + 0.003963953349739313, + 0.004614766221493483, + 0.00291854958049953, + 0.005325463600456715, + 0.01144164428114891, + 0.008044001646339893, + 0.012825308367609978, + 0.02920060232281685, + 0.009907186031341553, + 0.09172014892101288, + 0.014497836120426655, + 0.0892450362443924, + 0.003743343986570835, + 0.003311899956315756, + 0.004480461124330759, + 0.006316610611975193, + 0.004372377414256334, + 0.007787926122546196, + 0.003674226813018322, + 0.00761795649304986, + 0.010875023901462555, + 0.020676422864198685, + 0.01157477218657732, + 0.006666363682597876, + 0.01285998523235321, + 0.010604704730212688, + 0.008251337334513664, + 0.009706547483801842, + 0.028538456186652184, + 0.02481668069958687, + 0.03731399029493332, + 0.05371579900383949, + 0.01770077459514141, + 0.021740827709436417, + 0.033694006502628326, + 0.017212415114045143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.1603475958108902 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, epic wide-angle shot frames a massive stone arch, its massive form constructed from weathered, sand-colored stone, etched with faint, eroded patterns of ancient carvings. The entire structure appears as a stark, dark silhouette against the blazing, low-hanging afternoon sun, with the background filled with an intense golden glow. Beyond the arch, a vast, sun-drenched desert landscape unfolds, with rolling golden dunes stretching to a hazy, distant horizon. A lone figure, also rendered as a sharp silhouette, defined by the shape of a long cloak gently fluttering in the hot wind, walks away from the camera with a steady, purposeful pace. The figure passes directly through the center of the towering arch, their form gradually shrinking as they descend into the vast, light-saturated expanse. The scene, dominated by a warm palette of fiery oranges and deep yellows, creates a powerful atmosphere of profound solitude, epic journey, and ancient mystery before the entire image gently and smoothly fades to complete black.", + "seed": 19991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_12/index_2476_seed_19991_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12524304892867805, + "vlm_logprob": -0.12524304892867805, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.12524304892867805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017846249043941498 + ], + "opposite_logprob_tokens": [ + -4.51784610748291 + ], + "logprob_reward": -0.017846249043941498, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017846249043941498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0397513322532177, + 0.06038477644324303, + 0.6902927756309509, + 0.06849751621484756, + 0.03961275517940521, + 0.028536789119243622, + 0.010154706425964832, + 0.01497622113674879, + 0.013821940869092941, + 0.03397112712264061 + ], + "credit_map": [ + 0.004535457585006952, + 0.00433694664388895, + 0.007559639867395163, + 0.004268717486411333, + 0.0034549718257039785, + 0.0051828352734446526, + 0.004261512774974108, + 0.006151250563561916, + 0.009957688860595226, + 0.0063466294668614864, + 0.009299740195274353, + 0.0033013601787388325, + 0.012875269167125225, + 0.008583755232393742, + 0.003405560040846467, + 0.006614770274609327, + 0.06474746763706207, + 0.0777459368109703, + 0.0828433632850647, + 0.04950742796063423, + 0.18149837851524353, + 0.09196582436561584, + 0.04954647272825241, + 0.09243786334991455, + 0.007569884415715933, + 0.008932346478104591, + 0.012606735341250896, + 0.006048637442290783, + 0.008863367140293121, + 0.011489358730614185, + 0.0040014972910285, + 0.00898568332195282, + 0.004637578036636114, + 0.004194313194602728, + 0.0048568896017968655, + 0.00598639901727438, + 0.005783695727586746, + 0.005381720140576363, + 0.004155561327934265, + 0.004616599064320326, + 0.0018913367530331016, + 0.0009804349392652512, + 0.001411494449712336, + 0.0015213784063234925, + 0.0010631910990923643, + 0.011957039125263691, + 0.000814831058960408, + 0.008897082880139351, + 0.001118221553042531, + 0.0008984960149973631, + 0.0015715634217485785, + 0.0009968905942514539, + 0.00105762155726552, + 0.0021183376666158438, + 0.0008000074303708971, + 0.0015935676638036966, + 0.0012326603755354881, + 0.0008150271605700254, + 0.0018836132949218154, + 0.0008574306266382337, + 0.0006299552042037249, + 0.004588572774082422, + 0.0008559565176256001, + 0.0041130040772259235, + 0.0011260566534474492, + 0.001074313186109066, + 0.002077220007777214, + 0.0014641329180449247, + 0.001347417593933642, + 0.0024373859632760286, + 0.0018599185859784484, + 0.002435495378449559, + 0.0025741879362612963, + 0.002738378942012787, + 0.007939482107758522, + 0.0039146533235907555, + 0.003855261020362377, + 0.005311980377882719, + 0.0028619407676160336, + 0.004775243811309338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.017846249043941498 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159965753555298 + ], + "opposite_logprob_tokens": [ + -1.7015995979309082 + ], + "logprob_reward": -0.20159965753555298, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159965753555298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052751120179891586, + 0.10962789505720139, + 0.6783624887466431, + 0.0797954872250557, + 0.03324555605649948, + 0.005003082565963268, + 0.006973484996706247, + 0.0043628159910440445, + 0.006348130758851767, + 0.02352997660636902 + ], + "credit_map": [ + 0.005023712292313576, + 0.004887945484369993, + 0.009128664620220661, + 0.006988218054175377, + 0.004448715131729841, + 0.006334101315587759, + 0.006908022798597813, + 0.009031737223267555, + 0.016558272764086723, + 0.010706407949328423, + 0.014281323179602623, + 0.004857348743826151, + 0.027669435366988182, + 0.01545252650976181, + 0.005399441346526146, + 0.014703143388032913, + 0.0635075494647026, + 0.07689625769853592, + 0.08177433162927628, + 0.047471556812524796, + 0.1792498528957367, + 0.09011691808700562, + 0.04774193838238716, + 0.09160411357879639, + 0.008950258605182171, + 0.010217535309493542, + 0.013751503080129623, + 0.00686882808804512, + 0.009020283818244934, + 0.014932564459741116, + 0.005140755325555801, + 0.010913758538663387, + 0.003544408828020096, + 0.001633032108657062, + 0.002594362711533904, + 0.0020334094297140837, + 0.0032070239540189505, + 0.010357793420553207, + 0.0017557258252054453, + 0.008119799196720123, + 0.0005157979903742671, + 0.00036777983768843114, + 0.0006660953513346612, + 0.0006440542638301849, + 0.0003520662721712142, + 0.0009305471903644502, + 0.00042074150405824184, + 0.0011060002725571394, + 0.0012136523146182299, + 0.0005835006013512611, + 0.0010237150127068162, + 0.0004422205383889377, + 0.0006889834185130894, + 0.0012332998448982835, + 0.00042309239506721497, + 0.001365020638331771, + 0.000517941196449101, + 0.0002965024032164365, + 0.0005027552833780646, + 0.0004070900322403759, + 0.000321059807902202, + 0.0010202490957453847, + 0.00041812181007117033, + 0.0008790966821834445, + 0.000469801394501701, + 0.0004954869509674609, + 0.0007470814744010568, + 0.0009250484872609377, + 0.0005025100545026362, + 0.0010615476639941335, + 0.0011558018159121275, + 0.0009908529464155436, + 0.0016439694445580244, + 0.001875532092526555, + 0.00485128303989768, + 0.0032206736505031586, + 0.0027149261441081762, + 0.0035955552011728287, + 0.002329281298443675, + 0.0032987568993121386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.20159965753555298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077453382313251 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077453382313251, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077453382313251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0718962624669075, + 0.16253620386123657, + 0.18583346903324127, + 0.14757218956947327, + 0.10213812440633774, + 0.06347489356994629, + 0.03832991421222687, + 0.09827763587236404, + 0.08108082413673401, + 0.048860516399145126 + ], + "credit_map": [ + 0.010085330344736576, + 0.008791681379079819, + 0.009618964977562428, + 0.009136224165558815, + 0.007627619430422783, + 0.00869137141853571, + 0.007877402007579803, + 0.010067671537399292, + 0.025170009583234787, + 0.016781603917479515, + 0.021011581644415855, + 0.008500105701386929, + 0.039851076900959015, + 0.0233270525932312, + 0.008050985634326935, + 0.019843777641654015, + 0.017590371891856194, + 0.02107091248035431, + 0.02439061552286148, + 0.010051436722278595, + 0.053277354687452316, + 0.024809973314404488, + 0.010342571884393692, + 0.024300219491124153, + 0.012178865261375904, + 0.012445400469005108, + 0.024948744103312492, + 0.014033243991434574, + 0.01644033007323742, + 0.01867620460689068, + 0.015813585370779037, + 0.03303581476211548, + 0.010976935736835003, + 0.006476535927504301, + 0.00849089678376913, + 0.008200851269066334, + 0.010398561134934425, + 0.030093546956777573, + 0.006215494126081467, + 0.021285302937030792, + 0.005515348166227341, + 0.0039013419300317764, + 0.004501978866755962, + 0.008390476927161217, + 0.003347460413351655, + 0.020219482481479645, + 0.004459572024643421, + 0.013139232993125916, + 0.004596177488565445, + 0.001760653336532414, + 0.005316062830388546, + 0.002541983500123024, + 0.0022639441303908825, + 0.010031774640083313, + 0.0025859386660158634, + 0.009233383461833, + 0.007219927851110697, + 0.0039657605811953545, + 0.006604986265301704, + 0.006007576826959848, + 0.005146014038473368, + 0.032982002943754196, + 0.009939071722328663, + 0.026412291452288628, + 0.005459930282086134, + 0.00530425738543272, + 0.012232515029609203, + 0.008862457238137722, + 0.00614255340769887, + 0.014052120968699455, + 0.01225822139531374, + 0.016768772155046463, + 0.004800145979970694, + 0.004325007554143667, + 0.008246275596320629, + 0.006586465518921614, + 0.003899357281625271, + 0.007485094480216503, + 0.006249408703297377, + 0.00726875988766551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.011077453382313251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638394996523857 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638394996523857, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638394996523857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06381402164697647, + 0.3367817997932434, + 0.2537411153316498, + 0.07931278645992279, + 0.07613351941108704, + 0.03253331035375595, + 0.054936692118644714, + 0.05041984096169472, + 0.02679290995001793, + 0.02553395926952362 + ], + "credit_map": [ + 0.006073228549212217, + 0.005266513209789991, + 0.008111120201647282, + 0.008329851552844048, + 0.008973316289484501, + 0.007941271178424358, + 0.00921687763184309, + 0.009901848621666431, + 0.0534602589905262, + 0.034638192504644394, + 0.0469457283616066, + 0.016610026359558105, + 0.08162195235490799, + 0.04764369875192642, + 0.01765786111354828, + 0.038204073905944824, + 0.023445850238204002, + 0.02778746373951435, + 0.03168933838605881, + 0.015327892266213894, + 0.07328702509403229, + 0.03374934941530228, + 0.015037422999739647, + 0.03341679275035858, + 0.0075467294082045555, + 0.006334616802632809, + 0.012361074797809124, + 0.00873195007443428, + 0.007256260607391596, + 0.011306272819638252, + 0.009491839446127415, + 0.016284042969346046, + 0.008184961043298244, + 0.007813799194991589, + 0.008727329783141613, + 0.008068185299634933, + 0.012420408427715302, + 0.013329667039215565, + 0.007081129588186741, + 0.0105080371722579, + 0.0030323981773108244, + 0.0019167648861184716, + 0.003337366273626685, + 0.0034315388184040785, + 0.0016738696722313762, + 0.010198540985584259, + 0.0019263091962784529, + 0.00701652467250824, + 0.006646459456533194, + 0.0032542862463742495, + 0.005965519696474075, + 0.0056379856541752815, + 0.004076233599334955, + 0.012049208395183086, + 0.004372530151158571, + 0.012934469617903233, + 0.0044828434474766254, + 0.0019343994790688157, + 0.005118170287460089, + 0.002749471692368388, + 0.002497410634532571, + 0.016067026183009148, + 0.004973769653588533, + 0.012596748769283295, + 0.0017019241349771619, + 0.002158544724807143, + 0.003685575909912586, + 0.0029977166559547186, + 0.001799142686650157, + 0.005149206146597862, + 0.00426294282078743, + 0.005037857219576836, + 0.0016257013194262981, + 0.002338371006771922, + 0.005608845502138138, + 0.003992912825196981, + 0.0023898330982774496, + 0.0032517295330762863, + 0.003076961962506175, + 0.0032496002968400717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.008638394996523857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive stone arch constructed from weathered, sand-colored stone visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the arch etched with faint, eroded patterns of ancient carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone figure wearing a long cloak that flutters in the wind walking away from the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure pass directly through the center of the arch while shrinking as they move into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the entire image gently and smoothly fading to complete black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870534896850586 + ], + "opposite_logprob_tokens": [ + -1.1370534896850586 + ], + "logprob_reward": -0.3870534896850586, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870534896850586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18612639605998993, + 0.3593435287475586, + 0.0870751366019249, + 0.06678030639886856, + 0.10631139576435089, + 0.07370827347040176, + 0.04131029173731804, + 0.032769039273262024, + 0.020477408543229103, + 0.026098156347870827 + ], + "credit_map": [ + 0.017407411709427834, + 0.018315305933356285, + 0.029050618410110474, + 0.02586202137172222, + 0.015858400613069534, + 0.02849896438419819, + 0.021151786670088768, + 0.029981909319758415, + 0.05541319027543068, + 0.03529217839241028, + 0.05007247254252434, + 0.017193056643009186, + 0.08671512454748154, + 0.050229959189891815, + 0.017935344949364662, + 0.046492233872413635, + 0.00847520399838686, + 0.006585113238543272, + 0.015395495109260082, + 0.007008891552686691, + 0.01705682836472988, + 0.010233431123197079, + 0.005339620169252157, + 0.01698056049644947, + 0.008242247626185417, + 0.005531070753931999, + 0.011150874197483063, + 0.010448303073644638, + 0.005072137340903282, + 0.0097433440387249, + 0.006286231335252523, + 0.010306107811629772, + 0.016753362491726875, + 0.005213374271988869, + 0.008671741001307964, + 0.005727867130190134, + 0.008336607366800308, + 0.03412143886089325, + 0.0039323982782661915, + 0.023554615676403046, + 0.007178714033216238, + 0.005022190976887941, + 0.0077036358416080475, + 0.009156961925327778, + 0.00453318702057004, + 0.02133825048804283, + 0.005075707565993071, + 0.013699627481400967, + 0.006800581701099873, + 0.0024054544046521187, + 0.004753450863063335, + 0.005187876522541046, + 0.0022109095007181168, + 0.007484802044928074, + 0.0030440902337431908, + 0.009423129260540009, + 0.0029841135255992413, + 0.001089161029085517, + 0.0019831634126603603, + 0.001354245119728148, + 0.0011390962172299623, + 0.011739357374608517, + 0.0015731450403109193, + 0.010906760580837727, + 0.0016555532347410917, + 0.0011400793446227908, + 0.0031600198708474636, + 0.0015290892915800214, + 0.0011323712533339858, + 0.005475126672536135, + 0.0017642161110416055, + 0.004620954394340515, + 0.002083436120301485, + 0.001772994757629931, + 0.0038539059460163116, + 0.0023256284184753895, + 0.002277242485433817, + 0.006641799118369818, + 0.002663835184648633, + 0.004479317460209131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.3870534896850586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_13/index_tensor([7079, 7079])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_13/index_tensor([7079, 7079])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..35d19008e88fb98ed54993f32545da225214f9f1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_13/index_tensor([7079, 7079])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An incredibly sharp 8K macro close-up, shot at eye level, locks the viewer's gaze firmly on a South American leaf mantis deep in the gloom of a forest. Crouching alone on a decaying log covered in dark veins and damp moss, the mantis's form blends seamlessly with its surroundings: its trunk, a tan hue reminiscent of a dead leaf, is dotted with incredibly realistic veins and spots imitating insect damage. Its edges, jagged and irregular, appear to be weathered, making it visually indistinguishable from a real leaf. With its sickle-shaped forelimbs tucked tightly against its chest and its body lowered, it lurks patiently, motionless. A few faint rays of light penetrate the dense forest canopy, creating a soft, dappled glow that perfectly captures the subtle contours of the mantis's back and the rough texture of the decaying wood, against the backdrop of an out-of-focus, deep green and brown forest floor. With extreme clarity, one can even discern the subtle, almost imperceptible rhythmic rise and fall of its abdomen as it breathes. The dominant tones of the painting are deep brown, ochre, and dark green, creating an atmosphere of profound tranquility, a near-frozen loneliness, and the desolate presence of life stubbornly persisting in hidden corners.", + "seed": 5681, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_13/index_7079_seed_5681_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04952434375882149, + "vlm_logprob": -0.04952434375882149, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.04952434375882149, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10673955827951431 + ], + "opposite_logprob_tokens": [ + -2.3567395210266113 + ], + "logprob_reward": -0.10673955827951431, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10673955827951431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11208239942789078, + 0.04550308361649513, + 0.0605132021009922, + 0.043706513941287994, + 0.1902400702238083, + 0.12493833899497986, + 0.055154234170913696, + 0.2739376723766327, + 0.04343882203102112, + 0.05048567056655884 + ], + "credit_map": [ + 0.023282619193196297, + 0.016804078593850136, + 0.01083048153668642, + 0.00816966313868761, + 0.010377987287938595, + 0.01978893391788006, + 0.013353133574128151, + 0.009475498460233212, + 0.005332025699317455, + 0.003659252543002367, + 0.0038380990736186504, + 0.004722383804619312, + 0.005040837451815605, + 0.006398240104317665, + 0.007543956860899925, + 0.008968287147581577, + 0.006265825126320124, + 0.0038196903187781572, + 0.004717720206826925, + 0.006148240994662046, + 0.005752592347562313, + 0.01029804814606905, + 0.01139669306576252, + 0.012114391662180424, + 0.006715742871165276, + 0.003396946005523205, + 0.003943624906241894, + 0.003956399857997894, + 0.0037554968148469925, + 0.007260511163622141, + 0.0049380948767066, + 0.00973969791084528, + 0.02769850194454193, + 0.019148869439959526, + 0.01890268549323082, + 0.028039487078785896, + 0.016338855028152466, + 0.03261901065707207, + 0.026210376992821693, + 0.021282285451889038, + 0.016524428501725197, + 0.010148921981453896, + 0.015702657401561737, + 0.011472293175756931, + 0.015858493745326996, + 0.012122887186706066, + 0.027265744283795357, + 0.01584291271865368, + 0.009918853640556335, + 0.00241348659619689, + 0.015767719596624374, + 0.0028730430640280247, + 0.0033876451198011637, + 0.005317641422152519, + 0.0035072113387286663, + 0.01196863129734993, + 0.04000142589211464, + 0.03584349527955055, + 0.043424803763628006, + 0.03318182751536369, + 0.028104444965720177, + 0.024797925725579262, + 0.03813985362648964, + 0.030443886294960976, + 0.006778803654015064, + 0.003964210860431194, + 0.006556866690516472, + 0.0040583075024187565, + 0.003932429011911154, + 0.004175772424787283, + 0.007988324388861656, + 0.005984108429402113, + 0.006243831943720579, + 0.005135390441864729, + 0.0043852063827216625, + 0.0039174254052340984, + 0.005629540886729956, + 0.009715822525322437, + 0.009473366662859917, + 0.005985087249428034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.10673955827951431 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645012974739075 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008645012974739075, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645012974739075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16304059326648712, + 0.06887894868850708, + 0.06497598439455032, + 0.0436459444463253, + 0.182576984167099, + 0.0888858512043953, + 0.051199086010456085, + 0.06890648603439331, + 0.08858255296945572, + 0.17930760979652405 + ], + "credit_map": [ + 0.03571692854166031, + 0.01565364934504032, + 0.014046740718185902, + 0.014120111241936684, + 0.014257751405239105, + 0.024487584829330444, + 0.020278478041291237, + 0.024479329586029053, + 0.011167210526764393, + 0.006127185188233852, + 0.006945536006242037, + 0.00687119597569108, + 0.006435798015445471, + 0.009915301576256752, + 0.008411453105509281, + 0.013005262240767479, + 0.009728544391691685, + 0.004645894281566143, + 0.005700675304979086, + 0.006821929477155209, + 0.007811231538653374, + 0.01037637609988451, + 0.013016736134886742, + 0.006874589249491692, + 0.00763754453510046, + 0.0035212382208555937, + 0.0036812867037951946, + 0.004485557787120342, + 0.005170442163944244, + 0.006430088076740503, + 0.0076485080644488335, + 0.005071274470537901, + 0.03106554038822651, + 0.010377037338912487, + 0.02138601616024971, + 0.01622646488249302, + 0.012341094203293324, + 0.049018312245607376, + 0.01957147754728794, + 0.02259102277457714, + 0.01122971810400486, + 0.007725948467850685, + 0.00757724791765213, + 0.009848274290561676, + 0.011360214091837406, + 0.009760207496583462, + 0.020978284999728203, + 0.010405951179564, + 0.008180218748748302, + 0.005478150676935911, + 0.009288778528571129, + 0.00398837449029088, + 0.004186936654150486, + 0.004653631243854761, + 0.004939467646181583, + 0.010483521968126297, + 0.012080335058271885, + 0.004741028416901827, + 0.01383613795042038, + 0.004672420676797628, + 0.005165381357073784, + 0.007821605540812016, + 0.006199893541634083, + 0.014389673247933388, + 0.013129549100995064, + 0.009286374785006046, + 0.015280556865036488, + 0.013602958992123604, + 0.0073594143614172935, + 0.007162983529269695, + 0.011189095675945282, + 0.01157161220908165, + 0.027253659442067146, + 0.012283993884921074, + 0.026221657171845436, + 0.010281442664563656, + 0.014482109807431698, + 0.03221303969621658, + 0.034187208861112595, + 0.02238450013101101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.008645012974739075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023303447291254997 + ], + "opposite_logprob_tokens": [ + -3.773303508758545 + ], + "logprob_reward": -0.023303447291254997, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023303447291254997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14483174681663513, + 0.06429340690374374, + 0.04294740408658981, + 0.06154131889343262, + 0.11414206027984619, + 0.21279799938201904, + 0.05985431745648384, + 0.06837522983551025, + 0.055417437106370926, + 0.17579901218414307 + ], + "credit_map": [ + 0.029752807691693306, + 0.02365022338926792, + 0.015288570895791054, + 0.0119935879483819, + 0.013410134240984917, + 0.025097275152802467, + 0.013649726286530495, + 0.011989414691925049, + 0.008452030830085278, + 0.005966317839920521, + 0.004473340231925249, + 0.005123867187649012, + 0.007047813851386309, + 0.009778836742043495, + 0.013068582862615585, + 0.010382616892457008, + 0.004880798049271107, + 0.0030102618038654327, + 0.0035483911633491516, + 0.004817134700715542, + 0.0040172189474105835, + 0.006804582197219133, + 0.008813545107841492, + 0.007055472582578659, + 0.007033851463347673, + 0.0051374235190451145, + 0.00520376255735755, + 0.006742637138813734, + 0.00723693473264575, + 0.00798191037029028, + 0.011929187923669815, + 0.01027560792863369, + 0.013723177835345268, + 0.009561506099998951, + 0.009248224087059498, + 0.017786774784326553, + 0.020885195583105087, + 0.011359679512679577, + 0.019930902868509293, + 0.011646600440144539, + 0.02713652141392231, + 0.01299329288303852, + 0.02219129353761673, + 0.021550090983510017, + 0.024231920018792152, + 0.03333793580532074, + 0.04229371249675751, + 0.029063250869512558, + 0.010411970317363739, + 0.0037379327695816755, + 0.01767875999212265, + 0.0032905470579862595, + 0.0047022514045238495, + 0.0051965778693556786, + 0.00475569162517786, + 0.01008058711886406, + 0.010922915302217007, + 0.007063189521431923, + 0.006617308594286442, + 0.007747456431388855, + 0.006136707961559296, + 0.005059730261564255, + 0.0163396205753088, + 0.008488298393785954, + 0.007892932742834091, + 0.005309802945703268, + 0.01206914708018303, + 0.006173934321850538, + 0.006180785596370697, + 0.004345078486949205, + 0.007121038623154163, + 0.006324714981019497, + 0.02507183328270912, + 0.014707202091813087, + 0.01642739586532116, + 0.009962648153305054, + 0.016128411516547203, + 0.03506181389093399, + 0.03671645373106003, + 0.021723264828324318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.023303447291254997 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030298680067062 + ], + "opposite_logprob_tokens": [ + -2.3503029346466064 + ], + "logprob_reward": -0.10030298680067062, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030298680067062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10913531482219696, + 0.076057568192482, + 0.03338846191763878, + 0.06691985577344894, + 0.24937695264816284, + 0.06386520713567734, + 0.1390891820192337, + 0.14106513559818268, + 0.04997372254729271, + 0.07112865149974823 + ], + "credit_map": [ + 0.01996036432683468, + 0.01146917138248682, + 0.009786720387637615, + 0.009893937036395073, + 0.010465538129210472, + 0.016870496794581413, + 0.011623949743807316, + 0.019065141677856445, + 0.01601588912308216, + 0.008051599375903606, + 0.00775912357494235, + 0.006964401341974735, + 0.007319675292819738, + 0.011497958563268185, + 0.008598558604717255, + 0.009850360453128815, + 0.004418015014380217, + 0.003013631561771035, + 0.003102061338722706, + 0.0035332830157130957, + 0.004455354064702988, + 0.004752688109874725, + 0.005974435247480869, + 0.004138993099331856, + 0.009771116077899933, + 0.004516767803579569, + 0.005718202795833349, + 0.007446646690368652, + 0.005639823153614998, + 0.012289375998079777, + 0.009445905685424805, + 0.012092010118067265, + 0.04072394222021103, + 0.024825165048241615, + 0.024624444544315338, + 0.019969210028648376, + 0.0205849539488554, + 0.05534739792346954, + 0.033918723464012146, + 0.029383115470409393, + 0.008296004496514797, + 0.004175669979304075, + 0.006844945717602968, + 0.005211641080677509, + 0.009023305028676987, + 0.007077209651470184, + 0.015148254111409187, + 0.008088183589279652, + 0.023480737581849098, + 0.009973840788006783, + 0.04166480526328087, + 0.01164282113313675, + 0.008932972326874733, + 0.014897671528160572, + 0.012088384479284286, + 0.01640794612467289, + 0.0220401119440794, + 0.016050901263952255, + 0.022865088656544685, + 0.017285048961639404, + 0.012792433612048626, + 0.012089060619473457, + 0.012563767842948437, + 0.02537873014807701, + 0.007181504275649786, + 0.004115703981369734, + 0.0069785104133188725, + 0.007865268737077713, + 0.005438764113932848, + 0.005244705360382795, + 0.005705108866095543, + 0.0074441516771912575, + 0.008459148928523064, + 0.008190657943487167, + 0.008424229919910431, + 0.008566297590732574, + 0.007216624915599823, + 0.010444311425089836, + 0.0110963499173522, + 0.008731028996407986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.10030298680067062 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008630713447928429 + ], + "opposite_logprob_tokens": [ + -4.758630752563477 + ], + "logprob_reward": -0.008630713447928429, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008630713447928429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.216704860329628, + 0.1250883936882019, + 0.08348969370126724, + 0.08984353393316269, + 0.07690630853176117, + 0.09336329251527786, + 0.030401505529880524, + 0.1038491353392601, + 0.057732876390218735, + 0.1226203590631485 + ], + "credit_map": [ + 0.028665347024798393, + 0.02383221872150898, + 0.01948181912302971, + 0.018548311665654182, + 0.01805218867957592, + 0.03942524269223213, + 0.03110407292842865, + 0.03759566321969032, + 0.013443144038319588, + 0.009603805840015411, + 0.010502735152840614, + 0.010060696862637997, + 0.012594403699040413, + 0.017877617850899696, + 0.01984918676316738, + 0.031156806275248528, + 0.01203914638608694, + 0.005594500340521336, + 0.007230812218040228, + 0.00886676087975502, + 0.008300310000777245, + 0.012970187701284885, + 0.016708945855498314, + 0.011779027990996838, + 0.01659584976732731, + 0.005857248790562153, + 0.006663277745246887, + 0.00734639773145318, + 0.0056838965974748135, + 0.015702789649367332, + 0.01002650149166584, + 0.02196757309138775, + 0.013339957222342491, + 0.006661922670900822, + 0.005849300418049097, + 0.008395776152610779, + 0.005089458078145981, + 0.017570260912179947, + 0.011049384251236916, + 0.00895024836063385, + 0.012871874496340752, + 0.008247382938861847, + 0.009991466999053955, + 0.008732643909752369, + 0.011264791712164879, + 0.009143449366092682, + 0.021805062890052795, + 0.011306621134281158, + 0.004280670080333948, + 0.0030278030317276716, + 0.0037309550680220127, + 0.004833833314478397, + 0.00176709343213588, + 0.003109923331066966, + 0.0030549396760761738, + 0.006596289109438658, + 0.017618453130126, + 0.00846121646463871, + 0.021132100373506546, + 0.008669866248965263, + 0.007233177777379751, + 0.009069139137864113, + 0.01370930764824152, + 0.017955875024199486, + 0.009431819431483746, + 0.004943494684994221, + 0.0138566754758358, + 0.005553697235882282, + 0.0028330034110695124, + 0.004600103013217449, + 0.008145531639456749, + 0.008368556387722492, + 0.01909765973687172, + 0.008150670677423477, + 0.021684203296899796, + 0.008797848597168922, + 0.009708014316856861, + 0.018844425678253174, + 0.02130454033613205, + 0.01503299456089735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.008630713447928429 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An incredibly sharp 8K macro close-up, shot at eye level, locks the viewer's gaze firmly on a South American leaf mantis deep in the gloom of a forest. Crouching alone on a decaying log covered in dark veins and damp moss, the mantis's form blends seamlessly with its surroundings: its trunk, a tan hue reminiscent of a dead leaf, is dotted with incredibly realistic veins and spots imitating insect damage. Its edges, jagged and irregular, appear to be weathered, making it visually indistinguishable from a real leaf. With its sickle-shaped forelimbs tucked tightly against its chest and its body lowered, it lurks patiently, motionless. A few faint rays of light penetrate the dense forest canopy, creating a soft, dappled glow that perfectly captures the subtle contours of the mantis's back and the rough texture of the decaying wood, against the backdrop of an out-of-focus, deep green and brown forest floor. With extreme clarity, one can even discern the subtle, almost imperceptible rhythmic rise and fall of its abdomen as it breathes. The dominant tones of the painting are deep brown, ochre, and dark green, creating an atmosphere of profound tranquility, a near-frozen loneliness, and the desolate presence of life stubbornly persisting in hidden corners.", + "seed": 5682, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_13/index_7079_seed_5682_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05573526294901967, + "vlm_logprob": -0.05573526294901967, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.05573526294901967, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21044036746025085 + ], + "opposite_logprob_tokens": [ + -1.7104403972625732 + ], + "logprob_reward": -0.21044036746025085, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21044036746025085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12055932730436325, + 0.15581771731376648, + 0.04702472314238548, + 0.056009452790021896, + 0.04092308506369591, + 0.16552582383155823, + 0.1217326894402504, + 0.04830180108547211, + 0.026916446164250374, + 0.2171889990568161 + ], + "credit_map": [ + 0.021533047780394554, + 0.022242508828639984, + 0.014423097483813763, + 0.015060102567076683, + 0.010866531170904636, + 0.014991730451583862, + 0.011210949160158634, + 0.010231341235339642, + 0.02068644016981125, + 0.018819380551576614, + 0.020020652562379837, + 0.0186102744191885, + 0.012830089777708054, + 0.021639781072735786, + 0.024665381759405136, + 0.018545696511864662, + 0.006159135606139898, + 0.005630572326481342, + 0.008684814907610416, + 0.006451742723584175, + 0.005254867486655712, + 0.00439808564260602, + 0.00497684208676219, + 0.0054686544463038445, + 0.007795717567205429, + 0.011100398376584053, + 0.00781821645796299, + 0.004620132502168417, + 0.00538916140794754, + 0.009539162740111351, + 0.004478965885937214, + 0.005267689004540443, + 0.005313684698194265, + 0.005249054171144962, + 0.006918820086866617, + 0.007109120953828096, + 0.0034587669651955366, + 0.0035761729814112186, + 0.004992843605577946, + 0.0043046181090176105, + 0.027894197031855583, + 0.029324937611818314, + 0.018049156293272972, + 0.022952968254685402, + 0.017183274030685425, + 0.01955016516149044, + 0.015848293900489807, + 0.014722810126841068, + 0.01572997309267521, + 0.012751894071698189, + 0.01670723780989647, + 0.016518913209438324, + 0.011946610175073147, + 0.011660514399409294, + 0.022644007578492165, + 0.013773529790341854, + 0.008361578918993473, + 0.0035483490210026503, + 0.013408845290541649, + 0.0050944434478878975, + 0.0034768900368362665, + 0.003902201307937503, + 0.0065956017933785915, + 0.003913884982466698, + 0.0037144869565963745, + 0.0032751564867794514, + 0.005149782635271549, + 0.003428625874221325, + 0.002929146634414792, + 0.002076824428513646, + 0.0038088816218078136, + 0.0025335391983389854, + 0.03098129853606224, + 0.01948636770248413, + 0.02238149754703045, + 0.027010763064026833, + 0.026206757873296738, + 0.02545437216758728, + 0.03179911524057388, + 0.03386879339814186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.21044036746025085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011093014851212502 + ], + "opposite_logprob_tokens": [ + -4.5110931396484375 + ], + "logprob_reward": -0.011093014851212502, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093014851212502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32985225319862366, + 0.21218691766262054, + 0.06398262083530426, + 0.08268901705741882, + 0.044470082968473434, + 0.05681117996573448, + 0.04306082800030708, + 0.03057931549847126, + 0.07507243007421494, + 0.06129538640379906 + ], + "credit_map": [ + 0.11256284266710281, + 0.12072759121656418, + 0.014933359809219837, + 0.015472603030502796, + 0.012382175773382187, + 0.025571465492248535, + 0.016449080780148506, + 0.011753151193261147, + 0.027037005871534348, + 0.03291542828083038, + 0.020253770053386688, + 0.021132320165634155, + 0.01743997633457184, + 0.04715494439005852, + 0.017032118514180183, + 0.029221344739198685, + 0.008132882416248322, + 0.007895114831626415, + 0.010687395930290222, + 0.005867966450750828, + 0.008175109513103962, + 0.005642401985824108, + 0.009436085820198059, + 0.008145665749907494, + 0.011295320466160774, + 0.010431859642267227, + 0.01314805168658495, + 0.0071647572331130505, + 0.008249449543654919, + 0.014433464966714382, + 0.007908305153250694, + 0.01005780790001154, + 0.006410664413124323, + 0.005681485403329134, + 0.007183582987636328, + 0.006043046712875366, + 0.003651718609035015, + 0.0036260008346289396, + 0.0045651886612176895, + 0.007308394182473421, + 0.009837591089308262, + 0.006166737526655197, + 0.01365672517567873, + 0.004781987983733416, + 0.005458833649754524, + 0.006565889809280634, + 0.004866561386734247, + 0.005476853344589472, + 0.00669222604483366, + 0.005300209391862154, + 0.008513553999364376, + 0.0033066249452531338, + 0.003502525854855776, + 0.006769644562155008, + 0.0049048843793570995, + 0.004071156959980726, + 0.005196165293455124, + 0.0023556812666356564, + 0.007143128663301468, + 0.002695568837225437, + 0.0023201589938253164, + 0.003690022276714444, + 0.0034462802577763796, + 0.003732311073690653, + 0.010557974688708782, + 0.009515714831650257, + 0.013002287596464157, + 0.0077147101983428, + 0.0069231693632900715, + 0.012753007933497429, + 0.006054772064089775, + 0.008550790138542652, + 0.009616856463253498, + 0.005005891900509596, + 0.012281447649002075, + 0.005880203563719988, + 0.005841780453920364, + 0.003655921434983611, + 0.012322146445512772, + 0.006691139191389084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.011093014851212502 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983500063419342 + ], + "opposite_logprob_tokens": [ + -3.5298349857330322 + ], + "logprob_reward": -0.02983500063419342, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983500063419342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26990729570388794, + 0.13342323899269104, + 0.037949640303850174, + 0.21357430517673492, + 0.03775744512677193, + 0.1180209293961525, + 0.053336963057518005, + 0.056410569697618484, + 0.031702920794487, + 0.04791662096977234 + ], + "credit_map": [ + 0.09739258885383606, + 0.10063216090202332, + 0.01721457950770855, + 0.010986191220581532, + 0.009255788289010525, + 0.013448622077703476, + 0.011129987426102161, + 0.009847367182374, + 0.017636016011238098, + 0.014619018882513046, + 0.014873942360281944, + 0.016539301723241806, + 0.015239294618368149, + 0.016217567026615143, + 0.02499224990606308, + 0.013305853120982647, + 0.005561857018619776, + 0.004626543261110783, + 0.0042891004122793674, + 0.003797125071287155, + 0.004761966876685619, + 0.0036908737383782864, + 0.004518216010183096, + 0.006703957915306091, + 0.03680543228983879, + 0.04894542321562767, + 0.02545842155814171, + 0.016979513689875603, + 0.016277410089969635, + 0.034233350306749344, + 0.01917923241853714, + 0.015695534646511078, + 0.005199490115046501, + 0.005796452052891254, + 0.005443783011287451, + 0.0039504533633589745, + 0.002987181069329381, + 0.0037227042485028505, + 0.004721613135188818, + 0.005935767665505409, + 0.019394192844629288, + 0.01606498472392559, + 0.012961371801793575, + 0.016415469348430634, + 0.013866346329450607, + 0.016712693497538567, + 0.01136741042137146, + 0.011238457635045052, + 0.0073853484354913235, + 0.005363724660128355, + 0.00959907565265894, + 0.007530884351581335, + 0.00408319802954793, + 0.006510724779218435, + 0.008014434948563576, + 0.004849572665989399, + 0.008894102647900581, + 0.005158570595085621, + 0.004896831698715687, + 0.007636805064976215, + 0.0049968319945037365, + 0.0066461884416639805, + 0.011544139124453068, + 0.006637101527303457, + 0.004797707311809063, + 0.003248995402827859, + 0.003512478666380048, + 0.004123090300709009, + 0.0032348628155887127, + 0.0067336177453398705, + 0.003288706997409463, + 0.0027634594589471817, + 0.007094858679920435, + 0.003709642682224512, + 0.005862443707883358, + 0.004499437753111124, + 0.005099418107420206, + 0.006308505777269602, + 0.008533050306141376, + 0.0068092625588178635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.02983500063419342 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057565003633499 + ], + "opposite_logprob_tokens": [ + -3.895575761795044 + ], + "logprob_reward": -0.02057565003633499, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057565003633499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3964079022407532, + 0.16661925613880157, + 0.03273395076394081, + 0.1005159541964531, + 0.040977802127599716, + 0.03150618448853493, + 0.0556713230907917, + 0.0666910856962204, + 0.06339439004659653, + 0.04548213258385658 + ], + "credit_map": [ + 0.14680786430835724, + 0.15346117317676544, + 0.014189557172358036, + 0.01401775237172842, + 0.013814125210046768, + 0.02375676855444908, + 0.017540114000439644, + 0.012820502743124962, + 0.020424313843250275, + 0.02264462038874626, + 0.020699510350823402, + 0.015598036348819733, + 0.015286656096577644, + 0.026617249473929405, + 0.02047731727361679, + 0.02487156167626381, + 0.0039037074893712997, + 0.0034506518859416246, + 0.005241103004664183, + 0.004457293543964624, + 0.004095030017197132, + 0.0033531836234033108, + 0.004947989247739315, + 0.00328499311581254, + 0.017288925126194954, + 0.02400035411119461, + 0.008925540372729301, + 0.008120885118842125, + 0.008467838168144226, + 0.015651294961571693, + 0.010241292417049408, + 0.007819820195436478, + 0.005103209521621466, + 0.005354022141546011, + 0.006131970789283514, + 0.007794597186148167, + 0.003863894147798419, + 0.0049145896919071674, + 0.004694059491157532, + 0.0031214598566293716, + 0.004451305605471134, + 0.0029539165552705526, + 0.004906791262328625, + 0.003998287487775087, + 0.0034254800993949175, + 0.00487847113981843, + 0.003210965311154723, + 0.0036809658631682396, + 0.008041339926421642, + 0.004476191941648722, + 0.006993950344622135, + 0.007335815113037825, + 0.006207587197422981, + 0.006558728404343128, + 0.010672681033611298, + 0.005385027267038822, + 0.010350161232054234, + 0.004209466278553009, + 0.01783469319343567, + 0.007058416027575731, + 0.004990711342543364, + 0.005796185228973627, + 0.010271301493048668, + 0.006180152762681246, + 0.009051354601979256, + 0.009848827496170998, + 0.006897669285535812, + 0.004740293137729168, + 0.007991448044776917, + 0.012476031668484211, + 0.005852208472788334, + 0.006536555476486683, + 0.006425802130252123, + 0.006503594107925892, + 0.005086988210678101, + 0.005246900487691164, + 0.004907384514808655, + 0.0051584686152637005, + 0.006137649063020945, + 0.00601534266024828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02057565003633499 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067322817631065845 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.0067322817631065845, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067322817631065845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2237585484981537, + 0.13819636404514313, + 0.05191043019294739, + 0.13094620406627655, + 0.13586950302124023, + 0.05286674574017525, + 0.0323357991874218, + 0.0545986071228981, + 0.07797601073980331, + 0.10154185444116592 + ], + "credit_map": [ + 0.05977504327893257, + 0.05978003889322281, + 0.015225433744490147, + 0.012522631324827671, + 0.01339029148221016, + 0.02380368486046791, + 0.025035424157977104, + 0.01422598771750927, + 0.01863113045692444, + 0.02192513272166252, + 0.013197027146816254, + 0.012711425311863422, + 0.011234156787395477, + 0.03153124451637268, + 0.012575238943099976, + 0.01639101468026638, + 0.006460397969931364, + 0.006437904201447964, + 0.01105885487049818, + 0.00677443714812398, + 0.004689216613769531, + 0.005720873363316059, + 0.005267032887786627, + 0.005501710809767246, + 0.021390268579125404, + 0.031459905207157135, + 0.01560189202427864, + 0.009488007053732872, + 0.009953751228749752, + 0.0192062109708786, + 0.009795394726097584, + 0.01405076589435339, + 0.02031579427421093, + 0.026504937559366226, + 0.021099848672747612, + 0.015623827464878559, + 0.010740909725427628, + 0.01348934881389141, + 0.011778358370065689, + 0.016316482797265053, + 0.00950464978814125, + 0.007547575049102306, + 0.0077228550799191, + 0.005709094926714897, + 0.005967039614915848, + 0.007174338214099407, + 0.004021167289465666, + 0.005220025312155485, + 0.004255172796547413, + 0.0029362121131271124, + 0.004947287496179342, + 0.004383930470794439, + 0.002453787485137582, + 0.004413379821926355, + 0.0046249935403466225, + 0.004321034997701645, + 0.008079318329691887, + 0.005759420804679394, + 0.010952714830636978, + 0.005896711256355047, + 0.004686801694333553, + 0.005725475028157234, + 0.00702698016539216, + 0.006471187341958284, + 0.012288965284824371, + 0.010766037739813328, + 0.016448408365249634, + 0.007069809827953577, + 0.006700436118990183, + 0.008961336687207222, + 0.008394131436944008, + 0.0073468852788209915, + 0.01544627919793129, + 0.013392065651714802, + 0.014590498059988022, + 0.010041982866823673, + 0.00978170707821846, + 0.00861432682722807, + 0.019523898139595985, + 0.010151097550988197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0067322817631065845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An incredibly sharp 8K macro close-up, shot at eye level, locks the viewer's gaze firmly on a South American leaf mantis deep in the gloom of a forest. Crouching alone on a decaying log covered in dark veins and damp moss, the mantis's form blends seamlessly with its surroundings: its trunk, a tan hue reminiscent of a dead leaf, is dotted with incredibly realistic veins and spots imitating insect damage. Its edges, jagged and irregular, appear to be weathered, making it visually indistinguishable from a real leaf. With its sickle-shaped forelimbs tucked tightly against its chest and its body lowered, it lurks patiently, motionless. A few faint rays of light penetrate the dense forest canopy, creating a soft, dappled glow that perfectly captures the subtle contours of the mantis's back and the rough texture of the decaying wood, against the backdrop of an out-of-focus, deep green and brown forest floor. With extreme clarity, one can even discern the subtle, almost imperceptible rhythmic rise and fall of its abdomen as it breathes. The dominant tones of the painting are deep brown, ochre, and dark green, creating an atmosphere of profound tranquility, a near-frozen loneliness, and the desolate presence of life stubbornly persisting in hidden corners.", + "seed": 5683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_13/index_7079_seed_5683_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027780581917613744, + "vlm_logprob": -0.027780581917613744, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.027780581917613744, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07581259310245514 + ], + "opposite_logprob_tokens": [ + -2.700812578201294 + ], + "logprob_reward": -0.07581259310245514, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07581259310245514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10222022980451584, + 0.12839047610759735, + 0.036080051213502884, + 0.11728863418102264, + 0.09661659598350525, + 0.17052386701107025, + 0.14366906881332397, + 0.06885658949613571, + 0.07776506245136261, + 0.05858946964144707 + ], + "credit_map": [ + 0.01683717593550682, + 0.008369388990104198, + 0.010602048598229885, + 0.012343387119472027, + 0.014902007766067982, + 0.01472873892635107, + 0.015950264409184456, + 0.008487205021083355, + 0.015629388391971588, + 0.015021607279777527, + 0.012903358787298203, + 0.01408588420599699, + 0.022664833813905716, + 0.014242006465792656, + 0.024894127622246742, + 0.008949254639446735, + 0.0038784865755587816, + 0.003915031440556049, + 0.0041754343546926975, + 0.004946907050907612, + 0.004502493422478437, + 0.004090589936822653, + 0.00528075871989131, + 0.005290347151458263, + 0.015401721931993961, + 0.01170728076249361, + 0.011656217277050018, + 0.01791255921125412, + 0.014523811638355255, + 0.015658827498555183, + 0.018799789249897003, + 0.011628419160842896, + 0.018614863976836205, + 0.009480688720941544, + 0.013757921755313873, + 0.010670546442270279, + 0.010980640538036823, + 0.011780441738665104, + 0.010334959253668785, + 0.010996525175869465, + 0.021796541288495064, + 0.022949645295739174, + 0.025096407160162926, + 0.020597368478775024, + 0.018976960331201553, + 0.019865931943058968, + 0.019144663587212563, + 0.022096335887908936, + 0.015854211524128914, + 0.011990390717983246, + 0.023456694558262825, + 0.01716691255569458, + 0.01854701340198517, + 0.01587856374680996, + 0.016995785757899284, + 0.02377946674823761, + 0.010923621244728565, + 0.0122748464345932, + 0.007839201018214226, + 0.0074984245002269745, + 0.008531607687473297, + 0.006250175181776285, + 0.005326006095856428, + 0.01021269615739584, + 0.014204563573002815, + 0.013429481536149979, + 0.009263616055250168, + 0.009152501821517944, + 0.007171738892793655, + 0.008013754151761532, + 0.006855746731162071, + 0.0096736503764987, + 0.007277266122400761, + 0.009320579469203949, + 0.005397544242441654, + 0.004509517922997475, + 0.008133724331855774, + 0.006810735911130905, + 0.007556419353932142, + 0.009583673439919949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.07581259310245514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746135652065277 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746135652065277, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746135652065277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15589968860149384, + 0.08779913187026978, + 0.039279092103242874, + 0.1598348319530487, + 0.0898006483912468, + 0.15530508756637573, + 0.07409568876028061, + 0.08144214749336243, + 0.10143307596445084, + 0.05511061102151871 + ], + "credit_map": [ + 0.02334846556186676, + 0.012504718266427517, + 0.027053162455558777, + 0.02061540260910988, + 0.01039018202573061, + 0.020944586023688316, + 0.018839552998542786, + 0.022203627973794937, + 0.015213049948215485, + 0.008609933778643608, + 0.012378144077956676, + 0.009394367225468159, + 0.009551282972097397, + 0.011541616171598434, + 0.013285733759403229, + 0.00782499834895134, + 0.005886430386453867, + 0.004773969762027264, + 0.0058402386493980885, + 0.005147417075932026, + 0.003817567601799965, + 0.004382941406220198, + 0.004326253198087215, + 0.005104274023324251, + 0.024894503876566887, + 0.01674777641892433, + 0.019234977662563324, + 0.02342401072382927, + 0.016666652634739876, + 0.02021363377571106, + 0.023972170427441597, + 0.014681096188724041, + 0.01966787874698639, + 0.0098269646987319, + 0.01301773265004158, + 0.010910199955105782, + 0.008576310239732265, + 0.0116443932056427, + 0.008153488859534264, + 0.008003679104149342, + 0.019461307674646378, + 0.019540004432201385, + 0.02216402254998684, + 0.018321944400668144, + 0.019429218024015427, + 0.017784148454666138, + 0.018096011132001877, + 0.020508430898189545, + 0.008603145368397236, + 0.012416421435773373, + 0.011023481376469135, + 0.008685395121574402, + 0.008562730625271797, + 0.010222690179944038, + 0.0077012889087200165, + 0.0068805343471467495, + 0.010274451225996017, + 0.012623501941561699, + 0.010375060141086578, + 0.010126564651727676, + 0.010092271491885185, + 0.007755506318062544, + 0.008444476872682571, + 0.011750313453376293, + 0.017063738778233528, + 0.014648237265646458, + 0.011647187173366547, + 0.01300889067351818, + 0.010562275536358356, + 0.009792856872081757, + 0.011119095608592033, + 0.013590788468718529, + 0.006625395268201828, + 0.009354640729725361, + 0.0070754410699009895, + 0.004808179102838039, + 0.006913706660270691, + 0.006291853729635477, + 0.0067161438055336475, + 0.007325248792767525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.006746135652065277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982458658516407 + ], + "opposite_logprob_tokens": [ + -3.5298244953155518 + ], + "logprob_reward": -0.02982458658516407, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982458658516407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19435296952724457, + 0.07466837018728256, + 0.044174645096063614, + 0.08875475078821182, + 0.12426191568374634, + 0.10404369235038757, + 0.174912229180336, + 0.06788770854473114, + 0.07992949336767197, + 0.04701423645019531 + ], + "credit_map": [ + 0.03039678744971752, + 0.018567414954304695, + 0.033376336097717285, + 0.020470798015594482, + 0.014094463549554348, + 0.02576557546854019, + 0.01913345791399479, + 0.032548144459724426, + 0.010598869062960148, + 0.00801808200776577, + 0.00905426125973463, + 0.007682206109166145, + 0.011121425777673721, + 0.009491094388067722, + 0.013228426687419415, + 0.005473997909575701, + 0.005241685546934605, + 0.00434456579387188, + 0.007366048637777567, + 0.006415894255042076, + 0.005917710717767477, + 0.004397459793835878, + 0.005184757057577372, + 0.005306522361934185, + 0.01708659902215004, + 0.008502454496920109, + 0.011487334035336971, + 0.012610901147127151, + 0.008002985268831253, + 0.013842149637639523, + 0.010702503845095634, + 0.00651982706040144, + 0.02429051883518696, + 0.012675292789936066, + 0.01933451183140278, + 0.014482381753623486, + 0.013405892997980118, + 0.015876363962888718, + 0.012465791776776314, + 0.011731159873306751, + 0.012248672544956207, + 0.015199427492916584, + 0.013796795159578323, + 0.011691848747432232, + 0.013923048041760921, + 0.011991746723651886, + 0.011018342338502407, + 0.014173813164234161, + 0.01844368688762188, + 0.01655290648341179, + 0.021474339067935944, + 0.02116052247583866, + 0.024974649772047997, + 0.020349085330963135, + 0.02475619874894619, + 0.027200844138860703, + 0.009563584811985493, + 0.012645727023482323, + 0.008102512918412685, + 0.007871817797422409, + 0.008315082639455795, + 0.006550349295139313, + 0.005937718786299229, + 0.008900911547243595, + 0.009626117534935474, + 0.011097731068730354, + 0.0105467289686203, + 0.012323121540248394, + 0.00859773438423872, + 0.007101733703166246, + 0.007691231556236744, + 0.012945091351866722, + 0.005280317738652229, + 0.0060471356846392155, + 0.005034152418375015, + 0.004282809793949127, + 0.005864227656275034, + 0.0049230423755943775, + 0.006050494033843279, + 0.00953205581754446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.02982458658516407 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020576467737555504 + ], + "opposite_logprob_tokens": [ + -3.8955764770507812 + ], + "logprob_reward": -0.020576467737555504, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020576467737555504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18403677642345428, + 0.16221138834953308, + 0.06445978581905365, + 0.09739715605974197, + 0.0971352607011795, + 0.08082487434148788, + 0.0844540074467659, + 0.07225047051906586, + 0.052487004548311234, + 0.10474330931901932 + ], + "credit_map": [ + 0.024502737447619438, + 0.016727706417441368, + 0.03118736483156681, + 0.0231722854077816, + 0.01696617901325226, + 0.02262607216835022, + 0.024024948477745056, + 0.02482946589589119, + 0.015516885556280613, + 0.012131502851843834, + 0.024978702887892723, + 0.02119452692568302, + 0.025651218369603157, + 0.017852604389190674, + 0.029738951474428177, + 0.015146986581385136, + 0.007511891890317202, + 0.01021881215274334, + 0.007061426993459463, + 0.009125829674303532, + 0.007203828543424606, + 0.0075136045925319195, + 0.00784008763730526, + 0.00798430573195219, + 0.017446866258978844, + 0.009014835581183434, + 0.014101414009928703, + 0.012440867722034454, + 0.010519463568925858, + 0.014064854010939598, + 0.01046122144907713, + 0.009347628802061081, + 0.017202503979206085, + 0.010115403681993484, + 0.010072482749819756, + 0.01288394071161747, + 0.015191747806966305, + 0.011154184117913246, + 0.010846810415387154, + 0.009668190032243729, + 0.009197434410452843, + 0.0067607806995511055, + 0.01536084059625864, + 0.008738595992326736, + 0.00769278546795249, + 0.011129334568977356, + 0.007461088243871927, + 0.014484016224741936, + 0.008725033141672611, + 0.009935255162417889, + 0.010894919745624065, + 0.010009423829615116, + 0.012729841284453869, + 0.010636932216584682, + 0.011599243618547916, + 0.009923355653882027, + 0.011032926850020885, + 0.011682297103106976, + 0.00990724191069603, + 0.008741267025470734, + 0.006344152148813009, + 0.006822030525654554, + 0.006632561795413494, + 0.011087991297245026, + 0.006039245054125786, + 0.007916463539004326, + 0.006697504315525293, + 0.005513186100870371, + 0.005837003700435162, + 0.0061437515541911125, + 0.005940089467912912, + 0.008399756625294685, + 0.017833223566412926, + 0.024949830025434494, + 0.009175299666821957, + 0.010161628015339375, + 0.011229745112359524, + 0.009937561117112637, + 0.009660646319389343, + 0.01179537083953619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.020576467737555504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059431265108287334 + ], + "opposite_logprob_tokens": [ + -5.130943298339844 + ], + "logprob_reward": -0.0059431265108287334, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059431265108287334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09293406456708908, + 0.09700489044189453, + 0.04122956842184067, + 0.18601474165916443, + 0.03410351648926735, + 0.11396284401416779, + 0.1285635530948639, + 0.13968691229820251, + 0.13586339354515076, + 0.030636535957455635 + ], + "credit_map": [ + 0.01626250334084034, + 0.00918496586382389, + 0.01546070072799921, + 0.010175283066928387, + 0.007685866206884384, + 0.012775597162544727, + 0.009084044024348259, + 0.012305106036365032, + 0.010546485893428326, + 0.012276140041649342, + 0.012423614040017128, + 0.008764686062932014, + 0.01070475298911333, + 0.014295075088739395, + 0.015908751636743546, + 0.0120853865519166, + 0.005634698085486889, + 0.005990464705973864, + 0.005933922249823809, + 0.004271204117685556, + 0.005490161944180727, + 0.005029832478612661, + 0.004045954439789057, + 0.004833330865949392, + 0.02818894200026989, + 0.015433545224368572, + 0.02569710463285446, + 0.028736736625432968, + 0.024844564497470856, + 0.023623015731573105, + 0.02294149436056614, + 0.016549337655305862, + 0.004750974476337433, + 0.0036434109788388014, + 0.003923870623111725, + 0.004989129025489092, + 0.005083766300231218, + 0.0037264416459947824, + 0.004487145226448774, + 0.003498778911307454, + 0.014015116728842258, + 0.010790390893816948, + 0.01838938519358635, + 0.014403837732970715, + 0.014527794905006886, + 0.013279123231768608, + 0.01226053573191166, + 0.016296664252877235, + 0.01429423876106739, + 0.014967899769544601, + 0.015602482482790947, + 0.014077368192374706, + 0.017815781757235527, + 0.016114847734570503, + 0.018080735579133034, + 0.017610201612114906, + 0.022070445120334625, + 0.02321789041161537, + 0.018459105864167213, + 0.015866253525018692, + 0.013749079778790474, + 0.013809776864945889, + 0.011355573311448097, + 0.021158795803785324, + 0.02157699130475521, + 0.02801293507218361, + 0.022139329463243484, + 0.01567423902451992, + 0.010412418283522129, + 0.015232150442898273, + 0.008608346804976463, + 0.014206982217729092, + 0.007282864768058062, + 0.004838485270738602, + 0.0032685031183063984, + 0.0019930980633944273, + 0.0032915009651333094, + 0.0035429943818598986, + 0.002916610799729824, + 0.0035024757962673903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.0059431265108287334 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An incredibly sharp 8K macro close-up, shot at eye level, locks the viewer's gaze firmly on a South American leaf mantis deep in the gloom of a forest. Crouching alone on a decaying log covered in dark veins and damp moss, the mantis's form blends seamlessly with its surroundings: its trunk, a tan hue reminiscent of a dead leaf, is dotted with incredibly realistic veins and spots imitating insect damage. Its edges, jagged and irregular, appear to be weathered, making it visually indistinguishable from a real leaf. With its sickle-shaped forelimbs tucked tightly against its chest and its body lowered, it lurks patiently, motionless. A few faint rays of light penetrate the dense forest canopy, creating a soft, dappled glow that perfectly captures the subtle contours of the mantis's back and the rough texture of the decaying wood, against the backdrop of an out-of-focus, deep green and brown forest floor. With extreme clarity, one can even discern the subtle, almost imperceptible rhythmic rise and fall of its abdomen as it breathes. The dominant tones of the painting are deep brown, ochre, and dark green, creating an atmosphere of profound tranquility, a near-frozen loneliness, and the desolate presence of life stubbornly persisting in hidden corners.", + "seed": 5684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_13/index_7079_seed_5684_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03116120519116521, + "vlm_logprob": -0.03116120519116521, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03116120519116521, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09564165771007538 + ], + "opposite_logprob_tokens": [ + -2.470641613006592 + ], + "logprob_reward": -0.09564165771007538, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09564165771007538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39113497734069824, + 0.10152482241392136, + 0.06538064032793045, + 0.06746144592761993, + 0.0820433646440506, + 0.07034741342067719, + 0.028475385159254074, + 0.12258313596248627, + 0.03197892755270004, + 0.03906995803117752 + ], + "credit_map": [ + 0.11106864362955093, + 0.03592450171709061, + 0.02521841786801815, + 0.02223345637321472, + 0.014172742143273354, + 0.0275746900588274, + 0.034695204347372055, + 0.12024731934070587, + 0.015435568057000637, + 0.011960149742662907, + 0.01060838159173727, + 0.011612030677497387, + 0.010068248026072979, + 0.016608893871307373, + 0.01394182164222002, + 0.011289718560874462, + 0.008844107389450073, + 0.008635408245027065, + 0.006882003508508205, + 0.0071846735663712025, + 0.009110339917242527, + 0.006650012452155352, + 0.00850265845656395, + 0.009571432135999203, + 0.010118957608938217, + 0.006888567004352808, + 0.007500081788748503, + 0.009957505390048027, + 0.010532875545322895, + 0.007095661014318466, + 0.008349331095814705, + 0.007018457632511854, + 0.011801660060882568, + 0.01109478622674942, + 0.010835584253072739, + 0.01439625769853592, + 0.007463931571692228, + 0.007396217435598373, + 0.012232031673192978, + 0.006822890602052212, + 0.011206920258700848, + 0.008756917901337147, + 0.010268122889101505, + 0.0063322195783257484, + 0.004819396883249283, + 0.016060395166277885, + 0.005737250670790672, + 0.007166183553636074, + 0.0042107123881578445, + 0.0037274330388754606, + 0.0047611817717552185, + 0.004241099115461111, + 0.0027128932997584343, + 0.0030655532609671354, + 0.0027632524725049734, + 0.0029932563193142414, + 0.03147948533296585, + 0.00622277008369565, + 0.04070693626999855, + 0.008718185126781464, + 0.013060702942311764, + 0.0079047791659832, + 0.009118407033383846, + 0.0053718602284789085, + 0.004398110322654247, + 0.006090731825679541, + 0.004791795741766691, + 0.003976141102612019, + 0.0026709611993283033, + 0.0047266604378819466, + 0.0025132405571639538, + 0.002811283804476261, + 0.005202396307140589, + 0.0024528515059500933, + 0.0043368092738091946, + 0.0052641453221440315, + 0.003586792852729559, + 0.004414421506226063, + 0.008606025017797947, + 0.005206515546888113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.09564165771007538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008652104064822197 + ], + "opposite_logprob_tokens": [ + -4.758652210235596 + ], + "logprob_reward": -0.008652104064822197, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652104064822197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39895114302635193, + 0.07640301436185837, + 0.09034187346696854, + 0.09276407212018967, + 0.06118299812078476, + 0.0619407594203949, + 0.04818137362599373, + 0.09243509918451309, + 0.04293476790189743, + 0.03486485406756401 + ], + "credit_map": [ + 0.12471477687358856, + 0.031844768673181534, + 0.017506837844848633, + 0.02159896306693554, + 0.00986152607947588, + 0.024061178788542747, + 0.03370101749897003, + 0.13566210865974426, + 0.011213579215109348, + 0.007311017252504826, + 0.006905000656843185, + 0.008723205886781216, + 0.006692059803754091, + 0.013724054209887981, + 0.00786520168185234, + 0.013968896120786667, + 0.013446880504488945, + 0.0154885770753026, + 0.006393956486135721, + 0.008064492605626583, + 0.009472612291574478, + 0.008264610543847084, + 0.01144558284431696, + 0.01776515692472458, + 0.014198873192071915, + 0.008053087629377842, + 0.009837794117629528, + 0.01810349151492119, + 0.010296878404915333, + 0.009839191101491451, + 0.013674372807145119, + 0.008760390803217888, + 0.009115906432271004, + 0.009372072294354439, + 0.00894369836896658, + 0.008464043959975243, + 0.0062975273467600346, + 0.006208566948771477, + 0.007631317712366581, + 0.005149866919964552, + 0.01012718491256237, + 0.007789480034261942, + 0.008688164874911308, + 0.005301286466419697, + 0.004778939299285412, + 0.014881260693073273, + 0.005109838210046291, + 0.005264605861157179, + 0.008838809095323086, + 0.0040059769526124, + 0.010613507591187954, + 0.006536046974360943, + 0.006019429303705692, + 0.003970794379711151, + 0.005384952761232853, + 0.0028118586633354425, + 0.01959920860826969, + 0.009112818166613579, + 0.02314990945160389, + 0.01068843249231577, + 0.00817741360515356, + 0.006582357455044985, + 0.009553251788020134, + 0.0055716997012495995, + 0.006449274253100157, + 0.006442059762775898, + 0.007527950685471296, + 0.005730546545237303, + 0.0032893610186874866, + 0.005498691461980343, + 0.004241155926138163, + 0.0037557308096438646, + 0.0053567891009151936, + 0.0047391424886882305, + 0.006427923683077097, + 0.005304852966219187, + 0.0019469703547656536, + 0.0023455170448869467, + 0.005871365312486887, + 0.0028722931165248156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.008652104064822197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023312414065003395 + ], + "opposite_logprob_tokens": [ + -3.7733123302459717 + ], + "logprob_reward": -0.023312414065003395, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023312414065003395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5851878523826599, + 0.09816417098045349, + 0.04494694992899895, + 0.06756218522787094, + 0.05061059072613716, + 0.02590378373861313, + 0.03795017674565315, + 0.030686507001519203, + 0.030438272282481194, + 0.02854953147470951 + ], + "credit_map": [ + 0.18335436284542084, + 0.04590962454676628, + 0.02887520007789135, + 0.03343113884329796, + 0.01435240637511015, + 0.03318529576063156, + 0.048807527869939804, + 0.19727227091789246, + 0.014873798005282879, + 0.009653037413954735, + 0.013597637414932251, + 0.012809781357645988, + 0.007339858449995518, + 0.011405229568481445, + 0.015051763504743576, + 0.0134330615401268, + 0.0060980250127613544, + 0.005439919885247946, + 0.004131232853978872, + 0.005009117536246777, + 0.005124714691191912, + 0.004578659310936928, + 0.005926188081502914, + 0.008639094419777393, + 0.010277153924107552, + 0.007076641079038382, + 0.008089100010693073, + 0.009952053427696228, + 0.009117801673710346, + 0.006709096021950245, + 0.007944715209305286, + 0.008395622484385967, + 0.00722252344712615, + 0.007363826036453247, + 0.007396019995212555, + 0.00846803281456232, + 0.004922307096421719, + 0.005212321411818266, + 0.005806072615087032, + 0.004219489172101021, + 0.004066808614879847, + 0.0028371307998895645, + 0.003278308780863881, + 0.003226611763238907, + 0.0017619053833186626, + 0.005295648705214262, + 0.003136949148029089, + 0.0023004189133644104, + 0.005853119771927595, + 0.00485076941549778, + 0.0053649484179914, + 0.006533156614750624, + 0.003074072068557143, + 0.003921614959836006, + 0.004578766413033009, + 0.0037737255915999413, + 0.005279775708913803, + 0.0031032911501824856, + 0.005809232126921415, + 0.0035735787823796272, + 0.003184167668223381, + 0.003189863869920373, + 0.0039230212569236755, + 0.0026235757395625114, + 0.004149051383137703, + 0.006735735107213259, + 0.004480840638279915, + 0.0035075254272669554, + 0.0031928608659654856, + 0.002717159455642104, + 0.0030328412540256977, + 0.002622259547933936, + 0.004648814909160137, + 0.004388523753732443, + 0.005308083258569241, + 0.004626321140676737, + 0.0019226135918870568, + 0.0027580077294260263, + 0.0029200862627476454, + 0.0019770823419094086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023312414065003395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577052608132362 + ], + "opposite_logprob_tokens": [ + -3.8955769538879395 + ], + "logprob_reward": -0.020577052608132362, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577052608132362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10792749375104904, + 0.14402985572814941, + 0.10433609783649445, + 0.10743571817874908, + 0.19330552220344543, + 0.19557638466358185, + 0.03147942200303078, + 0.03399289399385452, + 0.05795435234904289, + 0.023962171748280525 + ], + "credit_map": [ + 0.03199814632534981, + 0.007470452226698399, + 0.005969229154288769, + 0.007769731804728508, + 0.004196077585220337, + 0.0070034656673669815, + 0.009046003222465515, + 0.03447439521551132, + 0.023241667076945305, + 0.018105557188391685, + 0.01375417597591877, + 0.018227528780698776, + 0.01426443550735712, + 0.022478889673948288, + 0.018093397840857506, + 0.01586422137916088, + 0.013054274953901768, + 0.011529439128935337, + 0.014040015637874603, + 0.012949290685355663, + 0.015383007004857063, + 0.012232341803610325, + 0.013911335729062557, + 0.0112364012748003, + 0.015853527933359146, + 0.011044174432754517, + 0.009970135055482388, + 0.020935211330652237, + 0.013762389309704304, + 0.014132238924503326, + 0.013563892804086208, + 0.008174157701432705, + 0.0277729369699955, + 0.029630886390805244, + 0.023907391354441643, + 0.02907349355518818, + 0.022184019908308983, + 0.020597433671355247, + 0.024260323494672775, + 0.015879064798355103, + 0.03244812786579132, + 0.02692415751516819, + 0.02674718014895916, + 0.017480313777923584, + 0.01348247192800045, + 0.04451656714081764, + 0.01666931062936783, + 0.017308278009295464, + 0.006044470705091953, + 0.0023893113248050213, + 0.007377303671091795, + 0.005193460267037153, + 0.0023064441047608852, + 0.0024097429122775793, + 0.0035045675467699766, + 0.0022541244979947805, + 0.00576081732288003, + 0.0034461773466318846, + 0.0046105836518108845, + 0.0034582398366183043, + 0.00573755893856287, + 0.0027551953680813313, + 0.004535404033958912, + 0.003688921919092536, + 0.009497114457190037, + 0.010826349258422852, + 0.013329870998859406, + 0.005831526126712561, + 0.005262223072350025, + 0.0044618952088057995, + 0.004769024904817343, + 0.003976349253207445, + 0.0033486918546259403, + 0.004479831084609032, + 0.003945339936763048, + 0.003416589926928282, + 0.0016093485755845904, + 0.0018846370512619615, + 0.003119378350675106, + 0.0021583556663244963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.020577052608132362 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a South American leaf mantis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mantis crouching on a decaying log covered in dark veins and damp moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mantis have a tan trunk with realistic veins and spots imitating insect damage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the mantis motionless with its sickle-shaped forelimbs tucked tightly against its chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by faint rays of light creating a soft, dappled glow against a deep green and brown out-of-focus forest floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622797507792711 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622797507792711, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622797507792711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24075451493263245, + 0.0967346727848053, + 0.04337962344288826, + 0.0973583534359932, + 0.12745237350463867, + 0.1815311759710312, + 0.03220595791935921, + 0.08471288532018661, + 0.042739830911159515, + 0.05313056334853172 + ], + "credit_map": [ + 0.07341942936182022, + 0.015564864501357079, + 0.011918101459741592, + 0.013449154794216156, + 0.009127163328230381, + 0.01699410378932953, + 0.019337406381964684, + 0.08094431459903717, + 0.013903949409723282, + 0.011319230310618877, + 0.00961230881512165, + 0.015813255682587624, + 0.009849713183939457, + 0.012142102234065533, + 0.009472415782511234, + 0.014621702022850513, + 0.006291295401751995, + 0.004980470519512892, + 0.00455944798886776, + 0.004134306684136391, + 0.004186984151601791, + 0.005263278726488352, + 0.006739748176187277, + 0.0072240945883095264, + 0.014076083898544312, + 0.007806691341102123, + 0.011863273568451405, + 0.01662726327776909, + 0.01209141407161951, + 0.0143349664285779, + 0.011766165494918823, + 0.008792512118816376, + 0.01936805061995983, + 0.014454314485192299, + 0.014609161764383316, + 0.016072116792201996, + 0.011332624591886997, + 0.022266460582613945, + 0.017885228618979454, + 0.011464416980743408, + 0.031022854149341583, + 0.023330768570303917, + 0.024448124691843987, + 0.014941055327653885, + 0.010637802071869373, + 0.04446570202708244, + 0.016695935279130936, + 0.01598895899951458, + 0.005102373193949461, + 0.0028774933889508247, + 0.0063475859351456165, + 0.005572992376983166, + 0.003049962455406785, + 0.003163788700476289, + 0.002677167532965541, + 0.0034145985264331102, + 0.013613604940474033, + 0.009370162151753902, + 0.015851743519306183, + 0.014440716244280338, + 0.008381103165447712, + 0.004540857393294573, + 0.011835369281470776, + 0.006679333746433258, + 0.007576021831482649, + 0.008298719301819801, + 0.00791222508996725, + 0.0034575709141790867, + 0.003151887096464634, + 0.004035547375679016, + 0.004484770819544792, + 0.0038230912759900093, + 0.0073521374724805355, + 0.00653238408267498, + 0.006295119412243366, + 0.006977238692343235, + 0.004735415801405907, + 0.006434784270823002, + 0.008577404543757439, + 0.0062260800041258335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.007622797507792711 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_14/index_tensor([6429, 6429])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_14/index_tensor([6429, 6429])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cb74f3466901de8d3aadf3980463a4e26201c3b6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_14/index_tensor([6429, 6429])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a high-angle, overhead shot, the initial image abstract and poetic. Raindrops pummel the deep, dark surface of a lake, stirring up concentric ripples that radiate outward. The water's surface simultaneously reflects the cold, strip-shaped lights within the laboratory, creating a distorted, interwoven tapestry of cool blue light and shadow. Then, the camera slowly and steadily zooms in, penetrating this illusion of light and shadow to focus on the protagonist, slumped over on the cold terrazzo floor. He leans limply against a metal wall covered in instruments. A hand, trembling violently with weakness, laboriously, almost spasmodically, unscrews a white plastic bottle of fluoxetine bearing a prescription label. As the cap slips loose, several small white pills tumble out, impacting precisely on the toes of a pair of rain-soaked, dark canvas shoes at his feet. Not far from the protagonist, a glass beaker of chemical solution overflows uncontrollably, producing a fine, viscous white foam that shimmers eerily under the cool blue light cast by the overhead instruments. Through the huge glass window covered with water vapor behind the character, you can see that the rain outside the window is weakening. At the same time, the waveform line on the screen of an electrocardiogram monitor tends to be flat, and the sharp, continuous beeping sound it emits gradually drowns out the sound of rain, injecting the entire picture with an extremely tense atmosphere of despair, loss of control and medical emergency.", + "seed": 1210, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_14/index_6429_seed_1210_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05753497295081615, + "vlm_logprob": -0.05753497295081615, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.05753497295081615, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03654509410262108 + ], + "opposite_logprob_tokens": [ + -3.4115450382232666 + ], + "logprob_reward": -0.03654509410262108, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03654509410262108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3313111960887909, + 0.06785408407449722, + 0.3540729284286499, + 0.07306738197803497, + 0.01815708540380001, + 0.02839699201285839, + 0.059124309569597244, + 0.019623741507530212, + 0.01647716946899891, + 0.031915199011564255 + ], + "credit_map": [ + 0.023876244202256203, + 0.04595491662621498, + 0.025954557582736015, + 0.08085987716913223, + 0.03331248462200165, + 0.024251431226730347, + 0.06363202631473541, + 0.03346966207027435, + 0.008275256492197514, + 0.009182113222777843, + 0.0075635043904185295, + 0.006164147052913904, + 0.010996785946190357, + 0.009384015575051308, + 0.008492128923535347, + 0.00779613247141242, + 0.018559424206614494, + 0.014030935242772102, + 0.09547021239995956, + 0.026598388329148293, + 0.02586272358894348, + 0.011637695133686066, + 0.13644057512283325, + 0.025472959503531456, + 0.0056449053809046745, + 0.010938304476439953, + 0.005137893836945295, + 0.01011297944933176, + 0.006690686102956533, + 0.00748470239341259, + 0.015123169869184494, + 0.011934742331504822, + 0.0017969777109101415, + 0.002952858339995146, + 0.001888750004582107, + 0.0027273050509393215, + 0.0020795962773263454, + 0.0016442934283986688, + 0.001984955044463277, + 0.0030823489651083946, + 0.002777551067993045, + 0.003323728684335947, + 0.003698711283504963, + 0.00403059646487236, + 0.0026248230133205652, + 0.0025206601712852716, + 0.005190882366150618, + 0.004230038728564978, + 0.004327978007495403, + 0.0021200920455157757, + 0.018253900110721588, + 0.006305955816060305, + 0.0030091109219938517, + 0.006249639205634594, + 0.006372452713549137, + 0.012485177256166935, + 0.001851472887210548, + 0.004119489341974258, + 0.0020646005868911743, + 0.0018352566985413432, + 0.0028001207392662764, + 0.002892531221732497, + 0.002132316352799535, + 0.0019279533298686147, + 0.0023753175046294928, + 0.002386971842497587, + 0.0015686299884691834, + 0.0021700426004827023, + 0.002687853993847966, + 0.0009561763727106154, + 0.002150941640138626, + 0.0021812350023537874, + 0.0036860888358205557, + 0.0033809675369411707, + 0.0038175825029611588, + 0.004415759816765785, + 0.004014686215668917, + 0.0024493024684488773, + 0.006694929674267769, + 0.0034558833576738834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.03654509410262108 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902193814516068 + ], + "opposite_logprob_tokens": [ + -2.57902193069458 + ], + "logprob_reward": -0.07902193814516068, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902193814516068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05156315863132477, + 0.094504214823246, + 0.19983263313770294, + 0.11005599796772003, + 0.04852915555238724, + 0.14073488116264343, + 0.10835593938827515, + 0.12141270190477371, + 0.07738908380270004, + 0.04762223735451698 + ], + "credit_map": [ + 0.005069965962320566, + 0.005746711511164904, + 0.00404125964269042, + 0.011484118178486824, + 0.005064353812485933, + 0.005629367660731077, + 0.008219392038881779, + 0.0063079907558858395, + 0.011276819743216038, + 0.011137877590954304, + 0.01262676902115345, + 0.008656715042889118, + 0.01686285249888897, + 0.011690890416502953, + 0.010991665534675121, + 0.011260629631578922, + 0.01037630531936884, + 0.010243105702102184, + 0.05400506407022476, + 0.015616115182638168, + 0.011720838025212288, + 0.007819085381925106, + 0.07545328885316849, + 0.014598838984966278, + 0.0067222542129457, + 0.01447291485965252, + 0.008788924664258957, + 0.01500686164945364, + 0.012912945821881294, + 0.011548579670488834, + 0.024248376488685608, + 0.01635514572262764, + 0.004579402506351471, + 0.004740018397569656, + 0.005503992550075054, + 0.008860926143825054, + 0.004362835083156824, + 0.004644173663109541, + 0.0070874691009521484, + 0.008750341832637787, + 0.010154146701097488, + 0.015642661601305008, + 0.018621109426021576, + 0.0196799635887146, + 0.014936933293938637, + 0.012489951215684414, + 0.026707714423537254, + 0.022502407431602478, + 0.007894277572631836, + 0.006224482320249081, + 0.028790434822440147, + 0.008630694821476936, + 0.006211841478943825, + 0.010164694860577583, + 0.011516502127051353, + 0.02892300672829151, + 0.014355015940964222, + 0.018177274614572525, + 0.01421597134321928, + 0.008945364505052567, + 0.022838380187749863, + 0.011046415194869041, + 0.019487489014863968, + 0.012346796691417694, + 0.008806015364825726, + 0.009423509240150452, + 0.007004186045378447, + 0.012995035387575626, + 0.01219203881919384, + 0.00513141555711627, + 0.013178600929677486, + 0.008658282458782196, + 0.004191902466118336, + 0.006509228143841028, + 0.0057366276159882545, + 0.008710648864507675, + 0.007823901250958443, + 0.0037449176888912916, + 0.006917994935065508, + 0.003987014293670654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.07902193814516068 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309023171663284 + ], + "opposite_logprob_tokens": [ + -3.168090343475342 + ], + "logprob_reward": -0.04309023171663284, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309023171663284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18059749901294708, + 0.057897619903087616, + 0.21920989453792572, + 0.03470688685774803, + 0.04498923942446709, + 0.099824920296669, + 0.07945777475833893, + 0.11265120655298233, + 0.027057930827140808, + 0.14360705018043518 + ], + "credit_map": [ + 0.014033039100468159, + 0.02461204305291176, + 0.014562524855136871, + 0.04394407942891121, + 0.017451170831918716, + 0.015116265043616295, + 0.03198239579796791, + 0.018895970657467842, + 0.007939280942082405, + 0.008714007213711739, + 0.006244612392038107, + 0.005067829042673111, + 0.009938911534845829, + 0.00713659543544054, + 0.007282007019966841, + 0.005574379116296768, + 0.011851954273879528, + 0.007971846498548985, + 0.06162353605031967, + 0.01679217629134655, + 0.013847325928509235, + 0.006487070582807064, + 0.08445455878973007, + 0.016181431710720062, + 0.002694574184715748, + 0.003948130179196596, + 0.0028145709075033665, + 0.004695299547165632, + 0.003735772566869855, + 0.004501149989664555, + 0.007740634493529797, + 0.0045767552219331264, + 0.006327381357550621, + 0.011934159323573112, + 0.00432614516466856, + 0.005582545418292284, + 0.005981115158647299, + 0.003068899968639016, + 0.004463830031454563, + 0.0033051634673029184, + 0.00803005788475275, + 0.011555127799510956, + 0.01301675383001566, + 0.011856453493237495, + 0.010808076709508896, + 0.008215060457587242, + 0.02000669576227665, + 0.016336696222424507, + 0.00509762903675437, + 0.006662233732640743, + 0.015864025801420212, + 0.007853723131120205, + 0.008083796128630638, + 0.009425976313650608, + 0.011213739402592182, + 0.01525664608925581, + 0.011001682840287685, + 0.017506880685687065, + 0.03021354228258133, + 0.014124065637588501, + 0.00801873579621315, + 0.0071638766676187515, + 0.013257630169391632, + 0.011364788748323917, + 0.0026439030189067125, + 0.0033219510223716497, + 0.002262015827000141, + 0.00442266371101141, + 0.004090803675353527, + 0.0018046855693683028, + 0.004894752521067858, + 0.0036171572282910347, + 0.015504607930779457, + 0.020389504730701447, + 0.018357284367084503, + 0.023537652567029, + 0.01997937448322773, + 0.011200100183486938, + 0.02274647168815136, + 0.011892060749232769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.04309023171663284 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294852942228317 + ], + "opposite_logprob_tokens": [ + -2.237948417663574 + ], + "logprob_reward": -0.11294852942228317, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294852942228317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054063573479652405, + 0.06774492561817169, + 0.3157083988189697, + 0.03396408632397652, + 0.051319610327482224, + 0.04143916070461273, + 0.08761213719844818, + 0.09360624104738235, + 0.032689668238162994, + 0.22185218334197998 + ], + "credit_map": [ + 0.004247373901307583, + 0.006486407946795225, + 0.005192762706428766, + 0.010080146603286266, + 0.005402487702667713, + 0.007051373366266489, + 0.009793996810913086, + 0.005809027235955, + 0.006831216625869274, + 0.007275063078850508, + 0.007261408492922783, + 0.006717856507748365, + 0.010762562043964863, + 0.010186594910919666, + 0.009394190274178982, + 0.009316029958426952, + 0.017536817118525505, + 0.01180938258767128, + 0.09068647772073746, + 0.024007083848118782, + 0.018217066302895546, + 0.008388163521885872, + 0.12141979485750198, + 0.0236436165869236, + 0.0026793130673468113, + 0.003916748333722353, + 0.003210462396964431, + 0.0041923196986317635, + 0.0033323983661830425, + 0.005056189838796854, + 0.007058445829898119, + 0.004518208093941212, + 0.004708883818238974, + 0.009273611940443516, + 0.005702964495867491, + 0.005933006294071674, + 0.007177406921982765, + 0.00527711259201169, + 0.007343679200857878, + 0.00590294785797596, + 0.003929264843463898, + 0.005207149311900139, + 0.006848566699773073, + 0.004480400588363409, + 0.0051580434665083885, + 0.004812953993678093, + 0.005231359042227268, + 0.005771425552666187, + 0.005866506602615118, + 0.006030707154422998, + 0.02526981197297573, + 0.009092581458389759, + 0.004485755693167448, + 0.009525963105261326, + 0.009963884018361568, + 0.01737692765891552, + 0.007507710251957178, + 0.016745716333389282, + 0.009454245679080486, + 0.005397201981395483, + 0.024499468505382538, + 0.013357799500226974, + 0.008960469625890255, + 0.00768363568931818, + 0.005447399336844683, + 0.0031091775745153427, + 0.003527160966768861, + 0.003971272148191929, + 0.004030387382954359, + 0.0023362301290035248, + 0.004905034322291613, + 0.005363007541745901, + 0.021252337843179703, + 0.026266468688845634, + 0.02653934806585312, + 0.033605918288230896, + 0.03031965345144272, + 0.021152829751372337, + 0.04054534435272217, + 0.022170284762978554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.11294852942228317 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069071367383003 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.016069071367383003, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069071367383003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31994199752807617, + 0.1008482351899147, + 0.3020589053630829, + 0.03918386995792389, + 0.051661569625139236, + 0.036565572023391724, + 0.042665883898735046, + 0.03067978471517563, + 0.025269044563174248, + 0.05112506076693535 + ], + "credit_map": [ + 0.023069702088832855, + 0.045056335628032684, + 0.024549931287765503, + 0.08045949786901474, + 0.032406143844127655, + 0.023211102932691574, + 0.05987481772899628, + 0.03131449967622757, + 0.00930566992610693, + 0.01953471451997757, + 0.00988958589732647, + 0.01229194924235344, + 0.012729262933135033, + 0.01480057556182146, + 0.012299560941755772, + 0.009996919892728329, + 0.016629112884402275, + 0.010434139519929886, + 0.08298537135124207, + 0.023628517985343933, + 0.019642552360892296, + 0.006989998277276754, + 0.11800310760736465, + 0.023746129125356674, + 0.003095217514783144, + 0.005919269751757383, + 0.002617495134472847, + 0.004831696394830942, + 0.00477594556286931, + 0.005172019824385643, + 0.00660741189494729, + 0.006164819002151489, + 0.003702369285747409, + 0.006850015372037888, + 0.005270213820040226, + 0.009227178990840912, + 0.004091312643140554, + 0.004492846317589283, + 0.008073326200246811, + 0.009954310953617096, + 0.003710194956511259, + 0.004605130758136511, + 0.004779310896992683, + 0.005646137520670891, + 0.0037221661768853664, + 0.003980980254709721, + 0.005339875351637602, + 0.0047817775048315525, + 0.003078503767028451, + 0.0036440202966332436, + 0.00741950748488307, + 0.005264485254883766, + 0.005606658756732941, + 0.008527595549821854, + 0.004588899202644825, + 0.004536215681582689, + 0.0033477391116321087, + 0.007182689383625984, + 0.003246943699195981, + 0.0023088152520358562, + 0.006244972813874483, + 0.0023016491904854774, + 0.0028832200914621353, + 0.0031637572683393955, + 0.0044164108112454414, + 0.0034347057808190584, + 0.002075087744742632, + 0.0030278859194368124, + 0.0032144091092050076, + 0.0020902762189507484, + 0.003433386329561472, + 0.0035768873058259487, + 0.005078925285488367, + 0.008283206261694431, + 0.006464342121034861, + 0.00684181647375226, + 0.0057888999581336975, + 0.005103281699120998, + 0.00624555628746748, + 0.007319034077227116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.016069071367383003 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a high-angle, overhead shot, the initial image abstract and poetic. Raindrops pummel the deep, dark surface of a lake, stirring up concentric ripples that radiate outward. The water's surface simultaneously reflects the cold, strip-shaped lights within the laboratory, creating a distorted, interwoven tapestry of cool blue light and shadow. Then, the camera slowly and steadily zooms in, penetrating this illusion of light and shadow to focus on the protagonist, slumped over on the cold terrazzo floor. He leans limply against a metal wall covered in instruments. A hand, trembling violently with weakness, laboriously, almost spasmodically, unscrews a white plastic bottle of fluoxetine bearing a prescription label. As the cap slips loose, several small white pills tumble out, impacting precisely on the toes of a pair of rain-soaked, dark canvas shoes at his feet. Not far from the protagonist, a glass beaker of chemical solution overflows uncontrollably, producing a fine, viscous white foam that shimmers eerily under the cool blue light cast by the overhead instruments. Through the huge glass window covered with water vapor behind the character, you can see that the rain outside the window is weakening. At the same time, the waveform line on the screen of an electrocardiogram monitor tends to be flat, and the sharp, continuous beeping sound it emits gradually drowns out the sound of rain, injecting the entire picture with an extremely tense atmosphere of despair, loss of control and medical emergency.", + "seed": 1211, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_14/index_6429_seed_1211_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0791904142126441, + "vlm_logprob": -0.0791904142126441, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.0791904142126441, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10992191731929779 + ], + "opposite_logprob_tokens": [ + -2.359921932220459 + ], + "logprob_reward": -0.10992191731929779, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10992191731929779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07748299837112427, + 0.10026025772094727, + 0.12299682199954987, + 0.05550370365381241, + 0.037096623331308365, + 0.08005673438310623, + 0.050773125141859055, + 0.06080438569188118, + 0.17605425417423248, + 0.2389710247516632 + ], + "credit_map": [ + 0.005976385902613401, + 0.011770022101700306, + 0.00873948261141777, + 0.013070967979729176, + 0.009696315042674541, + 0.008412614464759827, + 0.00981416366994381, + 0.010003059171140194, + 0.005166811402887106, + 0.031968288123607635, + 0.00570318941026926, + 0.00433734618127346, + 0.007998282089829445, + 0.0071396431885659695, + 0.03314606845378876, + 0.004800642374902964, + 0.015509969554841518, + 0.016124527901411057, + 0.009421234019100666, + 0.02096743695437908, + 0.016503633931279182, + 0.01877184771001339, + 0.014305445365607738, + 0.011392745189368725, + 0.0054549723863601685, + 0.006554099265486002, + 0.006910365540534258, + 0.005234161391854286, + 0.006452320609241724, + 0.0074287462048232555, + 0.010593566112220287, + 0.006875480059534311, + 0.004445885308086872, + 0.0035209115594625473, + 0.004430269356817007, + 0.00778216402977705, + 0.0036397469229996204, + 0.0039032623171806335, + 0.006303956266492605, + 0.0030704298987984657, + 0.008803888224065304, + 0.009236276149749756, + 0.006402421742677689, + 0.008039247244596481, + 0.009085072204470634, + 0.011121153831481934, + 0.0185848381370306, + 0.008783848024904728, + 0.006467614322900772, + 0.0053742979653179646, + 0.00586405536159873, + 0.005926609039306641, + 0.006682282313704491, + 0.004560557659715414, + 0.010651890188455582, + 0.005245824344456196, + 0.008730057626962662, + 0.00827635359019041, + 0.008363073691725731, + 0.005717045161873102, + 0.010340122506022453, + 0.007473979145288467, + 0.006939128506928682, + 0.004964632913470268, + 0.03412249684333801, + 0.022631505504250526, + 0.01958833634853363, + 0.014566520228981972, + 0.02123170904815197, + 0.019468966871500015, + 0.018723517656326294, + 0.02572123520076275, + 0.018028629943728447, + 0.045798636972904205, + 0.016658341512084007, + 0.015491342172026634, + 0.0857182964682579, + 0.021177813410758972, + 0.014566833153367043, + 0.021531157195568085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.10992191731929779 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006660848855972 + ], + "opposite_logprob_tokens": [ + -2.6950666904449463 + ], + "logprob_reward": -0.07006660848855972, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006660848855972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05935007333755493, + 0.11038955301046371, + 0.04959176480770111, + 0.03962411731481552, + 0.08121250569820404, + 0.159123957157135, + 0.15295684337615967, + 0.12908539175987244, + 0.11826367676258087, + 0.1004020944237709 + ], + "credit_map": [ + 0.004119676072150469, + 0.009552164003252983, + 0.003467144211754203, + 0.008049738593399525, + 0.00841966737061739, + 0.015060781501233578, + 0.006109017878770828, + 0.004571882542222738, + 0.007084050681442022, + 0.032724615186452866, + 0.006108155474066734, + 0.004791436716914177, + 0.007515397854149342, + 0.005504016298800707, + 0.03808698058128357, + 0.008574897423386574, + 0.005321556702256203, + 0.005951444152742624, + 0.0029229391366243362, + 0.0084456168115139, + 0.006960258353501558, + 0.007486474234610796, + 0.006725670304149389, + 0.0057778023183345795, + 0.002822519512847066, + 0.005096301902085543, + 0.002844909904524684, + 0.005039765499532223, + 0.006022651679813862, + 0.004481703042984009, + 0.0076902746222913265, + 0.005625992082059383, + 0.011049261316657066, + 0.008597847074270248, + 0.010534798726439476, + 0.015122083947062492, + 0.00861333217471838, + 0.009047581814229488, + 0.010715884156525135, + 0.007531726732850075, + 0.0191873237490654, + 0.015916690230369568, + 0.012853223830461502, + 0.011057778261601925, + 0.027406292036175728, + 0.017594926059246063, + 0.03806643560528755, + 0.017041292041540146, + 0.022352030500769615, + 0.0247142743319273, + 0.022006569430232048, + 0.028598854318261147, + 0.020052408799529076, + 0.010702250525355339, + 0.011168147437274456, + 0.013362312689423561, + 0.019855307415127754, + 0.016307847574353218, + 0.02228676527738571, + 0.015819968655705452, + 0.015376051887869835, + 0.012773443013429642, + 0.012670179829001427, + 0.013995830900967121, + 0.02361610345542431, + 0.015407806262373924, + 0.012240365147590637, + 0.012319928035140038, + 0.01239168643951416, + 0.011810226365923882, + 0.013202118687331676, + 0.017275439575314522, + 0.013536151498556137, + 0.017606645822525024, + 0.009941000491380692, + 0.01799989491701126, + 0.014238244853913784, + 0.009769473224878311, + 0.010454065166413784, + 0.006856616120785475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.07006660848855972 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547007761895657 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012547007761895657, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547007761895657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11802016943693161, + 0.037082377821207047, + 0.05729449540376663, + 0.02726322039961815, + 0.04633548855781555, + 0.2701139450073242, + 0.1447928249835968, + 0.17092067003250122, + 0.06521587818861008, + 0.06296093016862869 + ], + "credit_map": [ + 0.008591595105826855, + 0.01794486865401268, + 0.010192930698394775, + 0.02408195100724697, + 0.013060306198894978, + 0.01964479684829712, + 0.015396106988191605, + 0.009107617661356926, + 0.004657617304474115, + 0.0044746012426912785, + 0.0040141744539141655, + 0.0038650764618068933, + 0.004146317485719919, + 0.00669797183945775, + 0.0058090039528906345, + 0.0034176132176071405, + 0.008031425066292286, + 0.008173884823918343, + 0.00371363153681159, + 0.0076403203420341015, + 0.00990455411374569, + 0.007355412468314171, + 0.00841034110635519, + 0.004064925946295261, + 0.002426639199256897, + 0.0035467296838760376, + 0.0031232801266014576, + 0.003223212668672204, + 0.0032122263219207525, + 0.0021396807860583067, + 0.004749639425426722, + 0.004841810092329979, + 0.004485472105443478, + 0.0034406702034175396, + 0.005005722399801016, + 0.007097812369465828, + 0.006254232954233885, + 0.007112011779099703, + 0.008327209390699863, + 0.004612356889992952, + 0.014159963466227055, + 0.01974768377840519, + 0.015279307961463928, + 0.01581467129290104, + 0.07033434510231018, + 0.042931701987981796, + 0.045051369816064835, + 0.04679488018155098, + 0.015276257880032063, + 0.017008762806653976, + 0.01961642876267433, + 0.033362992107868195, + 0.020717736333608627, + 0.013144431635737419, + 0.013812201097607613, + 0.011854013428092003, + 0.021394265815615654, + 0.021394966170191765, + 0.022656647488474846, + 0.029856210574507713, + 0.025201858952641487, + 0.013015727512538433, + 0.019963322207331657, + 0.01743766851723194, + 0.008103874512016773, + 0.009396911598742008, + 0.007055323570966721, + 0.010521420277655125, + 0.008459566161036491, + 0.010008295066654682, + 0.006052529439330101, + 0.005617956630885601, + 0.004923750646412373, + 0.010774064809083939, + 0.004915266297757626, + 0.005325995851308107, + 0.017770685255527496, + 0.0055003659799695015, + 0.00863813515752554, + 0.005112662445753813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.012547007761895657 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035348176956177 + ], + "opposite_logprob_tokens": [ + -1.910353422164917 + ], + "logprob_reward": -0.16035348176956177, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035348176956177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08572406321763992, + 0.034023579210042953, + 0.03226694464683533, + 0.055194176733493805, + 0.06577768921852112, + 0.23188193142414093, + 0.18066518008708954, + 0.12778224050998688, + 0.11926786601543427, + 0.06741636991500854 + ], + "credit_map": [ + 0.008118866942822933, + 0.01566186733543873, + 0.009788598865270615, + 0.01281021535396576, + 0.01031477376818657, + 0.01158630009740591, + 0.009515800513327122, + 0.007927626371383667, + 0.0024802284315228462, + 0.008194925263524055, + 0.002925181295722723, + 0.002336408942937851, + 0.0029065643902868032, + 0.00500998180359602, + 0.006996789947152138, + 0.0031734968069940805, + 0.003609653329476714, + 0.004551220685243607, + 0.002172330627217889, + 0.004465727135539055, + 0.007479984313249588, + 0.0028970239218324423, + 0.004584718029946089, + 0.0025062840431928635, + 0.00456780893728137, + 0.007001532707363367, + 0.00554237887263298, + 0.0062430743128061295, + 0.006282048765569925, + 0.004562267102301121, + 0.013112341985106468, + 0.007882720790803432, + 0.010224850848317146, + 0.0065598380751907825, + 0.008536310866475105, + 0.012056206353008747, + 0.007215169724076986, + 0.00687250355258584, + 0.008960478007793427, + 0.005352329462766647, + 0.01164776086807251, + 0.01706814207136631, + 0.0164212416857481, + 0.016682801768183708, + 0.06672225892543793, + 0.03335588797926903, + 0.029379095882177353, + 0.04060472175478935, + 0.01977331005036831, + 0.021246174350380898, + 0.016676371917128563, + 0.029568690806627274, + 0.029629049822688103, + 0.01671728678047657, + 0.02707025222480297, + 0.019984019920229912, + 0.011726846918463707, + 0.012684586457908154, + 0.016131894662976265, + 0.012360078282654285, + 0.021242203190922737, + 0.013114155270159245, + 0.024336183443665504, + 0.01618628017604351, + 0.019617026671767235, + 0.014517098665237427, + 0.011942852288484573, + 0.010446532629430294, + 0.017829515039920807, + 0.014272998087108135, + 0.013449902646243572, + 0.01719193533062935, + 0.004144975915551186, + 0.011685812845826149, + 0.004363008309155703, + 0.002757198642939329, + 0.023071464151144028, + 0.005268399138003588, + 0.008779867552220821, + 0.007345634046941996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.16035348176956177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306305572390556 + ], + "opposite_logprob_tokens": [ + -3.168063163757324 + ], + "logprob_reward": -0.04306305572390556, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306305572390556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09464427828788757, + 0.09632223099470139, + 0.06624776870012283, + 0.083367258310318, + 0.11715930700302124, + 0.22861270606517792, + 0.08584218472242355, + 0.07659290730953217, + 0.07315131276845932, + 0.07806000858545303 + ], + "credit_map": [ + 0.007495519705116749, + 0.01212223805487156, + 0.012081921100616455, + 0.011433121748268604, + 0.015722285956144333, + 0.010259406641125679, + 0.01359062921255827, + 0.0119391568005085, + 0.005006816238164902, + 0.023750977590680122, + 0.008261949755251408, + 0.004910456947982311, + 0.007155056577175856, + 0.011027054861187935, + 0.026228301227092743, + 0.009981619194149971, + 0.007933088578283787, + 0.008418834768235683, + 0.0041047073900699615, + 0.013205713592469692, + 0.0068297660909593105, + 0.009908425621688366, + 0.009307929314672947, + 0.006539308000355959, + 0.007839926518499851, + 0.011757601983845234, + 0.007826480083167553, + 0.010229208506643772, + 0.00929900724440813, + 0.007405274081975222, + 0.014323997311294079, + 0.01468577142804861, + 0.010219183750450611, + 0.011774676851928234, + 0.011969560757279396, + 0.014353941194713116, + 0.017433887347579002, + 0.021798081696033478, + 0.019984988495707512, + 0.009624994359910488, + 0.019576165825128555, + 0.019615791738033295, + 0.019363312050700188, + 0.015482419170439243, + 0.06413128972053528, + 0.028396805748343468, + 0.02826371043920517, + 0.03378321975469589, + 0.006229248363524675, + 0.0061155506409704685, + 0.009794352576136589, + 0.015191462822258472, + 0.016592735424637794, + 0.009927517734467983, + 0.013266668654978275, + 0.008724657818675041, + 0.010106263682246208, + 0.01028447225689888, + 0.006182826589792967, + 0.005985391791909933, + 0.018598249182105064, + 0.009989484213292599, + 0.008752998895943165, + 0.006693223956972361, + 0.007869071327149868, + 0.008222769945859909, + 0.006288450211286545, + 0.009332646615803242, + 0.01529999915510416, + 0.008185138925909996, + 0.007868291810154915, + 0.010084951296448708, + 0.0061283716931939125, + 0.020647164434194565, + 0.005430812481790781, + 0.0032487036660313606, + 0.020999429747462273, + 0.007706304080784321, + 0.008930768817663193, + 0.004968455992639065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.04306305572390556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a high-angle, overhead shot, the initial image abstract and poetic. Raindrops pummel the deep, dark surface of a lake, stirring up concentric ripples that radiate outward. The water's surface simultaneously reflects the cold, strip-shaped lights within the laboratory, creating a distorted, interwoven tapestry of cool blue light and shadow. Then, the camera slowly and steadily zooms in, penetrating this illusion of light and shadow to focus on the protagonist, slumped over on the cold terrazzo floor. He leans limply against a metal wall covered in instruments. A hand, trembling violently with weakness, laboriously, almost spasmodically, unscrews a white plastic bottle of fluoxetine bearing a prescription label. As the cap slips loose, several small white pills tumble out, impacting precisely on the toes of a pair of rain-soaked, dark canvas shoes at his feet. Not far from the protagonist, a glass beaker of chemical solution overflows uncontrollably, producing a fine, viscous white foam that shimmers eerily under the cool blue light cast by the overhead instruments. Through the huge glass window covered with water vapor behind the character, you can see that the rain outside the window is weakening. At the same time, the waveform line on the screen of an electrocardiogram monitor tends to be flat, and the sharp, continuous beeping sound it emits gradually drowns out the sound of rain, injecting the entire picture with an extremely tense atmosphere of despair, loss of control and medical emergency.", + "seed": 1212, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_14/index_6429_seed_1212_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1239018727093935, + "vlm_logprob": -0.1239018727093935, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.1239018727093935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16340793669223785 + ], + "opposite_logprob_tokens": [ + -1.9134079217910767 + ], + "logprob_reward": -0.16340793669223785, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16340793669223785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17608284950256348, + 0.15040749311447144, + 0.11907374113798141, + 0.1193694919347763, + 0.09753243625164032, + 0.0474071130156517, + 0.09344138205051422, + 0.12251868844032288, + 0.02204597368836403, + 0.05212080106139183 + ], + "credit_map": [ + 0.023033389821648598, + 0.03557128831744194, + 0.013140852563083172, + 0.014599645510315895, + 0.026013387367129326, + 0.021244509145617485, + 0.015150262974202633, + 0.02732951194047928, + 0.013155942782759666, + 0.013271039351820946, + 0.018583964556455612, + 0.023490875959396362, + 0.02797188051044941, + 0.02291301265358925, + 0.016340577974915504, + 0.01468020211905241, + 0.01128330547362566, + 0.01822582818567753, + 0.008762864395976067, + 0.016204047948122025, + 0.017933892086148262, + 0.021232526749372482, + 0.010363486595451832, + 0.01506778970360756, + 0.01020190492272377, + 0.026622526347637177, + 0.012594291940331459, + 0.013355984352529049, + 0.018256746232509613, + 0.020450692623853683, + 0.008186910301446915, + 0.009700440801680088, + 0.014706107787787914, + 0.011500151827931404, + 0.015547341667115688, + 0.011522162705659866, + 0.010306659154593945, + 0.007917630486190319, + 0.011410084553062916, + 0.014622303657233715, + 0.005639181472361088, + 0.005123932845890522, + 0.00809229351580143, + 0.004564890172332525, + 0.005638432689011097, + 0.004834875464439392, + 0.005317717790603638, + 0.008195786736905575, + 0.012845233082771301, + 0.01730934903025627, + 0.016080379486083984, + 0.007939900271594524, + 0.010902594774961472, + 0.007429858669638634, + 0.01000004168599844, + 0.010934025049209595, + 0.022724078968167305, + 0.015203549526631832, + 0.012941932305693626, + 0.012375034391880035, + 0.017763780429959297, + 0.023333298042416573, + 0.009917394258081913, + 0.008259614929556847, + 0.002299197483807802, + 0.0023006706032902002, + 0.0021039419807493687, + 0.002002862747758627, + 0.004539490211755037, + 0.0031247471924871206, + 0.002520298818126321, + 0.00315476325340569, + 0.012694504112005234, + 0.007441044319421053, + 0.006834013387560844, + 0.006522144190967083, + 0.006496570073068142, + 0.004203696735203266, + 0.005141423083841801, + 0.002787406789138913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.16340793669223785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036865621805191 + ], + "opposite_logprob_tokens": [ + -2.3503687381744385 + ], + "logprob_reward": -0.10036865621805191, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036865621805191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13171733915805817, + 0.0720008909702301, + 0.11594259738922119, + 0.12116511911153793, + 0.1343831866979599, + 0.05137196183204651, + 0.15977680683135986, + 0.07769827544689178, + 0.021303270012140274, + 0.11464060842990875 + ], + "credit_map": [ + 0.017671355977654457, + 0.032784346491098404, + 0.011498513631522655, + 0.010544474236667156, + 0.015270350500941277, + 0.013890469446778297, + 0.01065063290297985, + 0.019407188519835472, + 0.007266013417392969, + 0.008192685432732105, + 0.010780241340398788, + 0.009548596106469631, + 0.008820012211799622, + 0.006800665985792875, + 0.011766808107495308, + 0.008825867436826229, + 0.015800563618540764, + 0.022122550755739212, + 0.01185369212180376, + 0.01335939858108759, + 0.017816882580518723, + 0.013304089196026325, + 0.010278338566422462, + 0.011407086625695229, + 0.01196710579097271, + 0.01774374209344387, + 0.01749047264456749, + 0.013375531882047653, + 0.0159504022449255, + 0.01832662709057331, + 0.013515510596334934, + 0.01279572956264019, + 0.009511888958513737, + 0.02140948176383972, + 0.012666006572544575, + 0.018944377079606056, + 0.016633452847599983, + 0.016406279057264328, + 0.01782197132706642, + 0.020989717915654182, + 0.004343416076153517, + 0.004801550414413214, + 0.006042228080332279, + 0.004281728994101286, + 0.007529756519943476, + 0.006183751858770847, + 0.008384309709072113, + 0.009805221110582352, + 0.02781393751502037, + 0.025085613131523132, + 0.017670758068561554, + 0.015423406846821308, + 0.02095191553235054, + 0.014926852658390999, + 0.02198200672864914, + 0.01592230424284935, + 0.014984059147536755, + 0.011202559806406498, + 0.01224331185221672, + 0.008955216966569424, + 0.008955484256148338, + 0.009141855873167515, + 0.006189678329974413, + 0.006026105489581823, + 0.0017768533434718847, + 0.0022761391010135412, + 0.002522852970287204, + 0.002343836473301053, + 0.003979396540671587, + 0.0033174699638038874, + 0.002069745445623994, + 0.0030169759411364794, + 0.026704300194978714, + 0.015391449443995953, + 0.01567244902253151, + 0.016894511878490448, + 0.012165459804236889, + 0.008715188130736351, + 0.01086348295211792, + 0.008233770728111267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.10036865621805191 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003684401512146 + ], + "opposite_logprob_tokens": [ + -2.3503684997558594 + ], + "logprob_reward": -0.1003684401512146, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003684401512146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1921626329421997, + 0.17791135609149933, + 0.1375540941953659, + 0.10513222962617874, + 0.1152958944439888, + 0.04841840639710426, + 0.0941598191857338, + 0.03166945278644562, + 0.03777375817298889, + 0.05992233008146286 + ], + "credit_map": [ + 0.02415062114596367, + 0.04042147845029831, + 0.018176095560193062, + 0.020606162026524544, + 0.023936398327350616, + 0.021014556288719177, + 0.019949989393353462, + 0.023907339200377464, + 0.01487402431666851, + 0.016711128875613213, + 0.016829615458846092, + 0.024403933435678482, + 0.026898037642240524, + 0.021547937765717506, + 0.0359797440469265, + 0.0206669382750988, + 0.014003635384142399, + 0.02756798267364502, + 0.012329726479947567, + 0.013862665742635727, + 0.012912258505821228, + 0.020502213388681412, + 0.012247204780578613, + 0.02412840537726879, + 0.00818967167288065, + 0.020127298310399055, + 0.011084177531301975, + 0.008586063981056213, + 0.015227348543703556, + 0.015682535246014595, + 0.010757869109511375, + 0.01547726709395647, + 0.010921256616711617, + 0.019565241411328316, + 0.010761828161776066, + 0.017539100721478462, + 0.012245364487171173, + 0.011497880332171917, + 0.014712660573422909, + 0.01805255562067032, + 0.005207529291510582, + 0.004503930453211069, + 0.006488040089607239, + 0.005337749607861042, + 0.007389442529529333, + 0.005449223332107067, + 0.0063254814594984055, + 0.007717011496424675, + 0.012660520151257515, + 0.01387298945337534, + 0.009127588011324406, + 0.012203143909573555, + 0.0157699566334486, + 0.007128389086574316, + 0.01326949056237936, + 0.010127737186849117, + 0.0036616961006075144, + 0.004563054069876671, + 0.00427272729575634, + 0.0033105798065662384, + 0.006150057539343834, + 0.0042751808650791645, + 0.002764613600447774, + 0.002671543974429369, + 0.006267523858696222, + 0.004761835094541311, + 0.005042911507189274, + 0.005388313438743353, + 0.005415185820311308, + 0.004132041707634926, + 0.0033419618848711252, + 0.0034239834640175104, + 0.008401104249060154, + 0.011771490797400475, + 0.008023575879633427, + 0.007978206500411034, + 0.006229114718735218, + 0.005214216187596321, + 0.006354167126119137, + 0.005950460210442543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.1003684401512146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255648523569107 + ], + "opposite_logprob_tokens": [ + -1.6005648374557495 + ], + "logprob_reward": -0.2255648523569107, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255648523569107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12500223517417908, + 0.15797503292560577, + 0.1943025141954422, + 0.16910989582538605, + 0.07572617381811142, + 0.030969897285103798, + 0.09223952889442444, + 0.04274209588766098, + 0.054457876831293106, + 0.0574747659265995 + ], + "credit_map": [ + 0.011063184589147568, + 0.007847352884709835, + 0.012999101541936398, + 0.011442643590271473, + 0.02167232520878315, + 0.015173261053860188, + 0.02160041406750679, + 0.023203961551189423, + 0.014003120362758636, + 0.01447075605392456, + 0.02006058767437935, + 0.027772871777415276, + 0.025195609778165817, + 0.019669800996780396, + 0.02267296612262726, + 0.014129336923360825, + 0.014833992347121239, + 0.016310153529047966, + 0.01705688051879406, + 0.017443152144551277, + 0.020222926512360573, + 0.021285206079483032, + 0.02439064532518387, + 0.06275954097509384, + 0.009532088413834572, + 0.018969135358929634, + 0.01973089575767517, + 0.016677388921380043, + 0.033015817403793335, + 0.03818453103303909, + 0.015786249190568924, + 0.017213810235261917, + 0.007457703817635775, + 0.007824581116437912, + 0.00641255546361208, + 0.007696592248976231, + 0.01094523910433054, + 0.012745855376124382, + 0.00881942268460989, + 0.013824221678078175, + 0.002558805514127016, + 0.0037122175563126802, + 0.0031665111891925335, + 0.003409756813198328, + 0.003557865973562002, + 0.0036079634446650743, + 0.006020784378051758, + 0.004935992415994406, + 0.012234105728566647, + 0.015507110394537449, + 0.011196274310350418, + 0.008898616768419743, + 0.008490209467709064, + 0.006861851084977388, + 0.01060878299176693, + 0.018442576751112938, + 0.004819955211132765, + 0.005100371781736612, + 0.006079569924622774, + 0.004362969193607569, + 0.006906399503350258, + 0.0072335973381996155, + 0.0043036965653300285, + 0.003935531713068485, + 0.005379912443459034, + 0.006364450789988041, + 0.007748567033559084, + 0.006038296036422253, + 0.009447673335671425, + 0.007059574127197266, + 0.006339086685329676, + 0.006080315914005041, + 0.013770118355751038, + 0.008293138816952705, + 0.008648410439491272, + 0.008084304630756378, + 0.0038656149990856647, + 0.004057527519762516, + 0.006678405683487654, + 0.004077242221683264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.2255648523569107 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799478128552437 + ], + "opposite_logprob_tokens": [ + -3.529799461364746 + ], + "logprob_reward": -0.029799478128552437, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799478128552437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15547697246074677, + 0.08980735391378403, + 0.09937284886837006, + 0.07796795666217804, + 0.1361677348613739, + 0.07783789932727814, + 0.14496523141860962, + 0.06092856079339981, + 0.10209827125072479, + 0.05537723749876022 + ], + "credit_map": [ + 0.019205011427402496, + 0.025474321097135544, + 0.012133019976317883, + 0.016012893989682198, + 0.024565398693084717, + 0.03412344679236412, + 0.011242125183343887, + 0.012720765545964241, + 0.007421964313834906, + 0.007299269549548626, + 0.008298782631754875, + 0.014951200224459171, + 0.01226765289902687, + 0.012892072089016438, + 0.01463301107287407, + 0.012043405324220657, + 0.009230238385498524, + 0.013243859633803368, + 0.009021216072142124, + 0.01027168519794941, + 0.011088607832789421, + 0.013407247141003609, + 0.010257978923618793, + 0.022852014750242233, + 0.006231565959751606, + 0.015031885355710983, + 0.006824771873652935, + 0.007141029927879572, + 0.014812620356678963, + 0.014156504534184933, + 0.007057332433760166, + 0.006712247151881456, + 0.012885397300124168, + 0.023986630141735077, + 0.012013762257993221, + 0.02597215585410595, + 0.01628323644399643, + 0.013560614548623562, + 0.01591361314058304, + 0.015552323311567307, + 0.004224770236760378, + 0.006841602735221386, + 0.006513462867587805, + 0.0051170457154512405, + 0.009792670607566833, + 0.0083664171397686, + 0.01806158572435379, + 0.01892034150660038, + 0.02686627395451069, + 0.02785169892013073, + 0.015966519713401794, + 0.012968925759196281, + 0.016736149787902832, + 0.01323549821972847, + 0.018015075474977493, + 0.01332508772611618, + 0.006782508920878172, + 0.007972458377480507, + 0.009151243604719639, + 0.007649053353816271, + 0.007405874785035849, + 0.01030650269240141, + 0.006286181043833494, + 0.00537473987787962, + 0.013128434307873249, + 0.01579766348004341, + 0.016050288453698158, + 0.013423974625766277, + 0.01361941546201706, + 0.01339807827025652, + 0.008561855182051659, + 0.00811855960637331, + 0.013155213557183743, + 0.007051313295960426, + 0.008106876164674759, + 0.0082970280200243, + 0.0049349041655659676, + 0.003775600576773286, + 0.0040284995920956135, + 0.006027805153280497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.029799478128552437 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a high-angle, overhead shot, the initial image abstract and poetic. Raindrops pummel the deep, dark surface of a lake, stirring up concentric ripples that radiate outward. The water's surface simultaneously reflects the cold, strip-shaped lights within the laboratory, creating a distorted, interwoven tapestry of cool blue light and shadow. Then, the camera slowly and steadily zooms in, penetrating this illusion of light and shadow to focus on the protagonist, slumped over on the cold terrazzo floor. He leans limply against a metal wall covered in instruments. A hand, trembling violently with weakness, laboriously, almost spasmodically, unscrews a white plastic bottle of fluoxetine bearing a prescription label. As the cap slips loose, several small white pills tumble out, impacting precisely on the toes of a pair of rain-soaked, dark canvas shoes at his feet. Not far from the protagonist, a glass beaker of chemical solution overflows uncontrollably, producing a fine, viscous white foam that shimmers eerily under the cool blue light cast by the overhead instruments. Through the huge glass window covered with water vapor behind the character, you can see that the rain outside the window is weakening. At the same time, the waveform line on the screen of an electrocardiogram monitor tends to be flat, and the sharp, continuous beeping sound it emits gradually drowns out the sound of rain, injecting the entire picture with an extremely tense atmosphere of despair, loss of control and medical emergency.", + "seed": 1213, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_14/index_6429_seed_1213_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16821762323379516, + "vlm_logprob": -0.16821762323379516, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.16821762323379516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4793355464935303 + ], + "opposite_logprob_tokens": [ + -0.9793355464935303 + ], + "logprob_reward": -0.4793355464935303, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4793355464935303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5798096060752869, + 0.08492948859930038, + 0.08631889522075653, + 0.028611602261662483, + 0.06735850125551224, + 0.0456806905567646, + 0.04822838306427002, + 0.02490224316716194, + 0.019632428884506226, + 0.014528169296681881 + ], + "credit_map": [ + 0.027212481945753098, + 0.022005382925271988, + 0.030600136145949364, + 0.04244633764028549, + 0.026743941009044647, + 0.020389670506119728, + 0.21840427815914154, + 0.1920074224472046, + 0.010516948997974396, + 0.005409318953752518, + 0.008354040794074535, + 0.017346996814012527, + 0.004955891519784927, + 0.004576433449983597, + 0.014931969344615936, + 0.01883789710700512, + 0.004182781092822552, + 0.012315968982875347, + 0.0050748540088534355, + 0.0193143542855978, + 0.006403767038136721, + 0.009949753992259502, + 0.012247753329575062, + 0.01682967320084572, + 0.0026378659531474113, + 0.0017346780514344573, + 0.0035594571381807327, + 0.004037268925458193, + 0.004944700747728348, + 0.003997837193310261, + 0.0035581993870437145, + 0.004141598008573055, + 0.007376724388450384, + 0.0041884020902216434, + 0.006487849168479443, + 0.012478355318307877, + 0.009608742780983448, + 0.00764452526345849, + 0.012107029557228088, + 0.00746688200160861, + 0.005498823244124651, + 0.002690283115953207, + 0.004388824570924044, + 0.005414681043475866, + 0.007496420294046402, + 0.005403523333370686, + 0.008776139467954636, + 0.00601199921220541, + 0.008435001596808434, + 0.002236135071143508, + 0.005031579174101353, + 0.0030686005484312773, + 0.004611274227499962, + 0.009049173444509506, + 0.007550259120762348, + 0.008246366865932941, + 0.002278280444443226, + 0.0021001501008868217, + 0.00221822876483202, + 0.004721850156784058, + 0.0026626281905919313, + 0.0026594498194754124, + 0.004234649240970612, + 0.004027009941637516, + 0.0020469774026423693, + 0.0017095195362344384, + 0.0018314343178644776, + 0.0037762855645269156, + 0.0015801353147253394, + 0.002539664041250944, + 0.0029948463197797537, + 0.003153566736727953, + 0.0014640818117186427, + 0.0010101358639076352, + 0.0018774080090224743, + 0.0016620492096990347, + 0.002061073435470462, + 0.0011093171779066324, + 0.0024649628903716803, + 0.0028791422955691814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.4793355464935303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055005524307489395 + ], + "opposite_logprob_tokens": [ + -2.9300055503845215 + ], + "logprob_reward": -0.055005524307489395, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055005524307489395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12339016795158386, + 0.06629293411970139, + 0.06352248042821884, + 0.09612976014614105, + 0.10552319139242172, + 0.05901151895523071, + 0.09139268100261688, + 0.23569262027740479, + 0.05153213068842888, + 0.10751251131296158 + ], + "credit_map": [ + 0.009547710418701172, + 0.010757805779576302, + 0.0135489571839571, + 0.01526437234133482, + 0.0119086392223835, + 0.009223688393831253, + 0.03191822022199631, + 0.021220767870545387, + 0.006287867669016123, + 0.004902779124677181, + 0.005245853681117296, + 0.009158754721283913, + 0.007754049729555845, + 0.00585448183119297, + 0.015302355401217937, + 0.011786790564656258, + 0.0038983484264463186, + 0.013817437924444675, + 0.0056466045789420605, + 0.005100927781313658, + 0.008440465666353703, + 0.006912466138601303, + 0.0058667357079684734, + 0.01383949164301157, + 0.0059384130872786045, + 0.006898285821080208, + 0.007460277993232012, + 0.01270239520817995, + 0.01860082894563675, + 0.016894416883587837, + 0.010789409279823303, + 0.01684572547674179, + 0.009447653777897358, + 0.00903009157627821, + 0.010074532590806484, + 0.018687525764107704, + 0.018875446170568466, + 0.013921204954385757, + 0.013918369077146053, + 0.011568346992135048, + 0.006942137144505978, + 0.003994500730186701, + 0.005712054669857025, + 0.008327300660312176, + 0.01146872341632843, + 0.006637896411120892, + 0.008959921076893806, + 0.006968978326767683, + 0.009027550928294659, + 0.0049597774632275105, + 0.005459189880639315, + 0.01197657361626625, + 0.013791353441774845, + 0.014034565538167953, + 0.01782512664794922, + 0.014318536967039108, + 0.01984292082488537, + 0.020430410280823708, + 0.02088116481900215, + 0.045816924422979355, + 0.026767082512378693, + 0.02211901918053627, + 0.042587533593177795, + 0.03724753111600876, + 0.006928217597305775, + 0.00504333944991231, + 0.006159602664411068, + 0.0049454215914011, + 0.006254358682781458, + 0.005187917035073042, + 0.009022647514939308, + 0.007990621030330658, + 0.008331299759447575, + 0.007049465086311102, + 0.008686364628374577, + 0.012307137250900269, + 0.018968407064676285, + 0.009306895546615124, + 0.02062375657260418, + 0.02223918028175831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.055005524307489395 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430806428194046 + ], + "opposite_logprob_tokens": [ + -3.1680805683135986 + ], + "logprob_reward": -0.0430806428194046, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430806428194046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24990352988243103, + 0.03313470259308815, + 0.054556284099817276, + 0.046058107167482376, + 0.08452747762203217, + 0.056681446731090546, + 0.10115555673837662, + 0.19859646260738373, + 0.09560578316450119, + 0.07978061586618423 + ], + "credit_map": [ + 0.010352831333875656, + 0.011817735619843006, + 0.014388345181941986, + 0.023272031918168068, + 0.016039421781897545, + 0.016293900087475777, + 0.08753536641597748, + 0.07020390778779984, + 0.003636881709098816, + 0.002132305409759283, + 0.003257080679759383, + 0.007877985946834087, + 0.002753159496933222, + 0.0025006523355841637, + 0.005754230543971062, + 0.005222405772656202, + 0.0037012011744081974, + 0.009601378813385963, + 0.004522449802607298, + 0.007956180721521378, + 0.005728892982006073, + 0.006438458804041147, + 0.006912033539265394, + 0.0096956891939044, + 0.0029393399599939585, + 0.0022033366840332747, + 0.003319067880511284, + 0.009450727142393589, + 0.006142341066151857, + 0.007204915396869183, + 0.006048274226486683, + 0.008750105276703835, + 0.011844037100672722, + 0.005987714044749737, + 0.00852548610419035, + 0.014430639334022999, + 0.011659865267574787, + 0.00880047120153904, + 0.014463373459875584, + 0.008815892972052097, + 0.0077232182957232, + 0.0042112767696380615, + 0.005670982412993908, + 0.009781367145478725, + 0.006871698424220085, + 0.006130707450211048, + 0.008985486812889576, + 0.007306704763323069, + 0.009957943111658096, + 0.006752471439540386, + 0.006318559870123863, + 0.014450039714574814, + 0.0170240830630064, + 0.01682923175394535, + 0.014684726484119892, + 0.015138503164052963, + 0.016960730776190758, + 0.017375215888023376, + 0.014994773082435131, + 0.03946840763092041, + 0.019690699875354767, + 0.017700346186757088, + 0.041797202080488205, + 0.03060908429324627, + 0.007094532251358032, + 0.005108993034809828, + 0.006109095178544521, + 0.008942607790231705, + 0.009616907685995102, + 0.01600068435072899, + 0.018648386001586914, + 0.024084584787487984, + 0.005270472727715969, + 0.0055750710889697075, + 0.004581247456371784, + 0.011307897046208382, + 0.00691098440438509, + 0.007755839265882969, + 0.01751863583922386, + 0.020860467106103897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0430806428194046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255626767873764 + ], + "opposite_logprob_tokens": [ + -1.6005626916885376 + ], + "logprob_reward": -0.2255626767873764, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255626767873764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2104506492614746, + 0.025120330974459648, + 0.0699169859290123, + 0.06483184546232224, + 0.18302176892757416, + 0.17827896773815155, + 0.09276702255010605, + 0.08466126024723053, + 0.04549067094922066, + 0.04546042159199715 + ], + "credit_map": [ + 0.015224187634885311, + 0.014131540432572365, + 0.020403197035193443, + 0.031059786677360535, + 0.017408447340130806, + 0.024974258616566658, + 0.054922353476285934, + 0.032326892018318176, + 0.0037690307945013046, + 0.0017777517205104232, + 0.0029389476403594017, + 0.0032916902564466, + 0.003647784236818552, + 0.0017528265016153455, + 0.004370654467493296, + 0.0035716465208679438, + 0.0038276533596217632, + 0.012291776947677135, + 0.004697661381214857, + 0.012752072885632515, + 0.0072619738057255745, + 0.0073921773582696915, + 0.007350134197622538, + 0.014343530870974064, + 0.003340934170410037, + 0.003839403623715043, + 0.004487250931560993, + 0.008575480431318283, + 0.009395475499331951, + 0.011954586021602154, + 0.008612329140305519, + 0.014626379124820232, + 0.013046512380242348, + 0.012558027170598507, + 0.011458085849881172, + 0.020442258566617966, + 0.04107357934117317, + 0.03007066436111927, + 0.030357368290424347, + 0.02401527389883995, + 0.015252713114023209, + 0.010683349333703518, + 0.012494049966335297, + 0.014696368016302586, + 0.04801773279905319, + 0.01953909359872341, + 0.03355501592159271, + 0.024040641263127327, + 0.014643991366028786, + 0.004280870780348778, + 0.0086738346144557, + 0.006148237735033035, + 0.015296876430511475, + 0.016699817031621933, + 0.011377094313502312, + 0.015646299347281456, + 0.00613793358206749, + 0.005942734889686108, + 0.006742238067090511, + 0.012182357721030712, + 0.01356265414506197, + 0.008528774604201317, + 0.016337981447577477, + 0.015226582065224648, + 0.0037326584570109844, + 0.0025822033639997244, + 0.0026596372481435537, + 0.004403117578476667, + 0.00912377517670393, + 0.004521271213889122, + 0.009741446934640408, + 0.008726559579372406, + 0.006859867833554745, + 0.003668074728921056, + 0.005115944426506758, + 0.0048044221475720406, + 0.004912990611046553, + 0.004480919800698757, + 0.007319645490497351, + 0.008298560045659542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.2255626767873764 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle, overhead shot of a dark lake surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the protagonist visible slumped over on a cold terrazzo floor leaning against a metal wall covered in instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a hand shown unscrewing a white plastic bottle of fluoxetine with small white pills tumbling out onto dark canvas shoes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a glass beaker of chemical solution visible overflowing with white foam near the protagonist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from the initial overhead shot to focus on the protagonist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038103725761175156 + ], + "opposite_logprob_tokens": [ + -3.2881038188934326 + ], + "logprob_reward": -0.038103725761175156, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038103725761175156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19443635642528534, + 0.0964481309056282, + 0.17069527506828308, + 0.09110628068447113, + 0.06523752212524414, + 0.08938401192426682, + 0.1536836177110672, + 0.06591849774122238, + 0.05378352850675583, + 0.019306736066937447 + ], + "credit_map": [ + 0.014594846405088902, + 0.013398464769124985, + 0.015534627251327038, + 0.022151000797748566, + 0.01432279497385025, + 0.015436389483511448, + 0.055515460669994354, + 0.04348278045654297, + 0.02021113596856594, + 0.006678476929664612, + 0.014955521561205387, + 0.014952138066291809, + 0.007783681619912386, + 0.006459709722548723, + 0.010085135698318481, + 0.015322335064411163, + 0.010863833129405975, + 0.019206728786230087, + 0.01116659864783287, + 0.05090919882059097, + 0.016076063737273216, + 0.01655956730246544, + 0.020048389211297035, + 0.02586491033434868, + 0.008304385468363762, + 0.004821379669010639, + 0.011229579336941242, + 0.013220476917922497, + 0.012186317704617977, + 0.014127643778920174, + 0.01118851825594902, + 0.016027992591261864, + 0.008045659400522709, + 0.004583463538438082, + 0.007853077724575996, + 0.00766719039529562, + 0.011004910804331303, + 0.007139927241951227, + 0.01038477011024952, + 0.008558530360460281, + 0.011502514593303204, + 0.00477087264880538, + 0.009848318062722683, + 0.0072302501648664474, + 0.013770049437880516, + 0.011628745123744011, + 0.018308276310563087, + 0.012324989773333073, + 0.02105291560292244, + 0.008133932016789913, + 0.011762956157326698, + 0.009870647452771664, + 0.016318304464221, + 0.036597006022930145, + 0.02265072613954544, + 0.02729714661836624, + 0.005917396396398544, + 0.004161459859460592, + 0.005095627158880234, + 0.004386465530842543, + 0.009928923100233078, + 0.012360656633973122, + 0.010334222577512264, + 0.013733753934502602, + 0.0048123630695044994, + 0.0034810472279787064, + 0.0034726010635495186, + 0.006472222972661257, + 0.00513782212510705, + 0.009119329042732716, + 0.009034798480570316, + 0.012253351509571075, + 0.001653975574299693, + 0.0013898138422518969, + 0.0014472658513113856, + 0.0019581206142902374, + 0.002935972297564149, + 0.0021677662152796984, + 0.00404568063095212, + 0.003708142787218094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.038103725761175156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_15/index_tensor([4916, 4916])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_15/index_tensor([4916, 4916])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c2af321bd2d2d4ba8f39f660457858ec401a189c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_15/index_tensor([4916, 4916])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot begins at a slightly upward angle, then steadily and slowly zooms out, initially focusing on a dragon boat struggling against the raging white rapids. The boat's hull is crafted from deep red polished wood, its sides intricately carved with golden dragon scales, glistening wetly under the splashing water. A fiery, lifelike dragon head adorns the prow, its maw gaping as if in a roar, its bright red eyes glaring at the churning waves ahead. Just then, a powerful surge jolts the boat upward, lurching it violently. The intricately carved prow is lifted high, nearly breaking the surface of the water, demonstrating its immense power against the current. As the camera slowly pulls away, the picture gradually reveals a wider scene: the dragon boat is not isolated, it is located in a turbulent river surrounded by steep, lush canyons. The slanting afternoon sunlight penetrates the thin mist, casting mottled light and shadows on the turbulent water. The dragon boat is like an arrow shot from a bow in this magnificent and wild natural landscape, full of a tense, majestic and epic atmosphere.", + "seed": 31052, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_15/index_4916_seed_31052_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09672905169427395, + "vlm_logprob": -0.09672905169427395, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.09672905169427395, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2598063051700592 + ], + "opposite_logprob_tokens": [ + -1.5098062753677368 + ], + "logprob_reward": -0.2598063051700592, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2598063051700592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037104539573192596, + 0.0654313787817955, + 0.05827454850077629, + 0.1655881553888321, + 0.021160490810871124, + 0.037684328854084015, + 0.08995792269706726, + 0.25977256894111633, + 0.17158471047878265, + 0.09344135224819183 + ], + "credit_map": [ + 0.00398638192564249, + 0.004144526552408934, + 0.0039831846952438354, + 0.0037383325397968292, + 0.005671580322086811, + 0.005074054468423128, + 0.004678895231336355, + 0.005827581975609064, + 0.008802907541394234, + 0.006617329549044371, + 0.013837769627571106, + 0.010099376551806927, + 0.005458566825836897, + 0.0048417262732982635, + 0.005175887607038021, + 0.01059782039374113, + 0.008989711292088032, + 0.0060516586527228355, + 0.006763301324099302, + 0.00618346594274044, + 0.007206355687230825, + 0.011140516959130764, + 0.006764578633010387, + 0.005174959544092417, + 0.05753105506300926, + 0.022436873987317085, + 0.01185618992894888, + 0.01159169152379036, + 0.009300429373979568, + 0.011889846995472908, + 0.009089422412216663, + 0.03189263865351677, + 0.003466981928795576, + 0.0024995235726237297, + 0.002789752557873726, + 0.0035423727240413427, + 0.002172689652070403, + 0.0020862827077507973, + 0.0021960102021694183, + 0.0024068783968687057, + 0.0041942717507481575, + 0.007143568713217974, + 0.003876635804772377, + 0.0032716935966163874, + 0.005201756022870541, + 0.006366487126797438, + 0.004378267098218203, + 0.0032516526989638805, + 0.0075743431225419044, + 0.008997459895908833, + 0.013466941192746162, + 0.011356654576957226, + 0.012495864182710648, + 0.01918107643723488, + 0.007671864703297615, + 0.009213718585669994, + 0.03176452964544296, + 0.0417703241109848, + 0.023750634863972664, + 0.025679191574454308, + 0.03934117779135704, + 0.03541701287031174, + 0.0295492522418499, + 0.03250046819448471, + 0.022571038454771042, + 0.012052919715642929, + 0.021647652611136436, + 0.018042610958218575, + 0.0244948361068964, + 0.023836052045226097, + 0.027720730751752853, + 0.02121887169778347, + 0.02075141854584217, + 0.014275780878961086, + 0.006149766501039267, + 0.007465527858585119, + 0.015511292964220047, + 0.016737649217247963, + 0.007310288958251476, + 0.005239622201770544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.2598063051700592 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867442324757576 + ], + "opposite_logprob_tokens": [ + -3.0486743450164795 + ], + "logprob_reward": -0.04867442324757576, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867442324757576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07023077458143234, + 0.3931415379047394, + 0.07453391700983047, + 0.08025645464658737, + 0.08167675882577896, + 0.0806378722190857, + 0.02884940430521965, + 0.06391172856092453, + 0.0608685240149498, + 0.06589297950267792 + ], + "credit_map": [ + 0.014512848109006882, + 0.005538804456591606, + 0.01212310791015625, + 0.006610526703298092, + 0.006141458172351122, + 0.009122670628130436, + 0.008431053720414639, + 0.00775030767545104, + 0.04255983605980873, + 0.033191826194524765, + 0.0839032232761383, + 0.05211827903985977, + 0.035380769520998, + 0.03386326879262924, + 0.03322407975792885, + 0.07890025526285172, + 0.009307291358709335, + 0.00848373956978321, + 0.01296534389257431, + 0.010324840433895588, + 0.0059075565077364445, + 0.007075687404721975, + 0.004420544486492872, + 0.016048921272158623, + 0.007668195758014917, + 0.00704374723136425, + 0.016128066927194595, + 0.019396869465708733, + 0.006316634360700846, + 0.006919080391526222, + 0.006059883162379265, + 0.010723983868956566, + 0.011318914592266083, + 0.00863010436296463, + 0.00924459844827652, + 0.011456981301307678, + 0.008232778869569302, + 0.011913367547094822, + 0.011784064583480358, + 0.009095961228013039, + 0.010796237736940384, + 0.016487043350934982, + 0.013847979716956615, + 0.007501903455704451, + 0.009526684880256653, + 0.007260361220687628, + 0.008228742517530918, + 0.006988926324993372, + 0.003303360426798463, + 0.0023418234195560217, + 0.004644032102078199, + 0.003379266941919923, + 0.004843292757868767, + 0.0059123653918504715, + 0.0026913932524621487, + 0.0017338752513751388, + 0.008162846788764, + 0.009031902067363262, + 0.005137261003255844, + 0.006176335271447897, + 0.009561463259160519, + 0.009027933701872826, + 0.009549169801175594, + 0.007264820858836174, + 0.006554148625582457, + 0.004470855463296175, + 0.005851789843291044, + 0.008970563299953938, + 0.0076493313536047935, + 0.007176232524216175, + 0.013964271172881126, + 0.006231337320059538, + 0.013919857330620289, + 0.008247967809438705, + 0.004214119631797075, + 0.006263894960284233, + 0.00982979591935873, + 0.011170770972967148, + 0.006607224699109793, + 0.005639354698359966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.04867442324757576 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188826739788055 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.018188826739788055, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188826739788055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17652419209480286, + 0.08386177569627762, + 0.05934891849756241, + 0.0924050509929657, + 0.049277596175670624, + 0.1965324431657791, + 0.04066897928714752, + 0.06115267425775528, + 0.1686372458934784, + 0.07159111648797989 + ], + "credit_map": [ + 0.02070855349302292, + 0.017745669931173325, + 0.02692548744380474, + 0.021304365247488022, + 0.015403318218886852, + 0.02668219618499279, + 0.02117210067808628, + 0.026582498103380203, + 0.012966031208634377, + 0.0090886065736413, + 0.01241263747215271, + 0.013718868605792522, + 0.00936840008944273, + 0.009352411143481731, + 0.00703453691676259, + 0.009920288808643818, + 0.00931790191680193, + 0.008139096200466156, + 0.007165995892137289, + 0.007250962778925896, + 0.006876632105559111, + 0.008886138908565044, + 0.0064549632370471954, + 0.005257230252027512, + 0.03010016866028309, + 0.016054604202508926, + 0.007372274063527584, + 0.006552685052156448, + 0.006330189760774374, + 0.00581118231639266, + 0.005925014615058899, + 0.014258929528295994, + 0.011884006671607494, + 0.004804788623005152, + 0.006324011832475662, + 0.005785787012428045, + 0.0039314148016273975, + 0.004112640395760536, + 0.0052758497186005116, + 0.007159099914133549, + 0.01857137493789196, + 0.021584520116448402, + 0.03945213556289673, + 0.014446398243308067, + 0.024506036192178726, + 0.05097971111536026, + 0.011821765452623367, + 0.015170497819781303, + 0.0055696298368275166, + 0.005927846301347017, + 0.005843834951519966, + 0.005319105461239815, + 0.005959231406450272, + 0.005473962519317865, + 0.0037814066745340824, + 0.002793962601572275, + 0.00784637313336134, + 0.008267193101346493, + 0.006162003148347139, + 0.009337970055639744, + 0.008949645794928074, + 0.00854365061968565, + 0.005917105823755264, + 0.0061287363059818745, + 0.027131792157888412, + 0.01414569653570652, + 0.015426279976963997, + 0.022339124232530594, + 0.02067357487976551, + 0.01899167336523533, + 0.035744961351156235, + 0.014184138737618923, + 0.01214404497295618, + 0.009317439049482346, + 0.00427471986040473, + 0.006307553965598345, + 0.016726423054933548, + 0.01373570691794157, + 0.0053675188682973385, + 0.0037177095655351877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.018188826739788055 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278216660022736 + ], + "opposite_logprob_tokens": [ + -2.017782211303711 + ], + "logprob_reward": -0.14278216660022736, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278216660022736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15083655714988708, + 0.09748327732086182, + 0.10976497828960419, + 0.1416301727294922, + 0.04428029805421829, + 0.033927664160728455, + 0.03864467889070511, + 0.11472761631011963, + 0.15266604721546173, + 0.1160387247800827 + ], + "credit_map": [ + 0.019498636946082115, + 0.015969302505254745, + 0.02424277551472187, + 0.018341226503252983, + 0.013965222053229809, + 0.024713104590773582, + 0.014300762675702572, + 0.019805528223514557, + 0.012624917551875114, + 0.010576272383332253, + 0.017344877123832703, + 0.015986137092113495, + 0.006274173967540264, + 0.00862803217023611, + 0.008690237998962402, + 0.017358621582388878, + 0.016483763232827187, + 0.011322780512273312, + 0.011594080366194248, + 0.012179866433143616, + 0.013089008629322052, + 0.01709073968231678, + 0.008111714385449886, + 0.019893020391464233, + 0.046053335070610046, + 0.024588624015450478, + 0.011630415916442871, + 0.011995027773082256, + 0.00910672266036272, + 0.008739233948290348, + 0.008336003869771957, + 0.021180808544158936, + 0.008265037089586258, + 0.004797878675162792, + 0.0068319751881062984, + 0.005674217827618122, + 0.0033943240996450186, + 0.0046967254020273685, + 0.0041656638495624065, + 0.0064544714987277985, + 0.0037761530838906765, + 0.004858947824686766, + 0.004242883995175362, + 0.006180830765515566, + 0.0033976943232119083, + 0.003208564594388008, + 0.004709623754024506, + 0.0035529653541743755, + 0.0056873406283557415, + 0.0044721197336912155, + 0.004579205997288227, + 0.005539879202842712, + 0.0049770851619541645, + 0.005454531870782375, + 0.004323496017605066, + 0.0036110158544033766, + 0.02163580432534218, + 0.013363641686737537, + 0.011336724273860455, + 0.007855230942368507, + 0.015649404376745224, + 0.01741096004843712, + 0.01037391647696495, + 0.01710193045437336, + 0.013325225561857224, + 0.014837558381259441, + 0.01287093572318554, + 0.02066316455602646, + 0.021455494686961174, + 0.018273139372467995, + 0.033445123583078384, + 0.017795391380786896, + 0.03168134018778801, + 0.015901951119303703, + 0.007653761655092239, + 0.0076265716925263405, + 0.014873139560222626, + 0.021256500855088234, + 0.008760800585150719, + 0.008284647017717361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.14278216660022736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193536713719368 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193536713719368, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193536713719368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11434736847877502, + 0.09840446710586548, + 0.13903409242630005, + 0.12674352526664734, + 0.05562647804617882, + 0.09914245456457138, + 0.05646929144859314, + 0.09759283810853958, + 0.09061950445175171, + 0.12201997637748718 + ], + "credit_map": [ + 0.013506381772458553, + 0.008805137127637863, + 0.01846279390156269, + 0.008695840835571289, + 0.009005725383758545, + 0.02270548976957798, + 0.011737609282135963, + 0.021428389474749565, + 0.0142765361815691, + 0.009578743949532509, + 0.015593812800943851, + 0.018272915855050087, + 0.008176577277481556, + 0.010230397805571556, + 0.011657805182039738, + 0.010617674328386784, + 0.016352610662579536, + 0.012395272962749004, + 0.0192815363407135, + 0.014066798612475395, + 0.012766358442604542, + 0.025558840483427048, + 0.01503206416964531, + 0.023580612614750862, + 0.04217996075749397, + 0.015744701027870178, + 0.010786132887005806, + 0.01618686504662037, + 0.0068831667304039, + 0.009699567221105099, + 0.006954493932425976, + 0.018308645114302635, + 0.014504248276352882, + 0.0063173649832606316, + 0.007213412784039974, + 0.006791504565626383, + 0.004526807460933924, + 0.00469615962356329, + 0.005582752637565136, + 0.005994230043143034, + 0.011197185143828392, + 0.01704656332731247, + 0.015913398936390877, + 0.011561155319213867, + 0.010867584496736526, + 0.012048885226249695, + 0.012045292183756828, + 0.008462396450340748, + 0.0063920398242771626, + 0.006086345762014389, + 0.009212451055645943, + 0.0047801765613257885, + 0.009366108104586601, + 0.012618140317499638, + 0.003564639948308468, + 0.004449385683983564, + 0.01844893768429756, + 0.012965366244316101, + 0.010288991034030914, + 0.010001737624406815, + 0.012066751718521118, + 0.012872070074081421, + 0.01076885499060154, + 0.010180126875638962, + 0.018367309123277664, + 0.007848371751606464, + 0.011452753096818924, + 0.008869237266480923, + 0.011267252266407013, + 0.01153749693185091, + 0.012959570623934269, + 0.008317510597407818, + 0.024216964840888977, + 0.01609990745782852, + 0.008571246638894081, + 0.013795508071780205, + 0.01703372225165367, + 0.01965165138244629, + 0.01272596511989832, + 0.009925021789968014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.014193536713719368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins at a slightly upward angle, then steadily and slowly zooms out, initially focusing on a dragon boat struggling against the raging white rapids. The boat's hull is crafted from deep red polished wood, its sides intricately carved with golden dragon scales, glistening wetly under the splashing water. A fiery, lifelike dragon head adorns the prow, its maw gaping as if in a roar, its bright red eyes glaring at the churning waves ahead. Just then, a powerful surge jolts the boat upward, lurching it violently. The intricately carved prow is lifted high, nearly breaking the surface of the water, demonstrating its immense power against the current. As the camera slowly pulls away, the picture gradually reveals a wider scene: the dragon boat is not isolated, it is located in a turbulent river surrounded by steep, lush canyons. The slanting afternoon sunlight penetrates the thin mist, casting mottled light and shadows on the turbulent water. The dragon boat is like an arrow shot from a bow in this magnificent and wild natural landscape, full of a tense, majestic and epic atmosphere.", + "seed": 31053, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_15/index_4916_seed_31053_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13709806837141514, + "vlm_logprob": -0.13709806837141514, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.13709806837141514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5806775093078613 + ], + "opposite_logprob_tokens": [ + -0.8306775093078613 + ], + "logprob_reward": -0.5806775093078613, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5806775093078613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09395348280668259, + 0.09144105017185211, + 0.08698687702417374, + 0.058585748076438904, + 0.05005399510264397, + 0.12385382503271103, + 0.09662280231714249, + 0.034323301166296005, + 0.10659628361463547, + 0.2575826644897461 + ], + "credit_map": [ + 0.007907332852482796, + 0.011574704200029373, + 0.01881844736635685, + 0.01580849662423134, + 0.0077020516619086266, + 0.00407553231343627, + 0.014122734777629375, + 0.013944180682301521, + 0.007637750823050737, + 0.01957213506102562, + 0.009176484309136868, + 0.013294456526637077, + 0.016276955604553223, + 0.005888147745281458, + 0.008183397352695465, + 0.011411726474761963, + 0.010643878020346165, + 0.021679949015378952, + 0.009642085991799831, + 0.007168189622461796, + 0.012289488688111305, + 0.011545549146831036, + 0.007179892156273127, + 0.006837845779955387, + 0.006373760290443897, + 0.0064092157408595085, + 0.008212615735828876, + 0.007397875189781189, + 0.006484922021627426, + 0.007843843661248684, + 0.009844779968261719, + 0.006018735934048891, + 0.004694866482168436, + 0.006050456315279007, + 0.007710288744419813, + 0.006362899672240019, + 0.00799510907381773, + 0.006919001694768667, + 0.006424245424568653, + 0.0038971281610429287, + 0.0169630479067564, + 0.011449537239968777, + 0.011219172738492489, + 0.016283905133605003, + 0.012965126894414425, + 0.01695593260228634, + 0.024417003616690636, + 0.013600095175206661, + 0.006434772163629532, + 0.011589682660996914, + 0.01129196584224701, + 0.009691537357866764, + 0.020584221929311752, + 0.014760092832148075, + 0.014466454274952412, + 0.00780407153069973, + 0.005090388469398022, + 0.0055455672554671764, + 0.003917974885553122, + 0.0033153847325593233, + 0.0037754056975245476, + 0.004178685136139393, + 0.00331995845772326, + 0.005179939791560173, + 0.013690014369785786, + 0.017546167597174644, + 0.00965049210935831, + 0.00800439715385437, + 0.022461412474513054, + 0.017561767250299454, + 0.010047688148915768, + 0.007634344045072794, + 0.037903014570474625, + 0.03678647801280022, + 0.02317393384873867, + 0.024094799533486366, + 0.0650646984577179, + 0.023012572899460793, + 0.024330787360668182, + 0.023216363042593002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.5806775093078613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307618737220764 + ], + "opposite_logprob_tokens": [ + -3.168076276779175 + ], + "logprob_reward": -0.04307618737220764, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307618737220764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12735658884048462, + 0.0727730542421341, + 0.12489785999059677, + 0.23611491918563843, + 0.13406021893024445, + 0.03254581615328789, + 0.06592313945293427, + 0.07850880920886993, + 0.08675792813301086, + 0.04106169939041138 + ], + "credit_map": [ + 0.010439638048410416, + 0.016373060643672943, + 0.024629691615700722, + 0.026957497000694275, + 0.008746047504246235, + 0.007384559139609337, + 0.013136972673237324, + 0.01968911476433277, + 0.00728571554645896, + 0.018414907157421112, + 0.007049889769405127, + 0.009099197573959827, + 0.013974991627037525, + 0.003246492939069867, + 0.005618487950414419, + 0.008083367720246315, + 0.013606984168291092, + 0.01522448007017374, + 0.014757789671421051, + 0.012608972378075123, + 0.01994696445763111, + 0.018279725685715675, + 0.016573723405599594, + 0.013899227604269981, + 0.0191032737493515, + 0.026820939034223557, + 0.028790703043341637, + 0.033524975180625916, + 0.039303772151470184, + 0.02984851785004139, + 0.033235833048820496, + 0.0254869032651186, + 0.009398745372891426, + 0.01594841666519642, + 0.016457483172416687, + 0.016597885638475418, + 0.013870824128389359, + 0.01577271707355976, + 0.027933906763792038, + 0.018080249428749084, + 0.0032781544141471386, + 0.004872384946793318, + 0.0036127702333033085, + 0.0036736971233040094, + 0.003935268148779869, + 0.004759053699672222, + 0.005711175035685301, + 0.002703314181417227, + 0.007530460599809885, + 0.009550543501973152, + 0.01057797484099865, + 0.007338980212807655, + 0.005791062489151955, + 0.01050411257892847, + 0.006571328267455101, + 0.008058680221438408, + 0.008589806966483593, + 0.013688352890312672, + 0.00688494136556983, + 0.007532036863267422, + 0.01225343532860279, + 0.010929746553301811, + 0.007233007811009884, + 0.01139748003333807, + 0.014405893161892891, + 0.016133051365613937, + 0.005970340222120285, + 0.005054842680692673, + 0.0173078253865242, + 0.010321018286049366, + 0.009942369535565376, + 0.007622584700584412, + 0.006053841672837734, + 0.00639639887958765, + 0.002781480085104704, + 0.004181664437055588, + 0.011193414218723774, + 0.003356338944286108, + 0.002985356142744422, + 0.004113204311579466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.04307618737220764 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020580321550369263 + ], + "opposite_logprob_tokens": [ + -3.895580291748047 + ], + "logprob_reward": -0.020580321550369263, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580321550369263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07304078340530396, + 0.5224976539611816, + 0.04918241873383522, + 0.04440157860517502, + 0.030639471486210823, + 0.04855365306138992, + 0.11857444792985916, + 0.04561799392104149, + 0.026624640449881554, + 0.040867287665605545 + ], + "credit_map": [ + 0.004830597434192896, + 0.008594871498644352, + 0.0108634727075696, + 0.016621144488453865, + 0.006643110886216164, + 0.005129410419613123, + 0.010381242260336876, + 0.009976946748793125, + 0.052709028124809265, + 0.14027680456638336, + 0.04954834654927254, + 0.06783197075128555, + 0.0971282571554184, + 0.021874049678444862, + 0.03525034338235855, + 0.05787893757224083, + 0.006188198458403349, + 0.011506077833473682, + 0.00669720396399498, + 0.005141052883118391, + 0.005623106844723225, + 0.006033783778548241, + 0.003929207567125559, + 0.004063793458044529, + 0.0037724252324551344, + 0.0042864540591835976, + 0.004053046926856041, + 0.005695558153092861, + 0.006768583320081234, + 0.007210605777800083, + 0.007459037471562624, + 0.005155871622264385, + 0.0029598716646432877, + 0.004429271444678307, + 0.003094404935836792, + 0.002979683456942439, + 0.0055076549760997295, + 0.004505641292780638, + 0.004164228681474924, + 0.002998718060553074, + 0.0065164463594555855, + 0.004656586796045303, + 0.0050093685276806355, + 0.005461011081933975, + 0.004602999426424503, + 0.006599804852157831, + 0.009844934567809105, + 0.005862506106495857, + 0.015178494155406952, + 0.01634431630373001, + 0.014265762642025948, + 0.013620498590171337, + 0.016744671389460564, + 0.01944766752421856, + 0.015415170229971409, + 0.0075578829273581505, + 0.006316537968814373, + 0.010032473132014275, + 0.003994818776845932, + 0.007531068753451109, + 0.004375332500785589, + 0.006422795820981264, + 0.003939493093639612, + 0.0030054764356464148, + 0.0038631281349807978, + 0.004781621973961592, + 0.002182664815336466, + 0.0015446085017174482, + 0.006581095978617668, + 0.0034245483111590147, + 0.0024534310214221478, + 0.0017935459036380053, + 0.006157690659165382, + 0.005301173776388168, + 0.002766778226941824, + 0.003497346304357052, + 0.007686729542911053, + 0.005741778761148453, + 0.0031431778334081173, + 0.006572618614882231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.020580321550369263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058207429945469 + ], + "opposite_logprob_tokens": [ + -3.8955819606781006 + ], + "logprob_reward": -0.02058207429945469, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058207429945469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.125712051987648, + 0.06984414160251617, + 0.1330127716064453, + 0.12900258600711823, + 0.06061616539955139, + 0.0951889380812645, + 0.07980085909366608, + 0.11342156678438187, + 0.1252337545156479, + 0.06816723942756653 + ], + "credit_map": [ + 0.009864416904747486, + 0.012770981527864933, + 0.023066148161888123, + 0.018954942002892494, + 0.011519434861838818, + 0.008780556730926037, + 0.02087041176855564, + 0.01988515816628933, + 0.007147228345274925, + 0.01483356487005949, + 0.00813873391598463, + 0.010246297344565392, + 0.011282713152468204, + 0.003941259812563658, + 0.0060345870442688465, + 0.008219754323363304, + 0.013835064135491848, + 0.015189855359494686, + 0.014223143458366394, + 0.011636842042207718, + 0.024571264162659645, + 0.0163690447807312, + 0.020842675119638443, + 0.016344895586371422, + 0.01272650994360447, + 0.017622441053390503, + 0.01540699228644371, + 0.017260434105992317, + 0.02205575257539749, + 0.01653968170285225, + 0.01596282236278057, + 0.011427941732108593, + 0.004593727178871632, + 0.007283850107342005, + 0.009068718180060387, + 0.007397770415991545, + 0.006849837023764849, + 0.007971656508743763, + 0.010741538368165493, + 0.006709069944918156, + 0.011101000010967255, + 0.013004804030060768, + 0.00944562442600727, + 0.022697538137435913, + 0.009596155025064945, + 0.01617540791630745, + 0.006993443239480257, + 0.0061749666929244995, + 0.013175410218536854, + 0.009199688211083412, + 0.01033130194991827, + 0.008429672569036484, + 0.00795076135545969, + 0.013068660162389278, + 0.010919473133981228, + 0.006725891027599573, + 0.014928339049220085, + 0.019123267382383347, + 0.017091386020183563, + 0.015718918293714523, + 0.013357212767004967, + 0.015553921461105347, + 0.01071853842586279, + 0.0069299847818911076, + 0.02068856917321682, + 0.018642935901880264, + 0.01267072930932045, + 0.016679871827363968, + 0.014201583340764046, + 0.01910357177257538, + 0.013819428160786629, + 0.009427059441804886, + 0.012682139873504639, + 0.00915574748069048, + 0.006201011594384909, + 0.00534044299274683, + 0.01399238221347332, + 0.006857300642877817, + 0.006465245969593525, + 0.007472964469343424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.02058207429945469 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057424932718277 + ], + "opposite_logprob_tokens": [ + -3.8955743312835693 + ], + "logprob_reward": -0.02057424932718277, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057424932718277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08713708072900772, + 0.1794419288635254, + 0.07426135987043381, + 0.04411010816693306, + 0.08015158772468567, + 0.04456940293312073, + 0.1748049110174179, + 0.08428912609815598, + 0.09778792411088943, + 0.13344651460647583 + ], + "credit_map": [ + 0.006123367231339216, + 0.008041806519031525, + 0.013522664085030556, + 0.013107437640428543, + 0.012548561207950115, + 0.008570842444896698, + 0.013296997174620628, + 0.01192540768533945, + 0.019526652991771698, + 0.04735390469431877, + 0.018452756106853485, + 0.02097235433757305, + 0.02587791159749031, + 0.00947204977273941, + 0.017185872420668602, + 0.020600421354174614, + 0.006520919967442751, + 0.008544167503714561, + 0.008094573393464088, + 0.007347099017351866, + 0.013848849572241306, + 0.007987952791154385, + 0.011437570676207542, + 0.010480228811502457, + 0.005471282172948122, + 0.0033928805496543646, + 0.008602974005043507, + 0.005011352710425854, + 0.006943457759916782, + 0.005093572195619345, + 0.006167598534375429, + 0.0034269874449819326, + 0.007712401449680328, + 0.01158370915800333, + 0.013876580633223057, + 0.008525408804416656, + 0.013168049976229668, + 0.01216974575072527, + 0.007408371660858393, + 0.005707326345145702, + 0.006806789431720972, + 0.004927315749228001, + 0.006932942196726799, + 0.007300044875591993, + 0.004299970343708992, + 0.006654472555965185, + 0.004933985415846109, + 0.0027138839941471815, + 0.020624220371246338, + 0.03226659819483757, + 0.027069328352808952, + 0.024098368361592293, + 0.014925248920917511, + 0.029278049245476723, + 0.01811177283525467, + 0.008431337773799896, + 0.011190896853804588, + 0.018619852140545845, + 0.008297333493828773, + 0.014481290243566036, + 0.007645230274647474, + 0.011515593156218529, + 0.007410600781440735, + 0.005128329619765282, + 0.018564127385616302, + 0.016893237829208374, + 0.008716188371181488, + 0.01210178341716528, + 0.011541751213371754, + 0.012692474760115147, + 0.009981611743569374, + 0.0072967493906617165, + 0.01841088756918907, + 0.017822623252868652, + 0.01200698409229517, + 0.012436331249773502, + 0.029487861320376396, + 0.012851784005761147, + 0.013808025047183037, + 0.016622019931674004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.02057424932718277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins at a slightly upward angle, then steadily and slowly zooms out, initially focusing on a dragon boat struggling against the raging white rapids. The boat's hull is crafted from deep red polished wood, its sides intricately carved with golden dragon scales, glistening wetly under the splashing water. A fiery, lifelike dragon head adorns the prow, its maw gaping as if in a roar, its bright red eyes glaring at the churning waves ahead. Just then, a powerful surge jolts the boat upward, lurching it violently. The intricately carved prow is lifted high, nearly breaking the surface of the water, demonstrating its immense power against the current. As the camera slowly pulls away, the picture gradually reveals a wider scene: the dragon boat is not isolated, it is located in a turbulent river surrounded by steep, lush canyons. The slanting afternoon sunlight penetrates the thin mist, casting mottled light and shadows on the turbulent water. The dragon boat is like an arrow shot from a bow in this magnificent and wild natural landscape, full of a tense, majestic and epic atmosphere.", + "seed": 31054, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_15/index_4916_seed_31054_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14133976977318524, + "vlm_logprob": -0.14133976977318524, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.14133976977318524, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6395247578620911 + ], + "opposite_logprob_tokens": [ + -0.7645247578620911 + ], + "logprob_reward": -0.6395247578620911, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6395247578620911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1315140277147293, + 0.09074347466230392, + 0.050487857311964035, + 0.046874385327100754, + 0.09891806542873383, + 0.3010239601135254, + 0.08333639055490494, + 0.052837859839200974, + 0.09073024988174438, + 0.05353379622101784 + ], + "credit_map": [ + 0.013678696937859058, + 0.025356778874993324, + 0.009804065339267254, + 0.0166010744869709, + 0.020062491297721863, + 0.019398733973503113, + 0.011195225641131401, + 0.015416947193443775, + 0.003528386587277055, + 0.007543825078755617, + 0.00936484057456255, + 0.03390856459736824, + 0.020658064633607864, + 0.004904267843812704, + 0.00603148527443409, + 0.0048040347173810005, + 0.006555883213877678, + 0.00442645326256752, + 0.004718594718724489, + 0.005565827712416649, + 0.01137047540396452, + 0.005525403190404177, + 0.004987746477127075, + 0.0073374672792851925, + 0.005588467698544264, + 0.004942645784467459, + 0.006488169077783823, + 0.0055899410508573055, + 0.00620600301772356, + 0.005236988887190819, + 0.006550561636686325, + 0.0062716081738471985, + 0.027586912736296654, + 0.021812211722135544, + 0.008583351038396358, + 0.012673495337367058, + 0.006522222422063351, + 0.005844854284077883, + 0.00824008509516716, + 0.007654928136616945, + 0.03577577695250511, + 0.0580965057015419, + 0.04534010589122772, + 0.08244150876998901, + 0.015317880548536777, + 0.015259390696883202, + 0.03524395823478699, + 0.013548820279538631, + 0.013415900990366936, + 0.00822304654866457, + 0.01932082697749138, + 0.009834304451942444, + 0.007240006234496832, + 0.005769211798906326, + 0.014013855718076229, + 0.0055192322470247746, + 0.005026608239859343, + 0.00612298771739006, + 0.009088446386158466, + 0.010606123134493828, + 0.005118613131344318, + 0.00438355840742588, + 0.005497072357684374, + 0.00699444767087698, + 0.00979229249060154, + 0.014291892759501934, + 0.012183588929474354, + 0.015618299134075642, + 0.012063651345670223, + 0.00800605583935976, + 0.008693731389939785, + 0.010080729611217976, + 0.010131516493856907, + 0.008366307243704796, + 0.004842570051550865, + 0.005712007638067007, + 0.011244011111557484, + 0.004207034129649401, + 0.003822216298431158, + 0.005208126734942198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.6395247578620911 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059970796108246 + ], + "opposite_logprob_tokens": [ + -3.895599603652954 + ], + "logprob_reward": -0.02059970796108246, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059970796108246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2863016128540039, + 0.09471355378627777, + 0.07445857673883438, + 0.06872030347585678, + 0.07842344045639038, + 0.13519780337810516, + 0.07087358087301254, + 0.05841638520359993, + 0.10297372192144394, + 0.029920976608991623 + ], + "credit_map": [ + 0.015627244487404823, + 0.022806944325566292, + 0.026371212676167488, + 0.06335458159446716, + 0.03002786450088024, + 0.041140366345644, + 0.0446789413690567, + 0.042294472455978394, + 0.0063186329789459705, + 0.013220316730439663, + 0.012541530653834343, + 0.009166235104203224, + 0.02175194025039673, + 0.009866525419056416, + 0.010610002093017101, + 0.011238384991884232, + 0.0073028793558478355, + 0.01129672210663557, + 0.006400438025593758, + 0.021182864904403687, + 0.00806419551372528, + 0.007376546040177345, + 0.004900582134723663, + 0.007934357970952988, + 0.007406457792967558, + 0.006706764921545982, + 0.01191565953195095, + 0.007865417748689651, + 0.007155274506658316, + 0.006374432239681482, + 0.009308886714279652, + 0.011987412348389626, + 0.01629721373319626, + 0.016394920647144318, + 0.011643778532743454, + 0.010442606173455715, + 0.005254662130028009, + 0.005026687402278185, + 0.006943264044821262, + 0.006420312449336052, + 0.011240417137742043, + 0.007229870650917292, + 0.04274752363562584, + 0.032582662999629974, + 0.006465019192546606, + 0.01375266257673502, + 0.01040054764598608, + 0.01077911164611578, + 0.008295864798128605, + 0.011270681396126747, + 0.00986882857978344, + 0.014671060256659985, + 0.004178821574896574, + 0.005140740890055895, + 0.010798710398375988, + 0.006648878566920757, + 0.004677488934248686, + 0.007712921593338251, + 0.00823882594704628, + 0.007354496978223324, + 0.006165384314954281, + 0.006545564625412226, + 0.009694487787783146, + 0.008027219213545322, + 0.00415238831192255, + 0.02638961933553219, + 0.01074299868196249, + 0.038774825632572174, + 0.005648306105285883, + 0.005176951177418232, + 0.005395440384745598, + 0.006693198811262846, + 0.004065379500389099, + 0.00400490453466773, + 0.004218646790832281, + 0.005283148493617773, + 0.004413219168782234, + 0.0034561010543257, + 0.00214343867264688, + 0.0023361400235444307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.02059970796108246 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193654991686344 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193654991686344, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193654991686344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10351132601499557, + 0.23270265758037567, + 0.19793207943439484, + 0.06764781475067139, + 0.06673327833414078, + 0.11205799877643585, + 0.026529619470238686, + 0.12720799446105957, + 0.034501392394304276, + 0.03117581643164158 + ], + "credit_map": [ + 0.008484059944748878, + 0.012197079136967659, + 0.008689108304679394, + 0.014369349926710129, + 0.013862453401088715, + 0.018083930015563965, + 0.016514819115400314, + 0.011310519650578499, + 0.022014982998371124, + 0.023574145510792732, + 0.029197681695222855, + 0.026432130485773087, + 0.03043902851641178, + 0.029524553567171097, + 0.025376426056027412, + 0.04614371061325073, + 0.030363908037543297, + 0.03134516626596451, + 0.010791795328259468, + 0.04194607585668564, + 0.03245268017053604, + 0.01349701453000307, + 0.010508591309189796, + 0.027026835829019547, + 0.008838641457259655, + 0.00801702681928873, + 0.01136693637818098, + 0.0063135623931884766, + 0.006374863442033529, + 0.006891308818012476, + 0.006550678052008152, + 0.013294804841279984, + 0.016655968502163887, + 0.01300899963825941, + 0.006475945468991995, + 0.007435887586325407, + 0.00502595491707325, + 0.004372842609882355, + 0.007441814057528973, + 0.0063158622942864895, + 0.00814088061451912, + 0.008665108121931553, + 0.022396298125386238, + 0.02121092565357685, + 0.006734343245625496, + 0.012865619733929634, + 0.01570725627243519, + 0.016337569802999496, + 0.0038273034151643515, + 0.0036156887654215097, + 0.003769163740798831, + 0.004365337546914816, + 0.002848654054105282, + 0.002593840006738901, + 0.0031154181342571974, + 0.0023942124098539352, + 0.009575708769261837, + 0.01661292091012001, + 0.01805487647652626, + 0.01656137779355049, + 0.01349314022809267, + 0.013304193504154682, + 0.02140389010310173, + 0.01820189878344536, + 0.004424989223480225, + 0.004612215328961611, + 0.004626366309821606, + 0.007567413151264191, + 0.00348911271430552, + 0.002549506491050124, + 0.0030423873104155064, + 0.004189400002360344, + 0.004544445313513279, + 0.0034148283302783966, + 0.00461653433740139, + 0.003983196336776018, + 0.004447081591933966, + 0.002858242020010948, + 0.002352389506995678, + 0.004959098994731903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.014193654991686344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194594696164131 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194594696164131, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194594696164131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1576574146747589, + 0.10993034392595291, + 0.06024182587862015, + 0.06732505559921265, + 0.12267706543207169, + 0.17557670176029205, + 0.07315128296613693, + 0.065679170191288, + 0.03030368685722351, + 0.137457475066185 + ], + "credit_map": [ + 0.009725202806293964, + 0.020752424374222755, + 0.013187376782298088, + 0.0396697111427784, + 0.015038730576634407, + 0.02196430042386055, + 0.01795731857419014, + 0.019362354651093483, + 0.007562032900750637, + 0.020449139177799225, + 0.011224977672100067, + 0.03110135719180107, + 0.009252953343093395, + 0.007778938394039869, + 0.010734442621469498, + 0.011826490052044392, + 0.007574939634650946, + 0.0060954624786973, + 0.004522828385233879, + 0.014263785444200039, + 0.015224920585751534, + 0.004845259711146355, + 0.004463948309421539, + 0.003250683890655637, + 0.0072776819579303265, + 0.008252228610217571, + 0.008605622686445713, + 0.007729803677648306, + 0.009038821794092655, + 0.006969085428863764, + 0.008431224152445793, + 0.011020584963262081, + 0.028707413002848625, + 0.027653545141220093, + 0.020141001790761948, + 0.01091301441192627, + 0.007661119103431702, + 0.007047766353935003, + 0.0119360676035285, + 0.008617142215371132, + 0.017138414084911346, + 0.012785752303898335, + 0.049102067947387695, + 0.0336909219622612, + 0.009016888216137886, + 0.020545266568660736, + 0.016714464873075485, + 0.016582924872636795, + 0.00959414429962635, + 0.007787217386066914, + 0.011894826777279377, + 0.007818290963768959, + 0.007150947116315365, + 0.0060936463996768, + 0.011654057539999485, + 0.011158155277371407, + 0.006428822409361601, + 0.007517807651311159, + 0.010518134571611881, + 0.014369582757353783, + 0.005763980094343424, + 0.006217985413968563, + 0.008555538021028042, + 0.006307319737970829, + 0.006468798033893108, + 0.005462041590362787, + 0.0030233284924179316, + 0.0034986725077033043, + 0.0029580602422356606, + 0.0021132095716893673, + 0.0027828318998217583, + 0.003996741957962513, + 0.02535131201148033, + 0.01478785090148449, + 0.014740080572664738, + 0.01475957129150629, + 0.02406780607998371, + 0.015332449227571487, + 0.011838791891932487, + 0.016579609364271164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.014194594696164131 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186133354902267 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018186133354902267, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186133354902267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11614194512367249, + 0.06362592428922653, + 0.0754447728395462, + 0.054469626396894455, + 0.10775777697563171, + 0.162838876247406, + 0.06024282053112984, + 0.18855369091033936, + 0.10262218117713928, + 0.06830234825611115 + ], + "credit_map": [ + 0.007458778563886881, + 0.01334140170365572, + 0.00846060924232006, + 0.019546953961253166, + 0.02189852111041546, + 0.018341878429055214, + 0.015041210688650608, + 0.012052604928612709, + 0.004377173725515604, + 0.012261097319424152, + 0.008843646384775639, + 0.006038000341504812, + 0.008058077655732632, + 0.00684947706758976, + 0.0061810496263206005, + 0.011017410084605217, + 0.00793447159230709, + 0.009544300846755505, + 0.0053389654494822025, + 0.023622289299964905, + 0.00949006900191307, + 0.005654522217810154, + 0.004273277707397938, + 0.009586881846189499, + 0.006822838447988033, + 0.0037649201694875956, + 0.0059720915742218494, + 0.0068097529001533985, + 0.006920351646840572, + 0.007369521539658308, + 0.008965683169662952, + 0.00784447230398655, + 0.01931346021592617, + 0.017227159813046455, + 0.0122554125264287, + 0.014607297256588936, + 0.008469940163195133, + 0.009523497894406319, + 0.016888471320271492, + 0.009472544305026531, + 0.02434254065155983, + 0.03719909489154816, + 0.027288686484098434, + 0.029896872118115425, + 0.009525598958134651, + 0.007782693486660719, + 0.017323726788163185, + 0.009479675441980362, + 0.009400666691362858, + 0.007838042452931404, + 0.012070132419466972, + 0.007339356932789087, + 0.004306808114051819, + 0.004345269408077002, + 0.00799588393419981, + 0.006946663372218609, + 0.018447820097208023, + 0.03606602922081947, + 0.02697300910949707, + 0.02717585116624832, + 0.019198140129446983, + 0.016620945185422897, + 0.02929036132991314, + 0.014781546778976917, + 0.009453133679926395, + 0.01904417760670185, + 0.013239915482699871, + 0.031114840880036354, + 0.008769485168159008, + 0.006835323292762041, + 0.005936645437031984, + 0.008228668943047523, + 0.01341345813125372, + 0.008629219606518745, + 0.007791725918650627, + 0.010729392059147358, + 0.008570502512156963, + 0.007153654005378485, + 0.004721979610621929, + 0.007292421068996191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.018186133354902267 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins at a slightly upward angle, then steadily and slowly zooms out, initially focusing on a dragon boat struggling against the raging white rapids. The boat's hull is crafted from deep red polished wood, its sides intricately carved with golden dragon scales, glistening wetly under the splashing water. A fiery, lifelike dragon head adorns the prow, its maw gaping as if in a roar, its bright red eyes glaring at the churning waves ahead. Just then, a powerful surge jolts the boat upward, lurching it violently. The intricately carved prow is lifted high, nearly breaking the surface of the water, demonstrating its immense power against the current. As the camera slowly pulls away, the picture gradually reveals a wider scene: the dragon boat is not isolated, it is located in a turbulent river surrounded by steep, lush canyons. The slanting afternoon sunlight penetrates the thin mist, casting mottled light and shadows on the turbulent water. The dragon boat is like an arrow shot from a bow in this magnificent and wild natural landscape, full of a tense, majestic and epic atmosphere.", + "seed": 31055, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_15/index_4916_seed_31055_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05081163626164198, + "vlm_logprob": -0.05081163626164198, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.05081163626164198, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05829325690865517 + ], + "opposite_logprob_tokens": [ + -3.058293342590332 + ], + "logprob_reward": -0.05829325690865517, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05829325690865517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16103678941726685, + 0.060946669429540634, + 0.04585159197449684, + 0.02671121433377266, + 0.08993899077177048, + 0.09284316748380661, + 0.3974400460720062, + 0.06762224435806274, + 0.027574332430958748, + 0.030034910887479782 + ], + "credit_map": [ + 0.01085011474788189, + 0.054293155670166016, + 0.023888802155852318, + 0.007270308677107096, + 0.006598975509405136, + 0.012630504556000233, + 0.0056419987231493, + 0.03986293822526932, + 0.008584592491388321, + 0.007739624008536339, + 0.0089025367051363, + 0.0071542817167937756, + 0.005451930221170187, + 0.005346572492271662, + 0.007493471261113882, + 0.010273661464452744, + 0.003725325455889106, + 0.004072212148457766, + 0.0031839958392083645, + 0.01241281908005476, + 0.005335996858775616, + 0.003419032320380211, + 0.00802552793174982, + 0.005676681641489267, + 0.002519066445529461, + 0.0025150610599666834, + 0.0028057873714715242, + 0.004339225124567747, + 0.0030207745730876923, + 0.0034500285983085632, + 0.004892272874712944, + 0.0031689980532974005, + 0.009689816273748875, + 0.009785016067326069, + 0.007757922634482384, + 0.012086603790521622, + 0.012326443567872047, + 0.012569162994623184, + 0.015795588493347168, + 0.009928431361913681, + 0.018358275294303894, + 0.00807269848883152, + 0.005135628394782543, + 0.006199899595230818, + 0.019629670307040215, + 0.010440988466143608, + 0.014486611820757389, + 0.01051939744502306, + 0.06350734829902649, + 0.11827437579631805, + 0.06466501951217651, + 0.019827451556921005, + 0.03635795786976814, + 0.037767987698316574, + 0.020579857751727104, + 0.03646005690097809, + 0.011004643514752388, + 0.02644338831305504, + 0.013746668584644794, + 0.002776922658085823, + 0.006161441095173359, + 0.0030081216245889664, + 0.0019752734806388617, + 0.0025057876482605934, + 0.005201374180614948, + 0.0057696388103067875, + 0.004081151448190212, + 0.001632574712857604, + 0.0031685386784374714, + 0.0025307617615908384, + 0.0031615050975233316, + 0.0020287856459617615, + 0.007099970709532499, + 0.003879975527524948, + 0.0021498643327504396, + 0.0019228049786761403, + 0.007032794877886772, + 0.0036775206681340933, + 0.002213814062997699, + 0.002058162819594145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.05829325690865517 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598655566573143 + ], + "opposite_logprob_tokens": [ + -3.8955986499786377 + ], + "logprob_reward": -0.020598655566573143, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598655566573143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.354143351316452, + 0.04812966287136078, + 0.09394795447587967, + 0.060535330325365067, + 0.07966286689043045, + 0.04939933493733406, + 0.16695581376552582, + 0.04994635283946991, + 0.0697864219546318, + 0.027492932975292206 + ], + "credit_map": [ + 0.01813584752380848, + 0.13240022957324982, + 0.030268091708421707, + 0.011332237161695957, + 0.014711497351527214, + 0.02256612479686737, + 0.008935472927987576, + 0.1157938539981842, + 0.0052025639452040195, + 0.007994563318789005, + 0.007170199416577816, + 0.00503749493509531, + 0.004748798441141844, + 0.004788522142916918, + 0.004914028104394674, + 0.008273495361208916, + 0.009755350649356842, + 0.006465346552431583, + 0.007742531597614288, + 0.0297567006200552, + 0.0053272792138159275, + 0.005131077952682972, + 0.016355738043785095, + 0.013413934037089348, + 0.004758614115417004, + 0.007392421830445528, + 0.010505862534046173, + 0.008203824050724506, + 0.00786492507904768, + 0.00569199351593852, + 0.009272634983062744, + 0.006845053751021624, + 0.009104486554861069, + 0.010434990748763084, + 0.007974006235599518, + 0.012024221941828728, + 0.012500817887485027, + 0.01191355288028717, + 0.009537971578538418, + 0.0061728209257125854, + 0.005479889921844006, + 0.003852901514619589, + 0.0019145094556733966, + 0.0023866714909672737, + 0.003891910891979933, + 0.004586337134242058, + 0.015857763588428497, + 0.011429349891841412, + 0.025147825479507446, + 0.04704708978533745, + 0.026920001953840256, + 0.008582720533013344, + 0.016058746725320816, + 0.015438436530530453, + 0.008800290524959564, + 0.01896071434020996, + 0.006310268305242062, + 0.013065611943602562, + 0.0073951780796051025, + 0.0046081929467618465, + 0.004582224879413843, + 0.004172646906226873, + 0.005110783968120813, + 0.004701446741819382, + 0.00927781406790018, + 0.01321390736848116, + 0.008559863083064556, + 0.004478808492422104, + 0.007578954100608826, + 0.007958333939313889, + 0.012449867092072964, + 0.00626888033002615, + 0.00438449252396822, + 0.004715105053037405, + 0.003642815398052335, + 0.0030529899522662163, + 0.004126921761780977, + 0.003885607933625579, + 0.0017826532712206244, + 0.001902348012663424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.020598655566573143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818695291876793 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818695291876793, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818695291876793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6221209764480591, + 0.05472395569086075, + 0.02732461504638195, + 0.017718028277158737, + 0.023680314421653748, + 0.036761291325092316, + 0.09552150219678879, + 0.03353177383542061, + 0.06978379935026169, + 0.018833696842193604 + ], + "credit_map": [ + 0.027670100331306458, + 0.23618648946285248, + 0.049443889409303665, + 0.014691651798784733, + 0.023339876905083656, + 0.03900732845067978, + 0.01286518108099699, + 0.21891649067401886, + 0.00515912938863039, + 0.008920684456825256, + 0.007793037686496973, + 0.004690730944275856, + 0.006019423715770245, + 0.007756895385682583, + 0.005195493809878826, + 0.009188567288219929, + 0.002066903281956911, + 0.003332379274070263, + 0.0028118896298110485, + 0.006289588753134012, + 0.002842553425580263, + 0.0024465431924909353, + 0.004462588578462601, + 0.0030721728689968586, + 0.0012760921381413937, + 0.0021908353082835674, + 0.002992936410009861, + 0.0023252288810908794, + 0.002007079776376486, + 0.002727511804550886, + 0.0022120520006865263, + 0.001986294286325574, + 0.002823194721713662, + 0.0032562073320150375, + 0.002628000220283866, + 0.0030098939314484596, + 0.0032351259142160416, + 0.0034568924456834793, + 0.003378604305908084, + 0.0018923989264294505, + 0.0035170840565115213, + 0.003707591909915209, + 0.0011994162341579795, + 0.0021245647221803665, + 0.00734372716397047, + 0.0032218422275036573, + 0.007445937022566795, + 0.008201133459806442, + 0.01455630362033844, + 0.02780306152999401, + 0.014424852095544338, + 0.004845845513045788, + 0.009406465105712414, + 0.009101543575525284, + 0.005012632813304663, + 0.01037080679088831, + 0.005139121785759926, + 0.011860128492116928, + 0.00645106378942728, + 0.0018502853345125914, + 0.0030000924598425627, + 0.0019924072548747063, + 0.0011718216119334102, + 0.002066858345642686, + 0.007423863280564547, + 0.013149570673704147, + 0.007679848000407219, + 0.005982269067317247, + 0.008749818429350853, + 0.009596292860805988, + 0.01146378181874752, + 0.005738360341638327, + 0.002354232594370842, + 0.0028844987973570824, + 0.001587956678122282, + 0.0017912917537614703, + 0.002882021712139249, + 0.0042950185015797615, + 0.0011110786581411958, + 0.0019275997765362263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.01818695291876793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278413355350494 + ], + "opposite_logprob_tokens": [ + -2.0177841186523438 + ], + "logprob_reward": -0.14278413355350494, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278413355350494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4090175926685333, + 0.03591236099600792, + 0.242228165268898, + 0.015932559967041016, + 0.030734451487660408, + 0.05431702360510826, + 0.06112746149301529, + 0.08365216106176376, + 0.0343957282602787, + 0.03268241509795189 + ], + "credit_map": [ + 0.02178126573562622, + 0.14968925714492798, + 0.03169989213347435, + 0.011511282995343208, + 0.015594336204230785, + 0.02924625761806965, + 0.010462955571711063, + 0.13903237879276276, + 0.004308030474931002, + 0.007544893771409988, + 0.0046842568553984165, + 0.004413604270666838, + 0.002376286778599024, + 0.00267808698117733, + 0.003664138726890087, + 0.00624306732788682, + 0.02372150868177414, + 0.016403602436184883, + 0.019574448466300964, + 0.078948475420475, + 0.010939981788396835, + 0.01213317271322012, + 0.044851452112197876, + 0.035655517131090164, + 0.00288226711563766, + 0.001552007277496159, + 0.0012770384782925248, + 0.0015116663416847587, + 0.0011333521688356996, + 0.002488616621121764, + 0.0032522170804440975, + 0.0018353963969275355, + 0.004217213019728661, + 0.0040077343583106995, + 0.0026612207293510437, + 0.0031960674095898867, + 0.003605669131502509, + 0.0060502770356833935, + 0.004285664297640324, + 0.0027106062043458223, + 0.005635593086481094, + 0.005765360314399004, + 0.0038373619318008423, + 0.003833089955151081, + 0.01567254401743412, + 0.005506598856300116, + 0.005724282935261726, + 0.008342196233570576, + 0.010562811978161335, + 0.01712782308459282, + 0.010750455781817436, + 0.0027759650256484747, + 0.005217552650719881, + 0.0051413546316325665, + 0.0035004988312721252, + 0.006051003932952881, + 0.013415569439530373, + 0.03257816657423973, + 0.01737464964389801, + 0.003343907417729497, + 0.007842665538191795, + 0.004001114517450333, + 0.002735541667789221, + 0.002360553713515401, + 0.009703133255243301, + 0.004878235049545765, + 0.00323368888348341, + 0.0030228905379772186, + 0.004593959543853998, + 0.003182929242029786, + 0.002522144466638565, + 0.0032587475143373013, + 0.008049686439335346, + 0.005353050772100687, + 0.002636027755215764, + 0.002427542582154274, + 0.007632165681570768, + 0.003112035570666194, + 0.0014430753653869033, + 0.002028832910582423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.14278413355350494 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dragon boat struggling against white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat's hull made of deep red polished wood with golden dragon scale carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a fiery lifelike dragon head at the prow with a gaping maw and bright red eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a slightly upward angle and steadily zoom out to reveal steep lush canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by slanting afternoon sunlight penetrating thin mist over turbulent water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195182360708714 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014195182360708714, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195182360708714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4108298122882843, + 0.040392592549324036, + 0.05337060987949371, + 0.032168734818696976, + 0.039701759815216064, + 0.1519511491060257, + 0.10720169544219971, + 0.0754457637667656, + 0.02687191031873226, + 0.06206599995493889 + ], + "credit_map": [ + 0.024228837341070175, + 0.14087699353694916, + 0.03909919410943985, + 0.018761994317173958, + 0.019128374755382538, + 0.026152366772294044, + 0.01202521100640297, + 0.13055682182312012, + 0.005217501427978277, + 0.007092483341693878, + 0.005639141891151667, + 0.003925731405615807, + 0.003710801713168621, + 0.004304111935198307, + 0.0035761771723628044, + 0.0069266450591385365, + 0.007849705405533314, + 0.00427599623799324, + 0.006198818329721689, + 0.01014934666454792, + 0.003464177716523409, + 0.00474737212061882, + 0.009416130371391773, + 0.007269063498824835, + 0.003134411061182618, + 0.0035202880389988422, + 0.0035412481520324945, + 0.004852305166423321, + 0.0037243859842419624, + 0.0038200297858566046, + 0.005835026036947966, + 0.0037410398945212364, + 0.007061971351504326, + 0.004224862437695265, + 0.0038702073507010937, + 0.004172140266746283, + 0.004142502322793007, + 0.006399258971214294, + 0.006302717607468367, + 0.0035280981101095676, + 0.010699654929339886, + 0.00968606024980545, + 0.006503926124423742, + 0.008107262663543224, + 0.020820816978812218, + 0.010743803344666958, + 0.039511196315288544, + 0.045878440141677856, + 0.018282121047377586, + 0.03130083903670311, + 0.01685604825615883, + 0.005222851876169443, + 0.009691622108221054, + 0.009752005338668823, + 0.0057144686579704285, + 0.010381734929978848, + 0.01267470046877861, + 0.029816381633281708, + 0.014881497249007225, + 0.0031126104295253754, + 0.006901941727846861, + 0.0034909488167613745, + 0.002047398127615452, + 0.002520289970561862, + 0.0068047200329601765, + 0.004140985663980246, + 0.0026342852506786585, + 0.002016092650592327, + 0.00326661579310894, + 0.0023955870419740677, + 0.003145274007692933, + 0.0024683515075594187, + 0.010318909771740437, + 0.01265785563737154, + 0.006115700118243694, + 0.008253941312432289, + 0.010397223755717278, + 0.006179363001137972, + 0.0038864915259182453, + 0.004256516229361296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.014195182360708714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_2/index_tensor([688, 688])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_2/index_tensor([688, 688])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cfa5be10a4bb41d49997c15f6726a707b530f0fb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_2/index_tensor([688, 688])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a series of close-ups imbued with the exquisite aesthetics of the new wuxia genre, the scene first focuses on a pair of exquisite black boots, gently and almost soundlessly stepping upon a ground covered in withered yellow leaves. Then, in extreme close-up, the camera captures a startling transformation: from the spot where the boots land, countless thorny black vines, shimmering with a dark sheen, erupt from the ground. Like living venomous snakes, they slither across the surface at a visible speed. A panning close-up follows the vines as they swiftly and gracefully cling to and entwine around the trunk of a nearby snow-white birch tree, their sharp barbs digging deep into the smooth bark. The stark contrast of black and white is striking in the cool, softly diffused light. The camera then slowly tilts upward, tracing the figure's flowing, inky silk robe, the sleeves of which flutter silently with his unhurried steps, finally framing his sharply defined face, a stern yet handsome expression. His eyes were deep and his expression was indifferent, as if the scene full of destructive vitality evoked by his footsteps was just an ordinary landscape. The whole scene, in an extreme interweaving of stillness and movement, beauty and danger, created an atmosphere of mystery, coldness and deadly elegance.", + "seed": 28838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_2/index_688_seed_28838_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01396921994164586, + "vlm_logprob": -0.01396921994164586, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.01396921994164586, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01628829725086689 + ], + "opposite_logprob_tokens": [ + -4.8912882804870605 + ], + "logprob_reward": -0.01628829725086689, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01628829725086689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20923911035060883, + 0.08090798556804657, + 0.15677477419376373, + 0.07121915370225906, + 0.05283486098051071, + 0.041467949748039246, + 0.07946839183568954, + 0.05752184987068176, + 0.12019158154726028, + 0.13037431240081787 + ], + "credit_map": [ + 0.027719246223568916, + 0.04719366133213043, + 0.019131271168589592, + 0.02047862857580185, + 0.03991109877824783, + 0.018992098048329353, + 0.02117461897432804, + 0.014638504013419151, + 0.010564816184341908, + 0.013937068171799183, + 0.01011055801063776, + 0.006423755548894405, + 0.013751181773841381, + 0.010547179728746414, + 0.007732936646789312, + 0.00784048717468977, + 0.014671256765723228, + 0.026222746819257736, + 0.01717175357043743, + 0.018861303105950356, + 0.02945641241967678, + 0.0138035723939538, + 0.02282695472240448, + 0.013760773465037346, + 0.005342363845556974, + 0.0061666094698011875, + 0.0074096559546887875, + 0.0072929514572024345, + 0.016630977392196655, + 0.005515970289707184, + 0.015637369826436043, + 0.007223263382911682, + 0.008029094897210598, + 0.004133663605898619, + 0.005588233470916748, + 0.0069544268772006035, + 0.007615601643919945, + 0.007901771925389767, + 0.007341329008340836, + 0.005270740482956171, + 0.0035964876879006624, + 0.003662893082946539, + 0.0041601997800171375, + 0.002943950705230236, + 0.0068814922124147415, + 0.004303299356251955, + 0.007597551681101322, + 0.008322075940668583, + 0.008789991028606892, + 0.012127491645514965, + 0.007527491077780724, + 0.008028678596019745, + 0.013673297129571438, + 0.006155284587293863, + 0.013568210415542126, + 0.009597948752343655, + 0.01288271602243185, + 0.004390173591673374, + 0.009352168068289757, + 0.007660535629838705, + 0.005403660703450441, + 0.006939691491425037, + 0.005055326968431473, + 0.005837578326463699, + 0.01953108422458172, + 0.01652400568127632, + 0.014813393354415894, + 0.015438586473464966, + 0.018276717513799667, + 0.010399557650089264, + 0.012717848643660545, + 0.012490383349359035, + 0.01522582396864891, + 0.01093866117298603, + 0.01596689596772194, + 0.014115015044808388, + 0.02106742188334465, + 0.020922984927892685, + 0.013416298665106297, + 0.01872120425105095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.01628829725086689 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865777675062418 + ], + "opposite_logprob_tokens": [ + -4.758657932281494 + ], + "logprob_reward": -0.00865777675062418, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865777675062418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16543686389923096, + 0.046810541301965714, + 0.2186783105134964, + 0.09899182617664337, + 0.21329662203788757, + 0.03684766963124275, + 0.027417512610554695, + 0.04540451616048813, + 0.04899350926280022, + 0.09812268614768982 + ], + "credit_map": [ + 0.017225267365574837, + 0.034476786851882935, + 0.013688596896827221, + 0.029305385425686836, + 0.027705714106559753, + 0.017272476106882095, + 0.013469191268086433, + 0.012293408624827862, + 0.007820293307304382, + 0.007847162894904613, + 0.005379420705139637, + 0.003591436194255948, + 0.006577988155186176, + 0.006172747351229191, + 0.004074115306138992, + 0.005347369238734245, + 0.019327973946928978, + 0.040074173361063004, + 0.020333148539066315, + 0.02484860271215439, + 0.030846543610095978, + 0.029121017083525658, + 0.0304715558886528, + 0.023655273020267487, + 0.006705146748572588, + 0.01139548048377037, + 0.007614580448716879, + 0.013083170168101788, + 0.021025603637099266, + 0.009612860158085823, + 0.01764960214495659, + 0.011905374005436897, + 0.024373851716518402, + 0.01928764395415783, + 0.024657733738422394, + 0.030377455055713654, + 0.03206076845526695, + 0.03228720277547836, + 0.03270397335290909, + 0.017547961324453354, + 0.004346056841313839, + 0.002673645969480276, + 0.005923573859035969, + 0.002464276272803545, + 0.006368989124894142, + 0.003441769164055586, + 0.006043761502951384, + 0.0055855875834822655, + 0.004030428361147642, + 0.0047364914789795876, + 0.0024183380883187056, + 0.0039537446573376656, + 0.004005809314548969, + 0.0017143009463325143, + 0.004368654452264309, + 0.0021897414699196815, + 0.007972867228090763, + 0.004408383741974831, + 0.007142504211515188, + 0.004722205456346273, + 0.00548042356967926, + 0.004883696790784597, + 0.006609573028981686, + 0.004184854216873646, + 0.006555099040269852, + 0.006416096352040768, + 0.0055238474160432816, + 0.007383918855339289, + 0.006322664674371481, + 0.005192968063056469, + 0.006626891903579235, + 0.004972013179212809, + 0.013167951256036758, + 0.011240726336836815, + 0.011523941531777382, + 0.014147674664855003, + 0.01273188367486, + 0.011831944808363914, + 0.011218812316656113, + 0.01225973665714264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.00865777675062418 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195776604115963 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.003195776604115963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195776604115963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17885585129261017, + 0.056107260286808014, + 0.09442932158708572, + 0.06890644133090973, + 0.17444027960300446, + 0.048557281494140625, + 0.055277206003665924, + 0.046852633357048035, + 0.13167694211006165, + 0.1448967456817627 + ], + "credit_map": [ + 0.020499462261795998, + 0.035356249660253525, + 0.03963762894272804, + 0.024255163967609406, + 0.01793999783694744, + 0.015678558498620987, + 0.010387318208813667, + 0.015101484954357147, + 0.009280824102461338, + 0.009327697567641735, + 0.005897252820432186, + 0.005756554193794727, + 0.008039139211177826, + 0.005505346227437258, + 0.007070233579725027, + 0.005230216309428215, + 0.0076533290557563305, + 0.015957992523908615, + 0.009823768399655819, + 0.008986311964690685, + 0.015144589357078075, + 0.015670830383896828, + 0.013167030178010464, + 0.00802548322826624, + 0.006198781076818705, + 0.009044046513736248, + 0.006615079008042812, + 0.007897489704191685, + 0.013777305372059345, + 0.005735570099204779, + 0.012902351096272469, + 0.006735815666615963, + 0.03158203139901161, + 0.014838684350252151, + 0.02636168897151947, + 0.024227550253272057, + 0.02299574203789234, + 0.02035980299115181, + 0.021206868812441826, + 0.012867916375398636, + 0.007469667587429285, + 0.005413114558905363, + 0.00850515253841877, + 0.004815569147467613, + 0.005740008316934109, + 0.004165591672062874, + 0.00657302513718605, + 0.005875155329704285, + 0.009690479375422001, + 0.009302625432610512, + 0.005585812963545322, + 0.0047481246292591095, + 0.0071814036928117275, + 0.0049572777934372425, + 0.007341512944549322, + 0.0064699687063694, + 0.007900864817202091, + 0.004726766142994165, + 0.007384380325675011, + 0.004956526216119528, + 0.005711470264941454, + 0.003610653802752495, + 0.007817675359547138, + 0.004744300618767738, + 0.02340373955667019, + 0.015792809426784515, + 0.01689922623336315, + 0.013711257837712765, + 0.012744843028485775, + 0.017231574282050133, + 0.01925819367170334, + 0.012635321356356144, + 0.01897537149488926, + 0.01952177658677101, + 0.021259209141135216, + 0.024665022268891335, + 0.018809638917446136, + 0.011207184754312038, + 0.0163281112909317, + 0.014130438677966595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.003195776604115963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808743134140968 + ], + "opposite_logprob_tokens": [ + -3.2880873680114746 + ], + "logprob_reward": -0.03808743134140968, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808743134140968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1868443787097931, + 0.10107750445604324, + 0.13264678418636322, + 0.08991511166095734, + 0.0753510594367981, + 0.03488955274224281, + 0.04926600679755211, + 0.027997590601444244, + 0.15660011768341064, + 0.1454118937253952 + ], + "credit_map": [ + 0.013377471826970577, + 0.02239394001662731, + 0.03300947695970535, + 0.03709977865219116, + 0.03181508928537369, + 0.015175753273069859, + 0.018795838579535484, + 0.015177029184997082, + 0.01780480332672596, + 0.014461911283433437, + 0.011156746186316013, + 0.007382792886346579, + 0.013027826324105263, + 0.013942818157374859, + 0.006979031953960657, + 0.0163215771317482, + 0.009823457337915897, + 0.023387616500258446, + 0.013297419995069504, + 0.01664692349731922, + 0.023348521441221237, + 0.012954901903867722, + 0.020915934816002846, + 0.012272011488676071, + 0.008811568841338158, + 0.01046357024461031, + 0.010041343979537487, + 0.00825398787856102, + 0.019593743607401848, + 0.007587258704006672, + 0.016387557610869408, + 0.008776087313890457, + 0.015034915879368782, + 0.00581629341468215, + 0.012481867335736752, + 0.010105928406119347, + 0.009532701224088669, + 0.009388982318341732, + 0.00816885195672512, + 0.00482151797041297, + 0.005217127967625856, + 0.0030502979643642902, + 0.005638084840029478, + 0.003339092480018735, + 0.00479274382814765, + 0.003243762068450451, + 0.005612446926534176, + 0.003995995502918959, + 0.008876316249370575, + 0.006269500590860844, + 0.004866885486990213, + 0.006373640149831772, + 0.005758092738687992, + 0.004376167897135019, + 0.006043309345841408, + 0.006702094804495573, + 0.003806023858487606, + 0.002600301057100296, + 0.0035647882614284754, + 0.004250561818480492, + 0.0029127737507224083, + 0.0034771638456732035, + 0.0043093059211969376, + 0.003076673485338688, + 0.020365985110402107, + 0.034296344965696335, + 0.01743367686867714, + 0.02013743855059147, + 0.016175471246242523, + 0.013279018923640251, + 0.018789740279316902, + 0.01612245664000511, + 0.019716767594218254, + 0.011672540567815304, + 0.02063433639705181, + 0.01438861433416605, + 0.02203349769115448, + 0.021635310724377632, + 0.011622142046689987, + 0.023708688095211983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03808743134140968 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036168177612125874 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036168177612125874, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036168177612125874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12710236012935638, + 0.0537133552134037, + 0.1908015012741089, + 0.06525001674890518, + 0.13086511194705963, + 0.05888514593243599, + 0.09973747283220291, + 0.08500754088163376, + 0.10849885642528534, + 0.08013866096735 + ], + "credit_map": [ + 0.010632332414388657, + 0.0167587548494339, + 0.02036009542644024, + 0.023073507472872734, + 0.023586120456457138, + 0.008461794815957546, + 0.014368741773068905, + 0.009861006401479244, + 0.007356560323387384, + 0.009262396953999996, + 0.004958820063620806, + 0.0036047089379280806, + 0.0077956560999155045, + 0.006358625367283821, + 0.0063964384607970715, + 0.007980148307979107, + 0.015630248934030533, + 0.045483145862817764, + 0.016950620338320732, + 0.014270453713834286, + 0.029583944007754326, + 0.03002792038023472, + 0.02255771867930889, + 0.016297435387969017, + 0.004692129325121641, + 0.008034467697143555, + 0.005227779503911734, + 0.009346597827970982, + 0.012666851282119751, + 0.005386631935834885, + 0.012905699200928211, + 0.006989854853600264, + 0.026477763429284096, + 0.009736771695315838, + 0.01918169856071472, + 0.01530430931597948, + 0.017252588644623756, + 0.01798444427549839, + 0.013651281595230103, + 0.011276237666606903, + 0.004903506953269243, + 0.0048533109948039055, + 0.00745176849886775, + 0.00583131518214941, + 0.009111786261200905, + 0.010044269263744354, + 0.00746841449290514, + 0.009220768697559834, + 0.011724449694156647, + 0.010656880214810371, + 0.008272728882730007, + 0.015623114071786404, + 0.018911516293883324, + 0.006983076222240925, + 0.017324944958090782, + 0.010240752249956131, + 0.010322988033294678, + 0.009838635101914406, + 0.010281738825142384, + 0.010955840349197388, + 0.009802683256566525, + 0.010324697941541672, + 0.012532312422990799, + 0.010948634706437588, + 0.017383305355906487, + 0.028902331367135048, + 0.013685435056686401, + 0.014128070324659348, + 0.009709373116493225, + 0.007720349822193384, + 0.009722774848341942, + 0.007247201167047024, + 0.00868732575327158, + 0.005857400596141815, + 0.0110582634806633, + 0.007977922447025776, + 0.014985473826527596, + 0.010357934050261974, + 0.010589133948087692, + 0.0106252022087574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.0036168177612125874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a series of close-ups imbued with the exquisite aesthetics of the new wuxia genre, the scene first focuses on a pair of exquisite black boots, gently and almost soundlessly stepping upon a ground covered in withered yellow leaves. Then, in extreme close-up, the camera captures a startling transformation: from the spot where the boots land, countless thorny black vines, shimmering with a dark sheen, erupt from the ground. Like living venomous snakes, they slither across the surface at a visible speed. A panning close-up follows the vines as they swiftly and gracefully cling to and entwine around the trunk of a nearby snow-white birch tree, their sharp barbs digging deep into the smooth bark. The stark contrast of black and white is striking in the cool, softly diffused light. The camera then slowly tilts upward, tracing the figure's flowing, inky silk robe, the sleeves of which flutter silently with his unhurried steps, finally framing his sharply defined face, a stern yet handsome expression. His eyes were deep and his expression was indifferent, as if the scene full of destructive vitality evoked by his footsteps was just an ordinary landscape. The whole scene, in an extreme interweaving of stillness and movement, beauty and danger, created an atmosphere of mystery, coldness and deadly elegance.", + "seed": 28839, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_2/index_688_seed_28839_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.00785201983526349, + "vlm_logprob": -0.00785201983526349, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.00785201983526349, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013297563418745995 + ], + "opposite_logprob_tokens": [ + -5.3882975578308105 + ], + "logprob_reward": -0.013297563418745995, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.013297563418745995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19613146781921387, + 0.10773132741451263, + 0.09224117547273636, + 0.06862857192754745, + 0.05578533187508583, + 0.06698934733867645, + 0.19679789245128632, + 0.03616728633642197, + 0.09516118466854095, + 0.08436640352010727 + ], + "credit_map": [ + 0.021453801542520523, + 0.030248746275901794, + 0.017728663980960846, + 0.0330205000936985, + 0.032899290323257446, + 0.017430542036890984, + 0.024716539308428764, + 0.018633374944329262, + 0.012967357411980629, + 0.02004777267575264, + 0.013490869663655758, + 0.017519855871796608, + 0.011278471909463406, + 0.011584443971514702, + 0.011770475655794144, + 0.009072077460587025, + 0.008435909636318684, + 0.010998967103660107, + 0.009953814558684826, + 0.011692695319652557, + 0.011468684300780296, + 0.008766072802245617, + 0.02074873261153698, + 0.010176298208534718, + 0.006540092173963785, + 0.014860407449305058, + 0.01811378076672554, + 0.008915029466152191, + 0.005353282205760479, + 0.005184631794691086, + 0.005094529595226049, + 0.004566822666674852, + 0.008102323859930038, + 0.01046278141438961, + 0.008915919810533524, + 0.008985319174826145, + 0.003680664347484708, + 0.004103977233171463, + 0.0040181418880820274, + 0.007516202051192522, + 0.0038770914543420076, + 0.003858230309560895, + 0.005819905083626509, + 0.003995021339505911, + 0.006600498221814632, + 0.008958953432738781, + 0.02102304808795452, + 0.012856601737439632, + 0.01231463998556137, + 0.0109592005610466, + 0.008036274462938309, + 0.05885776877403259, + 0.020356379449367523, + 0.006634043995290995, + 0.07355020940303802, + 0.006089388392865658, + 0.0027855841908603907, + 0.003460415406152606, + 0.0034781715366989374, + 0.003070476930588484, + 0.007354545872658491, + 0.005423319060355425, + 0.004472112748771906, + 0.0061226594261825085, + 0.008027180097997189, + 0.014246047474443913, + 0.008524657227098942, + 0.007735820021480322, + 0.017486706376075745, + 0.010460118763148785, + 0.013696115463972092, + 0.014984541572630405, + 0.007587882224470377, + 0.013878178782761097, + 0.00815570168197155, + 0.011250988580286503, + 0.013213272206485271, + 0.007192770950496197, + 0.014154784381389618, + 0.008932817727327347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.013297563418745995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00596374599263072 + ], + "opposite_logprob_tokens": [ + -5.1309638023376465 + ], + "logprob_reward": -0.00596374599263072, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00596374599263072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0531059131026268, + 0.11082902550697327, + 0.05010651797056198, + 0.48165297508239746, + 0.13354673981666565, + 0.038873784244060516, + 0.07605153322219849, + 0.017205681651830673, + 0.019719406962394714, + 0.018908469006419182 + ], + "credit_map": [ + 0.005152666009962559, + 0.009670250117778778, + 0.005204332526773214, + 0.00498975720256567, + 0.008409630507230759, + 0.0054818387143313885, + 0.008207990787923336, + 0.005989439785480499, + 0.008135601878166199, + 0.025622250512242317, + 0.00618708087131381, + 0.006640525534749031, + 0.009748151525855064, + 0.036461908370256424, + 0.005949568003416061, + 0.012083929032087326, + 0.004167142324149609, + 0.0048873587511479855, + 0.004693849477916956, + 0.006039909087121487, + 0.0077731674537062645, + 0.00558443320915103, + 0.010836707428097725, + 0.006123946979641914, + 0.0443846695125103, + 0.15348103642463684, + 0.15549473464488983, + 0.06259403377771378, + 0.021582022309303284, + 0.013914065435528755, + 0.010275584645569324, + 0.019926777109503746, + 0.011785287410020828, + 0.020184963941574097, + 0.020227713510394096, + 0.03561277687549591, + 0.014029281213879585, + 0.00921170599758625, + 0.012587620876729488, + 0.00990737322717905, + 0.004334249068051577, + 0.003316508373245597, + 0.008264380507171154, + 0.004053558222949505, + 0.0035446141846477985, + 0.0033819670788943768, + 0.006490217987447977, + 0.005488283466547728, + 0.00420237286016345, + 0.00577114662155509, + 0.006004033610224724, + 0.02351178601384163, + 0.007241858635097742, + 0.003667021868750453, + 0.022352520376443863, + 0.0033007850870490074, + 0.002023746259510517, + 0.002086973050609231, + 0.0016068639233708382, + 0.001736995647661388, + 0.002885566558688879, + 0.002330285729840398, + 0.0019837040454149246, + 0.002551544224843383, + 0.002081488026306033, + 0.003725663525983691, + 0.0014969325857236981, + 0.0017332076095044613, + 0.002330358838662505, + 0.0030270663555711508, + 0.002128083724528551, + 0.003196605248376727, + 0.001792341936379671, + 0.003970027901232243, + 0.0019738369155675173, + 0.0026938272640109062, + 0.0020342592615634203, + 0.0015331852482631803, + 0.0027595118153840303, + 0.0021514759864658117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.00596374599263072 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028228219598531723 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.0028228219598531723, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028228219598531723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034131523221731186, + 0.05277698487043381, + 0.035835105925798416, + 0.38233187794685364, + 0.06144026666879654, + 0.2247353196144104, + 0.1315249651670456, + 0.012450892478227615, + 0.045843400061130524, + 0.018929671496152878 + ], + "credit_map": [ + 0.0031019290909171104, + 0.006218841299414635, + 0.0038609125185757875, + 0.0039623151533305645, + 0.005755139980465174, + 0.0028293521609157324, + 0.005451721139252186, + 0.002951310947537422, + 0.004750325810164213, + 0.009505079127848148, + 0.005824498366564512, + 0.006391454953700304, + 0.004885134752839804, + 0.010832289233803749, + 0.0054535288363695145, + 0.0051346723921597, + 0.0025221635587513447, + 0.0037933087442070246, + 0.0034409749787300825, + 0.004728211555629969, + 0.005373775027692318, + 0.003791388124227524, + 0.007964135147631168, + 0.0042211455292999744, + 0.03500848636031151, + 0.12141279876232147, + 0.12146907299757004, + 0.05025707557797432, + 0.017052259296178818, + 0.012145035900175571, + 0.0090560894459486, + 0.01593106985092163, + 0.005449383519589901, + 0.008505276404321194, + 0.00828394666314125, + 0.015679266303777695, + 0.007818765938282013, + 0.004347800277173519, + 0.00759106082841754, + 0.0037647648714482784, + 0.011952468194067478, + 0.010339146479964256, + 0.028807180002331734, + 0.010082502849400043, + 0.027853550389409065, + 0.011385978199541569, + 0.08756241947412491, + 0.036752063781023026, + 0.007055511232465506, + 0.009737074375152588, + 0.008985833264887333, + 0.04158961772918701, + 0.012410574592649937, + 0.006992371287196875, + 0.04034123569726944, + 0.004412751644849777, + 0.0013141407398507, + 0.0014393882593140006, + 0.0007619920652359724, + 0.0009670690051279962, + 0.0012462841114029288, + 0.0026612465735524893, + 0.001439607352949679, + 0.0026211647782474756, + 0.003933857195079327, + 0.006064001005142927, + 0.0034526039380580187, + 0.0033985143527388573, + 0.008180996403098106, + 0.0072533623315393925, + 0.004749336279928684, + 0.008810729719698429, + 0.0016917523462325335, + 0.00400453619658947, + 0.0016788113862276077, + 0.002887614304199815, + 0.0021030970383435488, + 0.0016125441761687398, + 0.0027619164902716875, + 0.0021894003730267286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.0028228219598531723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253888476639986 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.01253888476639986, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253888476639986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051939330995082855, + 0.04763294383883476, + 0.03412845358252525, + 0.6540522575378418, + 0.0398300476372242, + 0.01711832545697689, + 0.020115479826927185, + 0.01469800528138876, + 0.0678492933511734, + 0.0526358038187027 + ], + "credit_map": [ + 0.005148991011083126, + 0.010477336123585701, + 0.004373888485133648, + 0.004217564128339291, + 0.009407388977706432, + 0.004931623116135597, + 0.008550690487027168, + 0.004831853322684765, + 0.004226237535476685, + 0.008728313259780407, + 0.004901798442006111, + 0.007232820149511099, + 0.00680142268538475, + 0.006599849555641413, + 0.004882126115262508, + 0.004260379821062088, + 0.0043298183009028435, + 0.005570774897933006, + 0.004583969246596098, + 0.005464452318847179, + 0.004197105765342712, + 0.003056697780266404, + 0.004009449388831854, + 0.002916188444942236, + 0.06054680421948433, + 0.20887869596481323, + 0.20966161787509918, + 0.0859951302409172, + 0.029409080743789673, + 0.019123749807476997, + 0.013916922733187675, + 0.026520315557718277, + 0.003062644973397255, + 0.004459021613001823, + 0.005485559813678265, + 0.007398323621600866, + 0.0039762877859175205, + 0.006929886527359486, + 0.0033547701314091682, + 0.005163558758795261, + 0.0017537602689117193, + 0.0015787777956575155, + 0.0033597717992961407, + 0.002050230046734214, + 0.0017946390435099602, + 0.001975712366402149, + 0.001577715273015201, + 0.003027719911187887, + 0.0018285596743226051, + 0.0023514418862760067, + 0.0026868090499192476, + 0.003677224740386009, + 0.002396082505583763, + 0.0025292178615927696, + 0.0025128601118922234, + 0.0021332844626158476, + 0.00265943119302392, + 0.0020724721252918243, + 0.001267824205569923, + 0.0016951676225289702, + 0.0018220177153125405, + 0.0017928280867636204, + 0.0016956751933321357, + 0.001692590187303722, + 0.007151986472308636, + 0.01188657060265541, + 0.005992016289383173, + 0.005866819992661476, + 0.009334693662822247, + 0.00803783256560564, + 0.0062242005951702595, + 0.013355174101889133, + 0.004563504364341497, + 0.010364934802055359, + 0.004251937847584486, + 0.007672331761568785, + 0.005860564764589071, + 0.004368525929749012, + 0.00981894414871931, + 0.005735065788030624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.01253888476639986 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637083038687706 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.004637083038687706, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637083038687706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07047836482524872, + 0.06451713293790817, + 0.03987093269824982, + 0.24613696336746216, + 0.04590034857392311, + 0.2518385946750641, + 0.054765086621046066, + 0.032895222306251526, + 0.14402827620506287, + 0.049569085240364075 + ], + "credit_map": [ + 0.007915272377431393, + 0.008753655478358269, + 0.007840658538043499, + 0.006853662896901369, + 0.012516353279352188, + 0.008505381643772125, + 0.011581211350858212, + 0.006512179039418697, + 0.00697912834584713, + 0.01309274323284626, + 0.008349164389073849, + 0.008054697886109352, + 0.007304863538593054, + 0.008326636627316475, + 0.006611696910113096, + 0.005798201076686382, + 0.004384118597954512, + 0.004118709824979305, + 0.004501958377659321, + 0.0059510706923902035, + 0.006362121552228928, + 0.004695798270404339, + 0.0050924248062074184, + 0.004764735233038664, + 0.022709893062710762, + 0.07328381389379501, + 0.07599558681249619, + 0.029716329649090767, + 0.012052207253873348, + 0.009357098489999771, + 0.009706762619316578, + 0.013315284624695778, + 0.004780014976859093, + 0.006542200688272715, + 0.006935188081115484, + 0.0074929846450686455, + 0.0038927269633859396, + 0.005981918890029192, + 0.00342831970192492, + 0.006846997886896133, + 0.01886771060526371, + 0.013739983551204205, + 0.025562459602952003, + 0.013667207211256027, + 0.019313519820570946, + 0.018007777631282806, + 0.07334396988153458, + 0.06933598220348358, + 0.005308741237968206, + 0.005679829511791468, + 0.008858921006321907, + 0.006252114661037922, + 0.007865950465202332, + 0.005006054881960154, + 0.012446719221770763, + 0.003346758894622326, + 0.0036602248437702656, + 0.003158555133268237, + 0.002556584542617202, + 0.00290053803473711, + 0.005259582307189703, + 0.005541195161640644, + 0.004654132295399904, + 0.005164413712918758, + 0.013482416048645973, + 0.018619354814291, + 0.012266187928617, + 0.010818980634212494, + 0.024761538952589035, + 0.021469606086611748, + 0.014998751692473888, + 0.027611443772912025, + 0.003968249075114727, + 0.010585512965917587, + 0.004501590505242348, + 0.006440593395382166, + 0.005489415489137173, + 0.004490356892347336, + 0.008292606100440025, + 0.005800765939056873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.004637083038687706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a series of close-ups imbued with the exquisite aesthetics of the new wuxia genre, the scene first focuses on a pair of exquisite black boots, gently and almost soundlessly stepping upon a ground covered in withered yellow leaves. Then, in extreme close-up, the camera captures a startling transformation: from the spot where the boots land, countless thorny black vines, shimmering with a dark sheen, erupt from the ground. Like living venomous snakes, they slither across the surface at a visible speed. A panning close-up follows the vines as they swiftly and gracefully cling to and entwine around the trunk of a nearby snow-white birch tree, their sharp barbs digging deep into the smooth bark. The stark contrast of black and white is striking in the cool, softly diffused light. The camera then slowly tilts upward, tracing the figure's flowing, inky silk robe, the sleeves of which flutter silently with his unhurried steps, finally framing his sharply defined face, a stern yet handsome expression. His eyes were deep and his expression was indifferent, as if the scene full of destructive vitality evoked by his footsteps was just an ordinary landscape. The whole scene, in an extreme interweaving of stillness and movement, beauty and danger, created an atmosphere of mystery, coldness and deadly elegance.", + "seed": 28840, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_2/index_688_seed_28840_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.00970629109069705, + "vlm_logprob": -0.00970629109069705, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.00970629109069705, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017823178321123123 + ], + "opposite_logprob_tokens": [ + -4.392823219299316 + ], + "logprob_reward": -0.017823178321123123, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017823178321123123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21168354153633118, + 0.043096110224723816, + 0.21936354041099548, + 0.19717557728290558, + 0.042942147701978683, + 0.04184552654623985, + 0.05247180908918381, + 0.10467367619276047, + 0.03204122185707092, + 0.05470672622323036 + ], + "credit_map": [ + 0.015208072029054165, + 0.02201494760811329, + 0.025502244010567665, + 0.023091483861207962, + 0.018598031252622604, + 0.014811690896749496, + 0.052511703222990036, + 0.03994540125131607, + 0.0042825667187571526, + 0.006997314747422934, + 0.006398194003850222, + 0.004672297276556492, + 0.0061454204842448235, + 0.00562320277094841, + 0.004285227041691542, + 0.004691894166171551, + 0.02377692610025406, + 0.031449634581804276, + 0.029026776552200317, + 0.026927262544631958, + 0.04018421098589897, + 0.02224220335483551, + 0.026735451072454453, + 0.01902109384536743, + 0.01511718612164259, + 0.035319793969392776, + 0.016254549846053123, + 0.01843120902776718, + 0.048560477793216705, + 0.036927178502082825, + 0.012460397556424141, + 0.014104812406003475, + 0.00485316151753068, + 0.008672798052430153, + 0.004975452087819576, + 0.010232304222881794, + 0.0035937740467488766, + 0.003078996669501066, + 0.003768413793295622, + 0.0037672542966902256, + 0.003615869442000985, + 0.0065217032097280025, + 0.005469180177897215, + 0.00445970892906189, + 0.0038789890240877867, + 0.009272556751966476, + 0.003532454604282975, + 0.005095068830996752, + 0.00623125908896327, + 0.015518859960138798, + 0.007141866721212864, + 0.006347960326820612, + 0.004167491104453802, + 0.002740301424637437, + 0.004767760168761015, + 0.005556314252316952, + 0.01229710690677166, + 0.018895955756306648, + 0.02112562395632267, + 0.014826798811554909, + 0.013430530205368996, + 0.008076343685388565, + 0.007347831502556801, + 0.008673486299812794, + 0.0032974767964333296, + 0.0036837239749729633, + 0.0043352036736905575, + 0.004141705110669136, + 0.006338178645819426, + 0.0029696316923946142, + 0.004730689339339733, + 0.00254461751319468, + 0.0037618568167090416, + 0.008393450640141964, + 0.0053581600077450275, + 0.009519310668110847, + 0.012192745693027973, + 0.004123201128095388, + 0.007952644489705563, + 0.003405363764613867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.017823178321123123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036281016655266285 + ], + "opposite_logprob_tokens": [ + -5.628628253936768 + ], + "logprob_reward": -0.0036281016655266285, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036281016655266285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09910211712121964, + 0.05775284022092819, + 0.08308325707912445, + 0.31002870202064514, + 0.051346130669116974, + 0.0893092229962349, + 0.08422873169183731, + 0.12188006937503815, + 0.053062744438648224, + 0.05020616576075554 + ], + "credit_map": [ + 0.011063058860599995, + 0.014851221814751625, + 0.02164868265390396, + 0.012042531743645668, + 0.008784143254160881, + 0.00928928330540657, + 0.009821276180446148, + 0.011601925827562809, + 0.004967495799064636, + 0.007665520999580622, + 0.012661846354603767, + 0.006591413170099258, + 0.006854828912764788, + 0.00627093156799674, + 0.007180582731962204, + 0.005560219753533602, + 0.008506614714860916, + 0.00902189314365387, + 0.009687836281955242, + 0.008129180409014225, + 0.010774473659694195, + 0.009214196354150772, + 0.012544164434075356, + 0.015204902738332748, + 0.03135228529572487, + 0.05390145629644394, + 0.028733661398291588, + 0.027448931708931923, + 0.05873497948050499, + 0.06610865890979767, + 0.019733868539333344, + 0.024014845490455627, + 0.008359847590327263, + 0.007443341426551342, + 0.007093147374689579, + 0.009547905065119267, + 0.005045527126640081, + 0.0047323559410870075, + 0.0035246761981397867, + 0.005599331110715866, + 0.00875296350568533, + 0.012825775891542435, + 0.014064745977520943, + 0.01111585833132267, + 0.006747322157025337, + 0.014823561534285545, + 0.008501673117280006, + 0.012477317824959755, + 0.008679943159222603, + 0.014121495187282562, + 0.008622391149401665, + 0.009954683482646942, + 0.01080416515469551, + 0.0055727409198880196, + 0.011891737580299377, + 0.014581572264432907, + 0.014214550144970417, + 0.020723940804600716, + 0.01881043054163456, + 0.016155226156115532, + 0.01727263629436493, + 0.01170334778726101, + 0.010575748048722744, + 0.012424192391335964, + 0.0045768641866743565, + 0.005481124389916658, + 0.007746150717139244, + 0.0069164554588496685, + 0.010492632165551186, + 0.004162794444710016, + 0.00935718696564436, + 0.00432953517884016, + 0.005073971580713987, + 0.0074462429620325565, + 0.005199109669774771, + 0.0051403106190264225, + 0.013048473745584488, + 0.00517130084335804, + 0.005681680049747229, + 0.00344507466070354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.0036281016655266285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634272798895836 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.007634272798895836, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634272798895836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15724731981754303, + 0.07366416603326797, + 0.04902297258377075, + 0.11403584480285645, + 0.09036771953105927, + 0.0775776356458664, + 0.0963682159781456, + 0.16112907230854034, + 0.10942163318395615, + 0.07116536796092987 + ], + "credit_map": [ + 0.014376322738826275, + 0.023160947486758232, + 0.03485807031393051, + 0.017589204013347626, + 0.013984974473714828, + 0.015455527231097221, + 0.021530164405703545, + 0.016292128711938858, + 0.007493991404771805, + 0.0068786246702075005, + 0.01521784532815218, + 0.010136143304407597, + 0.009684418328106403, + 0.008221135474741459, + 0.008733573369681835, + 0.007298443000763655, + 0.005344946403056383, + 0.0055092801339924335, + 0.010212436318397522, + 0.005814095959067345, + 0.006759195122867823, + 0.00419788109138608, + 0.005858898162841797, + 0.005326238926500082, + 0.008791300468146801, + 0.019753459841012955, + 0.008411127142608166, + 0.010370294563472271, + 0.027379386126995087, + 0.023159628733992577, + 0.0077062188647687435, + 0.008464446291327477, + 0.009423697367310524, + 0.014363497495651245, + 0.009707550518214703, + 0.016513777896761894, + 0.007998094893991947, + 0.011107495054602623, + 0.007565538864582777, + 0.01368808001279831, + 0.006698491517454386, + 0.013495278544723988, + 0.010672054253518581, + 0.009684049524366856, + 0.007898280397057533, + 0.009471134282648563, + 0.01256263442337513, + 0.007095715496689081, + 0.015008026733994484, + 0.02834276668727398, + 0.012655306607484818, + 0.009842832572758198, + 0.006183757912367582, + 0.006415058393031359, + 0.006979992147535086, + 0.010940482839941978, + 0.017478695139288902, + 0.0205623060464859, + 0.03291753679513931, + 0.021174263209104538, + 0.02377316914498806, + 0.01663866825401783, + 0.013897595927119255, + 0.014686844311654568, + 0.01483080256730318, + 0.016028184443712234, + 0.013460732065141201, + 0.019793037325143814, + 0.013618613593280315, + 0.010532696731388569, + 0.013355473056435585, + 0.0078020961955189705, + 0.005195964127779007, + 0.011493882164359093, + 0.00943113211542368, + 0.01220631692558527, + 0.016463080421090126, + 0.0064396606758236885, + 0.006573752500116825, + 0.003361582290381193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.007634272798895836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194242656230927 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014194242656230927, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194242656230927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11434473842382431, + 0.05526001378893852, + 0.20102714002132416, + 0.13625869154930115, + 0.05627235025167465, + 0.0604499951004982, + 0.05011162534356117, + 0.09708990901708603, + 0.09662707149982452, + 0.13255849480628967 + ], + "credit_map": [ + 0.013257022947072983, + 0.01670536771416664, + 0.026248633861541748, + 0.011569799855351448, + 0.0101326834410429, + 0.012233934365212917, + 0.012478701770305634, + 0.011718587018549442, + 0.005347858183085918, + 0.007295792922377586, + 0.007184581831097603, + 0.00743393087759614, + 0.00770587520673871, + 0.006813517305999994, + 0.006064634304493666, + 0.007413816638290882, + 0.020580146461725235, + 0.022511133924126625, + 0.023632889613509178, + 0.01740874908864498, + 0.028498444706201553, + 0.021306591108441353, + 0.036117009818553925, + 0.030972154811024666, + 0.018748285248875618, + 0.015482516027987003, + 0.02168911322951317, + 0.01635168306529522, + 0.021543683484196663, + 0.02182958647608757, + 0.009430143982172012, + 0.011183674447238445, + 0.008518348447978497, + 0.008694984018802643, + 0.006366562098264694, + 0.010489287786185741, + 0.006033265497535467, + 0.004490980878472328, + 0.004720848985016346, + 0.006958064157515764, + 0.006268367171287537, + 0.007275375071913004, + 0.008079944178462029, + 0.006500563118606806, + 0.007811184972524643, + 0.010028237476944923, + 0.007586903870105743, + 0.006899410393089056, + 0.00710864644497633, + 0.014109606854617596, + 0.006579938810318708, + 0.00465687969699502, + 0.0040689813904464245, + 0.003304490353912115, + 0.003906920552253723, + 0.006376155652105808, + 0.013927916064858437, + 0.01510909479111433, + 0.022850316017866135, + 0.013198561035096645, + 0.010476316325366497, + 0.007216558326035738, + 0.007133910898119211, + 0.00717721926048398, + 0.006361021194607019, + 0.008958798833191395, + 0.008116246201097965, + 0.011157178319990635, + 0.01765228621661663, + 0.017432991415262222, + 0.01529037207365036, + 0.011658165603876114, + 0.009086850099265575, + 0.020563744008541107, + 0.010824049822986126, + 0.01625424437224865, + 0.03470699489116669, + 0.012417500838637352, + 0.01848752796649933, + 0.0102175772190094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.014194242656230927 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251660011708736 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251660011708736, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251660011708736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09532025456428528, + 0.08190175890922546, + 0.0838058814406395, + 0.14569757878780365, + 0.15545415878295898, + 0.0663941502571106, + 0.038092996925115585, + 0.07853806018829346, + 0.11331362277269363, + 0.14148153364658356 + ], + "credit_map": [ + 0.012881883420050144, + 0.014326886273920536, + 0.018805820494890213, + 0.009944547899067402, + 0.008585171774029732, + 0.010074150748550892, + 0.009978817775845528, + 0.01072298176586628, + 0.006548314820975065, + 0.007192978635430336, + 0.018673744052648544, + 0.010856905952095985, + 0.008154644630849361, + 0.008796755224466324, + 0.007518221624195576, + 0.014160191640257835, + 0.009253675118088722, + 0.009621317498385906, + 0.010772556997835636, + 0.008564871735870838, + 0.012161116115748882, + 0.007711098995059729, + 0.013972918502986431, + 0.011748322285711765, + 0.01591852493584156, + 0.019361332058906555, + 0.02043212577700615, + 0.014880266040563583, + 0.025196734815835953, + 0.027491958811879158, + 0.009696817956864834, + 0.012719815596938133, + 0.039409976452589035, + 0.017657753080129623, + 0.02261137217283249, + 0.01530095562338829, + 0.019012294709682465, + 0.012649794109165668, + 0.01392217818647623, + 0.014889832586050034, + 0.0077446382492780685, + 0.00993911549448967, + 0.009154903702437878, + 0.007699117064476013, + 0.003925826866179705, + 0.012697177939116955, + 0.0056666238233447075, + 0.009566751308739185, + 0.005103574134409428, + 0.0046374318189918995, + 0.004028694238513708, + 0.004570511635392904, + 0.005384538322687149, + 0.0027785205747932196, + 0.006295397877693176, + 0.005294330883771181, + 0.01055870857089758, + 0.008806967176496983, + 0.013920797035098076, + 0.007676720153540373, + 0.011235839687287807, + 0.008336416445672512, + 0.010632405057549477, + 0.007370199076831341, + 0.010208473540842533, + 0.011899338103830814, + 0.01481239777058363, + 0.018289268016815186, + 0.015541629865765572, + 0.014054162427783012, + 0.01762172393500805, + 0.010886629112064838, + 0.011190536431968212, + 0.02449621632695198, + 0.015209836885333061, + 0.016337106004357338, + 0.03707709163427353, + 0.014091811142861843, + 0.015511450357735157, + 0.007567480206489563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.005251660011708736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a series of close-ups imbued with the exquisite aesthetics of the new wuxia genre, the scene first focuses on a pair of exquisite black boots, gently and almost soundlessly stepping upon a ground covered in withered yellow leaves. Then, in extreme close-up, the camera captures a startling transformation: from the spot where the boots land, countless thorny black vines, shimmering with a dark sheen, erupt from the ground. Like living venomous snakes, they slither across the surface at a visible speed. A panning close-up follows the vines as they swiftly and gracefully cling to and entwine around the trunk of a nearby snow-white birch tree, their sharp barbs digging deep into the smooth bark. The stark contrast of black and white is striking in the cool, softly diffused light. The camera then slowly tilts upward, tracing the figure's flowing, inky silk robe, the sleeves of which flutter silently with his unhurried steps, finally framing his sharply defined face, a stern yet handsome expression. His eyes were deep and his expression was indifferent, as if the scene full of destructive vitality evoked by his footsteps was just an ordinary landscape. The whole scene, in an extreme interweaving of stillness and movement, beauty and danger, created an atmosphere of mystery, coldness and deadly elegance.", + "seed": 28841, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_2/index_688_seed_28841_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.009245811030268669, + "vlm_logprob": -0.009245811030268669, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.009245811030268669, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020951267331838608 + ], + "opposite_logprob_tokens": [ + -4.270951271057129 + ], + "logprob_reward": -0.020951267331838608, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020951267331838608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19545842707157135, + 0.358701229095459, + 0.09957192093133926, + 0.11089305579662323, + 0.08741479367017746, + 0.03446955978870392, + 0.02656633034348488, + 0.020408576354384422, + 0.02549879066646099, + 0.041017305105924606 + ], + "credit_map": [ + 0.0221851859241724, + 0.027051329612731934, + 0.02069239318370819, + 0.023402459919452667, + 0.03128078207373619, + 0.014521922916173935, + 0.026874452829360962, + 0.02944990247488022, + 0.02701623924076557, + 0.05562994256615639, + 0.053280968219041824, + 0.04925183579325676, + 0.043196577578783035, + 0.045998238027095795, + 0.03921464458107948, + 0.04511279985308647, + 0.011321677826344967, + 0.01605503261089325, + 0.008379936218261719, + 0.010553795844316483, + 0.016756197437644005, + 0.01723724976181984, + 0.006811419501900673, + 0.012456615455448627, + 0.011666231788694859, + 0.01657191291451454, + 0.008088938891887665, + 0.01155332662165165, + 0.013019688427448273, + 0.01955358497798443, + 0.014209181070327759, + 0.016230188310146332, + 0.0034555825404822826, + 0.007439144887030125, + 0.004684407729655504, + 0.005420794244855642, + 0.007933584041893482, + 0.021030021831393242, + 0.027934584766626358, + 0.009516672231256962, + 0.001899995026178658, + 0.0028125157114118338, + 0.0015403396682813764, + 0.0064598252065479755, + 0.004932596813887358, + 0.005682236980646849, + 0.004964596591889858, + 0.006177456118166447, + 0.00238763727247715, + 0.004363596439361572, + 0.003276050789281726, + 0.0031702558044344187, + 0.003289526794105768, + 0.004300363827496767, + 0.002568154828622937, + 0.0032107429578900337, + 0.0022623126860708, + 0.002468437422066927, + 0.002095015486702323, + 0.003062753938138485, + 0.0025403816252946854, + 0.00284074735827744, + 0.002944807754829526, + 0.0021941193845123053, + 0.004583514295518398, + 0.00403343141078949, + 0.0025919179897755384, + 0.0035422486253082752, + 0.0034097773022949696, + 0.001793536590412259, + 0.002999443095177412, + 0.002544920891523361, + 0.0068257832899689674, + 0.007613909896463156, + 0.004992082715034485, + 0.004221717361360788, + 0.006286172196269035, + 0.0035360888577997684, + 0.0031858873553574085, + 0.004355662036687136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.020951267331838608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263162776827812 + ], + "opposite_logprob_tokens": [ + -5.255263328552246 + ], + "logprob_reward": -0.005263162776827812, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263162776827812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0821395143866539, + 0.18167366087436676, + 0.09037046134471893, + 0.08880506455898285, + 0.10734327137470245, + 0.2610726058483124, + 0.06865294277667999, + 0.03770747780799866, + 0.03976704925298691, + 0.04246792197227478 + ], + "credit_map": [ + 0.009240349754691124, + 0.010171797126531601, + 0.008449921384453773, + 0.010364452376961708, + 0.013789365068078041, + 0.008816580288112164, + 0.011958428658545017, + 0.009348620660603046, + 0.012412773445248604, + 0.029307818040251732, + 0.020418429747223854, + 0.02338823676109314, + 0.02481059730052948, + 0.024642176926136017, + 0.019122697412967682, + 0.027570929378271103, + 0.009789758361876011, + 0.014279000461101532, + 0.00801787432283163, + 0.011297364719212055, + 0.011426244862377644, + 0.017956489697098732, + 0.006573558785021305, + 0.011030170135200024, + 0.014128793962299824, + 0.009823252446949482, + 0.005864569451659918, + 0.009317216463387012, + 0.012083190493285656, + 0.01674814149737358, + 0.009238042868673801, + 0.011601858772337437, + 0.006893971469253302, + 0.015370799228549004, + 0.007361101917922497, + 0.011888476088643074, + 0.017550241202116013, + 0.017923440784215927, + 0.017655296251177788, + 0.01269993744790554, + 0.012326659634709358, + 0.017804831266403198, + 0.010455325245857239, + 0.016589771956205368, + 0.03949630260467529, + 0.07040221244096756, + 0.028350746259093285, + 0.06564673781394958, + 0.00727273803204298, + 0.006400859449058771, + 0.0077744401060044765, + 0.008469748310744762, + 0.009647269733250141, + 0.0106979263946414, + 0.009845219552516937, + 0.008544749580323696, + 0.004923083353787661, + 0.0037091707345098257, + 0.006150251254439354, + 0.0042959596030414104, + 0.003867141902446747, + 0.0043758065439760685, + 0.0041550262831151485, + 0.006231037899851799, + 0.010533079504966736, + 0.005781414452940226, + 0.005732730031013489, + 0.004066265653818846, + 0.0043366439640522, + 0.0028604110702872276, + 0.0029328500386327505, + 0.00352365686558187, + 0.0072234091348946095, + 0.007274003233760595, + 0.005474781151860952, + 0.0035424295347183943, + 0.006005926989018917, + 0.004729479551315308, + 0.004035524558275938, + 0.004182365257292986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.005263162776827812 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738912779837847 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006738912779837847, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738912779837847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11239323765039444, + 0.15982678532600403, + 0.10954467952251434, + 0.07969442009925842, + 0.0755038633942604, + 0.06407476216554642, + 0.04771437123417854, + 0.061945196241140366, + 0.08790873736143112, + 0.2013939619064331 + ], + "credit_map": [ + 0.014234043657779694, + 0.011074059642851353, + 0.013098184950649738, + 0.012201257050037384, + 0.011725101619958878, + 0.014659911394119263, + 0.02169993333518505, + 0.013700740411877632, + 0.014612776227295399, + 0.019252492114901543, + 0.026067091152071953, + 0.019344909116625786, + 0.023174790665507317, + 0.020139029249548912, + 0.017482861876487732, + 0.01975281722843647, + 0.014206761494278908, + 0.015224823728203773, + 0.008527569472789764, + 0.011811120435595512, + 0.024306150153279305, + 0.016103453934192657, + 0.008325139991939068, + 0.011039651930332184, + 0.010355433449149132, + 0.01730450801551342, + 0.00657319463789463, + 0.010235481895506382, + 0.00989046972244978, + 0.008032945916056633, + 0.008829023689031601, + 0.008473357185721397, + 0.003995351493358612, + 0.008425562642514706, + 0.005829848349094391, + 0.00772235868498683, + 0.009515934623777866, + 0.013921024277806282, + 0.017462920397520065, + 0.008630857802927494, + 0.004567131400108337, + 0.006204566452652216, + 0.004722068551927805, + 0.008098538964986801, + 0.007965609431266785, + 0.010284828022122383, + 0.01075587049126625, + 0.011476146057248116, + 0.00328681617975235, + 0.00913415476679802, + 0.0037328461185097694, + 0.006022270303219557, + 0.006776961497962475, + 0.005609283223748207, + 0.007882900536060333, + 0.00526913907378912, + 0.0070014274679124355, + 0.008296163752675056, + 0.006404588930308819, + 0.009034357033669949, + 0.007748785428702831, + 0.007748015690594912, + 0.0093040456995368, + 0.006407806184142828, + 0.026859819889068604, + 0.011538572609424591, + 0.012486881576478481, + 0.008709262125194073, + 0.0074846199713647366, + 0.006313111633062363, + 0.0076561737805604935, + 0.006860293447971344, + 0.058307114988565445, + 0.030812354758381844, + 0.03363243117928505, + 0.013641804456710815, + 0.02156665176153183, + 0.01500785443931818, + 0.014138550497591496, + 0.014287197031080723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.006738912779837847 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863910373300314 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.00863910373300314, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863910373300314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06603346019983292, + 0.11925158649682999, + 0.06812640279531479, + 0.08973749727010727, + 0.18570348620414734, + 0.12848500907421112, + 0.0797954797744751, + 0.048753779381513596, + 0.06835046410560608, + 0.1457628458738327 + ], + "credit_map": [ + 0.008320051245391369, + 0.008835365064442158, + 0.008411687798798084, + 0.007634918205440044, + 0.008097604848444462, + 0.008432193659245968, + 0.007459351327270269, + 0.008842295967042446, + 0.0072959489189088345, + 0.018835756927728653, + 0.012791511602699757, + 0.0154888229444623, + 0.020210200920701027, + 0.01812681183218956, + 0.010570690967142582, + 0.015931861475110054, + 0.008988172747194767, + 0.010974904522299767, + 0.005142229609191418, + 0.007935738191008568, + 0.012093696743249893, + 0.009560840204358101, + 0.005410938989371061, + 0.008019888773560524, + 0.01182682067155838, + 0.013225456699728966, + 0.005881450604647398, + 0.010190636850893497, + 0.011468442156910896, + 0.015364567749202251, + 0.009541062638163567, + 0.012239068746566772, + 0.006786658428609371, + 0.012923060916364193, + 0.005488687660545111, + 0.01362605020403862, + 0.01586001366376877, + 0.04443274810910225, + 0.06389980018138885, + 0.02268647775053978, + 0.006442733574658632, + 0.01229673158377409, + 0.005759525112807751, + 0.013427140191197395, + 0.013647687621414661, + 0.025865526869893074, + 0.02077355608344078, + 0.030272115021944046, + 0.008580406196415424, + 0.014276019297540188, + 0.007883409969508648, + 0.009176447987556458, + 0.010709819383919239, + 0.00995217077434063, + 0.00897311232984066, + 0.010244102217257023, + 0.006032405886799097, + 0.005685399752110243, + 0.0075299846939742565, + 0.005284611601382494, + 0.005307761020958424, + 0.005309023428708315, + 0.008406337350606918, + 0.005198261234909296, + 0.01652309112250805, + 0.008313141763210297, + 0.009351635351777077, + 0.006417949218302965, + 0.011035305447876453, + 0.005226244684308767, + 0.005634712520986795, + 0.005848392378538847, + 0.015749115496873856, + 0.019045615568757057, + 0.01601432077586651, + 0.011514309793710709, + 0.02672627568244934, + 0.032524850219488144, + 0.010954586789011955, + 0.013233781792223454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.00863910373300314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are black boots visible stepping on ground covered with withered yellow leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do thorny black vines with a dark sheen erupt from the ground where the boots land?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the vines slither and cling to the trunk of a snow-white birch tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a figure wearing an inky silk robe with a sharply defined face visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward to frame the figure's face after following the vines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636608529835939 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636608529835939, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636608529835939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19691401720046997, + 0.08697950839996338, + 0.09387964010238647, + 0.0656919926404953, + 0.08667697012424469, + 0.0887613520026207, + 0.09777641296386719, + 0.11176280677318573, + 0.07615029066801071, + 0.09540706872940063 + ], + "credit_map": [ + 0.021946128457784653, + 0.02514745481312275, + 0.026806196197867393, + 0.022395821288228035, + 0.019000345841050148, + 0.02237311191856861, + 0.03295702859759331, + 0.026287902146577835, + 0.006095890421420336, + 0.016985133290290833, + 0.006596246734261513, + 0.011469985358417034, + 0.013889304362237453, + 0.013033702969551086, + 0.006987554486840963, + 0.011921689845621586, + 0.014235164038836956, + 0.010971229523420334, + 0.006840963382273912, + 0.0088209742680192, + 0.02175317332148552, + 0.013700420036911964, + 0.007536422926932573, + 0.01002129353582859, + 0.011103448458015919, + 0.009397134184837341, + 0.004698536358773708, + 0.008132707327604294, + 0.008628924377262592, + 0.01148452702909708, + 0.0059647224843502045, + 0.006281993817538023, + 0.006827818229794502, + 0.008249628357589245, + 0.008929949253797531, + 0.010384336113929749, + 0.008843098767101765, + 0.015748105943202972, + 0.01964680850505829, + 0.008047210983932018, + 0.007026721723377705, + 0.006160583812743425, + 0.007409447804093361, + 0.012609894387423992, + 0.014591595157980919, + 0.013423481956124306, + 0.013259675353765488, + 0.014279941096901894, + 0.008932464756071568, + 0.024897754192352295, + 0.00858737900853157, + 0.013353069312870502, + 0.009662887081503868, + 0.010354626923799515, + 0.010478677228093147, + 0.011509545147418976, + 0.015345037914812565, + 0.020994234830141068, + 0.011165158823132515, + 0.019316798076033592, + 0.012397609651088715, + 0.009837057441473007, + 0.011551478877663612, + 0.011155428364872932, + 0.013760829344391823, + 0.009624951519072056, + 0.00869018118828535, + 0.011504587717354298, + 0.009646309539675713, + 0.006248720455914736, + 0.008748162537813187, + 0.007926542311906815, + 0.012532188557088375, + 0.014671383425593376, + 0.013061664998531342, + 0.007768341340124607, + 0.01825886219739914, + 0.01521243341267109, + 0.006011410616338253, + 0.007890776731073856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.004636608529835939 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_3/index_tensor([975, 975])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_3/index_tensor([975, 975])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b38b1a17f17c61cc887e79205cb61b9928de739d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_3/index_tensor([975, 975])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, static shot, taken from a high angle, captures a tranquil lake, its surface a deep sapphire, shimmering with layers of delicate, slowly spreading ripples. In the center of the frame, a bright red maple leaf, its edges slightly curled and its veins clearly visible, floats solitary and gracefully. Soft, golden morning light slants through the frame from the side, filtering through the leaf, giving it a translucent quality and a golden halo around its edges, casting long, shimmering paths across the shimmering water. Gently propelled by the ripples, the leaf drifts across the water with an almost imperceptible, steady, and leisurely motion, occasionally accompanied by the slightest rotation. The water reflects the blurred, vibrant outlines of the autumn trees on the distant shore. The entire scene is enveloped in a tranquil, contemplative atmosphere, imbued with warm, saturated tones and imbued with a poetic beauty.", + "seed": 31507, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_3/index_975_seed_31507_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.049704614514485004, + "vlm_logprob": -0.049704614514485004, + "vlm_yes_no_margin": 4.024999952316284, + "vlm_token_credit": -0.049704614514485004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08067098259925842 + ], + "opposite_logprob_tokens": [ + -2.9556710720062256 + ], + "logprob_reward": -0.08067098259925842, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08067098259925842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20382529497146606, + 0.06840743124485016, + 0.2746667265892029, + 0.148041769862175, + 0.04687316715717316, + 0.05498473718762398, + 0.06632369011640549, + 0.03751177713274956, + 0.03395017236471176, + 0.06541524082422256 + ], + "credit_map": [ + 0.04509429261088371, + 0.03570055589079857, + 0.02569105662405491, + 0.013568892143666744, + 0.017210735008120537, + 0.018407048657536507, + 0.016352863982319832, + 0.031799837946891785, + 0.00662324158474803, + 0.005921135190874338, + 0.008517234586179256, + 0.0112614119425416, + 0.008668207563459873, + 0.0061776405200362206, + 0.008492988534271717, + 0.012745572254061699, + 0.018602298572659492, + 0.015545200556516647, + 0.1044146865606308, + 0.02554197609424591, + 0.010606808587908745, + 0.010785912163555622, + 0.013182149268686771, + 0.07598767429590225, + 0.019188860431313515, + 0.013817265629768372, + 0.017825786024332047, + 0.01610700786113739, + 0.014532855711877346, + 0.013958035968244076, + 0.02437218651175499, + 0.028239775449037552, + 0.003867737017571926, + 0.003951333928853273, + 0.008180040866136551, + 0.006502983160316944, + 0.0048242430202662945, + 0.003399410517886281, + 0.006634872872382402, + 0.00951254740357399, + 0.006278147920966148, + 0.005966071039438248, + 0.009563709609210491, + 0.006473034620285034, + 0.007928314618766308, + 0.004792823921889067, + 0.007835815660655499, + 0.006146817933768034, + 0.00929952785372734, + 0.011308424174785614, + 0.0084069287404418, + 0.0063378214836120605, + 0.006208781152963638, + 0.0070492420345544815, + 0.009024449624121189, + 0.00868851039558649, + 0.006149406544864178, + 0.004786347970366478, + 0.00541008310392499, + 0.003611948573961854, + 0.003091803751885891, + 0.0034145095851272345, + 0.0038682688027620316, + 0.007179410196840763, + 0.005948911886662245, + 0.006462086923420429, + 0.004287298768758774, + 0.002524228999391198, + 0.00255430955439806, + 0.0034276435617357492, + 0.0028576164040714502, + 0.005888077430427074, + 0.008276868611574173, + 0.006020946428179741, + 0.01650816947221756, + 0.006733762100338936, + 0.004934432916343212, + 0.00534985912963748, + 0.006363820284605026, + 0.011227380484342575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.08067098259925842 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428370326757431 + ], + "opposite_logprob_tokens": [ + -2.0178370475769043 + ], + "logprob_reward": -0.1428370326757431, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428370326757431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21757349371910095, + 0.08983758091926575, + 0.17496302723884583, + 0.09940096735954285, + 0.053244441747665405, + 0.05459321662783623, + 0.08504360169172287, + 0.044662706553936005, + 0.05839037150144577, + 0.12229055911302567 + ], + "credit_map": [ + 0.04848743975162506, + 0.020750394091010094, + 0.06363166868686676, + 0.017578918486833572, + 0.013979349285364151, + 0.01145409606397152, + 0.013677439652383327, + 0.028014201670885086, + 0.010248580016195774, + 0.010100624524056911, + 0.016455676406621933, + 0.013258315622806549, + 0.00807222444564104, + 0.00822590570896864, + 0.010211171582341194, + 0.013265085406601429, + 0.011076436378061771, + 0.008694298565387726, + 0.06881079822778702, + 0.015188682824373245, + 0.004602938424795866, + 0.006519103888422251, + 0.008853215724229813, + 0.05121755599975586, + 0.014823367819190025, + 0.010372661054134369, + 0.022999558597803116, + 0.010512112639844418, + 0.00933010969310999, + 0.010446889325976372, + 0.009089868515729904, + 0.011826405301690102, + 0.006789303384721279, + 0.0069363052025437355, + 0.009149378165602684, + 0.007183491252362728, + 0.003925294149667025, + 0.003595741931349039, + 0.007616471499204636, + 0.008048458024859428, + 0.008113092742860317, + 0.006769934203475714, + 0.012440103106200695, + 0.00473951268941164, + 0.0038792812265455723, + 0.005086307413876057, + 0.005189063027501106, + 0.008375926874577999, + 0.015472142025828362, + 0.019555052742362022, + 0.008785916492342949, + 0.005856354720890522, + 0.005631895270198584, + 0.006875680759549141, + 0.007350154686719179, + 0.01551640685647726, + 0.007591704837977886, + 0.005271797999739647, + 0.0072294557467103004, + 0.004086829721927643, + 0.0048434436321258545, + 0.005715632811188698, + 0.0045776828192174435, + 0.005346162710338831, + 0.01176860649138689, + 0.01025991141796112, + 0.010036900639533997, + 0.0055328612215816975, + 0.003485019551590085, + 0.004344688728451729, + 0.004353196360170841, + 0.008609193377196789, + 0.015688246116042137, + 0.008654234930872917, + 0.026579707860946655, + 0.009030277840793133, + 0.010177410207688808, + 0.011347291059792042, + 0.0185870211571455, + 0.022226376459002495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.1428370326757431 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003204094711691141 + ], + "opposite_logprob_tokens": [ + -5.753203868865967 + ], + "logprob_reward": -0.003204094711691141, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003204094711691141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1425323188304901, + 0.09008152782917023, + 0.1833713799715042, + 0.059485893696546555, + 0.15440733730793, + 0.027293212711811066, + 0.0915098562836647, + 0.16577856242656708, + 0.04399654641747475, + 0.04154333099722862 + ], + "credit_map": [ + 0.014634573832154274, + 0.014513528905808926, + 0.03724450618028641, + 0.016452088952064514, + 0.013575613498687744, + 0.011174939572811127, + 0.012761720456182957, + 0.022175341844558716, + 0.010087965987622738, + 0.011938592419028282, + 0.011604070663452148, + 0.014456475153565407, + 0.0087961470708251, + 0.006443361286073923, + 0.011308904737234116, + 0.015446014702320099, + 0.013584832660853863, + 0.014352892525494099, + 0.03447223827242851, + 0.03544513136148453, + 0.015232071280479431, + 0.01160348579287529, + 0.021313095465302467, + 0.03736764192581177, + 0.0133771151304245, + 0.006825086195021868, + 0.006903446279466152, + 0.006385349202901125, + 0.006240146700292826, + 0.00569628830999136, + 0.007531323470175266, + 0.006527138873934746, + 0.010838886722922325, + 0.014213544316589832, + 0.01673239842057228, + 0.02585243619978428, + 0.01244161557406187, + 0.010673453100025654, + 0.03153977170586586, + 0.032115235924720764, + 0.004920945502817631, + 0.003892431268468499, + 0.004529588855803013, + 0.0031706001609563828, + 0.002655142918229103, + 0.002482340671122074, + 0.0029558453243225813, + 0.002686316380277276, + 0.014504720456898212, + 0.016751989722251892, + 0.009983920492231846, + 0.008484571240842342, + 0.009178479202091694, + 0.008240311406552792, + 0.010644961148500443, + 0.013720903545618057, + 0.012337478809058666, + 0.01115263719111681, + 0.014186404645442963, + 0.01669040508568287, + 0.012741617858409882, + 0.038308363407850266, + 0.016513222828507423, + 0.0438484251499176, + 0.00757908821105957, + 0.005806055851280689, + 0.007491762284189463, + 0.00382354948669672, + 0.003533824346959591, + 0.004245653282850981, + 0.0044413357973098755, + 0.00707527669146657, + 0.004759343806654215, + 0.003433868056163192, + 0.009445802308619022, + 0.00453932536765933, + 0.0030700191855430603, + 0.003317618742585182, + 0.006039143074303865, + 0.006938213016837835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.003204094711691141 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819058135151863 + ], + "opposite_logprob_tokens": [ + -4.018190383911133 + ], + "logprob_reward": -0.01819058135151863, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819058135151863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1813122034072876, + 0.18333859741687775, + 0.1577463001012802, + 0.060494374483823776, + 0.12401410937309265, + 0.03219248726963997, + 0.09948435425758362, + 0.02837035059928894, + 0.03811916336417198, + 0.09492804110050201 + ], + "credit_map": [ + 0.02256741374731064, + 0.025300776585936546, + 0.04051065444946289, + 0.015523947775363922, + 0.01682061143219471, + 0.01283173356205225, + 0.014720686711370945, + 0.03303638845682144, + 0.03097682073712349, + 0.030757451429963112, + 0.022826960310339928, + 0.022289631888270378, + 0.014006908982992172, + 0.014776498079299927, + 0.017247771844267845, + 0.030456548556685448, + 0.008296740241348743, + 0.009408373385667801, + 0.06000104546546936, + 0.013930189423263073, + 0.004680882673710585, + 0.006794482935220003, + 0.008480279706418514, + 0.04615429788827896, + 0.009619360789656639, + 0.00865420326590538, + 0.010586737655103207, + 0.005798554513603449, + 0.004976014606654644, + 0.006905192043632269, + 0.0056555382907390594, + 0.008298773318529129, + 0.017655597999691963, + 0.014764303341507912, + 0.01570133864879608, + 0.013909721747040749, + 0.008858554996550083, + 0.011247141286730766, + 0.01825089566409588, + 0.02362656034529209, + 0.006648898124694824, + 0.005097977817058563, + 0.005397004075348377, + 0.0029031638987362385, + 0.0025299619883298874, + 0.0031783045269548893, + 0.002607240341603756, + 0.003829935099929571, + 0.018850795924663544, + 0.02194211073219776, + 0.010685862973332405, + 0.007170806638896465, + 0.007299716584384441, + 0.008881739340722561, + 0.008361960761249065, + 0.016291361302137375, + 0.005685433279722929, + 0.003834420582279563, + 0.0027865248266607523, + 0.0021701687946915627, + 0.003027907805517316, + 0.0045987749472260475, + 0.002761765383183956, + 0.0035053561441600323, + 0.00683143874630332, + 0.006128891371190548, + 0.006134524010121822, + 0.0034039674792438745, + 0.002993127563968301, + 0.003338491078466177, + 0.0036825246643275023, + 0.005606199149042368, + 0.012211437337100506, + 0.006557847838848829, + 0.020538242533802986, + 0.007225723937153816, + 0.007703087758272886, + 0.010756607167422771, + 0.013648242689669132, + 0.016286853700876236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01819058135151863 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362038123421371 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.00362038123421371, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362038123421371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12561261653900146, + 0.06427564471960068, + 0.25402018427848816, + 0.05725136399269104, + 0.06694336980581284, + 0.04832746461033821, + 0.042225517332553864, + 0.04358268529176712, + 0.03847644105553627, + 0.259284645318985 + ], + "credit_map": [ + 0.01645061746239662, + 0.013787343166768551, + 0.03586648404598236, + 0.012244322337210178, + 0.009895284660160542, + 0.00831634271889925, + 0.009810895659029484, + 0.019241325557231903, + 0.009136704728007317, + 0.007919303141534328, + 0.00922943465411663, + 0.0075243194587528706, + 0.007269812747836113, + 0.0045979744754731655, + 0.008748632855713367, + 0.009849460795521736, + 0.025243354961276054, + 0.016404982656240463, + 0.07925683259963989, + 0.02987184375524521, + 0.020516065880656242, + 0.01576530560851097, + 0.01926756650209427, + 0.04769425466656685, + 0.011701593175530434, + 0.006589200813323259, + 0.006638659629970789, + 0.005745333153754473, + 0.004448358900845051, + 0.004598672967404127, + 0.006856195628643036, + 0.010673348791897297, + 0.0066984728910028934, + 0.008191955275833607, + 0.007496682461351156, + 0.010644366033375263, + 0.0054224226623773575, + 0.0037160899955779314, + 0.01280344370752573, + 0.01196993887424469, + 0.0054265703074634075, + 0.0051157413981854916, + 0.008572294376790524, + 0.005470330361276865, + 0.0070226965472102165, + 0.004148853477090597, + 0.006857240106910467, + 0.005713735707104206, + 0.005985258612781763, + 0.006445827428251505, + 0.005406307056546211, + 0.003689064644277096, + 0.003787735477089882, + 0.004520173650234938, + 0.005598645191639662, + 0.006792501080781221, + 0.006449979264289141, + 0.0054476941004395485, + 0.005449551157653332, + 0.004383299965411425, + 0.005333377048373222, + 0.005143195856362581, + 0.006322500295937061, + 0.005053091328591108, + 0.007545996457338333, + 0.005246841348707676, + 0.006906307768076658, + 0.0028230210300534964, + 0.003165947273373604, + 0.004719119984656572, + 0.0026905271224677563, + 0.005378680769354105, + 0.06050040200352669, + 0.025549152866005898, + 0.03057069703936577, + 0.016614211723208427, + 0.018072180449962616, + 0.017336511984467506, + 0.020557668060064316, + 0.07008382678031921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.00362038123421371 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, static shot, taken from a high angle, captures a tranquil lake, its surface a deep sapphire, shimmering with layers of delicate, slowly spreading ripples. In the center of the frame, a bright red maple leaf, its edges slightly curled and its veins clearly visible, floats solitary and gracefully. Soft, golden morning light slants through the frame from the side, filtering through the leaf, giving it a translucent quality and a golden halo around its edges, casting long, shimmering paths across the shimmering water. Gently propelled by the ripples, the leaf drifts across the water with an almost imperceptible, steady, and leisurely motion, occasionally accompanied by the slightest rotation. The water reflects the blurred, vibrant outlines of the autumn trees on the distant shore. The entire scene is enveloped in a tranquil, contemplative atmosphere, imbued with warm, saturated tones and imbued with a poetic beauty.", + "seed": 31508, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_3/index_975_seed_31508_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0511403227224946, + "vlm_logprob": -0.0511403227224946, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.0511403227224946, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08403513580560684 + ], + "opposite_logprob_tokens": [ + -2.9590351581573486 + ], + "logprob_reward": -0.08403513580560684, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08403513580560684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13939599692821503, + 0.049719471484422684, + 0.12152115255594254, + 0.39484843611717224, + 0.02335434965789318, + 0.023196760565042496, + 0.01897497847676277, + 0.1775769591331482, + 0.017082687467336655, + 0.034329112619161606 + ], + "credit_map": [ + 0.012159451842308044, + 0.00696683069691062, + 0.018786879256367683, + 0.013945044949650764, + 0.01809820719063282, + 0.02021087147295475, + 0.015547932125627995, + 0.03368079289793968, + 0.007041007746011019, + 0.0030016028322279453, + 0.007074847351759672, + 0.0034092396963387728, + 0.004878526087850332, + 0.01500630285590887, + 0.004777542781084776, + 0.004530410747975111, + 0.0068240598775446415, + 0.004337400663644075, + 0.019480843096971512, + 0.007668039761483669, + 0.009574665687978268, + 0.021758602932095528, + 0.01179265696555376, + 0.040084902197122574, + 0.025901641696691513, + 0.009561756625771523, + 0.1296241283416748, + 0.018598366528749466, + 0.018398193642497063, + 0.01975030079483986, + 0.030223913490772247, + 0.14279018342494965, + 0.006550791673362255, + 0.0020435473416000605, + 0.004036433529108763, + 0.0021703816018998623, + 0.0016326686600223184, + 0.0022203021217137575, + 0.0018727070419117808, + 0.00282752001658082, + 0.003708453616127372, + 0.0018501500599086285, + 0.0033367399591952562, + 0.0026909741573035717, + 0.001490588067099452, + 0.0022840863093733788, + 0.0019525206880643964, + 0.005883248522877693, + 0.0011562714353203773, + 0.0010134989861398935, + 0.004375356249511242, + 0.0027047654148191214, + 0.0015055405674502254, + 0.0025887100491672754, + 0.0033511428628116846, + 0.0022796953562647104, + 0.01194791030138731, + 0.011040161363780499, + 0.03330842778086662, + 0.01243972685188055, + 0.013599159196019173, + 0.017770059406757355, + 0.014012380503118038, + 0.06345914304256439, + 0.0018618232570588589, + 0.0012772061163559556, + 0.003763999789953232, + 0.0022737341932952404, + 0.0014663360780104995, + 0.002494033658877015, + 0.0018634841544553638, + 0.0020820728968828917, + 0.0024450316559523344, + 0.00161379121709615, + 0.010908189229667187, + 0.004597677383571863, + 0.0024526966735720634, + 0.008640793152153492, + 0.001379638910293579, + 0.0022912942804396152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.08403513580560684 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12711083889007568 + ], + "opposite_logprob_tokens": [ + -2.1271109580993652 + ], + "logprob_reward": -0.12711083889007568, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12711083889007568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08562906086444855, + 0.18021950125694275, + 0.5222396850585938, + 0.030803516507148743, + 0.049540694802999496, + 0.024970727041363716, + 0.017018627375364304, + 0.019257521256804466, + 0.01630537584424019, + 0.05401529371738434 + ], + "credit_map": [ + 0.010378159582614899, + 0.0057827685959637165, + 0.011194933205842972, + 0.009665431454777718, + 0.01467120461165905, + 0.016877030953764915, + 0.008382558822631836, + 0.008676961064338684, + 0.01567218452692032, + 0.005464192479848862, + 0.06160232797265053, + 0.010197550989687443, + 0.01288081519305706, + 0.02500702627003193, + 0.021824251860380173, + 0.027571149170398712, + 0.040330540388822556, + 0.022487109526991844, + 0.15321755409240723, + 0.026441384106874466, + 0.04759904742240906, + 0.093400739133358, + 0.03508637472987175, + 0.10367689281702042, + 0.0019513684092089534, + 0.0010432095732539892, + 0.009768546558916569, + 0.001631465507671237, + 0.0016590561717748642, + 0.002401681151241064, + 0.002375449985265732, + 0.009972738102078438, + 0.008147086016833782, + 0.003240538528189063, + 0.007781715597957373, + 0.002908559050410986, + 0.0048655057325959206, + 0.010137771256268024, + 0.0033683779183775187, + 0.009091134183108807, + 0.0037954419385641813, + 0.001996015664190054, + 0.0030092489905655384, + 0.0021623040083795786, + 0.003654100699350238, + 0.0033656901214271784, + 0.0018556839786469936, + 0.005132239777594805, + 0.0020963866263628006, + 0.001347749144770205, + 0.002689831890165806, + 0.0022434291895478964, + 0.0020773077849298716, + 0.0022337117698043585, + 0.001986467745155096, + 0.0023437433410435915, + 0.001389057724736631, + 0.0009853490628302097, + 0.004134384449571371, + 0.0013390803942456841, + 0.0013812328688800335, + 0.0014885506825521588, + 0.0017277227016165853, + 0.006812141742557287, + 0.0018667380791157484, + 0.001553152222186327, + 0.002999145071953535, + 0.0019947157707065344, + 0.0013558019418269396, + 0.001987665193155408, + 0.0021886168979108334, + 0.00235953857190907, + 0.003989111166447401, + 0.00222603278234601, + 0.01728014275431633, + 0.007885675877332687, + 0.004291154909878969, + 0.01323491707444191, + 0.0024435201194137335, + 0.0026647343765944242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.12711083889007568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041060671210289 + ], + "opposite_logprob_tokens": [ + -5.504106044769287 + ], + "logprob_reward": -0.0041060671210289, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041060671210289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08710363507270813, + 0.08664476126432419, + 0.2240891009569168, + 0.37457698583602905, + 0.042962606996297836, + 0.013162034563720226, + 0.014187217690050602, + 0.07507900148630142, + 0.015259222127497196, + 0.06693539023399353 + ], + "credit_map": [ + 0.010065061040222645, + 0.006120468024164438, + 0.011775151826441288, + 0.011835009790956974, + 0.014989521354436874, + 0.013396110385656357, + 0.0078706294298172, + 0.011051688343286514, + 0.005792155396193266, + 0.002795773558318615, + 0.025668896734714508, + 0.005871963687241077, + 0.007737932726740837, + 0.022655609995126724, + 0.011203272268176079, + 0.004919169005006552, + 0.016290415078401566, + 0.00787418708205223, + 0.06828363984823227, + 0.011497393250465393, + 0.01590900868177414, + 0.016672490164637566, + 0.01479512918740511, + 0.07276684790849686, + 0.025596708059310913, + 0.011708222329616547, + 0.11919606477022171, + 0.020400233566761017, + 0.018984558060765266, + 0.01746789552271366, + 0.027974039316177368, + 0.13324925303459167, + 0.012540627270936966, + 0.0038501236122101545, + 0.006378632504492998, + 0.003361775539815426, + 0.0032719061709940434, + 0.004866969771683216, + 0.0030008244793862104, + 0.005691747646778822, + 0.001967745367437601, + 0.0011535660596564412, + 0.0012966342037543654, + 0.0013790430966764688, + 0.002357506426051259, + 0.0020077365916222334, + 0.0013517154147848487, + 0.0016480874037370086, + 0.001117335050366819, + 0.0010487290564924479, + 0.0027731559239327908, + 0.002919054590165615, + 0.0011395139154046774, + 0.0016795232659205794, + 0.0019508472178131342, + 0.0015590592520311475, + 0.005870944354683161, + 0.004262191243469715, + 0.012618565000593662, + 0.00536060705780983, + 0.005829654633998871, + 0.00830624345690012, + 0.005999277345836163, + 0.026831530034542084, + 0.0019545629620552063, + 0.001670805853791535, + 0.0022276206873357296, + 0.0018866906175389886, + 0.0014095745282247663, + 0.0013742642477154732, + 0.002588718431070447, + 0.002146985614672303, + 0.0051681348122656345, + 0.002611030824482441, + 0.021473489701747894, + 0.00907949823886156, + 0.0044563873670995235, + 0.017672119662165642, + 0.002680198522284627, + 0.0037945362273603678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0041060671210289 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033710777759552 + ], + "opposite_logprob_tokens": [ + -3.4087107181549072 + ], + "logprob_reward": -0.033710777759552, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033710777759552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08728844672441483, + 0.10139372944831848, + 0.2031593769788742, + 0.08586527407169342, + 0.15620861947536469, + 0.03361843153834343, + 0.03159536421298981, + 0.1269465982913971, + 0.04233706369996071, + 0.13158711791038513 + ], + "credit_map": [ + 0.008272191509604454, + 0.006595881190150976, + 0.012027377262711525, + 0.009868641383945942, + 0.013943375088274479, + 0.00938438717275858, + 0.00879680085927248, + 0.018399782478809357, + 0.014991287142038345, + 0.004985172301530838, + 0.03221059590578079, + 0.00760452589020133, + 0.006504753604531288, + 0.008834418840706348, + 0.013103135861456394, + 0.013159836642444134, + 0.01541132852435112, + 0.007941358722746372, + 0.024257676675915718, + 0.008597909472882748, + 0.01778644137084484, + 0.04588242247700691, + 0.020843395963311195, + 0.06243886053562164, + 0.008204732090234756, + 0.004278762731701136, + 0.025341933593153954, + 0.005718620028346777, + 0.005132668651640415, + 0.007109188009053469, + 0.007589227054268122, + 0.02249014377593994, + 0.03283407911658287, + 0.011609571985900402, + 0.022443104535341263, + 0.01028386503458023, + 0.012144321575760841, + 0.02295047976076603, + 0.010785171762108803, + 0.033158015459775925, + 0.0050806947983801365, + 0.0028157224878668785, + 0.003504732623696327, + 0.0027598158922046423, + 0.005115104373544455, + 0.004149424843490124, + 0.002941214246675372, + 0.007251724600791931, + 0.0028447990771383047, + 0.0023085996508598328, + 0.006417811382561922, + 0.005265384446829557, + 0.002356290817260742, + 0.004392298869788647, + 0.0042111934162676334, + 0.0037989874836057425, + 0.009517163969576359, + 0.004448968917131424, + 0.03548630326986313, + 0.008560113608837128, + 0.006644349545240402, + 0.007805514615029097, + 0.008759340271353722, + 0.045724835246801376, + 0.008894147351384163, + 0.0049000028520822525, + 0.006934681907296181, + 0.003930237144231796, + 0.004302654415369034, + 0.005505850538611412, + 0.003762552049010992, + 0.0041069393046200275, + 0.00929163210093975, + 0.007595967501401901, + 0.03961149975657463, + 0.016658246517181396, + 0.012655164115130901, + 0.03246506303548813, + 0.005946538411080837, + 0.00736299529671669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.033710777759552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738794036209583 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738794036209583, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738794036209583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07981742173433304, + 0.12155488133430481, + 0.1402270346879959, + 0.3116251230239868, + 0.027192700654268265, + 0.012709365226328373, + 0.06340830028057098, + 0.20273663103580475, + 0.01608520746231079, + 0.024643301963806152 + ], + "credit_map": [ + 0.009596655145287514, + 0.004818814806640148, + 0.008750537410378456, + 0.00939185731112957, + 0.012814398854970932, + 0.01249503344297409, + 0.007652889937162399, + 0.014297250658273697, + 0.010614068247377872, + 0.003737415187060833, + 0.03963928297162056, + 0.007473306730389595, + 0.009667868725955486, + 0.027209816500544548, + 0.016942301765084267, + 0.006270826794207096, + 0.012321202084422112, + 0.005892821121960878, + 0.015239452011883259, + 0.005699455738067627, + 0.016458015888929367, + 0.04233786463737488, + 0.01479718741029501, + 0.027481045573949814, + 0.021015657112002373, + 0.009067323058843613, + 0.10206853598356247, + 0.01419721357524395, + 0.014733533374965191, + 0.01462389063090086, + 0.023495785892009735, + 0.11242320388555527, + 0.0036094537936151028, + 0.0014505886938422918, + 0.004667996894568205, + 0.0024630806874483824, + 0.0025842648465186357, + 0.007356110494583845, + 0.0019683085847646, + 0.0030929003842175007, + 0.0011292631970718503, + 0.0012466119369491935, + 0.0015038299607113004, + 0.0020319034811109304, + 0.0014257754664868116, + 0.0015364775899797678, + 0.0016699273837730289, + 0.0021655780728906393, + 0.007720564026385546, + 0.0044281817972660065, + 0.01165421400219202, + 0.007684785407036543, + 0.006951303221285343, + 0.008725461550056934, + 0.00964827835559845, + 0.006595509126782417, + 0.014323217794299126, + 0.00739959767088294, + 0.05494043603539467, + 0.013605470769107342, + 0.010796747170388699, + 0.01248775701969862, + 0.01366441510617733, + 0.07551901787519455, + 0.0025706745218485594, + 0.0019400258315727115, + 0.002703588455915451, + 0.0019758734852075577, + 0.0014024438569322228, + 0.0015396245289593935, + 0.002063308609649539, + 0.0018896696856245399, + 0.0025580620858818293, + 0.0020067363511770964, + 0.004490029998123646, + 0.0029006050899624825, + 0.0025917706079781055, + 0.005028876941651106, + 0.002188279991969466, + 0.0028789453208446503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.006738794036209583 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, static shot, taken from a high angle, captures a tranquil lake, its surface a deep sapphire, shimmering with layers of delicate, slowly spreading ripples. In the center of the frame, a bright red maple leaf, its edges slightly curled and its veins clearly visible, floats solitary and gracefully. Soft, golden morning light slants through the frame from the side, filtering through the leaf, giving it a translucent quality and a golden halo around its edges, casting long, shimmering paths across the shimmering water. Gently propelled by the ripples, the leaf drifts across the water with an almost imperceptible, steady, and leisurely motion, occasionally accompanied by the slightest rotation. The water reflects the blurred, vibrant outlines of the autumn trees on the distant shore. The entire scene is enveloped in a tranquil, contemplative atmosphere, imbued with warm, saturated tones and imbued with a poetic beauty.", + "seed": 31509, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_3/index_975_seed_31509_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12109271371737122, + "vlm_logprob": -0.12109271371737122, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.12109271371737122, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.37971699237823486 + ], + "opposite_logprob_tokens": [ + -1.2547169923782349 + ], + "logprob_reward": -0.37971699237823486, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.37971699237823486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09036630392074585, + 0.058401353657245636, + 0.1836308091878891, + 0.1259080320596695, + 0.27984198927879333, + 0.03658607229590416, + 0.014054698869585991, + 0.08173564821481705, + 0.10089664906263351, + 0.028578538447618484 + ], + "credit_map": [ + 0.017059408128261566, + 0.011362778954207897, + 0.013930541463196278, + 0.010409506969153881, + 0.007483192719519138, + 0.00902667734771967, + 0.007945946417748928, + 0.01314824540168047, + 0.008309779688715935, + 0.007065911777317524, + 0.007088422309607267, + 0.010891258716583252, + 0.0060877385549247265, + 0.004839170724153519, + 0.007566509302705526, + 0.0065525611862540245, + 0.02996121533215046, + 0.020363258197903633, + 0.0247343722730875, + 0.021093571558594704, + 0.01858210377395153, + 0.03063035197556019, + 0.011689421720802784, + 0.026576507836580276, + 0.013200495392084122, + 0.012296467088162899, + 0.029358824715018272, + 0.013024494051933289, + 0.009626821614801884, + 0.011464673094451427, + 0.008220668882131577, + 0.028715578839182854, + 0.01732039637863636, + 0.006505065131932497, + 0.11438944935798645, + 0.011587978340685368, + 0.007288177497684956, + 0.009174050763249397, + 0.011445624753832817, + 0.10213121771812439, + 0.0028821169398725033, + 0.003039728384464979, + 0.00744999386370182, + 0.005043318960815668, + 0.002806671429425478, + 0.0028699396643787622, + 0.004266570787876844, + 0.008227728307247162, + 0.0009018270648084581, + 0.0010392258409410715, + 0.0038911348674446344, + 0.001751755247823894, + 0.001096571795642376, + 0.0012026519980281591, + 0.0009467033669352531, + 0.0032248275820165873, + 0.006259738467633724, + 0.010308296419680119, + 0.008103091269731522, + 0.0028386530466377735, + 0.007192014716565609, + 0.025463925674557686, + 0.004982136655598879, + 0.01658777892589569, + 0.010700189508497715, + 0.015984004363417625, + 0.012784560211002827, + 0.01206208672374487, + 0.01052851602435112, + 0.01063559390604496, + 0.008147681131958961, + 0.020054003223776817, + 0.0027595949359238148, + 0.002739030634984374, + 0.005762005690485239, + 0.003816003445535898, + 0.003008575178682804, + 0.0024425385054200888, + 0.00443950854241848, + 0.0036112782545387745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.37971699237823486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015974223613739 + ], + "opposite_logprob_tokens": [ + -1.7015974521636963 + ], + "logprob_reward": -0.2015974223613739, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015974223613739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08356110751628876, + 0.17525897920131683, + 0.17880356311798096, + 0.09814475476741791, + 0.08796562254428864, + 0.07323522120714188, + 0.025037618353962898, + 0.03398490697145462, + 0.08539126813411713, + 0.15861694514751434 + ], + "credit_map": [ + 0.012905711308121681, + 0.012331117875874043, + 0.010191995650529861, + 0.01044438499957323, + 0.010434545576572418, + 0.00939653255045414, + 0.008680700324475765, + 0.009176130406558514, + 0.02564050629734993, + 0.02329409308731556, + 0.02676599659025669, + 0.01823895238339901, + 0.02408016100525856, + 0.016648657619953156, + 0.016409894451498985, + 0.024180730804800987, + 0.028865795582532883, + 0.0202792976051569, + 0.02200710028409958, + 0.022572733461856842, + 0.01944842003285885, + 0.02844780683517456, + 0.012499340809881687, + 0.02468308061361313, + 0.013095716945827007, + 0.009386981837451458, + 0.017909757792949677, + 0.00857654120773077, + 0.01117214560508728, + 0.01142670214176178, + 0.009097893722355366, + 0.01747901737689972, + 0.008685974404215813, + 0.0052147177048027515, + 0.023947879672050476, + 0.005727861542254686, + 0.00669242674484849, + 0.009835077449679375, + 0.004700636025518179, + 0.023161059245467186, + 0.00756429648026824, + 0.006976870354264975, + 0.01555395033210516, + 0.008018250577151775, + 0.006130537483841181, + 0.007571919355541468, + 0.0069571188651025295, + 0.014462294988334179, + 0.0022573021706193686, + 0.003458817256614566, + 0.003930207807570696, + 0.004278732929378748, + 0.002679613418877125, + 0.0027662639040499926, + 0.0025064987130463123, + 0.003160183085128665, + 0.00294799217954278, + 0.003597480244934559, + 0.00898753758519888, + 0.002843523398041725, + 0.0029340777546167374, + 0.0023847087286412716, + 0.0028475753497332335, + 0.007442015688866377, + 0.012117886915802956, + 0.010895581915974617, + 0.009291470982134342, + 0.006715396884828806, + 0.00773580651730299, + 0.021614674478769302, + 0.006632613018155098, + 0.010387844406068325, + 0.015435556881129742, + 0.015240388922393322, + 0.0307712871581316, + 0.016929421573877335, + 0.015239574015140533, + 0.02235228754580021, + 0.021548287943005562, + 0.021100157871842384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.2015974223613739 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013438966125249863 + ], + "opposite_logprob_tokens": [ + -6.626343727111816 + ], + "logprob_reward": -0.0013438966125249863, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013438966125249863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04506273940205574, + 0.03621494024991989, + 0.0718400776386261, + 0.1285199522972107, + 0.37502017617225647, + 0.029932690784335136, + 0.09130445867776871, + 0.12803761661052704, + 0.0396532341837883, + 0.054414112120866776 + ], + "credit_map": [ + 0.006094936281442642, + 0.004101872444152832, + 0.006676349323242903, + 0.004520144313573837, + 0.004232592880725861, + 0.0065889800898730755, + 0.005646802484989166, + 0.007201062049716711, + 0.004095715004950762, + 0.003445996902883053, + 0.004810556303709745, + 0.006572593003511429, + 0.003589159809052944, + 0.003346940502524376, + 0.0064797853119671345, + 0.003874195273965597, + 0.012063288129866123, + 0.00810861773788929, + 0.010060073807835579, + 0.008278234861791134, + 0.007253083400428295, + 0.010418160818517208, + 0.00518900016322732, + 0.010469621047377586, + 0.015598111785948277, + 0.018754824995994568, + 0.021600613370537758, + 0.015628835186362267, + 0.012043503113090992, + 0.01304197683930397, + 0.010283863171935081, + 0.02156822383403778, + 0.022788869217038155, + 0.009165515191853046, + 0.15175777673721313, + 0.015832863748073578, + 0.010695978067815304, + 0.013292908668518066, + 0.015514054335653782, + 0.13597218692302704, + 0.002914649900048971, + 0.0022447104565799236, + 0.007254452910274267, + 0.0023677728604525328, + 0.002486150013282895, + 0.0027177068404853344, + 0.002380100777372718, + 0.007567146793007851, + 0.004917088896036148, + 0.00813044048845768, + 0.018682366237044334, + 0.019020842388272285, + 0.008629933930933475, + 0.007423628121614456, + 0.00960267148911953, + 0.014897479675710201, + 0.007217295467853546, + 0.01238729152828455, + 0.023246146738529205, + 0.0031627502758055925, + 0.009843493811786175, + 0.03175719827413559, + 0.0066237482242286205, + 0.03379969298839569, + 0.004874644801020622, + 0.006042480003088713, + 0.0040187225677073, + 0.004629223141819239, + 0.0047453283332288265, + 0.0059144156984984875, + 0.0031387393828481436, + 0.006289678160101175, + 0.0051400624215602875, + 0.003899950534105301, + 0.012623141519725323, + 0.005021435208618641, + 0.004106421954929829, + 0.004979939199984074, + 0.011394018307328224, + 0.007249142974615097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.0013438966125249863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069423407316208 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069423407316208, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069423407316208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04256267473101616, + 0.028481634333729744, + 0.08445186167955399, + 0.06153717637062073, + 0.3813444972038269, + 0.05807898938655853, + 0.056307390332221985, + 0.16041933000087738, + 0.032612308859825134, + 0.09420420974493027 + ], + "credit_map": [ + 0.005695451516658068, + 0.004478236194700003, + 0.007576708681881428, + 0.003619732102379203, + 0.005278783850371838, + 0.005187542177736759, + 0.0037893455009907484, + 0.006936869118362665, + 0.004768076818436384, + 0.002665933920070529, + 0.004521744791418314, + 0.0045065609738230705, + 0.0029258879367262125, + 0.00265832943841815, + 0.003128788201138377, + 0.0033063082955777645, + 0.0134966429322958, + 0.009683702141046524, + 0.010344354435801506, + 0.007598572410643101, + 0.011159975081682205, + 0.014426940120756626, + 0.004786834120750427, + 0.01295483112335205, + 0.00895988754928112, + 0.007252243347465992, + 0.00967719592154026, + 0.007035417016595602, + 0.006240698043256998, + 0.007196218241006136, + 0.005241896957159042, + 0.00993360672146082, + 0.0273455660790205, + 0.009152512066066265, + 0.15301038324832916, + 0.016507545486092567, + 0.010780805721879005, + 0.013159171678125858, + 0.015028071589767933, + 0.13636040687561035, + 0.0058221956714987755, + 0.004396709613502026, + 0.013146902434527874, + 0.004827586468309164, + 0.004646084271371365, + 0.00712719326838851, + 0.004347902722656727, + 0.013764403760433197, + 0.004365039523690939, + 0.006129600573331118, + 0.008924533613026142, + 0.013346400111913681, + 0.0048907604068517685, + 0.004497788846492767, + 0.007030486594885588, + 0.007122781127691269, + 0.009300301782786846, + 0.016140276566147804, + 0.028197498992085457, + 0.004205390810966492, + 0.012050903402268887, + 0.04014156386256218, + 0.00852744746953249, + 0.041855935007333755, + 0.004136261530220509, + 0.00497052725404501, + 0.003557759802788496, + 0.0040709893219172955, + 0.004034356214106083, + 0.004672566894441843, + 0.0028502398636192083, + 0.004319604020565748, + 0.009829547256231308, + 0.00515375193208456, + 0.0233156755566597, + 0.008821818046271801, + 0.006808881182223558, + 0.006632740143686533, + 0.018855372443795204, + 0.01478640828281641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.016069423407316208 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735833827406168 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735833827406168, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735833827406168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05552109703421593, + 0.05406824126839638, + 0.07021354138851166, + 0.1399075984954834, + 0.28440552949905396, + 0.10929589718580246, + 0.0435662604868412, + 0.041249994188547134, + 0.07098589092493057, + 0.1307859718799591 + ], + "credit_map": [ + 0.011769241653382778, + 0.005827045068144798, + 0.008456612937152386, + 0.005571283865720034, + 0.005525389686226845, + 0.005400693044066429, + 0.0057512023486196995, + 0.0072196233086287975, + 0.00866622943431139, + 0.007303975522518158, + 0.006669261027127504, + 0.0072808777913451195, + 0.008363492786884308, + 0.005525701679289341, + 0.004844976589083672, + 0.005413723178207874, + 0.014259538613259792, + 0.007200723979622126, + 0.009903205558657646, + 0.006099492311477661, + 0.006814748048782349, + 0.010467488318681717, + 0.007132464554160833, + 0.008335870690643787, + 0.01956474967300892, + 0.015342853032052517, + 0.024789104238152504, + 0.014768451452255249, + 0.012749910354614258, + 0.018070369958877563, + 0.009611092507839203, + 0.025011051446199417, + 0.0186164490878582, + 0.006802184507250786, + 0.11584679782390594, + 0.012322242371737957, + 0.0075497073121368885, + 0.00904625654220581, + 0.01095564290881157, + 0.10326622426509857, + 0.009868589229881763, + 0.0073569416999816895, + 0.02815696783363819, + 0.008730125613510609, + 0.007166699040681124, + 0.010277938097715378, + 0.008931329473853111, + 0.028807295486330986, + 0.0038268493954092264, + 0.003932733554393053, + 0.007362083997577429, + 0.009756422601640224, + 0.0045908596366643906, + 0.003278175136074424, + 0.005332987755537033, + 0.005486144218593836, + 0.005278076510876417, + 0.009250986389815807, + 0.004144736565649509, + 0.005953026469796896, + 0.004725210834294558, + 0.00458484748378396, + 0.003118751337751746, + 0.004194358363747597, + 0.006806913763284683, + 0.012006602250039577, + 0.008904406800866127, + 0.009815067052841187, + 0.008332776837050915, + 0.005585005506873131, + 0.005995078943669796, + 0.013540029525756836, + 0.013349742628633976, + 0.0072400993667542934, + 0.033889561891555786, + 0.012756822630763054, + 0.009794875048100948, + 0.007655604742467403, + 0.02631337381899357, + 0.019785888493061066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.006735833827406168 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, static shot, taken from a high angle, captures a tranquil lake, its surface a deep sapphire, shimmering with layers of delicate, slowly spreading ripples. In the center of the frame, a bright red maple leaf, its edges slightly curled and its veins clearly visible, floats solitary and gracefully. Soft, golden morning light slants through the frame from the side, filtering through the leaf, giving it a translucent quality and a golden halo around its edges, casting long, shimmering paths across the shimmering water. Gently propelled by the ripples, the leaf drifts across the water with an almost imperceptible, steady, and leisurely motion, occasionally accompanied by the slightest rotation. The water reflects the blurred, vibrant outlines of the autumn trees on the distant shore. The entire scene is enveloped in a tranquil, contemplative atmosphere, imbued with warm, saturated tones and imbued with a poetic beauty.", + "seed": 31510, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_3/index_975_seed_31510_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09860436352901161, + "vlm_logprob": -0.09860436352901161, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.09860436352901161, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18447993695735931 + ], + "opposite_logprob_tokens": [ + -1.9344799518585205 + ], + "logprob_reward": -0.18447993695735931, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.18447993695735931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15118099749088287, + 0.04527902230620384, + 0.04090747982263565, + 0.17808516323566437, + 0.09448326379060745, + 0.23981183767318726, + 0.07287245243787766, + 0.07404158264398575, + 0.05261607840657234, + 0.05072209984064102 + ], + "credit_map": [ + 0.01734466850757599, + 0.013468759134411812, + 0.027562817558646202, + 0.015014716424047947, + 0.012110200710594654, + 0.03672380372881889, + 0.012435950338840485, + 0.016520075500011444, + 0.005571563262492418, + 0.004074696451425552, + 0.007720517925918102, + 0.005444783717393875, + 0.004059584345668554, + 0.008738991804420948, + 0.004588490817695856, + 0.005080395843833685, + 0.0045680576004087925, + 0.00417294492945075, + 0.006328359711915255, + 0.005168382078409195, + 0.004255359526723623, + 0.00745092798024416, + 0.004525330848991871, + 0.004438117146492004, + 0.0154152512550354, + 0.016882166266441345, + 0.04536466673016548, + 0.03269116207957268, + 0.016050564125180244, + 0.011018400080502033, + 0.023433690890669823, + 0.01722925528883934, + 0.011274637654423714, + 0.009386594407260418, + 0.021191002801060677, + 0.01337977685034275, + 0.007198047358542681, + 0.009502551518380642, + 0.012641367502510548, + 0.009909274987876415, + 0.026965148746967316, + 0.02802327089011669, + 0.05094682052731514, + 0.03738824278116226, + 0.025618817657232285, + 0.01932583563029766, + 0.025523558259010315, + 0.02602013386785984, + 0.008439858444035053, + 0.008885691873729229, + 0.014836222864687443, + 0.012012855149805546, + 0.0054484885185956955, + 0.008349954150617123, + 0.006525224540382624, + 0.008374163880944252, + 0.010286370292305946, + 0.008728559128940105, + 0.014239384792745113, + 0.010458996519446373, + 0.0057331775315105915, + 0.00811261311173439, + 0.009082082659006119, + 0.0074004014022648335, + 0.010464893653988838, + 0.007348479703068733, + 0.008407683111727238, + 0.00616066949442029, + 0.004755139350891113, + 0.006223189644515514, + 0.0046311551705002785, + 0.004624870140105486, + 0.009619085118174553, + 0.007750790566205978, + 0.006680633407086134, + 0.004774041939526796, + 0.0035291968379169703, + 0.009980620816349983, + 0.003481396706774831, + 0.004906335845589638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.18447993695735931 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814011573791504 + ], + "opposite_logprob_tokens": [ + -1.4064011573791504 + ], + "logprob_reward": -0.2814011573791504, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814011573791504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17412951588630676, + 0.18463531136512756, + 0.057355597615242004, + 0.16234983503818512, + 0.07687359303236008, + 0.13818268477916718, + 0.037635087966918945, + 0.07569000124931335, + 0.0507083497941494, + 0.04243997484445572 + ], + "credit_map": [ + 0.019796041771769524, + 0.01630181074142456, + 0.045582812279462814, + 0.017574789002537727, + 0.020134203135967255, + 0.024873297661542892, + 0.01294845063239336, + 0.0169181227684021, + 0.021299418061971664, + 0.016196278855204582, + 0.02755795419216156, + 0.028473487123847008, + 0.018110625445842743, + 0.03449974209070206, + 0.01835646480321884, + 0.020141350105404854, + 0.006217318121343851, + 0.0072629391215741634, + 0.01281378511339426, + 0.008231397718191147, + 0.006037500686943531, + 0.006845682393759489, + 0.004488183651119471, + 0.005458790808916092, + 0.015365837141871452, + 0.015626568347215652, + 0.03691745549440384, + 0.028561033308506012, + 0.018348803743720055, + 0.010061372071504593, + 0.022023096680641174, + 0.01544567197561264, + 0.009334917180240154, + 0.0062453714199364185, + 0.013846341520547867, + 0.011971213854849339, + 0.00986365508288145, + 0.00782228633761406, + 0.01039139460772276, + 0.007398413494229317, + 0.01630866341292858, + 0.015042031183838844, + 0.030513273552060127, + 0.020173465833067894, + 0.015696991235017776, + 0.011839537881314754, + 0.015060358680784702, + 0.013548377901315689, + 0.003407028503715992, + 0.003624814096838236, + 0.00877638440579176, + 0.007529043592512608, + 0.003001293633133173, + 0.002653001341968775, + 0.004642288200557232, + 0.004001236986368895, + 0.008290262892842293, + 0.007294174283742905, + 0.015065141022205353, + 0.010191473178565502, + 0.00553399883210659, + 0.014505957253277302, + 0.007234967779368162, + 0.00757402740418911, + 0.011455051600933075, + 0.008362273685634136, + 0.006250142119824886, + 0.00640038400888443, + 0.005087160039693117, + 0.004432883113622665, + 0.0039992243982851505, + 0.0047212326899170876, + 0.006092424970120192, + 0.006358257494866848, + 0.006753544323146343, + 0.006186518352478743, + 0.003905403893440962, + 0.0059564062394201756, + 0.0031611432787030935, + 0.00402627931907773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.2814011573791504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002209009835496545 + ], + "opposite_logprob_tokens": [ + -6.127209186553955 + ], + "logprob_reward": -0.002209009835496545, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002209009835496545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10566109418869019, + 0.05994360148906708, + 0.0649062916636467, + 0.17446961998939514, + 0.07270288467407227, + 0.10231991112232208, + 0.08748923987150192, + 0.07801318913698196, + 0.06407079845666885, + 0.19042333960533142 + ], + "credit_map": [ + 0.013399056158959866, + 0.009432956576347351, + 0.02332908660173416, + 0.01020861603319645, + 0.0095117362216115, + 0.021246694028377533, + 0.007016231305897236, + 0.011516728438436985, + 0.006229697726666927, + 0.0055406177416443825, + 0.009233224205672741, + 0.009763376787304878, + 0.0056852856650948524, + 0.009845213033258915, + 0.007387874647974968, + 0.006258315406739712, + 0.006554079707711935, + 0.006414402276277542, + 0.015299496240913868, + 0.007342384196817875, + 0.007382851559668779, + 0.01097601093351841, + 0.004673027899116278, + 0.006264045834541321, + 0.016506079584360123, + 0.01511175837367773, + 0.05524919927120209, + 0.03406183794140816, + 0.011844650842249393, + 0.010532815009355545, + 0.01217737514525652, + 0.01898592710494995, + 0.007748553995043039, + 0.006198981311172247, + 0.012975129298865795, + 0.012661756947636604, + 0.00848681852221489, + 0.0056915623135864735, + 0.011260457336902618, + 0.007679621689021587, + 0.009622715413570404, + 0.010187597945332527, + 0.010031674057245255, + 0.012767200358211994, + 0.018330834805965424, + 0.008869089186191559, + 0.02132721059024334, + 0.011183594353497028, + 0.008956714533269405, + 0.00887532252818346, + 0.01702239364385605, + 0.02040506899356842, + 0.006028881762176752, + 0.008781504817306995, + 0.00827507022768259, + 0.009144283831119537, + 0.00763916689902544, + 0.007702674251049757, + 0.011503905057907104, + 0.013644233345985413, + 0.0062999967485666275, + 0.014955580234527588, + 0.008503437973558903, + 0.007764200679957867, + 0.012800933793187141, + 0.009225182235240936, + 0.0079254861921072, + 0.006388370878994465, + 0.005754489451646805, + 0.008809917606413364, + 0.004961390979588032, + 0.008205033838748932, + 0.017576131969690323, + 0.02140078879892826, + 0.021915586665272713, + 0.025396423414349556, + 0.024808049201965332, + 0.02645600214600563, + 0.035107340663671494, + 0.017763040959835052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.002209009835496545 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819327473640442 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.01819327473640442, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819327473640442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1908223181962967, + 0.03040730580687523, + 0.04838256165385246, + 0.24026557803153992, + 0.05516651272773743, + 0.10136377066373825, + 0.05025003105401993, + 0.13581109046936035, + 0.03115592896938324, + 0.1163749247789383 + ], + "credit_map": [ + 0.024784458801150322, + 0.017674235627055168, + 0.04175322502851486, + 0.016399506479501724, + 0.02233615145087242, + 0.03496392443776131, + 0.013547016307711601, + 0.01936381123960018, + 0.0031210598535835743, + 0.0030835282523185015, + 0.0041872127912938595, + 0.004494394175708294, + 0.003314688103273511, + 0.004931706003844738, + 0.004096742253750563, + 0.0031779755372554064, + 0.005481046624481678, + 0.005157342180609703, + 0.011381976306438446, + 0.005872391629964113, + 0.004456488415598869, + 0.007063886150717735, + 0.00422104075551033, + 0.004748390521854162, + 0.02306658960878849, + 0.021975716575980186, + 0.07250995934009552, + 0.04303581640124321, + 0.02050909399986267, + 0.014069142751395702, + 0.020965786650776863, + 0.024133482947945595, + 0.007833924144506454, + 0.005714531522244215, + 0.011974372901022434, + 0.008927933871746063, + 0.004373288247734308, + 0.005397901404649019, + 0.005167842376977205, + 0.005776720587164164, + 0.01060307677835226, + 0.009232211858034134, + 0.022276049479842186, + 0.01590278372168541, + 0.012955276295542717, + 0.008598602376878262, + 0.011165151372551918, + 0.010630614124238491, + 0.004883485846221447, + 0.006006317213177681, + 0.009940912015736103, + 0.009724970906972885, + 0.004761777818202972, + 0.0053861672058701515, + 0.00436625350266695, + 0.005180146545171738, + 0.016959326341748238, + 0.015219162218272686, + 0.024720558896660805, + 0.019249776378273964, + 0.009852428920567036, + 0.02023020200431347, + 0.016189543530344963, + 0.013390100561082363, + 0.007023834157735109, + 0.005073934327811003, + 0.004027020186185837, + 0.0027906105387955904, + 0.0030412892811000347, + 0.0026362682692706585, + 0.0025971708819270134, + 0.003965803422033787, + 0.010193286463618279, + 0.011395166628062725, + 0.019535033032298088, + 0.012937016785144806, + 0.013264330103993416, + 0.01861027628183365, + 0.017921028658747673, + 0.012518789619207382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.01819327473640442 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface a deep sapphire color with visible ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright red maple leaf with visible veins and curled edges floating in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does soft golden morning light slant from the side, creating a translucent quality and golden halo on the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf drift with a steady, leisurely motion accompanied by slight rotation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067384387366473675 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.0067384387366473675, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067384387366473675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14659738540649414, + 0.11815638095140457, + 0.050645843148231506, + 0.19928699731826782, + 0.11680097132921219, + 0.06535226851701736, + 0.0838448777794838, + 0.04609398916363716, + 0.03333062306046486, + 0.139890655875206 + ], + "credit_map": [ + 0.014504404738545418, + 0.012438981793820858, + 0.022956375032663345, + 0.01583193987607956, + 0.013307033106684685, + 0.03476677089929581, + 0.016671795397996902, + 0.016120078042149544, + 0.01055083703249693, + 0.010453673079609871, + 0.014196149073541164, + 0.026118244975805283, + 0.010150058194994926, + 0.013913707807660103, + 0.019813841208815575, + 0.012959854677319527, + 0.005244515836238861, + 0.006034767720848322, + 0.0062891636043787, + 0.009202252142131329, + 0.0075636813417077065, + 0.005000926088541746, + 0.006256970576941967, + 0.005053563043475151, + 0.019676167517900467, + 0.021283533424139023, + 0.051444463431835175, + 0.035998690873384476, + 0.019799865782260895, + 0.012853325344622135, + 0.019924048334360123, + 0.01830688677728176, + 0.013633325695991516, + 0.010948735289275646, + 0.02456376701593399, + 0.017709307372570038, + 0.012343522161245346, + 0.01395140029489994, + 0.011016156524419785, + 0.012634755112230778, + 0.005744480527937412, + 0.007528833579272032, + 0.0077695720829069614, + 0.008860635571181774, + 0.00892096571624279, + 0.006371877156198025, + 0.013544145040214062, + 0.006611752789467573, + 0.008214197121560574, + 0.009936504065990448, + 0.015850801020860672, + 0.017264708876609802, + 0.007037496659904718, + 0.008011184632778168, + 0.008775693364441395, + 0.008754280395805836, + 0.00758395716547966, + 0.005522562190890312, + 0.01012678723782301, + 0.006172103341668844, + 0.0029344067443162203, + 0.005335621070116758, + 0.0036928446497768164, + 0.0047257025726139545, + 0.005682294722646475, + 0.004699057433754206, + 0.0037799335550516844, + 0.0043226876296103, + 0.003663286566734314, + 0.00442525465041399, + 0.00386532093398273, + 0.002892785705626011, + 0.02637510932981968, + 0.02179829776287079, + 0.020463112741708755, + 0.014842409640550613, + 0.013785511255264282, + 0.015435120090842247, + 0.015059798024594784, + 0.012131284922361374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.0067384387366473675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_4/index_tensor([171, 171])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_4/index_tensor([171, 171])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..55398f6f1c9126e6790e5265625935cb9b672e5b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_4/index_tensor([171, 171])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, high-angle macro shot, rendered in breathtakingly realistic detail, begins with sharp focus fixed on a perfectly formed, branching snowflake. This solitary crystal, with six intricate, symmetrical arms branching into delicate, feathery structures, sits precariously on the dark, grainy asphalt surface. The camera then executes a slow, deliberate rack focus, steadily pulling the plane of sharpness away from the delicate snowflake, which melts into a soft blur. The focus propagates downward, penetrating the surface to reveal the perilous crystalline structure of the black ice beneath. This newly focused layer, a thin, glossy flake, captures the dark hues of the road below, punctuated by shard-like cracks and frozen bubbles trapped within their transparent prisons. As this shift occurs, light snowflakes continue to gently move across the frame, appearing as soft, out-of-focus bokeh elements, adding depth and movement. The entire scene is bathed in the cold, unfocused light of an overcast sky, with countless icy surfaces shimmering with subtle light, creating a striking contrast between the rough asphalt texture and the smooth, treacherous ice. The cool, monochromatic palette of white, grey, and deep black creates a powerful atmosphere of both fragile, microscopic beauty and palpable, imminent danger.", + "seed": 8595, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_4/index_171_seed_8595_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01409244816750288, + "vlm_logprob": -0.01409244816750288, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.01409244816750288, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019767194986343384 + ], + "opposite_logprob_tokens": [ + -4.1447672843933105 + ], + "logprob_reward": -0.019767194986343384, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019767194986343384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3275772035121918, + 0.05634448677301407, + 0.12331609427928925, + 0.22121933102607727, + 0.04392410069704056, + 0.10017175227403641, + 0.021451210603117943, + 0.046677760779857635, + 0.028954561799764633, + 0.03036349080502987 + ], + "credit_map": [ + 0.06175054609775543, + 0.04045794904232025, + 0.03519096225500107, + 0.04144303873181343, + 0.032606277614831924, + 0.037110112607479095, + 0.038243021816015244, + 0.04077530279755592, + 0.006481332238763571, + 0.008276264183223248, + 0.0044297282584011555, + 0.0049684359692037106, + 0.008732699789106846, + 0.008494531735777855, + 0.008281190879642963, + 0.006680300924926996, + 0.006773088593035936, + 0.005851530469954014, + 0.03440634533762932, + 0.012412924319505692, + 0.006833880208432674, + 0.011044901795685291, + 0.009397998452186584, + 0.0365954264998436, + 0.023327305912971497, + 0.02772810496389866, + 0.019590795040130615, + 0.02494722232222557, + 0.05542965978384018, + 0.034195300191640854, + 0.017532479017972946, + 0.01846848428249359, + 0.005426391959190369, + 0.005138607695698738, + 0.004594151861965656, + 0.004795010667294264, + 0.005737598519772291, + 0.0043026236817240715, + 0.006543781608343124, + 0.00738593889400363, + 0.006632927339524031, + 0.03239187225699425, + 0.006501070689409971, + 0.005892087705433369, + 0.008136573247611523, + 0.028724167495965958, + 0.005273314192891121, + 0.006619738880544901, + 0.002996364375576377, + 0.0026470625307410955, + 0.002022998873144388, + 0.0027339435182511806, + 0.0036513765808194876, + 0.002741291420534253, + 0.002501225098967552, + 0.0021569496020674706, + 0.007131406106054783, + 0.005018661264330149, + 0.0047563123516738415, + 0.004568388219922781, + 0.0035988546442240477, + 0.01285287644714117, + 0.003978978842496872, + 0.004772278480231762, + 0.0034917080774903297, + 0.0032281370367854834, + 0.0032275987323373556, + 0.0030735761392861605, + 0.0035285879857838154, + 0.005019711330533028, + 0.0031093519646674395, + 0.004275889601558447, + 0.0038027914706617594, + 0.004301152657717466, + 0.004517509136348963, + 0.003944215830415487, + 0.003006428014487028, + 0.0031253965571522713, + 0.004064411856234074, + 0.003601586213335395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.019767194986343384 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813838213682175 + ], + "opposite_logprob_tokens": [ + -3.2881383895874023 + ], + "logprob_reward": -0.03813838213682175, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813838213682175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1310604363679886, + 0.21966905891895294, + 0.04711942747235298, + 0.06273540109395981, + 0.04649980738759041, + 0.14132621884346008, + 0.08298386633396149, + 0.2148633897304535, + 0.022499021142721176, + 0.031243404373526573 + ], + "credit_map": [ + 0.015470640733838081, + 0.01627602055668831, + 0.01668146811425686, + 0.015326505526900291, + 0.014674206264317036, + 0.02012685313820839, + 0.015075672417879105, + 0.017429068684577942, + 0.023469246923923492, + 0.03091093711555004, + 0.024468492716550827, + 0.02390740066766739, + 0.02394641563296318, + 0.027484647929668427, + 0.04026361554861069, + 0.0252182949334383, + 0.0028146032709628344, + 0.004246036987751722, + 0.009371254593133926, + 0.005266890395432711, + 0.004356940742582083, + 0.005020109470933676, + 0.003876935923472047, + 0.012166657485067844, + 0.006996180396527052, + 0.007458664011210203, + 0.006125281564891338, + 0.011313332244753838, + 0.007059326861053705, + 0.007843924686312675, + 0.01011121366173029, + 0.005827471613883972, + 0.0043660905212163925, + 0.004358526319265366, + 0.005334966350346804, + 0.005967424716800451, + 0.0041272277012467384, + 0.00607248954474926, + 0.008643233217298985, + 0.007629848551005125, + 0.011066191829741001, + 0.04602057486772537, + 0.007418843451887369, + 0.008970570750534534, + 0.008880672976374626, + 0.04081708565354347, + 0.010532890446484089, + 0.007619386073201895, + 0.009034281596541405, + 0.010419332422316074, + 0.0076568592339754105, + 0.012120847590267658, + 0.012455961667001247, + 0.009577510878443718, + 0.010156354866921902, + 0.011562720872461796, + 0.03143306076526642, + 0.022333472967147827, + 0.019000055268406868, + 0.01674390211701393, + 0.021359892562031746, + 0.06660620868206024, + 0.014558347873389721, + 0.02282845973968506, + 0.0030428324826061726, + 0.005359541159123182, + 0.002113955095410347, + 0.0022425497882068157, + 0.0020265730563551188, + 0.0024256273172795773, + 0.001703942776657641, + 0.003584000514820218, + 0.0030745395924896, + 0.003350227838382125, + 0.004112853202968836, + 0.005245410371571779, + 0.003855138784274459, + 0.003826071973890066, + 0.003477768274024129, + 0.00430139247328043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.03813838213682175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960191134363413 + ], + "opposite_logprob_tokens": [ + -5.130959987640381 + ], + "logprob_reward": -0.005960191134363413, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960191134363413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09742404520511627, + 0.05805737525224686, + 0.05786021426320076, + 0.16445863246917725, + 0.07669460773468018, + 0.29708606004714966, + 0.09624514728784561, + 0.07326396554708481, + 0.03477692976593971, + 0.04413299262523651 + ], + "credit_map": [ + 0.018211059272289276, + 0.012767408043146133, + 0.011355770751833916, + 0.011902205646038055, + 0.00917954184114933, + 0.011560052633285522, + 0.009046923369169235, + 0.0134010910987854, + 0.005773680284619331, + 0.008163553662598133, + 0.005791036412119865, + 0.00810085516422987, + 0.007033828180283308, + 0.006799479015171528, + 0.010357157327234745, + 0.006037787534296513, + 0.0044336942955851555, + 0.008188988082110882, + 0.006013320758938789, + 0.008681733161211014, + 0.00971581507474184, + 0.007074836641550064, + 0.007099812384694815, + 0.006652018986642361, + 0.019744085147976875, + 0.022582421079277992, + 0.014978021383285522, + 0.01731974072754383, + 0.033329788595438004, + 0.02221762016415596, + 0.02045578695833683, + 0.013831177726387978, + 0.008327268064022064, + 0.0078043569810688496, + 0.007354320026934147, + 0.01081706304103136, + 0.005774920806288719, + 0.007831539958715439, + 0.009054094552993774, + 0.019731052219867706, + 0.01736431196331978, + 0.09976667165756226, + 0.018857255578041077, + 0.017377670854330063, + 0.018887165933847427, + 0.09131179004907608, + 0.0142362080514431, + 0.019285013899207115, + 0.009811321273446083, + 0.01080744992941618, + 0.009998959489166737, + 0.01709853671491146, + 0.010525177232921124, + 0.012811743654310703, + 0.013648002408444881, + 0.011543959379196167, + 0.010544844903051853, + 0.007014095317572355, + 0.006631277501583099, + 0.006397255230695009, + 0.0071405027993023396, + 0.01852147839963436, + 0.006530307698994875, + 0.010484207421541214, + 0.0029575861990451813, + 0.005958007648587227, + 0.0033614628482609987, + 0.003291879314929247, + 0.008574653416872025, + 0.0045297108590602875, + 0.0030473012011498213, + 0.003056332701817155, + 0.0037037774454802275, + 0.0044814529828727245, + 0.007444960065186024, + 0.0066374228335917, + 0.005422303918749094, + 0.005669072736054659, + 0.005330819170922041, + 0.005443187430500984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.005960191134363413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002495629247277975 + ], + "opposite_logprob_tokens": [ + -6.002495765686035 + ], + "logprob_reward": -0.002495629247277975, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002495629247277975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23975719511508942, + 0.06313689053058624, + 0.03406883031129837, + 0.07847910374403, + 0.031635962426662445, + 0.25092947483062744, + 0.11446334421634674, + 0.08982770144939423, + 0.05471644550561905, + 0.04298511520028114 + ], + "credit_map": [ + 0.04584065079689026, + 0.02636735513806343, + 0.02422606758773327, + 0.024829264730215073, + 0.017231428995728493, + 0.036653950810432434, + 0.021202847361564636, + 0.043405622243881226, + 0.007550765294581652, + 0.008626826107501984, + 0.005922135431319475, + 0.0071939509361982346, + 0.008076738566160202, + 0.009021151810884476, + 0.008813554421067238, + 0.007931768894195557, + 0.002884256886318326, + 0.003989508841186762, + 0.005182261578738689, + 0.005139724351465702, + 0.003455342957749963, + 0.004257577937096357, + 0.0050762430764734745, + 0.0040839179418981075, + 0.006503131240606308, + 0.00648039486259222, + 0.011163544841110706, + 0.017547622323036194, + 0.009049647487699986, + 0.009221628308296204, + 0.01058829016983509, + 0.007924850098788738, + 0.0036585538182407618, + 0.003199812723323703, + 0.003638216760009527, + 0.0042850561439991, + 0.0017766341334208846, + 0.0030990794766694307, + 0.004991423338651657, + 0.006987186148762703, + 0.012072132900357246, + 0.09033426642417908, + 0.00889826100319624, + 0.012558670714497566, + 0.013601834885776043, + 0.08982405811548233, + 0.01120012253522873, + 0.012440122663974762, + 0.01043438259512186, + 0.013650398701429367, + 0.010539883747696877, + 0.01405786070972681, + 0.015261752530932426, + 0.020207779482007027, + 0.013241879642009735, + 0.01706940121948719, + 0.012640619650483131, + 0.008406086824834347, + 0.009930628351867199, + 0.007852708920836449, + 0.00902273878455162, + 0.0211385115981102, + 0.008381782099604607, + 0.012454619631171227, + 0.006028848700225353, + 0.007024912629276514, + 0.008200407959520817, + 0.0055098445154726505, + 0.008277898654341698, + 0.007892782799899578, + 0.004458526149392128, + 0.007323221303522587, + 0.0041593830101192, + 0.005425460636615753, + 0.0058785793371498585, + 0.005305486731231213, + 0.006338816601783037, + 0.004912912845611572, + 0.00463498942553997, + 0.006329486146569252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.002495629247277975 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100843332707882 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.004100843332707882, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100843332707882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21453583240509033, + 0.0488041453063488, + 0.04674430936574936, + 0.08783535659313202, + 0.08410106599330902, + 0.37327471375465393, + 0.0383114330470562, + 0.04211746156215668, + 0.025032147765159607, + 0.039243634790182114 + ], + "credit_map": [ + 0.04915660247206688, + 0.026036253198981285, + 0.0225235465914011, + 0.024354876950383186, + 0.01800294779241085, + 0.030299266800284386, + 0.02256566658616066, + 0.021596666425466537, + 0.0062231868505477905, + 0.006050868425518274, + 0.005584061611443758, + 0.005387457087635994, + 0.0053261336870491505, + 0.0059792953543365, + 0.007413532584905624, + 0.006839602719992399, + 0.0035824927035719156, + 0.0031028706580400467, + 0.012010300531983376, + 0.0054717641323804855, + 0.0030523096211254597, + 0.005081052891910076, + 0.004064236301928759, + 0.010379280894994736, + 0.008077154867351055, + 0.0069450694136321545, + 0.011585023254156113, + 0.022006409242749214, + 0.0075102029368281364, + 0.0111464262008667, + 0.010414701886475086, + 0.010150361806154251, + 0.008863922208547592, + 0.007182790897786617, + 0.007973918691277504, + 0.009053711779415607, + 0.009151904843747616, + 0.008802254684269428, + 0.011629425920546055, + 0.021443121135234833, + 0.017345676198601723, + 0.1421569585800171, + 0.014201619662344456, + 0.017570583149790764, + 0.02274603769183159, + 0.12893453240394592, + 0.013235977850854397, + 0.01708330027759075, + 0.0036818175576627254, + 0.004853647202253342, + 0.0039395918138325214, + 0.005722870118916035, + 0.005387160461395979, + 0.005909633822739124, + 0.004681451246142387, + 0.004135253373533487, + 0.005494432058185339, + 0.004420713055878878, + 0.003363610478118062, + 0.004010264296084642, + 0.00621796352788806, + 0.010496397502720356, + 0.003237130120396614, + 0.004876946099102497, + 0.0028951477725058794, + 0.0033336642663925886, + 0.002714412985369563, + 0.002845069859176874, + 0.003748843679204583, + 0.003934279549866915, + 0.0026704133488237858, + 0.0028903125785291195, + 0.003641944145783782, + 0.004434601869434118, + 0.006384596228599548, + 0.005844819359481335, + 0.004235736094415188, + 0.003994232974946499, + 0.005632996093481779, + 0.005074703600257635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.004100843332707882 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle macro shot, rendered in breathtakingly realistic detail, begins with sharp focus fixed on a perfectly formed, branching snowflake. This solitary crystal, with six intricate, symmetrical arms branching into delicate, feathery structures, sits precariously on the dark, grainy asphalt surface. The camera then executes a slow, deliberate rack focus, steadily pulling the plane of sharpness away from the delicate snowflake, which melts into a soft blur. The focus propagates downward, penetrating the surface to reveal the perilous crystalline structure of the black ice beneath. This newly focused layer, a thin, glossy flake, captures the dark hues of the road below, punctuated by shard-like cracks and frozen bubbles trapped within their transparent prisons. As this shift occurs, light snowflakes continue to gently move across the frame, appearing as soft, out-of-focus bokeh elements, adding depth and movement. The entire scene is bathed in the cold, unfocused light of an overcast sky, with countless icy surfaces shimmering with subtle light, creating a striking contrast between the rough asphalt texture and the smooth, treacherous ice. The cool, monochromatic palette of white, grey, and deep black creates a powerful atmosphere of both fragile, microscopic beauty and palpable, imminent danger.", + "seed": 8596, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_4/index_171_seed_8596_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.011868412722833455, + "vlm_logprob": -0.011868412722833455, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.011868412722833455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013593487441539764 + ], + "opposite_logprob_tokens": [ + -4.513593673706055 + ], + "logprob_reward": -0.013593487441539764, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013593487441539764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15500296652317047, + 0.07407832890748978, + 0.027974331751465797, + 0.1278086155653, + 0.04537798836827278, + 0.05770112946629524, + 0.04511362686753273, + 0.026859747245907784, + 0.04779743775725365, + 0.39228588342666626 + ], + "credit_map": [ + 0.022177835926413536, + 0.02528299391269684, + 0.018904564902186394, + 0.01770474761724472, + 0.022069962695240974, + 0.009744365699589252, + 0.028724774718284607, + 0.010393708944320679, + 0.011279724538326263, + 0.010016596876084805, + 0.005381298717111349, + 0.005663999821990728, + 0.016253946349024773, + 0.008379338309168816, + 0.00728895328938961, + 0.009814461693167686, + 0.003613996785134077, + 0.002866286551579833, + 0.0037994487211108208, + 0.0045393225736916065, + 0.002684962237253785, + 0.0033681513741612434, + 0.0035552431363612413, + 0.0035469175782054663, + 0.018724236637353897, + 0.009892554022371769, + 0.04332008957862854, + 0.012076649814844131, + 0.009640052914619446, + 0.010050232522189617, + 0.015679365023970604, + 0.008425431326031685, + 0.002984942402690649, + 0.003403489477932453, + 0.008576547726988792, + 0.0033119877334684134, + 0.002677071839570999, + 0.004011847078800201, + 0.015580248087644577, + 0.004831847734749317, + 0.003677734872326255, + 0.004143749363720417, + 0.014419570565223694, + 0.004133118316531181, + 0.007304154336452484, + 0.009047114290297031, + 0.00493402685970068, + 0.010041655972599983, + 0.003149721072986722, + 0.004147712141275406, + 0.0074513875879347324, + 0.006262156646698713, + 0.0035488675348460674, + 0.005312708672136068, + 0.00639008404687047, + 0.008850986137986183, + 0.0022341206204146147, + 0.0026166874449700117, + 0.0032741688191890717, + 0.004434607923030853, + 0.00402434216812253, + 0.0025877216830849648, + 0.0026254095137119293, + 0.005062684882432222, + 0.0035291833337396383, + 0.00480266846716404, + 0.006116978824138641, + 0.010948149487376213, + 0.008162974379956722, + 0.00426838081330061, + 0.00507856672629714, + 0.004890532232820988, + 0.019123230129480362, + 0.02634124830365181, + 0.011039568111300468, + 0.015959883108735085, + 0.16149339079856873, + 0.12581956386566162, + 0.021508224308490753, + 0.011000731959939003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.013593487441539764 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811003640294075 + ], + "opposite_logprob_tokens": [ + -3.2881100177764893 + ], + "logprob_reward": -0.03811003640294075, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811003640294075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0678754597902298, + 0.019885357469320297, + 0.01921614445745945, + 0.06452453136444092, + 0.051062390208244324, + 0.06524848192930222, + 0.025118038058280945, + 0.02526617795228958, + 0.032822877168655396, + 0.6289805173873901 + ], + "credit_map": [ + 0.009116902016103268, + 0.012628169730305672, + 0.004824556410312653, + 0.007334759458899498, + 0.011106787249445915, + 0.004506048280745745, + 0.013789941556751728, + 0.004568296484649181, + 0.0033550525549799204, + 0.002709815511479974, + 0.0018124234629794955, + 0.0021338164806365967, + 0.0024336681235581636, + 0.003386814147233963, + 0.0014445955166593194, + 0.002609173534438014, + 0.003346799872815609, + 0.002431220607832074, + 0.0023432292509824038, + 0.00308399205096066, + 0.0019046379020437598, + 0.0020308145321905613, + 0.0026480858214199543, + 0.0014273679116740823, + 0.008515255525708199, + 0.0041984315030276775, + 0.023366078734397888, + 0.00566754350438714, + 0.004855306353420019, + 0.003166396636515856, + 0.011562589555978775, + 0.003192938631400466, + 0.005135668441653252, + 0.007832104340195656, + 0.006700995843857527, + 0.0063279177993535995, + 0.003004412166774273, + 0.0059120203368365765, + 0.009861387312412262, + 0.006287887692451477, + 0.004808555357158184, + 0.006547888740897179, + 0.0116591677069664, + 0.004140265751630068, + 0.01904962584376335, + 0.009130671620368958, + 0.005005361046642065, + 0.0049069528467953205, + 0.0016935523599386215, + 0.002236497588455677, + 0.0027566549833863974, + 0.004086799453943968, + 0.003070040140300989, + 0.0029642628505825996, + 0.004927698988467455, + 0.0033825363498181105, + 0.0019867082592099905, + 0.002383897313848138, + 0.0033821763936430216, + 0.002895553596317768, + 0.0031646559946238995, + 0.00261008576489985, + 0.002616665093228221, + 0.0062264385633170605, + 0.002382502658292651, + 0.003999785520136356, + 0.003355380380526185, + 0.004150805529206991, + 0.005305906292051077, + 0.004192223772406578, + 0.005270238500088453, + 0.004166038241237402, + 0.028494657948613167, + 0.04992421716451645, + 0.012374958023428917, + 0.023375438526272774, + 0.28080034255981445, + 0.18673120439052582, + 0.02947414666414261, + 0.017805544659495354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.03811003640294075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040998938493430614 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.0040998938493430614, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040998938493430614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10320247709751129, + 0.06641442328691483, + 0.1006660982966423, + 0.10203566402196884, + 0.07014629989862442, + 0.14101749658584595, + 0.08366464823484421, + 0.05793972685933113, + 0.0917687714099884, + 0.18314436078071594 + ], + "credit_map": [ + 0.01349731907248497, + 0.01731695793569088, + 0.016945622861385345, + 0.00885017216205597, + 0.015273427590727806, + 0.011677581816911697, + 0.011960138566792011, + 0.007681252900511026, + 0.014693278819322586, + 0.004982607439160347, + 0.004894079640507698, + 0.004414327908307314, + 0.005982400383800268, + 0.007634551264345646, + 0.004626497160643339, + 0.019186677411198616, + 0.015908898785710335, + 0.010122671723365784, + 0.012976394966244698, + 0.02146054618060589, + 0.006350539159029722, + 0.009778431616723537, + 0.01323647890239954, + 0.01083214022219181, + 0.016030989587306976, + 0.006173380650579929, + 0.037432823330163956, + 0.009106803685426712, + 0.004520114045590162, + 0.006110484711825848, + 0.015794849023222923, + 0.006866225972771645, + 0.004760556388646364, + 0.011085801757872105, + 0.011352283880114555, + 0.0066817328333854675, + 0.003986796364188194, + 0.0034961532801389694, + 0.022298602387309074, + 0.0064843809232115746, + 0.01096578873693943, + 0.013258970342576504, + 0.021276423707604408, + 0.010982140898704529, + 0.03135542571544647, + 0.02334103360772133, + 0.010843933559954166, + 0.018993768841028214, + 0.0054129138588905334, + 0.006159187760204077, + 0.02421020157635212, + 0.008117424324154854, + 0.006955859251320362, + 0.00838342122733593, + 0.008247019723057747, + 0.016178615391254425, + 0.004504946526139975, + 0.007939307950437069, + 0.005980642046779394, + 0.0063635618425905704, + 0.0077534751035273075, + 0.008065816015005112, + 0.005215100944042206, + 0.012116873636841774, + 0.006979983299970627, + 0.010313973762094975, + 0.010623114183545113, + 0.019081516191363335, + 0.01463186927139759, + 0.00771138072013855, + 0.010800620540976524, + 0.011626306921243668, + 0.010299809277057648, + 0.017228880897164345, + 0.004977902863174677, + 0.008398812264204025, + 0.07515785843133926, + 0.026578543707728386, + 0.010050022043287754, + 0.03045252524316311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0040998938493430614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010449193650856614 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010449193650856614, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010449193650856614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06364437937736511, + 0.027956802397966385, + 0.022579791024327278, + 0.1457279771566391, + 0.09992269426584244, + 0.11936712265014648, + 0.014291657134890556, + 0.05349148064851761, + 0.10262278467416763, + 0.35039523243904114 + ], + "credit_map": [ + 0.00918611604720354, + 0.012475967407226562, + 0.005106756929308176, + 0.006577357184141874, + 0.01050602551549673, + 0.005227372050285339, + 0.01080059539526701, + 0.003764197463169694, + 0.0052813380025327206, + 0.0032497411593794823, + 0.0018830365734174848, + 0.0033003445714712143, + 0.0037359425332397223, + 0.002697448944672942, + 0.002548464573919773, + 0.005260487552732229, + 0.0031943642534315586, + 0.0034283006098121405, + 0.0027586515061557293, + 0.0026207834016531706, + 0.002115116221830249, + 0.003295997390523553, + 0.002674596616998315, + 0.0024919805582612753, + 0.021569307893514633, + 0.007338968571275473, + 0.05997500568628311, + 0.010994109325110912, + 0.006237276829779148, + 0.007233321201056242, + 0.025059737265110016, + 0.007320260629057884, + 0.01624736189842224, + 0.02139876037836075, + 0.013275121338665485, + 0.00734303705394268, + 0.004812778439372778, + 0.006630212999880314, + 0.020679505541920662, + 0.00953591987490654, + 0.00892354641109705, + 0.007772648241370916, + 0.018489371985197067, + 0.008798914030194283, + 0.02876710705459118, + 0.012605255469679832, + 0.010954427532851696, + 0.02305585704743862, + 0.0011557452380657196, + 0.0018812647322192788, + 0.002283734967932105, + 0.0018305135890841484, + 0.0025803211610764265, + 0.0014396540354937315, + 0.0017928434535861015, + 0.0013275803066790104, + 0.0041683935560286045, + 0.0068484083749353886, + 0.005818219389766455, + 0.005344923585653305, + 0.005035679787397385, + 0.0074724238365888596, + 0.0057252380065619946, + 0.013078199699521065, + 0.006885995157063007, + 0.010267742909491062, + 0.013598056510090828, + 0.0185402799397707, + 0.021365206688642502, + 0.011140421964228153, + 0.011825745925307274, + 0.008999339304864407, + 0.017838630825281143, + 0.02321276254951954, + 0.011379879899322987, + 0.01785302720963955, + 0.1353359967470169, + 0.11140795052051544, + 0.020217902958393097, + 0.013149070553481579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.0010449193650856614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024937265552580357 + ], + "opposite_logprob_tokens": [ + -6.002493858337402 + ], + "logprob_reward": -0.0024937265552580357, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024937265552580357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08475056290626526, + 0.03091944381594658, + 0.06867370009422302, + 0.2796444892883301, + 0.1153038740158081, + 0.07743322104215622, + 0.03232605382800102, + 0.06933876872062683, + 0.050070930272340775, + 0.1915389746427536 + ], + "credit_map": [ + 0.011440210044384003, + 0.013065045699477196, + 0.010737281292676926, + 0.009128302335739136, + 0.012645259499549866, + 0.008659272454679012, + 0.012011764571070671, + 0.007063429336994886, + 0.004206836689263582, + 0.004958402831107378, + 0.002646948676556349, + 0.003115375991910696, + 0.005973432213068008, + 0.0036208806559443474, + 0.002988381078466773, + 0.003409184282645583, + 0.010685192421078682, + 0.0075111896730959415, + 0.007115228101611137, + 0.016418959945440292, + 0.004823784809559584, + 0.0062164380215108395, + 0.008611265569925308, + 0.007291643880307674, + 0.037870194762945175, + 0.014208554290235043, + 0.10705316811800003, + 0.02082374319434166, + 0.020471511408686638, + 0.012608573772013187, + 0.05069715157151222, + 0.015911586582660675, + 0.012682455591857433, + 0.025864114984869957, + 0.013208807446062565, + 0.013188129290938377, + 0.006449807435274124, + 0.010708315297961235, + 0.025338541716337204, + 0.007863701321184635, + 0.003933087456971407, + 0.005496950354427099, + 0.023763081058859825, + 0.004814586136490107, + 0.011379485949873924, + 0.006765245925635099, + 0.007326336111873388, + 0.013954446651041508, + 0.002016712911427021, + 0.003691629273816943, + 0.004714378621429205, + 0.004414033610373735, + 0.004586091265082359, + 0.003646890865638852, + 0.0050449431873857975, + 0.004211375489830971, + 0.0057852258905768394, + 0.0064453487284481525, + 0.007649148348718882, + 0.011199111118912697, + 0.007295470684766769, + 0.006471508648246527, + 0.009416433982551098, + 0.015076521784067154, + 0.005215235985815525, + 0.005662370938807726, + 0.005029148887842894, + 0.008944174274802208, + 0.008608448319137096, + 0.005037302616983652, + 0.005438344553112984, + 0.006135907024145126, + 0.011411407962441444, + 0.009733629412949085, + 0.00538797490298748, + 0.009115248918533325, + 0.06732470542192459, + 0.06447619199752808, + 0.010977499186992645, + 0.013112327083945274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.0024937265552580357 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle macro shot, rendered in breathtakingly realistic detail, begins with sharp focus fixed on a perfectly formed, branching snowflake. This solitary crystal, with six intricate, symmetrical arms branching into delicate, feathery structures, sits precariously on the dark, grainy asphalt surface. The camera then executes a slow, deliberate rack focus, steadily pulling the plane of sharpness away from the delicate snowflake, which melts into a soft blur. The focus propagates downward, penetrating the surface to reveal the perilous crystalline structure of the black ice beneath. This newly focused layer, a thin, glossy flake, captures the dark hues of the road below, punctuated by shard-like cracks and frozen bubbles trapped within their transparent prisons. As this shift occurs, light snowflakes continue to gently move across the frame, appearing as soft, out-of-focus bokeh elements, adding depth and movement. The entire scene is bathed in the cold, unfocused light of an overcast sky, with countless icy surfaces shimmering with subtle light, creating a striking contrast between the rough asphalt texture and the smooth, treacherous ice. The cool, monochromatic palette of white, grey, and deep black creates a powerful atmosphere of both fragile, microscopic beauty and palpable, imminent danger.", + "seed": 8597, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_4/index_171_seed_8597_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016760986344888807, + "vlm_logprob": -0.016760986344888807, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.016760986344888807, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02471177838742733 + ], + "opposite_logprob_tokens": [ + -3.899711847305298 + ], + "logprob_reward": -0.02471177838742733, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02471177838742733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08683347702026367, + 0.07667774707078934, + 0.18715067207813263, + 0.06035200506448746, + 0.10877043008804321, + 0.03582964092493057, + 0.17228084802627563, + 0.15554092824459076, + 0.09427285939455032, + 0.02229144237935543 + ], + "credit_map": [ + 0.01306427177041769, + 0.012496021576225758, + 0.010802539996802807, + 0.008200611919164658, + 0.012618747539818287, + 0.011648546904325485, + 0.009111931547522545, + 0.008890794590115547, + 0.009399877861142159, + 0.007279608398675919, + 0.008209987543523312, + 0.011219893582165241, + 0.010571090504527092, + 0.010714356787502766, + 0.010268045589327812, + 0.009014876559376717, + 0.018448209390044212, + 0.014172066003084183, + 0.02417657896876335, + 0.02844465896487236, + 0.02168435789644718, + 0.019243251532316208, + 0.04297078028321266, + 0.01801075041294098, + 0.006312904879450798, + 0.0036661382764577866, + 0.011266039684414864, + 0.008565167896449566, + 0.006222932133823633, + 0.00697579700499773, + 0.008449498564004898, + 0.008893521502614021, + 0.009390206076204777, + 0.006339631974697113, + 0.014727558009326458, + 0.02216825634241104, + 0.008958209306001663, + 0.01228770986199379, + 0.018341058865189552, + 0.016557781025767326, + 0.003649656195193529, + 0.004701544530689716, + 0.004326007794588804, + 0.006818473804742098, + 0.003879685653373599, + 0.00440377090126276, + 0.004250446334481239, + 0.003800054779276252, + 0.01730431616306305, + 0.02189408428966999, + 0.01893145591020584, + 0.019488858059048653, + 0.021770525723695755, + 0.030979137867689133, + 0.022447310388088226, + 0.019465161487460136, + 0.016497887670993805, + 0.020659035071730614, + 0.03719811886548996, + 0.030849555507302284, + 0.009528600610792637, + 0.0126886498183012, + 0.013672652654349804, + 0.014446407556533813, + 0.009925485588610172, + 0.00858438853174448, + 0.013897563330829144, + 0.016082046553492546, + 0.007561552803963423, + 0.018072878941893578, + 0.012371201068162918, + 0.007777737453579903, + 0.0019205601420253515, + 0.002766887191683054, + 0.003155880607664585, + 0.003818563651293516, + 0.0025000986643135548, + 0.0030097949784249067, + 0.0033523186575621367, + 0.0017673377878963947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.02471177838742733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866136610507965 + ], + "opposite_logprob_tokens": [ + -3.048661470413208 + ], + "logprob_reward": -0.04866136610507965, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866136610507965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11195729672908783, + 0.06490680575370789, + 0.05816376209259033, + 0.03536422550678253, + 0.06442736834287643, + 0.06139151006937027, + 0.15056876838207245, + 0.134649857878685, + 0.22382649779319763, + 0.09474390000104904 + ], + "credit_map": [ + 0.0123754208907485, + 0.010837881825864315, + 0.01974659226834774, + 0.01411616150289774, + 0.01435953937470913, + 0.01313172560185194, + 0.013905265368521214, + 0.013484719209372997, + 0.0076095652766525745, + 0.007439953275024891, + 0.01326756365597248, + 0.00596679886803031, + 0.006438310258090496, + 0.00761514063924551, + 0.00798314344137907, + 0.008586333133280277, + 0.006106892600655556, + 0.005101184826344252, + 0.009511447511613369, + 0.008093188516795635, + 0.0073457700200378895, + 0.005501571577042341, + 0.010002264752984047, + 0.00650144275277853, + 0.0031492451671510935, + 0.002334565157070756, + 0.006994492840021849, + 0.00699416920542717, + 0.0019539576023817062, + 0.004170567262917757, + 0.004525685217231512, + 0.005241541191935539, + 0.005405571777373552, + 0.004073134623467922, + 0.009332200512290001, + 0.011857069097459316, + 0.005197572987526655, + 0.007280576042830944, + 0.011761108413338661, + 0.009520137682557106, + 0.004720366094261408, + 0.005725864786654711, + 0.008384752087295055, + 0.013268806040287018, + 0.005632741842418909, + 0.008027206175029278, + 0.008698356337845325, + 0.006933421827852726, + 0.01155008189380169, + 0.014875422231853008, + 0.02397266775369644, + 0.025783637538552284, + 0.012364199385046959, + 0.01863783784210682, + 0.021679239347577095, + 0.021705683320760727, + 0.013646397739648819, + 0.01968160830438137, + 0.02919136732816696, + 0.026117635890841484, + 0.008754301816225052, + 0.012798843905329704, + 0.011028513312339783, + 0.013431182131171227, + 0.01487586461007595, + 0.013987882062792778, + 0.06149327754974365, + 0.0302402526140213, + 0.013613050803542137, + 0.05315861478447914, + 0.02159390039741993, + 0.014863657765090466, + 0.008028595708310604, + 0.010881047695875168, + 0.0142144113779068, + 0.01722339168190956, + 0.009723355062305927, + 0.013466035947203636, + 0.013496499508619308, + 0.007710563018918037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.04866136610507965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619193332269788 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.003619193332269788, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619193332269788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051476288586854935, + 0.05993151664733887, + 0.06152164563536644, + 0.09293235093355179, + 0.08076706528663635, + 0.05589158087968826, + 0.0862128809094429, + 0.1280629187822342, + 0.31775739789009094, + 0.06544642150402069 + ], + "credit_map": [ + 0.009812872856855392, + 0.005880963988602161, + 0.00986186321824789, + 0.005273154936730862, + 0.00521027110517025, + 0.005301214288920164, + 0.004663716070353985, + 0.005472227931022644, + 0.007895896211266518, + 0.006373325828462839, + 0.007053472567349672, + 0.0064691212028265, + 0.00817362405359745, + 0.008884362876415253, + 0.007861684076488018, + 0.007220026105642319, + 0.005855027120560408, + 0.007093615829944611, + 0.007642090320587158, + 0.008328195661306381, + 0.010461627505719662, + 0.0049285669811069965, + 0.01015518605709076, + 0.007057330571115017, + 0.008812560699880123, + 0.007509919814765453, + 0.013706411235034466, + 0.015919718891382217, + 0.007086392026394606, + 0.010900397785007954, + 0.014709172770380974, + 0.014287772588431835, + 0.008256279863417149, + 0.004486808553338051, + 0.012417308054864407, + 0.013999631628394127, + 0.00561008183285594, + 0.00946673285216093, + 0.013588254339993, + 0.01294196117669344, + 0.00487608602270484, + 0.004473582375794649, + 0.007277818862348795, + 0.013870825991034508, + 0.0035147687885910273, + 0.007175872102379799, + 0.007889678701758385, + 0.006812942214310169, + 0.007265951484441757, + 0.011197191663086414, + 0.012065136805176735, + 0.010946720838546753, + 0.009862522594630718, + 0.01529844943434, + 0.0107704633846879, + 0.008806433528661728, + 0.01172424852848053, + 0.017314543947577477, + 0.03068438172340393, + 0.022800825536251068, + 0.007593246642500162, + 0.011418787762522697, + 0.010646498762071133, + 0.01588038168847561, + 0.027218136936426163, + 0.02405647374689579, + 0.06475845724344254, + 0.047407425940036774, + 0.019486868754029274, + 0.07066316157579422, + 0.035266876220703125, + 0.028899982571601868, + 0.006122109480202198, + 0.00764414481818676, + 0.008169925771653652, + 0.010638028383255005, + 0.00811285711824894, + 0.008864323608577251, + 0.008841962553560734, + 0.007053067442029715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.003619193332269788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003194232005625963 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.003194232005625963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003194232005625963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13451319932937622, + 0.06332173198461533, + 0.09191501885652542, + 0.03703643009066582, + 0.07297226786613464, + 0.07498987764120102, + 0.11441721022129059, + 0.15365570783615112, + 0.13745558261871338, + 0.11972300708293915 + ], + "credit_map": [ + 0.022063100710511208, + 0.01469811424612999, + 0.022980947047472, + 0.012292353436350822, + 0.0166753139346838, + 0.016433997079730034, + 0.013145966455340385, + 0.016223397105932236, + 0.012270432896912098, + 0.006452788133174181, + 0.009376917034387589, + 0.005871745757758617, + 0.00649708416312933, + 0.006560978014022112, + 0.0075253676623106, + 0.0087664145976305, + 0.011066244915127754, + 0.007268478628247976, + 0.009110061451792717, + 0.01591789722442627, + 0.011575978249311447, + 0.00919472984969616, + 0.019039014354348183, + 0.008742612786591053, + 0.003977528307586908, + 0.002820816123858094, + 0.004145808517932892, + 0.004394452087581158, + 0.0041099656373262405, + 0.005812963005155325, + 0.0064071244560182095, + 0.005367772188037634, + 0.006718209479004145, + 0.0044431062415242195, + 0.01194093283265829, + 0.012152942828834057, + 0.005524907726794481, + 0.008953351527452469, + 0.012539859861135483, + 0.010698959231376648, + 0.006161107216030359, + 0.011319143697619438, + 0.009833685122430325, + 0.010501427575945854, + 0.010277350433170795, + 0.010930592194199562, + 0.008605316281318665, + 0.007361257448792458, + 0.009904170408844948, + 0.013495231978595257, + 0.014121295884251595, + 0.017776241526007652, + 0.014020648784935474, + 0.018039757385849953, + 0.015240573324263096, + 0.011819287203252316, + 0.01596919447183609, + 0.02212553657591343, + 0.02890932559967041, + 0.026617204770445824, + 0.010531235486268997, + 0.02091149426996708, + 0.011598300188779831, + 0.016993416473269463, + 0.01339896023273468, + 0.010838325135409832, + 0.02343965508043766, + 0.027005823329091072, + 0.008869605138897896, + 0.024335647001862526, + 0.016046416014432907, + 0.013521159999072552, + 0.009819592349231243, + 0.012042173184454441, + 0.008976743556559086, + 0.020049357786774635, + 0.018366828560829163, + 0.015107918530702591, + 0.020549509674310684, + 0.014810879714787006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.003194232005625963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036183618940413 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036183618940413, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036183618940413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18602487444877625, + 0.11671926081180573, + 0.05596042051911354, + 0.10833632946014404, + 0.06603316962718964, + 0.05937414988875389, + 0.10733568668365479, + 0.12151027470827103, + 0.09097408503293991, + 0.0877317413687706 + ], + "credit_map": [ + 0.027879904955625534, + 0.02174176089465618, + 0.026394909247756004, + 0.01994280517101288, + 0.02723655104637146, + 0.02334439568221569, + 0.020090989768505096, + 0.019393544644117355, + 0.0215897336602211, + 0.014564357697963715, + 0.01587780937552452, + 0.010994459502398968, + 0.011045673862099648, + 0.012450859881937504, + 0.014943382702767849, + 0.015252978540956974, + 0.008009550161659718, + 0.005810072645545006, + 0.008967448025941849, + 0.005614709109067917, + 0.006881000939756632, + 0.008184134028851986, + 0.006096271798014641, + 0.006397234741598368, + 0.01197050604969263, + 0.007991133257746696, + 0.01738583669066429, + 0.013073774054646492, + 0.010230434127151966, + 0.016667263582348824, + 0.014384177513420582, + 0.016633203253149986, + 0.007796064019203186, + 0.004282019101083279, + 0.013462631031870842, + 0.009704695083200932, + 0.004154812078922987, + 0.006227758247405291, + 0.009501839987933636, + 0.010903356596827507, + 0.0048968009650707245, + 0.00753399170935154, + 0.010152329690754414, + 0.010010388679802418, + 0.006552273407578468, + 0.006471604108810425, + 0.0073000225238502026, + 0.0064567388035357, + 0.00967290997505188, + 0.011478574946522713, + 0.01615428552031517, + 0.014287949539721012, + 0.010455465875566006, + 0.01577981933951378, + 0.014229116961359978, + 0.01527756080031395, + 0.01239432580769062, + 0.015848686918616295, + 0.026588207110762596, + 0.022357473149895668, + 0.007813282310962677, + 0.014351774007081985, + 0.010040396824479103, + 0.012116136029362679, + 0.010087836533784866, + 0.010760128498077393, + 0.013953148387372494, + 0.01454766932874918, + 0.009177305735647678, + 0.010481135919690132, + 0.010840022936463356, + 0.01112684328109026, + 0.00832933746278286, + 0.011667372658848763, + 0.011170145124197006, + 0.013195974752306938, + 0.009487028233706951, + 0.011642059311270714, + 0.011321895755827427, + 0.010917927138507366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0036183618940413 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle macro shot, rendered in breathtakingly realistic detail, begins with sharp focus fixed on a perfectly formed, branching snowflake. This solitary crystal, with six intricate, symmetrical arms branching into delicate, feathery structures, sits precariously on the dark, grainy asphalt surface. The camera then executes a slow, deliberate rack focus, steadily pulling the plane of sharpness away from the delicate snowflake, which melts into a soft blur. The focus propagates downward, penetrating the surface to reveal the perilous crystalline structure of the black ice beneath. This newly focused layer, a thin, glossy flake, captures the dark hues of the road below, punctuated by shard-like cracks and frozen bubbles trapped within their transparent prisons. As this shift occurs, light snowflakes continue to gently move across the frame, appearing as soft, out-of-focus bokeh elements, adding depth and movement. The entire scene is bathed in the cold, unfocused light of an overcast sky, with countless icy surfaces shimmering with subtle light, creating a striking contrast between the rough asphalt texture and the smooth, treacherous ice. The cool, monochromatic palette of white, grey, and deep black creates a powerful atmosphere of both fragile, microscopic beauty and palpable, imminent danger.", + "seed": 8598, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_4/index_171_seed_8598_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10165891335345804, + "vlm_logprob": -0.10165891335345804, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.10165891335345804, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.44008949398994446 + ], + "opposite_logprob_tokens": [ + -1.065089464187622 + ], + "logprob_reward": -0.44008949398994446, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.44008949398994446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719580739736557, + 0.04534251242876053, + 0.20672005414962769, + 0.05074857920408249, + 0.06073024496436119, + 0.03199664503335953, + 0.11769463866949081, + 0.05463271215558052, + 0.060462843626737595, + 0.19971369206905365 + ], + "credit_map": [ + 0.01765325292944908, + 0.024271611124277115, + 0.015964170917868614, + 0.014349039644002914, + 0.028768429532647133, + 0.027043329551815987, + 0.01640916056931019, + 0.027499085292220116, + 0.004680435173213482, + 0.00800351519137621, + 0.003912299405783415, + 0.0039294129237532616, + 0.005791668314486742, + 0.006073692813515663, + 0.0033720226492732763, + 0.009579469449818134, + 0.016163744032382965, + 0.030293237417936325, + 0.015162630006670952, + 0.016325190663337708, + 0.037872638553380966, + 0.03837309032678604, + 0.014399143867194653, + 0.03813037648797035, + 0.011578536592423916, + 0.0068817161954939365, + 0.006551545578986406, + 0.005383490584790707, + 0.005449409130960703, + 0.006070298608392477, + 0.0038768923841416836, + 0.004956692922860384, + 0.006194766610860825, + 0.007446169387549162, + 0.009577523916959763, + 0.008419953286647797, + 0.006739573087543249, + 0.005132642108947039, + 0.010175078175961971, + 0.007044539786875248, + 0.004128560423851013, + 0.0038372206036001444, + 0.004439156502485275, + 0.00276169809512794, + 0.003700027707964182, + 0.0036924511659890413, + 0.006312975659966469, + 0.0031245555728673935, + 0.010273723863065243, + 0.013807499781250954, + 0.015971271321177483, + 0.009789003059267998, + 0.02480335161089897, + 0.014186721295118332, + 0.017584461718797684, + 0.011278613470494747, + 0.004853724502027035, + 0.007115976884961128, + 0.009991267696022987, + 0.008056340739130974, + 0.004002419300377369, + 0.005696364678442478, + 0.008545597083866596, + 0.00637102173641324, + 0.0051894644275307655, + 0.009007432498037815, + 0.004710533190518618, + 0.006294214632362127, + 0.010506105609238148, + 0.005183469504117966, + 0.011104214005172253, + 0.008467406965792179, + 0.01078645046800375, + 0.06700334697961807, + 0.0074798655696213245, + 0.008702072314918041, + 0.08166228979825974, + 0.011679606512188911, + 0.006640489678829908, + 0.005759573075920343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.44008949398994446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867362603545189 + ], + "opposite_logprob_tokens": [ + -3.048673629760742 + ], + "logprob_reward": -0.04867362603545189, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867362603545189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0525236539542675, + 0.18870656192302704, + 0.036193378269672394, + 0.025179794058203697, + 0.04765425994992256, + 0.012069919146597385, + 0.025591444224119186, + 0.013984068296849728, + 0.06592822819948196, + 0.5321686863899231 + ], + "credit_map": [ + 0.007311783265322447, + 0.006888319738209248, + 0.006367977242916822, + 0.003944758791476488, + 0.008119029924273491, + 0.009975778870284557, + 0.004109181463718414, + 0.005806824192404747, + 0.024255070835351944, + 0.060244739055633545, + 0.013615645468235016, + 0.007656192407011986, + 0.016443409025669098, + 0.021556798368692398, + 0.0057159834541380405, + 0.0392187237739563, + 0.002989264205098152, + 0.0033469032496213913, + 0.00217319093644619, + 0.0024685340467840433, + 0.003820628160610795, + 0.008266248740255833, + 0.0022658223751932383, + 0.010862788185477257, + 0.0032679173164069653, + 0.0025969729758799076, + 0.0027628776151686907, + 0.004968825727701187, + 0.002559782937169075, + 0.002785260556265712, + 0.0032276371493935585, + 0.0030105209443718195, + 0.006395293399691582, + 0.00618574395775795, + 0.008898208849132061, + 0.00568938534706831, + 0.003524493658915162, + 0.003319165436550975, + 0.007351451553404331, + 0.0062905168160796165, + 0.0012778625823557377, + 0.0017666058847680688, + 0.0015730962622910738, + 0.001715335645712912, + 0.0015197629109025002, + 0.0012096997816115618, + 0.001428473275154829, + 0.0015790830366313457, + 0.0038687207270413637, + 0.002875932725146413, + 0.002858846914023161, + 0.0033994868863373995, + 0.0025530518032610416, + 0.002824651775881648, + 0.0036236154846847057, + 0.003587140003219247, + 0.0019052511779591441, + 0.0017432819586247206, + 0.0016284161247313023, + 0.0019659067038446665, + 0.0013776139821857214, + 0.002188656246289611, + 0.0012977345613762736, + 0.0018772074254229665, + 0.009922131896018982, + 0.008667097426950932, + 0.005719321314245462, + 0.006447767373174429, + 0.007839331403374672, + 0.008476590737700462, + 0.007984449155628681, + 0.01087153423577547, + 0.023988228291273117, + 0.21230299770832062, + 0.007797435857355595, + 0.009572363458573818, + 0.2359248399734497, + 0.024657398462295532, + 0.009444141760468483, + 0.008481272496283054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.04867362603545189 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085116304457188 + ], + "opposite_logprob_tokens": [ + -4.511085033416748 + ], + "logprob_reward": -0.011085116304457188, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085116304457188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13583016395568848, + 0.07925263047218323, + 0.05368073657155037, + 0.0723508670926094, + 0.07474527508020401, + 0.06861667335033417, + 0.060143887996673584, + 0.039899516850709915, + 0.05099349841475487, + 0.36448678374290466 + ], + "credit_map": [ + 0.016658877953886986, + 0.01736528053879738, + 0.012204321101307869, + 0.013918486423790455, + 0.02271939069032669, + 0.026604633778333664, + 0.010965218767523766, + 0.015393957495689392, + 0.009381793439388275, + 0.023199282586574554, + 0.00602011987939477, + 0.004560956731438637, + 0.007671671453863382, + 0.01791914738714695, + 0.004548065364360809, + 0.005951594095677137, + 0.007927103899419308, + 0.006497787777334452, + 0.0062966481782495975, + 0.005949303507804871, + 0.00611853739246726, + 0.008100277744233608, + 0.004725244827568531, + 0.008065836504101753, + 0.012186202220618725, + 0.010006655007600784, + 0.00996541976928711, + 0.008304281160235405, + 0.009599081240594387, + 0.0076524801552295685, + 0.0055887470953166485, + 0.009048000909388065, + 0.010004784911870956, + 0.009342053905129433, + 0.014434248208999634, + 0.006591737270355225, + 0.007450886536389589, + 0.006570616737008095, + 0.008755465969443321, + 0.011595486663281918, + 0.009671113453805447, + 0.012312235310673714, + 0.009607153944671154, + 0.006808461155742407, + 0.009095189161598682, + 0.00534402159973979, + 0.007387637626379728, + 0.008390858769416809, + 0.006682887207716703, + 0.008551470004022121, + 0.006075448356568813, + 0.006951938383281231, + 0.008486822247505188, + 0.007794010452926159, + 0.009184126742184162, + 0.006417182739824057, + 0.004018292762339115, + 0.005275337491184473, + 0.006253991276025772, + 0.006173103116452694, + 0.003754399484023452, + 0.004328809212893248, + 0.004594170022755861, + 0.005501410458236933, + 0.006926536560058594, + 0.006949584931135178, + 0.005275824572890997, + 0.009124412201344967, + 0.004720551893115044, + 0.004807703197002411, + 0.008309598080813885, + 0.004879286978393793, + 0.016599219292402267, + 0.14153091609477997, + 0.008860486559569836, + 0.007343995850533247, + 0.15770114958286285, + 0.017630409449338913, + 0.008424505591392517, + 0.006396102719008923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.011085116304457188 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024927754420787096 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.0024927754420787096, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024927754420787096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0924510732293129, + 0.06016147881746292, + 0.08984068781137466, + 0.031865980476140976, + 0.1296047866344452, + 0.09358978271484375, + 0.05192141979932785, + 0.03771170601248741, + 0.028828635811805725, + 0.3840245008468628 + ], + "credit_map": [ + 0.015571922063827515, + 0.008901978842914104, + 0.007974197156727314, + 0.00726652704179287, + 0.02232370525598526, + 0.013278883881866932, + 0.00862510409206152, + 0.008508742786943913, + 0.005602279212325811, + 0.004934640135616064, + 0.0057145687751472, + 0.004358846694231033, + 0.005370184313505888, + 0.01439616084098816, + 0.0035811620764434338, + 0.0162036269903183, + 0.005180654115974903, + 0.008198605850338936, + 0.00627942243590951, + 0.004877020604908466, + 0.012230412103235722, + 0.019014140591025352, + 0.007023384794592857, + 0.027037031948566437, + 0.0038296300917863846, + 0.004813592415302992, + 0.0033635154832154512, + 0.004707068670541048, + 0.005578918382525444, + 0.0032600064296275377, + 0.0028166458941996098, + 0.0034965984523296356, + 0.016094975173473358, + 0.01721194013953209, + 0.02156090922653675, + 0.013990812003612518, + 0.010235585272312164, + 0.011128933168947697, + 0.022270843386650085, + 0.017110776156187057, + 0.007869108580052853, + 0.010646519251167774, + 0.019332820549607277, + 0.013380271382629871, + 0.008887728676199913, + 0.009378241375088692, + 0.015389867126941681, + 0.008705204352736473, + 0.006787803024053574, + 0.006606605369597673, + 0.005512372590601444, + 0.005566657520830631, + 0.009521903470158577, + 0.0060167391784489155, + 0.0059779975563287735, + 0.0059313299134373665, + 0.004187901969999075, + 0.003959954250603914, + 0.006636841222643852, + 0.0053895702585577965, + 0.0034025683999061584, + 0.005235700402408838, + 0.005412514321506023, + 0.0034866479691118, + 0.003255113260820508, + 0.00458945520222187, + 0.0025525419041514397, + 0.0037907615769654512, + 0.0032802619971334934, + 0.005012039095163345, + 0.003336223540827632, + 0.00301223318092525, + 0.04306311905384064, + 0.09205717593431473, + 0.01762288436293602, + 0.02084405906498432, + 0.14388808608055115, + 0.030889280140399933, + 0.018437424674630165, + 0.017222436144948006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.0024927754420787096 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up macro view of a single snowflake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snowflake sitting on a dark, grainy asphalt surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a slow rack focus that pulls sharpness away from the snowflake to reveal black ice beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed black ice layer thin, glossy, and containing shard-like cracks and frozen bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool, monochromatic light (white, grey, deep black) with falling snow appearing as soft bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595355499535799 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.00595355499535799, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595355499535799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08131192624568939, + 0.04387259483337402, + 0.07939460128545761, + 0.053527869284152985, + 0.12670260667800903, + 0.08834796398878098, + 0.01858445815742016, + 0.014724133536219597, + 0.016173722222447395, + 0.47736015915870667 + ], + "credit_map": [ + 0.0099338972941041, + 0.008977090008556843, + 0.008561093360185623, + 0.004104215651750565, + 0.010734867304563522, + 0.014852042309939861, + 0.0045257266610860825, + 0.01962299644947052, + 0.005826612003147602, + 0.012126922607421875, + 0.003993041813373566, + 0.0033044381998479366, + 0.003994264639914036, + 0.0063518681563436985, + 0.0026045995764434338, + 0.005670848768204451, + 0.0055190217681229115, + 0.005270142108201981, + 0.006000283639878035, + 0.0055316658690571785, + 0.005625647027045488, + 0.018568048253655434, + 0.004946304950863123, + 0.02793348766863346, + 0.009897456504404545, + 0.007104587275534868, + 0.005763581953942776, + 0.005968885030597448, + 0.006026591174304485, + 0.006761451251804829, + 0.0037677274085581303, + 0.008237591944634914, + 0.014378193765878677, + 0.016669776290655136, + 0.02377227321267128, + 0.01328569371253252, + 0.010331748984754086, + 0.010244684293866158, + 0.018840763717889786, + 0.01917947828769684, + 0.007691206876188517, + 0.008372613228857517, + 0.02142108604311943, + 0.011055413633584976, + 0.007872440852224827, + 0.007567473221570253, + 0.01664729416370392, + 0.007720437832176685, + 0.002694037975743413, + 0.002698510652408004, + 0.0022668330930173397, + 0.0017396501498296857, + 0.002889166120439768, + 0.002008383395150304, + 0.002381246769800782, + 0.0019066294189542532, + 0.0015125565696507692, + 0.0018916932167485356, + 0.0017978241667151451, + 0.0025867074728012085, + 0.0017571646021679044, + 0.0017578114056959748, + 0.0016610239399597049, + 0.001759352977387607, + 0.0013157721841707826, + 0.0027884594164788723, + 0.0015100204618647695, + 0.0023951237089931965, + 0.0021689520217478275, + 0.0016342212911695242, + 0.002272573299705982, + 0.0020885998383164406, + 0.02929873578250408, + 0.16925524175167084, + 0.009460941888391972, + 0.011109347455203533, + 0.20921140909194946, + 0.026898745447397232, + 0.011626341380178928, + 0.010499387979507446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.00595355499535799 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_5/index_tensor([2763, 2763])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_5/index_tensor([2763, 2763])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2be7e101ac70a0c4d0b445854f63c6713617c22a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_5/index_tensor([2763, 2763])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic and powerful low-angle, tilted shot (from a Dutch angle) closely follows, focusing on a matte-black tactical assault rifle. Atop the rifle, a meticulously carved golden dragon coils, its scales reflecting a cool metallic luster against the dim backdrop, while tiny rubies embedded in its eyes gleam fiercely. With the pull of the trigger, a blinding burst of flame erupts from the muzzle. In slow motion, the scorching brass cartridge case is ejected sideways at high speed, and the gun violently vibrates backwards from the immense recoil. With unstoppable force, the rapidly spinning bullet slams into the rough, cracked concrete wall ahead, instantly exploding at the point of impact. Countless concrete fragments of varying sizes and gray dust, briefly illuminated by the muzzle flash, explode in all directions like a torrential rain. The entire scene, amidst the sheer chaos, evokes a violent aesthetic, creating an atmosphere of tension, intensity, and destructive power.", + "seed": 5681, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_5/index_2763_seed_5681_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04364905133843422, + "vlm_logprob": -0.04364905133843422, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04364905133843422, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12239442020654678 + ], + "opposite_logprob_tokens": [ + -2.247394323348999 + ], + "logprob_reward": -0.12239442020654678, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12239442020654678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2199976146221161, + 0.04760263115167618, + 0.037249524146318436, + 0.07750696688890457, + 0.05016718804836273, + 0.0550772063434124, + 0.07633668929338455, + 0.06929266452789307, + 0.11000865697860718, + 0.2567608654499054 + ], + "credit_map": [ + 0.04769930988550186, + 0.011238401755690575, + 0.01743997260928154, + 0.019845882430672646, + 0.015975654125213623, + 0.009069555439054966, + 0.07922476530075073, + 0.01950407586991787, + 0.006297006271779537, + 0.0042954059317708015, + 0.006206902675330639, + 0.006375902332365513, + 0.003620131639763713, + 0.0038435233291238546, + 0.00822027213871479, + 0.008743484504520893, + 0.002884353045374155, + 0.00340920127928257, + 0.0049254694022238255, + 0.005584417376667261, + 0.004635137040168047, + 0.003904549404978752, + 0.005470530595630407, + 0.006435863673686981, + 0.0076029663905501366, + 0.013653364963829517, + 0.005291687790304422, + 0.00927942804992199, + 0.012947255745530128, + 0.008058243431150913, + 0.01256814319640398, + 0.008105880580842495, + 0.0048039755783975124, + 0.003779809223487973, + 0.00903883296996355, + 0.004941185005009174, + 0.0054642315953969955, + 0.009661405347287655, + 0.004936947952955961, + 0.007540799677371979, + 0.010608837008476257, + 0.008194880560040474, + 0.005502622574567795, + 0.00658672209829092, + 0.00561120267957449, + 0.008047862909734249, + 0.005281603429466486, + 0.005243469960987568, + 0.0063697099685668945, + 0.0056406971998512745, + 0.006360553205013275, + 0.005883201025426388, + 0.017275754362344742, + 0.008633586578071117, + 0.007672616746276617, + 0.018500568345189095, + 0.008219802752137184, + 0.00840642862021923, + 0.008037134073674679, + 0.0094028664752841, + 0.00933137908577919, + 0.008855631574988365, + 0.008850257843732834, + 0.00818916130810976, + 0.011294770985841751, + 0.01129962783306837, + 0.005439520347863436, + 0.009181692264974117, + 0.031514972448349, + 0.017957789823412895, + 0.00914833601564169, + 0.014171947725117207, + 0.042121678590774536, + 0.032216012477874756, + 0.01670254021883011, + 0.013765832409262657, + 0.06728975474834442, + 0.04946526512503624, + 0.01685543730854988, + 0.018344350159168243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.12239442020654678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371411934494972 + ], + "opposite_logprob_tokens": [ + -3.4087140560150146 + ], + "logprob_reward": -0.03371411934494972, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371411934494972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1585588902235031, + 0.08935211598873138, + 0.12651033699512482, + 0.06367017328739166, + 0.03519866243004799, + 0.04754773899912834, + 0.0607527419924736, + 0.08070535212755203, + 0.15966975688934326, + 0.1780342012643814 + ], + "credit_map": [ + 0.0213694516569376, + 0.00922671239823103, + 0.009586850181221962, + 0.026771247386932373, + 0.008887364529073238, + 0.005152283236384392, + 0.06397878378629684, + 0.013586190529167652, + 0.010343685746192932, + 0.008325677365064621, + 0.007426135241985321, + 0.010668862611055374, + 0.011751275509595871, + 0.011261897161602974, + 0.014690474607050419, + 0.014884110540151596, + 0.01006510853767395, + 0.011936526745557785, + 0.01699230447411537, + 0.021237287670373917, + 0.010458610020577908, + 0.008788707666099072, + 0.020842919126152992, + 0.026188883930444717, + 0.004775086883455515, + 0.00841322261840105, + 0.006738599389791489, + 0.008562726899981499, + 0.009161009453237057, + 0.007169791031628847, + 0.010371721349656582, + 0.008478017523884773, + 0.004828450735658407, + 0.004677027463912964, + 0.0043998658657073975, + 0.004713656846433878, + 0.0038233459927141666, + 0.0043845875188708305, + 0.004855048377066851, + 0.0035166796296834946, + 0.0043073357082903385, + 0.006551908329129219, + 0.00620791781693697, + 0.008631073869764805, + 0.005353277549147606, + 0.00720091862604022, + 0.005097068380564451, + 0.0041982391849160194, + 0.006408403627574444, + 0.009429587982594967, + 0.005195366218686104, + 0.0042579565197229385, + 0.006248923484236002, + 0.012891752645373344, + 0.008664784952998161, + 0.007655968889594078, + 0.009816744364798069, + 0.011727811768651009, + 0.006882131565362215, + 0.008706727996468544, + 0.009417102672159672, + 0.013260211795568466, + 0.010205057449638844, + 0.01068957056850195, + 0.014690465293824673, + 0.018640097230672836, + 0.011583928018808365, + 0.012538236565887928, + 0.04095841199159622, + 0.025095639750361443, + 0.01502841804176569, + 0.02113456092774868, + 0.028437631204724312, + 0.020404061302542686, + 0.018825603649020195, + 0.023168571293354034, + 0.02426721714437008, + 0.020643483847379684, + 0.019257644191384315, + 0.02302999049425125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.03371411934494972 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745543330907822 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745543330907822, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745543330907822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19726157188415527, + 0.03371591493487358, + 0.08755601942539215, + 0.02941528521478176, + 0.12508709728717804, + 0.03673331439495087, + 0.04337837174534798, + 0.13435247540473938, + 0.22395627200603485, + 0.08854366838932037 + ], + "credit_map": [ + 0.041729994118213654, + 0.014541726559400558, + 0.015537762083113194, + 0.021533165127038956, + 0.00772869074717164, + 0.007266677916049957, + 0.07047884166240692, + 0.018444731831550598, + 0.006594648119062185, + 0.003335896646603942, + 0.0034099428448826075, + 0.004393096547573805, + 0.003557164454832673, + 0.0032734216656535864, + 0.004251770209521055, + 0.004899975378066301, + 0.0070902565494179726, + 0.007718471810221672, + 0.010202277451753616, + 0.010829461738467216, + 0.014815771952271461, + 0.009361326694488525, + 0.012326707132160664, + 0.01521175168454647, + 0.002717594150453806, + 0.004567664582282305, + 0.002291698008775711, + 0.003792965319007635, + 0.004647380206733942, + 0.002985486527904868, + 0.0050776656717062, + 0.003334831679239869, + 0.013920553959906101, + 0.013541563414037228, + 0.015813060104846954, + 0.01267527136951685, + 0.011538876220583916, + 0.0214556697756052, + 0.013652686960995197, + 0.022489409893751144, + 0.005720765795558691, + 0.005619807168841362, + 0.0038865951355546713, + 0.005101222079247236, + 0.003764171153306961, + 0.005832403898239136, + 0.0028918019961565733, + 0.0039165508933365345, + 0.0043546948581933975, + 0.004257913213223219, + 0.005151387769728899, + 0.00411742739379406, + 0.005360400769859552, + 0.008246907964348793, + 0.006858303211629391, + 0.005031334701925516, + 0.014877130277454853, + 0.017075834795832634, + 0.018761059269309044, + 0.018032992258667946, + 0.01565202884376049, + 0.017734147608280182, + 0.016513435170054436, + 0.015705842524766922, + 0.020113641396164894, + 0.026658810675144196, + 0.016160136088728905, + 0.016036350280046463, + 0.05922927334904671, + 0.035057432949543, + 0.020988358184695244, + 0.029712265357375145, + 0.011813266202807426, + 0.010612082667648792, + 0.008002405054867268, + 0.011087103746831417, + 0.012800346128642559, + 0.014716577716171741, + 0.007617567200213671, + 0.01189432293176651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.006745543330907822 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865830019116402 + ], + "opposite_logprob_tokens": [ + -3.0486583709716797 + ], + "logprob_reward": -0.04865830019116402, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865830019116402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09778737276792526, + 0.04481269791722298, + 0.07758413255214691, + 0.03791477531194687, + 0.07740410417318344, + 0.06125510856509209, + 0.06139882281422615, + 0.16262909770011902, + 0.22496184706687927, + 0.15425200760364532 + ], + "credit_map": [ + 0.01487084198743105, + 0.01214324589818716, + 0.0106869637966156, + 0.015747694298624992, + 0.0066392612643539906, + 0.005393192172050476, + 0.023367449641227722, + 0.008938713930547237, + 0.006512444466352463, + 0.0028702139388769865, + 0.003143368987366557, + 0.004153009038418531, + 0.0029641096480190754, + 0.003483422566205263, + 0.011436119675636292, + 0.010250011458992958, + 0.00558925187215209, + 0.006579704117029905, + 0.010115931741893291, + 0.009173889644443989, + 0.009497958235442638, + 0.008309301920235157, + 0.011215580627322197, + 0.017102515324950218, + 0.004494229331612587, + 0.005698648747056723, + 0.0026299774181097746, + 0.005787390749901533, + 0.004807735793292522, + 0.004563719034194946, + 0.005745138972997665, + 0.004187933634966612, + 0.00878746435046196, + 0.012250499799847603, + 0.00836035143584013, + 0.013791140168905258, + 0.006462514400482178, + 0.00816255807876587, + 0.009316104464232922, + 0.01027347706258297, + 0.009888870641589165, + 0.007097708992660046, + 0.006966706365346909, + 0.007043547462671995, + 0.006194194313138723, + 0.010806959122419357, + 0.00704049551859498, + 0.006216630805283785, + 0.004924322012811899, + 0.003214209573343396, + 0.005298831965774298, + 0.003897082293406129, + 0.013695999048650265, + 0.011203223839402199, + 0.004525979049503803, + 0.014639176428318024, + 0.019930865615606308, + 0.021924838423728943, + 0.014553277753293514, + 0.01704149693250656, + 0.021264471113681793, + 0.025027863681316376, + 0.02225269190967083, + 0.020633580163121223, + 0.022201497107744217, + 0.027966732159256935, + 0.01696057990193367, + 0.01689535565674305, + 0.05669437721371651, + 0.033657122403383255, + 0.021961810067296028, + 0.02862437441945076, + 0.024351421743631363, + 0.018181486055254936, + 0.010271874256432056, + 0.012030503712594509, + 0.03859145939350128, + 0.030259715393185616, + 0.009846112690865993, + 0.010719427838921547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.04865830019116402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732873618602753 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732873618602753, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732873618602753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719662994146347, + 0.05470992624759674, + 0.06322254240512848, + 0.018467001616954803, + 0.07342983037233353, + 0.05322983115911484, + 0.03128573298454285, + 0.08664077520370483, + 0.035630080848932266, + 0.41141796112060547 + ], + "credit_map": [ + 0.03266545385122299, + 0.010593347251415253, + 0.007870059460401535, + 0.022363483905792236, + 0.0068882121704518795, + 0.0071187433786690235, + 0.07020988315343857, + 0.014257125556468964, + 0.005464527290314436, + 0.004107609856873751, + 0.004044282250106335, + 0.0052291033789515495, + 0.004616561811417341, + 0.004561331123113632, + 0.014904741197824478, + 0.011781774461269379, + 0.004725159611552954, + 0.0052462415769696236, + 0.009554361924529076, + 0.008264792151749134, + 0.006808871403336525, + 0.006576632149517536, + 0.009149003773927689, + 0.012897483073174953, + 0.002102862112224102, + 0.002488258760422468, + 0.001476011355407536, + 0.002373448107391596, + 0.002392418449744582, + 0.0023814626038074493, + 0.002852708799764514, + 0.0023998322430998087, + 0.005902889650315046, + 0.009041296318173409, + 0.006335833575576544, + 0.011888863518834114, + 0.004692968912422657, + 0.013842967338860035, + 0.008341539651155472, + 0.013383472338318825, + 0.004683753941208124, + 0.008137197233736515, + 0.00526835210621357, + 0.00761584285646677, + 0.007943723350763321, + 0.006754111498594284, + 0.004934933502227068, + 0.00789191760122776, + 0.0029024153482168913, + 0.003936412278562784, + 0.003502885578200221, + 0.0024577800650149584, + 0.005113301333039999, + 0.0036861903499811888, + 0.003052694955840707, + 0.00663405517116189, + 0.010836202651262283, + 0.013433056883513927, + 0.008169561624526978, + 0.009017663076519966, + 0.010158141143620014, + 0.012459985911846161, + 0.011376159265637398, + 0.011190001852810383, + 0.005982223432511091, + 0.0043291314505040646, + 0.0031195965129882097, + 0.004114234820008278, + 0.005964541807770729, + 0.004334192257374525, + 0.0035090744495391846, + 0.004277086351066828, + 0.06758858263492584, + 0.04990566149353981, + 0.027087798342108727, + 0.025317728519439697, + 0.10454334318637848, + 0.08098628371953964, + 0.0261775404214859, + 0.0298110693693161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.006732873618602753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic and powerful low-angle, tilted shot (from a Dutch angle) closely follows, focusing on a matte-black tactical assault rifle. Atop the rifle, a meticulously carved golden dragon coils, its scales reflecting a cool metallic luster against the dim backdrop, while tiny rubies embedded in its eyes gleam fiercely. With the pull of the trigger, a blinding burst of flame erupts from the muzzle. In slow motion, the scorching brass cartridge case is ejected sideways at high speed, and the gun violently vibrates backwards from the immense recoil. With unstoppable force, the rapidly spinning bullet slams into the rough, cracked concrete wall ahead, instantly exploding at the point of impact. Countless concrete fragments of varying sizes and gray dust, briefly illuminated by the muzzle flash, explode in all directions like a torrential rain. The entire scene, amidst the sheer chaos, evokes a violent aesthetic, creating an atmosphere of tension, intensity, and destructive power.", + "seed": 5682, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_5/index_2763_seed_5682_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.036001383978873494, + "vlm_logprob": -0.036001383978873494, + "vlm_yes_no_margin": 4.324999928474426, + "vlm_token_credit": -0.036001383978873494, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14057667553424835 + ], + "opposite_logprob_tokens": [ + -2.1405766010284424 + ], + "logprob_reward": -0.14057667553424835, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.14057667553424835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05957277491688728, + 0.018022144213318825, + 0.026572152972221375, + 0.7061750292778015, + 0.026205111294984818, + 0.06393036991357803, + 0.027168000116944313, + 0.02244776114821434, + 0.02001158893108368, + 0.029895059764385223 + ], + "credit_map": [ + 0.008169478736817837, + 0.005942656192928553, + 0.008254670538008213, + 0.00467737065628171, + 0.008233627304434776, + 0.00708408560603857, + 0.00730142230167985, + 0.009909464046359062, + 0.0019950938876718283, + 0.0016152845928445458, + 0.0015159997856244445, + 0.0025092025753110647, + 0.0030753668397665024, + 0.00159204110968858, + 0.0030140019953250885, + 0.0027051528450101614, + 0.002904947381466627, + 0.0036349999718368053, + 0.003174236509948969, + 0.003220005426555872, + 0.00311253871768713, + 0.0021893575321882963, + 0.004531201906502247, + 0.0038048664573580027, + 0.07619453221559525, + 0.0719878152012825, + 0.09830111265182495, + 0.1276935636997223, + 0.05188050493597984, + 0.12272874265909195, + 0.08798189461231232, + 0.06940687447786331, + 0.0023467587307095528, + 0.003176442813128233, + 0.0030312868766486645, + 0.0034609846770763397, + 0.002611555391922593, + 0.0028169716242700815, + 0.004739357624202967, + 0.004021754488348961, + 0.006360627245157957, + 0.006450246088206768, + 0.004451379645615816, + 0.00788290798664093, + 0.00812230072915554, + 0.011207099072635174, + 0.010512353852391243, + 0.008943457156419754, + 0.0027098944410681725, + 0.003575073555111885, + 0.002831500954926014, + 0.003207240253686905, + 0.003937720321118832, + 0.0030399749521166086, + 0.004693689290434122, + 0.003172905184328556, + 0.0017506042495369911, + 0.0032959969248622656, + 0.001672086538746953, + 0.0025375564582645893, + 0.004457083530724049, + 0.0021632923744618893, + 0.0024083261378109455, + 0.004162814933806658, + 0.0023938959930092096, + 0.0022494238801300526, + 0.0017507700249552727, + 0.0016365026822313666, + 0.004769162740558386, + 0.002897783415392041, + 0.0017540734261274338, + 0.002559977350756526, + 0.002975249895825982, + 0.0040305377915501595, + 0.0039001470431685448, + 0.0048417928628623486, + 0.0036104449536651373, + 0.003926449920982122, + 0.003867154475301504, + 0.002743283985182643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.14057667553424835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012558426707983017 + ], + "opposite_logprob_tokens": [ + -4.387558460235596 + ], + "logprob_reward": -0.012558426707983017, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558426707983017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04501355439424515, + 0.02658204734325409, + 0.06147262454032898, + 0.6607814431190491, + 0.04466499760746956, + 0.04389432817697525, + 0.018643071874976158, + 0.017649415880441666, + 0.02760552428662777, + 0.05369304493069649 + ], + "credit_map": [ + 0.0063980864360928535, + 0.0047403909265995026, + 0.0041237641125917435, + 0.003595461370423436, + 0.004856938496232033, + 0.007081060670316219, + 0.006566437892615795, + 0.007651412859559059, + 0.002722326200455427, + 0.002449221443384886, + 0.0017892350442707539, + 0.004318307153880596, + 0.0051556178368628025, + 0.002967064268887043, + 0.003307386301457882, + 0.0038728886283934116, + 0.004195520654320717, + 0.007042394019663334, + 0.007288628723472357, + 0.009150071069598198, + 0.008720139972865582, + 0.00618949206545949, + 0.008816007524728775, + 0.010070370510220528, + 0.07329218089580536, + 0.06557906419038773, + 0.09046592563390732, + 0.12393493950366974, + 0.04540317878127098, + 0.1063191369175911, + 0.0862644612789154, + 0.06952250003814697, + 0.004838180728256702, + 0.006392693147063255, + 0.004354572854936123, + 0.005242038983851671, + 0.0052727144211530685, + 0.004761639051139355, + 0.0071523310616612434, + 0.006650829222053289, + 0.006399130914360285, + 0.0048012929037213326, + 0.003906369674950838, + 0.004190909676253796, + 0.00529909273609519, + 0.007070489227771759, + 0.006749492138624191, + 0.005477551836520433, + 0.002152221743017435, + 0.002324418630450964, + 0.0018721692031249404, + 0.0017093466594815254, + 0.002036498161032796, + 0.0026328556705266237, + 0.003103599650785327, + 0.002811963204294443, + 0.0014071351615712047, + 0.00178115488961339, + 0.0011244203196838498, + 0.0017560914857313037, + 0.0027359258383512497, + 0.002917849225923419, + 0.0026297462172806263, + 0.0032970926258713007, + 0.0036902097053825855, + 0.002267615171149373, + 0.0029519402887672186, + 0.0034511543344706297, + 0.004478768445551395, + 0.003877118695527315, + 0.002574967686086893, + 0.004313749261200428, + 0.006355364341288805, + 0.008093489333987236, + 0.00889830756932497, + 0.006906288210302591, + 0.0059355865232646465, + 0.00647192494943738, + 0.006111737806349993, + 0.0049203489907085896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.012558426707983017 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674506975337863 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.00674506975337863, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674506975337863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1487564891576767, + 0.020982159301638603, + 0.0361790731549263, + 0.3847971260547638, + 0.043645940721035004, + 0.058432381600141525, + 0.0388362854719162, + 0.01717907004058361, + 0.02867268957197666, + 0.22251877188682556 + ], + "credit_map": [ + 0.019649619236588478, + 0.0220304224640131, + 0.013282282277941704, + 0.016268039122223854, + 0.025690609589219093, + 0.01278339046984911, + 0.01834489405155182, + 0.02070724032819271, + 0.0028921121265739202, + 0.002183901611715555, + 0.001864537363871932, + 0.0025057268794625998, + 0.003832727437838912, + 0.0019807268399745226, + 0.00270015187561512, + 0.0030222765635699034, + 0.0040827346965670586, + 0.004632158204913139, + 0.004701214376837015, + 0.005586986895650625, + 0.002880249870941043, + 0.002469636034220457, + 0.005816289223730564, + 0.006009806878864765, + 0.0425715334713459, + 0.03834205120801926, + 0.05539014935493469, + 0.06954537332057953, + 0.02761252038180828, + 0.06421118229627609, + 0.04758571833372116, + 0.03953859582543373, + 0.004321543965488672, + 0.0033976053819060326, + 0.003956088330596685, + 0.0044273375533521175, + 0.004715203773230314, + 0.0060953847132623196, + 0.006026314105838537, + 0.010706464760005474, + 0.0062641240656375885, + 0.006648380309343338, + 0.004051662981510162, + 0.008483801037073135, + 0.008143595419824123, + 0.009433452039957047, + 0.008132845163345337, + 0.007274521980434656, + 0.0033291506115347147, + 0.005030919332057238, + 0.003805909538641572, + 0.004111067857593298, + 0.005837241653352976, + 0.0049138376489281654, + 0.007624101359397173, + 0.004184056539088488, + 0.0014557283138856292, + 0.00204171659424901, + 0.0013533385936170816, + 0.0016661110566928983, + 0.0035537052899599075, + 0.0018129395321011543, + 0.002144336234778166, + 0.00315119419246912, + 0.0028655908536165953, + 0.0033126629423350096, + 0.002395809628069401, + 0.003354040440171957, + 0.008399521000683308, + 0.0035333356354385614, + 0.0021006714086979628, + 0.002711057895794511, + 0.02074667252600193, + 0.03384628891944885, + 0.025750961154699326, + 0.0276216808706522, + 0.030406279489398003, + 0.03787853941321373, + 0.02564166858792305, + 0.020626692101359367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.00674506975337863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187187612056732 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.018187187612056732, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018187187612056732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05476929247379303, + 0.023019684478640556, + 0.05090250447392464, + 0.2545623481273651, + 0.10387042909860611, + 0.10155411064624786, + 0.09925442188978195, + 0.03324798122048378, + 0.04472653940320015, + 0.23409268260002136 + ], + "credit_map": [ + 0.007949947379529476, + 0.006053306628018618, + 0.008493658155202866, + 0.005153764504939318, + 0.0067577357403934, + 0.005397326312959194, + 0.00667940080165863, + 0.00828415248543024, + 0.002274664118885994, + 0.002043134532868862, + 0.0024090935476124287, + 0.0035953649785369635, + 0.002806582022458315, + 0.0027370809111744165, + 0.003943357616662979, + 0.0032104062847793102, + 0.00430338503792882, + 0.006022053305059671, + 0.006502434145659208, + 0.007154448423534632, + 0.003969712648540735, + 0.0039835283532738686, + 0.00908463355153799, + 0.009882313199341297, + 0.02760113589465618, + 0.026314768940210342, + 0.033258479088544846, + 0.05171126499772072, + 0.017864402383565903, + 0.03883858770132065, + 0.03299549221992493, + 0.0259782113134861, + 0.00993671640753746, + 0.00796919409185648, + 0.007298522628843784, + 0.011397880502045155, + 0.009633591398596764, + 0.01453368179500103, + 0.018042396754026413, + 0.025058448314666748, + 0.013054302893579006, + 0.012646198272705078, + 0.007710128091275692, + 0.01464187540113926, + 0.011959324590861797, + 0.015398334711790085, + 0.01345739047974348, + 0.012686556205153465, + 0.008048409596085548, + 0.01285682711750269, + 0.00836129393428564, + 0.01156650111079216, + 0.01694422774016857, + 0.012162627652287483, + 0.016866644844412804, + 0.012447886168956757, + 0.0028696968220174313, + 0.003244349965825677, + 0.0025688353925943375, + 0.0035270838998258114, + 0.007418949156999588, + 0.003086481709033251, + 0.0035039400681853294, + 0.007028643973171711, + 0.007168475538492203, + 0.0036206101067364216, + 0.005615358240902424, + 0.0038443475496023893, + 0.006527829449623823, + 0.006110407877713442, + 0.003816980170086026, + 0.00802252721041441, + 0.02002648264169693, + 0.030399521812796593, + 0.03212909772992134, + 0.03020397759974003, + 0.02850356325507164, + 0.03745120391249657, + 0.03189700469374657, + 0.023481832817196846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.018187187612056732 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019395602867007256 + ], + "opposite_logprob_tokens": [ + -6.25193977355957 + ], + "logprob_reward": -0.0019395602867007256, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019395602867007256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04430536925792694, + 0.035929594188928604, + 0.02839730679988861, + 0.38709756731987, + 0.06900341063737869, + 0.13985925912857056, + 0.052295196801424026, + 0.01832960918545723, + 0.0664750412106514, + 0.15830761194229126 + ], + "credit_map": [ + 0.0063033695332705975, + 0.003985923249274492, + 0.004426358733326197, + 0.0038778092712163925, + 0.005200060084462166, + 0.0060037472285330296, + 0.006681634578853846, + 0.007826467044651508, + 0.0034003141336143017, + 0.003352774539962411, + 0.002932545030489564, + 0.006008035037666559, + 0.007134740706533194, + 0.003403573529794812, + 0.004340931307524443, + 0.0053566824644804, + 0.0031062227208167315, + 0.0030220886692404747, + 0.003555666422471404, + 0.0044656419195234776, + 0.0024424116127192974, + 0.0019743882585316896, + 0.00493995426222682, + 0.004890931770205498, + 0.040465205907821655, + 0.04009966924786568, + 0.05175841972231865, + 0.07266345620155334, + 0.0260098148137331, + 0.0634794533252716, + 0.05075180530548096, + 0.04186973348259926, + 0.005533620715141296, + 0.005729930009692907, + 0.006398853380233049, + 0.007411684840917587, + 0.0058016120456159115, + 0.010251258499920368, + 0.009705728851258755, + 0.01817071996629238, + 0.012125873006880283, + 0.012729841284453869, + 0.007856595329940319, + 0.015598134137690067, + 0.02034788578748703, + 0.029494917020201683, + 0.021258318796753883, + 0.02044769749045372, + 0.0038008911069482565, + 0.004297495819628239, + 0.003857652423903346, + 0.0047439998015761375, + 0.010070993565022945, + 0.00714897969737649, + 0.00972436461597681, + 0.008650820702314377, + 0.0017991038039326668, + 0.0024820028338581324, + 0.0014261254109442234, + 0.001977192237973213, + 0.002423944417387247, + 0.003394655417650938, + 0.0028400500304996967, + 0.001986535033211112, + 0.008942583575844765, + 0.007289106957614422, + 0.005230344366282225, + 0.004579187370836735, + 0.01729642227292061, + 0.008160708472132683, + 0.005752794910222292, + 0.009223897010087967, + 0.011585641652345657, + 0.02006739191710949, + 0.020453382283449173, + 0.01998899318277836, + 0.018519345670938492, + 0.017036760225892067, + 0.026160474866628647, + 0.024495625868439674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.0019395602867007256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic and powerful low-angle, tilted shot (from a Dutch angle) closely follows, focusing on a matte-black tactical assault rifle. Atop the rifle, a meticulously carved golden dragon coils, its scales reflecting a cool metallic luster against the dim backdrop, while tiny rubies embedded in its eyes gleam fiercely. With the pull of the trigger, a blinding burst of flame erupts from the muzzle. In slow motion, the scorching brass cartridge case is ejected sideways at high speed, and the gun violently vibrates backwards from the immense recoil. With unstoppable force, the rapidly spinning bullet slams into the rough, cracked concrete wall ahead, instantly exploding at the point of impact. Countless concrete fragments of varying sizes and gray dust, briefly illuminated by the muzzle flash, explode in all directions like a torrential rain. The entire scene, amidst the sheer chaos, evokes a violent aesthetic, creating an atmosphere of tension, intensity, and destructive power.", + "seed": 5683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_5/index_2763_seed_5683_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01926972633227706, + "vlm_logprob": -0.01926972633227706, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.01926972633227706, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037393614649772644 + ], + "opposite_logprob_tokens": [ + -3.662393569946289 + ], + "logprob_reward": -0.037393614649772644, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.037393614649772644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49258723855018616, + 0.0490548238158226, + 0.04973820224404335, + 0.03455159813165665, + 0.03327434882521629, + 0.15134789049625397, + 0.09302542358636856, + 0.026448456570506096, + 0.04131925106048584, + 0.02865278720855713 + ], + "credit_map": [ + 0.06835225969552994, + 0.056458454579114914, + 0.03870094195008278, + 0.062691330909729, + 0.047547053545713425, + 0.06270694732666016, + 0.0754534900188446, + 0.08067672699689865, + 0.003257515374571085, + 0.007119067944586277, + 0.002989518456161022, + 0.0033511908259242773, + 0.007772583980113268, + 0.009146381169557571, + 0.010076030157506466, + 0.005342536140233278, + 0.004463208373636007, + 0.00340023054741323, + 0.006437492091208696, + 0.008467229083180428, + 0.010657766833901405, + 0.005863361060619354, + 0.004663898143917322, + 0.005785017739981413, + 0.0042465259321033955, + 0.004002787172794342, + 0.0036836538929492235, + 0.003617155132815242, + 0.006669077556580305, + 0.004262979608029127, + 0.003555714851245284, + 0.004513704217970371, + 0.006642496678978205, + 0.003912095446139574, + 0.004362036008387804, + 0.004021671134978533, + 0.0029643417801707983, + 0.003064075019210577, + 0.00436010304838419, + 0.003947526682168245, + 0.026298072189092636, + 0.025099661201238632, + 0.02263965643942356, + 0.01739582046866417, + 0.013898640871047974, + 0.013625331223011017, + 0.017069535329937935, + 0.01532119233161211, + 0.01504940539598465, + 0.011599579825997353, + 0.009150541387498379, + 0.011735532432794571, + 0.017504176124930382, + 0.009690566919744015, + 0.006979560945183039, + 0.01131606288254261, + 0.003985450137406588, + 0.0026990235783159733, + 0.0027476593386381865, + 0.0024517064448446035, + 0.00334914680570364, + 0.00394983496516943, + 0.003466567024588585, + 0.0037990680430084467, + 0.007805142551660538, + 0.00403344351798296, + 0.004214275628328323, + 0.006610753946006298, + 0.004964467603713274, + 0.006263113580644131, + 0.003036916721612215, + 0.004391137044876814, + 0.0036510159261524677, + 0.002857854589819908, + 0.0028895994182676077, + 0.004595552105456591, + 0.0027506365440785885, + 0.0036545032635331154, + 0.0037913077976554632, + 0.004462314769625664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.037393614649772644 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649542272090912 + ], + "opposite_logprob_tokens": [ + -5.379649639129639 + ], + "logprob_reward": -0.004649542272090912, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649542272090912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14469493925571442, + 0.1454385370016098, + 0.08391648530960083, + 0.0823831707239151, + 0.04562695696949959, + 0.11412492394447327, + 0.21457333862781525, + 0.054847922176122665, + 0.06941907852888107, + 0.04497465491294861 + ], + "credit_map": [ + 0.020803671330213547, + 0.015379427000880241, + 0.013102598488330841, + 0.01743168570101261, + 0.01618579961359501, + 0.022996481508016586, + 0.020256534218788147, + 0.018538735806941986, + 0.013038385659456253, + 0.013415813446044922, + 0.01142363715916872, + 0.012539057992398739, + 0.015561922453343868, + 0.029324499890208244, + 0.03513738885521889, + 0.014997833408415318, + 0.011489422991871834, + 0.016419658437371254, + 0.008616333827376366, + 0.007817800156772137, + 0.014507198706269264, + 0.009250137023627758, + 0.007844054140150547, + 0.00797188188880682, + 0.005344685632735491, + 0.008116825483739376, + 0.006226740311831236, + 0.019510546699166298, + 0.012357928790152073, + 0.006999480538070202, + 0.009992866776883602, + 0.013834098353981972, + 0.005095430184155703, + 0.005131635349243879, + 0.004727592691779137, + 0.00983482412993908, + 0.0029529561288654804, + 0.00543900765478611, + 0.004735413007438183, + 0.00771009735763073, + 0.016594184562563896, + 0.021858112886548042, + 0.01614256575703621, + 0.009996161796152592, + 0.008321142755448818, + 0.00932291429489851, + 0.014687027782201767, + 0.017202820628881454, + 0.02531687542796135, + 0.028516188263893127, + 0.016274619847536087, + 0.022704720497131348, + 0.051257677376270294, + 0.02797752246260643, + 0.020347973331809044, + 0.022177761420607567, + 0.01034118328243494, + 0.006946833338588476, + 0.005471962504088879, + 0.0038598275277763605, + 0.006807810161262751, + 0.006327358540147543, + 0.006073976866900921, + 0.0090189753100276, + 0.011000181548297405, + 0.006593513768166304, + 0.006509015336632729, + 0.009511378593742847, + 0.009474791586399078, + 0.011905566789209843, + 0.005456734914332628, + 0.008967895060777664, + 0.0050521609373390675, + 0.010283345356583595, + 0.0031731908675283194, + 0.004309186711907387, + 0.010534132830798626, + 0.0046117547899484634, + 0.0038027630653232336, + 0.0032081196550279856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.004649542272090912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036227568052709103 + ], + "opposite_logprob_tokens": [ + -5.628622531890869 + ], + "logprob_reward": -0.0036227568052709103, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036227568052709103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10359307378530502, + 0.1741655319929123, + 0.03757830709218979, + 0.03639211133122444, + 0.03361436724662781, + 0.38067758083343506, + 0.07071027159690857, + 0.037891194224357605, + 0.08167697489261627, + 0.04370062053203583 + ], + "credit_map": [ + 0.013349343091249466, + 0.013125940226018429, + 0.008225472643971443, + 0.011050679720938206, + 0.014082696288824081, + 0.01215642411261797, + 0.014463411644101143, + 0.01713911071419716, + 0.017954429611563683, + 0.02141817845404148, + 0.01368138287216425, + 0.013602322898805141, + 0.024684425443410873, + 0.04038473218679428, + 0.024139730259776115, + 0.01830032281577587, + 0.003166642040014267, + 0.004731676075607538, + 0.005069967359304428, + 0.005642345175147057, + 0.007816527970135212, + 0.0042308662086725235, + 0.0030067493207752705, + 0.003913532476872206, + 0.002779033500701189, + 0.0034093102440238, + 0.002413483103737235, + 0.00905245915055275, + 0.004469070117920637, + 0.0033689644187688828, + 0.004466643091291189, + 0.0064331479370594025, + 0.004616945516318083, + 0.004602931439876556, + 0.0029017222113907337, + 0.005666614975780249, + 0.0025663867127150297, + 0.003952086437493563, + 0.0030819124076515436, + 0.006225766614079475, + 0.08237414062023163, + 0.07899029552936554, + 0.045881833881139755, + 0.037768829613924026, + 0.03610038384795189, + 0.029691632837057114, + 0.035614509135484695, + 0.03425594046711922, + 0.008359065279364586, + 0.009272616356611252, + 0.005489431321620941, + 0.00803050771355629, + 0.016801605001091957, + 0.008908814750611782, + 0.007278048433363438, + 0.006570181809365749, + 0.007575612515211105, + 0.005285507533699274, + 0.003206859575584531, + 0.002835123334079981, + 0.006665375549346209, + 0.0036270366981625557, + 0.0038782518822699785, + 0.0048174248076975346, + 0.013697084039449692, + 0.008222193457186222, + 0.00782474223524332, + 0.011945007368922234, + 0.009851295500993729, + 0.013225494883954525, + 0.0058311098255217075, + 0.011080041527748108, + 0.006772313266992569, + 0.010300799272954464, + 0.005149749107658863, + 0.0022256074007600546, + 0.009071523323655128, + 0.00450925063341856, + 0.0031107168179005384, + 0.0025606609415262938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.0036227568052709103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305814579129219 + ], + "opposite_logprob_tokens": [ + -3.168058156967163 + ], + "logprob_reward": -0.04305814579129219, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305814579129219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0685100182890892, + 0.1105952262878418, + 0.05281585827469826, + 0.05362819880247116, + 0.06565708667039871, + 0.1575930416584015, + 0.26502808928489685, + 0.042764607816934586, + 0.10651873052120209, + 0.07688914984464645 + ], + "credit_map": [ + 0.0076614199206233025, + 0.006448442582041025, + 0.005959199275821447, + 0.00958301778882742, + 0.007792723830789328, + 0.008717122487723827, + 0.011425381526350975, + 0.010922706685960293, + 0.008634072728455067, + 0.014319421723484993, + 0.006370697636157274, + 0.006027081049978733, + 0.015231345780193806, + 0.02725839614868164, + 0.022985879331827164, + 0.009768330492079258, + 0.006985923275351524, + 0.010220135562121868, + 0.0036882914137095213, + 0.0060858591459691525, + 0.006802918389439583, + 0.004501146264374256, + 0.006123023573309183, + 0.008408554829657078, + 0.007549289148300886, + 0.007731135003268719, + 0.004726996645331383, + 0.005861601792275906, + 0.0070726629346609116, + 0.008030066266655922, + 0.006751955021172762, + 0.0059044924564659595, + 0.010022031143307686, + 0.006717688869684935, + 0.006643742322921753, + 0.007145265117287636, + 0.007350084837526083, + 0.00809138547629118, + 0.008732198737561703, + 0.010954687371850014, + 0.023295536637306213, + 0.025952495634555817, + 0.019587887451052666, + 0.015197798609733582, + 0.01840413548052311, + 0.017510781064629555, + 0.01760503277182579, + 0.02003936469554901, + 0.030586304143071175, + 0.03327517211437225, + 0.016889821738004684, + 0.034116651862859726, + 0.062273237854242325, + 0.03231748193502426, + 0.024170871824026108, + 0.03139856085181236, + 0.008807488717138767, + 0.005620274227112532, + 0.0037167586851865053, + 0.0025961578357964754, + 0.006249777507036924, + 0.004866074305027723, + 0.003919694572687149, + 0.006988380569964647, + 0.020405301824212074, + 0.011257627047598362, + 0.01099413726478815, + 0.015941577032208443, + 0.010748088359832764, + 0.016437772661447525, + 0.007120110560208559, + 0.01361411064863205, + 0.01799659989774227, + 0.008401562459766865, + 0.009804736822843552, + 0.004592281300574541, + 0.013142897747457027, + 0.009821978397667408, + 0.006223222706466913, + 0.0069058723747730255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.04305814579129219 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624572142958641 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624572142958641, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624572142958641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10292412340641022, + 0.08156341314315796, + 0.09739908576011658, + 0.028403695672750473, + 0.04400080069899559, + 0.10437473654747009, + 0.3435112237930298, + 0.05189567804336548, + 0.06463584303855896, + 0.08129142224788666 + ], + "credit_map": [ + 0.014845711179077625, + 0.01445326954126358, + 0.0071880146861076355, + 0.015448136255145073, + 0.012172767892479897, + 0.012940088286995888, + 0.014885802753269672, + 0.010990332812070847, + 0.00950081367045641, + 0.01037928368896246, + 0.007412716280668974, + 0.007656076457351446, + 0.011725228279829025, + 0.012488344684243202, + 0.012672611512243748, + 0.00972833577543497, + 0.011285600252449512, + 0.013782057911157608, + 0.007768751587718725, + 0.01646503433585167, + 0.010645047761499882, + 0.009598383679986, + 0.01026761531829834, + 0.017586594447493553, + 0.0022027771919965744, + 0.0037454848643392324, + 0.001962886890396476, + 0.006322554778307676, + 0.0027348892763257027, + 0.002915741642937064, + 0.00423808116465807, + 0.004281278699636459, + 0.006222816184163094, + 0.004367087967693806, + 0.004829314537346363, + 0.005201536696404219, + 0.003887835657224059, + 0.006095836870372295, + 0.005632556043565273, + 0.007763816975057125, + 0.02141650766134262, + 0.02332915924489498, + 0.010902103036642075, + 0.007568200118839741, + 0.008853025734424591, + 0.007362561300396919, + 0.01162962056696415, + 0.013313554227352142, + 0.0373140312731266, + 0.04755609482526779, + 0.025677897036075592, + 0.034735057502985, + 0.08488621562719345, + 0.04460183531045914, + 0.03263110667467117, + 0.036108970642089844, + 0.005943719297647476, + 0.006453028880059719, + 0.004480232484638691, + 0.0035712888929992914, + 0.011873232200741768, + 0.006119866855442524, + 0.00696451123803854, + 0.0064898026175796986, + 0.007643773220479488, + 0.0064788637682795525, + 0.006795833352953196, + 0.010670759715139866, + 0.01037816796451807, + 0.009601956233382225, + 0.0067503503523766994, + 0.006316141691058874, + 0.015028746798634529, + 0.011712503619492054, + 0.009324933402240276, + 0.005256384611129761, + 0.013030529022216797, + 0.009124324657022953, + 0.008705945685505867, + 0.009108051657676697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.007624572142958641 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic and powerful low-angle, tilted shot (from a Dutch angle) closely follows, focusing on a matte-black tactical assault rifle. Atop the rifle, a meticulously carved golden dragon coils, its scales reflecting a cool metallic luster against the dim backdrop, while tiny rubies embedded in its eyes gleam fiercely. With the pull of the trigger, a blinding burst of flame erupts from the muzzle. In slow motion, the scorching brass cartridge case is ejected sideways at high speed, and the gun violently vibrates backwards from the immense recoil. With unstoppable force, the rapidly spinning bullet slams into the rough, cracked concrete wall ahead, instantly exploding at the point of impact. Countless concrete fragments of varying sizes and gray dust, briefly illuminated by the muzzle flash, explode in all directions like a torrential rain. The entire scene, amidst the sheer chaos, evokes a violent aesthetic, creating an atmosphere of tension, intensity, and destructive power.", + "seed": 5684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_5/index_2763_seed_5684_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03879146138206124, + "vlm_logprob": -0.03879146138206124, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03879146138206124, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11937694251537323 + ], + "opposite_logprob_tokens": [ + -2.2443768978118896 + ], + "logprob_reward": -0.11937694251537323, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11937694251537323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07479876279830933, + 0.059263959527015686, + 0.2965135872364044, + 0.031076569110155106, + 0.04349488392472267, + 0.16212992370128632, + 0.06671781092882156, + 0.08590974658727646, + 0.07879646122455597, + 0.10129832476377487 + ], + "credit_map": [ + 0.006514750886708498, + 0.01191779412329197, + 0.007272022310644388, + 0.010726586915552616, + 0.008231437765061855, + 0.009005390107631683, + 0.011054109781980515, + 0.010076663456857204, + 0.00789960939437151, + 0.004711289424449205, + 0.009056764654815197, + 0.007611211854964495, + 0.00744598126038909, + 0.008147421292960644, + 0.007034084293991327, + 0.007357594091445208, + 0.03872287645936012, + 0.039762429893016815, + 0.04184814170002937, + 0.04223918914794922, + 0.03476105257868767, + 0.024182073771953583, + 0.04948510229587555, + 0.025512687861919403, + 0.005294557195156813, + 0.0031213010661303997, + 0.005148452706634998, + 0.0031389817595481873, + 0.003446461632847786, + 0.004209431353956461, + 0.0033897152170538902, + 0.003327667713165283, + 0.006299386266618967, + 0.0071353004314005375, + 0.007246100809425116, + 0.00522548146545887, + 0.006256333086639643, + 0.00360797718167305, + 0.004191212356090546, + 0.0035330920945852995, + 0.058092955499887466, + 0.008134917356073856, + 0.058234430849552155, + 0.009258871898055077, + 0.007057853043079376, + 0.006674077361822128, + 0.010030828416347504, + 0.004645981825888157, + 0.016662558540701866, + 0.016485951840877533, + 0.009943419136106968, + 0.005723200272768736, + 0.005046548321843147, + 0.004569195210933685, + 0.004188125021755695, + 0.00409881304949522, + 0.005400603637099266, + 0.019065355882048607, + 0.008237234316766262, + 0.005578591953963041, + 0.0071015143766999245, + 0.008958610706031322, + 0.01625259406864643, + 0.015315230935811996, + 0.008521306328475475, + 0.016023432835936546, + 0.00955239124596119, + 0.008102091029286385, + 0.009798217564821243, + 0.008184426464140415, + 0.008528235368430614, + 0.01008635014295578, + 0.013450872153043747, + 0.012943904846906662, + 0.015559825114905834, + 0.009559541009366512, + 0.013886142522096634, + 0.008588922210037708, + 0.015014426782727242, + 0.01229468546807766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.11937694251537323 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810739889740944 + ], + "opposite_logprob_tokens": [ + -3.288107395172119 + ], + "logprob_reward": -0.03810739889740944, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810739889740944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07711216807365417, + 0.07077687978744507, + 0.28587067127227783, + 0.060050755739212036, + 0.04352201893925667, + 0.10999568551778793, + 0.09286728501319885, + 0.16515688598155975, + 0.06396937370300293, + 0.030678240582346916 + ], + "credit_map": [ + 0.010064424946904182, + 0.00933133065700531, + 0.006773090921342373, + 0.007500039413571358, + 0.011788545176386833, + 0.008594905026257038, + 0.01254767645150423, + 0.01051215548068285, + 0.008194571360945702, + 0.007838893681764603, + 0.011640273034572601, + 0.00939954537898302, + 0.007625930942595005, + 0.008701059967279434, + 0.008613022975623608, + 0.008763578720390797, + 0.03910531476140022, + 0.03957494720816612, + 0.04067324474453926, + 0.04028870910406113, + 0.03419671207666397, + 0.02336295135319233, + 0.045180048793554306, + 0.023488754406571388, + 0.009964617900550365, + 0.005471567623317242, + 0.009869933128356934, + 0.00751760508865118, + 0.009123829193413258, + 0.004441019147634506, + 0.009093998931348324, + 0.004568187054246664, + 0.006421447265893221, + 0.005237119272351265, + 0.006788525730371475, + 0.008333747275173664, + 0.0051626297645270824, + 0.004464715253561735, + 0.0035154172219336033, + 0.0035984166897833347, + 0.040039438754320145, + 0.005435251165181398, + 0.03898245841264725, + 0.00616617314517498, + 0.005346023011952639, + 0.004034372512251139, + 0.0071211885660886765, + 0.002870782744139433, + 0.029043637216091156, + 0.02676096372306347, + 0.008883733302354813, + 0.005661884788423777, + 0.007145383395254612, + 0.0062131634913384914, + 0.005158692132681608, + 0.0039998325519263744, + 0.010067255236208439, + 0.04049043357372284, + 0.01169705018401146, + 0.009353980422019958, + 0.012550856918096542, + 0.016241488978266716, + 0.0292789489030838, + 0.03547687083482742, + 0.007096029352396727, + 0.010895251296460629, + 0.00888090394437313, + 0.007184935733675957, + 0.008223610930144787, + 0.006737311836332083, + 0.0056885890662670135, + 0.009262735955417156, + 0.003884006291627884, + 0.005104999523609877, + 0.004756327252835035, + 0.0030216246377676725, + 0.0025966225657612085, + 0.002559379441663623, + 0.003677431493997574, + 0.005077849607914686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.03810739889740944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639833260327578 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.007639833260327578, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639833260327578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10586617141962051, + 0.14663708209991455, + 0.05584309622645378, + 0.1903507262468338, + 0.06757248193025589, + 0.17029984295368195, + 0.07294891029596329, + 0.12976591289043427, + 0.04004225879907608, + 0.020673435181379318 + ], + "credit_map": [ + 0.009487085044384003, + 0.019796542823314667, + 0.008631601929664612, + 0.01381931733340025, + 0.015403138473629951, + 0.01159840077161789, + 0.013627687469124794, + 0.013502408750355244, + 0.018849318847060204, + 0.015623644925653934, + 0.020354323089122772, + 0.014799441210925579, + 0.021113527938723564, + 0.023615943267941475, + 0.017329614609479904, + 0.014951267279684544, + 0.008168309926986694, + 0.008203914389014244, + 0.007671174127608538, + 0.008442889899015427, + 0.0060020918026566505, + 0.005125056020915508, + 0.00693289702758193, + 0.005296764895319939, + 0.03071514330804348, + 0.018520986661314964, + 0.0499376505613327, + 0.022645657882094383, + 0.021968061104416847, + 0.012669689953327179, + 0.021601691842079163, + 0.012291855178773403, + 0.007693324703723192, + 0.009177235886454582, + 0.00867603998631239, + 0.009306417778134346, + 0.01292747724801302, + 0.00714061688631773, + 0.006745053455233574, + 0.005906321108341217, + 0.05904889479279518, + 0.01087301317602396, + 0.057388801127672195, + 0.00911299604922533, + 0.011948688887059689, + 0.006988930515944958, + 0.010920084081590176, + 0.004018446896225214, + 0.021935952827334404, + 0.020340047776699066, + 0.0060628922656178474, + 0.005063319578766823, + 0.006565704941749573, + 0.0049274992197752, + 0.0039307670667767525, + 0.004122728947550058, + 0.007489376235753298, + 0.027813849970698357, + 0.011684777215123177, + 0.00912291556596756, + 0.009391420520842075, + 0.01201616320759058, + 0.025669069960713387, + 0.02657836303114891, + 0.003661771072074771, + 0.0073396279476583, + 0.006869686767458916, + 0.004994319751858711, + 0.003648711834102869, + 0.004084760323166847, + 0.004614595789462328, + 0.0048287855461239815, + 0.003260871395468712, + 0.0027699896600097418, + 0.0020950092002749443, + 0.0017117318930104375, + 0.0022044661454856396, + 0.0022072261199355125, + 0.0027485962491482496, + 0.003675545332953334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.007639833260327578 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346659287810326 + ], + "opposite_logprob_tokens": [ + -3.6513466835021973 + ], + "logprob_reward": -0.026346659287810326, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346659287810326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0397610068321228, + 0.026629749685525894, + 0.15005691349506378, + 0.029080474749207497, + 0.053675584495067596, + 0.23755775392055511, + 0.0762513056397438, + 0.22812192142009735, + 0.08952300250530243, + 0.069342240691185 + ], + "credit_map": [ + 0.003799900645390153, + 0.0044647459872066975, + 0.0038817080203443766, + 0.00518043152987957, + 0.004756588954478502, + 0.004871092736721039, + 0.0069756400771439075, + 0.005830899812281132, + 0.0037061593029648066, + 0.0026260572485625744, + 0.0031617730855941772, + 0.003061809577047825, + 0.0034646333660930395, + 0.0042227753438055515, + 0.0033314169850200415, + 0.0030551257077604532, + 0.02050662972033024, + 0.020588049665093422, + 0.022585511207580566, + 0.021801801398396492, + 0.017302410677075386, + 0.011873181909322739, + 0.02302505634725094, + 0.012374279089272022, + 0.003906374331563711, + 0.002737757982686162, + 0.005823488812893629, + 0.0038405656814575195, + 0.00335319759324193, + 0.002466210164129734, + 0.0046000173315405846, + 0.00235286308452487, + 0.008689050562679768, + 0.006757538765668869, + 0.008189770393073559, + 0.010576135478913784, + 0.007037011906504631, + 0.003995773382484913, + 0.0053480383940041065, + 0.003082266775891185, + 0.09209387004375458, + 0.011968281120061874, + 0.08763384819030762, + 0.011537776328623295, + 0.008874176070094109, + 0.006988581735640764, + 0.014181045815348625, + 0.004280175548046827, + 0.010223159566521645, + 0.010013657622039318, + 0.013587462715804577, + 0.009641322307288647, + 0.009862800128757954, + 0.008323087356984615, + 0.007951047271490097, + 0.0066487668082118034, + 0.01329752430319786, + 0.04476626217365265, + 0.023318113759160042, + 0.017553653568029404, + 0.018311960622668266, + 0.021826008334755898, + 0.04557635635137558, + 0.0434720441699028, + 0.007457846775650978, + 0.014542452991008759, + 0.013220780529081821, + 0.011006062850356102, + 0.0060571664944291115, + 0.00941622257232666, + 0.010790718719363213, + 0.017031753435730934, + 0.006670376751571894, + 0.008900301530957222, + 0.009760315530002117, + 0.007236256264150143, + 0.011231069453060627, + 0.008734493516385555, + 0.00760498596355319, + 0.009204443544149399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.026346659287810326 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with a tilted (Dutch) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a matte-black tactical assault rifle the central focus of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden dragon carved on top of the rifle with ruby eyes and metallic scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a brass cartridge case ejecting sideways and a bullet hitting a cracked concrete wall in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are concrete fragments and gray dust exploding in all directions illuminated by a muzzle flash?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002486472949385643 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.002486472949385643, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002486472949385643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05001842975616455, + 0.04999313876032829, + 0.046548448503017426, + 0.02324836328625679, + 0.03733684495091438, + 0.06281972676515579, + 0.1462695151567459, + 0.4970734715461731, + 0.05565320700407028, + 0.031038811430335045 + ], + "credit_map": [ + 0.004693996626883745, + 0.005770454648882151, + 0.004554622806608677, + 0.005740490276366472, + 0.006172890774905682, + 0.007157681044191122, + 0.007791795302182436, + 0.008136505261063576, + 0.006514445878565311, + 0.00681289890781045, + 0.005909843370318413, + 0.007518498692661524, + 0.00579593563452363, + 0.005141531582921743, + 0.006718178745359182, + 0.005581808276474476, + 0.006024131551384926, + 0.006227435544133186, + 0.006848159711807966, + 0.006763914134353399, + 0.005154149606823921, + 0.003597373841330409, + 0.00802976917475462, + 0.0039035172667354345, + 0.0037360722199082375, + 0.002469707280397415, + 0.0032712763641029596, + 0.002220355672761798, + 0.0037234274204820395, + 0.002587210852652788, + 0.0029118547681719065, + 0.0023284610360860825, + 0.005466820672154427, + 0.005413249600678682, + 0.006495223846286535, + 0.005903078243136406, + 0.004629325121641159, + 0.0023783999495208263, + 0.004714129492640495, + 0.0023366196546703577, + 0.016999060288071632, + 0.003285755170509219, + 0.020179687067866325, + 0.00393943814560771, + 0.004310885444283485, + 0.006598333362489939, + 0.004936442710459232, + 0.002570133423432708, + 0.042421095073223114, + 0.03903927281498909, + 0.01818869635462761, + 0.010709194466471672, + 0.011746958829462528, + 0.007881643250584602, + 0.00842937920242548, + 0.007853282615542412, + 0.029333217069506645, + 0.11301738768815994, + 0.041695307940244675, + 0.034516576677560806, + 0.03579127788543701, + 0.0440676175057888, + 0.09892300516366959, + 0.09972913563251495, + 0.005833741743117571, + 0.011923065409064293, + 0.008053788915276527, + 0.007225853390991688, + 0.005844083148986101, + 0.005665399134159088, + 0.005429547745734453, + 0.005677729845046997, + 0.004461376462131739, + 0.004886961076408625, + 0.003533707931637764, + 0.0028431909158825874, + 0.0028295889496803284, + 0.002533884719014168, + 0.004267891403287649, + 0.0056822155602276325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.002486472949385643 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_6/index_tensor([7743, 7743])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_6/index_tensor([7743, 7743])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f99c1e50a2b4384976973de71e7747cb58779ef1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_6/index_tensor([7743, 7743])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, statically shot, slightly tilted from above, captures a scene with the dramatic tension and high-contrast lighting of a classic martial arts film. The frame is almost entirely filled by the sushi master's hands, his weathered yet spotless fingers revealing the focused precision of a seasoned warrior. Both hands hold a clean, hand-rolled roll of crisp, dark green nori, its texture nuanced under the sharp spotlight. Inside the cone, pearly grains of vinegar-flavored rice are perfectly arranged around a core of premium, crimson tuna, its surface gleaming with natural oils. Before making the final, decisive move, the master's other hand remains poised and silent, holding a delicate brush smothered in a drop of thick, iridescent, dark sauce that swirls over the tuna. The sauce, almost black, catches the key light, revealing hues of deep amber and mahogany. In a moment of suspended action, the drop is meticulously applied, landing with controlled slowness on the center of the tuna, where it instantly solidifies, its surface tension forming a perfect, shimmering dome – a solid jewel of flavor that catches the light, creating a powerful visual climax that elevates the act of cooking to a ritual of profound care and artistry.", + "seed": 5835, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_6/index_7743_seed_5835_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.22832278870046138, + "vlm_logprob": -0.22832278870046138, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.22832278870046138, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06258569657802582 + ], + "opposite_logprob_tokens": [ + -2.9375855922698975 + ], + "logprob_reward": -0.06258569657802582, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06258569657802582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0824332907795906, + 0.06843265891075134, + 0.20599444210529327, + 0.07014864683151245, + 0.05786912143230438, + 0.13524651527404785, + 0.068747878074646, + 0.17239591479301453, + 0.10641643404960632, + 0.03231510519981384 + ], + "credit_map": [ + 0.008815770037472248, + 0.011049400083720684, + 0.008915786631405354, + 0.00759304640814662, + 0.007065425161272287, + 0.010868268087506294, + 0.01021008100360632, + 0.0179155133664608, + 0.017749018967151642, + 0.006001942791044712, + 0.010247529484331608, + 0.0074005345813930035, + 0.006974715739488602, + 0.004368337336927652, + 0.008011179976165295, + 0.007679401431232691, + 0.02607731707394123, + 0.037041764706373215, + 0.019079342484474182, + 0.018417395651340485, + 0.02094963751733303, + 0.025164201855659485, + 0.029377592727541924, + 0.02988719381392002, + 0.008893406018614769, + 0.006293364334851503, + 0.007242731750011444, + 0.016040775924921036, + 0.008005991578102112, + 0.0050728959031403065, + 0.010300748981535435, + 0.008298731409013271, + 0.008185173384845257, + 0.006133280694484711, + 0.012075289152562618, + 0.006303523201495409, + 0.008229386992752552, + 0.0033835237845778465, + 0.007104154210537672, + 0.006454785820096731, + 0.015121638774871826, + 0.016394060105085373, + 0.014332205057144165, + 0.021369557827711105, + 0.011801544576883316, + 0.015603462234139442, + 0.021345611661672592, + 0.01927843876183033, + 0.009830296970903873, + 0.0077624754048883915, + 0.009942565113306046, + 0.00739523209631443, + 0.007298354525119066, + 0.006916305050253868, + 0.010403217747807503, + 0.009199429303407669, + 0.02293304167687893, + 0.015351919457316399, + 0.02862764708697796, + 0.021180488169193268, + 0.01453875657171011, + 0.025001028552651405, + 0.024254193529486656, + 0.020508842542767525, + 0.014620298519730568, + 0.011038494296371937, + 0.016569551080465317, + 0.020937584340572357, + 0.0076569439843297005, + 0.008194097317755222, + 0.01240698155015707, + 0.0149924885481596, + 0.004885692149400711, + 0.0030710098799318075, + 0.004484664648771286, + 0.0066574811935424805, + 0.002694417955353856, + 0.0022279564291238785, + 0.003716228064149618, + 0.0045776572078466415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.06258569657802582 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498092621564865 + ], + "opposite_logprob_tokens": [ + -2.929980993270874 + ], + "logprob_reward": -0.05498092621564865, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498092621564865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1387416571378708, + 0.16924308240413666, + 0.14432595670223236, + 0.03920602425932884, + 0.14213696122169495, + 0.10608948767185211, + 0.04777415841817856, + 0.05988189950585365, + 0.05571666732430458, + 0.09688419848680496 + ], + "credit_map": [ + 0.02237892337143421, + 0.019034694880247116, + 0.030619850382208824, + 0.010348990559577942, + 0.011571278795599937, + 0.009155387058854103, + 0.017268121242523193, + 0.018364407122135162, + 0.0306604765355587, + 0.017636675387620926, + 0.02317412942647934, + 0.021548572927713394, + 0.01675984635949135, + 0.01426900178194046, + 0.020894881337881088, + 0.0242994986474514, + 0.020676229149103165, + 0.01527748629450798, + 0.022770250216126442, + 0.027502333745360374, + 0.012733353301882744, + 0.010716177523136139, + 0.01930188201367855, + 0.015348237939178944, + 0.0061004809103906155, + 0.0032374428119510412, + 0.006829618476331234, + 0.004277538973838091, + 0.004402238875627518, + 0.0027583374176174402, + 0.007395118474960327, + 0.004205246921628714, + 0.018603913486003876, + 0.0141912791877985, + 0.023400086909532547, + 0.024925969541072845, + 0.00907295010983944, + 0.013964938931167126, + 0.0158514566719532, + 0.02212636172771454, + 0.015572003088891506, + 0.011013534851372242, + 0.018382016569375992, + 0.014996856451034546, + 0.007888884283602238, + 0.008289476856589317, + 0.012255970388650894, + 0.017690744251012802, + 0.005834329407662153, + 0.003951816353946924, + 0.005816975142806768, + 0.010306927375495434, + 0.0035128623712807894, + 0.005063822027295828, + 0.006129443645477295, + 0.007157983724027872, + 0.008183363825082779, + 0.004632977768778801, + 0.010662497021257877, + 0.008067633956670761, + 0.00457898760214448, + 0.005196477752178907, + 0.007224971428513527, + 0.01133499015122652, + 0.006638067774474621, + 0.007136382162570953, + 0.010528236627578735, + 0.007918712683022022, + 0.005557996220886707, + 0.004214228596538305, + 0.008136791177093983, + 0.00558624928817153, + 0.014541400596499443, + 0.008435581810772419, + 0.01428022887557745, + 0.01802472025156021, + 0.008048679679632187, + 0.00839967466890812, + 0.010195858776569366, + 0.014958050101995468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.05498092621564865 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02639310248196125 + ], + "opposite_logprob_tokens": [ + -3.651393175125122 + ], + "logprob_reward": -0.02639310248196125, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02639310248196125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10756354033946991, + 0.36876603960990906, + 0.12569686770439148, + 0.03177570924162865, + 0.06644299626350403, + 0.025822529569268227, + 0.03944924473762512, + 0.1050707995891571, + 0.05924682691693306, + 0.0701654702425003 + ], + "credit_map": [ + 0.012122215703129768, + 0.01689395122230053, + 0.017491867765784264, + 0.009218624792993069, + 0.009866409935057163, + 0.009889417327940464, + 0.010900494642555714, + 0.021180573850870132, + 0.09466251730918884, + 0.03335230052471161, + 0.05498135834932327, + 0.033194441348314285, + 0.030298959463834763, + 0.02853342704474926, + 0.052962951362133026, + 0.04078010469675064, + 0.02006278559565544, + 0.014276117086410522, + 0.010728621855378151, + 0.023897355422377586, + 0.009300896897912025, + 0.013024892657995224, + 0.01790769211947918, + 0.016498509794473648, + 0.0050018723122775555, + 0.002941124141216278, + 0.003979943227022886, + 0.0037717975210398436, + 0.003524567000567913, + 0.001840250683017075, + 0.006372819654643536, + 0.0043433355167508125, + 0.008940516039729118, + 0.007319985888898373, + 0.010770740918815136, + 0.013753536157310009, + 0.005040274932980537, + 0.004207461606711149, + 0.007876191288232803, + 0.00853428989648819, + 0.0033905685413628817, + 0.0023806015960872173, + 0.0030411879997700453, + 0.004587436094880104, + 0.0029101092368364334, + 0.0029199423734098673, + 0.0036689843982458115, + 0.0029236990958452225, + 0.005335938651114702, + 0.007206835318356752, + 0.006215597037225962, + 0.00312018278054893, + 0.0037963702343404293, + 0.00455530546605587, + 0.005030362866818905, + 0.0041886563412845135, + 0.014828115701675415, + 0.011914948001503944, + 0.01833789236843586, + 0.012776059098541737, + 0.009152197279036045, + 0.014787804335355759, + 0.011729615740478039, + 0.011544172652065754, + 0.008566273376345634, + 0.006291165482252836, + 0.009034393355250359, + 0.011933164671063423, + 0.004847842268645763, + 0.0043959347531199455, + 0.00688146473839879, + 0.007296591065824032, + 0.010549587197601795, + 0.006118088494986296, + 0.010377258993685246, + 0.014217265881597996, + 0.005366424098610878, + 0.005369874648749828, + 0.007545588072389364, + 0.010621391236782074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.02639310248196125 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331264689564705 + ], + "opposite_logprob_tokens": [ + -3.773312568664551 + ], + "logprob_reward": -0.02331264689564705, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331264689564705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13579143583774567, + 0.20137955248355865, + 0.1860949844121933, + 0.07257670909166336, + 0.11137506365776062, + 0.05240016430616379, + 0.032490961253643036, + 0.1127195879817009, + 0.04528459534049034, + 0.049886979162693024 + ], + "credit_map": [ + 0.014662589877843857, + 0.014133933931589127, + 0.020975079387426376, + 0.014555699191987514, + 0.013080169446766376, + 0.013126428239047527, + 0.014926305040717125, + 0.030331235378980637, + 0.048573970794677734, + 0.018895134329795837, + 0.030553901568055153, + 0.018805140629410744, + 0.014544333331286907, + 0.014092818833887577, + 0.03234601020812988, + 0.02356824278831482, + 0.027721717953681946, + 0.029051097109913826, + 0.03530162200331688, + 0.036488793790340424, + 0.01209493912756443, + 0.009637405164539814, + 0.019360607489943504, + 0.0164388045668602, + 0.009511184878647327, + 0.007628218270838261, + 0.005756601225584745, + 0.015186212956905365, + 0.006390298251062632, + 0.0056560286320745945, + 0.009488738141953945, + 0.012959429062902927, + 0.015134547837078571, + 0.014091532677412033, + 0.01891816407442093, + 0.02313501574099064, + 0.00925485696643591, + 0.007788461167365313, + 0.011782549321651459, + 0.011269940994679928, + 0.007681309711188078, + 0.006240257527679205, + 0.009349115192890167, + 0.00677252933382988, + 0.004574851132929325, + 0.004495664965361357, + 0.006424291990697384, + 0.006862143985927105, + 0.004323966335505247, + 0.005799123086035252, + 0.003020624164491892, + 0.003925526514649391, + 0.002339803148061037, + 0.0030682648066431284, + 0.004770382773131132, + 0.005243272520601749, + 0.013541033491492271, + 0.020739255473017693, + 0.017032915726304054, + 0.011024810373783112, + 0.011795178055763245, + 0.0064381761476397514, + 0.017415786162018776, + 0.014732436276972294, + 0.0058480179868638515, + 0.0032325556967407465, + 0.0049741268157958984, + 0.006544942036271095, + 0.0027400809340178967, + 0.006479655858129263, + 0.006606344133615494, + 0.008858872577548027, + 0.006938050966709852, + 0.004244350828230381, + 0.0065795970149338245, + 0.008610048331320286, + 0.0033611406106501818, + 0.004534879699349403, + 0.005552056711167097, + 0.010066856630146503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.02331264689564705 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743415713310242 + ], + "opposite_logprob_tokens": [ + -0.47434157133102417 + ], + "logprob_reward": -0.9743415713310242, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743415713310242, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11440452933311462, + 0.19978947937488556, + 0.11034118384122849, + 0.10462316125631332, + 0.22310124337673187, + 0.07410021126270294, + 0.050056103616952896, + 0.06111865118145943, + 0.03454966843128204, + 0.02791576460003853 + ], + "credit_map": [ + 0.010518183000385761, + 0.014016947709023952, + 0.01765107922255993, + 0.009467137977480888, + 0.014396538026630878, + 0.013174145482480526, + 0.014825987629592419, + 0.020354507490992546, + 0.05497284606099129, + 0.020128704607486725, + 0.02972172386944294, + 0.017205966636538506, + 0.013115211389958858, + 0.01653578318655491, + 0.02292880229651928, + 0.025180436670780182, + 0.014608760364353657, + 0.0174773670732975, + 0.017203206196427345, + 0.013724887743592262, + 0.008161346428096294, + 0.008891524747014046, + 0.018550364300608635, + 0.011723734438419342, + 0.013596837408840656, + 0.008406754583120346, + 0.010871779173612595, + 0.026331672444939613, + 0.007819053716957569, + 0.010104844346642494, + 0.012005899101495743, + 0.015486323274672031, + 0.03269033133983612, + 0.028257619589567184, + 0.04524097964167595, + 0.042790401726961136, + 0.02053670771420002, + 0.013623170554637909, + 0.020509377121925354, + 0.019452668726444244, + 0.010541100054979324, + 0.008873744867742062, + 0.014620840549468994, + 0.0061187539249658585, + 0.00821694079786539, + 0.006720955017954111, + 0.010843049734830856, + 0.008164821192622185, + 0.006912548094987869, + 0.008423471823334694, + 0.006461743265390396, + 0.00415869290009141, + 0.005040631163865328, + 0.004049242008477449, + 0.008049394935369492, + 0.0069603766314685345, + 0.006874560844153166, + 0.007766254711896181, + 0.0067589059472084045, + 0.013244936242699623, + 0.006986597552895546, + 0.005439420230686665, + 0.007728636264801025, + 0.00631934218108654, + 0.005019496660679579, + 0.0032148377504199743, + 0.003389721969142556, + 0.005344379227608442, + 0.0039306264370679855, + 0.002859419910237193, + 0.006144657265394926, + 0.00464653130620718, + 0.0038617332465946674, + 0.0027292289305478334, + 0.0033389199525117874, + 0.0036956570111215115, + 0.0030731705483049154, + 0.003125617513433099, + 0.0033720755018293858, + 0.004719362128525972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.9743415713310242 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, statically shot, slightly tilted from above, captures a scene with the dramatic tension and high-contrast lighting of a classic martial arts film. The frame is almost entirely filled by the sushi master's hands, his weathered yet spotless fingers revealing the focused precision of a seasoned warrior. Both hands hold a clean, hand-rolled roll of crisp, dark green nori, its texture nuanced under the sharp spotlight. Inside the cone, pearly grains of vinegar-flavored rice are perfectly arranged around a core of premium, crimson tuna, its surface gleaming with natural oils. Before making the final, decisive move, the master's other hand remains poised and silent, holding a delicate brush smothered in a drop of thick, iridescent, dark sauce that swirls over the tuna. The sauce, almost black, catches the key light, revealing hues of deep amber and mahogany. In a moment of suspended action, the drop is meticulously applied, landing with controlled slowness on the center of the tuna, where it instantly solidifies, its surface tension forming a perfect, shimmering dome – a solid jewel of flavor that catches the light, creating a powerful visual climax that elevates the act of cooking to a ritual of profound care and artistry.", + "seed": 5836, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_6/index_7743_seed_5836_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10122242774814368, + "vlm_logprob": -0.10122242774814368, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.10122242774814368, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028243018314242363 + ], + "opposite_logprob_tokens": [ + -4.903243064880371 + ], + "logprob_reward": -0.028243018314242363, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.028243018314242363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06795225292444229, + 0.05182226374745369, + 0.11481934785842896, + 0.07291493564844131, + 0.05250687897205353, + 0.03171573579311371, + 0.08969923108816147, + 0.2217230498790741, + 0.22657795250415802, + 0.07026828825473785 + ], + "credit_map": [ + 0.0038675570394843817, + 0.005564376711845398, + 0.005549601744860411, + 0.007712231017649174, + 0.007393281906843185, + 0.014640792272984982, + 0.01166506577283144, + 0.011559344828128815, + 0.00562359718605876, + 0.004801445174962282, + 0.00816081278026104, + 0.0037375055253505707, + 0.004779461305588484, + 0.006608886644244194, + 0.00709944823756814, + 0.011011107824742794, + 0.0032493658363819122, + 0.0034187762066721916, + 0.00709273898974061, + 0.006567183416336775, + 0.008322804234921932, + 0.04453355818986893, + 0.0068941120989620686, + 0.03474081680178642, + 0.005357616581022739, + 0.0054304697550833225, + 0.008390967734158039, + 0.008218830451369286, + 0.01704946905374527, + 0.015179416164755821, + 0.005309592466801405, + 0.007978569716215134, + 0.003111449768766761, + 0.003640835639089346, + 0.008116590790450573, + 0.007242633029818535, + 0.011597630567848682, + 0.011192958801984787, + 0.0036174894776195288, + 0.003987292293459177, + 0.0036660826299339533, + 0.0037487053778022528, + 0.006844443269073963, + 0.0030772495083510876, + 0.005844705272465944, + 0.003637246321886778, + 0.0022192387841641903, + 0.0026780643966048956, + 0.004644228145480156, + 0.0033139584120362997, + 0.00429203687235713, + 0.0037494292482733727, + 0.029065987095236778, + 0.010640377178788185, + 0.00690426817163825, + 0.02708894945681095, + 0.010519388131797314, + 0.0055575440637767315, + 0.005627085454761982, + 0.009518292732536793, + 0.02229311875998974, + 0.08053777366876602, + 0.013293297961354256, + 0.07437656074762344, + 0.014942904934287071, + 0.007170108146965504, + 0.012110498733818531, + 0.012333802878856659, + 0.012245002202689648, + 0.07357606291770935, + 0.0155753493309021, + 0.07862422615289688, + 0.0042149703949689865, + 0.0034642645623534918, + 0.00555670540779829, + 0.009390552528202534, + 0.004291078075766563, + 0.017421497032046318, + 0.009351280517876148, + 0.016577936708927155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.028243018314242363 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208055093884468 + ], + "opposite_logprob_tokens": [ + -2.8120806217193604 + ], + "logprob_reward": -0.06208055093884468, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208055093884468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07777208089828491, + 0.04347827658057213, + 0.03540937229990959, + 0.10069083422422409, + 0.027652231976389885, + 0.012637173756957054, + 0.21276986598968506, + 0.12266174703836441, + 0.33102861046791077, + 0.03589976578950882 + ], + "credit_map": [ + 0.006349046248942614, + 0.008474405854940414, + 0.008191174827516079, + 0.005684136878699064, + 0.004791933111846447, + 0.010360434651374817, + 0.020500699058175087, + 0.013420246541500092, + 0.00497785210609436, + 0.008667663671076298, + 0.006157011725008488, + 0.0048479740507900715, + 0.00457360502332449, + 0.00404158653691411, + 0.0048430124297738075, + 0.005369570571929216, + 0.0036788044963032007, + 0.0026793740689754486, + 0.0034254202619194984, + 0.0022503375075757504, + 0.0035627197939902544, + 0.00837213546037674, + 0.003618597984313965, + 0.007821982726454735, + 0.006534735206514597, + 0.005310620181262493, + 0.006504559889435768, + 0.007513111457228661, + 0.02425112947821617, + 0.03647486865520477, + 0.005172138102352619, + 0.008929671719670296, + 0.0027645286172628403, + 0.0022606763523072004, + 0.0038552337791770697, + 0.0037068233359605074, + 0.005637237336486578, + 0.004016179591417313, + 0.002641428727656603, + 0.0027701242361217737, + 0.0011255844729021192, + 0.0012778642121702433, + 0.002200057730078697, + 0.0020465045236051083, + 0.001816505566239357, + 0.001343250973150134, + 0.0013561465311795473, + 0.0014712599804624915, + 0.008452636189758778, + 0.007238066755235195, + 0.007346386555582285, + 0.007026421371847391, + 0.0730052962899208, + 0.026514574885368347, + 0.015780379995703697, + 0.06740611046552658, + 0.005522910971194506, + 0.0021774768829345703, + 0.0038547676522284746, + 0.00512062618508935, + 0.013168304227292538, + 0.0472724623978138, + 0.006176798604428768, + 0.0393683947622776, + 0.015252193436026573, + 0.004279100801795721, + 0.00639735721051693, + 0.015889635309576988, + 0.018789377063512802, + 0.12498660385608673, + 0.016881274059414864, + 0.12855306267738342, + 0.003705487586557865, + 0.0022486925590783358, + 0.006056353449821472, + 0.004680844955146313, + 0.002949122339487076, + 0.005202176980674267, + 0.006202194839715958, + 0.00485489284619689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.06208055093884468 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012569610960781574 + ], + "opposite_logprob_tokens": [ + -4.387569427490234 + ], + "logprob_reward": -0.012569610960781574, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012569610960781574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06247950717806816, + 0.07005370408296585, + 0.039153698831796646, + 0.027653446421027184, + 0.03444567695260048, + 0.02597769722342491, + 0.0972113162279129, + 0.16823358833789825, + 0.43838176131248474, + 0.036409586668014526 + ], + "credit_map": [ + 0.0056169405579566956, + 0.0072564356960356236, + 0.007668268401175737, + 0.006154786795377731, + 0.005463809240609407, + 0.012971637770533562, + 0.007591350469738245, + 0.009756283834576607, + 0.0050103021785616875, + 0.009605796076357365, + 0.008057937026023865, + 0.006522570736706257, + 0.00400855066254735, + 0.017698561772704124, + 0.005040958523750305, + 0.014109030365943909, + 0.001410751254297793, + 0.001975569175556302, + 0.0029220273718237877, + 0.0026217419654130936, + 0.0044722226448357105, + 0.014354289509356022, + 0.0020663475152105093, + 0.009330755099654198, + 0.002226408338174224, + 0.0021311682648956776, + 0.0032574704382568598, + 0.003448910778388381, + 0.006011130288243294, + 0.006778071168810129, + 0.0014291913248598576, + 0.0023711000103503466, + 0.003741917898878455, + 0.001957597443833947, + 0.006057058926671743, + 0.003792999777942896, + 0.007031256332993507, + 0.00685247452929616, + 0.002532593673095107, + 0.002479781396687031, + 0.0019915977027267218, + 0.0021078113932162523, + 0.005734665784984827, + 0.003735301550477743, + 0.005062389187514782, + 0.003277717623859644, + 0.0018935427069664001, + 0.0021746759302914143, + 0.007949849590659142, + 0.003955668769776821, + 0.007383083458989859, + 0.005069200415164232, + 0.027685491368174553, + 0.0102359214797616, + 0.008058837614953518, + 0.0268732737749815, + 0.008075518533587456, + 0.0038270277436822653, + 0.004393021110445261, + 0.0070709544233977795, + 0.017638206481933594, + 0.06321887671947479, + 0.00909157283604145, + 0.054918449372053146, + 0.02089834213256836, + 0.00617531081661582, + 0.011922463774681091, + 0.020597418770194054, + 0.02463296614587307, + 0.16410277783870697, + 0.022488106042146683, + 0.16756445169448853, + 0.0036192089319229126, + 0.00393285509198904, + 0.005980110261589289, + 0.004265685100108385, + 0.0039946832694113255, + 0.004243527539074421, + 0.004499616101384163, + 0.005873902700841427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.012569610960781574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608831062912941 + ], + "opposite_logprob_tokens": [ + -4.141088485717773 + ], + "logprob_reward": -0.01608831062912941, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608831062912941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11682873964309692, + 0.06571999937295914, + 0.025972193107008934, + 0.055547911673784256, + 0.048864249140024185, + 0.016339777037501335, + 0.0345674604177475, + 0.17364197969436646, + 0.3752596378326416, + 0.08725805580615997 + ], + "credit_map": [ + 0.00954490341246128, + 0.009896831586956978, + 0.006873089820146561, + 0.008688698522746563, + 0.007525098510086536, + 0.021031849086284637, + 0.02899337187409401, + 0.02427489310503006, + 0.003403193550184369, + 0.0028518252074718475, + 0.006843390874564648, + 0.0032177811954170465, + 0.0034362366423010826, + 0.02222089096903801, + 0.0030428876634687185, + 0.020703790709376335, + 0.0010601402027532458, + 0.0013058383483439684, + 0.0030040880665183067, + 0.0013074518647044897, + 0.0023937702644616365, + 0.007959337905049324, + 0.0018083994509652257, + 0.0071331653743982315, + 0.003927391022443771, + 0.00464978301897645, + 0.0073575424030423164, + 0.005527934990823269, + 0.009226751513779163, + 0.014643318019807339, + 0.003315284149721265, + 0.006899903994053602, + 0.003749182214960456, + 0.0033923385199159384, + 0.008439485915005207, + 0.006148658227175474, + 0.009922010824084282, + 0.010495275259017944, + 0.003274316433817148, + 0.003442981746047735, + 0.0015089588705450296, + 0.002331840805709362, + 0.0025334618985652924, + 0.0019916261080652475, + 0.002216554945334792, + 0.002050230745226145, + 0.0019180080853402615, + 0.0017890953458845615, + 0.005891812965273857, + 0.001941076130606234, + 0.005768922623246908, + 0.003156716702505946, + 0.0074682459235191345, + 0.004381912760436535, + 0.0022957241162657738, + 0.003663049777969718, + 0.00795227475464344, + 0.0035087266005575657, + 0.0031827690545469522, + 0.006456664297729731, + 0.017877643927931786, + 0.0653771162033081, + 0.010349237360060215, + 0.05893755704164505, + 0.016322394832968712, + 0.0059626176953315735, + 0.013982033357024193, + 0.018456939607858658, + 0.021176451817154884, + 0.13806186616420746, + 0.01968909241259098, + 0.14160825312137604, + 0.005023077595978975, + 0.0033389022573828697, + 0.004383402410894632, + 0.004334405530244112, + 0.003793506184592843, + 0.03008021041750908, + 0.005255912896245718, + 0.031048642471432686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.01608831062912941 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38713064789772034 + ], + "opposite_logprob_tokens": [ + -1.137130618095398 + ], + "logprob_reward": -0.38713064789772034, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38713064789772034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08001234382390976, + 0.015691544860601425, + 0.05113231763243675, + 0.042074818164110184, + 0.0370672382414341, + 0.021661637350916862, + 0.024958377704024315, + 0.06302368640899658, + 0.5834371447563171, + 0.08094083517789841 + ], + "credit_map": [ + 0.011359951458871365, + 0.00907802488654852, + 0.006680890917778015, + 0.008089683018624783, + 0.005546144675463438, + 0.014759211800992489, + 0.010707216337323189, + 0.013791226781904697, + 0.001918273395858705, + 0.0020304257050156593, + 0.002063289051875472, + 0.0018500047735869884, + 0.001332686166279018, + 0.0019116151379421353, + 0.002502238843590021, + 0.002083013067021966, + 0.001696974504739046, + 0.0016322452574968338, + 0.0032667978666722775, + 0.002682693302631378, + 0.004168075509369373, + 0.02004251442849636, + 0.002614235505461693, + 0.015028784982860088, + 0.0031453517731279135, + 0.0038092543836683035, + 0.005787102971225977, + 0.003457298967987299, + 0.007754408288747072, + 0.011864865198731422, + 0.002681155689060688, + 0.003575384384021163, + 0.0023023001849651337, + 0.0017965375445783138, + 0.005293041933327913, + 0.003640747396275401, + 0.008658760227262974, + 0.009990118443965912, + 0.002356118755415082, + 0.0030296221375465393, + 0.0021245847456157207, + 0.0034379411954432726, + 0.004071043338626623, + 0.0025981783401221037, + 0.0027959258295595646, + 0.0021817400120198727, + 0.0025251691695302725, + 0.0019270547199994326, + 0.0033322181552648544, + 0.0015986969228833914, + 0.004126558545976877, + 0.003054667031392455, + 0.00357788591645658, + 0.0019247743766754866, + 0.003123144619166851, + 0.004220432601869106, + 0.005083165131509304, + 0.004970721434801817, + 0.005365932360291481, + 0.007378258742392063, + 0.0056029134429991245, + 0.012075825594365597, + 0.011136040091514587, + 0.011410829611122608, + 0.02742915041744709, + 0.006840662099421024, + 0.015783097594976425, + 0.027815930545330048, + 0.03300553560256958, + 0.219058558344841, + 0.028873490169644356, + 0.22463075816631317, + 0.004928813315927982, + 0.003658011555671692, + 0.006866248790174723, + 0.00583186699077487, + 0.004406323656439781, + 0.02399764209985733, + 0.008409875445067883, + 0.022842062637209892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.38713064789772034 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, statically shot, slightly tilted from above, captures a scene with the dramatic tension and high-contrast lighting of a classic martial arts film. The frame is almost entirely filled by the sushi master's hands, his weathered yet spotless fingers revealing the focused precision of a seasoned warrior. Both hands hold a clean, hand-rolled roll of crisp, dark green nori, its texture nuanced under the sharp spotlight. Inside the cone, pearly grains of vinegar-flavored rice are perfectly arranged around a core of premium, crimson tuna, its surface gleaming with natural oils. Before making the final, decisive move, the master's other hand remains poised and silent, holding a delicate brush smothered in a drop of thick, iridescent, dark sauce that swirls over the tuna. The sauce, almost black, catches the key light, revealing hues of deep amber and mahogany. In a moment of suspended action, the drop is meticulously applied, landing with controlled slowness on the center of the tuna, where it instantly solidifies, its surface tension forming a perfect, shimmering dome – a solid jewel of flavor that catches the light, creating a powerful visual climax that elevates the act of cooking to a ritual of profound care and artistry.", + "seed": 5837, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_6/index_7743_seed_5837_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20074635092169046, + "vlm_logprob": -0.20074635092169046, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.20074635092169046, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08462362736463547 + ], + "opposite_logprob_tokens": [ + -2.9596235752105713 + ], + "logprob_reward": -0.08462362736463547, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08462362736463547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07743960618972778, + 0.05541684851050377, + 0.0903463289141655, + 0.07864140719175339, + 0.09650354087352753, + 0.3245292007923126, + 0.06667736172676086, + 0.0897703543305397, + 0.028770139440894127, + 0.09190521389245987 + ], + "credit_map": [ + 0.013429722748696804, + 0.011055265553295612, + 0.01030869223177433, + 0.009316300973296165, + 0.006219492759555578, + 0.0071818772703409195, + 0.006847039330750704, + 0.01308121532201767, + 0.010609371587634087, + 0.0060016922652721405, + 0.007530309725552797, + 0.008325202390551567, + 0.004733555018901825, + 0.0046950592659413815, + 0.004832036793231964, + 0.008689621463418007, + 0.013197219930589199, + 0.008099716156721115, + 0.009963545016944408, + 0.012838970869779587, + 0.015809644013643265, + 0.008511101827025414, + 0.00781513936817646, + 0.014110998250544071, + 0.013598446734249592, + 0.007425698451697826, + 0.009922773577272892, + 0.012401627376675606, + 0.005856845527887344, + 0.007097467314451933, + 0.010128555819392204, + 0.012209990993142128, + 0.016244225203990936, + 0.011669861152768135, + 0.01347198523581028, + 0.014835088513791561, + 0.008774129673838615, + 0.007750395219773054, + 0.010180532932281494, + 0.013577315956354141, + 0.05112786218523979, + 0.03592752292752266, + 0.040821533650159836, + 0.053454168140888214, + 0.029087146744132042, + 0.026209350675344467, + 0.043023042380809784, + 0.044878579676151276, + 0.01077444851398468, + 0.006491499021649361, + 0.010721590369939804, + 0.009798438288271427, + 0.009321626275777817, + 0.005181404761970043, + 0.007648096885532141, + 0.006740263197571039, + 0.014059765264391899, + 0.011937522329390049, + 0.009708722122013569, + 0.012475336901843548, + 0.009952116757631302, + 0.00786077044904232, + 0.01038371678441763, + 0.013392404653131962, + 0.004905350040644407, + 0.003651315113529563, + 0.0029176496900618076, + 0.00396315474063158, + 0.0033873808570206165, + 0.0025811109226197004, + 0.0036856611259281635, + 0.0036785167176276445, + 0.015086286701261997, + 0.009579270146787167, + 0.011234572157263756, + 0.012507913634181023, + 0.008553819730877876, + 0.009997482411563396, + 0.008390696719288826, + 0.0165551770478487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.08462362736463547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062092650681734085 + ], + "opposite_logprob_tokens": [ + -2.8120925426483154 + ], + "logprob_reward": -0.062092650681734085, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062092650681734085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.112512968480587, + 0.06866729259490967, + 0.05754319205880165, + 0.06520658731460571, + 0.10826605558395386, + 0.17663758993148804, + 0.018737848848104477, + 0.2578293979167938, + 0.03710811585187912, + 0.09749090671539307 + ], + "credit_map": [ + 0.01913038268685341, + 0.01580909639596939, + 0.018181895837187767, + 0.012520628049969673, + 0.01107843965291977, + 0.011364259757101536, + 0.009740360081195831, + 0.014687903225421906, + 0.011949559673666954, + 0.006189414300024509, + 0.009852273389697075, + 0.008179128170013428, + 0.004790263716131449, + 0.004722826648503542, + 0.007316847797483206, + 0.01566697657108307, + 0.008740535005927086, + 0.005355892702937126, + 0.007256354205310345, + 0.008798073045909405, + 0.00899048987776041, + 0.006037166807800531, + 0.0056518083438277245, + 0.006712873466312885, + 0.00985816027969122, + 0.009155334904789925, + 0.009624297730624676, + 0.00889518391340971, + 0.0052019101567566395, + 0.0070464289747178555, + 0.007669989950954914, + 0.007755282334983349, + 0.019270190969109535, + 0.010584642179310322, + 0.014579389244318008, + 0.019136035814881325, + 0.009972963482141495, + 0.00975365936756134, + 0.01224152185022831, + 0.012727655470371246, + 0.027701154351234436, + 0.019406024366617203, + 0.02203289419412613, + 0.028706392273306847, + 0.01573358289897442, + 0.01462772861123085, + 0.024344367906451225, + 0.024085456505417824, + 0.0029573726933449507, + 0.002091107191517949, + 0.0020093002822250128, + 0.002859259257093072, + 0.002844878938049078, + 0.0023609017953276634, + 0.0017589993076398969, + 0.001856029499322176, + 0.03993918001651764, + 0.030216049402952194, + 0.026745451614260674, + 0.03829726576805115, + 0.026443136855959892, + 0.030747953802347183, + 0.02714933454990387, + 0.03829103708267212, + 0.006849410478025675, + 0.0030660592019557953, + 0.004016366321593523, + 0.005986894015222788, + 0.002210053615272045, + 0.005299314856529236, + 0.005489264614880085, + 0.004190756473690271, + 0.014746465720236301, + 0.013417544774711132, + 0.01410314254462719, + 0.017040692269802094, + 0.008473360911011696, + 0.008727364242076874, + 0.009358231909573078, + 0.011624103412032127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.062092650681734085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018229445442557335 + ], + "opposite_logprob_tokens": [ + -4.0182294845581055 + ], + "logprob_reward": -0.018229445442557335, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018229445442557335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10028638690710068, + 0.07795212417840958, + 0.06047029048204422, + 0.04144757613539696, + 0.04968653619289398, + 0.5253072381019592, + 0.020560424774885178, + 0.04600683972239494, + 0.02517705224454403, + 0.053105540573596954 + ], + "credit_map": [ + 0.016967155039310455, + 0.010535388253629208, + 0.010835003107786179, + 0.01289706863462925, + 0.013823327608406544, + 0.010689388029277325, + 0.011760760098695755, + 0.012778286822140217, + 0.013533554971218109, + 0.007977093569934368, + 0.013300337828695774, + 0.01338923443108797, + 0.005811788607388735, + 0.0067244418896734715, + 0.00772686256095767, + 0.00948879774659872, + 0.008236601948738098, + 0.005593756679445505, + 0.007789903320372105, + 0.008417107164859772, + 0.010914058424532413, + 0.005466441158205271, + 0.0065012224949896336, + 0.007551191840320826, + 0.006519683636724949, + 0.00681150471791625, + 0.00635980861261487, + 0.005428649019449949, + 0.0032733504194766283, + 0.005156803410500288, + 0.0035777126904577017, + 0.0043200599029660225, + 0.0066293952986598015, + 0.004349692724645138, + 0.005332990549504757, + 0.006557302083820105, + 0.007597734685987234, + 0.004490358754992485, + 0.005152457859367132, + 0.009576600976288319, + 0.08157090842723846, + 0.05503877252340317, + 0.06245457008481026, + 0.08632780611515045, + 0.04887666925787926, + 0.045362688601017, + 0.07195413112640381, + 0.07372164726257324, + 0.0030254682060331106, + 0.0034610358998179436, + 0.003010930959135294, + 0.0027949875220656395, + 0.002392098307609558, + 0.0017836794722825289, + 0.0017481149407103658, + 0.0023441046942025423, + 0.007045416627079248, + 0.005564246326684952, + 0.004527474287897348, + 0.00699725653976202, + 0.005516743753105402, + 0.0051895142532885075, + 0.004789278376847506, + 0.006376903969794512, + 0.004073928575962782, + 0.0034625260159373283, + 0.0030844530556350946, + 0.0035606692545115948, + 0.002592132892459631, + 0.0026328398380428553, + 0.0028778675477951765, + 0.002892630873247981, + 0.008324643597006798, + 0.005314153619110584, + 0.0062381369061768055, + 0.007846544496715069, + 0.0053659589029848576, + 0.005578105337917805, + 0.00591434258967638, + 0.008523653261363506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.018229445442557335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553836219012737 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553836219012737, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553836219012737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18574288487434387, + 0.0866103544831276, + 0.055012915283441544, + 0.056329451501369476, + 0.15656138956546783, + 0.0807933434844017, + 0.029661258682608604, + 0.09925545006990433, + 0.04760267585515976, + 0.20243024826049805 + ], + "credit_map": [ + 0.028849443420767784, + 0.017631027847528458, + 0.021318215876817703, + 0.024632975459098816, + 0.02377140522003174, + 0.016910487785935402, + 0.016439713537693024, + 0.03618961572647095, + 0.014881441369652748, + 0.011775671504437923, + 0.012506330385804176, + 0.01050954032689333, + 0.008644623681902885, + 0.008767500519752502, + 0.008425005711615086, + 0.011100242845714092, + 0.008957368321716785, + 0.0060681188479065895, + 0.008469223976135254, + 0.007619061507284641, + 0.005869083572179079, + 0.005001965910196304, + 0.004299907945096493, + 0.008728181943297386, + 0.010414828546345234, + 0.004226289223879576, + 0.004927508067339659, + 0.005520255770534277, + 0.006140520330518484, + 0.004579484462738037, + 0.005749203730374575, + 0.014771359972655773, + 0.022487429901957512, + 0.01748676598072052, + 0.019612988457083702, + 0.03044290468096733, + 0.02203364297747612, + 0.012433091178536415, + 0.01635960303246975, + 0.01570497825741768, + 0.012194217182695866, + 0.008572723716497421, + 0.009909632615745068, + 0.012861885130405426, + 0.009101394563913345, + 0.006816984154284, + 0.010219617746770382, + 0.01111688930541277, + 0.0042506936006248, + 0.005368982441723347, + 0.00483328104019165, + 0.0039377156645059586, + 0.003929623868316412, + 0.0022167020943015814, + 0.0023397558834403753, + 0.002784504322335124, + 0.015219558030366898, + 0.013935272581875324, + 0.009795060381293297, + 0.012843123637139797, + 0.008639948442578316, + 0.012292596511542797, + 0.011796573176980019, + 0.014733314514160156, + 0.008912274613976479, + 0.005073621403425932, + 0.005051409360021353, + 0.0064163729548454285, + 0.0054268804378807545, + 0.005663599353283644, + 0.006168115884065628, + 0.00489040045067668, + 0.03141506016254425, + 0.027421990409493446, + 0.028523830696940422, + 0.03370562568306923, + 0.020914273336529732, + 0.01765511929988861, + 0.018543636426329613, + 0.024250714108347893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.012553836219012737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262321949005127 + ], + "opposite_logprob_tokens": [ + -0.5762321949005127 + ], + "logprob_reward": -0.8262321949005127, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262321949005127, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10690479725599289, + 0.10253072530031204, + 0.10862976312637329, + 0.05986299365758896, + 0.15865929424762726, + 0.1529085487127304, + 0.02578859031200409, + 0.07070863991975784, + 0.06973110139369965, + 0.14427553117275238 + ], + "credit_map": [ + 0.017983442172408104, + 0.010295473970472813, + 0.011272652074694633, + 0.014462659135460854, + 0.012659767642617226, + 0.010873298160731792, + 0.011194910854101181, + 0.018162591382861137, + 0.017775164917111397, + 0.01213861908763647, + 0.01883559301495552, + 0.017527777701616287, + 0.008116418495774269, + 0.008676714263856411, + 0.009308247826993465, + 0.010152180679142475, + 0.017724866047501564, + 0.010130531154572964, + 0.012361587956547737, + 0.015451625920832157, + 0.018610302358865738, + 0.009405777789652348, + 0.009920218959450722, + 0.015024849213659763, + 0.010069622658193111, + 0.005475426092743874, + 0.0071679940447211266, + 0.009059895761311054, + 0.006264294032007456, + 0.006827216129750013, + 0.005144353490322828, + 0.009854190051555634, + 0.02548668533563614, + 0.019522665068507195, + 0.022633180022239685, + 0.022397847846150398, + 0.01727939397096634, + 0.013246772810816765, + 0.016473382711410522, + 0.021619362756609917, + 0.023395303636789322, + 0.015642711892724037, + 0.018615705892443657, + 0.023919126018881798, + 0.01437351480126381, + 0.013568154536187649, + 0.020235570147633553, + 0.02315845526754856, + 0.004221249837428331, + 0.0048211608082056046, + 0.005301847122609615, + 0.002843582071363926, + 0.0020165343303233385, + 0.0017731909174472094, + 0.0024798947852104902, + 0.002331129275262356, + 0.0126950116828084, + 0.006051710806787014, + 0.009561223909258842, + 0.00837365910410881, + 0.005100111942738295, + 0.010193648748099804, + 0.009892517700791359, + 0.008840756490826607, + 0.013454163447022438, + 0.008424272760748863, + 0.006030289456248283, + 0.008788657374680042, + 0.0074378433637320995, + 0.008910484611988068, + 0.0077360039576888084, + 0.008949389681220055, + 0.023590054363012314, + 0.01554072741419077, + 0.017837481573224068, + 0.0207563117146492, + 0.015546985901892185, + 0.01368003711104393, + 0.0182592011988163, + 0.019064728170633316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.8262321949005127 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, statically shot, slightly tilted from above, captures a scene with the dramatic tension and high-contrast lighting of a classic martial arts film. The frame is almost entirely filled by the sushi master's hands, his weathered yet spotless fingers revealing the focused precision of a seasoned warrior. Both hands hold a clean, hand-rolled roll of crisp, dark green nori, its texture nuanced under the sharp spotlight. Inside the cone, pearly grains of vinegar-flavored rice are perfectly arranged around a core of premium, crimson tuna, its surface gleaming with natural oils. Before making the final, decisive move, the master's other hand remains poised and silent, holding a delicate brush smothered in a drop of thick, iridescent, dark sauce that swirls over the tuna. The sauce, almost black, catches the key light, revealing hues of deep amber and mahogany. In a moment of suspended action, the drop is meticulously applied, landing with controlled slowness on the center of the tuna, where it instantly solidifies, its surface tension forming a perfect, shimmering dome – a solid jewel of flavor that catches the light, creating a powerful visual climax that elevates the act of cooking to a ritual of profound care and artistry.", + "seed": 5838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_6/index_7743_seed_5838_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.23767392188310624, + "vlm_logprob": -0.23767392188310624, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.23767392188310624, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04975161701440811 + ], + "opposite_logprob_tokens": [ + -3.4247515201568604 + ], + "logprob_reward": -0.04975161701440811, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04975161701440811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16991238296031952, + 0.0337197370827198, + 0.36572206020355225, + 0.06545652449131012, + 0.01892426796257496, + 0.08237918466329575, + 0.1367974430322647, + 0.03210750222206116, + 0.044396284967660904, + 0.050584644079208374 + ], + "credit_map": [ + 0.02407952770590782, + 0.02068423107266426, + 0.02700556069612503, + 0.0195597093552351, + 0.017574401572346687, + 0.0163127388805151, + 0.016485080122947693, + 0.028211131691932678, + 0.00447476701810956, + 0.0030609597451984882, + 0.006521299947053194, + 0.002665234263986349, + 0.00362766464240849, + 0.003778092795982957, + 0.003016470465809107, + 0.006575246807187796, + 0.03297543153166771, + 0.006456949282437563, + 0.1397646814584732, + 0.008868821896612644, + 0.007018380332738161, + 0.026331325992941856, + 0.006694172043353319, + 0.13761231303215027, + 0.005465835332870483, + 0.0030378736555576324, + 0.02005009911954403, + 0.002673652023077011, + 0.006261065136641264, + 0.004550142213702202, + 0.0022777297999709845, + 0.021140120923519135, + 0.002614238066598773, + 0.0027888326440006495, + 0.0026037413626909256, + 0.0025053401477634907, + 0.0027408648747950792, + 0.0016001448966562748, + 0.0014259106246754527, + 0.002645194996148348, + 0.010719181038439274, + 0.010522473603487015, + 0.014139557257294655, + 0.009626798331737518, + 0.009898904711008072, + 0.00712927570566535, + 0.010119394399225712, + 0.010223598219454288, + 0.018224241212010384, + 0.015251575037837029, + 0.018617551773786545, + 0.016252506524324417, + 0.017123237252235413, + 0.017119113355875015, + 0.010182937607169151, + 0.024026282131671906, + 0.005699101369827986, + 0.004177679307758808, + 0.004903297405689955, + 0.0024603509809821844, + 0.0025481318589299917, + 0.004515224602073431, + 0.0023048529401421547, + 0.005498865619301796, + 0.008544864133000374, + 0.008554792031645775, + 0.005974726751446724, + 0.003099258290603757, + 0.0033254814334213734, + 0.004632667172700167, + 0.00299608102068305, + 0.007268413435667753, + 0.008234366774559021, + 0.006927958223968744, + 0.007832365110516548, + 0.003933127969503403, + 0.004156461916863918, + 0.007508872542530298, + 0.004023524932563305, + 0.007967963814735413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.04975161701440811 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311158508062363 + ], + "opposite_logprob_tokens": [ + -3.168111562728882 + ], + "logprob_reward": -0.04311158508062363, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311158508062363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19616590440273285, + 0.02393505908548832, + 0.509567379951477, + 0.07395265251398087, + 0.01429701317101717, + 0.03676033765077591, + 0.03698558732867241, + 0.05394411087036133, + 0.027379216626286507, + 0.027012765407562256 + ], + "credit_map": [ + 0.021486209705471992, + 0.024119695648550987, + 0.0288913082331419, + 0.021924663335084915, + 0.02368130162358284, + 0.018234822899103165, + 0.02527346834540367, + 0.032554443925619125, + 0.002610487164929509, + 0.003021690994501114, + 0.004109353292733431, + 0.0032963554840534925, + 0.002081869402900338, + 0.0025015342980623245, + 0.002994310110807419, + 0.0033194597344845533, + 0.04554640129208565, + 0.009104837663471699, + 0.19308631122112274, + 0.012254002504050732, + 0.010064450092613697, + 0.03751075267791748, + 0.008977120742201805, + 0.19302350282669067, + 0.0058682700619101524, + 0.002248445525765419, + 0.02622511051595211, + 0.0027838980313390493, + 0.002466977806761861, + 0.005253937095403671, + 0.002222257200628519, + 0.02688375487923622, + 0.0015070142690092325, + 0.0010136390337720513, + 0.003599158488214016, + 0.001265200087800622, + 0.0011403521057218313, + 0.000807038857601583, + 0.0008359737112186849, + 0.004128636326640844, + 0.0038570815231651068, + 0.005282857920974493, + 0.00445450097322464, + 0.005912228487432003, + 0.004896977916359901, + 0.0035201001446694136, + 0.004684912506490946, + 0.004151678644120693, + 0.003212584415450692, + 0.003603219287469983, + 0.005928108934313059, + 0.004455088172107935, + 0.004426619969308376, + 0.003035716712474823, + 0.005199695937335491, + 0.007124550640583038, + 0.013189880177378654, + 0.006866616662591696, + 0.011165631003677845, + 0.003347711870446801, + 0.002091063652187586, + 0.0066650849767029285, + 0.002719650976359844, + 0.007898470386862755, + 0.00567633705213666, + 0.003337770001962781, + 0.004436133895069361, + 0.002171177417039871, + 0.0017075169598683715, + 0.003194201271981001, + 0.0016762890154495835, + 0.0051797921769320965, + 0.004924384877085686, + 0.004345618654042482, + 0.0036946674808859825, + 0.001714782090857625, + 0.0017348218243569136, + 0.0033758990466594696, + 0.002183580072596669, + 0.005039009265601635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.04311158508062363 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018230030313134193 + ], + "opposite_logprob_tokens": [ + -4.018229961395264 + ], + "logprob_reward": -0.018230030313134193, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018230030313134193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19488486647605896, + 0.05082320049405098, + 0.2303842157125473, + 0.14863194525241852, + 0.04272732138633728, + 0.1280898004770279, + 0.048021361231803894, + 0.06479141116142273, + 0.04662828519940376, + 0.04501757770776749 + ], + "credit_map": [ + 0.03106854483485222, + 0.03487597033381462, + 0.03115369938313961, + 0.016987653449177742, + 0.013084297999739647, + 0.0164953600615263, + 0.01604885421693325, + 0.035170476883649826, + 0.007489717099815607, + 0.004442220088094473, + 0.01197994314134121, + 0.006526629440486431, + 0.004333628341555595, + 0.0037879494484514, + 0.006282363552600145, + 0.005980750545859337, + 0.020253490656614304, + 0.004997094161808491, + 0.08503077924251556, + 0.005970761179924011, + 0.0052599431946873665, + 0.016108907759189606, + 0.005001164972782135, + 0.08776205778121948, + 0.012023385614156723, + 0.0043968199752271175, + 0.052041083574295044, + 0.005568106193095446, + 0.006478512194007635, + 0.009924110025167465, + 0.004131984896957874, + 0.0540679506957531, + 0.004358875565230846, + 0.005291956476867199, + 0.009611409157514572, + 0.005937340669333935, + 0.0028109827544540167, + 0.0029878977220505476, + 0.0025846557691693306, + 0.00914420373737812, + 0.014629194512963295, + 0.017071083188056946, + 0.015152599662542343, + 0.019217655062675476, + 0.01805749349296093, + 0.014990178868174553, + 0.013891054317355156, + 0.015080542303621769, + 0.005657549947500229, + 0.004463996738195419, + 0.006825053598731756, + 0.008164891041815281, + 0.003950833342969418, + 0.006342697888612747, + 0.0044708275236189365, + 0.008145513013005257, + 0.012985690496861935, + 0.007276454009115696, + 0.009311581030488014, + 0.007757486775517464, + 0.0048299673944711685, + 0.007254299707710743, + 0.0066426824778318405, + 0.008733251132071018, + 0.007813678123056889, + 0.005975318141281605, + 0.007353080436587334, + 0.004896756261587143, + 0.003457834245637059, + 0.006278508342802525, + 0.004557297099381685, + 0.0062958127819001675, + 0.00570951821282506, + 0.00666277389973402, + 0.006037168670445681, + 0.004259245935827494, + 0.0033152038231492043, + 0.006339810788631439, + 0.004673932678997517, + 0.008019921369850636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.018230030313134193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233097355812788 + ], + "opposite_logprob_tokens": [ + -3.7733097076416016 + ], + "logprob_reward": -0.0233097355812788, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233097355812788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11491382867097855, + 0.03682006523013115, + 0.25730398297309875, + 0.3304517865180969, + 0.012459762394428253, + 0.08245205134153366, + 0.029190722852945328, + 0.04423397034406662, + 0.04094649851322174, + 0.05122733488678932 + ], + "credit_map": [ + 0.023577973246574402, + 0.01827477291226387, + 0.016989275813102722, + 0.012636931613087654, + 0.006958300713449717, + 0.008224858902394772, + 0.013487391173839569, + 0.014764323830604553, + 0.005621945019811392, + 0.0055322363041341305, + 0.0063795531168580055, + 0.00460910564288497, + 0.002841735491529107, + 0.003536531003192067, + 0.003781835315749049, + 0.004517124500125647, + 0.022817589342594147, + 0.004659476224333048, + 0.09686669707298279, + 0.0069666230119764805, + 0.005960979964584112, + 0.01792651042342186, + 0.005057561211287975, + 0.09704853594303131, + 0.02791830524802208, + 0.006954425014555454, + 0.1262694001197815, + 0.008917356841266155, + 0.00923192873597145, + 0.021648379042744637, + 0.007127503864467144, + 0.12238450348377228, + 0.00142563134431839, + 0.0009368071332573891, + 0.0022267289459705353, + 0.0013379496522247791, + 0.001381309237331152, + 0.0012983415508642793, + 0.0007423177012242377, + 0.0031106758397072554, + 0.006876120809465647, + 0.010357091203331947, + 0.010843779891729355, + 0.012702150270342827, + 0.01148237008601427, + 0.008161245845258236, + 0.01120345015078783, + 0.010825845412909985, + 0.004568117670714855, + 0.0028611442539840937, + 0.004519713576883078, + 0.002959441626444459, + 0.0034967074170708656, + 0.002813822589814663, + 0.0042074439115822315, + 0.003764333436265588, + 0.0061379168182611465, + 0.005199130158871412, + 0.005990137346088886, + 0.00692888954654336, + 0.0044561405666172504, + 0.004374208394438028, + 0.005080040078610182, + 0.00606750650331378, + 0.005491779651492834, + 0.006807343102991581, + 0.0047341338358819485, + 0.0053781489841639996, + 0.003973723389208317, + 0.005226870998740196, + 0.004099228885024786, + 0.005235269200056791, + 0.005346359219402075, + 0.005433869082480669, + 0.006331831216812134, + 0.007151543162763119, + 0.005236814729869366, + 0.00525401346385479, + 0.00718259159475565, + 0.009290310554206371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0233097355812788 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up where the frame is almost entirely filled by the sushi master's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands holding a clean, hand-rolled sushi roll with dark green nori, pearly rice, and a core of crimson tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second hand visible holding a delicate brush with a drop of thick, iridescent dark sauce poised over the tuna?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the drop of sauce landing on the tuna and instantly solidifying into a perfect, shimmering dome?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539666414260864 + ], + "opposite_logprob_tokens": [ + -0.4289666414260864 + ], + "logprob_reward": -1.0539666414260864, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539666414260864, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3161007761955261, + 0.053822774440050125, + 0.23395445942878723, + 0.12000653147697449, + 0.03684956952929497, + 0.0469612255692482, + 0.024814197793602943, + 0.03529300540685654, + 0.09935019910335541, + 0.03284718841314316 + ], + "credit_map": [ + 0.05274711176753044, + 0.04823349416255951, + 0.04798372462391853, + 0.04409138485789299, + 0.0230715312063694, + 0.02608184702694416, + 0.033757343888282776, + 0.04013437032699585, + 0.005627342965453863, + 0.006506825797259808, + 0.009711585007607937, + 0.008819947950541973, + 0.0045290119014680386, + 0.005169824697077274, + 0.006673191208392382, + 0.006785046774893999, + 0.020867474377155304, + 0.004480887204408646, + 0.08629268407821655, + 0.007111790124326944, + 0.005171319004148245, + 0.016637133434414864, + 0.004914627410471439, + 0.08847855031490326, + 0.009664763696491718, + 0.0035591796040534973, + 0.0440361462533474, + 0.004322055261582136, + 0.003742367494851351, + 0.008487063460052013, + 0.002829781034961343, + 0.043365176767110825, + 0.005005026236176491, + 0.005503058433532715, + 0.0055289240553975105, + 0.006459480617195368, + 0.0033732850570231676, + 0.0036316332407295704, + 0.0030416022054851055, + 0.004306559916585684, + 0.00598136754706502, + 0.006272002123296261, + 0.005427927244454622, + 0.006887126713991165, + 0.005432609934359789, + 0.003742712549865246, + 0.006205535028129816, + 0.007011943496763706, + 0.003448182949796319, + 0.003104612696915865, + 0.0038203930016607046, + 0.0033552153035998344, + 0.002377455122768879, + 0.002217591041699052, + 0.003332071704789996, + 0.0031586741097271442, + 0.006935986690223217, + 0.004020045045763254, + 0.0059695797972381115, + 0.0032246585469692945, + 0.0024887416511774063, + 0.005024325102567673, + 0.002173439599573612, + 0.005456228274852037, + 0.012415871024131775, + 0.016117479652166367, + 0.011285658925771713, + 0.01408039964735508, + 0.010675722733139992, + 0.01240403763949871, + 0.010365919210016727, + 0.012005115859210491, + 0.004980254452675581, + 0.00465673953294754, + 0.004930146969854832, + 0.003052130341529846, + 0.0031931728590279818, + 0.003324531950056553, + 0.0029410412535071373, + 0.005769170355051756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -1.0539666414260864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_tensor([1458, 1458])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_tensor([1458, 1458])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5401e746ef9ae5652ee234bf7d38fb058b059b4c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_tensor([1458, 1458])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly low-angle close-up captures a scene of quiet reverence. A slender, pale hand, its movements steady and deliberate, holds a vintage hardcover bound in dark green leather, its surface showing the subtle, yet noble wear of time. Warm, focused light from an unseen source gleams on the ornate, almost illegible gilded lettering and filigree on the book's spine. The hand carefully guides the book toward a narrow opening in a towering, floor-to-ceiling bookshelf crafted from dark polished mahogany. The backdrop is a dense, richly textured mosaic of countless other antique books, their leather-bound spines forming a rich tapestry of crimson, brown, and blue. With meticulous precision, the hand lowers the book into place, the spine sliding smoothly into its neighbor until it lies perfectly flush, adding another volume to the vast, silent collection. The fingers linger for a moment, gently touching the spine, then slowly withdraw, immersing the scene in a soft glow and the subtle scent of old paper, a tranquil, scholarly atmosphere.", + "seed": 1167, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_1458_seed_1167_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01487744809128344, + "vlm_logprob": -0.01487744809128344, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.01487744809128344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04780462384223938 + ], + "opposite_logprob_tokens": [ + -4.297804832458496 + ], + "logprob_reward": -0.04780462384223938, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.04780462384223938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06990871578454971, + 0.0475308820605278, + 0.365072637796402, + 0.11939535290002823, + 0.0435669906437397, + 0.05780684947967529, + 0.01870652474462986, + 0.04067312180995941, + 0.07868772745132446, + 0.15865111351013184 + ], + "credit_map": [ + 0.00908061396330595, + 0.010922054760158062, + 0.00998274702578783, + 0.009370864368975163, + 0.007853315211832523, + 0.009128205478191376, + 0.007486371323466301, + 0.006084543187171221, + 0.004420544020831585, + 0.005352491978555918, + 0.007176525425165892, + 0.0063108885660767555, + 0.005265166983008385, + 0.005027489736676216, + 0.006648675072938204, + 0.007329100742936134, + 0.014007394202053547, + 0.009713873267173767, + 0.007433405611664057, + 0.00645674392580986, + 0.14453917741775513, + 0.10537584871053696, + 0.05123415216803551, + 0.026312053203582764, + 0.016404367983341217, + 0.007904959842562675, + 0.006678742356598377, + 0.007996967993676662, + 0.008384288288652897, + 0.02288263477385044, + 0.008352055214345455, + 0.040791336447000504, + 0.005614637862890959, + 0.005446211900562048, + 0.00623629754409194, + 0.0037032051477581263, + 0.003804786829277873, + 0.0075842877849936485, + 0.0028973068110644817, + 0.00828025583177805, + 0.01664172299206257, + 0.007105331867933273, + 0.007462795823812485, + 0.002850203774869442, + 0.007669589016586542, + 0.004815556108951569, + 0.0034979963675141335, + 0.007763656787574291, + 0.003353079780936241, + 0.001954648643732071, + 0.0021312085445970297, + 0.0013775720726698637, + 0.00405068090185523, + 0.0019247413147240877, + 0.001486692694015801, + 0.0024278999771922827, + 0.009341423399746418, + 0.003798741614446044, + 0.004841346759349108, + 0.0032104465644806623, + 0.006920567248016596, + 0.003916717134416103, + 0.0032722391188144684, + 0.005371640436351299, + 0.015791714191436768, + 0.00642439303919673, + 0.009353268891572952, + 0.009582320228219032, + 0.011719119735062122, + 0.00855921022593975, + 0.006772909313440323, + 0.010484788566827774, + 0.0417645126581192, + 0.01097922958433628, + 0.027493610978126526, + 0.011324923485517502, + 0.009970225393772125, + 0.01800474151968956, + 0.009958362206816673, + 0.029155509546399117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.04780462384223938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036607650108635426 + ], + "opposite_logprob_tokens": [ + -5.628660678863525 + ], + "logprob_reward": -0.0036607650108635426, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036607650108635426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04065345227718353, + 0.07553545385599136, + 0.28124773502349854, + 0.06721629202365875, + 0.04919241741299629, + 0.01883990690112114, + 0.026066536083817482, + 0.049926236271858215, + 0.30098167061805725, + 0.09034030139446259 + ], + "credit_map": [ + 0.0035568124148994684, + 0.005245555192232132, + 0.004394413437694311, + 0.006019356194883585, + 0.005309743341058493, + 0.005222908221185207, + 0.007371740881353617, + 0.003532921429723501, + 0.005953827407211065, + 0.005734490696340799, + 0.009477839805185795, + 0.007192453369498253, + 0.00963404681533575, + 0.011371880769729614, + 0.010162437334656715, + 0.01600848138332367, + 0.014210710301995277, + 0.009597059339284897, + 0.007534031290560961, + 0.00648536579683423, + 0.11897959560155869, + 0.05282866209745407, + 0.043532390147447586, + 0.028079906478524208, + 0.008788827806711197, + 0.004173701629042625, + 0.004113189876079559, + 0.003449906362220645, + 0.006412796676158905, + 0.012399611063301563, + 0.005486029665917158, + 0.022392230108380318, + 0.011533232405781746, + 0.0035811099223792553, + 0.0065434654243290424, + 0.0024746591225266457, + 0.0026751402765512466, + 0.0059730117209255695, + 0.0026366296224296093, + 0.013775170780718327, + 0.005286831874400377, + 0.002361506223678589, + 0.002493866952136159, + 0.0009409361518919468, + 0.002490123501047492, + 0.001626739394851029, + 0.0011790511198341846, + 0.0024608515668660402, + 0.0022222513798624277, + 0.0027570321690291166, + 0.002962120110169053, + 0.002671334194019437, + 0.0065060025081038475, + 0.0026947539299726486, + 0.002673145616427064, + 0.003579898504540324, + 0.014780308119952679, + 0.005433985497802496, + 0.007213547360152006, + 0.0031274722423404455, + 0.0029672905802726746, + 0.0067734853364527225, + 0.0023963465355336666, + 0.007233801297843456, + 0.05539479851722717, + 0.028048258274793625, + 0.0337086021900177, + 0.03783917427062988, + 0.05099990963935852, + 0.03521984443068504, + 0.02591683343052864, + 0.033854249864816666, + 0.024272728711366653, + 0.006104696542024612, + 0.01613467000424862, + 0.006617299746721983, + 0.005703686270862818, + 0.01092472393065691, + 0.005834528710693121, + 0.014747965149581432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.0036607650108635426 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032045699190348387 + ], + "opposite_logprob_tokens": [ + -5.753204345703125 + ], + "logprob_reward": -0.0032045699190348387, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032045699190348387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0984220951795578, + 0.10876007378101349, + 0.4993935525417328, + 0.09343855828046799, + 0.0375375859439373, + 0.053014080971479416, + 0.029229246079921722, + 0.024914715439081192, + 0.02617855928838253, + 0.029111545532941818 + ], + "credit_map": [ + 0.008604757487773895, + 0.015549436211585999, + 0.009166215546429157, + 0.01329168863594532, + 0.014253851026296616, + 0.014568294398486614, + 0.012549160048365593, + 0.010438693687319756, + 0.00805853120982647, + 0.010332719422876835, + 0.015140715055167675, + 0.008839745074510574, + 0.015616961754858494, + 0.013713677413761616, + 0.015046504326164722, + 0.022011226043105125, + 0.021826254203915596, + 0.01840771920979023, + 0.01567426323890686, + 0.01222925540059805, + 0.20796892046928406, + 0.12389267235994339, + 0.07164101302623749, + 0.027753472328186035, + 0.012791446410119534, + 0.007550446316599846, + 0.006014012731611729, + 0.004890267737209797, + 0.007491276599466801, + 0.018075380474328995, + 0.006701286416500807, + 0.029924456030130386, + 0.004703095648437738, + 0.004007987678050995, + 0.005552566144615412, + 0.0032301759347319603, + 0.004735787399113178, + 0.005945687182247639, + 0.003332820488139987, + 0.0060294694267213345, + 0.015096048824489117, + 0.006162709556519985, + 0.006960125640034676, + 0.002456687856465578, + 0.006554243620485067, + 0.00555952824652195, + 0.0030345250852406025, + 0.007190218195319176, + 0.004440854303538799, + 0.002615587320178747, + 0.0031038960441946983, + 0.001967793330550194, + 0.006785145495086908, + 0.0029545538127422333, + 0.0033939569257199764, + 0.00396746164187789, + 0.002790850820019841, + 0.002282151486724615, + 0.0024753252509981394, + 0.0027030156925320625, + 0.005164152476936579, + 0.002818380482494831, + 0.0029159937985241413, + 0.003764847991988063, + 0.004418901167809963, + 0.0022676032967865467, + 0.0032830298878252506, + 0.0033613855484873056, + 0.004610451404005289, + 0.003264713566750288, + 0.0019120908109471202, + 0.003060385584831238, + 0.005811343435198069, + 0.0023531520273536444, + 0.004476033616811037, + 0.0022070971317589283, + 0.002394114388152957, + 0.004518122877925634, + 0.002449839608743787, + 0.004901844542473555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0032045699190348387 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645959198474884 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008645959198474884, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645959198474884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11004897207021713, + 0.14292705059051514, + 0.13086684048175812, + 0.21909505128860474, + 0.05010508745908737, + 0.035234302282333374, + 0.09492349624633789, + 0.06864183396100998, + 0.08375657349824905, + 0.06440076977014542 + ], + "credit_map": [ + 0.013629436492919922, + 0.015308460220694542, + 0.01713745854794979, + 0.019025424495339394, + 0.0118954973295331, + 0.012257128953933716, + 0.012293647043406963, + 0.00850193202495575, + 0.014034719206392765, + 0.012483833357691765, + 0.023774748668074608, + 0.012501669116318226, + 0.017874376848340034, + 0.018072787672281265, + 0.018454324454069138, + 0.025730585679411888, + 0.004619741812348366, + 0.004892607685178518, + 0.005642528645694256, + 0.004424085840582848, + 0.03941246494650841, + 0.04157707095146179, + 0.012562415562570095, + 0.01773594319820404, + 0.02912358194589615, + 0.019818825647234917, + 0.013532537966966629, + 0.007039635442197323, + 0.0114120002835989, + 0.04518217220902443, + 0.015190637670457363, + 0.07779567688703537, + 0.006869526579976082, + 0.004983567167073488, + 0.007996700704097748, + 0.0040533277206122875, + 0.005050149746239185, + 0.008201015181839466, + 0.004149528220295906, + 0.008801278658211231, + 0.007082679774612188, + 0.004078995436429977, + 0.004647464957088232, + 0.0022205894347280264, + 0.004582683090120554, + 0.004246562719345093, + 0.003521814476698637, + 0.004853515885770321, + 0.01921835169196129, + 0.008368493989109993, + 0.015546194277703762, + 0.005727427080273628, + 0.015146367251873016, + 0.01237051747739315, + 0.007657311391085386, + 0.010888835415244102, + 0.016519665718078613, + 0.007992494851350784, + 0.008473332971334457, + 0.007351808715611696, + 0.005727093666791916, + 0.008049169555306435, + 0.004986193962395191, + 0.009542073123157024, + 0.012870932929217815, + 0.007978932932019234, + 0.008357596583664417, + 0.01067456603050232, + 0.01662631705403328, + 0.009808545932173729, + 0.00722276559099555, + 0.010216924361884594, + 0.010639334097504616, + 0.005474422592669725, + 0.009799950756132603, + 0.005273766815662384, + 0.005814633332192898, + 0.007729747798293829, + 0.008281806483864784, + 0.0113871144130826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.008645959198474884 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071322485804558 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071322485804558, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071322485804558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2683083415031433, + 0.05400809273123741, + 0.3096325993537903, + 0.048747047781944275, + 0.03254266455769539, + 0.07475506514310837, + 0.046960484236478806, + 0.1006915420293808, + 0.03163214027881622, + 0.03272196650505066 + ], + "credit_map": [ + 0.02646205574274063, + 0.03556006774306297, + 0.02032272145152092, + 0.03775513544678688, + 0.029596315696835518, + 0.052948351949453354, + 0.03394724801182747, + 0.0317164771258831, + 0.0043659512884914875, + 0.004932729061692953, + 0.00601620739325881, + 0.003820698941126466, + 0.007491055876016617, + 0.009003620594739914, + 0.006837224587798119, + 0.011540612205862999, + 0.010186348110437393, + 0.009532690048217773, + 0.0062119257636368275, + 0.0065263547003269196, + 0.12032755464315414, + 0.09325078129768372, + 0.04006001725792885, + 0.02353697270154953, + 0.005986787844449282, + 0.004610246047377586, + 0.0031997284386307, + 0.0020488679874688387, + 0.004798403475433588, + 0.008555087260901928, + 0.00398486340418458, + 0.015563069842755795, + 0.003776348428800702, + 0.004608507268130779, + 0.006490100175142288, + 0.00263605872169137, + 0.0023349274415522814, + 0.005773779936134815, + 0.0024859250988811255, + 0.0044370233081281185, + 0.023677906021475792, + 0.008979063481092453, + 0.010797061026096344, + 0.0032022977247834206, + 0.006477201357483864, + 0.006680845748633146, + 0.0034628910943865776, + 0.011477812193334103, + 0.006461255252361298, + 0.004455069545656443, + 0.006081629544496536, + 0.0038646163884550333, + 0.01053389348089695, + 0.004468856379389763, + 0.0048447055742144585, + 0.006250462494790554, + 0.027294568717479706, + 0.009477490559220314, + 0.013143270276486874, + 0.005507359746843576, + 0.011909245513379574, + 0.01074914075434208, + 0.007157651707530022, + 0.015452822670340538, + 0.0029852730222046375, + 0.0031885895878076553, + 0.002968929708003998, + 0.004124053753912449, + 0.005414734128862619, + 0.0054835663177073, + 0.0034471109975129366, + 0.00401988485828042, + 0.007443705108016729, + 0.0024256394244730473, + 0.004837935324758291, + 0.002885236404836178, + 0.002689816290512681, + 0.004646118730306625, + 0.002905012806877494, + 0.004888505209237337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.011071322485804558 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up captures a scene of quiet reverence. A slender, pale hand, its movements steady and deliberate, holds a vintage hardcover bound in dark green leather, its surface showing the subtle, yet noble wear of time. Warm, focused light from an unseen source gleams on the ornate, almost illegible gilded lettering and filigree on the book's spine. The hand carefully guides the book toward a narrow opening in a towering, floor-to-ceiling bookshelf crafted from dark polished mahogany. The backdrop is a dense, richly textured mosaic of countless other antique books, their leather-bound spines forming a rich tapestry of crimson, brown, and blue. With meticulous precision, the hand lowers the book into place, the spine sliding smoothly into its neighbor until it lies perfectly flush, adding another volume to the vast, silent collection. The fingers linger for a moment, gently touching the spine, then slowly withdraw, immersing the scene in a soft glow and the subtle scent of old paper, a tranquil, scholarly atmosphere.", + "seed": 1168, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_1458_seed_1168_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012899736128747463, + "vlm_logprob": -0.012899736128747463, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.012899736128747463, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02794664539396763 + ], + "opposite_logprob_tokens": [ + -4.652946472167969 + ], + "logprob_reward": -0.02794664539396763, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.02794664539396763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07580378651618958, + 0.056372128427028656, + 0.043301187455654144, + 0.04341565817594528, + 0.07268274575471878, + 0.013886003755033016, + 0.21390895545482635, + 0.36581477522850037, + 0.09289539605379105, + 0.02191941998898983 + ], + "credit_map": [ + 0.010196655057370663, + 0.0070646703243255615, + 0.006052748765796423, + 0.00946018099784851, + 0.004444746766239405, + 0.00523665314540267, + 0.016375238075852394, + 0.01697288639843464, + 0.0032784624490886927, + 0.006870192475616932, + 0.0047751604579389095, + 0.005698622204363346, + 0.012090764008462429, + 0.008436677046120167, + 0.008627400733530521, + 0.006594850216060877, + 0.0028781499713659286, + 0.0066225347109138966, + 0.0042334189638495445, + 0.007066414691507816, + 0.009029206819832325, + 0.002519926754757762, + 0.00790639128535986, + 0.00304514542222023, + 0.0035933349281549454, + 0.00568317249417305, + 0.0067075882107019424, + 0.004021117929369211, + 0.007440596353262663, + 0.003912854008376598, + 0.007088206242769957, + 0.004968787077814341, + 0.008449502289295197, + 0.0036015105433762074, + 0.01362711749970913, + 0.0040166317485272884, + 0.023711565881967545, + 0.004629755392670631, + 0.006618317682296038, + 0.008028346113860607, + 0.0012699958169832826, + 0.0011036201613023877, + 0.001585811492986977, + 0.001238058670423925, + 0.0019817266147583723, + 0.003358265617862344, + 0.0015692615415900946, + 0.0017792636062949896, + 0.015172467567026615, + 0.012192215770483017, + 0.027373893186450005, + 0.01696254126727581, + 0.07137376815080643, + 0.012523743323981762, + 0.029011670500040054, + 0.029298653826117516, + 0.03185167536139488, + 0.01684396155178547, + 0.10317467898130417, + 0.015185878612101078, + 0.12680423259735107, + 0.02781529538333416, + 0.02391812391579151, + 0.020220933482050896, + 0.009856397286057472, + 0.0063584474846720695, + 0.029246676713228226, + 0.00455453572794795, + 0.008572355844080448, + 0.022337276488542557, + 0.0041631413623690605, + 0.007806563284248114, + 0.003018095390871167, + 0.002500882837921381, + 0.003494899719953537, + 0.002595541998744011, + 0.0020336671732366085, + 0.003034190973266959, + 0.002239622175693512, + 0.003002519952133298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.02794664539396763 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001739775761961937 + ], + "opposite_logprob_tokens": [ + -6.376739978790283 + ], + "logprob_reward": -0.001739775761961937, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001739775761961937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06272060424089432, + 0.07856912910938263, + 0.06437835097312927, + 0.04295671358704567, + 0.23577943444252014, + 0.053012292832136154, + 0.22826449573040009, + 0.13211579620838165, + 0.07223927974700928, + 0.029963871464133263 + ], + "credit_map": [ + 0.008833878673613071, + 0.005495130084455013, + 0.00621689110994339, + 0.005309926345944405, + 0.004633716773241758, + 0.005171885248273611, + 0.011125022545456886, + 0.015934152528643608, + 0.005410932004451752, + 0.012903256341814995, + 0.006537499371916056, + 0.008619144558906555, + 0.014749754220247269, + 0.0072748661041259766, + 0.011687717400491238, + 0.011385957710444927, + 0.006090208888053894, + 0.010662499815225601, + 0.00585149833932519, + 0.009795138612389565, + 0.01198737695813179, + 0.00422388780862093, + 0.009145382791757584, + 0.006622358690947294, + 0.00379447964951396, + 0.004004586488008499, + 0.005365019664168358, + 0.005952191073447466, + 0.008920086547732353, + 0.0042772660963237286, + 0.006844880059361458, + 0.0037982056383043528, + 0.021195746958255768, + 0.010056355968117714, + 0.02723751589655876, + 0.009610534645617008, + 0.09122323244810104, + 0.013037854805588722, + 0.025006331503391266, + 0.038411859422922134, + 0.003200087696313858, + 0.0023006945848464966, + 0.004526203032582998, + 0.00383702851831913, + 0.015861550346016884, + 0.014998970553278923, + 0.0035972674377262592, + 0.004690490197390318, + 0.014996085315942764, + 0.011623052880167961, + 0.028791477903723717, + 0.019021103158593178, + 0.07619801163673401, + 0.013392200693488121, + 0.030668148770928383, + 0.033574409782886505, + 0.011915909126400948, + 0.006130266468971968, + 0.0376146100461483, + 0.005678999703377485, + 0.04299342259764671, + 0.010435926727950573, + 0.008724303916096687, + 0.008622368797659874, + 0.007838774472475052, + 0.00608566589653492, + 0.022422805428504944, + 0.004033336415886879, + 0.006538668647408485, + 0.015580892562866211, + 0.003663316834717989, + 0.006075821351259947, + 0.0053182062692940235, + 0.004242454655468464, + 0.003957007080316544, + 0.00414396682754159, + 0.003129031276330352, + 0.0026102124247699976, + 0.0033820595126599073, + 0.0031809331849217415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.001739775761961937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100012592971325 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.004100012592971325, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100012592971325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032031916081905365, + 0.040486179292201996, + 0.11145496368408203, + 0.021820567548274994, + 0.2572287321090698, + 0.22164082527160645, + 0.10025312006473541, + 0.17278148233890533, + 0.020413372665643692, + 0.02188878506422043 + ], + "credit_map": [ + 0.004863115027546883, + 0.003056479385122657, + 0.003352202707901597, + 0.0032700549345463514, + 0.003054489614441991, + 0.004384675528854132, + 0.005216874182224274, + 0.004834020975977182, + 0.002720954827964306, + 0.0065714335069060326, + 0.0027965407352894545, + 0.005179440136998892, + 0.008293962106108665, + 0.0035099824890494347, + 0.006765798665583134, + 0.004648062866181135, + 0.012488113716244698, + 0.015449121594429016, + 0.011099737137556076, + 0.016145896166563034, + 0.02229957841336727, + 0.006073396652936935, + 0.01864120364189148, + 0.009257908910512924, + 0.0018422226421535015, + 0.002195701701566577, + 0.002833720063790679, + 0.0038644936867058277, + 0.0033169465605169535, + 0.0022852388210594654, + 0.0033647920936346054, + 0.0021174531430006027, + 0.022727645933628082, + 0.010593239217996597, + 0.028552982956171036, + 0.008745308965444565, + 0.09966737776994705, + 0.015498449094593525, + 0.026350969448685646, + 0.04509276896715164, + 0.012195716612040997, + 0.006829690653830767, + 0.016400136053562164, + 0.007351655513048172, + 0.07225625962018967, + 0.07731727510690689, + 0.011671803891658783, + 0.01761828176677227, + 0.006304269190877676, + 0.005836502183228731, + 0.013137436471879482, + 0.014261324889957905, + 0.025029463693499565, + 0.006625934969633818, + 0.017973357811570168, + 0.011084827594459057, + 0.014057672582566738, + 0.007160621229559183, + 0.05001365765929222, + 0.007748420350253582, + 0.062001783400774, + 0.01231120340526104, + 0.011971837840974331, + 0.007516295183449984, + 0.0018758532823994756, + 0.0036397255025804043, + 0.003980631474405527, + 0.002421692246571183, + 0.0021626322995871305, + 0.00255522970110178, + 0.0018110659439116716, + 0.0019665430299937725, + 0.004284128081053495, + 0.0021869551856070757, + 0.004065439570695162, + 0.002524252515286207, + 0.0019141569500789046, + 0.0017436095513403416, + 0.0017481613904237747, + 0.003422081470489502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.004100012592971325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253676600754261 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253676600754261, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253676600754261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08971544355154037, + 0.07106170803308487, + 0.18737030029296875, + 0.02752097137272358, + 0.11312390863895416, + 0.12911321222782135, + 0.04979250207543373, + 0.1447598785161972, + 0.14118528366088867, + 0.04635681211948395 + ], + "credit_map": [ + 0.007288276683539152, + 0.006298074033111334, + 0.008257497102022171, + 0.008052443154156208, + 0.0036388603039085865, + 0.010396435856819153, + 0.017678389325737953, + 0.02810545638203621, + 0.004113184753805399, + 0.005031703505665064, + 0.0059354910627007484, + 0.004763557109981775, + 0.0111464848741889, + 0.016840653494000435, + 0.010033599101006985, + 0.013197029009461403, + 0.009476925246417522, + 0.028254494071006775, + 0.013800832442939281, + 0.03372689336538315, + 0.046193383634090424, + 0.008349995128810406, + 0.03542691841721535, + 0.012140849605202675, + 0.0025440051686018705, + 0.001976834610104561, + 0.0038259439170360565, + 0.0034481652546674013, + 0.0051918113604187965, + 0.0028585123363882303, + 0.004582442808896303, + 0.0030932538211345673, + 0.009133230894804, + 0.005059044808149338, + 0.015656305477023125, + 0.0050009386613965034, + 0.04461907222867012, + 0.006350693292915821, + 0.011913025751709938, + 0.015391599386930466, + 0.010454759933054447, + 0.005537942983210087, + 0.01109224557876587, + 0.004693901631981134, + 0.03512610122561455, + 0.043805189430713654, + 0.00648852763697505, + 0.01191454753279686, + 0.004533200059086084, + 0.003774621058255434, + 0.006393770687282085, + 0.004944091197103262, + 0.0094526382163167, + 0.0064172339625656605, + 0.007423350587487221, + 0.006853592582046986, + 0.013104897923767567, + 0.006662028841674328, + 0.0418473556637764, + 0.006198063958436251, + 0.048305124044418335, + 0.010661093518137932, + 0.009637610986828804, + 0.008343701250851154, + 0.013709633611142635, + 0.008227914571762085, + 0.046602215617895126, + 0.007740074303001165, + 0.011751880869269371, + 0.03461100161075592, + 0.006495954003185034, + 0.012046598829329014, + 0.006284733302891254, + 0.005953300278633833, + 0.006775057874619961, + 0.007728999480605125, + 0.004401442129164934, + 0.005816859658807516, + 0.00479257432743907, + 0.004603842739015818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.01253676600754261 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018175480887293816 + ], + "opposite_logprob_tokens": [ + -4.0181756019592285 + ], + "logprob_reward": -0.018175480887293816, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018175480887293816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09678718447685242, + 0.046787016093730927, + 0.07939964532852173, + 0.04331262409687042, + 0.26220133900642395, + 0.03543635830283165, + 0.2558227479457855, + 0.07326383888721466, + 0.047175083309412, + 0.0598142072558403 + ], + "credit_map": [ + 0.010577946901321411, + 0.006346941459923983, + 0.00940505787730217, + 0.006115752272307873, + 0.003976906184107065, + 0.008176694624125957, + 0.02120921202003956, + 0.030978653579950333, + 0.003937169909477234, + 0.007579972967505455, + 0.004813244566321373, + 0.004793359432369471, + 0.007706779986619949, + 0.005002843216061592, + 0.006766797974705696, + 0.006186843384057283, + 0.007274227682501078, + 0.01113803405314684, + 0.006823094096034765, + 0.015104114077985287, + 0.015212388709187508, + 0.005333641078323126, + 0.013236645609140396, + 0.005277490243315697, + 0.004129444248974323, + 0.003402378410100937, + 0.006599988788366318, + 0.003654561936855316, + 0.0058785779401659966, + 0.0063431537710130215, + 0.004825154785066843, + 0.008479360491037369, + 0.02284053899347782, + 0.011065499857068062, + 0.024189503863453865, + 0.008587929420173168, + 0.10242525488138199, + 0.015034954063594341, + 0.028804868459701538, + 0.04925275593996048, + 0.0022695984225720167, + 0.0020147375762462616, + 0.0034437000285834074, + 0.0018812898779287934, + 0.009631498716771603, + 0.011246994137763977, + 0.0019169643055647612, + 0.003031573724001646, + 0.017389848828315735, + 0.013061984442174435, + 0.031007282435894012, + 0.013779236935079098, + 0.09299518167972565, + 0.015925001353025436, + 0.028445200994610786, + 0.04321899637579918, + 0.005982253234833479, + 0.0037633238825947046, + 0.019152147695422173, + 0.0033185516949743032, + 0.0251313429325819, + 0.0062891216948628426, + 0.005268935114145279, + 0.0043581631034612656, + 0.004431509412825108, + 0.004482904449105263, + 0.013589209876954556, + 0.004511683247983456, + 0.004332078155130148, + 0.006299388129264116, + 0.005551245529204607, + 0.003977058455348015, + 0.008178879506886005, + 0.006965912878513336, + 0.009702137671411037, + 0.006782949436455965, + 0.005866696126759052, + 0.0076719545759260654, + 0.006741053890436888, + 0.007904613390564919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.018175480887293816 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up captures a scene of quiet reverence. A slender, pale hand, its movements steady and deliberate, holds a vintage hardcover bound in dark green leather, its surface showing the subtle, yet noble wear of time. Warm, focused light from an unseen source gleams on the ornate, almost illegible gilded lettering and filigree on the book's spine. The hand carefully guides the book toward a narrow opening in a towering, floor-to-ceiling bookshelf crafted from dark polished mahogany. The backdrop is a dense, richly textured mosaic of countless other antique books, their leather-bound spines forming a rich tapestry of crimson, brown, and blue. With meticulous precision, the hand lowers the book into place, the spine sliding smoothly into its neighbor until it lies perfectly flush, adding another volume to the vast, silent collection. The fingers linger for a moment, gently touching the spine, then slowly withdraw, immersing the scene in a soft glow and the subtle scent of old paper, a tranquil, scholarly atmosphere.", + "seed": 1169, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_1458_seed_1169_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01348609805572778, + "vlm_logprob": -0.01348609805572778, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.01348609805572778, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025201745331287384 + ], + "opposite_logprob_tokens": [ + -5.400201797485352 + ], + "logprob_reward": -0.025201745331287384, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.025201745331287384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10103214532136917, + 0.2913479804992676, + 0.10158845037221909, + 0.09373839199542999, + 0.1285816729068756, + 0.04134132340550423, + 0.06844551116228104, + 0.07657433301210403, + 0.039094600826501846, + 0.05825559049844742 + ], + "credit_map": [ + 0.015286749228835106, + 0.006226467899978161, + 0.007740802597254515, + 0.005637556314468384, + 0.010299021378159523, + 0.019129179418087006, + 0.012374293990433216, + 0.024338064715266228, + 0.037361349910497665, + 0.011573146097362041, + 0.012514936737716198, + 0.03447715565562248, + 0.01350042037665844, + 0.05333198979496956, + 0.11016207933425903, + 0.018426910042762756, + 0.011923587881028652, + 0.018262332305312157, + 0.013803894631564617, + 0.016576506197452545, + 0.009783524088561535, + 0.009698454290628433, + 0.01127791777253151, + 0.010262222029268742, + 0.009574384428560734, + 0.011392110027372837, + 0.017498889937996864, + 0.019023174419999123, + 0.00818493403494358, + 0.010973772965371609, + 0.010323185473680496, + 0.006767926272004843, + 0.010802377946674824, + 0.012036968022584915, + 0.0278834979981184, + 0.03497691452503204, + 0.013934859074652195, + 0.009745270945131779, + 0.00894557498395443, + 0.01025619450956583, + 0.005409192759543657, + 0.0038041325751692057, + 0.007767301518470049, + 0.006371671333909035, + 0.0058107334189116955, + 0.004750712309032679, + 0.0027617882005870342, + 0.00466578733175993, + 0.006514071952551603, + 0.005687678698450327, + 0.012656058184802532, + 0.008240999653935432, + 0.010410299524664879, + 0.008175692521035671, + 0.0070113143883645535, + 0.009749389253556728, + 0.01243231724947691, + 0.008145458996295929, + 0.011565417051315308, + 0.009333482012152672, + 0.008159150369465351, + 0.007489365991204977, + 0.011615579947829247, + 0.007833551615476608, + 0.005815995391458273, + 0.0038409950211644173, + 0.007086345925927162, + 0.006857459433376789, + 0.003466040827333927, + 0.0038481007795780897, + 0.0037230481393635273, + 0.004456612281501293, + 0.005602140910923481, + 0.006149339489638805, + 0.009588229469954967, + 0.01668541319668293, + 0.004541076719760895, + 0.006432897876948118, + 0.005232879426330328, + 0.004023610148578882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.025201745331287384 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015328098088502884 + ], + "opposite_logprob_tokens": [ + -6.501533031463623 + ], + "logprob_reward": -0.0015328098088502884, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015328098088502884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20305456221103668, + 0.29290375113487244, + 0.1477321833372116, + 0.10018783062696457, + 0.02474515326321125, + 0.019490370526909828, + 0.07803554832935333, + 0.041307803243398666, + 0.040824953466653824, + 0.05171786621212959 + ], + "credit_map": [ + 0.017438653856515884, + 0.01726648211479187, + 0.0164469163864851, + 0.01522803958505392, + 0.025345714762806892, + 0.02817283570766449, + 0.04762272536754608, + 0.03553316742181778, + 0.03632336109876633, + 0.014306306838989258, + 0.019881248474121094, + 0.033246323466300964, + 0.01610521599650383, + 0.054760344326496124, + 0.09875120222568512, + 0.01952974498271942, + 0.019962361082434654, + 0.03172881156206131, + 0.01913028582930565, + 0.021504387259483337, + 0.00882171094417572, + 0.01506934966892004, + 0.014813481830060482, + 0.016701780259609222, + 0.010181216523051262, + 0.013145128265023232, + 0.01733674854040146, + 0.021456897258758545, + 0.008375708945095539, + 0.011233736760914326, + 0.011686190962791443, + 0.006772194989025593, + 0.002886792179197073, + 0.0033604030031710863, + 0.0037805375177413225, + 0.0035577351227402687, + 0.0025742189027369022, + 0.0032485786359757185, + 0.002071356400847435, + 0.0032655284740030766, + 0.0022096061147749424, + 0.0017839687643572688, + 0.003338203299790621, + 0.002203743439167738, + 0.0017043761909008026, + 0.0023842635564506054, + 0.002116710878908634, + 0.0037494958378374577, + 0.008642325177788734, + 0.008611378259956837, + 0.010464196093380451, + 0.009361770935356617, + 0.015941619873046875, + 0.008172824047505856, + 0.009479849599301815, + 0.0073615796864032745, + 0.006849185097962618, + 0.005524825770407915, + 0.005634047091007233, + 0.006977096665650606, + 0.004040371160954237, + 0.003412992926314473, + 0.004795631393790245, + 0.00407364871352911, + 0.004282908979803324, + 0.003799845464527607, + 0.005162267945706844, + 0.006040267646312714, + 0.004239897709339857, + 0.005124109331518412, + 0.006112445145845413, + 0.006063205190002918, + 0.004974063020199537, + 0.005167542956769466, + 0.006809983868151903, + 0.0059053897857666016, + 0.004843554459512234, + 0.0073056938126683235, + 0.008470525965094566, + 0.008241106756031513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0015328098088502884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019440815085545182 + ], + "opposite_logprob_tokens": [ + -6.251944065093994 + ], + "logprob_reward": -0.0019440815085545182, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019440815085545182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1260245442390442, + 0.23392875492572784, + 0.1865944266319275, + 0.091975636780262, + 0.0641283243894577, + 0.03576456010341644, + 0.09255468845367432, + 0.03387053310871124, + 0.07509475946426392, + 0.060063865035772324 + ], + "credit_map": [ + 0.013802562840282917, + 0.009462135843932629, + 0.009375791065394878, + 0.008394308388233185, + 0.01673034392297268, + 0.020047470927238464, + 0.021059885621070862, + 0.027152007445693016, + 0.0219383854418993, + 0.013253840617835522, + 0.01843469962477684, + 0.0466005802154541, + 0.0122239263728261, + 0.04045497998595238, + 0.06267690658569336, + 0.01834539696574211, + 0.024153439328074455, + 0.03790998086333275, + 0.023615539073944092, + 0.03459962457418442, + 0.008180640637874603, + 0.02288651466369629, + 0.011688393540680408, + 0.02356026880443096, + 0.010186866857111454, + 0.011596212163567543, + 0.017543425783514977, + 0.017689501866698265, + 0.007541092112660408, + 0.00998876802623272, + 0.01111734751611948, + 0.006312401499599218, + 0.005804851185530424, + 0.004216313362121582, + 0.015694890171289444, + 0.013190194964408875, + 0.007217509672045708, + 0.006945447996258736, + 0.003703340422362089, + 0.0073557645082473755, + 0.004078212194144726, + 0.004754798021167517, + 0.006459852214902639, + 0.004151118453592062, + 0.0039026320446282625, + 0.0043777357786893845, + 0.0030971551313996315, + 0.004943051841109991, + 0.010349814780056477, + 0.009424611926078796, + 0.011924432590603828, + 0.010164393112063408, + 0.01956595480442047, + 0.00982279796153307, + 0.010742506943643093, + 0.010560157708823681, + 0.0037867145147174597, + 0.004225927870720625, + 0.004797407425940037, + 0.00536609161645174, + 0.0034559292253106833, + 0.003818738041445613, + 0.004164782352745533, + 0.004254936706274748, + 0.006373126059770584, + 0.0059967245906591415, + 0.009382134303450584, + 0.008385163731873035, + 0.0084255775436759, + 0.010291066020727158, + 0.013535374775528908, + 0.012705574743449688, + 0.004960409831255674, + 0.004763649310916662, + 0.007605687249451876, + 0.007123260758817196, + 0.005873057059943676, + 0.008938835002481937, + 0.010094787925481796, + 0.010704166255891323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0019440815085545182 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184494227170944 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184494227170944, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184494227170944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07727218419313431, + 0.2830766439437866, + 0.05643728747963905, + 0.07869630306959152, + 0.08409589529037476, + 0.03531891107559204, + 0.06479538232088089, + 0.1763552725315094, + 0.08209934830665588, + 0.06185271590948105 + ], + "credit_map": [ + 0.00876578874886036, + 0.005570854060351849, + 0.006512487772852182, + 0.0054003591649234295, + 0.011900846846401691, + 0.01202694233506918, + 0.012520468793809414, + 0.014574432745575905, + 0.027772551402449608, + 0.011935217306017876, + 0.01324347872287035, + 0.044959235936403275, + 0.014288079924881458, + 0.0522594153881073, + 0.10056989639997482, + 0.018048785626888275, + 0.006242954637855291, + 0.011153924278914928, + 0.008544698357582092, + 0.009325715713202953, + 0.004474244080483913, + 0.006156747695058584, + 0.005232944618910551, + 0.005306054372340441, + 0.008845497854053974, + 0.010159943252801895, + 0.014879364520311356, + 0.013774238526821136, + 0.006044627632945776, + 0.008095207624137402, + 0.011147357523441315, + 0.005750066135078669, + 0.009937655180692673, + 0.0150372926145792, + 0.017940398305654526, + 0.014866421930491924, + 0.006099794525653124, + 0.006511256098747253, + 0.006556106731295586, + 0.0071469745598733425, + 0.0035956327337771654, + 0.0028190165758132935, + 0.0077033257111907005, + 0.007097045425325632, + 0.004057773854583502, + 0.0029046316631138325, + 0.0038644855376333, + 0.0032770007383078337, + 0.007622415665537119, + 0.006638705730438232, + 0.010123740881681442, + 0.00836991798132658, + 0.006978262215852737, + 0.0068290578201413155, + 0.009165303781628609, + 0.009067981503903866, + 0.023227188736200333, + 0.026268245652318, + 0.02383604273200035, + 0.03485652059316635, + 0.016397804021835327, + 0.014760454185307026, + 0.019015060737729073, + 0.017993958666920662, + 0.009655202738940716, + 0.006539445836097002, + 0.012964148074388504, + 0.008883736096322536, + 0.008205060847103596, + 0.010714841075241566, + 0.013538890518248081, + 0.011598021723330021, + 0.005921691190451384, + 0.005680008325725794, + 0.008858911693096161, + 0.009030393324792385, + 0.005122586153447628, + 0.008857826702296734, + 0.007616951130330563, + 0.0107643511146307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.018184494227170944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020567359402775764 + ], + "opposite_logprob_tokens": [ + -3.8955674171447754 + ], + "logprob_reward": -0.020567359402775764, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020567359402775764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07484366744756699, + 0.16348227858543396, + 0.059495165944099426, + 0.12913106381893158, + 0.06000367924571037, + 0.04929181933403015, + 0.10993023961782455, + 0.05556919053196907, + 0.11697755753993988, + 0.18127527832984924 + ], + "credit_map": [ + 0.007845725864171982, + 0.0047034574672579765, + 0.005243239924311638, + 0.004753366112709045, + 0.007686840370297432, + 0.014651562087237835, + 0.011947620660066605, + 0.018011854961514473, + 0.019834956154227257, + 0.006459798198193312, + 0.00807288009673357, + 0.021088209003210068, + 0.007854709401726723, + 0.03076658956706524, + 0.059348996728658676, + 0.010056141763925552, + 0.00793510489165783, + 0.012025113217532635, + 0.007678827736526728, + 0.00904055405408144, + 0.004196682013571262, + 0.006440992001444101, + 0.006254452280700207, + 0.005923441145569086, + 0.013634649105370045, + 0.015985431149601936, + 0.027059728279709816, + 0.02133607305586338, + 0.010625891387462616, + 0.014438082464039326, + 0.016180548816919327, + 0.009870666079223156, + 0.006574009079486132, + 0.0063916780054569244, + 0.013644176535308361, + 0.011523650027811527, + 0.005188358016312122, + 0.005790715105831623, + 0.00408190255984664, + 0.006809195037931204, + 0.00596278440207243, + 0.005943220108747482, + 0.008254596032202244, + 0.00638650031760335, + 0.006481450516730547, + 0.0056313276290893555, + 0.004550499841570854, + 0.006081445142626762, + 0.012514960952103138, + 0.007787678390741348, + 0.014648772776126862, + 0.016149472445249557, + 0.015505149029195309, + 0.011914722621440887, + 0.014822320081293583, + 0.01658717356622219, + 0.007815137505531311, + 0.006406249012798071, + 0.008425123058259487, + 0.00549688283354044, + 0.006276687607169151, + 0.006387512199580669, + 0.007469080854207277, + 0.007292520720511675, + 0.01055152714252472, + 0.009168773889541626, + 0.017022233456373215, + 0.017385587096214294, + 0.011832916177809238, + 0.014154430478811264, + 0.020781826227903366, + 0.01608026959002018, + 0.017541486769914627, + 0.013859792612493038, + 0.02330349013209343, + 0.020039569586515427, + 0.018346013501286507, + 0.026700256392359734, + 0.03250013291835785, + 0.028984548524022102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.020567359402775764 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up captures a scene of quiet reverence. A slender, pale hand, its movements steady and deliberate, holds a vintage hardcover bound in dark green leather, its surface showing the subtle, yet noble wear of time. Warm, focused light from an unseen source gleams on the ornate, almost illegible gilded lettering and filigree on the book's spine. The hand carefully guides the book toward a narrow opening in a towering, floor-to-ceiling bookshelf crafted from dark polished mahogany. The backdrop is a dense, richly textured mosaic of countless other antique books, their leather-bound spines forming a rich tapestry of crimson, brown, and blue. With meticulous precision, the hand lowers the book into place, the spine sliding smoothly into its neighbor until it lies perfectly flush, adding another volume to the vast, silent collection. The fingers linger for a moment, gently touching the spine, then slowly withdraw, immersing the scene in a soft glow and the subtle scent of old paper, a tranquil, scholarly atmosphere.", + "seed": 1170, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_7/index_1458_seed_1170_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013105953496415168, + "vlm_logprob": -0.013105953496415168, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.013105953496415168, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03188342973589897 + ], + "opposite_logprob_tokens": [ + -4.781883239746094 + ], + "logprob_reward": -0.03188342973589897, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.03188342973589897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10852950811386108, + 0.09436420351266861, + 0.02148415707051754, + 0.07377889752388, + 0.058722998946905136, + 0.08510187268257141, + 0.27772608399391174, + 0.05928785726428032, + 0.1504339873790741, + 0.0705704614520073 + ], + "credit_map": [ + 0.013395797461271286, + 0.018607890233397484, + 0.012247861362993717, + 0.009578977711498737, + 0.01565425656735897, + 0.014478661119937897, + 0.013357442803680897, + 0.011208624579012394, + 0.01023140549659729, + 0.007920317351818085, + 0.012114856392145157, + 0.004148379899561405, + 0.024685049429535866, + 0.014019228518009186, + 0.009336313232779503, + 0.011908656917512417, + 0.003434630110859871, + 0.001961342990398407, + 0.0027186593506485224, + 0.0018014025408774614, + 0.002770585473626852, + 0.002948039211332798, + 0.0024353666231036186, + 0.0034141300711780787, + 0.004606278613209724, + 0.00653927493840456, + 0.014595014974474907, + 0.006835069041699171, + 0.009331283159554005, + 0.009425770491361618, + 0.007172708865255117, + 0.015273495577275753, + 0.004758367780596018, + 0.009971041232347488, + 0.009655725210905075, + 0.0031922669149935246, + 0.007013946771621704, + 0.011624271050095558, + 0.0034523182548582554, + 0.009055061265826225, + 0.021875375881791115, + 0.006139957811683416, + 0.02089511603116989, + 0.004478523973375559, + 0.006514287553727627, + 0.012559309601783752, + 0.005870820954442024, + 0.006768481805920601, + 0.015721390023827553, + 0.011369886808097363, + 0.027390791103243828, + 0.008844130672514439, + 0.016037248075008392, + 0.06308243423700333, + 0.020253291353583336, + 0.11502691358327866, + 0.004697589669376612, + 0.004353134427219629, + 0.017034387215971947, + 0.0023400175850838423, + 0.010462143458425999, + 0.006183583289384842, + 0.003813345916569233, + 0.010403653606772423, + 0.007255575619637966, + 0.008389487862586975, + 0.00753540825098753, + 0.005037585273385048, + 0.004214464221149683, + 0.056895796209573746, + 0.004860287066549063, + 0.056245386600494385, + 0.0045763710513710976, + 0.006604765076190233, + 0.007962768897414207, + 0.0060667609795928, + 0.005304206162691116, + 0.01792379468679428, + 0.004836468491703272, + 0.017295323312282562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.03188342973589897 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009467886411584914 + ], + "opposite_logprob_tokens": [ + -7.000946998596191 + ], + "logprob_reward": -0.0009467886411584914, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009467886411584914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06158483028411865, + 0.07307710498571396, + 0.02335446886718273, + 0.24819254875183105, + 0.04530587047338486, + 0.12596920132637024, + 0.24905681610107422, + 0.10184004157781601, + 0.03550393506884575, + 0.036115143448114395 + ], + "credit_map": [ + 0.010211844928562641, + 0.009073397144675255, + 0.010004149749875069, + 0.005307255778461695, + 0.006138852797448635, + 0.007016118615865707, + 0.00927986204624176, + 0.004553347360342741, + 0.008667386136949062, + 0.007406368386000395, + 0.010576575994491577, + 0.004045465029776096, + 0.01604650542140007, + 0.009608074091374874, + 0.007162187248468399, + 0.009564544074237347, + 0.0025538508780300617, + 0.00156830542255193, + 0.0042395745404064655, + 0.0021865693852305412, + 0.002572031691670418, + 0.0034397097770124674, + 0.002581813372671604, + 0.004212615080177784, + 0.013000798411667347, + 0.014144297689199448, + 0.06160851567983627, + 0.016011422500014305, + 0.022072341293096542, + 0.0354851596057415, + 0.01297017466276884, + 0.07289982587099075, + 0.003517994424328208, + 0.006297317333519459, + 0.0054462323896586895, + 0.0014995241072028875, + 0.006699542049318552, + 0.011665699072182178, + 0.002820729510858655, + 0.007358829956501722, + 0.022205468267202377, + 0.003857880597934127, + 0.04581905156373978, + 0.004726434126496315, + 0.008356292732059956, + 0.006976715289056301, + 0.004094945266842842, + 0.029932420700788498, + 0.013587875291705132, + 0.01005235593765974, + 0.024708691984415054, + 0.008042674511671066, + 0.014470883645117283, + 0.056371551007032394, + 0.018310319632291794, + 0.1035124734044075, + 0.006755591835826635, + 0.008532477542757988, + 0.008077542297542095, + 0.004214409273117781, + 0.028565123677253723, + 0.011260360479354858, + 0.006154212635010481, + 0.02828032337129116, + 0.0024938031565397978, + 0.0029103776905685663, + 0.0019417096627876163, + 0.001842489349655807, + 0.0012871433282271028, + 0.011837477795779705, + 0.0018999899039044976, + 0.01129094511270523, + 0.004258070141077042, + 0.00415618484839797, + 0.005271198693662882, + 0.0033531414810568094, + 0.0036993050016462803, + 0.006727141328155994, + 0.0032206890173256397, + 0.005429412703961134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.0009467886411584914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410274276509881 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.00410274276509881, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410274276509881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11270471662282944, + 0.08886506408452988, + 0.025229686871170998, + 0.1461118459701538, + 0.05518980324268341, + 0.13289137184619904, + 0.19571028649806976, + 0.18422216176986694, + 0.027933409437537193, + 0.031141651794314384 + ], + "credit_map": [ + 0.01271600928157568, + 0.015186739154160023, + 0.01386286225169897, + 0.007244995795190334, + 0.020310865715146065, + 0.01622246950864792, + 0.015752699226140976, + 0.011408074758946896, + 0.01001150906085968, + 0.008683186955749989, + 0.01506556011736393, + 0.005287930369377136, + 0.017662785947322845, + 0.011779461987316608, + 0.00800714548677206, + 0.012367486953735352, + 0.0030556851997971535, + 0.0018251603469252586, + 0.0036735718604177237, + 0.002299970481544733, + 0.0022084384690970182, + 0.004049993120133877, + 0.002431327011436224, + 0.00568554038181901, + 0.007518092170357704, + 0.007953392341732979, + 0.03403463587164879, + 0.009537897072732449, + 0.014707973226904869, + 0.021409697830677032, + 0.009512163698673248, + 0.04143797978758812, + 0.005065041594207287, + 0.007033712230622768, + 0.014850616455078125, + 0.003032820299267769, + 0.0060315439477562904, + 0.005413513630628586, + 0.0021675950847566128, + 0.011594957672059536, + 0.01015026867389679, + 0.005630943458527327, + 0.04153703525662422, + 0.005538088735193014, + 0.010433553718030453, + 0.012104117311537266, + 0.006479551550000906, + 0.0410178117454052, + 0.010696737095713615, + 0.008351524360477924, + 0.018965713679790497, + 0.006334691308438778, + 0.01161634549498558, + 0.04508667811751366, + 0.01423999760299921, + 0.08041860163211823, + 0.01505624782294035, + 0.01223782915621996, + 0.04579377919435501, + 0.008427309803664684, + 0.009778446517884731, + 0.024157408624887466, + 0.009933561086654663, + 0.058837588876485825, + 0.0015674304449930787, + 0.001968804746866226, + 0.0018169094109907746, + 0.0014807499246671796, + 0.0012453383533284068, + 0.008865004405379295, + 0.0013333966489881277, + 0.009655778296291828, + 0.00381325907073915, + 0.0033198499586433172, + 0.0041330233216285706, + 0.002463066950440407, + 0.003929339814931154, + 0.0057146502658724785, + 0.003315169829875231, + 0.004453291185200214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.00410274276509881 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253711897879839 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253711897879839, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253711897879839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13625656068325043, + 0.04973936453461647, + 0.01885387673974037, + 0.032718248665332794, + 0.02502010017633438, + 0.19293352961540222, + 0.11560051143169403, + 0.314574658870697, + 0.09148867428302765, + 0.022814538329839706 + ], + "credit_map": [ + 0.016900328919291496, + 0.018858784809708595, + 0.01516125351190567, + 0.007781490217894316, + 0.023315444588661194, + 0.020755691453814507, + 0.020082727074623108, + 0.013400834985077381, + 0.005173433572053909, + 0.004324367735534906, + 0.0075784712098538876, + 0.0032158021349459887, + 0.010917691513895988, + 0.006495651789009571, + 0.004633334930986166, + 0.00740060955286026, + 0.002503197407349944, + 0.000858485174831003, + 0.002569779986515641, + 0.0009745429269969463, + 0.0011792551958933473, + 0.002989451866596937, + 0.0024615854490548372, + 0.005317577626556158, + 0.0021265882533043623, + 0.0016498398035764694, + 0.008236279711127281, + 0.0018269340507686138, + 0.002218531910330057, + 0.00445798272266984, + 0.0024449308402836323, + 0.009757159277796745, + 0.0022289915941655636, + 0.0033234695438295603, + 0.0049390788190066814, + 0.0017938511446118355, + 0.002800622023642063, + 0.0037817060947418213, + 0.0018680954817682505, + 0.004284284543246031, + 0.0404314287006855, + 0.007298353128135204, + 0.0719924122095108, + 0.007505768910050392, + 0.013964798301458359, + 0.009048079140484333, + 0.005387321580201387, + 0.037305381149053574, + 0.006745968014001846, + 0.006017231848090887, + 0.011148042976856232, + 0.0052678557112813, + 0.007154862862080336, + 0.023790987208485603, + 0.00875784456729889, + 0.04671771079301834, + 0.028602736070752144, + 0.013802568428218365, + 0.10050314664840698, + 0.011983080767095089, + 0.039710626006126404, + 0.02976967766880989, + 0.019182555377483368, + 0.07102025300264359, + 0.0045747351832687855, + 0.005806989502161741, + 0.0068616801872849464, + 0.0035469054710119963, + 0.002935175783932209, + 0.03168802708387375, + 0.0031762055587023497, + 0.032898955047130585, + 0.00211144145578146, + 0.0024263954255729914, + 0.0021646281238645315, + 0.0014015526976436377, + 0.0019409856759011745, + 0.005317159928381443, + 0.001675816485658288, + 0.005776556674391031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.01253711897879839 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly low-angle close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a slender, pale hand holding a vintage hardcover book bound in dark green leather visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the book spine feature ornate, gilded lettering and filigree illuminated by warm, focused light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hand guiding the book into a narrow opening in a floor-to-ceiling bookshelf made of dark polished mahogany?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dense mosaic of antique books with spines in crimson, brown, and blue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059687361121178 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.016059687361121178, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059687361121178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11530237644910812, + 0.03525669872760773, + 0.018652474507689476, + 0.03051629476249218, + 0.053543999791145325, + 0.15353234112262726, + 0.2719266712665558, + 0.25623419880867004, + 0.014847598038613796, + 0.05018734559416771 + ], + "credit_map": [ + 0.01429822202771902, + 0.013119084760546684, + 0.014852783642709255, + 0.0070712873712182045, + 0.02037222497165203, + 0.018711650744080544, + 0.015025230124592781, + 0.01185188814997673, + 0.004191996064037085, + 0.003063588635995984, + 0.006683193612843752, + 0.0019205338321626186, + 0.007087213918566704, + 0.0039857071824371815, + 0.0033539091236889362, + 0.004970554728060961, + 0.00229071662761271, + 0.001503518782556057, + 0.00271235266700387, + 0.001340779010206461, + 0.002184817101806402, + 0.002412403468042612, + 0.0025018316227942705, + 0.0037060549948364496, + 0.002130689565092325, + 0.00505743408575654, + 0.004904803354293108, + 0.0021574890706688166, + 0.0037297778762876987, + 0.005123881623148918, + 0.002493153791874647, + 0.004919066093862057, + 0.004059955477714539, + 0.008867013268172741, + 0.01197674497961998, + 0.0029910781886428595, + 0.00459607969969511, + 0.008686112239956856, + 0.0027054683305323124, + 0.009661544114351273, + 0.030953414738178253, + 0.004141743760555983, + 0.05799911543726921, + 0.005729980766773224, + 0.011821605265140533, + 0.007133515086025, + 0.0041270069777965546, + 0.03162596747279167, + 0.014909191057085991, + 0.012980429455637932, + 0.025966055691242218, + 0.009527976624667645, + 0.016115443781018257, + 0.06052222102880478, + 0.019887933507561684, + 0.112017422914505, + 0.019294919446110725, + 0.01872030273079872, + 0.060173988342285156, + 0.009931782260537148, + 0.08351024985313416, + 0.01866474375128746, + 0.017827356234192848, + 0.02811085246503353, + 0.0013365476625040174, + 0.001478207646869123, + 0.0017723151249811053, + 0.0011183415772393346, + 0.0010164231061935425, + 0.002979844342917204, + 0.0015172784915193915, + 0.0036286392714828253, + 0.005081694107502699, + 0.00738138984888792, + 0.006477981340140104, + 0.0030961555894464254, + 0.005347386468201876, + 0.007781824097037315, + 0.004655151627957821, + 0.010365760885179043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.016059687361121178 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_8/index_tensor([9253, 9253])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_8/index_tensor([9253, 9253])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..48ff98c7231a398289a857c904b9bb49ebff7563 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_8/index_tensor([9253, 9253])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, beginning from a slight overhead angle, slowly pans forward, ultimately transforming into an extreme macro, captures the moment a sharp, angular white sugar cube is carefully dropped into a mirror-smooth cup of hot coffee, nestled within a dark ceramic mug. Soft, warm indoor light streams down from above and sideways, creating a bright highlight on the rim of the cup, while the coffee itself, a deep, almost black color, emanates delicate wisps of steam. The sugar cube gently disrupts the calm surface of the liquid, creating a tiny ripple before sinking. As the camera approaches, countless tiny bubbles can be clearly seen surging from the pores of the sugar cube, its hard edges quickly rounding and disintegrating. The dissolved sugar transforms into a milky, viscous, warm current that swirls and diffuses within the dark coffee liquid, forming a dynamic and graceful vortex. The warm tones of the entire scene create an atmosphere of tranquility, healing, and ritualistic concentration.", + "seed": 31785, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_8/index_9253_seed_31785_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04790337160229683, + "vlm_logprob": -0.04790337160229683, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.04790337160229683, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018738040700554848 + ], + "opposite_logprob_tokens": [ + -4.893738269805908 + ], + "logprob_reward": -0.018738040700554848, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.018738040700554848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03592179715633392, + 0.06267741322517395, + 0.03292526304721832, + 0.4177595376968384, + 0.017682000994682312, + 0.03213704749941826, + 0.024996638298034668, + 0.03425118327140808, + 0.2852564752101898, + 0.05639273673295975 + ], + "credit_map": [ + 0.004036521073430777, + 0.0029848446138203144, + 0.010735626332461834, + 0.002517521847039461, + 0.003125261515378952, + 0.0030630058608949184, + 0.002337884157896042, + 0.0071211280301213264, + 0.006626591552048922, + 0.009223168715834618, + 0.007347648032009602, + 0.0068889702670276165, + 0.009962114505469799, + 0.006036482751369476, + 0.011313453316688538, + 0.0052789803594350815, + 0.0033047697506844997, + 0.004865047987550497, + 0.0032637258991599083, + 0.0022705665323883295, + 0.006109967827796936, + 0.00519917905330658, + 0.003364271717146039, + 0.004547731950879097, + 0.029475504532456398, + 0.011305131018161774, + 0.1787516474723816, + 0.01977115124464035, + 0.010617557913064957, + 0.02545679733157158, + 0.014477629214525223, + 0.12790407240390778, + 0.00263908039778471, + 0.001715976744890213, + 0.0031827050261199474, + 0.0015777407679706812, + 0.0013709486229345202, + 0.0020381384529173374, + 0.0018805606523528695, + 0.0032768500968813896, + 0.0029619000852108, + 0.0027881506830453873, + 0.003915559966117144, + 0.002141903154551983, + 0.002908910857513547, + 0.007035107351839542, + 0.0024106146302074194, + 0.007974897511303425, + 0.001950336154550314, + 0.004650887567549944, + 0.003281776560470462, + 0.0029699639417231083, + 0.004978009033948183, + 0.003419262822717428, + 0.001986169721931219, + 0.0017602306324988604, + 0.0022293319925665855, + 0.007140778936445713, + 0.002312003169208765, + 0.004722777288407087, + 0.00685547012835741, + 0.005970687605440617, + 0.0026399758644402027, + 0.002380155259743333, + 0.012880872935056686, + 0.01506908517330885, + 0.011323453858494759, + 0.024967186152935028, + 0.0141121381893754, + 0.11020679771900177, + 0.015655288472771645, + 0.0810416117310524, + 0.005466242786496878, + 0.007677287794649601, + 0.010810882784426212, + 0.006148996762931347, + 0.005630496423691511, + 0.006644692737609148, + 0.007065657526254654, + 0.006948478985577822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.018738040700554848 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048669882118701935 + ], + "opposite_logprob_tokens": [ + -3.0486698150634766 + ], + "logprob_reward": -0.048669882118701935, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048669882118701935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07603384554386139, + 0.061800044029951096, + 0.022802777588367462, + 0.050289086997509, + 0.015517204999923706, + 0.08256734907627106, + 0.019699586555361748, + 0.035352230072021484, + 0.23537559807300568, + 0.4005622863769531 + ], + "credit_map": [ + 0.009831066243350506, + 0.006148910615593195, + 0.015955699607729912, + 0.007516807410866022, + 0.0072059244848787785, + 0.007192390505224466, + 0.006647671572864056, + 0.015535370446741581, + 0.006244870368391275, + 0.01067107543349266, + 0.00519819650799036, + 0.0052913278341293335, + 0.015817785635590553, + 0.005239225924015045, + 0.0066839237697422504, + 0.0066536348313093185, + 0.002459777519106865, + 0.0030409034807235003, + 0.0029279105365276337, + 0.0019354663090780377, + 0.00265864166431129, + 0.003938029985874891, + 0.0020827397238463163, + 0.003759305225685239, + 0.00649223243817687, + 0.005880400072783232, + 0.006666796747595072, + 0.00591266481205821, + 0.007790053728967905, + 0.005754780489951372, + 0.005531334318220615, + 0.006260821595788002, + 0.0021677471231669188, + 0.0016937883337959647, + 0.001882535987533629, + 0.001647262368351221, + 0.0024239523336291313, + 0.0021303794346749783, + 0.0014508619206026196, + 0.0021206766832619905, + 0.008045580238103867, + 0.008707862347364426, + 0.00992095097899437, + 0.0062951743602752686, + 0.009015848860144615, + 0.01480791438370943, + 0.006761603988707066, + 0.019012412056326866, + 0.0017682306934148073, + 0.0036368013825267553, + 0.0030418310780078173, + 0.001767558977007866, + 0.003487073816359043, + 0.0027731922455132008, + 0.001950897858478129, + 0.0012739994563162327, + 0.002328592585399747, + 0.004806705750524998, + 0.0042599644511938095, + 0.003541664918884635, + 0.006798911839723587, + 0.007162080612033606, + 0.003176296828314662, + 0.0032780119217932224, + 0.010499668307602406, + 0.01078851893544197, + 0.012152634561061859, + 0.023253869265317917, + 0.009309572167694569, + 0.08921675384044647, + 0.014491789974272251, + 0.06566279381513596, + 0.03859572112560272, + 0.053375739604234695, + 0.0751626193523407, + 0.05286049097776413, + 0.04373078793287277, + 0.04364187270402908, + 0.056455545127391815, + 0.03673950582742691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.048669882118701935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296737939119339 + ], + "opposite_logprob_tokens": [ + -2.2379674911499023 + ], + "logprob_reward": -0.11296737939119339, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296737939119339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12658695876598358, + 0.04793098196387291, + 0.10119344294071198, + 0.0857551321387291, + 0.07127903401851654, + 0.12500974535942078, + 0.0652799978852272, + 0.11474716663360596, + 0.060341306030750275, + 0.2018762230873108 + ], + "credit_map": [ + 0.011201097629964352, + 0.010544665157794952, + 0.02298448234796524, + 0.013074117712676525, + 0.011142798699438572, + 0.013524966314435005, + 0.012394970282912254, + 0.031719863414764404, + 0.0059133474715054035, + 0.00569618446752429, + 0.005207186099141836, + 0.0049564954824745655, + 0.009374643675982952, + 0.004107113461941481, + 0.007171198725700378, + 0.005504812113940716, + 0.008929746225476265, + 0.014952038414776325, + 0.012289903126657009, + 0.007987807504832745, + 0.018410230055451393, + 0.015071914531290531, + 0.010563146322965622, + 0.012988652102649212, + 0.008122197352349758, + 0.008097724989056587, + 0.018176322802901268, + 0.006558305583894253, + 0.013000906445086002, + 0.008600085973739624, + 0.007698044180870056, + 0.015501538291573524, + 0.008142143487930298, + 0.007548817433416843, + 0.010242143645882607, + 0.007988662458956242, + 0.009346739389002323, + 0.009042241610586643, + 0.009065329097211361, + 0.009902957826852798, + 0.011410505510866642, + 0.010760573670268059, + 0.013535733334720135, + 0.00788869708776474, + 0.012281947769224644, + 0.028860608115792274, + 0.007208261638879776, + 0.033063411712646484, + 0.005811234004795551, + 0.011357439681887627, + 0.007076595444232225, + 0.006791511084884405, + 0.01273296494036913, + 0.009014721028506756, + 0.006918442901223898, + 0.005577094852924347, + 0.00840118620544672, + 0.0195985846221447, + 0.011904824525117874, + 0.024769969284534454, + 0.016281090676784515, + 0.01561732031404972, + 0.010341803543269634, + 0.007832381874322891, + 0.003339836373925209, + 0.005371692590415478, + 0.007038837298750877, + 0.008234902285039425, + 0.00988497119396925, + 0.01020647119730711, + 0.008694279938936234, + 0.0075703151524066925, + 0.017529314383864403, + 0.02459125407040119, + 0.036606352776288986, + 0.031191756948828697, + 0.02180691435933113, + 0.02114277333021164, + 0.02956961654126644, + 0.019438236951828003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.11296737939119339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306579753756523 + ], + "opposite_logprob_tokens": [ + -3.1680657863616943 + ], + "logprob_reward": -0.04306579753756523, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306579753756523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07082779705524445, + 0.03729374334216118, + 0.08961132913827896, + 0.43133455514907837, + 0.03699064627289772, + 0.032123252749443054, + 0.027959899976849556, + 0.03392975032329559, + 0.13173672556877136, + 0.1081923246383667 + ], + "credit_map": [ + 0.005231792572885752, + 0.0090138865634799, + 0.00565115874633193, + 0.01176314428448677, + 0.008274276740849018, + 0.0073349433951079845, + 0.009960643947124481, + 0.013597950339317322, + 0.00403513154014945, + 0.006348852068185806, + 0.00404934911057353, + 0.004394896794110537, + 0.005707950796931982, + 0.0033728666603565216, + 0.005663792137056589, + 0.003720906563103199, + 0.009637405164539814, + 0.011962451040744781, + 0.010016249492764473, + 0.006095987278968096, + 0.014744735322892666, + 0.016693221405148506, + 0.008465618826448917, + 0.011995655484497547, + 0.0310673788189888, + 0.011942806653678417, + 0.18233469128608704, + 0.021027954295277596, + 0.011691534891724586, + 0.026612447574734688, + 0.01514058094471693, + 0.13151715695858002, + 0.004812897648662329, + 0.004385341424494982, + 0.004286196082830429, + 0.0037930773105472326, + 0.004225506447255611, + 0.007383841555565596, + 0.003360228380188346, + 0.004743556957691908, + 0.002881593070924282, + 0.003273790003731847, + 0.0037460767198354006, + 0.0024408360477536917, + 0.0036379473749548197, + 0.006690839305520058, + 0.0022653057239949703, + 0.007186862640082836, + 0.0023873934987932444, + 0.005233705975115299, + 0.0030863587744534016, + 0.0035855588503181934, + 0.004893073346465826, + 0.0045454734936356544, + 0.0024358397349715233, + 0.0017924952553585172, + 0.0024413547944277525, + 0.007317036390304565, + 0.003036250825971365, + 0.004791928455233574, + 0.006356791127473116, + 0.005175773054361343, + 0.0025728268083184958, + 0.002237789100036025, + 0.00649383245036006, + 0.007013918831944466, + 0.007429130375385284, + 0.009044463746249676, + 0.010709677822887897, + 0.04670868441462517, + 0.008596814237535, + 0.0357402078807354, + 0.009007559157907963, + 0.012515051290392876, + 0.01605280116200447, + 0.013833396136760712, + 0.010800465941429138, + 0.01679479144513607, + 0.018197359517216682, + 0.010990897193551064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.04306579753756523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075758263468742 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.016075758263468742, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075758263468742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09800205379724503, + 0.16429905593395233, + 0.08495213836431503, + 0.26259663701057434, + 0.036489758640527725, + 0.04113951325416565, + 0.03972148522734642, + 0.04298427328467369, + 0.1254121959209442, + 0.10440286993980408 + ], + "credit_map": [ + 0.008431466296315193, + 0.010087721981108189, + 0.008938325569033623, + 0.01668475568294525, + 0.008566750213503838, + 0.009461709298193455, + 0.015369502827525139, + 0.02046181447803974, + 0.019394736737012863, + 0.027117008343338966, + 0.01710905134677887, + 0.019044961780309677, + 0.027403291314840317, + 0.014483155682682991, + 0.026546746492385864, + 0.01320009957998991, + 0.00953871849924326, + 0.009790375828742981, + 0.014359910041093826, + 0.00742837181314826, + 0.011199447326362133, + 0.010365456342697144, + 0.006913685239851475, + 0.015356178395450115, + 0.018507860600948334, + 0.00741097005084157, + 0.111300989985466, + 0.011900480836629868, + 0.006858938839286566, + 0.015357266180217266, + 0.009251857176423073, + 0.08200826495885849, + 0.00568140996620059, + 0.0043493336997926235, + 0.004660043865442276, + 0.0044411467388272285, + 0.003893541870638728, + 0.004713450092822313, + 0.0036772002931684256, + 0.005073629319667816, + 0.004646435845643282, + 0.00598987378180027, + 0.004138318356126547, + 0.0045175375416874886, + 0.006205426063388586, + 0.006353297270834446, + 0.003453023498877883, + 0.005835598800331354, + 0.003982045687735081, + 0.006241183262318373, + 0.005609381478279829, + 0.004045425448566675, + 0.0076659987680613995, + 0.005323286633938551, + 0.0031248568557202816, + 0.003729308722540736, + 0.002869349904358387, + 0.007255641743540764, + 0.0056395214051008224, + 0.004727860912680626, + 0.008814094588160515, + 0.006541512440890074, + 0.003974468447268009, + 0.003161821747198701, + 0.005775326397269964, + 0.014533088542521, + 0.012559747323393822, + 0.016299230977892876, + 0.01320670172572136, + 0.031500980257987976, + 0.012010600417852402, + 0.019526513293385506, + 0.008868713863193989, + 0.014751574024558067, + 0.017751993611454964, + 0.01283240970224142, + 0.012560507282614708, + 0.011550729162991047, + 0.015190964564681053, + 0.010895980522036552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.016075758263468742 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, beginning from a slight overhead angle, slowly pans forward, ultimately transforming into an extreme macro, captures the moment a sharp, angular white sugar cube is carefully dropped into a mirror-smooth cup of hot coffee, nestled within a dark ceramic mug. Soft, warm indoor light streams down from above and sideways, creating a bright highlight on the rim of the cup, while the coffee itself, a deep, almost black color, emanates delicate wisps of steam. The sugar cube gently disrupts the calm surface of the liquid, creating a tiny ripple before sinking. As the camera approaches, countless tiny bubbles can be clearly seen surging from the pores of the sugar cube, its hard edges quickly rounding and disintegrating. The dissolved sugar transforms into a milky, viscous, warm current that swirls and diffuses within the dark coffee liquid, forming a dynamic and graceful vortex. The warm tones of the entire scene create an atmosphere of tranquility, healing, and ritualistic concentration.", + "seed": 31786, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_8/index_9253_seed_31786_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.033150287345051765, + "vlm_logprob": -0.033150287345051765, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.033150287345051765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013617478311061859 + ], + "opposite_logprob_tokens": [ + -6.013617515563965 + ], + "logprob_reward": -0.013617478311061859, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.013617478311061859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5812134146690369, + 0.03888481855392456, + 0.014719867147505283, + 0.2449071705341339, + 0.016187254339456558, + 0.021586569026112556, + 0.018242672085762024, + 0.02702951617538929, + 0.016692819073796272, + 0.020535847172141075 + ], + "credit_map": [ + 0.01790759153664112, + 0.21235434710979462, + 0.019899247214198112, + 0.21395234763622284, + 0.0550854466855526, + 0.005250090267509222, + 0.04276552051305771, + 0.013998819515109062, + 0.006273190025240183, + 0.0032347182277590036, + 0.0020095081999897957, + 0.013699069619178772, + 0.0034050429239869118, + 0.0012114251730963588, + 0.003235672367736697, + 0.005816193763166666, + 0.0017665590858086944, + 0.002099700504913926, + 0.0018036620458588004, + 0.0031101226340979338, + 0.0016325219767168164, + 0.0012197390897199512, + 0.0018097684951499104, + 0.001277792383916676, + 0.09721091389656067, + 0.06862644851207733, + 0.012438983656466007, + 0.009177488274872303, + 0.030053777620196342, + 0.006165388040244579, + 0.01404505129903555, + 0.0071891117841005325, + 0.0036679457407444715, + 0.0019321145955473185, + 0.0032048895955085754, + 0.0014351272257044911, + 0.0023272705730050802, + 0.0009471425437368453, + 0.001829872839152813, + 0.0008428914006799459, + 0.003983999136835337, + 0.0038690976798534393, + 0.00490636145696044, + 0.001808290253393352, + 0.0023797352332621813, + 0.0010025555966421962, + 0.0018481167498975992, + 0.0017884133849292994, + 0.0027841953560709953, + 0.002659342484548688, + 0.002357015386223793, + 0.0025184352416545153, + 0.0021320153027772903, + 0.0012696095509454608, + 0.0033444208092987537, + 0.0011776374885812402, + 0.0025120784994214773, + 0.00294329971075058, + 0.002841810928657651, + 0.0036041683051735163, + 0.005873085465282202, + 0.0024681163486093283, + 0.005154366604983807, + 0.0016325903125107288, + 0.0016831525135785341, + 0.002064330969005823, + 0.0017464072443544865, + 0.0018367882585152984, + 0.0027163829654455185, + 0.0019063708605244756, + 0.0031177960336208344, + 0.0016215903451666236, + 0.002202296629548073, + 0.002929337788373232, + 0.002705973805859685, + 0.0017885537818074226, + 0.0034871522802859545, + 0.0024016008246690035, + 0.002458334667608142, + 0.0025625983253121376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.013617478311061859 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372691571712494 + ], + "opposite_logprob_tokens": [ + -3.408726930618286 + ], + "logprob_reward": -0.03372691571712494, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372691571712494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4594147205352783, + 0.050098128616809845, + 0.024294190108776093, + 0.2681851089000702, + 0.02476741373538971, + 0.052111126482486725, + 0.01371738500893116, + 0.032692763954401016, + 0.04072621837258339, + 0.03399292379617691 + ], + "credit_map": [ + 0.014063248410820961, + 0.1697455495595932, + 0.014926565811038017, + 0.16797514259815216, + 0.04450595751404762, + 0.003902065334841609, + 0.033502157777547836, + 0.010794026777148247, + 0.017372190952301025, + 0.004620683845132589, + 0.0022237650118768215, + 0.0037274665664881468, + 0.00455600768327713, + 0.002000865526497364, + 0.002622791565954685, + 0.012974359095096588, + 0.002742402022704482, + 0.002866775030270219, + 0.002506547374650836, + 0.0066625517792999744, + 0.003005972597748041, + 0.0016450390685349703, + 0.0026286954525858164, + 0.0022362060844898224, + 0.10709456354379654, + 0.07527042180299759, + 0.013113332912325859, + 0.009001025930047035, + 0.03663056716322899, + 0.005881643854081631, + 0.01547771506011486, + 0.005715848412364721, + 0.002686193445697427, + 0.004058916587382555, + 0.004772948566824198, + 0.0037031853571534157, + 0.002871084026992321, + 0.0019045410444959998, + 0.0027080639265477657, + 0.002062481828033924, + 0.013127907179296017, + 0.0037611303851008415, + 0.02040814980864525, + 0.003568114945665002, + 0.004354424308985472, + 0.0019444972276687622, + 0.0035618485417217016, + 0.001385050592944026, + 0.002291173906996846, + 0.0014654089463874698, + 0.0017168456688523293, + 0.00166364258620888, + 0.0018889884231612086, + 0.0011221792083233595, + 0.0027359947562217712, + 0.0008331519202329218, + 0.0034314757212996483, + 0.0033208539243787527, + 0.0032964590936899185, + 0.00391239020973444, + 0.006611734628677368, + 0.0030526146292686462, + 0.006761983036994934, + 0.002305251080542803, + 0.003448021598160267, + 0.005392446648329496, + 0.005661136470735073, + 0.004052365198731422, + 0.007215400226414204, + 0.0037818679120391607, + 0.007293674163520336, + 0.0038813052233308554, + 0.0036087301559746265, + 0.004225378856062889, + 0.004059603903442621, + 0.0036942469887435436, + 0.0056989071890711784, + 0.0032908881548792124, + 0.005693916697055101, + 0.0037212520837783813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.03372691571712494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901466637849808 + ], + "opposite_logprob_tokens": [ + -2.579014778137207 + ], + "logprob_reward": -0.07901466637849808, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901466637849808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25600385665893555, + 0.13710135221481323, + 0.05649112910032272, + 0.2895090878009796, + 0.06491697579622269, + 0.07596949487924576, + 0.012335935607552528, + 0.039911240339279175, + 0.04763789474964142, + 0.02012295462191105 + ], + "credit_map": [ + 0.00979246012866497, + 0.07897146046161652, + 0.010321784764528275, + 0.09258495271205902, + 0.017445188015699387, + 0.015109059400856495, + 0.019966714084148407, + 0.01181222964078188, + 0.04929465800523758, + 0.01141316257417202, + 0.005779692903161049, + 0.0166262648999691, + 0.01224224641919136, + 0.004437567200511694, + 0.006980287842452526, + 0.030327482149004936, + 0.007646243553608656, + 0.0070239948108792305, + 0.006643437780439854, + 0.01435414981096983, + 0.006324468646198511, + 0.004109047818928957, + 0.0062041818164289, + 0.004185604862868786, + 0.10933974385261536, + 0.10699814558029175, + 0.012079990468919277, + 0.009433467872440815, + 0.019467424601316452, + 0.005603159312158823, + 0.018614618107676506, + 0.007972554303705692, + 0.021659787744283676, + 0.005543917417526245, + 0.02091880515217781, + 0.0034989104606211185, + 0.0037679430097341537, + 0.002592931967228651, + 0.0034285751171410084, + 0.0035061107482761145, + 0.01643313840031624, + 0.008746223524212837, + 0.024628551676869392, + 0.006799160968512297, + 0.006509722676128149, + 0.0028834484983235598, + 0.006540520116686821, + 0.003428726689890027, + 0.0015735934721305966, + 0.0024661968927830458, + 0.0010498587507754564, + 0.0017779875779524446, + 0.0017741608899086714, + 0.0011970772175118327, + 0.0017533107893541455, + 0.0007437503081746399, + 0.004185722675174475, + 0.004298670683056116, + 0.0045605371706187725, + 0.005025653168559074, + 0.007492304313927889, + 0.0033433253411203623, + 0.007409112993627787, + 0.0035959132947027683, + 0.004168584942817688, + 0.006381266284734011, + 0.006558610592037439, + 0.005627536214888096, + 0.008921658620238304, + 0.003369611455127597, + 0.009001622907817364, + 0.003609000239521265, + 0.0020567127503454685, + 0.0026556081138551235, + 0.002282608300447464, + 0.002475770190358162, + 0.0031683340203016996, + 0.0021201251074671745, + 0.0033775686752051115, + 0.001986226998269558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.07901466637849808 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607704907655716 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.01607704907655716, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607704907655716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6712303161621094, + 0.018218204379081726, + 0.01088869571685791, + 0.1140204444527626, + 0.01879259943962097, + 0.016872312873601913, + 0.016299061477184296, + 0.026205549016594887, + 0.052129168063402176, + 0.0553436316549778 + ], + "credit_map": [ + 0.02024390734732151, + 0.25461217761039734, + 0.019956445321440697, + 0.2399989813566208, + 0.0685720443725586, + 0.005304009187966585, + 0.046935226768255234, + 0.015607555396854877, + 0.0019046436063945293, + 0.0018169437535107136, + 0.001847529667429626, + 0.005737866275012493, + 0.0017144978046417236, + 0.0010167210130020976, + 0.0018201080383732915, + 0.002359892474487424, + 0.0014009269652888179, + 0.0012917349813506007, + 0.00146157864946872, + 0.0027369980234652758, + 0.001308318693190813, + 0.0009696679771877825, + 0.001068399171344936, + 0.000651070789899677, + 0.03624763339757919, + 0.04554909095168114, + 0.004912093281745911, + 0.004117595963180065, + 0.011484815739095211, + 0.0021220978815108538, + 0.005975437816232443, + 0.0036116845440119505, + 0.004060317762196064, + 0.0020558438263833523, + 0.004007916431874037, + 0.0018732427852228284, + 0.0018357997760176659, + 0.00146116700489074, + 0.0019302131840959191, + 0.0015680985525250435, + 0.003633460495620966, + 0.0017187230987474322, + 0.005201999563723803, + 0.0018449979834258556, + 0.0015471919905394316, + 0.0008472900371998549, + 0.0012248910497874022, + 0.0008537574904039502, + 0.0032790040131658316, + 0.0027212132699787617, + 0.0019113069865852594, + 0.00196745153516531, + 0.001771944691427052, + 0.0005473980563692749, + 0.0031832922250032425, + 0.0009174497681669891, + 0.0030841382686048746, + 0.0034056215081363916, + 0.0032060432713478804, + 0.004021969623863697, + 0.0038758800365030766, + 0.0024602466728538275, + 0.0034901718609035015, + 0.0026614770758897066, + 0.004101071041077375, + 0.007082647178322077, + 0.006765445694327354, + 0.006889666896313429, + 0.010957658290863037, + 0.0043123820796608925, + 0.008567656390368938, + 0.0034526423551142216, + 0.005431577563285828, + 0.008648738265037537, + 0.007910527288913727, + 0.005643285810947418, + 0.007520757615566254, + 0.005610131658613682, + 0.007773826364427805, + 0.006804785691201687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.01607704907655716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023315327242016792 + ], + "opposite_logprob_tokens": [ + -3.7733154296875 + ], + "logprob_reward": -0.023315327242016792, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023315327242016792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6928769946098328, + 0.06774096935987473, + 0.0059363385662436485, + 0.12452804297208786, + 0.010029377415776253, + 0.05006546899676323, + 0.010270590893924236, + 0.01858234964311123, + 0.010312685742974281, + 0.009657146409153938 + ], + "credit_map": [ + 0.020573927089571953, + 0.2553927004337311, + 0.02202341891825199, + 0.25317463278770447, + 0.0656861811876297, + 0.009055271744728088, + 0.05022166669368744, + 0.016749193891882896, + 0.008956308476626873, + 0.008181494660675526, + 0.008039689622819424, + 0.013388083316385746, + 0.011637764051556587, + 0.005581722129136324, + 0.005829973611980677, + 0.006125940475612879, + 0.0009020605939440429, + 0.0007658908725716174, + 0.000620038656052202, + 0.0012115180725231767, + 0.0008614693651907146, + 0.0004544830007944256, + 0.0006018711137585342, + 0.0005190066294744611, + 0.04986577481031418, + 0.03031223453581333, + 0.0062625110149383545, + 0.0034731097985059023, + 0.022136427462100983, + 0.002849344164133072, + 0.007572314701974392, + 0.0020563313737511635, + 0.00218744738958776, + 0.0009650031570345163, + 0.0024860366247594357, + 0.0007437543827109039, + 0.0016676528612151742, + 0.0005295452428981662, + 0.0009316586656495929, + 0.0005182782770134509, + 0.013461736030876637, + 0.003547659143805504, + 0.020829204469919205, + 0.0017363169463351369, + 0.0042473329231143, + 0.001540238270536065, + 0.003291643690317869, + 0.0014113348443061113, + 0.0017706021899357438, + 0.0010984553955495358, + 0.0009681789670139551, + 0.0011692935368046165, + 0.0014540795236825943, + 0.0008762822835706174, + 0.0022625033743679523, + 0.000671195681206882, + 0.0019019984174519777, + 0.0013963659293949604, + 0.0015713406028226018, + 0.0022607294376939535, + 0.003615949535742402, + 0.0018272725865244865, + 0.004301352892071009, + 0.00170734035782516, + 0.0009642061195336282, + 0.0012460632715374231, + 0.0014277625596150756, + 0.0009068853687494993, + 0.0019293061923235655, + 0.0012547068763524294, + 0.0016498318873345852, + 0.0009339225362055004, + 0.0011665361234918237, + 0.0014434801414608955, + 0.0011144945165142417, + 0.0010248655453324318, + 0.0015950251836329699, + 0.0008119747508317232, + 0.0011677434667944908, + 0.0013330267975106835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.023315327242016792 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, beginning from a slight overhead angle, slowly pans forward, ultimately transforming into an extreme macro, captures the moment a sharp, angular white sugar cube is carefully dropped into a mirror-smooth cup of hot coffee, nestled within a dark ceramic mug. Soft, warm indoor light streams down from above and sideways, creating a bright highlight on the rim of the cup, while the coffee itself, a deep, almost black color, emanates delicate wisps of steam. The sugar cube gently disrupts the calm surface of the liquid, creating a tiny ripple before sinking. As the camera approaches, countless tiny bubbles can be clearly seen surging from the pores of the sugar cube, its hard edges quickly rounding and disintegrating. The dissolved sugar transforms into a milky, viscous, warm current that swirls and diffuses within the dark coffee liquid, forming a dynamic and graceful vortex. The warm tones of the entire scene create an atmosphere of tranquility, healing, and ritualistic concentration.", + "seed": 31787, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_8/index_9253_seed_31787_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.039704371243715286, + "vlm_logprob": -0.039704371243715286, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.039704371243715286, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013918010517954826 + ], + "opposite_logprob_tokens": [ + -5.888917922973633 + ], + "logprob_reward": -0.013918010517954826, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.013918010517954826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22626040875911713, + 0.09811405092477798, + 0.02172217145562172, + 0.07209000736474991, + 0.08502247184515, + 0.05541633814573288, + 0.06855417042970657, + 0.09071485698223114, + 0.16812549531459808, + 0.11398006975650787 + ], + "credit_map": [ + 0.035433873534202576, + 0.019151106476783752, + 0.02040606364607811, + 0.014019644819200039, + 0.0829940214753151, + 0.01949354074895382, + 0.018957190215587616, + 0.015804968774318695, + 0.034301698207855225, + 0.024676255881786346, + 0.005624419543892145, + 0.005798524245619774, + 0.007867556996643543, + 0.004924547392874956, + 0.004243152681738138, + 0.010677899233996868, + 0.004406461957842112, + 0.003155086189508438, + 0.002712059998884797, + 0.001953632803633809, + 0.003166943322867155, + 0.002842447254806757, + 0.0014690188691020012, + 0.002016522455960512, + 0.011518951505422592, + 0.0065447501838207245, + 0.007816447876393795, + 0.006452185567468405, + 0.010285815224051476, + 0.00806137453764677, + 0.009944138117134571, + 0.011466345749795437, + 0.010689299553632736, + 0.011176230385899544, + 0.010343074798583984, + 0.0091096805408597, + 0.011584230698645115, + 0.007702036760747433, + 0.01003003679215908, + 0.014387879520654678, + 0.005142880603671074, + 0.006228508893400431, + 0.007937201298773289, + 0.010814226232469082, + 0.007978987880051136, + 0.00618214113637805, + 0.0053560021333396435, + 0.0057763890363276005, + 0.005375799722969532, + 0.010751046240329742, + 0.00980882253497839, + 0.01335990708321333, + 0.009884598664939404, + 0.006129010114818811, + 0.006192435510456562, + 0.0070525468327105045, + 0.009271659888327122, + 0.009746327996253967, + 0.012768772430717945, + 0.012664025649428368, + 0.011991080828011036, + 0.011686739511787891, + 0.012002943083643913, + 0.010583310388028622, + 0.013362200930714607, + 0.013395051471889019, + 0.015769969671964645, + 0.010675488971173763, + 0.038837820291519165, + 0.036476969718933105, + 0.022084752097725868, + 0.017523234710097313, + 0.019419578835368156, + 0.010538648813962936, + 0.01682082563638687, + 0.010839076712727547, + 0.014507568441331387, + 0.016603516414761543, + 0.012673224322497845, + 0.012577632442116737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.013918010517954826 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602043718099594 + ], + "opposite_logprob_tokens": [ + -3.895601987838745 + ], + "logprob_reward": -0.020602043718099594, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602043718099594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09309734404087067, + 0.06531064957380295, + 0.021053031086921692, + 0.03663881868124008, + 0.06841986626386642, + 0.04003720358014107, + 0.08182492107152939, + 0.1121620163321495, + 0.27701377868652344, + 0.20444247126579285 + ], + "credit_map": [ + 0.011275896802544594, + 0.020871810615062714, + 0.007497678976505995, + 0.009726225398480892, + 0.016881631687283516, + 0.008911722339689732, + 0.007807428482919931, + 0.010124942287802696, + 0.011940806172788143, + 0.023445725440979004, + 0.003281264565885067, + 0.004688195418566465, + 0.008972195908427238, + 0.0041328370571136475, + 0.0028076113667339087, + 0.006042009219527245, + 0.005345484707504511, + 0.0022233740892261267, + 0.0019098588963970542, + 0.0011089487234130502, + 0.005099526606500149, + 0.0011051399633288383, + 0.001107432646676898, + 0.0031532638240605593, + 0.005834060721099377, + 0.004498177208006382, + 0.002991259330883622, + 0.0031068972311913967, + 0.0070051513612270355, + 0.00627909554168582, + 0.002341252751648426, + 0.004582923371344805, + 0.006740447599440813, + 0.01081901229918003, + 0.0068978904746472836, + 0.009175743907690048, + 0.009191734716296196, + 0.0078038303181529045, + 0.012212964706122875, + 0.005578233860433102, + 0.00403921352699399, + 0.0052316817454993725, + 0.005353007465600967, + 0.006584442686289549, + 0.005369154736399651, + 0.0043225958943367004, + 0.004711131565272808, + 0.004425973631441593, + 0.006173529662191868, + 0.010259448550641537, + 0.01107768714427948, + 0.011666610836982727, + 0.010411474853754044, + 0.009643393568694592, + 0.012618675827980042, + 0.009974096901714802, + 0.015065710991621017, + 0.009772084653377533, + 0.021466229110956192, + 0.014297551475465298, + 0.013966969214379787, + 0.012907344847917557, + 0.014622202143073082, + 0.010063908994197845, + 0.030712217092514038, + 0.024730687960982323, + 0.028431003913283348, + 0.019977346062660217, + 0.049510132521390915, + 0.05066589638590813, + 0.03948264196515083, + 0.03350382670760155, + 0.03425503522157669, + 0.022828448563814163, + 0.031475555151700974, + 0.024152955040335655, + 0.024370549246668816, + 0.029133323580026627, + 0.021673347800970078, + 0.016553247347474098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.020602043718099594 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034342855215073 + ], + "opposite_logprob_tokens": [ + -2.3503434658050537 + ], + "logprob_reward": -0.10034342855215073, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034342855215073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2074984312057495, + 0.0875442698597908, + 0.10224410891532898, + 0.14951689541339874, + 0.08328128606081009, + 0.11307962238788605, + 0.03812769427895546, + 0.05185339227318764, + 0.10698147118091583, + 0.0598728246986866 + ], + "credit_map": [ + 0.025182675570249557, + 0.034205738455057144, + 0.017169717699289322, + 0.013926368206739426, + 0.05973523110151291, + 0.02019590511918068, + 0.017662828788161278, + 0.019419968128204346, + 0.015716543421149254, + 0.014746775850653648, + 0.006381653249263763, + 0.011207212693989277, + 0.014154432341456413, + 0.007614573463797569, + 0.0067247855477035046, + 0.010998290963470936, + 0.020581848919391632, + 0.012115289457142353, + 0.007683856412768364, + 0.00889712292701006, + 0.0251199658960104, + 0.010583759285509586, + 0.007283419836312532, + 0.009978844784200191, + 0.029270177707076073, + 0.018247904255986214, + 0.01530150230973959, + 0.013809793628752232, + 0.02379319816827774, + 0.012582322582602501, + 0.012426434084773064, + 0.02408556640148163, + 0.009177598170936108, + 0.011769097298383713, + 0.009346438571810722, + 0.011973482556641102, + 0.011614999733865261, + 0.009442828595638275, + 0.01022045686841011, + 0.009736386127769947, + 0.008033680729568005, + 0.011462935246527195, + 0.018970968201756477, + 0.021618559956550598, + 0.017651522532105446, + 0.014812500216066837, + 0.00946593564003706, + 0.011063522659242153, + 0.003766526933759451, + 0.006317798513919115, + 0.0048306724056601524, + 0.0056326985359191895, + 0.00547422980889678, + 0.005292346701025963, + 0.003499746322631836, + 0.0033136773854494095, + 0.006767231039702892, + 0.0060056047514081, + 0.01068126317113638, + 0.006731375586241484, + 0.006855496671050787, + 0.004324055276811123, + 0.005947666708379984, + 0.004540703725069761, + 0.009875337593257427, + 0.00902494229376316, + 0.007170877885073423, + 0.0070045944303274155, + 0.02703806944191456, + 0.02403920330107212, + 0.00950452871620655, + 0.013323924504220486, + 0.008866708725690842, + 0.005127540323883295, + 0.00630818260833621, + 0.0041284807957708836, + 0.010867123492062092, + 0.010656824335455894, + 0.005374846048653126, + 0.008543119765818119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.10034342855215073 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306625574827194 + ], + "opposite_logprob_tokens": [ + -3.1680662631988525 + ], + "logprob_reward": -0.04306625574827194, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306625574827194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18587596714496613, + 0.04189927503466606, + 0.03010360337793827, + 0.055377326905727386, + 0.08083632588386536, + 0.1273117959499359, + 0.09685533493757248, + 0.04557599127292633, + 0.09849411994218826, + 0.23767025768756866 + ], + "credit_map": [ + 0.029257265850901604, + 0.01582801342010498, + 0.015775669366121292, + 0.010249706916511059, + 0.06624949723482132, + 0.021302323788404465, + 0.014203721657395363, + 0.013009769842028618, + 0.00732837151736021, + 0.01588059403002262, + 0.002301001688465476, + 0.003282802877947688, + 0.005685095675289631, + 0.0017264679772779346, + 0.0016309992643073201, + 0.004063940607011318, + 0.005319452844560146, + 0.004346271511167288, + 0.0023941630497574806, + 0.003012998728081584, + 0.006367554888129234, + 0.0034958641044795513, + 0.00210047559812665, + 0.0030668217223137617, + 0.010473651811480522, + 0.00566997891291976, + 0.0043532829731702805, + 0.003667867975309491, + 0.010150684043765068, + 0.00892451498657465, + 0.0044640167616307735, + 0.007673326879739761, + 0.009878682903945446, + 0.013173251412808895, + 0.006830476690083742, + 0.00926733948290348, + 0.01160387508571148, + 0.007474691607058048, + 0.011704125441610813, + 0.010903876274824142, + 0.009624749422073364, + 0.012592717073857784, + 0.020019369199872017, + 0.031032824888825417, + 0.015435993671417236, + 0.014550448395311832, + 0.013277064077556133, + 0.010778623633086681, + 0.006674896460026503, + 0.012023654766380787, + 0.01396462507545948, + 0.01431320235133171, + 0.013653675094246864, + 0.011292492970824242, + 0.015537831000983715, + 0.009394960477948189, + 0.0049651507288217545, + 0.004633339587599039, + 0.007060580886900425, + 0.0074737537652254105, + 0.006166459526866674, + 0.0061949328519403934, + 0.005300343036651611, + 0.003781429724767804, + 0.008397101424634457, + 0.0070987893268466, + 0.009991275146603584, + 0.008409381844103336, + 0.026899149641394615, + 0.021987536922097206, + 0.007320023607462645, + 0.008390857838094234, + 0.0315069854259491, + 0.026230886578559875, + 0.03222919628024101, + 0.02751918137073517, + 0.03717384487390518, + 0.03710469231009483, + 0.027501236647367477, + 0.018404237926006317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.04306625574827194 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592117682099342 + ], + "opposite_logprob_tokens": [ + -3.895592212677002 + ], + "logprob_reward": -0.020592117682099342, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592117682099342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11815367639064789, + 0.0734667032957077, + 0.16692499816417694, + 0.06453724205493927, + 0.04481663182377815, + 0.09807027131319046, + 0.06136675179004669, + 0.08871497958898544, + 0.11753256618976593, + 0.16641612350940704 + ], + "credit_map": [ + 0.020203247666358948, + 0.016386792063713074, + 0.010618461295962334, + 0.010085529647767544, + 0.024093035608530045, + 0.015665285289287567, + 0.010400108993053436, + 0.010701213032007217, + 0.015428232960402966, + 0.022012600675225258, + 0.0048674969002604485, + 0.005323327146470547, + 0.008896558545529842, + 0.003207910805940628, + 0.004837337881326675, + 0.008893236517906189, + 0.036084432154893875, + 0.0262883473187685, + 0.012584112584590912, + 0.01601376198232174, + 0.03471609577536583, + 0.014738239347934723, + 0.01046160701662302, + 0.016038406640291214, + 0.01226139161735773, + 0.007327030412852764, + 0.00844398234039545, + 0.0062309009954333305, + 0.009958365932106972, + 0.005425266921520233, + 0.006346267182379961, + 0.00854403991252184, + 0.004785057157278061, + 0.0051778811030089855, + 0.005186610389500856, + 0.006137306336313486, + 0.0066275522112846375, + 0.004403040744364262, + 0.006833615712821484, + 0.005665568169206381, + 0.007358937989920378, + 0.011599443852901459, + 0.009257123805582523, + 0.012155069038271904, + 0.015125252306461334, + 0.01450230460613966, + 0.01598973758518696, + 0.012082407251000404, + 0.003970737103372812, + 0.004643506370484829, + 0.009177049621939659, + 0.005709859076887369, + 0.008105549030005932, + 0.008016920648515224, + 0.012760376557707787, + 0.008982756175100803, + 0.013404684141278267, + 0.009774919599294662, + 0.01808151789009571, + 0.010822907090187073, + 0.011086407117545605, + 0.008527517318725586, + 0.010384637862443924, + 0.0066323913633823395, + 0.016468750312924385, + 0.01291025523096323, + 0.013377334922552109, + 0.010997600853443146, + 0.015733014792203903, + 0.018352268263697624, + 0.016192764043807983, + 0.013500570319592953, + 0.02592051401734352, + 0.01463398989289999, + 0.02271255850791931, + 0.015275917947292328, + 0.01876463182270527, + 0.02291616052389145, + 0.01848055236041546, + 0.02771179750561714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.020592117682099342 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, beginning from a slight overhead angle, slowly pans forward, ultimately transforming into an extreme macro, captures the moment a sharp, angular white sugar cube is carefully dropped into a mirror-smooth cup of hot coffee, nestled within a dark ceramic mug. Soft, warm indoor light streams down from above and sideways, creating a bright highlight on the rim of the cup, while the coffee itself, a deep, almost black color, emanates delicate wisps of steam. The sugar cube gently disrupts the calm surface of the liquid, creating a tiny ripple before sinking. As the camera approaches, countless tiny bubbles can be clearly seen surging from the pores of the sugar cube, its hard edges quickly rounding and disintegrating. The dissolved sugar transforms into a milky, viscous, warm current that swirls and diffuses within the dark coffee liquid, forming a dynamic and graceful vortex. The warm tones of the entire scene create an atmosphere of tranquility, healing, and ritualistic concentration.", + "seed": 31788, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_8/index_9253_seed_31788_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04864399917423725, + "vlm_logprob": -0.04864399917423725, + "vlm_yes_no_margin": 3.7750000238418577, + "vlm_token_credit": -0.04864399917423725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015055594965815544 + ], + "opposite_logprob_tokens": [ + -6.0150556564331055 + ], + "logprob_reward": -0.015055594965815544, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.015055594965815544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1666320264339447, + 0.12523554265499115, + 0.1254378855228424, + 0.04981966316699982, + 0.045720137655735016, + 0.036076851189136505, + 0.11244365572929382, + 0.04926283285021782, + 0.16792312264442444, + 0.12144831568002701 + ], + "credit_map": [ + 0.012637482024729252, + 0.029953064396977425, + 0.01501589734107256, + 0.012402490712702274, + 0.04582635313272476, + 0.013083580881357193, + 0.017895285040140152, + 0.019817883148789406, + 0.004768555052578449, + 0.006657104939222336, + 0.012765960767865181, + 0.03705429285764694, + 0.004178640898317099, + 0.005077609792351723, + 0.010103132575750351, + 0.04463023692369461, + 0.009853661991655827, + 0.007430907338857651, + 0.03977106511592865, + 0.006603909656405449, + 0.005161253735423088, + 0.006394165568053722, + 0.007454061880707741, + 0.04276884347200394, + 0.00955266784876585, + 0.010002100840210915, + 0.0038697165437042713, + 0.004132688976824284, + 0.005117656663060188, + 0.0036766279954463243, + 0.007545212749391794, + 0.005922991316765547, + 0.00433078920468688, + 0.0167655348777771, + 0.0026085153222084045, + 0.003752416232600808, + 0.010952367447316647, + 0.002150955842807889, + 0.0024689496494829655, + 0.0026906048879027367, + 0.0054663196206092834, + 0.00434141606092453, + 0.003127248492091894, + 0.003318507457152009, + 0.005781063809990883, + 0.0036617836449295282, + 0.007011023350059986, + 0.003369487589225173, + 0.024212546646595, + 0.010462010279297829, + 0.005146587733179331, + 0.010063670575618744, + 0.013835776597261429, + 0.008797393180429935, + 0.03006894327700138, + 0.009856713935732841, + 0.005658020731061697, + 0.006365825887769461, + 0.0038446420803666115, + 0.004807985387742519, + 0.009236909449100494, + 0.003970233723521233, + 0.012383732944726944, + 0.0029954782221466303, + 0.01705602928996086, + 0.019217967987060547, + 0.049649156630039215, + 0.035997144877910614, + 0.017753267660737038, + 0.007573210634291172, + 0.013053919188678265, + 0.007622424978762865, + 0.014171641319990158, + 0.01475772075355053, + 0.025766899809241295, + 0.014589142054319382, + 0.008078373037278652, + 0.009143143892288208, + 0.012016626074910164, + 0.0229247584939003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.015055594965815544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016085024923086166 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.016085024923086166, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016085024923086166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060787491500377655, + 0.19244039058685303, + 0.08746185898780823, + 0.02171907015144825, + 0.03716028109192848, + 0.1597864180803299, + 0.1777660846710205, + 0.03303305804729462, + 0.16918687522411346, + 0.06065845116972923 + ], + "credit_map": [ + 0.005419699475169182, + 0.009797899052500725, + 0.008549236692488194, + 0.007667218334972858, + 0.011281391605734825, + 0.004906396381556988, + 0.007186420727521181, + 0.005979232024401426, + 0.006645756773650646, + 0.00971759669482708, + 0.022025300189852715, + 0.054512329399585724, + 0.008062883280217648, + 0.007701885886490345, + 0.015169523656368256, + 0.06860511004924774, + 0.011744753457605839, + 0.006745730992406607, + 0.029097680002450943, + 0.004192663822323084, + 0.004378292243927717, + 0.004759042523801327, + 0.004452972672879696, + 0.022090720012784004, + 0.004896225873380899, + 0.004065867513418198, + 0.0019416736904531717, + 0.001296374131925404, + 0.002884147223085165, + 0.002115268725901842, + 0.001729196752421558, + 0.00279031484387815, + 0.0058674984611570835, + 0.006890537682920694, + 0.0024177010636776686, + 0.0037491891998797655, + 0.005998317617923021, + 0.003307271981611848, + 0.0036433630157262087, + 0.005286401137709618, + 0.026329686865210533, + 0.030576352030038834, + 0.012795369140803814, + 0.018203221261501312, + 0.020121220499277115, + 0.013448677957057953, + 0.020190246403217316, + 0.018121642991900444, + 0.04053817689418793, + 0.009310265071690083, + 0.007258548866957426, + 0.013752608560025692, + 0.013498781248927116, + 0.009038710035383701, + 0.07176172733306885, + 0.012607264332473278, + 0.00794833805412054, + 0.003765149274840951, + 0.0031766286119818687, + 0.003295480040833354, + 0.004480272531509399, + 0.00392333185300231, + 0.0036078142002224922, + 0.0028360430151224136, + 0.012820959091186523, + 0.020398477092385292, + 0.04524152725934982, + 0.048813991248607635, + 0.013331412337720394, + 0.006252215709537268, + 0.012236903421580791, + 0.010091402567923069, + 0.008079673163592815, + 0.007188953924924135, + 0.01319555938243866, + 0.006070026196539402, + 0.005296322517096996, + 0.005718570668250322, + 0.005695698317140341, + 0.009413652122020721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.016085024923086166 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282836019992828 + ], + "opposite_logprob_tokens": [ + -2.0178284645080566 + ], + "logprob_reward": -0.14282836019992828, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14282836019992828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08042997866868973, + 0.15013085305690765, + 0.0829373300075531, + 0.032844845205545425, + 0.036305565387010574, + 0.29633185267448425, + 0.12687671184539795, + 0.03145275264978409, + 0.12182030081748962, + 0.040869832038879395 + ], + "credit_map": [ + 0.005554330535233021, + 0.02470373921096325, + 0.009566506370902061, + 0.007623543497174978, + 0.014962812885642052, + 0.005071726161986589, + 0.007929778657853603, + 0.005017539020627737, + 0.0059921713545918465, + 0.010541358031332493, + 0.010009455494582653, + 0.047753795981407166, + 0.011711079627275467, + 0.006085213739424944, + 0.012510654516518116, + 0.04552712291479111, + 0.011068773455917835, + 0.012317555025219917, + 0.026062987744808197, + 0.005511379800736904, + 0.007553524803370237, + 0.005787353962659836, + 0.006846662610769272, + 0.007789089344441891, + 0.008167117834091187, + 0.0058273375034332275, + 0.001991285476833582, + 0.002030791249126196, + 0.004467132035642862, + 0.0025647657457739115, + 0.003541678888723254, + 0.0042547364719212055, + 0.008657842874526978, + 0.012255476787686348, + 0.0016811757814139128, + 0.0021578017622232437, + 0.004984740167856216, + 0.001899577910080552, + 0.0019103195518255234, + 0.0027586310170590878, + 0.04919455572962761, + 0.054711148142814636, + 0.02301972731947899, + 0.03376992419362068, + 0.03492646664381027, + 0.024711623787879944, + 0.04267466813325882, + 0.033323734998703, + 0.03173213079571724, + 0.007991029880940914, + 0.005174691788852215, + 0.009644091129302979, + 0.0095775555819273, + 0.006241894327104092, + 0.0490584522485733, + 0.007456866558641195, + 0.004078818950802088, + 0.0039281090721488, + 0.0025530834682285786, + 0.0025530504062771797, + 0.006748584099113941, + 0.003188490867614746, + 0.006134185008704662, + 0.002268429845571518, + 0.03627568483352661, + 0.010547707788646221, + 0.029106173664331436, + 0.016643203794956207, + 0.008296440355479717, + 0.006160019896924496, + 0.007035619579255581, + 0.0077554588206112385, + 0.0058447085320949554, + 0.005092903971672058, + 0.005063088610768318, + 0.007654696702957153, + 0.003749195486307144, + 0.0036876900121569633, + 0.004937711171805859, + 0.0048398361541330814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.14282836019992828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865773022174835 + ], + "opposite_logprob_tokens": [ + -3.0486576557159424 + ], + "logprob_reward": -0.04865773022174835, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865773022174835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08326207846403122, + 0.030325645580887794, + 0.06406979262828827, + 0.02071143314242363, + 0.037242818623781204, + 0.07674463093280792, + 0.06679720431566238, + 0.03819238767027855, + 0.5299960970878601, + 0.05265793576836586 + ], + "credit_map": [ + 0.012727629393339157, + 0.01078889798372984, + 0.013051548041403294, + 0.005627183243632317, + 0.017798641696572304, + 0.0071665383875370026, + 0.006835102569311857, + 0.009266536682844162, + 0.0026206434704363346, + 0.0049957213923335075, + 0.0029403313528746367, + 0.004891413729637861, + 0.006415144540369511, + 0.0017172748921439052, + 0.0019670617766678333, + 0.00477805407717824, + 0.005273381248116493, + 0.007230383809655905, + 0.018064267933368683, + 0.0042497506365180016, + 0.003725379705429077, + 0.003601296106353402, + 0.004784050397574902, + 0.017141282558441162, + 0.0027579045854508877, + 0.005218456964939833, + 0.0015724346740171313, + 0.001652497798204422, + 0.002206074772402644, + 0.001486720284447074, + 0.003011435503140092, + 0.0028059084434062243, + 0.005156913306564093, + 0.007464369758963585, + 0.0020816426258534193, + 0.007044278550893068, + 0.006693090777844191, + 0.0024468658957630396, + 0.0032585698645561934, + 0.0030970852822065353, + 0.011444754898548126, + 0.014336970634758472, + 0.005621104035526514, + 0.009166029281914234, + 0.010124200955033302, + 0.006766092963516712, + 0.013316674157977104, + 0.005968810524791479, + 0.016372378915548325, + 0.003908978309482336, + 0.0031908631790429354, + 0.0054265353828668594, + 0.0051228320226073265, + 0.003491026582196355, + 0.025049149990081787, + 0.004235442262142897, + 0.00520716467872262, + 0.005234844982624054, + 0.003227002453058958, + 0.004792420193552971, + 0.007796516641974449, + 0.004191600251942873, + 0.004633639939129353, + 0.003109199460595846, + 0.0753345936536789, + 0.047125209122896194, + 0.1883418709039688, + 0.10950085520744324, + 0.03530731052160263, + 0.019544238224625587, + 0.02980649657547474, + 0.02503550611436367, + 0.008181530982255936, + 0.006969335954636335, + 0.004522038158029318, + 0.012069636955857277, + 0.004842595662921667, + 0.0047757006250321865, + 0.006540960632264614, + 0.004756137728691101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.04865773022174835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up that begins from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan forward and transform into an extreme macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sharp, angular white sugar cube dropped into a mirror-smooth cup of dark coffee inside a dark ceramic mug?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny bubbles surge from the sugar cube's pores as it disintegrates, creating a milky, viscous current that swirls into a vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, warm indoor light streaming from above and sideways, creating a bright highlight on the cup's rim?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059328556060791 + ], + "opposite_logprob_tokens": [ + -3.8955931663513184 + ], + "logprob_reward": -0.02059328556060791, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059328556060791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11012744158506393, + 0.06964111328125, + 0.1381470412015915, + 0.04339457303285599, + 0.04974818974733353, + 0.14460468292236328, + 0.08187124878168106, + 0.027203969657421112, + 0.2751200795173645, + 0.06014172360301018 + ], + "credit_map": [ + 0.007265838794410229, + 0.03157436102628708, + 0.014450998976826668, + 0.00891965813934803, + 0.018740138038992882, + 0.006658029276877642, + 0.012665841728448868, + 0.009852578863501549, + 0.003140910528600216, + 0.0062208776362240314, + 0.006174711976200342, + 0.017694123089313507, + 0.006513651926070452, + 0.002949503483250737, + 0.004647768568247557, + 0.022299563512206078, + 0.013076895847916603, + 0.007321762386709452, + 0.04788060113787651, + 0.007674151100218296, + 0.005077837500721216, + 0.00610698014497757, + 0.008566734381020069, + 0.04244207218289375, + 0.0049156304448843, + 0.01009411457926035, + 0.002947542117908597, + 0.003526638261973858, + 0.0046738870441913605, + 0.0025981587823480368, + 0.01105316448956728, + 0.003585440805181861, + 0.005299486219882965, + 0.01512821577489376, + 0.002625594846904278, + 0.006506725214421749, + 0.009921520948410034, + 0.0024486861657351255, + 0.005129466764628887, + 0.0026884968392550945, + 0.023748591542243958, + 0.027757376432418823, + 0.010449959896504879, + 0.017189063131809235, + 0.01842610538005829, + 0.011790974996984005, + 0.021327469497919083, + 0.013915129005908966, + 0.021383821964263916, + 0.0047919610515236855, + 0.0038266892079263926, + 0.006221798248589039, + 0.005946002434939146, + 0.00462174229323864, + 0.029820207506418228, + 0.005259022116661072, + 0.0051721748895943165, + 0.003274190239608288, + 0.0024256452452391386, + 0.002576051978394389, + 0.0044571710750460625, + 0.0029476473573595285, + 0.0033545554615557194, + 0.002996534574776888, + 0.035406701266765594, + 0.025021450594067574, + 0.09681056439876556, + 0.06120169162750244, + 0.017767302691936493, + 0.011163626797497272, + 0.015556123107671738, + 0.012192624621093273, + 0.00726762181147933, + 0.006153132300823927, + 0.013095797970890999, + 0.011765819042921066, + 0.003966445568948984, + 0.004217270761728287, + 0.006379167549312115, + 0.007296469993889332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.02059328556060791 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_9/index_tensor([9345, 9345])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_9/index_tensor([9345, 9345])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..799a2b35bb4016f86c65817f94fff0c71a4b2047 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_9/index_tensor([9345, 9345])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium, static shot, shot from a slightly upward angle, focuses on a thin, middle-aged strategist with a neatly trimmed goatee, nestled within a dimly lit military tent. He wears a loose, flowing dark blue silk robe, his long hair held back with a simple wooden hairpin. He reclines leisurely on an antique carved wooden chair. The only illumination comes from a bronze oil lamp on a table beside him. Its flickering, warm yellow light casts deep, undulating shadows on his face, emphasizing his pensive expression and illuminating the folds of his silk robe with a soft sheen. In his right hand, he grasps a feather fan made of pure white crane feathers, gently waving it with a steady, leisurely rhythm, the breeze occasionally causing the lamplight to flicker. In the background, a military map hangs on the thick dark canvas tent wall, and a wooden weapons rack with unfolded bamboo slips and weapons can be vaguely seen next to it. The whole scene has a warm tone, mainly deep dark colors, creating an atmosphere of extreme concentration, tranquility and strategic planning against the backdrop of military tension.", + "seed": 1223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_9/index_9345_seed_1223_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.29199293730780485, + "vlm_logprob": -0.29199293730780485, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.29199293730780485, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.074587881565094 + ], + "opposite_logprob_tokens": [ + -2.699587821960449 + ], + "logprob_reward": -0.074587881565094, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.074587881565094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2822725176811218, + 0.08728143572807312, + 0.060473620891571045, + 0.14367520809173584, + 0.09407734870910645, + 0.0696486085653305, + 0.08811046183109283, + 0.05257117748260498, + 0.07525893300771713, + 0.04663066565990448 + ], + "credit_map": [ + 0.021659500896930695, + 0.027516771107912064, + 0.029968764632940292, + 0.03344592824578285, + 0.04530888423323631, + 0.04177645593881607, + 0.030900586396455765, + 0.051695629954338074, + 0.010708019137382507, + 0.012221597135066986, + 0.00828568171709776, + 0.008026049472391605, + 0.009472634643316269, + 0.015006990171968937, + 0.009198572486639023, + 0.014361902140080929, + 0.007106543518602848, + 0.006751832086592913, + 0.006065430119633675, + 0.006216104608029127, + 0.007033834233880043, + 0.009614132344722748, + 0.00565463537350297, + 0.012031115591526031, + 0.01554368156939745, + 0.018237974494695663, + 0.011823142878711224, + 0.01802029460668564, + 0.018359987065196037, + 0.019882675260305405, + 0.019169379025697708, + 0.022638089954853058, + 0.01275859959423542, + 0.007247405592352152, + 0.007920457981526852, + 0.009986316785216331, + 0.0090800691395998, + 0.014774106442928314, + 0.009521422907710075, + 0.022788984701037407, + 0.009286627173423767, + 0.0076574101112782955, + 0.008840896189212799, + 0.008821157738566399, + 0.007697702385485172, + 0.009146134369075298, + 0.008372177369892597, + 0.009826511144638062, + 0.009622661396861076, + 0.009831641800701618, + 0.007090591825544834, + 0.011303704231977463, + 0.013302760198712349, + 0.011505176313221455, + 0.013199429959058762, + 0.0122545026242733, + 0.005673735868185759, + 0.007198800798505545, + 0.005222564097493887, + 0.004657550249248743, + 0.00705679040402174, + 0.007986683398485184, + 0.007063570898026228, + 0.007711485493928194, + 0.007969751954078674, + 0.009452158585190773, + 0.006913542281836271, + 0.008721708320081234, + 0.011048026382923126, + 0.01091300044208765, + 0.01093310583382845, + 0.009307649917900562, + 0.005049970932304859, + 0.005582361947745085, + 0.005624840501695871, + 0.004466830752789974, + 0.005853734444826841, + 0.005389556288719177, + 0.006470763124525547, + 0.008192610926926136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.074587881565094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960191134363413 + ], + "opposite_logprob_tokens": [ + -5.130959987640381 + ], + "logprob_reward": -0.005960191134363413, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960191134363413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18861007690429688, + 0.08474615216255188, + 0.07003764808177948, + 0.06044391915202141, + 0.12498172372579575, + 0.09962163865566254, + 0.19692763686180115, + 0.045462220907211304, + 0.08284570276737213, + 0.04632328823208809 + ], + "credit_map": [ + 0.021909167990088463, + 0.017010189592838287, + 0.027358610183000565, + 0.025339119136333466, + 0.020132631063461304, + 0.026266416534781456, + 0.025184964761137962, + 0.02540896274149418, + 0.008847558870911598, + 0.008554874919354916, + 0.0074900672771036625, + 0.008134361356496811, + 0.017488745972514153, + 0.011138181202113628, + 0.011617601849138737, + 0.01147475652396679, + 0.00793167669326067, + 0.008988668210804462, + 0.009353257715702057, + 0.007773193065077066, + 0.0075224475003778934, + 0.010582538321614265, + 0.006548477336764336, + 0.011337386444211006, + 0.0068794479593634605, + 0.007376738358289003, + 0.005467639770358801, + 0.00676885899156332, + 0.008326854556798935, + 0.008446368388831615, + 0.007604554761201143, + 0.009573450312018394, + 0.013690273277461529, + 0.01004155445843935, + 0.009640022180974483, + 0.017295949161052704, + 0.015470348298549652, + 0.01666494458913803, + 0.021958358585834503, + 0.020220262929797173, + 0.013541975989937782, + 0.008325557224452496, + 0.006867929361760616, + 0.009420830756425858, + 0.00964145828038454, + 0.015747778117656708, + 0.011751041747629642, + 0.024325057864189148, + 0.018294965848326683, + 0.025079751387238503, + 0.023437337949872017, + 0.019900323823094368, + 0.028664611279964447, + 0.0281243696808815, + 0.025484340265393257, + 0.027941938489675522, + 0.00557774817571044, + 0.004116876050829887, + 0.0039049021434038877, + 0.004458366893231869, + 0.007405328098684549, + 0.006635424215346575, + 0.005804680287837982, + 0.007558892481029034, + 0.007019946817308664, + 0.008891034871339798, + 0.005297141149640083, + 0.011879541911184788, + 0.012324572540819645, + 0.011403406970202923, + 0.01732887513935566, + 0.008701184764504433, + 0.0041879708878695965, + 0.004758664406836033, + 0.0036756473127752542, + 0.005891328677535057, + 0.005201802123337984, + 0.006283169146627188, + 0.010058313608169556, + 0.006266389042139053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.005960191134363413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673630740493536 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.00673630740493536, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673630740493536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26308339834213257, + 0.09644889831542969, + 0.16231173276901245, + 0.07053713500499725, + 0.14681652188301086, + 0.06165513023734093, + 0.04960641264915466, + 0.05867116153240204, + 0.05910010635852814, + 0.03176955506205559 + ], + "credit_map": [ + 0.04316151142120361, + 0.029972631484270096, + 0.036790043115615845, + 0.025633495301008224, + 0.03247766196727753, + 0.033629730343818665, + 0.022505104541778564, + 0.038913220167160034, + 0.012019963003695011, + 0.014168528839945793, + 0.015615914948284626, + 0.009347254410386086, + 0.010258298367261887, + 0.012700937688350677, + 0.010086992755532265, + 0.012251006439328194, + 0.020761171355843544, + 0.019186560064554214, + 0.021314583718776703, + 0.016007913276553154, + 0.01915481872856617, + 0.02523287758231163, + 0.012997156009078026, + 0.02765665389597416, + 0.00653205718845129, + 0.011037320829927921, + 0.006473534274846315, + 0.010623911395668983, + 0.008301688358187675, + 0.008343652822077274, + 0.009290219284594059, + 0.009934748522937298, + 0.020211748778820038, + 0.01243719644844532, + 0.012685746885836124, + 0.016596021130681038, + 0.015366312116384506, + 0.024239709600806236, + 0.014787225052714348, + 0.030492549762129784, + 0.006871949415653944, + 0.007753994781523943, + 0.005917737726122141, + 0.005624597892165184, + 0.006797974929213524, + 0.009748993441462517, + 0.005649547558277845, + 0.01329033449292183, + 0.005246924236416817, + 0.005285961553454399, + 0.005502644926309586, + 0.0047986735589802265, + 0.009093176573514938, + 0.007211728487163782, + 0.006777750328183174, + 0.005689556710422039, + 0.006426002364605665, + 0.005087051074951887, + 0.006336715072393417, + 0.0055358437821269035, + 0.010592741891741753, + 0.008235308341681957, + 0.007782694883644581, + 0.008674802258610725, + 0.008020474575459957, + 0.007708120159804821, + 0.0048707821406424046, + 0.004823809489607811, + 0.006639545317739248, + 0.009293033741414547, + 0.005367615260183811, + 0.012376725673675537, + 0.003935352433472872, + 0.003653668100014329, + 0.0034586950205266476, + 0.003380573121830821, + 0.004177556838840246, + 0.004173988476395607, + 0.0037004926707595587, + 0.00528922863304615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.00673630740493536 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.474315881729126 + ], + "opposite_logprob_tokens": [ + -0.974315881729126 + ], + "logprob_reward": -0.474315881729126, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.474315881729126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3433471918106079, + 0.09717321395874023, + 0.0955333560705185, + 0.06103682145476341, + 0.07040724903345108, + 0.07800715416669846, + 0.04229072108864784, + 0.04866991192102432, + 0.1360248476266861, + 0.027509525418281555 + ], + "credit_map": [ + 0.032822947949171066, + 0.021989945322275162, + 0.02813015505671501, + 0.028370557352900505, + 0.040363967418670654, + 0.07333549112081528, + 0.04178929328918457, + 0.07654483616352081, + 0.011214742437005043, + 0.010585934855043888, + 0.011202465742826462, + 0.009454742074012756, + 0.015108861029148102, + 0.012262612581253052, + 0.014162098057568073, + 0.013181755319237709, + 0.009013948030769825, + 0.011699872091412544, + 0.010425611399114132, + 0.01276956032961607, + 0.0133149279281497, + 0.012334508821368217, + 0.012061964720487595, + 0.013912963680922985, + 0.00558024225756526, + 0.006381141021847725, + 0.005656673572957516, + 0.008814006112515926, + 0.007748698815703392, + 0.007877256721258163, + 0.009985535405576229, + 0.008993270806968212, + 0.008432621136307716, + 0.008936718106269836, + 0.008522024378180504, + 0.007105211727321148, + 0.008057272993028164, + 0.00935971736907959, + 0.009202828630805016, + 0.010790854692459106, + 0.012820090167224407, + 0.014191126450896263, + 0.0058502620086073875, + 0.0072788349352777, + 0.006779313087463379, + 0.011918585747480392, + 0.006281523033976555, + 0.012887413613498211, + 0.004531137645244598, + 0.00454071955755353, + 0.004810205660760403, + 0.004132850095629692, + 0.006620138417929411, + 0.0056740231812000275, + 0.00549524649977684, + 0.006486397702246904, + 0.005430229939520359, + 0.00508072366937995, + 0.0049732886254787445, + 0.005386383272707462, + 0.005811033770442009, + 0.006652298383414745, + 0.00814682524651289, + 0.007189126219600439, + 0.011253749020397663, + 0.012490264140069485, + 0.007037443574517965, + 0.0212499238550663, + 0.016598740592598915, + 0.018646497279405594, + 0.03256946802139282, + 0.01617874763906002, + 0.0030503487214446068, + 0.0031285651493817568, + 0.0028782878071069717, + 0.0027832267805933952, + 0.0029475928749889135, + 0.003611211897805333, + 0.0038071228191256523, + 0.005303168669342995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.474315881729126 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983644247055054 + ], + "opposite_logprob_tokens": [ + -0.5233644247055054 + ], + "logprob_reward": -0.8983644247055054, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983644247055054, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20672228932380676, + 0.11284253746271133, + 0.10351702570915222, + 0.10096877813339233, + 0.17702697217464447, + 0.052280910313129425, + 0.03904779627919197, + 0.08186161518096924, + 0.047381963580846786, + 0.07835017889738083 + ], + "credit_map": [ + 0.02500648982822895, + 0.028437737375497818, + 0.028608977794647217, + 0.025097794830799103, + 0.02501317486166954, + 0.024387981742620468, + 0.025519059970974922, + 0.02465103380382061, + 0.013820803724229336, + 0.013033770956099033, + 0.01484009064733982, + 0.009351502172648907, + 0.011280514299869537, + 0.0192122682929039, + 0.01125565730035305, + 0.02004791423678398, + 0.012713891454041004, + 0.012506178580224514, + 0.012774597853422165, + 0.012233796529471874, + 0.010463242419064045, + 0.014717266894876957, + 0.010909002274274826, + 0.01719902828335762, + 0.011371620930731297, + 0.013780680485069752, + 0.008762484416365623, + 0.01293645054101944, + 0.010275695472955704, + 0.013509091921150684, + 0.00874494668096304, + 0.02158779464662075, + 0.03083239123225212, + 0.01273294910788536, + 0.016252394765615463, + 0.016301343217492104, + 0.020877409726381302, + 0.022157184779644012, + 0.02183002606034279, + 0.03604322671890259, + 0.00933003332465887, + 0.006492725573480129, + 0.004418470431119204, + 0.005274311639368534, + 0.0049078986048698425, + 0.006848610006272793, + 0.008014842867851257, + 0.00699401181191206, + 0.0041108629666268826, + 0.003294222289696336, + 0.004190734587609768, + 0.005611130967736244, + 0.004768616985529661, + 0.004817350767552853, + 0.006710848771035671, + 0.0055440254509449005, + 0.009981968440115452, + 0.01649068109691143, + 0.0062599265947937965, + 0.006425608415156603, + 0.010352172888815403, + 0.011454327031970024, + 0.0064769103191792965, + 0.014420009218156338, + 0.0056604682467877865, + 0.005841961596161127, + 0.0042201620526611805, + 0.0053878589533269405, + 0.005947785452008247, + 0.007605487015098333, + 0.005844090599566698, + 0.006874144542962313, + 0.01096292957663536, + 0.010211589746177197, + 0.011888855136930943, + 0.005112107843160629, + 0.00890006497502327, + 0.011627895757555962, + 0.006701345555484295, + 0.012945376336574554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.8983644247055054 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium, static shot, shot from a slightly upward angle, focuses on a thin, middle-aged strategist with a neatly trimmed goatee, nestled within a dimly lit military tent. He wears a loose, flowing dark blue silk robe, his long hair held back with a simple wooden hairpin. He reclines leisurely on an antique carved wooden chair. The only illumination comes from a bronze oil lamp on a table beside him. Its flickering, warm yellow light casts deep, undulating shadows on his face, emphasizing his pensive expression and illuminating the folds of his silk robe with a soft sheen. In his right hand, he grasps a feather fan made of pure white crane feathers, gently waving it with a steady, leisurely rhythm, the breeze occasionally causing the lamplight to flicker. In the background, a military map hangs on the thick dark canvas tent wall, and a wooden weapons rack with unfolded bamboo slips and weapons can be vaguely seen next to it. The whole scene has a warm tone, mainly deep dark colors, creating an atmosphere of extreme concentration, tranquility and strategic planning against the backdrop of military tension.", + "seed": 1224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_9/index_9345_seed_1224_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13525333600118755, + "vlm_logprob": -0.13525333600118755, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.13525333600118755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031513284891843796 + ], + "opposite_logprob_tokens": [ + -3.531513214111328 + ], + "logprob_reward": -0.031513284891843796, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.031513284891843796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21061281859874725, + 0.44545304775238037, + 0.020239703357219696, + 0.04037934169173241, + 0.0228742528706789, + 0.12253689020872116, + 0.02972663752734661, + 0.010893217287957668, + 0.04923848807811737, + 0.04804566875100136 + ], + "credit_map": [ + 0.032694652676582336, + 0.01759650930762291, + 0.020786622539162636, + 0.005043873563408852, + 0.01422115322202444, + 0.08879324048757553, + 0.007791085168719292, + 0.023685673251748085, + 0.05533578619360924, + 0.052270181477069855, + 0.041813258081674576, + 0.01048809103667736, + 0.02447708696126938, + 0.19633644819259644, + 0.01612200029194355, + 0.048610150814056396, + 0.0030580582097172737, + 0.0032553686760365963, + 0.0033025387674570084, + 0.0019628554582595825, + 0.001966647570952773, + 0.0026758320163935423, + 0.0016107055125758052, + 0.002407695632427931, + 0.004181471653282642, + 0.014157201163470745, + 0.004890637472271919, + 0.0017078875098377466, + 0.0020735065918415785, + 0.006864840630441904, + 0.0016430763062089682, + 0.004860714543610811, + 0.0036318539641797543, + 0.0016042726347222924, + 0.002376277931034565, + 0.000692369940225035, + 0.0013562573585659266, + 0.009171800687909126, + 0.0013573667965829372, + 0.0026840525679290295, + 0.03402440994977951, + 0.005923776887357235, + 0.010779881849884987, + 0.002683222759515047, + 0.006078766658902168, + 0.04595290496945381, + 0.005265181418508291, + 0.011828736402094364, + 0.0038625779561698437, + 0.0017891948809847236, + 0.004703715443611145, + 0.001021760399453342, + 0.0011921536643058062, + 0.009837237186729908, + 0.0015350269386544824, + 0.005784968845546246, + 0.001016705995425582, + 0.0027593905106186867, + 0.0017879067454487085, + 0.0005649646045640111, + 0.0006457791314460337, + 0.0015871054492890835, + 0.0008273445419035852, + 0.0017040196107700467, + 0.014675209298729897, + 0.0022916377056390047, + 0.005591254681348801, + 0.0013216359075158834, + 0.0017898075748234987, + 0.006293053738772869, + 0.0018180971965193748, + 0.015457785688340664, + 0.014192081987857819, + 0.002409768057987094, + 0.003115636296570301, + 0.0013871019473299384, + 0.002361180493608117, + 0.006628239527344704, + 0.0018841024721041322, + 0.01606755144894123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.031513284891843796 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674400432035327 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.00674400432035327, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674400432035327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30145159363746643, + 0.20333382487297058, + 0.038926370441913605, + 0.05518752336502075, + 0.005714376922696829, + 0.02572842501103878, + 0.044949598610401154, + 0.012854526750743389, + 0.030221587046980858, + 0.2816321551799774 + ], + "credit_map": [ + 0.03630530834197998, + 0.026446446776390076, + 0.02539844438433647, + 0.0059904553927481174, + 0.02167954295873642, + 0.13527387380599976, + 0.010674618184566498, + 0.03968289494514465, + 0.02594599686563015, + 0.023868635296821594, + 0.018236862495541573, + 0.004899204708635807, + 0.011210293509066105, + 0.08945416659116745, + 0.007352778222411871, + 0.022365879267454147, + 0.004088780377060175, + 0.0035656928084790707, + 0.011735408566892147, + 0.002523180330172181, + 0.0025107755791395903, + 0.008214861154556274, + 0.002050624927505851, + 0.00423704506829381, + 0.0056678117252886295, + 0.01822337508201599, + 0.00517689436674118, + 0.0018595537403598428, + 0.0026576032396405935, + 0.014702720567584038, + 0.001853334833867848, + 0.005046228878200054, + 0.0008008468430489302, + 0.0005751887802034616, + 0.0007529655122198164, + 0.0004403984930831939, + 0.0007663918077014387, + 0.000918464851565659, + 0.0006326731527224183, + 0.0008274473948404193, + 0.0032368625979870558, + 0.0013871223200112581, + 0.0012912987731397152, + 0.0004992959438823164, + 0.001034189946949482, + 0.010830128565430641, + 0.0012247508857399225, + 0.006224776152521372, + 0.011509614996612072, + 0.0022854262497276068, + 0.009343474172055721, + 0.0013141537783667445, + 0.001995420316234231, + 0.006485152989625931, + 0.0028272666968405247, + 0.009189088828861713, + 0.001801490900106728, + 0.003990719094872475, + 0.0016947995172813535, + 0.0006313549820333719, + 0.0007917691254988313, + 0.0017512220656499267, + 0.0008165111648850143, + 0.0013766594929620624, + 0.004177636932581663, + 0.0021868159528821707, + 0.008213978260755539, + 0.0018314376939088106, + 0.0016482684295624495, + 0.003320760326460004, + 0.0018024215241894126, + 0.007040268275886774, + 0.06111597642302513, + 0.010779788717627525, + 0.022704560309648514, + 0.005543872248381376, + 0.010491170920431614, + 0.047887757420539856, + 0.011100920848548412, + 0.11200813949108124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.00674400432035327 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110772168263793 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.0110772168263793, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110772168263793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31714728474617004, + 0.34464848041534424, + 0.07185132801532745, + 0.027411509305238724, + 0.01318933442234993, + 0.06053876131772995, + 0.029836099594831467, + 0.00924660824239254, + 0.03079547919332981, + 0.09533512592315674 + ], + "credit_map": [ + 0.036859024316072464, + 0.029009856283664703, + 0.02608303725719452, + 0.007571336347609758, + 0.022880403324961662, + 0.14106246829032898, + 0.010584509931504726, + 0.04309667646884918, + 0.03326708823442459, + 0.04148656502366066, + 0.04166695848107338, + 0.008354139514267445, + 0.018503347411751747, + 0.15084527432918549, + 0.012593797408044338, + 0.03793134167790413, + 0.007421231362968683, + 0.006870059762150049, + 0.019620046019554138, + 0.005897786933928728, + 0.0038364192005246878, + 0.016633866354823112, + 0.004248928744345903, + 0.00732299592345953, + 0.001938801840879023, + 0.00816537905484438, + 0.0022359180729836226, + 0.0008470158791169524, + 0.0014066356234252453, + 0.009473900310695171, + 0.0010166065767407417, + 0.0023272568359971046, + 0.002293448196724057, + 0.0018819133983924985, + 0.0020370809361338615, + 0.0010566450655460358, + 0.0010686431778594851, + 0.0024654276203364134, + 0.0007486332324333489, + 0.0016375449486076832, + 0.0041144839487969875, + 0.003373674349859357, + 0.004191030748188496, + 0.0016192342154681683, + 0.0026423337403684855, + 0.024671442806720734, + 0.0031367649789899588, + 0.01678980514407158, + 0.004036539234220982, + 0.0013920925557613373, + 0.0013197959633544087, + 0.0005494055221788585, + 0.0009129785466939211, + 0.009150231257081032, + 0.0014074931386858225, + 0.011067565530538559, + 0.0025699776597321033, + 0.0016094130696728826, + 0.0017164688324555755, + 0.0005620611482299864, + 0.0004321612068451941, + 0.0009273833711631596, + 0.0004507045669015497, + 0.000978439231403172, + 0.008752630092203617, + 0.0015635862946510315, + 0.003066748846322298, + 0.0009167221141979098, + 0.001231072237715125, + 0.004029492847621441, + 0.0010701726423576474, + 0.010165059007704258, + 0.025405243039131165, + 0.0038528020959347486, + 0.005556269083172083, + 0.0018015825189650059, + 0.0035093724727630615, + 0.01474045030772686, + 0.002929250244051218, + 0.037540167570114136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0110772168263793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134816884994507 + ], + "opposite_logprob_tokens": [ + -1.3134816884994507 + ], + "logprob_reward": -0.3134816884994507, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134816884994507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5212172865867615, + 0.0340350903570652, + 0.04324379935860634, + 0.018497902899980545, + 0.014163616113364697, + 0.0766138806939125, + 0.06409906595945358, + 0.017056332901120186, + 0.13203951716423035, + 0.07903344184160233 + ], + "credit_map": [ + 0.06502015888690948, + 0.045086588710546494, + 0.04704840108752251, + 0.011306743137538433, + 0.03725399449467659, + 0.2322940081357956, + 0.01847720518708229, + 0.0647301971912384, + 0.006022103596478701, + 0.0038356429431587458, + 0.002176583744585514, + 0.0010291499784216285, + 0.002074315445497632, + 0.013908935710787773, + 0.0014184651663526893, + 0.003569894703105092, + 0.005505949258804321, + 0.005257521755993366, + 0.01117057353258133, + 0.003500947030261159, + 0.004122843965888023, + 0.006289032753556967, + 0.003150864504277706, + 0.004246068652719259, + 0.001826238352805376, + 0.0045765056274831295, + 0.001301351236179471, + 0.0009633235749788582, + 0.00122739071957767, + 0.004588065203279257, + 0.0009764137212187052, + 0.0030386154539883137, + 0.002330842660740018, + 0.001462986576370895, + 0.00184072635602206, + 0.0007927033584564924, + 0.0011194541584700346, + 0.0036116591654717922, + 0.0009905125480145216, + 0.0020147308241575956, + 0.022164316847920418, + 0.003950618673115969, + 0.006361367646604776, + 0.001851393491961062, + 0.0037776161916553974, + 0.0277568269520998, + 0.003365771844983101, + 0.007385976612567902, + 0.022285714745521545, + 0.0037435051053762436, + 0.005517050623893738, + 0.0019506714306771755, + 0.0026368247345089912, + 0.016033625230193138, + 0.0033910830970853567, + 0.008540593087673187, + 0.002862523775547743, + 0.0034708636812865734, + 0.001783084822818637, + 0.00092245580162853, + 0.0012817594688385725, + 0.002687513828277588, + 0.0009987674420699477, + 0.003049364546313882, + 0.027424562722444534, + 0.007283926475793123, + 0.017194565385580063, + 0.004299914930015802, + 0.004713223781436682, + 0.01910601183772087, + 0.004752371460199356, + 0.047264937311410904, + 0.023944828659296036, + 0.006755263544619083, + 0.012009766884148121, + 0.004807871300727129, + 0.003667803481221199, + 0.008081602863967419, + 0.003712156554684043, + 0.016054147854447365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.3134816884994507 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31345048546791077 + ], + "opposite_logprob_tokens": [ + -1.3134504556655884 + ], + "logprob_reward": -0.31345048546791077, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31345048546791077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5263336300849915, + 0.2851378917694092, + 0.0651712641119957, + 0.019983891397714615, + 0.02561277337372303, + 0.036833662539720535, + 0.0110729755833745, + 0.0032171804923564196, + 0.013008669950067997, + 0.013628091663122177 + ], + "credit_map": [ + 0.07263806462287903, + 0.04473276063799858, + 0.04864158853888512, + 0.011634847149252892, + 0.03656341880559921, + 0.23307542502880096, + 0.01816723681986332, + 0.06088029965758324, + 0.035473790019750595, + 0.03316955640912056, + 0.02675996907055378, + 0.006590622011572123, + 0.015626877546310425, + 0.12557469308376312, + 0.010519646108150482, + 0.03142275661230087, + 0.009603952057659626, + 0.007727273739874363, + 0.018917199224233627, + 0.003675999352708459, + 0.0031296166125684977, + 0.01222547609359026, + 0.00330623472109437, + 0.006585516966879368, + 0.0018098409054800868, + 0.006439423654228449, + 0.00295997248031199, + 0.0008547635516151786, + 0.0010199231328442693, + 0.003952071536332369, + 0.0009591776179149747, + 0.0019887180533260107, + 0.003918711561709642, + 0.0019734809175133705, + 0.002837359206750989, + 0.0007155398488976061, + 0.001139747560955584, + 0.010781797580420971, + 0.0012214524904266, + 0.0030246837995946407, + 0.0062186866998672485, + 0.0018073651008307934, + 0.004110538866370916, + 0.0008037976222112775, + 0.0018251790897920728, + 0.015052800066769123, + 0.0016978851053863764, + 0.005317412782460451, + 0.0025264949072152376, + 0.0005500199622474611, + 0.0026951865293085575, + 0.00040347888716496527, + 0.00045829001464881003, + 0.0015356664080172777, + 0.0006625623209401965, + 0.0022412759717553854, + 0.0005994464154355228, + 0.00039286265382543206, + 0.0006498967995867133, + 0.00019935953605454415, + 0.00023119323304854333, + 0.0004431278503034264, + 0.0002333934826310724, + 0.00046790039050392807, + 0.0037385867908596992, + 0.0006766148726455867, + 0.0007889608386904001, + 0.00032258991268463433, + 0.0005390858859755099, + 0.001763186533935368, + 0.0004179620009381324, + 0.004761682823300362, + 0.00439986540004611, + 0.0010554672917351127, + 0.0023243469186127186, + 0.0005734590231440961, + 0.0005904264398850501, + 0.0013772343518212438, + 0.0007085348479449749, + 0.002598757157102227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.31345048546791077 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, static shot, shot from a slightly upward angle, focuses on a thin, middle-aged strategist with a neatly trimmed goatee, nestled within a dimly lit military tent. He wears a loose, flowing dark blue silk robe, his long hair held back with a simple wooden hairpin. He reclines leisurely on an antique carved wooden chair. The only illumination comes from a bronze oil lamp on a table beside him. Its flickering, warm yellow light casts deep, undulating shadows on his face, emphasizing his pensive expression and illuminating the folds of his silk robe with a soft sheen. In his right hand, he grasps a feather fan made of pure white crane feathers, gently waving it with a steady, leisurely rhythm, the breeze occasionally causing the lamplight to flicker. In the background, a military map hangs on the thick dark canvas tent wall, and a wooden weapons rack with unfolded bamboo slips and weapons can be vaguely seen next to it. The whole scene has a warm tone, mainly deep dark colors, creating an atmosphere of extreme concentration, tranquility and strategic planning against the backdrop of military tension.", + "seed": 1225, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_9/index_9345_seed_1225_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17502068281173705, + "vlm_logprob": -0.17502068281173705, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.17502068281173705, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04332028329372406 + ], + "opposite_logprob_tokens": [ + -3.2933201789855957 + ], + "logprob_reward": -0.04332028329372406, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04332028329372406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12793117761611938, + 0.11087916046380997, + 0.03110707737505436, + 0.11164059489965439, + 0.04018488526344299, + 0.024142613634467125, + 0.38324403762817383, + 0.03516046330332756, + 0.08735942840576172, + 0.04835059866309166 + ], + "credit_map": [ + 0.019680310040712357, + 0.01848662830889225, + 0.018042199313640594, + 0.015997815877199173, + 0.012857072986662388, + 0.012398422695696354, + 0.015080435201525688, + 0.015388287603855133, + 0.009925072081387043, + 0.020035987719893456, + 0.012393214739859104, + 0.012703576125204563, + 0.012483357451856136, + 0.017804663628339767, + 0.008511165156960487, + 0.017022114247083664, + 0.005889470223337412, + 0.0044357324950397015, + 0.0034326105378568172, + 0.0029699369333684444, + 0.00279307016171515, + 0.00432862713932991, + 0.00338913407176733, + 0.0038684913888573647, + 0.014120200648903847, + 0.009682164527475834, + 0.015671683475375175, + 0.009018182754516602, + 0.016546733677387238, + 0.02258222922682762, + 0.010198760777711868, + 0.013820636086165905, + 0.0056149037554860115, + 0.005466760136187077, + 0.003678127657622099, + 0.0035727873910218477, + 0.006239182781428099, + 0.005966543685644865, + 0.004323713015764952, + 0.005322862882167101, + 0.002590734278783202, + 0.0023898237850517035, + 0.002326050540432334, + 0.002970942296087742, + 0.0037432084791362286, + 0.0035679927095770836, + 0.003547631436958909, + 0.0030062305741012096, + 0.023935314267873764, + 0.01857439987361431, + 0.12432271987199783, + 0.016812963411211967, + 0.012232540175318718, + 0.1478818953037262, + 0.010360331274569035, + 0.029123861342668533, + 0.005013491027057171, + 0.003267740597948432, + 0.0034365965984761715, + 0.0031274580396711826, + 0.003468415467068553, + 0.005981456488370895, + 0.007005426567047834, + 0.003859876422211528, + 0.006356910336762667, + 0.007498251274228096, + 0.006162173114717007, + 0.010756959207355976, + 0.01147837471216917, + 0.01577465422451496, + 0.014426951296627522, + 0.014905155636370182, + 0.006742860656231642, + 0.004377786070108414, + 0.005261050537228584, + 0.004166832193732262, + 0.005866261664777994, + 0.008779932744801044, + 0.006073964759707451, + 0.007081907242536545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.04332028329372406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210578985512257 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.014210578985512257, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210578985512257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18923963606357574, + 0.0994274914264679, + 0.06385988742113113, + 0.09393027424812317, + 0.06031983345746994, + 0.04157751053571701, + 0.26689064502716064, + 0.06784289330244064, + 0.04148260876536369, + 0.07542926073074341 + ], + "credit_map": [ + 0.014331853948533535, + 0.039428289979696274, + 0.02739850990474224, + 0.029565030708909035, + 0.016358470544219017, + 0.01435514260083437, + 0.027684153988957405, + 0.020118189975619316, + 0.008590546436607838, + 0.017822979018092155, + 0.011449797078967094, + 0.008660230785608292, + 0.01163504645228386, + 0.016492361202836037, + 0.010571991093456745, + 0.014204539358615875, + 0.009516216814517975, + 0.00534588610753417, + 0.006294123828411102, + 0.0077512310817837715, + 0.007779653649777174, + 0.010073505342006683, + 0.010208702646195889, + 0.006890567950904369, + 0.00827853474766016, + 0.010437189601361752, + 0.01421199832111597, + 0.005231485236436129, + 0.017503369599580765, + 0.018738996237516403, + 0.009718077257275581, + 0.009810629300773144, + 0.007940633222460747, + 0.007952762767672539, + 0.006167688872665167, + 0.006694068666547537, + 0.006268756929785013, + 0.010144765488803387, + 0.007876281626522541, + 0.007274874020367861, + 0.0032921696547418833, + 0.0032396623864769936, + 0.002922190586104989, + 0.006377601996064186, + 0.0032306425273418427, + 0.007603430189192295, + 0.005714788567274809, + 0.009197021834552288, + 0.017789343371987343, + 0.013748519122600555, + 0.08361977338790894, + 0.01213561836630106, + 0.010280259884893894, + 0.09852264821529388, + 0.010219985619187355, + 0.020574504509568214, + 0.012606549076735973, + 0.007903527468442917, + 0.006597617641091347, + 0.008842791430652142, + 0.00663694366812706, + 0.008955082856118679, + 0.008973237127065659, + 0.007327142637223005, + 0.003890544641762972, + 0.00413354579359293, + 0.0032055021729320288, + 0.005061738658696413, + 0.004910435527563095, + 0.007251773960888386, + 0.0069837188348174095, + 0.006045350339263678, + 0.005289129447191954, + 0.006334403995424509, + 0.0049853879027068615, + 0.009779834188520908, + 0.01575242541730404, + 0.012994163669645786, + 0.01016377191990614, + 0.010130144655704498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.014210578985512257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107992883771658 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.01107992883771658, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107992883771658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10726077109575272, + 0.1821199506521225, + 0.11463054269552231, + 0.11120399832725525, + 0.033892273902893066, + 0.0288272462785244, + 0.259211003780365, + 0.040055178105831146, + 0.06191029027104378, + 0.060888681560754776 + ], + "credit_map": [ + 0.01270921342074871, + 0.017810309305787086, + 0.01587194949388504, + 0.018112387508153915, + 0.01033326331526041, + 0.010134673677384853, + 0.01276723574846983, + 0.009521754458546638, + 0.018537573516368866, + 0.04175408184528351, + 0.02638212963938713, + 0.01772480644285679, + 0.018964089453220367, + 0.030739961192011833, + 0.013285310938954353, + 0.014732008799910545, + 0.014782794751226902, + 0.028449401259422302, + 0.01259060762822628, + 0.012454992160201073, + 0.008448431268334389, + 0.018948225304484367, + 0.008413231000304222, + 0.010542874224483967, + 0.008784960955381393, + 0.010930323041975498, + 0.019456591457128525, + 0.009035611525177956, + 0.016156502068042755, + 0.022978294640779495, + 0.01058711763471365, + 0.013274605385959148, + 0.005816212855279446, + 0.0040373667143285275, + 0.0033368768636137247, + 0.0034350419882684946, + 0.004021223168820143, + 0.00543236406520009, + 0.0034701144322752953, + 0.004343078471720219, + 0.0034132509026676416, + 0.005141880363225937, + 0.0029641950968652964, + 0.0033195202704519033, + 0.0026413185987621546, + 0.004828726872801781, + 0.0035699186846613884, + 0.0029484385158866644, + 0.016387658193707466, + 0.014224213548004627, + 0.08098181337118149, + 0.011167858727276325, + 0.008784073404967785, + 0.10015188157558441, + 0.007597103249281645, + 0.019916433840990067, + 0.0046288445591926575, + 0.004975425079464912, + 0.005058481357991695, + 0.00484180822968483, + 0.004033830016851425, + 0.008625583723187447, + 0.0033101800363510847, + 0.004581029526889324, + 0.008095907978713512, + 0.009259549900889397, + 0.008013773709535599, + 0.008632340468466282, + 0.0051590446382761, + 0.010347522795200348, + 0.006233546417206526, + 0.006168609485030174, + 0.010216427966952324, + 0.005318209063261747, + 0.006419997662305832, + 0.005286716390401125, + 0.006521007977426052, + 0.011125554330646992, + 0.007069178391247988, + 0.008931593969464302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.01107992883771658 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864978417754173 + ], + "opposite_logprob_tokens": [ + -3.048649787902832 + ], + "logprob_reward": -0.04864978417754173, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864978417754173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13443244993686676, + 0.039423905313014984, + 0.030746567994356155, + 0.2624453604221344, + 0.046287499368190765, + 0.050510454922914505, + 0.1483057588338852, + 0.22289018332958221, + 0.031092308461666107, + 0.03386550024151802 + ], + "credit_map": [ + 0.015572702512145042, + 0.02581300586462021, + 0.021328197792172432, + 0.02345939725637436, + 0.011783439666032791, + 0.011395762674510479, + 0.01431899145245552, + 0.0107609573751688, + 0.0041407872922718525, + 0.004166465252637863, + 0.00320475734770298, + 0.004985434468835592, + 0.006375757046043873, + 0.006008465774357319, + 0.004871771205216646, + 0.005670466925948858, + 0.004851421341300011, + 0.0027753806207329035, + 0.0029818471521139145, + 0.0034217967186123133, + 0.005152091383934021, + 0.00458860257640481, + 0.004054609220474958, + 0.0029208173509687185, + 0.027071479707956314, + 0.02902683988213539, + 0.040600311011075974, + 0.02022477053105831, + 0.03491608053445816, + 0.051602642983198166, + 0.025560058653354645, + 0.03344316780567169, + 0.00836490374058485, + 0.004879079293459654, + 0.004002769943326712, + 0.006091982126235962, + 0.0045068333856761456, + 0.007316471543163061, + 0.004818927962332964, + 0.006306531839072704, + 0.008174898102879524, + 0.00827882532030344, + 0.004127101041376591, + 0.006283157970756292, + 0.005682904273271561, + 0.007180496584624052, + 0.005050384905189276, + 0.005732686724513769, + 0.01197119615972042, + 0.011813637800514698, + 0.038514696061611176, + 0.0074962470680475235, + 0.00931570678949356, + 0.05139416828751564, + 0.006695778574794531, + 0.011104324832558632, + 0.04634407162666321, + 0.021804265677928925, + 0.028309568762779236, + 0.030474016442894936, + 0.018668195232748985, + 0.03122570365667343, + 0.02100359834730625, + 0.025060761719942093, + 0.005705166142433882, + 0.0020388357806950808, + 0.0037143363151699305, + 0.0035673463717103004, + 0.0030608500819653273, + 0.005516755860298872, + 0.003351879771798849, + 0.0041371360421180725, + 0.0019496432505548, + 0.0035990739706903696, + 0.002295280108228326, + 0.003785171778872609, + 0.0051805623807013035, + 0.006231155712157488, + 0.0054521700367331505, + 0.005372443702071905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.04864978417754173 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578428387641907 + ], + "opposite_logprob_tokens": [ + -0.6328428387641907 + ], + "logprob_reward": -0.7578428387641907, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578428387641907, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09487654268741608, + 0.18795578181743622, + 0.06412718445062637, + 0.0673220306634903, + 0.1016872227191925, + 0.029251042753458023, + 0.2440194934606552, + 0.0787723958492279, + 0.0631600171327591, + 0.06882831454277039 + ], + "credit_map": [ + 0.012023565359413624, + 0.012857625260949135, + 0.012753968127071857, + 0.012421727180480957, + 0.010646023787558079, + 0.008732017129659653, + 0.01459574792534113, + 0.010845869779586792, + 0.018525755032896996, + 0.044998835772275925, + 0.02203388325870037, + 0.02273123897612095, + 0.016688209027051926, + 0.03135792165994644, + 0.017689429223537445, + 0.013930492103099823, + 0.007457931060343981, + 0.014322194270789623, + 0.007957182824611664, + 0.0063048917800188065, + 0.005399773363023996, + 0.011561309918761253, + 0.004934802185744047, + 0.00618909252807498, + 0.015141531825065613, + 0.0078045763075351715, + 0.010854306630790234, + 0.006788153667002916, + 0.004963281098753214, + 0.00705541530624032, + 0.0057179625146090984, + 0.008996803313493729, + 0.017175205051898956, + 0.021356675773859024, + 0.013196002691984177, + 0.010091770440340042, + 0.008802876807749271, + 0.01671430468559265, + 0.006190056446939707, + 0.00816032849252224, + 0.006717733107507229, + 0.004198156762868166, + 0.003807632951065898, + 0.002996152499690652, + 0.0018113821279257536, + 0.0048154923133552074, + 0.0026501775719225407, + 0.002254313323646784, + 0.014241973869502544, + 0.014170239679515362, + 0.07634995132684708, + 0.010954555124044418, + 0.006606126204133034, + 0.09603479504585266, + 0.006102730520069599, + 0.0195591039955616, + 0.018816780298948288, + 0.0061815096996724606, + 0.009071140550076962, + 0.010515379719436169, + 0.00666537182405591, + 0.010504193603992462, + 0.008098331280052662, + 0.008919688872992992, + 0.007213455159217119, + 0.004917789716273546, + 0.0071970927529037, + 0.007181020453572273, + 0.009443005546927452, + 0.011318187229335308, + 0.0074850101955235004, + 0.00840445701032877, + 0.01263889018446207, + 0.01121197734028101, + 0.010842190124094486, + 0.0065033650025725365, + 0.004500044044107199, + 0.01182196568697691, + 0.0041686841286718845, + 0.007141197100281715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.7578428387641907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "medium, static shot, shot from a slightly upward angle, focuses on a thin, middle-aged strategist with a neatly trimmed goatee, nestled within a dimly lit military tent. He wears a loose, flowing dark blue silk robe, his long hair held back with a simple wooden hairpin. He reclines leisurely on an antique carved wooden chair. The only illumination comes from a bronze oil lamp on a table beside him. Its flickering, warm yellow light casts deep, undulating shadows on his face, emphasizing his pensive expression and illuminating the folds of his silk robe with a soft sheen. In his right hand, he grasps a feather fan made of pure white crane feathers, gently waving it with a steady, leisurely rhythm, the breeze occasionally causing the lamplight to flicker. In the background, a military map hangs on the thick dark canvas tent wall, and a wooden weapons rack with unfolded bamboo slips and weapons can be vaguely seen next to it. The whole scene has a warm tone, mainly deep dark colors, creating an atmosphere of extreme concentration, tranquility and strategic planning against the backdrop of military tension.", + "seed": 1226, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000117/group_9/index_9345_seed_1226_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.29546896647661924, + "vlm_logprob": -0.29546896647661924, + "vlm_yes_no_margin": 2.400000011920929, + "vlm_token_credit": -0.29546896647661924, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06019279360771179 + ], + "opposite_logprob_tokens": [ + -2.935192823410034 + ], + "logprob_reward": -0.06019279360771179, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06019279360771179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17895379662513733, + 0.05366693437099457, + 0.10755354166030884, + 0.11937786638736725, + 0.11741078644990921, + 0.10490217059850693, + 0.0988522320985794, + 0.10823836922645569, + 0.020869866013526917, + 0.09017439186573029 + ], + "credit_map": [ + 0.03040214255452156, + 0.03416045382618904, + 0.025584084913134575, + 0.02258160524070263, + 0.01432257890701294, + 0.025191694498062134, + 0.012010529637336731, + 0.0147006930783391, + 0.007454182021319866, + 0.008655001409351826, + 0.0050994656048715115, + 0.007538828533142805, + 0.007376619148999453, + 0.007187969982624054, + 0.005380347836762667, + 0.004974518436938524, + 0.012242544442415237, + 0.014153285883367062, + 0.014100683853030205, + 0.011026601307094097, + 0.018992507830262184, + 0.015332876704633236, + 0.008742311969399452, + 0.012962727807462215, + 0.015993382781744003, + 0.013058663345873356, + 0.019913773983716965, + 0.01746257022023201, + 0.011097059585154057, + 0.016639284789562225, + 0.010701564140617847, + 0.01451156660914421, + 0.0085481321439147, + 0.01384406816214323, + 0.011788738891482353, + 0.01337722223252058, + 0.01785348542034626, + 0.019651250913739204, + 0.018849719315767288, + 0.013498182408511639, + 0.014249597676098347, + 0.00974930077791214, + 0.009092860855162144, + 0.015859639272093773, + 0.012965550646185875, + 0.015606582164764404, + 0.009193561971187592, + 0.018185071647167206, + 0.013015251606702805, + 0.010944831185042858, + 0.009606403298676014, + 0.014703083783388138, + 0.014636124484241009, + 0.016513770446181297, + 0.006578847300261259, + 0.01285392977297306, + 0.012486709281802177, + 0.010498620569705963, + 0.012713395059108734, + 0.01826741173863411, + 0.012869241647422314, + 0.015380654484033585, + 0.00803951546549797, + 0.017982827499508858, + 0.0029324116185307503, + 0.002612752141430974, + 0.002669948386028409, + 0.0025092707946896553, + 0.002627452602609992, + 0.0029206443578004837, + 0.001688280957750976, + 0.0029091041069477797, + 0.012544949539005756, + 0.008005276322364807, + 0.011641362681984901, + 0.009719040244817734, + 0.00968589261174202, + 0.012515990994870663, + 0.010489393025636673, + 0.015572487376630306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.06019279360771179 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012558662332594395 + ], + "opposite_logprob_tokens": [ + -4.387558460235596 + ], + "logprob_reward": -0.012558662332594395, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558662332594395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14945369958877563, + 0.0985836386680603, + 0.06929684430360794, + 0.19343920052051544, + 0.14185042679309845, + 0.06445135921239853, + 0.10572497546672821, + 0.06816112250089645, + 0.032629404217004776, + 0.07640931755304337 + ], + "credit_map": [ + 0.021988799795508385, + 0.01844455674290657, + 0.01464320532977581, + 0.022822465747594833, + 0.0156160993501544, + 0.02208966761827469, + 0.013084406033158302, + 0.020764505490660667, + 0.009647568687796593, + 0.012712966650724411, + 0.00945417582988739, + 0.0140306381508708, + 0.01262398436665535, + 0.014715665020048618, + 0.009762706235051155, + 0.01563592441380024, + 0.01253694761544466, + 0.007219168357551098, + 0.00865250639617443, + 0.008925245143473148, + 0.008216479793190956, + 0.01004992425441742, + 0.0050759511068463326, + 0.00862062070518732, + 0.026123497635126114, + 0.022151915356516838, + 0.01816990040242672, + 0.026674583554267883, + 0.0264880433678627, + 0.027164064347743988, + 0.019543617963790894, + 0.027123576030135155, + 0.010525781661272049, + 0.01815660297870636, + 0.015096621587872505, + 0.017365746200084686, + 0.022466866299510002, + 0.023063767701387405, + 0.017887331545352936, + 0.017287708818912506, + 0.008798777125775814, + 0.006544108502566814, + 0.010347561910748482, + 0.008843506686389446, + 0.0065986597910523415, + 0.009961416013538837, + 0.004794476553797722, + 0.008562851697206497, + 0.009596654213964939, + 0.012308484874665737, + 0.010363094508647919, + 0.017546918243169785, + 0.016016582027077675, + 0.0197767224162817, + 0.006532405968755484, + 0.013584116473793983, + 0.008340866304934025, + 0.006929546128958464, + 0.007479438558220863, + 0.01140341255813837, + 0.008248877711594105, + 0.00973887462168932, + 0.005088163539767265, + 0.010931942611932755, + 0.003091977909207344, + 0.0030226006638258696, + 0.0038639253471046686, + 0.0037728992756456137, + 0.0034568547271192074, + 0.0056411270052194595, + 0.00346588809043169, + 0.006314132362604141, + 0.012848683632910252, + 0.006649905815720558, + 0.010607810690999031, + 0.009035637602210045, + 0.008935484103858471, + 0.010424574837088585, + 0.007276289630681276, + 0.010630936361849308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.012558662332594395 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589781925082207 + ], + "opposite_logprob_tokens": [ + -3.895589828491211 + ], + "logprob_reward": -0.020589781925082207, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589781925082207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18090537190437317, + 0.03348354622721672, + 0.06985490024089813, + 0.12579810619354248, + 0.08438609540462494, + 0.2652228772640228, + 0.14094901084899902, + 0.011133291758596897, + 0.02342221513390541, + 0.06484457850456238 + ], + "credit_map": [ + 0.029568342491984367, + 0.026151252910494804, + 0.02203678898513317, + 0.023982034996151924, + 0.019117115065455437, + 0.024412812665104866, + 0.01733635738492012, + 0.018300674855709076, + 0.0037984768860042095, + 0.004684996325522661, + 0.0036295060999691486, + 0.005151564255356789, + 0.0034681737888604403, + 0.004890083335340023, + 0.0028290702030062675, + 0.005031677894294262, + 0.008424398489296436, + 0.007889514788985252, + 0.009719444438815117, + 0.009315221570432186, + 0.009065544232726097, + 0.00961408019065857, + 0.005831532645970583, + 0.009995165280997753, + 0.019260717555880547, + 0.014447401277720928, + 0.02531455084681511, + 0.014182968996465206, + 0.013349256478250027, + 0.017516613006591797, + 0.008971422910690308, + 0.012755172327160835, + 0.007635811343789101, + 0.012044727802276611, + 0.008550679311156273, + 0.011455927975475788, + 0.010942446999251842, + 0.012458132579922676, + 0.012539027258753777, + 0.008759339340031147, + 0.03461233898997307, + 0.024074727669358253, + 0.0419178381562233, + 0.040975432842969894, + 0.019796399399638176, + 0.04465804621577263, + 0.017897039651870728, + 0.04129106178879738, + 0.018111487850546837, + 0.015443997457623482, + 0.02343812957406044, + 0.017337001860141754, + 0.018630564212799072, + 0.02236836403608322, + 0.008183405734598637, + 0.017436053603887558, + 0.001121143577620387, + 0.0013189261080697179, + 0.0013908070977777243, + 0.0015198905020952225, + 0.001655236235819757, + 0.0015356414951384068, + 0.001206012675538659, + 0.0013856334844604135, + 0.0027028019540011883, + 0.002804323099553585, + 0.002436046488583088, + 0.0028831965755671263, + 0.0019789596553891897, + 0.004041401203721762, + 0.0020798994228243828, + 0.004495585337281227, + 0.010304688476026058, + 0.006300922948867083, + 0.006264970637857914, + 0.01018770132213831, + 0.009588364511728287, + 0.009144006296992302, + 0.007151763886213303, + 0.005902156233787537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.020589781925082207 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037459671497345 + ], + "opposite_logprob_tokens": [ + -1.910374641418457 + ], + "logprob_reward": -0.16037459671497345, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037459671497345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1774459034204483, + 0.027274239808321, + 0.035550110042095184, + 0.07920629531145096, + 0.20260727405548096, + 0.09080850332975388, + 0.05947455018758774, + 0.17964418232440948, + 0.04163524881005287, + 0.10635367780923843 + ], + "credit_map": [ + 0.021158624440431595, + 0.024026187136769295, + 0.02433927170932293, + 0.022806406021118164, + 0.024947015568614006, + 0.025256680324673653, + 0.01944378763437271, + 0.015467917546629906, + 0.003414708888158202, + 0.004321759566664696, + 0.0032760901376605034, + 0.003623228520154953, + 0.0035193697549402714, + 0.004089350812137127, + 0.0021602422930300236, + 0.002869491232559085, + 0.004678741563111544, + 0.005174796096980572, + 0.005552016664296389, + 0.0034698909148573875, + 0.005125119350850582, + 0.004565912764519453, + 0.0024041959550231695, + 0.004579437430948019, + 0.01102515123784542, + 0.008548278361558914, + 0.007455634884536266, + 0.013495957478880882, + 0.009635563008487225, + 0.010583111084997654, + 0.008070358075201511, + 0.010392236523330212, + 0.015979580581188202, + 0.027406111359596252, + 0.017965950071811676, + 0.022143792361021042, + 0.030631881207227707, + 0.032422419637441635, + 0.033399779349565506, + 0.022657746449112892, + 0.015673665329813957, + 0.009959474205970764, + 0.012432410381734371, + 0.013526495546102524, + 0.008293240331113338, + 0.012842233292758465, + 0.006530942395329475, + 0.011550042778253555, + 0.00657928641885519, + 0.005271465051919222, + 0.009504313580691814, + 0.00769925769418478, + 0.008581207133829594, + 0.010438438504934311, + 0.0034790830686688423, + 0.007921497337520123, + 0.02056829258799553, + 0.01729150302708149, + 0.02563738077878952, + 0.02790762111544609, + 0.024785146117210388, + 0.023582499474287033, + 0.013751454651355743, + 0.02612028457224369, + 0.0051628602668643, + 0.004045410547405481, + 0.004849358461797237, + 0.006202271208167076, + 0.00523242587223649, + 0.006545042619109154, + 0.003861322533339262, + 0.005736560095101595, + 0.009369262494146824, + 0.011488266289234161, + 0.012443603947758675, + 0.014679606072604656, + 0.0171149130910635, + 0.015338757075369358, + 0.01282432395964861, + 0.013094945810735226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.16037459671497345 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed goatee and long hair held back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, flowing dark blue silk robe and reclining on an antique carved wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated primarily by a bronze oil lamp casting warm yellow light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a white crane feather fan in his right hand and waving it with a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background set inside a dimly lit military tent with a map on the wall and a weapons rack visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236289978027344 + ], + "opposite_logprob_tokens": [ + -0.34862905740737915 + ], + "logprob_reward": -1.2236289978027344, + "yes_no_margin_reward": -0.8749999403953552, + "token_credit_reward": -1.2236289978027344, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18654406070709229, + 0.06394632905721664, + 0.12550270557403564, + 0.1296698898077011, + 0.1129375770688057, + 0.10924692451953888, + 0.14049606025218964, + 0.04463348910212517, + 0.01984386332333088, + 0.06717909872531891 + ], + "credit_map": [ + 0.020360508933663368, + 0.02281217649579048, + 0.024974532425403595, + 0.02697562985122204, + 0.020226482301950455, + 0.029956264421343803, + 0.019162675365805626, + 0.02207581326365471, + 0.008232680149376392, + 0.009519296698272228, + 0.008041569031774998, + 0.007851241156458855, + 0.008617240004241467, + 0.01039536576718092, + 0.004632263444364071, + 0.006656683515757322, + 0.020529014989733696, + 0.01401677168905735, + 0.016017168760299683, + 0.017504923045635223, + 0.014732650481164455, + 0.017189987003803253, + 0.010198711417615414, + 0.015313487499952316, + 0.018532508984208107, + 0.013510802760720253, + 0.01728624850511551, + 0.01752789132297039, + 0.017210256308317184, + 0.017209984362125397, + 0.013077082112431526, + 0.015315121039748192, + 0.009832913056015968, + 0.012615716084837914, + 0.012062449008226395, + 0.011977512389421463, + 0.016861610114574432, + 0.01824386604130268, + 0.018057262524962425, + 0.013286249712109566, + 0.018027909100055695, + 0.010480690747499466, + 0.011912520974874496, + 0.018455948680639267, + 0.009335777722299099, + 0.017907695844769478, + 0.007272631395608187, + 0.015853755176067352, + 0.011713670566678047, + 0.017097368836402893, + 0.012966305017471313, + 0.02059231325984001, + 0.021539393812417984, + 0.027461176738142967, + 0.0089495200663805, + 0.020176339894533157, + 0.004737378563731909, + 0.00507001718506217, + 0.004915135912597179, + 0.008479875512421131, + 0.005474521778523922, + 0.0066547691822052, + 0.0039484621956944466, + 0.0053533343598246574, + 0.002357248682528734, + 0.0017687858780846, + 0.002046796726062894, + 0.0031020918395370245, + 0.0030442848801612854, + 0.002919950755313039, + 0.0018581937765702605, + 0.002746512647718191, + 0.009136856533586979, + 0.008312867023050785, + 0.007355523761361837, + 0.012576334178447723, + 0.009124552831053734, + 0.009488839656114578, + 0.005611334461718798, + 0.005572793073952198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -1.2236289978027344 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_1/index_tensor([2899, 2899])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_1/index_tensor([2899, 2899])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fcf37c7b5bd083b14a762a30deca5ac956b0296d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_1/index_tensor([2899, 2899])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide shot, the camera tracks steadily sideways, maintaining a constant distance from a single wooden crate as it moves through a vast, cavernous, empty warehouse. Cool, ethereal light filters through a high, grimy skylight, filtering through the dusty air and casting long, twisting shadows across the polished but grimy concrete floor. Composed of dark, weathered wood planks held together by rusted iron sheets, the crate is the only object in the vast space. Propelled relentlessly and silently across the smooth surface, its movement unsettlingly fluid and steady, driven by an entirely unseen force. The camera's parallel movement emphasizes the crate's frictionless gliding, creating an atmosphere of tension, mystery, and deep unease within the cold, monochrome setting.", + "seed": 6159, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_1/index_2899_seed_6159_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017484007216989995, + "vlm_logprob": -0.017484007216989995, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.017484007216989995, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010363809764385223 + ], + "opposite_logprob_tokens": [ + -5.010363578796387 + ], + "logprob_reward": -0.010363809764385223, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010363809764385223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1484042853116989, + 0.12919503450393677, + 0.11808351427316666, + 0.07341597229242325, + 0.07798754423856735, + 0.07785546779632568, + 0.11151432245969772, + 0.18927845358848572, + 0.03353342413902283, + 0.04073197394609451 + ], + "credit_map": [ + 0.016563331708312035, + 0.02101026102900505, + 0.02131889946758747, + 0.024478768929839134, + 0.016870003193616867, + 0.01455303467810154, + 0.02126811258494854, + 0.012341869063675404, + 0.01754434034228325, + 0.009317794814705849, + 0.027705363929271698, + 0.026368608698248863, + 0.009934918023645878, + 0.009445074945688248, + 0.015102848410606384, + 0.01377608347684145, + 0.012949301861226559, + 0.016146069392561913, + 0.015594563446938992, + 0.013150522485375404, + 0.02629639022052288, + 0.00967014953494072, + 0.01600549928843975, + 0.008271022699773312, + 0.00733756460249424, + 0.006487972103059292, + 0.011470825411379337, + 0.011600903235375881, + 0.007047575898468494, + 0.0071049281395971775, + 0.012740791775286198, + 0.009625410661101341, + 0.011385725811123848, + 0.005833026021718979, + 0.010467009618878365, + 0.013184326700866222, + 0.007550171576440334, + 0.007012438494712114, + 0.009268210269510746, + 0.013286630623042583, + 0.007540255319327116, + 0.007831879891455173, + 0.011116084642708302, + 0.011913280002772808, + 0.009213893674314022, + 0.006898493506014347, + 0.0113510238006711, + 0.011990553699433804, + 0.00897892750799656, + 0.006945739965885878, + 0.015409777872264385, + 0.02617597207427025, + 0.0071775829419493675, + 0.016452137380838394, + 0.018929073587059975, + 0.011445114389061928, + 0.0214068740606308, + 0.016640616580843925, + 0.037965018302202225, + 0.01888568326830864, + 0.02886980213224888, + 0.022948771715164185, + 0.025156361982226372, + 0.017405331134796143, + 0.0035298820585012436, + 0.003080241149291396, + 0.006960478611290455, + 0.002848235424607992, + 0.003528309054672718, + 0.0038716010749340057, + 0.006157427094876766, + 0.0035572522319853306, + 0.005320086609572172, + 0.004853506572544575, + 0.005310944747179747, + 0.005898077506572008, + 0.003282588440924883, + 0.0057211085222661495, + 0.004991818685084581, + 0.005353845655918121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.010363809764385223 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026395773515105247 + ], + "opposite_logprob_tokens": [ + -3.651395797729492 + ], + "logprob_reward": -0.026395773515105247, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026395773515105247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12785246968269348, + 0.08324083685874939, + 0.08022107928991318, + 0.06901433318853378, + 0.10894937813282013, + 0.11553963273763657, + 0.09718609601259232, + 0.2598109245300293, + 0.016101235523819923, + 0.0420839823782444 + ], + "credit_map": [ + 0.0209296066313982, + 0.01654447428882122, + 0.020658642053604126, + 0.0145603958517313, + 0.01184329017996788, + 0.009825161658227444, + 0.016533387824892998, + 0.016957523301243782, + 0.007509914226830006, + 0.005485809873789549, + 0.017802922055125237, + 0.02252872660756111, + 0.004723701626062393, + 0.00542205385863781, + 0.00810819212347269, + 0.01165953278541565, + 0.010836015455424786, + 0.008939981460571289, + 0.012000764720141888, + 0.010800744406878948, + 0.009464392438530922, + 0.007856810465455055, + 0.011431312188506126, + 0.008891062811017036, + 0.006738310679793358, + 0.007227959111332893, + 0.010442860424518585, + 0.010918986983597279, + 0.007446580100804567, + 0.007619147188961506, + 0.008771847933530807, + 0.009848649613559246, + 0.01363485585898161, + 0.007470628246665001, + 0.016401248052716255, + 0.017828552052378654, + 0.009840050712227821, + 0.010770726948976517, + 0.012911243364214897, + 0.020092086866497993, + 0.010340360924601555, + 0.010461819358170033, + 0.017615823075175285, + 0.016982143744826317, + 0.013936544768512249, + 0.010590063408017159, + 0.017328746616840363, + 0.018284140154719353, + 0.0076680658385157585, + 0.009491398930549622, + 0.011193870566785336, + 0.022598721086978912, + 0.008470525965094566, + 0.013697165995836258, + 0.01525900512933731, + 0.008807361125946045, + 0.030941370874643326, + 0.025102507323026657, + 0.049005355685949326, + 0.02609824202954769, + 0.03956259787082672, + 0.03295883908867836, + 0.03299471363425255, + 0.023147299885749817, + 0.0009744833223521709, + 0.0010404543718323112, + 0.0028328534681349993, + 0.0014882279792800546, + 0.002319919178262353, + 0.0026696347631514072, + 0.002638867823407054, + 0.00213679694570601, + 0.004436757415533066, + 0.0040583061054348946, + 0.004409183748066425, + 0.005703780800104141, + 0.0035975677892565727, + 0.0065908730030059814, + 0.006252915132790804, + 0.007034600712358952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.026395773515105247 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086412705481052 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086412705481052, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086412705481052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10902806371450424, + 0.07372945547103882, + 0.14360421895980835, + 0.06588400155305862, + 0.10612578690052032, + 0.10643631964921951, + 0.10584572702646255, + 0.26909971237182617, + 0.00813289638608694, + 0.012113785371184349 + ], + "credit_map": [ + 0.018353847786784172, + 0.008327017538249493, + 0.015753407031297684, + 0.01618056371808052, + 0.009406954050064087, + 0.007395060732960701, + 0.01371737290173769, + 0.019893836230039597, + 0.007261484861373901, + 0.00632883794605732, + 0.011440606787800789, + 0.009155233390629292, + 0.009420448914170265, + 0.007571735884994268, + 0.01587621681392193, + 0.0066748932003974915, + 0.019669184461236, + 0.02307378500699997, + 0.02188166230916977, + 0.01798386313021183, + 0.018064381554722786, + 0.013105658814311028, + 0.014069294556975365, + 0.0157563965767622, + 0.006370931398123503, + 0.007971981540322304, + 0.009403424337506294, + 0.008253684267401695, + 0.008635768666863441, + 0.008416369557380676, + 0.009408937767148018, + 0.007422903086990118, + 0.013973912224173546, + 0.007242370862513781, + 0.01496431790292263, + 0.02023506723344326, + 0.008713574148714542, + 0.009020381607115269, + 0.013799872249364853, + 0.018176289275288582, + 0.009276608936488628, + 0.009640779346227646, + 0.016202177852392197, + 0.016079843044281006, + 0.013392128981649876, + 0.009703031741082668, + 0.015967916697263718, + 0.016173826530575752, + 0.008287249132990837, + 0.009323175996541977, + 0.012433310970664024, + 0.024829620495438576, + 0.008820539340376854, + 0.01577398180961609, + 0.01661228947341442, + 0.009765553288161755, + 0.03136951103806496, + 0.024913497269153595, + 0.048439767211675644, + 0.026374876499176025, + 0.04604669287800789, + 0.03281316161155701, + 0.03448139503598213, + 0.024660810828208923, + 0.0009375534136779606, + 0.000556402956135571, + 0.001135380589403212, + 0.0010657600359991193, + 0.0010019650217145681, + 0.0011052675545215607, + 0.001130404183641076, + 0.0012001628056168556, + 0.0020173878874629736, + 0.0010881469352170825, + 0.0016234047943726182, + 0.0012175575830042362, + 0.0010087552946060896, + 0.0017958013340830803, + 0.0017251592362299562, + 0.0016375720733776689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.011086412705481052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793577268719673 + ], + "opposite_logprob_tokens": [ + -3.5297935009002686 + ], + "logprob_reward": -0.029793577268719673, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793577268719673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08850344270467758, + 0.08195909857749939, + 0.05588224530220032, + 0.06447721272706985, + 0.06200186908245087, + 0.07256465405225754, + 0.29980385303497314, + 0.19355560839176178, + 0.06605977565050125, + 0.015192277729511261 + ], + "credit_map": [ + 0.012067173607647419, + 0.008283374831080437, + 0.012111012823879719, + 0.011322073638439178, + 0.008601277135312557, + 0.006810573861002922, + 0.012058928608894348, + 0.017249036580324173, + 0.0058634462766349316, + 0.004900884814560413, + 0.014610799960792065, + 0.021846437826752663, + 0.007324332371354103, + 0.005953967105597258, + 0.01238358486443758, + 0.009075638838112354, + 0.007234120741486549, + 0.006205013487488031, + 0.008949399925768375, + 0.008022831752896309, + 0.00686284527182579, + 0.005015858449041843, + 0.00837795902043581, + 0.005214213859289885, + 0.005456530023366213, + 0.008042613044381142, + 0.00836153794080019, + 0.006444679573178291, + 0.009924064390361309, + 0.01055510900914669, + 0.008325347676873207, + 0.007367328740656376, + 0.008390900678932667, + 0.003986608702689409, + 0.00894836150109768, + 0.011155933141708374, + 0.004757117014378309, + 0.005534172058105469, + 0.008252940140664577, + 0.010975833050906658, + 0.0064290170557796955, + 0.006712350528687239, + 0.011118076741695404, + 0.011905315332114697, + 0.00846212450414896, + 0.006598829757422209, + 0.01069504115730524, + 0.010643897578120232, + 0.012971681542694569, + 0.01328213606029749, + 0.04972238466143608, + 0.06572094559669495, + 0.024397188797593117, + 0.05656888708472252, + 0.05342120677232742, + 0.02371942438185215, + 0.022225651890039444, + 0.01726151816546917, + 0.03547203168272972, + 0.01963922753930092, + 0.03268246725201607, + 0.023596275597810745, + 0.025028133764863014, + 0.01765030436217785, + 0.00442242668941617, + 0.0033217009622603655, + 0.014345810748636723, + 0.005315243732184172, + 0.00867694802582264, + 0.007889571599662304, + 0.013764990493655205, + 0.008323082700371742, + 0.0013944635866209865, + 0.000883361033629626, + 0.0035065398551523685, + 0.0015853974036872387, + 0.0011456144275143743, + 0.002380798803642392, + 0.0021216205786913633, + 0.002174482448026538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.029793577268719673 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780462831258774 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780462831258774, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780462831258774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10135628283023834, + 0.0734485387802124, + 0.06466381996870041, + 0.07073050737380981, + 0.09232928603887558, + 0.12376284599304199, + 0.1389474719762802, + 0.2860603332519531, + 0.026191115379333496, + 0.022509759292006493 + ], + "credit_map": [ + 0.0170858483761549, + 0.010009015910327435, + 0.017793715000152588, + 0.012778027914464474, + 0.009122188203036785, + 0.00835972186177969, + 0.011567499488592148, + 0.014640269801020622, + 0.006531610153615475, + 0.0049150800332427025, + 0.01332471426576376, + 0.014984876848757267, + 0.006488756276667118, + 0.005477321334183216, + 0.012171780690550804, + 0.009554401971399784, + 0.008650976233184338, + 0.007907569408416748, + 0.009513472206890583, + 0.00945678073912859, + 0.007505101151764393, + 0.005718471482396126, + 0.009326942265033722, + 0.006584506947547197, + 0.006028236821293831, + 0.007233345881104469, + 0.0106426440179348, + 0.009639154188334942, + 0.008063049986958504, + 0.009565407410264015, + 0.010059699416160583, + 0.009498966857790947, + 0.012545079924166203, + 0.006270696874707937, + 0.013595347292721272, + 0.016185082495212555, + 0.007851505652070045, + 0.0075753736309707165, + 0.010679593309760094, + 0.01762661337852478, + 0.011023919098079205, + 0.011129094287753105, + 0.018270323053002357, + 0.018841542303562164, + 0.015513352118432522, + 0.011350555345416069, + 0.01873154006898403, + 0.018902530893683434, + 0.00792631134390831, + 0.009902648627758026, + 0.022877659648656845, + 0.03021269291639328, + 0.018609678372740746, + 0.01637130044400692, + 0.02309487946331501, + 0.009952318854629993, + 0.03225259855389595, + 0.026159297674894333, + 0.050851669162511826, + 0.028060900047421455, + 0.04999465495347977, + 0.035189438611269, + 0.03685629367828369, + 0.026695484295487404, + 0.001678827335126698, + 0.0014167657354846597, + 0.005411940161138773, + 0.002325852168723941, + 0.003493195865303278, + 0.003211722709238529, + 0.005335901398211718, + 0.0033169130329042673, + 0.003126078750938177, + 0.002850324846804142, + 0.0037652747705578804, + 0.0023051933385431767, + 0.002076944801956415, + 0.0026902786921709776, + 0.0029236464761197567, + 0.0027720192447304726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.009780462831258774 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot, the camera tracks steadily sideways, maintaining a constant distance from a single wooden crate as it moves through a vast, cavernous, empty warehouse. Cool, ethereal light filters through a high, grimy skylight, filtering through the dusty air and casting long, twisting shadows across the polished but grimy concrete floor. Composed of dark, weathered wood planks held together by rusted iron sheets, the crate is the only object in the vast space. Propelled relentlessly and silently across the smooth surface, its movement unsettlingly fluid and steady, driven by an entirely unseen force. The camera's parallel movement emphasizes the crate's frictionless gliding, creating an atmosphere of tension, mystery, and deep unease within the cold, monochrome setting.", + "seed": 6160, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_1/index_2899_seed_6160_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.032174562010914086, + "vlm_logprob": -0.032174562010914086, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.032174562010914086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052696969360113144 + ], + "opposite_logprob_tokens": [ + -3.052696943283081 + ], + "logprob_reward": -0.052696969360113144, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.052696969360113144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1517014056444168, + 0.024000301957130432, + 0.03512870520353317, + 0.042995188385248184, + 0.021821890026330948, + 0.06474226713180542, + 0.5791852474212646, + 0.02534870058298111, + 0.030680391937494278, + 0.024395950138568878 + ], + "credit_map": [ + 0.01576286368072033, + 0.01866256073117256, + 0.017509840428829193, + 0.03142012655735016, + 0.01827079989016056, + 0.01827085018157959, + 0.018637269735336304, + 0.013167074881494045, + 0.0023028149735182524, + 0.0038388383109122515, + 0.002620202023535967, + 0.003475590143352747, + 0.0037242474500089884, + 0.003300322452560067, + 0.0019315071403980255, + 0.0028067766688764095, + 0.005079164169728756, + 0.00575988320633769, + 0.004788560792803764, + 0.002891869517043233, + 0.005931262392550707, + 0.004787007346749306, + 0.002555784536525607, + 0.0033351690508425236, + 0.004987874533981085, + 0.007840064354240894, + 0.009056111797690392, + 0.0038875481113791466, + 0.0046838074922561646, + 0.0028984486125409603, + 0.003002095501869917, + 0.006639237515628338, + 0.0030771498568356037, + 0.002940189791843295, + 0.002949247369542718, + 0.0024933505337685347, + 0.004037644248455763, + 0.0020749971736222506, + 0.0027892105281352997, + 0.001460099359974265, + 0.010861063376069069, + 0.018592089414596558, + 0.00508901895955205, + 0.007958809845149517, + 0.005597177427262068, + 0.006250923033803701, + 0.006723909638822079, + 0.003669271245598793, + 0.0912286713719368, + 0.11364316940307617, + 0.08062788844108582, + 0.06868702173233032, + 0.05803762376308441, + 0.06906244158744812, + 0.06507541984319687, + 0.03282291814684868, + 0.002725789323449135, + 0.0039139422588050365, + 0.0029302379116415977, + 0.003289424581453204, + 0.0021800424437969923, + 0.005265168845653534, + 0.0026488269213587046, + 0.002395268063992262, + 0.003773418953642249, + 0.004009102936834097, + 0.002359328093007207, + 0.004543437622487545, + 0.0028904646169394255, + 0.005529511719942093, + 0.004173492547124624, + 0.0034016319550573826, + 0.0017792262369766831, + 0.00193898961879313, + 0.005604154895991087, + 0.00399800855666399, + 0.0019018658204004169, + 0.003069418016821146, + 0.0044057960622012615, + 0.0016984888352453709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.052696969360113144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006773487199097872 + ], + "opposite_logprob_tokens": [ + -5.006773471832275 + ], + "logprob_reward": -0.006773487199097872, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006773487199097872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11411462724208832, + 0.13342349231243134, + 0.05507338047027588, + 0.03373367339372635, + 0.0499701201915741, + 0.07023520767688751, + 0.3958342969417572, + 0.05756879597902298, + 0.041654348373413086, + 0.048392098397016525 + ], + "credit_map": [ + 0.010700064711272717, + 0.017218658700585365, + 0.011214596219360828, + 0.018912022933363914, + 0.020996147766709328, + 0.015210951678454876, + 0.00957381073385477, + 0.010288371704518795, + 0.012371544726192951, + 0.030966943129897118, + 0.008961349725723267, + 0.01386062428355217, + 0.02694118581712246, + 0.02078491635620594, + 0.009270815178751945, + 0.010266109369695187, + 0.0064737219363451, + 0.010052376426756382, + 0.007839149795472622, + 0.007180899381637573, + 0.0073639629408717155, + 0.0063171410001814365, + 0.004647703841328621, + 0.0051984223537147045, + 0.0039437878876924515, + 0.011589973233640194, + 0.002079266356304288, + 0.003359188325703144, + 0.006347802001982927, + 0.0020220454316586256, + 0.002102076541632414, + 0.002289534080773592, + 0.00824471190571785, + 0.008788948878645897, + 0.006508160382509232, + 0.008570089936256409, + 0.003454676829278469, + 0.002259226283058524, + 0.00783164706081152, + 0.004312656819820404, + 0.011859690770506859, + 0.00759490579366684, + 0.005603892263025045, + 0.012268458493053913, + 0.005505803506821394, + 0.007565626408904791, + 0.014643730595707893, + 0.00519309937953949, + 0.06172959879040718, + 0.07741880416870117, + 0.05402204766869545, + 0.04731769487261772, + 0.03871186822652817, + 0.047612350434064865, + 0.045962195843458176, + 0.02305975928902626, + 0.005057919770479202, + 0.005877413786947727, + 0.00447232648730278, + 0.011570875532925129, + 0.005926144775003195, + 0.007636582013219595, + 0.012296156957745552, + 0.004731376655399799, + 0.004635294899344444, + 0.003922664560377598, + 0.00448131887242198, + 0.009575729258358479, + 0.0029262546449899673, + 0.005060192663222551, + 0.00625458313152194, + 0.00479830801486969, + 0.004487793892621994, + 0.005417834036052227, + 0.0073254345916211605, + 0.009945938363671303, + 0.0067285033874213696, + 0.005337925162166357, + 0.005666837561875582, + 0.0034818316344171762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.006773487199097872 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078807413578033 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016078807413578033, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078807413578033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10025765746831894, + 0.034692756831645966, + 0.05088800564408302, + 0.046683717519044876, + 0.05484115704894066, + 0.06686560809612274, + 0.525617241859436, + 0.06607452034950256, + 0.022619975730776787, + 0.03145931288599968 + ], + "credit_map": [ + 0.011400829069316387, + 0.01300447341054678, + 0.011245261877775192, + 0.020554009824991226, + 0.012039906345307827, + 0.01297040656208992, + 0.01085080485790968, + 0.008191966451704502, + 0.0037037848960608244, + 0.0050450460985302925, + 0.0035501071251928806, + 0.003072780556976795, + 0.00791133102029562, + 0.005060770083218813, + 0.0029241028241813183, + 0.0034248344600200653, + 0.006156781688332558, + 0.013854737393558025, + 0.005021885968744755, + 0.004298764746636152, + 0.0077240667305886745, + 0.0054930648766458035, + 0.003588980296626687, + 0.004749726969748735, + 0.0065080225467681885, + 0.01131245493888855, + 0.006596168503165245, + 0.0035691752564162016, + 0.007298436481505632, + 0.003218208672478795, + 0.002756390254944563, + 0.005424861330538988, + 0.009215353056788445, + 0.010920901782810688, + 0.009318946860730648, + 0.0038796537555754185, + 0.006062852684408426, + 0.005289755761623383, + 0.004836035426706076, + 0.005317660514265299, + 0.01128939539194107, + 0.014820415526628494, + 0.005518734455108643, + 0.005397454369813204, + 0.00930368434637785, + 0.007950281724333763, + 0.00653478130698204, + 0.0060508581809699535, + 0.08202983438968658, + 0.10370204597711563, + 0.07052607834339142, + 0.06290596723556519, + 0.053063008934259415, + 0.06373969465494156, + 0.059485871344804764, + 0.030164774507284164, + 0.007424627896398306, + 0.012178092263638973, + 0.005462117027491331, + 0.009741727262735367, + 0.00678569171577692, + 0.010339383967220783, + 0.007787137292325497, + 0.0063557433895766735, + 0.0027620166074484587, + 0.0026978724636137486, + 0.0021184973884373903, + 0.003389599733054638, + 0.0019135202746838331, + 0.003338603302836418, + 0.0031581546645611525, + 0.003241711063310504, + 0.0028524529188871384, + 0.0025602306704968214, + 0.006933434866368771, + 0.003581034019589424, + 0.0033046177122741938, + 0.003430528799071908, + 0.006027097348123789, + 0.0027699170168489218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.016078807413578033 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204245984554291 + ], + "opposite_logprob_tokens": [ + -2.812042474746704 + ], + "logprob_reward": -0.06204245984554291, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204245984554291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08927883207798004, + 0.04541098698973656, + 0.06433500349521637, + 0.06075456738471985, + 0.06784398853778839, + 0.061500608921051025, + 0.2773960530757904, + 0.049002084881067276, + 0.23733507096767426, + 0.047142863273620605 + ], + "credit_map": [ + 0.009172871708869934, + 0.011841891333460808, + 0.010495864786207676, + 0.012538542971014977, + 0.01245738472789526, + 0.011038342490792274, + 0.008639891631901264, + 0.013094036839902401, + 0.004718141630291939, + 0.009524385444819927, + 0.003669550409540534, + 0.0062407259829342365, + 0.0077729607000947, + 0.006733889691531658, + 0.0037533771246671677, + 0.0029979513492435217, + 0.008362755179405212, + 0.009933464229106903, + 0.008380912244319916, + 0.010839187540113926, + 0.007071211934089661, + 0.007927482016384602, + 0.005597466137260199, + 0.0062225149013102055, + 0.0075328778475522995, + 0.009445689618587494, + 0.00991665106266737, + 0.00580718694254756, + 0.00874403491616249, + 0.004404763225466013, + 0.0049484362825751305, + 0.009954922832548618, + 0.010457362048327923, + 0.007805489469319582, + 0.006122411694377661, + 0.013549383729696274, + 0.005421580281108618, + 0.006496293004602194, + 0.01116990391165018, + 0.006821562070399523, + 0.010211941786110401, + 0.012538406066596508, + 0.003920361399650574, + 0.008800133131444454, + 0.004985805135220289, + 0.005389773286879063, + 0.0113359484821558, + 0.0043182335793972015, + 0.04465940594673157, + 0.04992571845650673, + 0.04321523383259773, + 0.03300875797867775, + 0.026388922706246376, + 0.031743381172418594, + 0.0325097031891346, + 0.015944909304380417, + 0.004575635306537151, + 0.005115841515362263, + 0.0037181249354034662, + 0.008475864306092262, + 0.004224596545100212, + 0.0069989534094929695, + 0.012236653827130795, + 0.003656409913673997, + 0.022076856344938278, + 0.025238139554858208, + 0.025106577202677727, + 0.05434463918209076, + 0.018005957826972008, + 0.025585414841771126, + 0.033642206341028214, + 0.033335257321596146, + 0.0032571638002991676, + 0.003177665639668703, + 0.006429146975278854, + 0.0072554792277514935, + 0.005029001273214817, + 0.0057836794294416904, + 0.011669017374515533, + 0.004541703499853611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.06204245984554291 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328108623623848 + ], + "opposite_logprob_tokens": [ + -3.7732810974121094 + ], + "logprob_reward": -0.02328108623623848, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328108623623848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18430182337760925, + 0.048963237553834915, + 0.0281185545027256, + 0.07039054483175278, + 0.04348903149366379, + 0.104532890021801, + 0.3603129982948303, + 0.044610995799303055, + 0.06611078232526779, + 0.04916917905211449 + ], + "credit_map": [ + 0.016021734103560448, + 0.021333182230591774, + 0.018778137862682343, + 0.04317507892847061, + 0.02709098905324936, + 0.02423972636461258, + 0.020754605531692505, + 0.012908367440104485, + 0.004124621395021677, + 0.011023879051208496, + 0.0035792100243270397, + 0.00808495283126831, + 0.007612362504005432, + 0.007316268514841795, + 0.0035478780046105385, + 0.0036740629002451897, + 0.0032362432684749365, + 0.003917542286217213, + 0.0031654576305299997, + 0.004621272906661034, + 0.003926355857402086, + 0.0027021868154406548, + 0.0037013839464634657, + 0.0028481115587055683, + 0.005877373740077019, + 0.01472011860460043, + 0.008603168651461601, + 0.005400508642196655, + 0.01647486910223961, + 0.004809477366507053, + 0.004410772584378719, + 0.01009425800293684, + 0.008670501410961151, + 0.005933811422437429, + 0.005218802485615015, + 0.005171995609998703, + 0.003358541987836361, + 0.003197356592863798, + 0.007901163771748543, + 0.004036858677864075, + 0.0202716663479805, + 0.033485185354948044, + 0.006356940604746342, + 0.009359940886497498, + 0.008152611553668976, + 0.011428358033299446, + 0.010599915869534016, + 0.004878274165093899, + 0.055072229355573654, + 0.07219238579273224, + 0.045989569276571274, + 0.04363425448536873, + 0.036254674196243286, + 0.0447220653295517, + 0.04132677987217903, + 0.021121008321642876, + 0.004512927494943142, + 0.004815896041691303, + 0.004446381237357855, + 0.006895009893923998, + 0.0037979125045239925, + 0.006480053998529911, + 0.008120759390294552, + 0.005542058032006025, + 0.005591148976236582, + 0.007314411457628012, + 0.007444037590175867, + 0.014987504109740257, + 0.003934010397642851, + 0.008253397420048714, + 0.010749805718660355, + 0.007836470380425453, + 0.006111931521445513, + 0.006953595671802759, + 0.006510176230221987, + 0.005315210670232773, + 0.005780257284641266, + 0.007255125790834427, + 0.007037451956421137, + 0.004205432254821062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.02328108623623848 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot, the camera tracks steadily sideways, maintaining a constant distance from a single wooden crate as it moves through a vast, cavernous, empty warehouse. Cool, ethereal light filters through a high, grimy skylight, filtering through the dusty air and casting long, twisting shadows across the polished but grimy concrete floor. Composed of dark, weathered wood planks held together by rusted iron sheets, the crate is the only object in the vast space. Propelled relentlessly and silently across the smooth surface, its movement unsettlingly fluid and steady, driven by an entirely unseen force. The camera's parallel movement emphasizes the crate's frictionless gliding, creating an atmosphere of tension, mystery, and deep unease within the cold, monochrome setting.", + "seed": 6161, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_1/index_2899_seed_6161_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014948364160954953, + "vlm_logprob": -0.014948364160954953, + "vlm_yes_no_margin": 4.3249999523162845, + "vlm_token_credit": -0.014948364160954953, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023443426936864853 + ], + "opposite_logprob_tokens": [ + -4.023443222045898 + ], + "logprob_reward": -0.023443426936864853, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.023443426936864853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07116599380970001, + 0.040907617658376694, + 0.06809554994106293, + 0.04437359794974327, + 0.0445973239839077, + 0.05111083760857582, + 0.06041073426604271, + 0.056019868701696396, + 0.5339526534080505, + 0.029365791007876396 + ], + "credit_map": [ + 0.01030142791569233, + 0.01289744395762682, + 0.004974368028342724, + 0.00608655484393239, + 0.0123292813077569, + 0.009372496046125889, + 0.008288596756756306, + 0.006915820296853781, + 0.006277457810938358, + 0.004002459347248077, + 0.002882234053686261, + 0.004141352139413357, + 0.009404828771948814, + 0.005402018316090107, + 0.0044633131474256516, + 0.00433395616710186, + 0.0126305827870965, + 0.008598556742072105, + 0.003994406666606665, + 0.004961870145052671, + 0.013374845497310162, + 0.010557889007031918, + 0.009530387818813324, + 0.004447010811418295, + 0.0076205055229365826, + 0.004649103153496981, + 0.0029219589196145535, + 0.004519529640674591, + 0.008649597875773907, + 0.006930929608643055, + 0.005951691884547472, + 0.003130281111225486, + 0.006989570800215006, + 0.009969708509743214, + 0.002876217244192958, + 0.0037646787241101265, + 0.005828531924635172, + 0.006829786580055952, + 0.005928535480052233, + 0.0024102963507175446, + 0.004599494859576225, + 0.01269725151360035, + 0.001939807552844286, + 0.004198663402348757, + 0.010889536701142788, + 0.007102923933416605, + 0.006422796286642551, + 0.0032603610306978226, + 0.0098158810287714, + 0.009957834146916866, + 0.003205228131264448, + 0.005505463574081659, + 0.012666931375861168, + 0.009226592257618904, + 0.004833574406802654, + 0.005199233070015907, + 0.004663438070565462, + 0.008715278469026089, + 0.0038975898642092943, + 0.006967576686292887, + 0.009031704626977444, + 0.007331574335694313, + 0.00837620534002781, + 0.007036502007395029, + 0.0792999193072319, + 0.07317642867565155, + 0.03655393421649933, + 0.05269525572657585, + 0.0576675720512867, + 0.10294955223798752, + 0.07688942551612854, + 0.05472058057785034, + 0.004598744213581085, + 0.0057235886342823505, + 0.001797226257622242, + 0.0028842214960604906, + 0.0030341316014528275, + 0.0032786645460873842, + 0.004435340873897076, + 0.0036138752475380898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.023443426936864853 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020644783973693848 + ], + "opposite_logprob_tokens": [ + -3.8956446647644043 + ], + "logprob_reward": -0.020644783973693848, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020644783973693848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06719762831926346, + 0.06115277484059334, + 0.10711974650621414, + 0.09192913770675659, + 0.08052950352430344, + 0.06666615605354309, + 0.1596490740776062, + 0.06689143180847168, + 0.1759256273508072, + 0.12293887883424759 + ], + "credit_map": [ + 0.01052633672952652, + 0.009436067193746567, + 0.005975913722068071, + 0.008742768317461014, + 0.010478639043867588, + 0.007719056680798531, + 0.0077492790296673775, + 0.006569570861756802, + 0.011714144609868526, + 0.00618790602311492, + 0.0038767345249652863, + 0.005290161352604628, + 0.01021481305360794, + 0.00980033166706562, + 0.009182285517454147, + 0.004886399954557419, + 0.018438948318362236, + 0.01791064441204071, + 0.004521654918789864, + 0.006621657405048609, + 0.01952292025089264, + 0.014596342109143734, + 0.01822139322757721, + 0.007286189589649439, + 0.008340814150869846, + 0.00699638994410634, + 0.006232918240129948, + 0.009200743399560452, + 0.018579769879579544, + 0.014184297062456608, + 0.020351139828562737, + 0.008043065667152405, + 0.008566537871956825, + 0.010299178771674633, + 0.005443233996629715, + 0.006621127016842365, + 0.01630634441971779, + 0.009813681244850159, + 0.01616848260164261, + 0.007310926914215088, + 0.005779283121228218, + 0.014773325063288212, + 0.002851195400580764, + 0.00556340953335166, + 0.015157906338572502, + 0.00997348502278328, + 0.009313421323895454, + 0.0032541370019316673, + 0.017320794984698296, + 0.023195015266537666, + 0.007717681583017111, + 0.017588205635547638, + 0.04237920045852661, + 0.024888837710022926, + 0.015520553104579449, + 0.011038793250918388, + 0.008745670318603516, + 0.009371903724968433, + 0.0069425576366484165, + 0.005796224810183048, + 0.008352884091436863, + 0.014469152316451073, + 0.006324829533696175, + 0.006888213101774454, + 0.023766718804836273, + 0.02445642277598381, + 0.013617313466966152, + 0.012595567852258682, + 0.022243810817599297, + 0.036162685602903366, + 0.020162219181656837, + 0.022920897230505943, + 0.009886075742542744, + 0.016749585047364235, + 0.007606086321175098, + 0.012390035204589367, + 0.021488947793841362, + 0.014342022128403187, + 0.025522423908114433, + 0.014953714795410633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.020644783973693848 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979746226221323 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.00979746226221323, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979746226221323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05411522462964058, + 0.05338296666741371, + 0.0706477239727974, + 0.07855942100286484, + 0.037151165306568146, + 0.03592708334326744, + 0.16214346885681152, + 0.24939768016338348, + 0.1981877237558365, + 0.0604875348508358 + ], + "credit_map": [ + 0.013317336328327656, + 0.005595500115305185, + 0.0028926583472639322, + 0.003753569209948182, + 0.00808181893080473, + 0.008150088600814342, + 0.006416039541363716, + 0.005908217281103134, + 0.008952676318585873, + 0.00570875545963645, + 0.003991624340415001, + 0.004711836110800505, + 0.011491759680211544, + 0.00704225292429328, + 0.004398362711071968, + 0.007085698656737804, + 0.01164428610354662, + 0.0085679329931736, + 0.0055326614528894424, + 0.005639988463371992, + 0.011789264157414436, + 0.009518994018435478, + 0.011956487782299519, + 0.0059981029480695724, + 0.016953254118561745, + 0.006632079835981131, + 0.004902750253677368, + 0.006502439733594656, + 0.0177670456469059, + 0.011425883509218693, + 0.008474183268845081, + 0.005901780910789967, + 0.0041191717609763145, + 0.0043222722597420216, + 0.002094552619382739, + 0.0035374967847019434, + 0.006872214376926422, + 0.005415847525000572, + 0.00821120198816061, + 0.002578406361863017, + 0.005259962752461433, + 0.0050773718394339085, + 0.002197167370468378, + 0.0030301474034786224, + 0.007240870501846075, + 0.005127374082803726, + 0.005744445603340864, + 0.0022497440222650766, + 0.0300325695425272, + 0.03584583103656769, + 0.009835812263190746, + 0.013328460976481438, + 0.02601180039346218, + 0.020055977627635002, + 0.015469864010810852, + 0.011563163250684738, + 0.037050869315862656, + 0.03136807680130005, + 0.026022648438811302, + 0.026663489639759064, + 0.03252739459276199, + 0.04647783935070038, + 0.02563859522342682, + 0.023648755624890327, + 0.02737647481262684, + 0.027918469160795212, + 0.013186754658818245, + 0.01954405941069126, + 0.022749874740839005, + 0.03817405179142952, + 0.02798963338136673, + 0.021248411387205124, + 0.008229341357946396, + 0.007885788567364216, + 0.005274875555187464, + 0.004941044375300407, + 0.00633806549012661, + 0.008283738978207111, + 0.011067437008023262, + 0.008467244915664196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.00979746226221323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780462831258774 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780462831258774, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780462831258774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12724900245666504, + 0.09157214313745499, + 0.0806787759065628, + 0.10269121080636978, + 0.1743452250957489, + 0.08736946433782578, + 0.10441617667675018, + 0.06741895526647568, + 0.1120833307504654, + 0.05217570438981056 + ], + "credit_map": [ + 0.026404647156596184, + 0.01423303596675396, + 0.007151898927986622, + 0.012005764991044998, + 0.024613214656710625, + 0.018615804612636566, + 0.014048279263079166, + 0.010176355019211769, + 0.015808669850230217, + 0.012271370738744736, + 0.00507647730410099, + 0.011754333041608334, + 0.01904263161122799, + 0.012279302813112736, + 0.009028715081512928, + 0.006310647819191217, + 0.01246863417327404, + 0.01107828225940466, + 0.005715847481042147, + 0.008502278476953506, + 0.013607686385512352, + 0.010822857730090618, + 0.013056904077529907, + 0.005426285322755575, + 0.017823101952672005, + 0.008126200176775455, + 0.00725057628005743, + 0.009591699577867985, + 0.017470011487603188, + 0.014582859352231026, + 0.01827831193804741, + 0.009568446315824986, + 0.01578366383910179, + 0.0243863295763731, + 0.0077430796809494495, + 0.016221266239881516, + 0.03561883792281151, + 0.027196142822504044, + 0.03600577637553215, + 0.01139013096690178, + 0.010764269158244133, + 0.019515544176101685, + 0.003356121713295579, + 0.008409758098423481, + 0.0142587935552001, + 0.01326628215610981, + 0.012558781541883945, + 0.005239919759333134, + 0.007858785800635815, + 0.013033987022936344, + 0.005731125827878714, + 0.010703567415475845, + 0.028961464762687683, + 0.017545374110341072, + 0.013304767198860645, + 0.007277111988514662, + 0.007968319579958916, + 0.010334284044802189, + 0.005861235782504082, + 0.008949518203735352, + 0.010539558716118336, + 0.008669213391840458, + 0.008844816125929356, + 0.0062520140781998634, + 0.013823173940181732, + 0.01633201166987419, + 0.007234535180032253, + 0.012846448458731174, + 0.009752238169312477, + 0.01648448035120964, + 0.020902497693896294, + 0.014707951806485653, + 0.004744411911815405, + 0.008593031205236912, + 0.0026080652605742216, + 0.006262119859457016, + 0.007348998449742794, + 0.00631370022892952, + 0.010584373958408833, + 0.005721001885831356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.009780462831258774 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075684800744057 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075684800744057, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075684800744057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25332707166671753, + 0.05996107682585716, + 0.05945461243391037, + 0.05080355703830719, + 0.09320336580276489, + 0.12058959156274796, + 0.075416699051857, + 0.07756565511226654, + 0.17580431699752808, + 0.03387409448623657 + ], + "credit_map": [ + 0.04240594059228897, + 0.029932130128145218, + 0.016954949125647545, + 0.030600016936659813, + 0.051253750920295715, + 0.03290608897805214, + 0.029334688559174538, + 0.019939487800002098, + 0.008951229974627495, + 0.007651757914572954, + 0.005800262093544006, + 0.004779196809977293, + 0.012219495140016079, + 0.0071345120668411255, + 0.008102628402411938, + 0.005321993492543697, + 0.013341621495783329, + 0.005299042910337448, + 0.003179449588060379, + 0.003728247247636318, + 0.009954136796295643, + 0.008661682717502117, + 0.010804873891174793, + 0.004485553130507469, + 0.006671049166470766, + 0.004536548629403114, + 0.004259391222149134, + 0.0042715962044894695, + 0.010476103983819485, + 0.0072213732637465, + 0.009854410775005817, + 0.003513079136610031, + 0.011778331361711025, + 0.015182950533926487, + 0.005000947508960962, + 0.00801236554980278, + 0.015245013870298862, + 0.014441246166825294, + 0.016578663140535355, + 0.006963837891817093, + 0.01689399592578411, + 0.023704158142209053, + 0.004819029942154884, + 0.010850823484361172, + 0.026800980791449547, + 0.01712474413216114, + 0.011838545091450214, + 0.008557305671274662, + 0.009659557603299618, + 0.01455591432750225, + 0.0035511725582182407, + 0.006581700872629881, + 0.01652478612959385, + 0.01183124165982008, + 0.008438311517238617, + 0.004274014383554459, + 0.011050347238779068, + 0.008322963491082191, + 0.006560278125107288, + 0.007904834114015102, + 0.011939954943954945, + 0.014480339363217354, + 0.008643976412713528, + 0.008662964217364788, + 0.02113555744290352, + 0.0261542871594429, + 0.010113367810845375, + 0.018328024074435234, + 0.019460121169686317, + 0.03044971451163292, + 0.03182656690478325, + 0.018336670473217964, + 0.005360142793506384, + 0.00410479074344039, + 0.002245580544695258, + 0.0027043758891522884, + 0.004111274611204863, + 0.005316633265465498, + 0.006175530608743429, + 0.0038557674270123243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.011075684800744057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot, the camera tracks steadily sideways, maintaining a constant distance from a single wooden crate as it moves through a vast, cavernous, empty warehouse. Cool, ethereal light filters through a high, grimy skylight, filtering through the dusty air and casting long, twisting shadows across the polished but grimy concrete floor. Composed of dark, weathered wood planks held together by rusted iron sheets, the crate is the only object in the vast space. Propelled relentlessly and silently across the smooth surface, its movement unsettlingly fluid and steady, driven by an entirely unseen force. The camera's parallel movement emphasizes the crate's frictionless gliding, creating an atmosphere of tension, mystery, and deep unease within the cold, monochrome setting.", + "seed": 6162, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_1/index_2899_seed_6162_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030309633910655977, + "vlm_logprob": -0.030309633910655977, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.030309633910655977, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08348357677459717 + ], + "opposite_logprob_tokens": [ + -2.5834836959838867 + ], + "logprob_reward": -0.08348357677459717, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08348357677459717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08389516174793243, + 0.06850816309452057, + 0.03428370878100395, + 0.10836571455001831, + 0.2204262763261795, + 0.029751690104603767, + 0.11795169860124588, + 0.3016796410083771, + 0.017017696052789688, + 0.018120242282748222 + ], + "credit_map": [ + 0.00862776767462492, + 0.011010725982487202, + 0.01736932434141636, + 0.010966095142066479, + 0.008767358027398586, + 0.0109304990619421, + 0.010464146733283997, + 0.005759242922067642, + 0.006747621577233076, + 0.009609592147171497, + 0.012281833216547966, + 0.009938460774719715, + 0.00620682630687952, + 0.006853864993900061, + 0.005867497995495796, + 0.011002461425960064, + 0.005528343375772238, + 0.0077742296271026134, + 0.0020294543355703354, + 0.005766838323324919, + 0.00410737469792366, + 0.004198223818093538, + 0.0027639525942504406, + 0.0021152934059500694, + 0.015876414254307747, + 0.021257802844047546, + 0.005309712141752243, + 0.008287273347377777, + 0.01341684628278017, + 0.011262606829404831, + 0.009515466168522835, + 0.023439593613147736, + 0.03300270438194275, + 0.010070166550576687, + 0.005387493409216404, + 0.08392246812582016, + 0.007504898589104414, + 0.014122594147920609, + 0.012881341390311718, + 0.053534600883722305, + 0.005009081214666367, + 0.0035328606609255075, + 0.0031156069599092007, + 0.0042216721922159195, + 0.003688849275931716, + 0.004698844626545906, + 0.0032673869282007217, + 0.0022173887118697166, + 0.016948305070400238, + 0.008330809883773327, + 0.014262189157307148, + 0.020119063556194305, + 0.01055101677775383, + 0.019436508417129517, + 0.021497368812561035, + 0.006806424353271723, + 0.028723226860165596, + 0.03440176323056221, + 0.025612249970436096, + 0.06585423648357391, + 0.042132534086704254, + 0.032367877662181854, + 0.04729649797081947, + 0.025291262194514275, + 0.0011775760212913156, + 0.0018561170436441898, + 0.0008043342386372387, + 0.001672354992479086, + 0.0031461843755096197, + 0.003540105652064085, + 0.0037554982118308544, + 0.0010655256919562817, + 0.0009654788300395012, + 0.0016071292338892817, + 0.0011999475536867976, + 0.0014822061639279127, + 0.0030136757995933294, + 0.003079985501244664, + 0.005128137301653624, + 0.001643682480789721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.08348357677459717 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011114707216620445 + ], + "opposite_logprob_tokens": [ + -4.511114597320557 + ], + "logprob_reward": -0.011114707216620445, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011114707216620445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014204133301973343, + 0.010353080928325653, + 0.021240120753645897, + 0.02407781034708023, + 0.7754456996917725, + 0.02144336886703968, + 0.07736651599407196, + 0.03572554513812065, + 0.011251626536250114, + 0.008892025798559189 + ], + "credit_map": [ + 0.0016559205250814557, + 0.0017094288486987352, + 0.0030716038309037685, + 0.002271236153319478, + 0.0014284466160461307, + 0.0013873224379494786, + 0.0017716757720336318, + 0.0009084995836019516, + 0.001157013000920415, + 0.0011135434033349156, + 0.0028010832611471415, + 0.00098432379309088, + 0.0008448807639069855, + 0.0007507835980504751, + 0.0007598209194839001, + 0.0019416322465986013, + 0.0041553424671292305, + 0.0028230580501258373, + 0.0009631164721213281, + 0.006022974383085966, + 0.001478524412959814, + 0.0014522065175697207, + 0.0032097287476062775, + 0.0011351704597473145, + 0.0035490773152559996, + 0.004671170376241207, + 0.0017944780411198735, + 0.004023168236017227, + 0.0020645533222705126, + 0.0020059728994965553, + 0.0037493479903787374, + 0.002220040885731578, + 0.11585225909948349, + 0.035162489861249924, + 0.019299261271953583, + 0.2977185845375061, + 0.02533005177974701, + 0.04941415414214134, + 0.044505663216114044, + 0.18816326558589935, + 0.002081202110275626, + 0.0021410805638879538, + 0.0019933355506509542, + 0.004267124459147453, + 0.0024305693805217743, + 0.0032478251960128546, + 0.0025492750573903322, + 0.002732956549152732, + 0.010122536681592464, + 0.006704537197947502, + 0.009051112458109856, + 0.013266755267977715, + 0.007154513616114855, + 0.011019189842045307, + 0.014783586375415325, + 0.005264279432594776, + 0.003101915353909135, + 0.004184543155133724, + 0.003067800309509039, + 0.008406325243413448, + 0.004056519828736782, + 0.003998872824013233, + 0.006160193122923374, + 0.0027493739034980536, + 0.0006112281698733568, + 0.0010109025752171874, + 0.0005198509315960109, + 0.0014452488394454122, + 0.0018666393589228392, + 0.002126280451193452, + 0.0029011638835072517, + 0.0007703119772486389, + 0.0006892999517731369, + 0.0007635317160747945, + 0.0003568491374608129, + 0.001292593078687787, + 0.0010057748295366764, + 0.0016441517509520054, + 0.002671079244464636, + 0.0004687460605055094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.011114707216620445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607939414680004 + ], + "opposite_logprob_tokens": [ + -4.141079425811768 + ], + "logprob_reward": -0.01607939414680004, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607939414680004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12440691143274307, + 0.03529655188322067, + 0.02513873018324375, + 0.1610780656337738, + 0.03900528326630592, + 0.016538752242922783, + 0.2806132137775421, + 0.20135122537612915, + 0.05180087685585022, + 0.06477031111717224 + ], + "credit_map": [ + 0.018741048872470856, + 0.014860969968140125, + 0.03246929123997688, + 0.01541906502097845, + 0.01074796449393034, + 0.012817857787013054, + 0.012054990977048874, + 0.007295738905668259, + 0.00541112432256341, + 0.006877502426505089, + 0.0024499676655977964, + 0.0053467340767383575, + 0.004534102510660887, + 0.004510770086199045, + 0.0033592048566788435, + 0.002807152457535267, + 0.004188163671642542, + 0.0034550365526229143, + 0.0009703256073407829, + 0.006457343231886625, + 0.002859194064512849, + 0.0027406634762883186, + 0.0032142596319317818, + 0.0012537461007013917, + 0.019952986389398575, + 0.017119551077485085, + 0.010411148890852928, + 0.012355400249361992, + 0.03251298516988754, + 0.032502688467502594, + 0.01787961833178997, + 0.018343718722462654, + 0.0057775056920945644, + 0.0021794645581394434, + 0.0018275062320753932, + 0.013464318588376045, + 0.0022578518837690353, + 0.003021238837391138, + 0.0025966432876884937, + 0.007880756631493568, + 0.0013856716686859727, + 0.0015621754573658109, + 0.0017916866345331073, + 0.003004753962159157, + 0.001754809869453311, + 0.0033699264749884605, + 0.0022515493910759687, + 0.0014181811129674315, + 0.04126537963747978, + 0.019344929605722427, + 0.03444669768214226, + 0.05236457660794258, + 0.02462696097791195, + 0.044093187898397446, + 0.04895639419555664, + 0.015515110455453396, + 0.018926722928881645, + 0.022743813693523407, + 0.01701487973332405, + 0.044001370668411255, + 0.02713797427713871, + 0.021355433389544487, + 0.032762352377176285, + 0.0174087043851614, + 0.0036893251817673445, + 0.005492433439940214, + 0.002385474508628249, + 0.0049129510298371315, + 0.009586852043867111, + 0.008517607115209103, + 0.012690606527030468, + 0.004525633063167334, + 0.002247412921860814, + 0.003402032656595111, + 0.002400984289124608, + 0.00909553375095129, + 0.006952765863388777, + 0.0127848070114851, + 0.024722207337617874, + 0.0031645798590034246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.01607939414680004 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797164723277092 + ], + "opposite_logprob_tokens": [ + -3.529797077178955 + ], + "logprob_reward": -0.029797164723277092, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797164723277092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04441496729850769, + 0.06611485034227371, + 0.035071682184934616, + 0.07969977706670761, + 0.41695326566696167, + 0.027480799704790115, + 0.21068425476551056, + 0.0349695198237896, + 0.055825505405664444, + 0.02878537029027939 + ], + "credit_map": [ + 0.006539893336594105, + 0.005527636501938105, + 0.0060651227831840515, + 0.005771453492343426, + 0.0069901044480502605, + 0.004756574518978596, + 0.005611888598650694, + 0.0031522964127361774, + 0.007298450451344252, + 0.009387442842125893, + 0.009206765331327915, + 0.011864923872053623, + 0.0067740557715296745, + 0.004530314356088638, + 0.007103550247848034, + 0.00994934979826212, + 0.005750909447669983, + 0.007388146594166756, + 0.0015815314836800098, + 0.00649500684812665, + 0.0038930957671254873, + 0.0029359024483710527, + 0.005097083281725645, + 0.0019300068961456418, + 0.013721490278840065, + 0.023579299449920654, + 0.0031288459431380033, + 0.007601401302963495, + 0.007735913153737783, + 0.006085600703954697, + 0.008243681862950325, + 0.009603540413081646, + 0.06216537207365036, + 0.018526578322052956, + 0.0104222996160388, + 0.1599225401878357, + 0.013524056412279606, + 0.026356011629104614, + 0.023731181398034096, + 0.10230521112680435, + 0.0024718453641980886, + 0.003454994410276413, + 0.0025248387828469276, + 0.005328189581632614, + 0.003473132848739624, + 0.003972295206040144, + 0.003134588012471795, + 0.003120915964245796, + 0.029664883390069008, + 0.013587411493062973, + 0.02459968626499176, + 0.04132634028792381, + 0.016809063032269478, + 0.035648949444293976, + 0.03746125474572182, + 0.011586652137339115, + 0.003989247605204582, + 0.004062774125486612, + 0.002777331741526723, + 0.006379642523825169, + 0.006139211356639862, + 0.0042658341117203236, + 0.004662637133151293, + 0.0026928407605737448, + 0.003940001130104065, + 0.006083478219807148, + 0.0026513757184147835, + 0.00566578283905983, + 0.009498369880020618, + 0.008966971188783646, + 0.01388752181082964, + 0.005132003221660852, + 0.0013572756433859468, + 0.0012177217286080122, + 0.0016902723582461476, + 0.002915675053372979, + 0.0036805281415581703, + 0.005958525463938713, + 0.010734651237726212, + 0.001230720430612564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.029797164723277092 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cavernous, empty warehouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single wooden crate composed of dark weathered wood planks and rusted iron sheets visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the floor polished but grimy concrete with long, twisting shadows cast by cool, ethereal light from a high grimy skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the crate moving across the floor in a frictionless, fluid, and steady gliding motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a low-angle, wide shot that tracks steadily sideways at a constant distance parallel to the crate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107332669198513 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.01107332669198513, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107332669198513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013675565831363201, + 0.048994001001119614, + 0.017230387777090073, + 0.014738966710865498, + 0.7973045706748962, + 0.026792122051119804, + 0.03829289600253105, + 0.005367543548345566, + 0.027661465108394623, + 0.009942512959241867 + ], + "credit_map": [ + 0.0017743074567988515, + 0.0025507903192192316, + 0.0018114279955625534, + 0.0018101116875186563, + 0.0018449132330715656, + 0.001286344020627439, + 0.0017108672764152288, + 0.0008868034929037094, + 0.005361836403608322, + 0.005701013840734959, + 0.009305237792432308, + 0.007281855680048466, + 0.004313129931688309, + 0.003148850752040744, + 0.005009196698665619, + 0.008872882463037968, + 0.0035062944516539574, + 0.0037024999037384987, + 0.0006772225024178624, + 0.0030277946498245, + 0.002143226796761155, + 0.0013749063946306705, + 0.002023440320044756, + 0.0007750022923573852, + 0.001966881798580289, + 0.002676932839676738, + 0.0006459829164668918, + 0.0010803240584209561, + 0.0024984886404126883, + 0.0019861951004713774, + 0.002289606025442481, + 0.0015945549821481109, + 0.11927588284015656, + 0.03601754084229469, + 0.01989814080297947, + 0.3059252202510834, + 0.025960003957152367, + 0.05074315145611763, + 0.045662909746170044, + 0.1938217431306839, + 0.0038069800939410925, + 0.0026217703707516193, + 0.002533768070861697, + 0.0043347557075321674, + 0.002430008491501212, + 0.004962962586432695, + 0.0025789295323193073, + 0.0035229481291025877, + 0.00625627301633358, + 0.0032073981128633022, + 0.004154517315328121, + 0.005948591511696577, + 0.004973115399479866, + 0.00485860975459218, + 0.005986270494759083, + 0.0029081199318170547, + 0.0006173086003400385, + 0.0007742990856058896, + 0.0004639208491425961, + 0.0006672609597444534, + 0.0009402443538419902, + 0.0008661149768158793, + 0.0005567114567384124, + 0.0004816835280507803, + 0.0020539602264761925, + 0.002935027237981558, + 0.001302886987105012, + 0.002856023609638214, + 0.004951943643391132, + 0.004447920713573694, + 0.006532523315399885, + 0.002581179141998291, + 0.0021506750490516424, + 0.0014066307339817286, + 0.0011794199235737324, + 0.0008139620185829699, + 0.001350984675809741, + 0.001121699227951467, + 0.000987116596661508, + 0.0009320246172137558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.01107332669198513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_11/index_tensor([4715, 4715])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_11/index_tensor([4715, 4715])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2d399839116940e2c79757cc65faee56c559a53d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_11/index_tensor([4715, 4715])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fluid, low-angle, cinematic tracking shot, rendered in a vibrant anime style, dynamically follows a dramatic chase across the mossy forest floor. The pursuer, a massive, centipede-like creature, is stunningly depicted. Its segmented body is encased in a vibrant reddish-brown carapace, its chitinous sheen reflecting the dappled forest light. A multitude of sharp, electric-yellow legs tumbles in a coordinated, undulating blur, scurrying at terrifying speed over fuzzy emerald-green moss and decaying leaves. It relentlessly pursues a much smaller insect, an iridescent beetle, its delicate, translucent wings beating frantically as it desperately zigzags between gnarled tree roots that, from this low angle, resemble massive canyon walls. Spotted sunlight penetrates the dense tree canopy above, casting dramatic, shifting spotlights and deep shadows that heighten the tension. The camera stays close to the ground, emphasizing the creatures' sheer speed and predatory single-mindedness, while subtle motion blur and speed lines emphasize the swift movements, creating a visually stunning and high-stakes atmosphere filled with a naturalistic yet dreamlike struggle between life and death.", + "seed": 16559, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_11/index_4715_seed_16559_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0067198773380368945, + "vlm_logprob": -0.0067198773380368945, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.0067198773380368945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0026950249448418617 + ], + "opposite_logprob_tokens": [ + -6.752695083618164 + ], + "logprob_reward": -0.0026950249448418617, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0026950249448418617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07837262749671936, + 0.14712977409362793, + 0.08389204740524292, + 0.0530618280172348, + 0.04680633544921875, + 0.11251217871904373, + 0.06803736835718155, + 0.029303474351763725, + 0.09335996210575104, + 0.28752434253692627 + ], + "credit_map": [ + 0.0072681172750890255, + 0.011838467791676521, + 0.008493272587656975, + 0.009398088790476322, + 0.008218370378017426, + 0.009342598728835583, + 0.016453953459858894, + 0.007359764073044062, + 0.005568096414208412, + 0.005810406059026718, + 0.007950430735945702, + 0.009297040291130543, + 0.012639177963137627, + 0.045443590730428696, + 0.010863053612411022, + 0.04955798014998436, + 0.006610363721847534, + 0.007647898979485035, + 0.008641533553600311, + 0.011366826482117176, + 0.01075152587145567, + 0.009031356312334538, + 0.019522687420248985, + 0.01031985692679882, + 0.004782597068697214, + 0.006796549540013075, + 0.006284359842538834, + 0.007177021354436874, + 0.008300351910293102, + 0.00492503959685564, + 0.007416502106934786, + 0.007379408925771713, + 0.0036706915125250816, + 0.004593468736857176, + 0.005127044394612312, + 0.006124379578977823, + 0.0037301459815353155, + 0.005317185074090958, + 0.009986381977796555, + 0.008257037959992886, + 0.010899964720010757, + 0.01817796751856804, + 0.014407115988433361, + 0.009676441550254822, + 0.019440682604908943, + 0.011475685983896255, + 0.01255310233682394, + 0.015881208702921867, + 0.012216637842357159, + 0.006284598261117935, + 0.009030592627823353, + 0.010091733187437057, + 0.005497732199728489, + 0.01211946178227663, + 0.0056527648121118546, + 0.007143851835280657, + 0.0025462154299020767, + 0.003138954984024167, + 0.007397040259093046, + 0.005291002336889505, + 0.0026723863556981087, + 0.002366018248721957, + 0.0035851642023772, + 0.002306691836565733, + 0.005908467806875706, + 0.0059750815853476524, + 0.02283099666237831, + 0.016798051074147224, + 0.006746563129127026, + 0.007767604198306799, + 0.013064850121736526, + 0.014268350787460804, + 0.012422974221408367, + 0.020299168303608894, + 0.01361902616918087, + 0.012423037551343441, + 0.1185094341635704, + 0.02569560334086418, + 0.05990110710263252, + 0.02465399168431759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.0026950249448418617 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551363557577133 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.012551363557577133, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551363557577133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06301488727331161, + 0.09716134518384933, + 0.13783681392669678, + 0.08332501351833344, + 0.034248072654008865, + 0.0682358518242836, + 0.0431843139231205, + 0.035454005002975464, + 0.08365100622177124, + 0.35388869047164917 + ], + "credit_map": [ + 0.0035743280313909054, + 0.008416201919317245, + 0.006114257499575615, + 0.01007044780999422, + 0.006027933210134506, + 0.007424105890095234, + 0.01476266048848629, + 0.006624951958656311, + 0.004359722137451172, + 0.0063451300375163555, + 0.006450864486396313, + 0.006652927026152611, + 0.007613713853061199, + 0.02848128415644169, + 0.00713075278326869, + 0.030126959085464478, + 0.012712291441857815, + 0.014814846217632294, + 0.013848812319338322, + 0.01876211352646351, + 0.016186818480491638, + 0.023852305486798286, + 0.022845754399895668, + 0.014813861809670925, + 0.006033965386450291, + 0.012676620855927467, + 0.010149272158741951, + 0.010532810352742672, + 0.01008861418813467, + 0.012418155558407307, + 0.012425671331584454, + 0.008999906480312347, + 0.0038334927521646023, + 0.003860104363411665, + 0.004037711303681135, + 0.003677119966596365, + 0.0030337499920278788, + 0.0030126904603093863, + 0.006110548973083496, + 0.006682652980089188, + 0.006691054441034794, + 0.0071733747608959675, + 0.007111451122909784, + 0.005904339253902435, + 0.011169915087521076, + 0.013296309858560562, + 0.007315175607800484, + 0.009574231691658497, + 0.013214820995926857, + 0.00330346985720098, + 0.005437076557427645, + 0.004671172704547644, + 0.0032245484180748463, + 0.004325182642787695, + 0.0038724280893802643, + 0.0051356167532503605, + 0.004965227097272873, + 0.005203098524361849, + 0.0049818819388747215, + 0.004192287568002939, + 0.005977833177894354, + 0.004093117080628872, + 0.0033626880031079054, + 0.0026778709143400192, + 0.005657248664647341, + 0.006158194039016962, + 0.012502072378993034, + 0.019963884726166725, + 0.007563281804323196, + 0.007399145048111677, + 0.01592699997127056, + 0.008480175398290157, + 0.03330744057893753, + 0.028900831937789917, + 0.04537893831729889, + 0.03245134651660919, + 0.07905474305152893, + 0.048511803150177, + 0.04719887673854828, + 0.039084695279598236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.012551363557577133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203292317688465 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203292317688465, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203292317688465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06022268533706665, + 0.18864384293556213, + 0.21585486829280853, + 0.08142098039388657, + 0.04256638512015343, + 0.04443513602018356, + 0.04831967130303383, + 0.023749278858304024, + 0.026016756892204285, + 0.26877036690711975 + ], + "credit_map": [ + 0.0036034644581377506, + 0.007904959842562675, + 0.00551790464669466, + 0.008286142721772194, + 0.006400806363672018, + 0.008838688023388386, + 0.012178934179246426, + 0.007491780910640955, + 0.008446473628282547, + 0.012342449277639389, + 0.013776564039289951, + 0.02091732621192932, + 0.015270908363163471, + 0.051216818392276764, + 0.013536806218326092, + 0.05313648656010628, + 0.016778629273176193, + 0.016908762976527214, + 0.021346641704440117, + 0.03349102661013603, + 0.023688212037086487, + 0.029553759843111038, + 0.04685298725962639, + 0.02723483182489872, + 0.00788203440606594, + 0.008672046475112438, + 0.008633087389171124, + 0.011207505129277706, + 0.01046603824943304, + 0.00771810207515955, + 0.012839865870773792, + 0.01400230173021555, + 0.004841696936637163, + 0.005543487146496773, + 0.006108943372964859, + 0.005340311676263809, + 0.004761228803545237, + 0.004928954411298037, + 0.005302242934703827, + 0.0057395221665501595, + 0.005440704990178347, + 0.004831547848880291, + 0.005489588249474764, + 0.004580108914524317, + 0.006491823587566614, + 0.006382600404322147, + 0.004408699925988913, + 0.006810062099248171, + 0.008065475150942802, + 0.005123545415699482, + 0.0062634581699967384, + 0.00505078537389636, + 0.005654930602759123, + 0.008954701945185661, + 0.004382497165352106, + 0.004824278876185417, + 0.0024783730041235685, + 0.0022845896892249584, + 0.00815308652818203, + 0.004924628883600235, + 0.0018956022104248405, + 0.0013087796978652477, + 0.0015826806193217635, + 0.0011215385748073459, + 0.001750490628182888, + 0.0019088971894234419, + 0.005546971224248409, + 0.004876282997429371, + 0.0021585659123957157, + 0.0031612375751137733, + 0.0022853596601635218, + 0.004328951705247164, + 0.013011429458856583, + 0.01885356940329075, + 0.019773978739976883, + 0.01550834160298109, + 0.09856988489627838, + 0.02859942801296711, + 0.0476653128862381, + 0.026788419112563133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.014203292317688465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815095242112875 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.002815095242112875, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815095242112875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12567004561424255, + 0.09594985842704773, + 0.11804834753274918, + 0.24282991886138916, + 0.021789420396089554, + 0.03852809593081474, + 0.05641916021704674, + 0.06245451420545578, + 0.05818723887205124, + 0.180123433470726 + ], + "credit_map": [ + 0.013295944780111313, + 0.012056258507072926, + 0.02063850313425064, + 0.01961654983460903, + 0.009833394549787045, + 0.0159214548766613, + 0.01877680793404579, + 0.01553112082183361, + 0.0045846267603337765, + 0.006278970744460821, + 0.0099160922691226, + 0.009588992223143578, + 0.008418754674494267, + 0.023204201832413673, + 0.008103372529149055, + 0.025854844599962234, + 0.009949359111487865, + 0.008303136564791203, + 0.01040336862206459, + 0.016985638067126274, + 0.011699874885380268, + 0.022657394409179688, + 0.02073615975677967, + 0.01731342077255249, + 0.01514055673032999, + 0.015521816909313202, + 0.021135544404387474, + 0.04206757992506027, + 0.038407642394304276, + 0.02666977234184742, + 0.04415120556950569, + 0.0397358164191246, + 0.0025729567278176546, + 0.0021845761220902205, + 0.0025400619488209486, + 0.0029487586580216885, + 0.001966934883967042, + 0.0024003302678465843, + 0.00432107737287879, + 0.002854723483324051, + 0.00432634400203824, + 0.006296111736446619, + 0.004385802429169416, + 0.004853572230786085, + 0.005658380687236786, + 0.0037767544854432344, + 0.004429380875080824, + 0.004801750648766756, + 0.007294973824173212, + 0.005082907620817423, + 0.00951826199889183, + 0.007463161833584309, + 0.005292190704494715, + 0.010769981890916824, + 0.0044803316704928875, + 0.006517353001981974, + 0.006221062503755093, + 0.007645656820386648, + 0.011251244693994522, + 0.010115541517734528, + 0.007701589725911617, + 0.005364090204238892, + 0.0075174616649746895, + 0.006637862883508205, + 0.0034286959562450647, + 0.0036636628210544586, + 0.009277593344449997, + 0.015923259779810905, + 0.003634429071098566, + 0.003979548346251249, + 0.01203317753970623, + 0.006246872711926699, + 0.008677735924720764, + 0.01300828717648983, + 0.01280503161251545, + 0.014881816692650318, + 0.05887917801737785, + 0.02126947045326233, + 0.03417075052857399, + 0.016431152820587158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.002815095242112875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001334610627964139 + ], + "opposite_logprob_tokens": [ + -6.6263346672058105 + ], + "logprob_reward": -0.001334610627964139, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001334610627964139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12366175651550293, + 0.0723661258816719, + 0.13184648752212524, + 0.05294571444392204, + 0.06006472185254097, + 0.06784017384052277, + 0.07898400723934174, + 0.05365406721830368, + 0.09791234880685806, + 0.26072463393211365 + ], + "credit_map": [ + 0.00814539473503828, + 0.012479125522077084, + 0.010633249767124653, + 0.015299445949494839, + 0.012908593751490116, + 0.027487145736813545, + 0.017408262938261032, + 0.01930052787065506, + 0.0045729828998446465, + 0.006390983704477549, + 0.011451960541307926, + 0.010361837223172188, + 0.00998157449066639, + 0.010374964214861393, + 0.008303026668727398, + 0.010928790085017681, + 0.014607245102524757, + 0.015405322425067425, + 0.01227032020688057, + 0.019148387014865875, + 0.021378351375460625, + 0.015681467950344086, + 0.02128264121711254, + 0.012072724290192127, + 0.007371260318905115, + 0.006375919561833143, + 0.006393385585397482, + 0.006448071915656328, + 0.0065824962221086025, + 0.00550212524831295, + 0.007199132815003395, + 0.0070733134634792805, + 0.005109161138534546, + 0.006674541626125574, + 0.0072650485672056675, + 0.00679105194285512, + 0.005335437133908272, + 0.008598796091973782, + 0.010855596512556076, + 0.009435081854462624, + 0.007345128804445267, + 0.008375274948775768, + 0.007586234249174595, + 0.00677140336483717, + 0.011742936447262764, + 0.00805721990764141, + 0.007734860293567181, + 0.010227103717625141, + 0.018915044143795967, + 0.0059027099050581455, + 0.011255569756031036, + 0.009210219606757164, + 0.006700979545712471, + 0.010514109395444393, + 0.007176187355071306, + 0.009309173561632633, + 0.006540872622281313, + 0.008510713465511799, + 0.00969880074262619, + 0.005693397950381041, + 0.007235548924654722, + 0.005555197596549988, + 0.006176370661705732, + 0.004243148025125265, + 0.005924716591835022, + 0.004675851669162512, + 0.024057134985923767, + 0.01724846474826336, + 0.005482042673975229, + 0.006758098490536213, + 0.016455532982945442, + 0.0173104889690876, + 0.021584607660770416, + 0.021177321672439575, + 0.02633776143193245, + 0.023010041564702988, + 0.07053056359291077, + 0.026146220043301582, + 0.04999539256095886, + 0.02194269374012947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.001334610627964139 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, low-angle, cinematic tracking shot, rendered in a vibrant anime style, dynamically follows a dramatic chase across the mossy forest floor. The pursuer, a massive, centipede-like creature, is stunningly depicted. Its segmented body is encased in a vibrant reddish-brown carapace, its chitinous sheen reflecting the dappled forest light. A multitude of sharp, electric-yellow legs tumbles in a coordinated, undulating blur, scurrying at terrifying speed over fuzzy emerald-green moss and decaying leaves. It relentlessly pursues a much smaller insect, an iridescent beetle, its delicate, translucent wings beating frantically as it desperately zigzags between gnarled tree roots that, from this low angle, resemble massive canyon walls. Spotted sunlight penetrates the dense tree canopy above, casting dramatic, shifting spotlights and deep shadows that heighten the tension. The camera stays close to the ground, emphasizing the creatures' sheer speed and predatory single-mindedness, while subtle motion blur and speed lines emphasize the swift movements, creating a visually stunning and high-stakes atmosphere filled with a naturalistic yet dreamlike struggle between life and death.", + "seed": 16560, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_11/index_4715_seed_16560_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.005423168186098337, + "vlm_logprob": -0.005423168186098337, + "vlm_yes_no_margin": 5.7, + "vlm_token_credit": -0.005423168186098337, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002894737757742405 + ], + "opposite_logprob_tokens": [ + -6.377894878387451 + ], + "logprob_reward": -0.002894737757742405, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.002894737757742405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49049127101898193, + 0.023129787296056747, + 0.01729705184698105, + 0.02501724474132061, + 0.03130558505654335, + 0.08254563808441162, + 0.025057708844542503, + 0.057325053960084915, + 0.022222768515348434, + 0.2256079465150833 + ], + "credit_map": [ + 0.048349086195230484, + 0.014010869897902012, + 0.16795682907104492, + 0.022809311747550964, + 0.014051024802029133, + 0.033045679330825806, + 0.009833129122853279, + 0.18043528497219086, + 0.0017054416239261627, + 0.0031752611976116896, + 0.003352314932271838, + 0.002860430860891938, + 0.002810021163895726, + 0.0032949408050626516, + 0.003265530103817582, + 0.0026658428832888603, + 0.0012111981632187963, + 0.0018379483371973038, + 0.0033499060664325953, + 0.0017962646670639515, + 0.0019457789603620768, + 0.0013461918570101261, + 0.0017609273781999946, + 0.004048834089189768, + 0.0021998125594109297, + 0.0022998079657554626, + 0.003269397420808673, + 0.0037601508665829897, + 0.0021522012539207935, + 0.004321940243244171, + 0.003180929459631443, + 0.003833002643659711, + 0.001606006408110261, + 0.0025340463034808636, + 0.00477988924831152, + 0.006028330884873867, + 0.002071472117677331, + 0.002528339857235551, + 0.006485200021415949, + 0.005272298119962215, + 0.0046957116574049, + 0.006353424862027168, + 0.013892987743020058, + 0.019792310893535614, + 0.008002134039998055, + 0.013499846681952477, + 0.011834299191832542, + 0.004474914167076349, + 0.0019786569755524397, + 0.002931491704657674, + 0.004710127133876085, + 0.002677410375326872, + 0.002711992245167494, + 0.002805622760206461, + 0.0040145921520888805, + 0.0032278108410537243, + 0.005026367958635092, + 0.0033375024795532227, + 0.011355618014931679, + 0.005366044584661722, + 0.003835321171209216, + 0.006982383783906698, + 0.009059525094926357, + 0.012362288311123848, + 0.0022792944218963385, + 0.0027368778828531504, + 0.002280165208503604, + 0.0027128246147185564, + 0.0021639131009578705, + 0.003054441651329398, + 0.004112526308745146, + 0.0028827213682234287, + 0.006530286278575659, + 0.004798914771527052, + 0.010439403355121613, + 0.007905228063464165, + 0.015604836866259575, + 0.007309988606721163, + 0.09507773071527481, + 0.07794153690338135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.002894737757742405 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747201085090637 + ], + "opposite_logprob_tokens": [ + -5.006747245788574 + ], + "logprob_reward": -0.006747201085090637, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747201085090637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21272709965705872, + 0.07157223671674728, + 0.022128988057374954, + 0.02441359870135784, + 0.02053971402347088, + 0.049615226686000824, + 0.05804013833403587, + 0.040877290070056915, + 0.055360130965709686, + 0.44472554326057434 + ], + "credit_map": [ + 0.020751046016812325, + 0.006536340340971947, + 0.06963782012462616, + 0.01094956323504448, + 0.007473572622984648, + 0.0160210020840168, + 0.009396511130034924, + 0.07196124643087387, + 0.003615358378738165, + 0.0030018400866538286, + 0.01883508265018463, + 0.00683693028986454, + 0.002497877227142453, + 0.0041722790338099, + 0.027674537152051926, + 0.004938333295285702, + 0.0015352441696450114, + 0.0021719413343816996, + 0.004759648814797401, + 0.0030705758836120367, + 0.0017746856901794672, + 0.002650590380653739, + 0.0027164684142917395, + 0.003449834417551756, + 0.0016526210820302367, + 0.0034340592101216316, + 0.0032160000409930944, + 0.0038989251479506493, + 0.0034047970548272133, + 0.0030488744378089905, + 0.0034417882561683655, + 0.0023165338207036257, + 0.0011958979303017259, + 0.0013092274311929941, + 0.002730540232732892, + 0.0036633643321692944, + 0.0013347903732210398, + 0.0016968792770057917, + 0.00452649500221014, + 0.004082519561052322, + 0.002787783509120345, + 0.0021231488790363073, + 0.014705078676342964, + 0.016761699691414833, + 0.003187937429174781, + 0.002920373110100627, + 0.004780482966452837, + 0.0023487224243581295, + 0.004846394993364811, + 0.004064467735588551, + 0.008982662111520767, + 0.005658709444105625, + 0.004887829069048166, + 0.011032124049961567, + 0.010780123062431812, + 0.0077878269366919994, + 0.003598848357796669, + 0.0032856015022844076, + 0.005880675744265318, + 0.0036441460251808167, + 0.0028586036060005426, + 0.00621239747852087, + 0.00694718211889267, + 0.008449836634099483, + 0.004337002988904715, + 0.006928382907062769, + 0.0057193320244550705, + 0.009943794459104538, + 0.006525668315589428, + 0.006713752634823322, + 0.009044653736054897, + 0.006147544831037521, + 0.013510636985301971, + 0.00850633904337883, + 0.01418000552803278, + 0.016073625534772873, + 0.03159458190202713, + 0.015462414361536503, + 0.18455129861831665, + 0.16084663569927216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.006747201085090637 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202586375176907 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202586375176907, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202586375176907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6035491824150085, + 0.06016656383872032, + 0.0200300682336092, + 0.02467050962150097, + 0.027168050408363342, + 0.075796939432621, + 0.01148155890405178, + 0.0848298966884613, + 0.019206753000617027, + 0.07310046255588531 + ], + "credit_map": [ + 0.05969124659895897, + 0.015602943487465382, + 0.2066361904144287, + 0.027821935713291168, + 0.017955197021365166, + 0.04032348468899727, + 0.016883397474884987, + 0.21863481402397156, + 0.003478962928056717, + 0.006112333387136459, + 0.013997271656990051, + 0.005742244888097048, + 0.003318247850984335, + 0.0031418988946825266, + 0.020386841148138046, + 0.003988763317465782, + 0.001432627672329545, + 0.002063923282548785, + 0.004156742710620165, + 0.0027797843795269728, + 0.0022706210147589445, + 0.0014515307266265154, + 0.0015304567059502006, + 0.004344382788985968, + 0.0018792424816638231, + 0.0034700515680015087, + 0.0032194240484386683, + 0.004888975527137518, + 0.002812739461660385, + 0.0026786127127707005, + 0.003425705013796687, + 0.0022957578767091036, + 0.0012248295824974775, + 0.002149764681234956, + 0.004185451660305262, + 0.00443377997726202, + 0.0020247811917215586, + 0.0024681685026735067, + 0.005765662994235754, + 0.004915611818432808, + 0.004018484149128199, + 0.005011280532926321, + 0.01418114174157381, + 0.02164783887565136, + 0.006843539420515299, + 0.009628535248339176, + 0.010058180429041386, + 0.004407943692058325, + 0.0010579742956906557, + 0.001320571405813098, + 0.0016612558392807841, + 0.0018495334079489112, + 0.0009550274116918445, + 0.001358584500849247, + 0.0015259800711646676, + 0.0017526322044432163, + 0.005322469398379326, + 0.00409414479508996, + 0.020288053900003433, + 0.007821617648005486, + 0.00532602658495307, + 0.009563254192471504, + 0.012184686958789825, + 0.02022964134812355, + 0.0018489441135898232, + 0.0024028054904192686, + 0.0016345118638128042, + 0.002425852231681347, + 0.0022115830797702074, + 0.0019542390946298838, + 0.004337783437222242, + 0.0023910340387374163, + 0.002696745563298464, + 0.0019681896083056927, + 0.00483807735145092, + 0.002774597378447652, + 0.005446288734674454, + 0.0033351294696331024, + 0.025040719658136368, + 0.027000712230801582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.014202586375176907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019363479223102331 + ], + "opposite_logprob_tokens": [ + -6.251936435699463 + ], + "logprob_reward": -0.0019363479223102331, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019363479223102331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34046316146850586, + 0.056578028947114944, + 0.027648506686091423, + 0.015289987437427044, + 0.05259191244840622, + 0.06078098714351654, + 0.07397695630788803, + 0.03726310655474663, + 0.010173889808356762, + 0.32523345947265625 + ], + "credit_map": [ + 0.03381340205669403, + 0.008425815030932426, + 0.11630674451589584, + 0.015885068103671074, + 0.00997303705662489, + 0.024682609364390373, + 0.006831798702478409, + 0.12454470247030258, + 0.002694207476451993, + 0.0023800013586878777, + 0.014048988930881023, + 0.005420667119324207, + 0.0022248595487326384, + 0.0030560351442545652, + 0.022348951548337936, + 0.004404323175549507, + 0.0014962120912969112, + 0.001579193864017725, + 0.007792947813868523, + 0.003200715873390436, + 0.0016423927154392004, + 0.0017362086800858378, + 0.0023823040537536144, + 0.007818534970283508, + 0.0013873367570340633, + 0.0013506541727110744, + 0.0022586688864976168, + 0.001781272585503757, + 0.001196286641061306, + 0.002138707088306546, + 0.002696060808375478, + 0.0024810016620904207, + 0.0039307065308094025, + 0.0059806364588439465, + 0.006518372800201178, + 0.008347894065082073, + 0.005615870002657175, + 0.00499072065576911, + 0.00870643462985754, + 0.008501277305185795, + 0.0020875795744359493, + 0.002342639025300741, + 0.019345615059137344, + 0.023196544498205185, + 0.001906018820591271, + 0.0030649742111563683, + 0.00541320163756609, + 0.0034244165290147066, + 0.005730075761675835, + 0.005626798607409, + 0.011276938952505589, + 0.007755108177661896, + 0.005044687073677778, + 0.016207074746489525, + 0.013053324073553085, + 0.009282947517931461, + 0.0029224311001598835, + 0.0019077470060437918, + 0.00861445814371109, + 0.0034965320955961943, + 0.0023236977867782116, + 0.0037551417481154203, + 0.00533715495839715, + 0.008905945345759392, + 0.0009254051255993545, + 0.0012953926343470812, + 0.0011140390997752547, + 0.0016928353579714894, + 0.0011582447914406657, + 0.0010274192318320274, + 0.0019145810510963202, + 0.0010459732729941607, + 0.00905705988407135, + 0.0055164736695587635, + 0.015962904319167137, + 0.011139120906591415, + 0.02316933125257492, + 0.010124711319804192, + 0.14086666703224182, + 0.10939718782901764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.0019363479223102331 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001334967790171504 + ], + "opposite_logprob_tokens": [ + -6.626335144042969 + ], + "logprob_reward": -0.001334967790171504, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001334967790171504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49675482511520386, + 0.0548589825630188, + 0.028907200321555138, + 0.022677715867757797, + 0.019324585795402527, + 0.032423000782728195, + 0.0757405161857605, + 0.08900419622659683, + 0.011787230148911476, + 0.16852174699306488 + ], + "credit_map": [ + 0.0492154099047184, + 0.012522571720182896, + 0.17157362401485443, + 0.022382982075214386, + 0.013813496567308903, + 0.0337739959359169, + 0.008846542797982693, + 0.1846262365579605, + 0.003154005156829953, + 0.0032451453153043985, + 0.014214745722711086, + 0.005225143861025572, + 0.0027300575748085976, + 0.003873852314427495, + 0.018655382096767426, + 0.003760650986805558, + 0.0018748522270470858, + 0.0024251723662018776, + 0.00743732787668705, + 0.0032814284786581993, + 0.002557875122874975, + 0.002441830700263381, + 0.003237572032958269, + 0.005651141982525587, + 0.0023624564055353403, + 0.0027202200144529343, + 0.0030878069810569286, + 0.00261185341514647, + 0.003046611091122031, + 0.002681909129023552, + 0.004071318544447422, + 0.0020955391228199005, + 0.001373256091028452, + 0.00236713420599699, + 0.00225830078125, + 0.0029040437657386065, + 0.001732907840050757, + 0.001694818725809455, + 0.003782597603276372, + 0.003211527829989791, + 0.0019782742019742727, + 0.0026127323508262634, + 0.005701194982975721, + 0.008782388642430305, + 0.0026070517487823963, + 0.004818125162273645, + 0.0038801843766123056, + 0.0020430509466677904, + 0.006138762459158897, + 0.0082168560475111, + 0.011313695460557938, + 0.007143858354538679, + 0.004935473669320345, + 0.011670385487377644, + 0.016026148572564125, + 0.010295329615473747, + 0.0060579292476177216, + 0.004694684408605099, + 0.021781953051686287, + 0.008096272125840187, + 0.005567731335759163, + 0.009409462101757526, + 0.012229418382048607, + 0.021166741847991943, + 0.0015856069512665272, + 0.0013526855036616325, + 0.0011698392918333411, + 0.0015219270717352629, + 0.0010095882462337613, + 0.0018603055505082011, + 0.001630916609428823, + 0.0016563605749979615, + 0.005071674007922411, + 0.0030406105797737837, + 0.004826545715332031, + 0.006003737449645996, + 0.011801491491496563, + 0.005898778792470694, + 0.07123655825853348, + 0.06064235046505928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.001334967790171504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, low-angle, cinematic tracking shot, rendered in a vibrant anime style, dynamically follows a dramatic chase across the mossy forest floor. The pursuer, a massive, centipede-like creature, is stunningly depicted. Its segmented body is encased in a vibrant reddish-brown carapace, its chitinous sheen reflecting the dappled forest light. A multitude of sharp, electric-yellow legs tumbles in a coordinated, undulating blur, scurrying at terrifying speed over fuzzy emerald-green moss and decaying leaves. It relentlessly pursues a much smaller insect, an iridescent beetle, its delicate, translucent wings beating frantically as it desperately zigzags between gnarled tree roots that, from this low angle, resemble massive canyon walls. Spotted sunlight penetrates the dense tree canopy above, casting dramatic, shifting spotlights and deep shadows that heighten the tension. The camera stays close to the ground, emphasizing the creatures' sheer speed and predatory single-mindedness, while subtle motion blur and speed lines emphasize the swift movements, creating a visually stunning and high-stakes atmosphere filled with a naturalistic yet dreamlike struggle between life and death.", + "seed": 16561, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_11/index_4715_seed_16561_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0053511701291427015, + "vlm_logprob": -0.0053511701291427015, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.0053511701291427015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003911583684384823 + ], + "opposite_logprob_tokens": [ + -5.75391149520874 + ], + "logprob_reward": -0.003911583684384823, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003911583684384823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12203166633844376, + 0.04816003888845444, + 0.15843747556209564, + 0.08643665909767151, + 0.22128243744373322, + 0.19624102115631104, + 0.044429417699575424, + 0.05614137649536133, + 0.04931560531258583, + 0.017524318769574165 + ], + "credit_map": [ + 0.01174829714000225, + 0.018048057332634926, + 0.015609221532940865, + 0.011083059012889862, + 0.011056588962674141, + 0.022590478882193565, + 0.013097904622554779, + 0.01879805326461792, + 0.004022079054266214, + 0.005373951513320208, + 0.010357128456234932, + 0.0045224823988974094, + 0.0028555940371006727, + 0.0054268729873001575, + 0.005394143518060446, + 0.010207785293459892, + 0.007823864929378033, + 0.011279208585619926, + 0.04428384453058243, + 0.016227861866354942, + 0.010184301063418388, + 0.013444045558571815, + 0.008593250066041946, + 0.04660109058022499, + 0.006616789847612381, + 0.00544852577149868, + 0.016463974490761757, + 0.01063243206590414, + 0.005292558576911688, + 0.013279207050800323, + 0.009045528247952461, + 0.019657636061310768, + 0.019459011033177376, + 0.017329219728708267, + 0.037059102207422256, + 0.019649947062134743, + 0.010683181695640087, + 0.062012992799282074, + 0.02095690183341503, + 0.03413206711411476, + 0.027344724163413048, + 0.022987721487879753, + 0.03951207920908928, + 0.019739454612135887, + 0.036373645067214966, + 0.014408893883228302, + 0.021198280155658722, + 0.014676208607852459, + 0.00683700293302536, + 0.0036432151682674885, + 0.0037511170376092196, + 0.00439749751240015, + 0.005754407029598951, + 0.006989311892539263, + 0.007036880124360323, + 0.006019980646669865, + 0.007629431784152985, + 0.008789093233644962, + 0.0059990962035954, + 0.008924823254346848, + 0.004245202522724867, + 0.005734323523938656, + 0.007161544170230627, + 0.007657854352146387, + 0.006052019540220499, + 0.005183863919228315, + 0.006718359421938658, + 0.006269190926104784, + 0.004298848565667868, + 0.007854628376662731, + 0.0069342139177024364, + 0.006004479248076677, + 0.002538650529459119, + 0.002224612981081009, + 0.0023098427336663008, + 0.0021289698779582977, + 0.0018388277385383844, + 0.0024711377918720245, + 0.0018940838053822517, + 0.002118191681802273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.003911583684384823 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790497832000256 + ], + "opposite_logprob_tokens": [ + -4.634790420532227 + ], + "logprob_reward": -0.009790497832000256, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790497832000256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08601292222738266, + 0.038289401680231094, + 0.2101481407880783, + 0.10660199075937271, + 0.11142570525407791, + 0.18772266805171967, + 0.07543768733739853, + 0.04959394782781601, + 0.11450493335723877, + 0.020262643694877625 + ], + "credit_map": [ + 0.006215366069227457, + 0.01018165610730648, + 0.008881598711013794, + 0.006583936512470245, + 0.010153418406844139, + 0.018587876111268997, + 0.008390766568481922, + 0.017018305137753487, + 0.002452082931995392, + 0.003944260999560356, + 0.006683379411697388, + 0.0022630863822996616, + 0.0034980906639248133, + 0.004130977671593428, + 0.005118002649396658, + 0.01019951980561018, + 0.013944488018751144, + 0.019466813653707504, + 0.05298401415348053, + 0.021810006350278854, + 0.01605393923819065, + 0.014957479201257229, + 0.014607138931751251, + 0.05632426217198372, + 0.007028729654848576, + 0.006898115389049053, + 0.021576810628175735, + 0.011673715896904469, + 0.006287758704274893, + 0.017187049612402916, + 0.009841745719313622, + 0.026108065620064735, + 0.009598251432180405, + 0.00957573764026165, + 0.01951582543551922, + 0.009060092270374298, + 0.006020164117217064, + 0.03247619420289993, + 0.009171593002974987, + 0.01600785367190838, + 0.020974840968847275, + 0.029118824750185013, + 0.03174934536218643, + 0.024806486442685127, + 0.025248020887374878, + 0.019054224714636803, + 0.02152836136519909, + 0.015242558903992176, + 0.012304599396884441, + 0.005075851920992136, + 0.006305410526692867, + 0.006591848563402891, + 0.007794368080794811, + 0.012232341803610325, + 0.013366486877202988, + 0.011766783893108368, + 0.006092892028391361, + 0.006946542300283909, + 0.006506406236439943, + 0.006271112244576216, + 0.003097804496064782, + 0.007429840508848429, + 0.006264801137149334, + 0.00698454724624753, + 0.013109543360769749, + 0.010369266383349895, + 0.0131102679297328, + 0.013194753788411617, + 0.008342103101313114, + 0.018999557942152023, + 0.021786075085401535, + 0.015593362040817738, + 0.0024299381766468287, + 0.001897681737318635, + 0.0025487139355391264, + 0.0034756248351186514, + 0.0014757466269657016, + 0.0028913705609738827, + 0.0029084074776619673, + 0.002635160693898797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.009790497832000256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784476831555367 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784476831555367, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784476831555367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07884660363197327, + 0.04201945289969444, + 0.09086210280656815, + 0.10375712811946869, + 0.3854987919330597, + 0.06629561632871628, + 0.06800822168588638, + 0.03508855029940605, + 0.1011284738779068, + 0.028495024889707565 + ], + "credit_map": [ + 0.0073504033498466015, + 0.013201097957789898, + 0.01035529375076294, + 0.007821742445230484, + 0.01029381062835455, + 0.0097089484333992, + 0.011026646010577679, + 0.009088673628866673, + 0.0023189398925751448, + 0.004671583883464336, + 0.006205581594258547, + 0.003561874385923147, + 0.0033372915349900723, + 0.0053758141584694386, + 0.00715389009565115, + 0.009394480846822262, + 0.006939687766134739, + 0.005903233308345079, + 0.020629214122891426, + 0.009852520190179348, + 0.007214623503386974, + 0.009946657344698906, + 0.007180652115494013, + 0.023195533081889153, + 0.007615544367581606, + 0.006960292346775532, + 0.022529596462845802, + 0.009976865723729134, + 0.0063556162640452385, + 0.017293062061071396, + 0.009596257470548153, + 0.023429911583662033, + 0.02899305336177349, + 0.028957244008779526, + 0.068862184882164, + 0.030352676287293434, + 0.020417669788002968, + 0.10991170257329941, + 0.039980340749025345, + 0.05802394449710846, + 0.00764312781393528, + 0.00688213761895895, + 0.01008667703717947, + 0.005470458883792162, + 0.009291669353842735, + 0.010723398998379707, + 0.009075416252017021, + 0.007122731301933527, + 0.009851188398897648, + 0.009378000162541866, + 0.006043982692062855, + 0.007936256006360054, + 0.006940322928130627, + 0.008110231719911098, + 0.011562184430658817, + 0.008186058141291142, + 0.004038259852677584, + 0.003362659364938736, + 0.006001739762723446, + 0.003031262895092368, + 0.0024357314687222242, + 0.007384417112916708, + 0.003546431427821517, + 0.005288050044327974, + 0.009956194087862968, + 0.006319782696664333, + 0.015763936564326286, + 0.010213903151452541, + 0.006770357955247164, + 0.016997603699564934, + 0.018864193931221962, + 0.016242505982518196, + 0.004629236180335283, + 0.0028731331694871187, + 0.0038797021843492985, + 0.0029995443765074015, + 0.0024923530872911215, + 0.004412509500980377, + 0.0035296983551234007, + 0.0036788496654480696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.009784476831555367 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019357530400156975 + ], + "opposite_logprob_tokens": [ + -6.251935958862305 + ], + "logprob_reward": -0.0019357530400156975, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019357530400156975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16000321507453918, + 0.07596295326948166, + 0.1692028045654297, + 0.05568710342049599, + 0.16002300381660461, + 0.07132860273122787, + 0.08334526419639587, + 0.13399215042591095, + 0.057609956711530685, + 0.03284495323896408 + ], + "credit_map": [ + 0.008936119265854359, + 0.01648140326142311, + 0.018736088648438454, + 0.01435199473053217, + 0.01737302727997303, + 0.03079080395400524, + 0.01985025592148304, + 0.03348352760076523, + 0.005127331707626581, + 0.007819294929504395, + 0.013828146271407604, + 0.008089146576821804, + 0.006271846126765013, + 0.009850548580288887, + 0.010887560434639454, + 0.01408908236771822, + 0.00959453359246254, + 0.010654892772436142, + 0.04773932695388794, + 0.017117351293563843, + 0.01052907109260559, + 0.013637728989124298, + 0.00980909913778305, + 0.05012081190943718, + 0.0043917251750826836, + 0.004303210414946079, + 0.009397556073963642, + 0.007123188115656376, + 0.0040491060353815556, + 0.013709907419979572, + 0.005121621768921614, + 0.007590793073177338, + 0.025543322786688805, + 0.01826709322631359, + 0.025035100057721138, + 0.01858152076601982, + 0.01267511211335659, + 0.0200814176350832, + 0.02202432230114937, + 0.017815126106142998, + 0.007902287878096104, + 0.006932270713150501, + 0.008286802098155022, + 0.00935329869389534, + 0.007049159612506628, + 0.011462980881333351, + 0.012965063564479351, + 0.007376743946224451, + 0.012528080493211746, + 0.005961974151432514, + 0.00782718975096941, + 0.010422672145068645, + 0.00660072872415185, + 0.012315248139202595, + 0.016691133379936218, + 0.01099824346601963, + 0.014920972287654877, + 0.009919309988617897, + 0.02969447150826454, + 0.0076486533507704735, + 0.008510574698448181, + 0.029444079846143723, + 0.014116020873188972, + 0.01973806880414486, + 0.006481542717665434, + 0.004733086563646793, + 0.012816445901989937, + 0.005342741962522268, + 0.0042297388426959515, + 0.008752942085266113, + 0.0062942542135715485, + 0.008959205821156502, + 0.003952404484152794, + 0.0036391981411725283, + 0.005636565387248993, + 0.003508336842060089, + 0.002883167704567313, + 0.004552882164716721, + 0.004231792874634266, + 0.004440607503056526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.0019357530400156975 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013335392577573657 + ], + "opposite_logprob_tokens": [ + -6.626333713531494 + ], + "logprob_reward": -0.0013335392577573657, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013335392577573657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13998572528362274, + 0.04006990045309067, + 0.15338879823684692, + 0.06878580152988434, + 0.19715915620326996, + 0.07477057725191116, + 0.14493972063064575, + 0.07476642727851868, + 0.06370813399553299, + 0.04242578148841858 + ], + "credit_map": [ + 0.009397466666996479, + 0.011494451202452183, + 0.01758734881877899, + 0.014447759836912155, + 0.014310702681541443, + 0.030707674100995064, + 0.016518279910087585, + 0.025522030889987946, + 0.005098642315715551, + 0.006091848015785217, + 0.00554852606728673, + 0.003718240186572075, + 0.003824828425422311, + 0.004724257625639439, + 0.004380870144814253, + 0.006682688370347023, + 0.006266366690397263, + 0.005485336296260357, + 0.05318571627140045, + 0.009646495804190636, + 0.004865758121013641, + 0.008012516424059868, + 0.008204723708331585, + 0.05772188678383827, + 0.005267865490168333, + 0.005272650625556707, + 0.014076083898544312, + 0.007176567800343037, + 0.004718204028904438, + 0.011009188368916512, + 0.006248972378671169, + 0.015016268007457256, + 0.0178658626973629, + 0.012103123590350151, + 0.033858489245176315, + 0.01602787710726261, + 0.010463614016771317, + 0.05404462292790413, + 0.021507758647203445, + 0.03128780797123909, + 0.009657724760472775, + 0.009062061086297035, + 0.008925085887312889, + 0.009507541544735432, + 0.0071698147803545, + 0.013290933333337307, + 0.009112624451518059, + 0.008044792339205742, + 0.01903866045176983, + 0.008847001940011978, + 0.013820682652294636, + 0.016359737142920494, + 0.010958320461213589, + 0.02253676950931549, + 0.031843025237321854, + 0.021535523235797882, + 0.010063810274004936, + 0.009754438884556293, + 0.009148695506155491, + 0.012839418835937977, + 0.00515164528042078, + 0.0064057521522045135, + 0.011257154867053032, + 0.010145511478185654, + 0.006252999417483807, + 0.00724255433306098, + 0.008667961694300175, + 0.007003962993621826, + 0.004027420189231634, + 0.01067116018384695, + 0.010646064765751362, + 0.00919601321220398, + 0.005775484722107649, + 0.006651860661804676, + 0.004766061902046204, + 0.0067411623895168304, + 0.003454248420894146, + 0.004919956438243389, + 0.0047866771928966045, + 0.005330329295247793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.0013335392577573657 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, low-angle, cinematic tracking shot, rendered in a vibrant anime style, dynamically follows a dramatic chase across the mossy forest floor. The pursuer, a massive, centipede-like creature, is stunningly depicted. Its segmented body is encased in a vibrant reddish-brown carapace, its chitinous sheen reflecting the dappled forest light. A multitude of sharp, electric-yellow legs tumbles in a coordinated, undulating blur, scurrying at terrifying speed over fuzzy emerald-green moss and decaying leaves. It relentlessly pursues a much smaller insect, an iridescent beetle, its delicate, translucent wings beating frantically as it desperately zigzags between gnarled tree roots that, from this low angle, resemble massive canyon walls. Spotted sunlight penetrates the dense tree canopy above, casting dramatic, shifting spotlights and deep shadows that heighten the tension. The camera stays close to the ground, emphasizing the creatures' sheer speed and predatory single-mindedness, while subtle motion blur and speed lines emphasize the swift movements, creating a visually stunning and high-stakes atmosphere filled with a naturalistic yet dreamlike struggle between life and death.", + "seed": 16562, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_11/index_4715_seed_16562_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.003950726706534624, + "vlm_logprob": -0.003950726706534624, + "vlm_yes_no_margin": 5.875, + "vlm_token_credit": -0.003950726706534624, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0018549631349742413 + ], + "opposite_logprob_tokens": [ + -6.87685489654541 + ], + "logprob_reward": -0.0018549631349742413, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0018549631349742413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2151622325181961, + 0.06938111037015915, + 0.08075137436389923, + 0.07069208472967148, + 0.04908460006117821, + 0.05060983821749687, + 0.06100985035300255, + 0.04833017662167549, + 0.1515049785375595, + 0.2034737765789032 + ], + "credit_map": [ + 0.015630420297384262, + 0.014178430661559105, + 0.03307799994945526, + 0.031014706939458847, + 0.031218010932207108, + 0.023253997787833214, + 0.03495577722787857, + 0.03183288499712944, + 0.007931297644972801, + 0.005005293525755405, + 0.00872459914535284, + 0.008569912984967232, + 0.005878429859876633, + 0.008316715247929096, + 0.009604561142623425, + 0.015350301750004292, + 0.008618704974651337, + 0.006909300573170185, + 0.010022865608334541, + 0.009511887095868587, + 0.011798212304711342, + 0.008648477494716644, + 0.01588556356728077, + 0.009356364607810974, + 0.009480498731136322, + 0.006719094235450029, + 0.008880983106791973, + 0.013274840079247952, + 0.008409839123487473, + 0.006303728092461824, + 0.00959850288927555, + 0.00802459754049778, + 0.006215816363692284, + 0.006018890533596277, + 0.004891202785074711, + 0.006716161035001278, + 0.007539116311818361, + 0.0051002781838178635, + 0.006392765790224075, + 0.0062103718519210815, + 0.006299606058746576, + 0.005297686904668808, + 0.0054639531299471855, + 0.005064260680228472, + 0.006701268721371889, + 0.006653218530118465, + 0.008259622380137444, + 0.006870220880955458, + 0.006665833760052919, + 0.0034537361934781075, + 0.004527943208813667, + 0.004344834014773369, + 0.005549395922571421, + 0.008062350563704967, + 0.015666114166378975, + 0.012739642523229122, + 0.004843052942305803, + 0.0052084238268435, + 0.0069054788909852505, + 0.006993803195655346, + 0.006458700634539127, + 0.004702593665570021, + 0.006843884941190481, + 0.006374236661940813, + 0.013115662150084972, + 0.021849127486348152, + 0.013336105272173882, + 0.01756276935338974, + 0.017149658873677254, + 0.0342116616666317, + 0.018574180081486702, + 0.015705807134509087, + 0.0199639480561018, + 0.021293405443429947, + 0.031628385186195374, + 0.026859961450099945, + 0.036858223378658295, + 0.022703038528561592, + 0.021042469888925552, + 0.023124342784285545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0018549631349742413 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645840920507908 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008645840920507908, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645840920507908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.138699010014534, + 0.04882483929395676, + 0.05964132025837898, + 0.025725968182086945, + 0.032501544803380966, + 0.02026238478720188, + 0.18856418132781982, + 0.01632428728044033, + 0.3887315094470978, + 0.08072494715452194 + ], + "credit_map": [ + 0.010232244618237019, + 0.007637624628841877, + 0.023368854075670242, + 0.022093098610639572, + 0.011694683693349361, + 0.020482676103711128, + 0.02480771206319332, + 0.018382104113698006, + 0.0045273201540112495, + 0.004225111100822687, + 0.0053015211597085, + 0.006047764327377081, + 0.008262896910309792, + 0.0053023966029286385, + 0.006942087318748236, + 0.008215741254389286, + 0.006754443049430847, + 0.006780008785426617, + 0.00687094870954752, + 0.0060300277546048164, + 0.009003601036965847, + 0.009101304225623608, + 0.00825818907469511, + 0.006842794828116894, + 0.0034661919344216585, + 0.0020687952637672424, + 0.003176341764628887, + 0.0036136163398623466, + 0.0027073209639638662, + 0.002452227985486388, + 0.0037584982346743345, + 0.004482977092266083, + 0.00477893790230155, + 0.00317234150134027, + 0.003295950824394822, + 0.005391189828515053, + 0.004461066797375679, + 0.0032288527581840754, + 0.0034140876960009336, + 0.004759117495268583, + 0.003154381411150098, + 0.0020546673331409693, + 0.002925961511209607, + 0.001756289741024375, + 0.002737764734774828, + 0.0019215388456359506, + 0.0029258395079523325, + 0.0027859406545758247, + 0.009865372441709042, + 0.0052711134776473045, + 0.009026026353240013, + 0.00805289763957262, + 0.007749346550554037, + 0.012123104184865952, + 0.07805689424276352, + 0.05841943249106407, + 0.0015446323668584228, + 0.0015540318563580513, + 0.0020862261299043894, + 0.0016787733184173703, + 0.0032965843565762043, + 0.0014444254338741302, + 0.002243291586637497, + 0.002476321766152978, + 0.03055611066520214, + 0.03251423314213753, + 0.039937250316143036, + 0.03100975602865219, + 0.04569337144494057, + 0.12196280062198639, + 0.04723358154296875, + 0.03982442617416382, + 0.007104246411472559, + 0.00886570941656828, + 0.011536364443600178, + 0.009074331261217594, + 0.014788948930799961, + 0.008961418643593788, + 0.010019049048423767, + 0.010374873876571655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.008645840920507908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257826764136553 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257826764136553, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257826764136553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16362497210502625, + 0.20173753798007965, + 0.09767867624759674, + 0.06392080336809158, + 0.023382563143968582, + 0.02804327942430973, + 0.09654843062162399, + 0.01933199353516102, + 0.2176477164030075, + 0.08808403462171555 + ], + "credit_map": [ + 0.013349431566894054, + 0.010552458465099335, + 0.02712552435696125, + 0.02943168208003044, + 0.016548868268728256, + 0.016421934589743614, + 0.024058284237980843, + 0.026136774569749832, + 0.020123207941651344, + 0.011809426359832287, + 0.02332328073680401, + 0.031127167865633965, + 0.0273688156157732, + 0.01846807636320591, + 0.050413936376571655, + 0.01910361833870411, + 0.01268948707729578, + 0.007312825880944729, + 0.011533073149621487, + 0.009776199236512184, + 0.014030379243195057, + 0.015209555625915527, + 0.016151010990142822, + 0.010976135730743408, + 0.009421875700354576, + 0.008968845009803772, + 0.009278330951929092, + 0.010286572389304638, + 0.008399025537073612, + 0.007555053569376469, + 0.004897001199424267, + 0.005114090628921986, + 0.0029889713041484356, + 0.0027658489998430014, + 0.002186625264585018, + 0.002814768347889185, + 0.004084030166268349, + 0.0029565885197371244, + 0.002623877953737974, + 0.00296184909529984, + 0.0028855782002210617, + 0.003432030789554119, + 0.0031656783539801836, + 0.0033838581293821335, + 0.003336455440148711, + 0.0035083023831248283, + 0.004656106699258089, + 0.003675266867503524, + 0.00819411315023899, + 0.005697589833289385, + 0.007522311061620712, + 0.008784390054643154, + 0.006502162199467421, + 0.008833101950585842, + 0.028545716777443886, + 0.022469045594334602, + 0.002438971772789955, + 0.0015037738485261798, + 0.002839877735823393, + 0.0025609189178794622, + 0.00271560437977314, + 0.0022142906673252583, + 0.002425124868750572, + 0.002633430529385805, + 0.016704795882105827, + 0.019128946587443352, + 0.01928265579044819, + 0.01697247289121151, + 0.024222156032919884, + 0.0718604326248169, + 0.026347054168581963, + 0.02312919683754444, + 0.00664459029212594, + 0.00941008422523737, + 0.00826966855674982, + 0.010903041809797287, + 0.017231419682502747, + 0.011942850425839424, + 0.011361073702573776, + 0.012321295216679573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.005257826764136553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024845702573657036 + ], + "opposite_logprob_tokens": [ + -6.0024847984313965 + ], + "logprob_reward": -0.0024845702573657036, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024845702573657036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2738698422908783, + 0.1612061709165573, + 0.05670490860939026, + 0.08250772953033447, + 0.14209294319152832, + 0.05679088085889816, + 0.06319274753332138, + 0.03980725258588791, + 0.07701555639505386, + 0.04681198298931122 + ], + "credit_map": [ + 0.019895434379577637, + 0.01523573324084282, + 0.04753800109028816, + 0.04263710230588913, + 0.024593865498900414, + 0.042914409190416336, + 0.04310516268014908, + 0.037950124591588974, + 0.0160248801112175, + 0.013538588769733906, + 0.021700670942664146, + 0.026374321430921555, + 0.01850944757461548, + 0.023582084104418755, + 0.022096943110227585, + 0.01937924325466156, + 0.006090300157666206, + 0.004767334554344416, + 0.010636509396135807, + 0.008455042727291584, + 0.007790598552674055, + 0.006490418221801519, + 0.00565625773742795, + 0.006818452384322882, + 0.011982831172645092, + 0.00962880626320839, + 0.0093244519084692, + 0.01111972238868475, + 0.015406826511025429, + 0.009822843596339226, + 0.007951094768941402, + 0.007271156646311283, + 0.01546043436974287, + 0.016632473096251488, + 0.01771448366343975, + 0.018957318738102913, + 0.012801985256373882, + 0.01135749276727438, + 0.029103970155119896, + 0.020064791664481163, + 0.007000319194048643, + 0.008838463574647903, + 0.006385229527950287, + 0.005164750386029482, + 0.005690092220902443, + 0.007089280057698488, + 0.008576438762247562, + 0.00804630946367979, + 0.006332214921712875, + 0.005610508844256401, + 0.008025278337299824, + 0.006642699707299471, + 0.0048292893916368484, + 0.01414098497480154, + 0.008726137690246105, + 0.008885633200407028, + 0.003988795448094606, + 0.003122789552435279, + 0.0059474315494298935, + 0.005451061762869358, + 0.005200160201638937, + 0.004832109436392784, + 0.004931922536343336, + 0.006332981400191784, + 0.0063662901520729065, + 0.006400836631655693, + 0.008224012330174446, + 0.006397820543497801, + 0.009551734663546085, + 0.02272365801036358, + 0.0097046522423625, + 0.00764655414968729, + 0.0038737624417990446, + 0.005861842073500156, + 0.004888321738690138, + 0.006754777394235134, + 0.007767318282276392, + 0.006171223241835833, + 0.005896728951483965, + 0.005598008166998625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0024845702573657036 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive centipede-like creature with a reddish-brown segmented body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the creature's legs electric-yellow and depicted as a coordinated, undulating blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the creature chasing a smaller iridescent beetle with translucent wings that is zigzagging between tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a mossy forest floor with fuzzy emerald-green moss and decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle cinematic tracking shot that keeps the camera close to the ground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001510432455688715 + ], + "opposite_logprob_tokens": [ + -6.5015106201171875 + ], + "logprob_reward": -0.001510432455688715, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001510432455688715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1340574026107788, + 0.25431111454963684, + 0.0475788488984108, + 0.06075480952858925, + 0.0362008772790432, + 0.03332224488258362, + 0.257578581571579, + 0.0252771507948637, + 0.10091618448495865, + 0.05000278353691101 + ], + "credit_map": [ + 0.013384501449763775, + 0.009623019024729729, + 0.022070137783885002, + 0.02256009541451931, + 0.013441618531942368, + 0.01744673028588295, + 0.01710520312190056, + 0.018426088616251945, + 0.025614837184548378, + 0.022931968793272972, + 0.03319002315402031, + 0.03678647056221962, + 0.028930064290761948, + 0.039215851575136185, + 0.03726564720273018, + 0.030376240611076355, + 0.0034763459116220474, + 0.004293304868042469, + 0.007884050719439983, + 0.006755970884114504, + 0.008272375911474228, + 0.005273777991533279, + 0.005653157830238342, + 0.005969862919300795, + 0.006221514195203781, + 0.006097456905990839, + 0.0077982801012694836, + 0.013185023330152035, + 0.00615053391084075, + 0.005037636961787939, + 0.008373604156076908, + 0.007890760898590088, + 0.006664075888693333, + 0.005330581218004227, + 0.0025494624860584736, + 0.004172361455857754, + 0.004287182819098234, + 0.002983368467539549, + 0.0037374887615442276, + 0.0064763533882796764, + 0.003519704332575202, + 0.004823815543204546, + 0.004593479912728071, + 0.003708659904077649, + 0.0037722205743193626, + 0.0039800237864255905, + 0.00508821289986372, + 0.003836126532405615, + 0.011439074762165546, + 0.006640464998781681, + 0.014085708186030388, + 0.01015239767730236, + 0.010052508674561977, + 0.020639711990952492, + 0.10540927201509476, + 0.07915943115949631, + 0.0024220687337219715, + 0.002548761432990432, + 0.002640676451846957, + 0.0025612888857722282, + 0.004422906320542097, + 0.0029247025959193707, + 0.0041898926720023155, + 0.003566853469237685, + 0.013972509652376175, + 0.005875915288925171, + 0.01655626855790615, + 0.013079543597996235, + 0.007621712516993284, + 0.014376791194081306, + 0.01650470308959484, + 0.012928741984069347, + 0.00474235275760293, + 0.005080580711364746, + 0.007417186629027128, + 0.005807117093354464, + 0.007108970079571009, + 0.00589493615552783, + 0.007839955389499664, + 0.006111685186624527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.001510432455688715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_15/index_tensor([95, 95])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_15/index_tensor([95, 95])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c1cbaf0caa1efd300b5933f1b2e18aaef4737bc8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_15/index_tensor([95, 95])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot begins at a low, ground-level angle, then slowly pulls back and rises slightly, capturing a field of ripe, golden wheat. In the foreground, several heavy ears of wheat are magnified to the utmost, each grain appearing golden and plump, gracefully bent by its own weight, their dry surfaces revealing rich textured details in the light. A gentle breeze sweeps by, and these ears of wheat, leading the endless waves of wheat behind them, begin to sway rhythmically, rising and falling. The entire field resembles a golden ocean, creating layers of gentle, majestic waves that stretch to the horizon where the azure sky meets. The warm, soft evening sun slants across the scene, glaring the entire image with a rich amber hue. The tips of the wheat awns shimmer with dazzling light in the backlight, and fine particles of light seem to float in the air. The scene creates an atmosphere of tranquility, abundance, and vitality, a silent celebration of the bountiful harvest.", + "seed": 21125, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_15/index_95_seed_21125_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04537575845024548, + "vlm_logprob": -0.04537575845024548, + "vlm_yes_no_margin": 4.724999904632568, + "vlm_token_credit": -0.04537575845024548, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029237205162644386 + ], + "opposite_logprob_tokens": [ + -4.029237270355225 + ], + "logprob_reward": -0.029237205162644386, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.029237205162644386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09083252400159836, + 0.031242525205016136, + 0.11549675464630127, + 0.04536385089159012, + 0.046040285378694534, + 0.19059351086616516, + 0.05873429402709007, + 0.1323229968547821, + 0.27108800411224365, + 0.018285254016518593 + ], + "credit_map": [ + 0.005839705001562834, + 0.007527660112828016, + 0.022808237001299858, + 0.010876396670937538, + 0.012834385968744755, + 0.007690537720918655, + 0.01041796337813139, + 0.01283762976527214, + 0.0022705262526869774, + 0.0038726588245481253, + 0.004674640018492937, + 0.00498043280094862, + 0.003960596397519112, + 0.004506220575422049, + 0.002614768221974373, + 0.0043626814149320126, + 0.011004033498466015, + 0.013352491892874241, + 0.007749220822006464, + 0.00587864313274622, + 0.04668978229165077, + 0.008753827773034573, + 0.004961400292813778, + 0.017107361927628517, + 0.0032418572809547186, + 0.0022518434561789036, + 0.015834996476769447, + 0.008713972754776478, + 0.003731965087354183, + 0.001895564142614603, + 0.004445096477866173, + 0.005248556844890118, + 0.002957763848826289, + 0.014740620739758015, + 0.002597934566438198, + 0.01138184405863285, + 0.0059826308861374855, + 0.0023483752738684416, + 0.0019528795965015888, + 0.00407823920249939, + 0.01700201816856861, + 0.06402555108070374, + 0.01338143553584814, + 0.007236000150442123, + 0.01781991496682167, + 0.010138217359781265, + 0.013068840838968754, + 0.047921523451805115, + 0.006030337885022163, + 0.008672144263982773, + 0.00680658919736743, + 0.007600044831633568, + 0.007099466398358345, + 0.009074431844055653, + 0.0052704522386193275, + 0.008180828765034676, + 0.0155337518081069, + 0.015286204405128956, + 0.015639523044228554, + 0.015409606508910656, + 0.01569407247006893, + 0.016244471073150635, + 0.023992439731955528, + 0.01452292688190937, + 0.03120385855436325, + 0.04766549542546272, + 0.036527324467897415, + 0.046558406203985214, + 0.01987117901444435, + 0.02459198608994484, + 0.023802390322089195, + 0.040867384523153305, + 0.0013398414012044668, + 0.0018186134984716773, + 0.0026627597399055958, + 0.0037391206715255976, + 0.0013835523277521133, + 0.002018567407503724, + 0.0017124414443969727, + 0.0036103578750044107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.029237205162644386 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421516202390194 + ], + "opposite_logprob_tokens": [ + -4.264214992523193 + ], + "logprob_reward": -0.01421516202390194, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421516202390194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10619871318340302, + 0.02648172527551651, + 0.2616904675960541, + 0.04691706597805023, + 0.03971012309193611, + 0.2998621165752411, + 0.10184525698423386, + 0.05494772270321846, + 0.042543865740299225, + 0.019802916795015335 + ], + "credit_map": [ + 0.006936775520443916, + 0.007611926645040512, + 0.025492029264569283, + 0.011726978234946728, + 0.011714352294802666, + 0.009378471411764622, + 0.019563617184758186, + 0.013774563558399677, + 0.0020306939259171486, + 0.0031328515615314245, + 0.004340586252510548, + 0.004531116224825382, + 0.002305446658283472, + 0.003450250718742609, + 0.002370565664023161, + 0.004320215899497271, + 0.017505494877696037, + 0.014074609614908695, + 0.03774024918675423, + 0.017441526055336, + 0.05869676172733307, + 0.013987899757921696, + 0.014822485856711864, + 0.0874214693903923, + 0.003279539057984948, + 0.0022675087675452232, + 0.016431624069809914, + 0.008823801763355732, + 0.0036943794693797827, + 0.0021576792933046818, + 0.004211075603961945, + 0.006051459349691868, + 0.0029520308598876, + 0.010002542287111282, + 0.0026364929508417845, + 0.008859035558998585, + 0.004841968417167664, + 0.0038461070507764816, + 0.0019155563786625862, + 0.004656392615288496, + 0.026706106960773468, + 0.09663994610309601, + 0.022304611280560493, + 0.012631496414542198, + 0.02605205960571766, + 0.016347099095582962, + 0.022320402786135674, + 0.07686040550470352, + 0.01001915242522955, + 0.012494664639234543, + 0.014429578557610512, + 0.012693366035819054, + 0.010658657178282738, + 0.009229379706084728, + 0.013847716152668, + 0.018472740426659584, + 0.005919954273849726, + 0.006359837017953396, + 0.0065595549531280994, + 0.005753866396844387, + 0.007778482977300882, + 0.006187974940985441, + 0.010190093889832497, + 0.006197957322001457, + 0.004685245919972658, + 0.006854902487248182, + 0.005863843485713005, + 0.006189052015542984, + 0.0036023769062012434, + 0.004827897530049086, + 0.003333867061883211, + 0.007186680566519499, + 0.0025791043881326914, + 0.0018141388427466154, + 0.00263025751337409, + 0.003720667213201523, + 0.00206861924380064, + 0.0022026796359568834, + 0.0018359344685450196, + 0.002951514208689332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.01421516202390194 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003493413969408721 + ], + "opposite_logprob_tokens": [ + -8.000349044799805 + ], + "logprob_reward": -0.0003493413969408721, + "yes_no_margin_reward": 7.999999523162842, + "token_credit_reward": -0.0003493413969408721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06691205501556396, + 0.057518962770700455, + 0.4532344937324524, + 0.0355452224612236, + 0.048881448805332184, + 0.18281865119934082, + 0.024651888757944107, + 0.0162269938737154, + 0.09885792434215546, + 0.015352331101894379 + ], + "credit_map": [ + 0.004303947556763887, + 0.0029115895740687847, + 0.019225874915719032, + 0.005957567598670721, + 0.011328529566526413, + 0.003982821013778448, + 0.006023784633725882, + 0.013177949003875256, + 0.004361364059150219, + 0.008309785276651382, + 0.006206832360476255, + 0.009535234421491623, + 0.006750958971679211, + 0.008450419642031193, + 0.005233976058661938, + 0.008670399896800518, + 0.026384875178337097, + 0.03474532067775726, + 0.05846809968352318, + 0.02441537380218506, + 0.13595189154148102, + 0.026419678702950478, + 0.021576805040240288, + 0.1252724677324295, + 0.0025252175983041525, + 0.0031060941983014345, + 0.009816629812121391, + 0.005849029403179884, + 0.005027042701840401, + 0.0028989450074732304, + 0.0024971673265099525, + 0.0038250957150012255, + 0.003944074735045433, + 0.018758568912744522, + 0.002235386986285448, + 0.006881894078105688, + 0.007200147490948439, + 0.0027214479632675648, + 0.0021794233471155167, + 0.004960506688803434, + 0.01612413115799427, + 0.06223888322710991, + 0.012317025102674961, + 0.006263983901590109, + 0.017409054562449455, + 0.009589708410203457, + 0.01287777628749609, + 0.04599810764193535, + 0.0022080636117607355, + 0.00315466639585793, + 0.002991327317431569, + 0.003412543563172221, + 0.0025391126982867718, + 0.0036110857035964727, + 0.002361443592235446, + 0.004373645875602961, + 0.001635496853850782, + 0.0016833548434078693, + 0.002457320922985673, + 0.002009218791499734, + 0.001491851406171918, + 0.0018141571199521422, + 0.0031532554421573877, + 0.001982338959351182, + 0.011518658138811588, + 0.015362905338406563, + 0.011638950556516647, + 0.019364187493920326, + 0.007020088378340006, + 0.009222892113029957, + 0.009416061453521252, + 0.015314191579818726, + 0.0017284348141402006, + 0.0017840963555499911, + 0.0017297841841354966, + 0.003069167258217931, + 0.0013368272921070457, + 0.0016470064874738455, + 0.001543730846606195, + 0.0025132852606475353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.0003493413969408721 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988677322864532 + ], + "opposite_logprob_tokens": [ + -1.804886817932129 + ], + "logprob_reward": -0.17988677322864532, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988677322864532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08448545634746552, + 0.016797102987766266, + 0.043634261935949326, + 0.05540075898170471, + 0.013451757840812206, + 0.5829809904098511, + 0.03284100443124771, + 0.049839749932289124, + 0.09020493924617767, + 0.030364060774445534 + ], + "credit_map": [ + 0.006137391086667776, + 0.004218106623739004, + 0.02436659298837185, + 0.007296827621757984, + 0.014819573611021042, + 0.006393012590706348, + 0.012613200582563877, + 0.008640747517347336, + 0.0011571255745366216, + 0.001576733193360269, + 0.0032888527493923903, + 0.0026214448735117912, + 0.002133219037204981, + 0.0015094607369974256, + 0.0016899387119337916, + 0.0028203283436596394, + 0.004025529604405165, + 0.0021976972930133343, + 0.009563078172504902, + 0.003984641749411821, + 0.0019572856836020947, + 0.0030177158769220114, + 0.002787836128845811, + 0.016100479289889336, + 0.0033063122536987066, + 0.004188850522041321, + 0.01645808108150959, + 0.009699743241071701, + 0.00967872142791748, + 0.003282213816419244, + 0.0036819728557020426, + 0.005104859825223684, + 0.0010417293524369597, + 0.0025945028755813837, + 0.0019340140279382467, + 0.0018958623986691236, + 0.0013195432256907225, + 0.0011125601595267653, + 0.0011870736489072442, + 0.0023664722684770823, + 0.05205667391419411, + 0.19410008192062378, + 0.04074134677648544, + 0.020168036222457886, + 0.053211312741041183, + 0.03160719946026802, + 0.0427577905356884, + 0.14833852648735046, + 0.003530327696353197, + 0.004151481669396162, + 0.004645089153200388, + 0.0037603946402668953, + 0.0035569234751164913, + 0.0033183384221047163, + 0.003789161331951618, + 0.006089288741350174, + 0.004658218938857317, + 0.005422517657279968, + 0.005142012611031532, + 0.005333134438842535, + 0.005261718761175871, + 0.007898341864347458, + 0.010245412588119507, + 0.005878392606973648, + 0.010384284891188145, + 0.018967391923069954, + 0.010397979989647865, + 0.01478993333876133, + 0.007033413741737604, + 0.009143978357315063, + 0.0072634522803127766, + 0.012224502861499786, + 0.002505610464140773, + 0.002692781388759613, + 0.0043295761570334435, + 0.004157213028520346, + 0.0058043478056788445, + 0.003903089789673686, + 0.002795922104269266, + 0.0041755205020308495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.17988677322864532 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190310439094901 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.003190310439094901, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190310439094901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07864285260438919, + 0.023686854168772697, + 0.35801035165786743, + 0.0811106488108635, + 0.04388715326786041, + 0.2241850197315216, + 0.029191453009843826, + 0.038935694843530655, + 0.08276297897100449, + 0.03958702087402344 + ], + "credit_map": [ + 0.005237617529928684, + 0.004162285476922989, + 0.02640216425061226, + 0.005787363741546869, + 0.013429365120828152, + 0.0041482895612716675, + 0.007411427330225706, + 0.012064342387020588, + 0.0018835054943338037, + 0.003946973942220211, + 0.0028829327784478664, + 0.0033905375748872757, + 0.003229324473068118, + 0.0030364578124135733, + 0.002093364717438817, + 0.003223756793886423, + 0.025204339995980263, + 0.035330209881067276, + 0.03469272330403328, + 0.017633872106671333, + 0.12736383080482483, + 0.023849280551075935, + 0.01539420336484909, + 0.07854186743497849, + 0.004744970239698887, + 0.005931281019002199, + 0.024910150095820427, + 0.014298518188297749, + 0.015228877775371075, + 0.004426288418471813, + 0.005011335946619511, + 0.006559227593243122, + 0.00360690476372838, + 0.014194861985743046, + 0.002326118992641568, + 0.004525146912783384, + 0.005739716347306967, + 0.0034200057853013277, + 0.003040648065507412, + 0.007033750414848328, + 0.019551550969481468, + 0.07590775191783905, + 0.014390809461474419, + 0.007524214219301939, + 0.021944526582956314, + 0.012056308798491955, + 0.015921223908662796, + 0.056888628751039505, + 0.003097225446254015, + 0.0033535854890942574, + 0.003702044952660799, + 0.003694895189255476, + 0.0039235432632267475, + 0.002911515301093459, + 0.0033942721784114838, + 0.005114372819662094, + 0.003885894315317273, + 0.004615870304405689, + 0.004455635789781809, + 0.0042252372950315475, + 0.004777821246534586, + 0.005673076957464218, + 0.006860898807644844, + 0.004441262222826481, + 0.009604177437722683, + 0.01215914636850357, + 0.009679628536105156, + 0.014817526564002037, + 0.005842484999448061, + 0.00890924222767353, + 0.007654169574379921, + 0.01409660279750824, + 0.004314537160098553, + 0.003392931539565325, + 0.007773429621011019, + 0.007134087383747101, + 0.004263470880687237, + 0.0036516778636723757, + 0.0030172900296747684, + 0.006039598025381565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.003190310439094901 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins at a low, ground-level angle, then slowly pulls back and rises slightly, capturing a field of ripe, golden wheat. In the foreground, several heavy ears of wheat are magnified to the utmost, each grain appearing golden and plump, gracefully bent by its own weight, their dry surfaces revealing rich textured details in the light. A gentle breeze sweeps by, and these ears of wheat, leading the endless waves of wheat behind them, begin to sway rhythmically, rising and falling. The entire field resembles a golden ocean, creating layers of gentle, majestic waves that stretch to the horizon where the azure sky meets. The warm, soft evening sun slants across the scene, glaring the entire image with a rich amber hue. The tips of the wheat awns shimmer with dazzling light in the backlight, and fine particles of light seem to float in the air. The scene creates an atmosphere of tranquility, abundance, and vitality, a silent celebration of the bountiful harvest.", + "seed": 21126, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_15/index_95_seed_21126_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02028998915629927, + "vlm_logprob": -0.02028998915629927, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.02028998915629927, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00886670034378767 + ], + "opposite_logprob_tokens": [ + -5.633866786956787 + ], + "logprob_reward": -0.00886670034378767, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00886670034378767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19168917834758759, + 0.25325316190719604, + 0.04813304543495178, + 0.06350511312484741, + 0.25080814957618713, + 0.028711384162306786, + 0.027859043329954147, + 0.05346754938364029, + 0.03780258446931839, + 0.04477078095078468 + ], + "credit_map": [ + 0.009376810863614082, + 0.018736427649855614, + 0.058499742299318314, + 0.06495551764965057, + 0.01086464710533619, + 0.005945310462266207, + 0.015439390204846859, + 0.007871335372328758, + 0.03093506209552288, + 0.039062514901161194, + 0.03184735029935837, + 0.07642263174057007, + 0.009583123959600925, + 0.012509318068623543, + 0.022222811356186867, + 0.030670352280139923, + 0.004384978208690882, + 0.005721999332308769, + 0.005134737119078636, + 0.010059362277388573, + 0.004408066160976887, + 0.005404411815106869, + 0.0074869925156235695, + 0.00553249754011631, + 0.005897228606045246, + 0.007217069156467915, + 0.009308229200541973, + 0.009069644846022129, + 0.0074467845261096954, + 0.006169039290398359, + 0.012582707218825817, + 0.005814408417791128, + 0.0337614081799984, + 0.02432214468717575, + 0.036459628492593765, + 0.03376976028084755, + 0.016276642680168152, + 0.015905480831861496, + 0.03838912025094032, + 0.0519239604473114, + 0.003201419021934271, + 0.00292807980440557, + 0.004758948925882578, + 0.003650667844340205, + 0.0023249343503266573, + 0.0026445258408784866, + 0.005931515246629715, + 0.0032712940592318773, + 0.0021326676942408085, + 0.002620323561131954, + 0.005492737516760826, + 0.00405700271949172, + 0.002323910826817155, + 0.0024306310806423426, + 0.004391993395984173, + 0.00440977793186903, + 0.003858485957607627, + 0.004188516177237034, + 0.006518069189041853, + 0.012470771558582783, + 0.0034202621318399906, + 0.004479067400097847, + 0.00900974590331316, + 0.009522632695734501, + 0.0020681233145296574, + 0.0064774323254823685, + 0.006049189250916243, + 0.004286215174943209, + 0.006111849565058947, + 0.003178717801347375, + 0.005188908893615007, + 0.004442145582288504, + 0.005953595507889986, + 0.0037431346718221903, + 0.00678081251680851, + 0.006797807291150093, + 0.00398059980943799, + 0.005216314923018217, + 0.0048147039487957954, + 0.007483812980353832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.00886670034378767 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354903355240822 + ], + "opposite_logprob_tokens": [ + -3.6513547897338867 + ], + "logprob_reward": -0.026354903355240822, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354903355240822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18180719017982483, + 0.159567728638649, + 0.05481552705168724, + 0.03582380339503288, + 0.10280460864305496, + 0.03819435462355614, + 0.0411672480404377, + 0.1387467086315155, + 0.17035476863384247, + 0.0767180547118187 + ], + "credit_map": [ + 0.013230135664343834, + 0.021332375705242157, + 0.044944263994693756, + 0.05251007527112961, + 0.012868991121649742, + 0.009831803850829601, + 0.01639670692384243, + 0.010692853480577469, + 0.010732604190707207, + 0.03234153985977173, + 0.01880778558552265, + 0.047929905354976654, + 0.011016963981091976, + 0.010231065563857555, + 0.011913090944290161, + 0.016594775021076202, + 0.00688687851652503, + 0.009749527089297771, + 0.005337370093911886, + 0.008173689246177673, + 0.0042669810354709625, + 0.005680944304913282, + 0.007798768579959869, + 0.006921372376382351, + 0.004298051819205284, + 0.006364168133586645, + 0.004012934863567352, + 0.005116937682032585, + 0.0035869982093572617, + 0.0037774895317852497, + 0.004870332777500153, + 0.0037968880496919155, + 0.010437794961035252, + 0.010808818973600864, + 0.013570845127105713, + 0.015280614607036114, + 0.009857477620244026, + 0.01043775025755167, + 0.010953043587505817, + 0.021458256989717484, + 0.005181869026273489, + 0.004630505573004484, + 0.005978830624371767, + 0.0044844006188213825, + 0.0029750890098512173, + 0.004259834531694651, + 0.006871253252029419, + 0.003812575014308095, + 0.003273983718827367, + 0.003620975650846958, + 0.004877422470599413, + 0.0060747163370251656, + 0.0050040544010698795, + 0.006091385614126921, + 0.007170800119638443, + 0.005053909029811621, + 0.011674252338707447, + 0.009128603152930737, + 0.017276030033826828, + 0.034921795129776, + 0.008093387819826603, + 0.012013535015285015, + 0.02337755262851715, + 0.022261539474129677, + 0.009004754014313221, + 0.024782640859484673, + 0.01415812037885189, + 0.014863334596157074, + 0.013874630443751812, + 0.011822286061942577, + 0.02916976995766163, + 0.05267924815416336, + 0.008822483010590076, + 0.0101212989538908, + 0.00838429108262062, + 0.007348076906055212, + 0.009524292312562466, + 0.011112787760794163, + 0.009145884774625301, + 0.012258944101631641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.026354903355240822 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00021217002358753234 + ], + "opposite_logprob_tokens": [ + -8.500211715698242 + ], + "logprob_reward": -0.00021217002358753234, + "yes_no_margin_reward": 8.5, + "token_credit_reward": -0.00021217002358753234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26276031136512756, + 0.3294493854045868, + 0.03938714414834976, + 0.02196984551846981, + 0.1046908050775528, + 0.011270487681031227, + 0.03396386653184891, + 0.08345923572778702, + 0.07473188638687134, + 0.0383169986307621 + ], + "credit_map": [ + 0.01134426798671484, + 0.01573329232633114, + 0.09419199824333191, + 0.09406865388154984, + 0.008676976896822453, + 0.007342054042965174, + 0.019647331908345222, + 0.011755760759115219, + 0.02363617531955242, + 0.055468954145908356, + 0.04375116899609566, + 0.11913367360830307, + 0.011069647036492825, + 0.009123735129833221, + 0.02691848762333393, + 0.04034755006432533, + 0.0032325845677405596, + 0.006402940489351749, + 0.003893346292898059, + 0.007596207782626152, + 0.003255401737987995, + 0.0042166258208453655, + 0.006294221617281437, + 0.004495818633586168, + 0.003302321070805192, + 0.003673272207379341, + 0.0025104002561420202, + 0.0028771907091140747, + 0.001675273641012609, + 0.002213905332610011, + 0.003701093839481473, + 0.002016389975324273, + 0.013222568668425083, + 0.01009835209697485, + 0.013979297131299973, + 0.014789871871471405, + 0.008886451832950115, + 0.008476235903799534, + 0.01201435923576355, + 0.023223675787448883, + 0.0010162506951019168, + 0.0015281678643077612, + 0.002041031140834093, + 0.001404919777996838, + 0.0007714981911703944, + 0.0012020058929920197, + 0.002004360780119896, + 0.001302253920584917, + 0.002694482449442148, + 0.0033105395268648863, + 0.005056993570178747, + 0.004320850595831871, + 0.0026839543133974075, + 0.0037459833547472954, + 0.006155993789434433, + 0.005995072424411774, + 0.006822013296186924, + 0.0076365042477846146, + 0.009683015756309032, + 0.02006293088197708, + 0.005849200766533613, + 0.008803005330264568, + 0.015582002699375153, + 0.009020576253533363, + 0.005370893515646458, + 0.010358725674450397, + 0.00688394857570529, + 0.005256779957562685, + 0.007400698494166136, + 0.009412392973899841, + 0.012633876875042915, + 0.017414573580026627, + 0.004926260095089674, + 0.004435769747942686, + 0.005245521664619446, + 0.0045968638733029366, + 0.0036843097768723965, + 0.00559229077771306, + 0.004155880305916071, + 0.0056801061145961285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.00021217002358753234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494673550128937 + ], + "opposite_logprob_tokens": [ + -2.9299466609954834 + ], + "logprob_reward": -0.05494673550128937, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494673550128937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07358642667531967, + 0.22925424575805664, + 0.04854803904891014, + 0.10816989094018936, + 0.09475330263376236, + 0.021797602996230125, + 0.05044195428490639, + 0.25495144724845886, + 0.09104987233877182, + 0.027447205036878586 + ], + "credit_map": [ + 0.007459240034222603, + 0.009874152950942516, + 0.011669504456222057, + 0.010675078257918358, + 0.00906447321176529, + 0.0064652240835130215, + 0.008555459789931774, + 0.009823297150433064, + 0.015077590942382812, + 0.04184507951140404, + 0.02771654911339283, + 0.0775298997759819, + 0.010479521937668324, + 0.010851463302969933, + 0.01863463595509529, + 0.027119504287838936, + 0.004819404799491167, + 0.008955912664532661, + 0.0038962692487984896, + 0.010855388827621937, + 0.0035555509384721518, + 0.004796416964381933, + 0.005852419883012772, + 0.005816679447889328, + 0.008119447156786919, + 0.00821553822606802, + 0.014032531529664993, + 0.015201800502836704, + 0.01222368236631155, + 0.00968865118920803, + 0.028167881071567535, + 0.01252035889774561, + 0.011455067433416843, + 0.009753350168466568, + 0.013871930539608002, + 0.011564334854483604, + 0.006268606521189213, + 0.00584174320101738, + 0.01676337420940399, + 0.01923489011824131, + 0.0027628184761852026, + 0.002564364345744252, + 0.002847107592970133, + 0.002625280525535345, + 0.0031387601047754288, + 0.0025894269347190857, + 0.002676315838471055, + 0.002593530109152198, + 0.005564676132053137, + 0.005967562086880207, + 0.007022321689873934, + 0.007124241441488266, + 0.005645877681672573, + 0.006591469515115023, + 0.005530714523047209, + 0.006995094008743763, + 0.022333769127726555, + 0.02175869606435299, + 0.02838602289557457, + 0.05586060881614685, + 0.017243215814232826, + 0.021180573850870132, + 0.05091254040598869, + 0.0372760035097599, + 0.004184403456747532, + 0.012000689283013344, + 0.011220450513064861, + 0.008497076109051704, + 0.005902405362576246, + 0.004749521147459745, + 0.014233596622943878, + 0.030261732637882233, + 0.003445191541686654, + 0.004886104725301266, + 0.0036637033335864544, + 0.0028828654903918505, + 0.003539029508829117, + 0.0027041048742830753, + 0.002871252130717039, + 0.0034549543634057045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.05494673550128937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069436557590961 + ], + "opposite_logprob_tokens": [ + -4.511069297790527 + ], + "logprob_reward": -0.011069436557590961, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069436557590961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.249123677611351, + 0.1241544783115387, + 0.03330273553729057, + 0.08909095078706741, + 0.19107301533222198, + 0.04890480637550354, + 0.03376103565096855, + 0.06311087310314178, + 0.05874039605259895, + 0.10873793065547943 + ], + "credit_map": [ + 0.02757330983877182, + 0.03940514102578163, + 0.04387805610895157, + 0.0452493354678154, + 0.02430226281285286, + 0.02246118150651455, + 0.022969165816903114, + 0.0232852715998888, + 0.01682359166443348, + 0.021379917860031128, + 0.01204692292958498, + 0.0315755158662796, + 0.006241823546588421, + 0.010523565113544464, + 0.009935510344803333, + 0.01562766171991825, + 0.0041482821106910706, + 0.004687066655606031, + 0.004511482547968626, + 0.005094532854855061, + 0.0031283877324312925, + 0.003996788524091244, + 0.004144984297454357, + 0.0035912147723138332, + 0.01504204049706459, + 0.016719752922654152, + 0.009743163362145424, + 0.01149855088442564, + 0.008097676560282707, + 0.00817213486880064, + 0.011503324843943119, + 0.008314318023622036, + 0.0319083034992218, + 0.016858579590916634, + 0.031965117901563644, + 0.02429245598614216, + 0.010639447718858719, + 0.014299456030130386, + 0.016932278871536255, + 0.04417741298675537, + 0.006490926258265972, + 0.0051701064221560955, + 0.008168141357600689, + 0.0077646654099226, + 0.004401867743581533, + 0.00445201201364398, + 0.007549186237156391, + 0.004907912574708462, + 0.004624510649591684, + 0.003921284340322018, + 0.0046843248419463634, + 0.003657406195998192, + 0.0027967598289251328, + 0.003190231742337346, + 0.003969836048781872, + 0.006916687358170748, + 0.006457394454628229, + 0.0050575812347233295, + 0.007454517297446728, + 0.012632385827600956, + 0.0037974214646965265, + 0.005428399424999952, + 0.010234430432319641, + 0.01204875111579895, + 0.004257471300661564, + 0.0034534346777945757, + 0.008117273449897766, + 0.006444561295211315, + 0.004794334061443806, + 0.005172034725546837, + 0.0103958984836936, + 0.016105394810438156, + 0.014469578862190247, + 0.013393159955739975, + 0.013631515204906464, + 0.012165352702140808, + 0.016470657661557198, + 0.013638835400342941, + 0.010146125219762325, + 0.014822719618678093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.011069436557590961 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins at a low, ground-level angle, then slowly pulls back and rises slightly, capturing a field of ripe, golden wheat. In the foreground, several heavy ears of wheat are magnified to the utmost, each grain appearing golden and plump, gracefully bent by its own weight, their dry surfaces revealing rich textured details in the light. A gentle breeze sweeps by, and these ears of wheat, leading the endless waves of wheat behind them, begin to sway rhythmically, rising and falling. The entire field resembles a golden ocean, creating layers of gentle, majestic waves that stretch to the horizon where the azure sky meets. The warm, soft evening sun slants across the scene, glaring the entire image with a rich amber hue. The tips of the wheat awns shimmer with dazzling light in the backlight, and fine particles of light seem to float in the air. The scene creates an atmosphere of tranquility, abundance, and vitality, a silent celebration of the bountiful harvest.", + "seed": 21127, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_15/index_95_seed_21127_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016770420875400305, + "vlm_logprob": -0.016770420875400305, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.016770420875400305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013860991224646568 + ], + "opposite_logprob_tokens": [ + -5.513861179351807 + ], + "logprob_reward": -0.013860991224646568, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.013860991224646568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09357668459415436, + 0.062132418155670166, + 0.019187889993190765, + 0.017138822004199028, + 0.5060484409332275, + 0.03529889136552811, + 0.0665905773639679, + 0.04448894038796425, + 0.038256172090768814, + 0.11728128045797348 + ], + "credit_map": [ + 0.0064608873799443245, + 0.010186462663114071, + 0.017401734367012978, + 0.013936403207480907, + 0.01071038469672203, + 0.005795618519186974, + 0.016806580126285553, + 0.012278605252504349, + 0.0031601502560079098, + 0.00338608305901289, + 0.00465727923437953, + 0.017118623480200768, + 0.002272728132084012, + 0.0030266970861703157, + 0.008187164552509785, + 0.02032369002699852, + 0.0015538186999037862, + 0.0016239366959780455, + 0.003292427398264408, + 0.003249747445806861, + 0.0019938929472118616, + 0.0018116686260327697, + 0.0034412979148328304, + 0.00222109816968441, + 0.0014789270935580134, + 0.0010465405648574233, + 0.003811520291492343, + 0.00299114384688437, + 0.0010959886712953448, + 0.0019389991648495197, + 0.0022034470457583666, + 0.002572252182289958, + 0.021833786740899086, + 0.034015554934740067, + 0.03018168918788433, + 0.19100536406040192, + 0.008767453022301197, + 0.012598257511854172, + 0.025343064218759537, + 0.18230320513248444, + 0.0032761686015874147, + 0.003684258321300149, + 0.004943098872900009, + 0.006458591669797897, + 0.0031256908550858498, + 0.003574093570932746, + 0.005598438438028097, + 0.004638546612113714, + 0.004118495620787144, + 0.005437527317553759, + 0.0067328051663935184, + 0.006815304979681969, + 0.013210793025791645, + 0.003563585691154003, + 0.02094542793929577, + 0.0057666306383907795, + 0.0038092220202088356, + 0.0036575528793036938, + 0.0074628181755542755, + 0.006325855385512114, + 0.005835367366671562, + 0.005124697461724281, + 0.005792871583253145, + 0.0064805541187524796, + 0.002553202910348773, + 0.002795255510136485, + 0.005365239456295967, + 0.0049838656559586525, + 0.0053209527395665646, + 0.005564325023442507, + 0.005356610752642155, + 0.006316716317087412, + 0.01141532976180315, + 0.01923339255154133, + 0.007862460799515247, + 0.00925036147236824, + 0.020211365073919296, + 0.026653097942471504, + 0.009657426737248898, + 0.012997833080589771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.013860991224646568 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264822859317064 + ], + "opposite_logprob_tokens": [ + -5.255264759063721 + ], + "logprob_reward": -0.005264822859317064, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264822859317064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03929557278752327, + 0.19043470919132233, + 0.0592113733291626, + 0.03712712228298187, + 0.2827405333518982, + 0.04860849678516388, + 0.050096433609724045, + 0.06238573044538498, + 0.03762652352452278, + 0.19247350096702576 + ], + "credit_map": [ + 0.0032076328061521053, + 0.003289544954895973, + 0.005759734660387039, + 0.006717917043715715, + 0.003393398132175207, + 0.004551846068352461, + 0.005068110767751932, + 0.007307387888431549, + 0.007182965986430645, + 0.008145864121615887, + 0.013699375092983246, + 0.06464168429374695, + 0.0037206958513706923, + 0.007859387435019016, + 0.019698450341820717, + 0.06548626720905304, + 0.00600472604855895, + 0.003979475237429142, + 0.013137091882526875, + 0.011045289225876331, + 0.00621330039575696, + 0.0056172567419707775, + 0.0068630799651145935, + 0.006351157557219267, + 0.0039000045508146286, + 0.0035654425155371428, + 0.008883979171514511, + 0.006312661804258823, + 0.0022779915016144514, + 0.003759177168831229, + 0.004488883540034294, + 0.0039389790035784245, + 0.012273802421987057, + 0.019353099167346954, + 0.016704145818948746, + 0.10557804256677628, + 0.005221947096288204, + 0.007102629169821739, + 0.014764374122023582, + 0.10174249857664108, + 0.0038913788739591837, + 0.006569516845047474, + 0.006050500087440014, + 0.008648871444165707, + 0.004662081133574247, + 0.005781400483101606, + 0.006774695590138435, + 0.006230051629245281, + 0.006713171489536762, + 0.007832901552319527, + 0.005421563051640987, + 0.005792639218270779, + 0.007209517993032932, + 0.004964040592312813, + 0.005286381114274263, + 0.00687621533870697, + 0.005508206784725189, + 0.008958213031291962, + 0.007124339696019888, + 0.008485331200063229, + 0.007639196701347828, + 0.008779093623161316, + 0.00893083680421114, + 0.0069605098105967045, + 0.0038446765393018723, + 0.007260242477059364, + 0.0025049946270883083, + 0.004156999755650759, + 0.00773968268185854, + 0.006650265771895647, + 0.002669493667781353, + 0.002800165908411145, + 0.019841911271214485, + 0.03973892703652382, + 0.013436838984489441, + 0.0172352846711874, + 0.033724069595336914, + 0.034663472324609756, + 0.013481422327458858, + 0.020351581275463104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.005264822859317064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004996005445718765 + ], + "opposite_logprob_tokens": [ + -7.625499725341797 + ], + "logprob_reward": -0.0004996005445718765, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0004996005445718765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12082289159297943, + 0.0411103330552578, + 0.04443887248635292, + 0.06118857488036156, + 0.047156624495983124, + 0.0314374603331089, + 0.3607756197452545, + 0.06501980125904083, + 0.04628203064203262, + 0.18176774680614471 + ], + "credit_map": [ + 0.00729777105152607, + 0.010933993384242058, + 0.023453157395124435, + 0.014969817362725735, + 0.01315494254231453, + 0.00716034322977066, + 0.022959692403674126, + 0.02089318074285984, + 0.002889445284381509, + 0.0026192227378487587, + 0.003890102729201317, + 0.01011898647993803, + 0.0021378768142312765, + 0.0024390334729105234, + 0.004506751429289579, + 0.012508911080658436, + 0.004835265222936869, + 0.004715261049568653, + 0.006616159807890654, + 0.006928652990609407, + 0.0037705351132899523, + 0.004225017037242651, + 0.007405173499137163, + 0.005942806601524353, + 0.00496795354411006, + 0.0042732167057693005, + 0.015147214755415916, + 0.011532269418239594, + 0.0037035210989415646, + 0.004784180782735348, + 0.009518303908407688, + 0.007261916529387236, + 0.006335369776934385, + 0.005646376870572567, + 0.006025572773069143, + 0.006427478976547718, + 0.005230630747973919, + 0.005902206525206566, + 0.005474332720041275, + 0.006114656571298838, + 0.0021956050768494606, + 0.003269857494160533, + 0.003670166712254286, + 0.005352930631488562, + 0.00337198912166059, + 0.0057067410089075565, + 0.00380658358335495, + 0.004063587635755539, + 0.020152797922492027, + 0.03021078370511532, + 0.034541837871074677, + 0.035564567893743515, + 0.07626378536224365, + 0.01941368728876114, + 0.11383713036775589, + 0.03079104609787464, + 0.005685707088559866, + 0.007318427320569754, + 0.008795535191893578, + 0.009413951076567173, + 0.007230300456285477, + 0.0072683957405388355, + 0.011957132257521152, + 0.0073503595776855946, + 0.0040073273703455925, + 0.004155958071351051, + 0.005086174700409174, + 0.005747929681092501, + 0.0056765214540064335, + 0.005642288830131292, + 0.009021174162626266, + 0.006944658700376749, + 0.017310230061411858, + 0.03201492503285408, + 0.012739886529743671, + 0.015569331124424934, + 0.03185999020934105, + 0.037850845605134964, + 0.014035838656127453, + 0.020386699587106705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0004996005445718765 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203147768974304 + ], + "opposite_logprob_tokens": [ + -2.8120315074920654 + ], + "logprob_reward": -0.06203147768974304, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203147768974304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06252004206180573, + 0.3386610448360443, + 0.06657196581363678, + 0.02398686297237873, + 0.05454712733626366, + 0.01910633035004139, + 0.15943603217601776, + 0.023310458287596703, + 0.05541696026921272, + 0.19644315540790558 + ], + "credit_map": [ + 0.005304085556417704, + 0.0052448599599301815, + 0.009720614179968834, + 0.010015418753027916, + 0.0058267065323889256, + 0.0034362489823251963, + 0.00902719795703888, + 0.013944928534328938, + 0.012510761618614197, + 0.012274261564016342, + 0.02388804219663143, + 0.11425159126520157, + 0.006563616916537285, + 0.014139430597424507, + 0.03486592322587967, + 0.12016742676496506, + 0.007343098986893892, + 0.007617028895765543, + 0.01165524497628212, + 0.010799224488437176, + 0.007143337279558182, + 0.005766322836279869, + 0.00728621007874608, + 0.008961507119238377, + 0.0019573913887143135, + 0.0017556784441694617, + 0.004211220424622297, + 0.0039517139084637165, + 0.0018058110726997256, + 0.00281509873457253, + 0.0039549837820231915, + 0.003534967079758644, + 0.004187650978565216, + 0.0036044572480022907, + 0.006320360582321882, + 0.01173268910497427, + 0.004169633146375418, + 0.0037835335824638605, + 0.005600569769740105, + 0.015148240141570568, + 0.001602046424522996, + 0.0022904262878000736, + 0.003190521150827408, + 0.0024295030161738396, + 0.001835110830143094, + 0.0033274719025939703, + 0.002196213696151972, + 0.0022350384388118982, + 0.008480259217321873, + 0.014698386192321777, + 0.014566655270755291, + 0.015022261999547482, + 0.03608720377087593, + 0.009375995956361294, + 0.04859181493520737, + 0.012613466940820217, + 0.0023173135705292225, + 0.0027036748360842466, + 0.003525384236127138, + 0.004568623844534159, + 0.002722395583987236, + 0.0023738741874694824, + 0.0029632053337991238, + 0.0021359894890338182, + 0.003968187142163515, + 0.006224791053682566, + 0.005994083359837532, + 0.008273879066109657, + 0.007514434400945902, + 0.006283518858253956, + 0.008414308540523052, + 0.008743769489228725, + 0.019692223519086838, + 0.038136597722768784, + 0.013390063308179379, + 0.016726108267903328, + 0.0343376100063324, + 0.038282569497823715, + 0.01432306319475174, + 0.021554945036768913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.06203147768974304 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195212058722973 + ], + "opposite_logprob_tokens": [ + -6.127195358276367 + ], + "logprob_reward": -0.002195212058722973, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195212058722973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10759402811527252, + 0.11210447549819946, + 0.050377652049064636, + 0.026088790968060493, + 0.11272340267896652, + 0.09058473259210587, + 0.3149312436580658, + 0.026195459067821503, + 0.043363139033317566, + 0.1160370483994484 + ], + "credit_map": [ + 0.008516918867826462, + 0.010464278049767017, + 0.016140280291438103, + 0.016536740586161613, + 0.015605772845447063, + 0.010914121754467487, + 0.015056515112519264, + 0.01435940247029066, + 0.005364356562495232, + 0.00836656428873539, + 0.009533414617180824, + 0.03757519647479057, + 0.0042852982878685, + 0.007040433585643768, + 0.012020932510495186, + 0.027918288484215736, + 0.005137632135301828, + 0.007317844312638044, + 0.007223042659461498, + 0.0076211728155612946, + 0.005035131238400936, + 0.004328147042542696, + 0.006054025609046221, + 0.007660658098757267, + 0.002525572432205081, + 0.0026152695063501596, + 0.004555624444037676, + 0.005350323859602213, + 0.0017095202347263694, + 0.001996039180085063, + 0.004091266542673111, + 0.0032451741863042116, + 0.011014383286237717, + 0.007265312131494284, + 0.012848255224525928, + 0.02412991225719452, + 0.011746651493012905, + 0.011077520437538624, + 0.013290716335177422, + 0.021350648254156113, + 0.00799110159277916, + 0.008843932300806046, + 0.016468524932861328, + 0.014658709987998009, + 0.007109533064067364, + 0.01088277529925108, + 0.011013304814696312, + 0.013616853393614292, + 0.019249236211180687, + 0.03349793702363968, + 0.026961946859955788, + 0.028031498193740845, + 0.07384885847568512, + 0.019698798656463623, + 0.08658960461616516, + 0.02705334685742855, + 0.0027386383153498173, + 0.004008632153272629, + 0.004122976679354906, + 0.003061301074922085, + 0.0027711489237844944, + 0.0027627237141132355, + 0.003943918272852898, + 0.002786120167002082, + 0.005000164732336998, + 0.010522262193262577, + 0.0029862099327147007, + 0.0044061848893761635, + 0.005906531121581793, + 0.006646741181612015, + 0.0037408648058772087, + 0.004154175985604525, + 0.01239748764783144, + 0.015100942924618721, + 0.008348489180207253, + 0.008452784270048141, + 0.01996808685362339, + 0.028727304190397263, + 0.012137500569224358, + 0.010904454626142979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.002195212058722973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins at a low, ground-level angle, then slowly pulls back and rises slightly, capturing a field of ripe, golden wheat. In the foreground, several heavy ears of wheat are magnified to the utmost, each grain appearing golden and plump, gracefully bent by its own weight, their dry surfaces revealing rich textured details in the light. A gentle breeze sweeps by, and these ears of wheat, leading the endless waves of wheat behind them, begin to sway rhythmically, rising and falling. The entire field resembles a golden ocean, creating layers of gentle, majestic waves that stretch to the horizon where the azure sky meets. The warm, soft evening sun slants across the scene, glaring the entire image with a rich amber hue. The tips of the wheat awns shimmer with dazzling light in the backlight, and fine particles of light seem to float in the air. The scene creates an atmosphere of tranquility, abundance, and vitality, a silent celebration of the bountiful harvest.", + "seed": 21128, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_15/index_95_seed_21128_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019975972606334835, + "vlm_logprob": -0.019975972606334835, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.019975972606334835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02678615413606167 + ], + "opposite_logprob_tokens": [ + -4.0267863273620605 + ], + "logprob_reward": -0.02678615413606167, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02678615413606167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09192869812250137, + 0.02669915370643139, + 0.03032621182501316, + 0.35965320467948914, + 0.052085358649492264, + 0.09634161740541458, + 0.03323478624224663, + 0.05665199086070061, + 0.23472702503204346, + 0.018351944163441658 + ], + "credit_map": [ + 0.009843739680945873, + 0.006710289046168327, + 0.01817670650780201, + 0.012992947362363338, + 0.011302989907562733, + 0.0063092634081840515, + 0.016026034951210022, + 0.010566730052232742, + 0.0020693165715783834, + 0.002433391520753503, + 0.004292181693017483, + 0.006298284512013197, + 0.0021157064475119114, + 0.0018927034689113498, + 0.0042528994381427765, + 0.003344670869410038, + 0.002507115015760064, + 0.0032727026846259832, + 0.005047294311225414, + 0.0049724034033715725, + 0.002231036312878132, + 0.0040369704365730286, + 0.004048067145049572, + 0.004210624378174543, + 0.01651756651699543, + 0.011253848671913147, + 0.1236899122595787, + 0.031289637088775635, + 0.01607978716492653, + 0.015173162333667278, + 0.0193626768887043, + 0.1262866109609604, + 0.008505488745868206, + 0.005040489137172699, + 0.006172616966068745, + 0.00664543779566884, + 0.005620546173304319, + 0.006109341513365507, + 0.0074974228627979755, + 0.006494016852229834, + 0.0073304492980241776, + 0.007932883687317371, + 0.013174904510378838, + 0.023455042392015457, + 0.006955814082175493, + 0.009843237698078156, + 0.01703517511487007, + 0.010614115744829178, + 0.005819957237690687, + 0.0032030523288995028, + 0.005148338153958321, + 0.0036052283830940723, + 0.0024382546544075012, + 0.004957353696227074, + 0.0029465951956808567, + 0.005116008222103119, + 0.010050958953797817, + 0.009626979939639568, + 0.00518938759341836, + 0.005987103097140789, + 0.005523025523871183, + 0.008705763146281242, + 0.005389539059251547, + 0.006179237738251686, + 0.04007550701498985, + 0.08075910806655884, + 0.019386079162359238, + 0.011420106515288353, + 0.02285517007112503, + 0.020467521622776985, + 0.007206020411103964, + 0.03255751356482506, + 0.0037077884189784527, + 0.002222727984189987, + 0.002199807669967413, + 0.002302404260262847, + 0.0013807419454678893, + 0.002628298243507743, + 0.0017035732744261622, + 0.002206602832302451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.02678615413606167 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556896544992924 + ], + "opposite_logprob_tokens": [ + -4.387557029724121 + ], + "logprob_reward": -0.012556896544992924, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556896544992924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015327977016568184, + 0.013323829509317875, + 0.005903969053179026, + 0.7225779294967651, + 0.006098894402384758, + 0.008520655333995819, + 0.025988779962062836, + 0.016116106882691383, + 0.18411846458911896, + 0.0020233409013599157 + ], + "credit_map": [ + 0.0012844284065067768, + 0.0008783240336924791, + 0.0032037829514592886, + 0.0024625323712825775, + 0.0017453463515266776, + 0.001197708654217422, + 0.002787258941680193, + 0.0017685953062027693, + 0.0011385256657376885, + 0.0015910754445940256, + 0.0022034449502825737, + 0.001570091932080686, + 0.0014308540849015117, + 0.0022439395543187857, + 0.0015137827722355723, + 0.001632115337997675, + 0.0005317915347404778, + 0.0007769989897496998, + 0.0010373570257797837, + 0.0008831827435642481, + 0.0005680343019776046, + 0.0007723173475824296, + 0.000680615718010813, + 0.0006536711007356644, + 0.02752249874174595, + 0.01687663234770298, + 0.26580649614334106, + 0.059015676379203796, + 0.01606176421046257, + 0.02595837041735649, + 0.03684251755475998, + 0.2744940519332886, + 0.0007013100548647344, + 0.0005823945975862443, + 0.0008362655062228441, + 0.0009554506395943463, + 0.0005232621333561838, + 0.0004545044794213027, + 0.001063677715137601, + 0.0009820291306823492, + 0.0006686497363261878, + 0.0005469037569127977, + 0.0019929702393710613, + 0.0017853950848802924, + 0.0005151065997779369, + 0.0006115518626756966, + 0.0012408545007929206, + 0.001159224077127874, + 0.0022698305547237396, + 0.003109230659902096, + 0.0052880784496665, + 0.004375218413770199, + 0.002112401882186532, + 0.002643026877194643, + 0.0027920145075768232, + 0.0033989776857197285, + 0.0024645384401082993, + 0.002614617347717285, + 0.0015621364582329988, + 0.0017891016323119402, + 0.0015129087259992957, + 0.002805886324495077, + 0.0015639049233868718, + 0.0018030130304396152, + 0.023071160539984703, + 0.06231178343296051, + 0.016126219183206558, + 0.009469470009207726, + 0.017395049333572388, + 0.014649013057351112, + 0.005751930642873049, + 0.03534384071826935, + 0.0004854114667978138, + 0.0002167916391044855, + 0.00018969048687722534, + 0.00022849359083920717, + 0.0001469277049181983, + 0.0002686817606445402, + 0.00023166253231465816, + 0.00025568174896761775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.012556896544992924 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000816131301689893 + ], + "opposite_logprob_tokens": [ + -7.125816345214844 + ], + "logprob_reward": -0.000816131301689893, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.000816131301689893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03574525937438011, + 0.016435697674751282, + 0.02518446370959282, + 0.6938989758491516, + 0.020567018538713455, + 0.018417449668049812, + 0.011866726912558079, + 0.024111753329634666, + 0.12030148506164551, + 0.03347109630703926 + ], + "credit_map": [ + 0.0039586350321769714, + 0.002905994886532426, + 0.0068452502600848675, + 0.004346049390733242, + 0.0035278520081192255, + 0.0024612152483314276, + 0.006992735899984837, + 0.0047075264155864716, + 0.0018923634197562933, + 0.002124358667060733, + 0.0022484883666038513, + 0.002911945804953575, + 0.0014065360883250833, + 0.001994436141103506, + 0.002144428202882409, + 0.001713139470666647, + 0.00168151396792382, + 0.0017002202803269029, + 0.00562282744795084, + 0.0027540281880646944, + 0.0024211520794779062, + 0.004751562140882015, + 0.002925618551671505, + 0.0033275384921580553, + 0.024474071338772774, + 0.01577780395746231, + 0.2580218017101288, + 0.05647111311554909, + 0.01254220400005579, + 0.02443663217127323, + 0.035085927695035934, + 0.26708948612213135, + 0.002188390353694558, + 0.001628309953957796, + 0.0033647343516349792, + 0.0031548456754535437, + 0.0019249533070251346, + 0.0027180842589586973, + 0.003005643142387271, + 0.0025820566806942225, + 0.0017892925534397364, + 0.0015074220718815923, + 0.003036918118596077, + 0.003480523359030485, + 0.0014056515647098422, + 0.0022633769549429417, + 0.002790495753288269, + 0.0021437681280076504, + 0.001593923196196556, + 0.0011429422302171588, + 0.0020011262968182564, + 0.0019239595858380198, + 0.0010751570807769895, + 0.001185266301035881, + 0.0014016153290867805, + 0.0015427365433424711, + 0.003982902504503727, + 0.003967999015003443, + 0.0024148139636963606, + 0.0032781308982521296, + 0.00260011269710958, + 0.002790574450045824, + 0.0021945852786302567, + 0.002882634522393346, + 0.014989782124757767, + 0.04044764116406441, + 0.010643771849572659, + 0.006593409925699234, + 0.011418341659009457, + 0.009770343080163002, + 0.0038424101658165455, + 0.022595787420868874, + 0.0033616817090660334, + 0.005385386757552624, + 0.0055296109057962894, + 0.00379046518355608, + 0.0027395340148359537, + 0.006118090357631445, + 0.0030428976751863956, + 0.0035034287720918655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.000816131301689893 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048651598393917084 + ], + "opposite_logprob_tokens": [ + -3.048651695251465 + ], + "logprob_reward": -0.048651598393917084, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048651598393917084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07882817089557648, + 0.029763836413621902, + 0.05872902274131775, + 0.1286976933479309, + 0.06739383935928345, + 0.1314883530139923, + 0.030816441401839256, + 0.07440263032913208, + 0.3842178285121918, + 0.015662172809243202 + ], + "credit_map": [ + 0.012157746590673923, + 0.00791846215724945, + 0.010002097114920616, + 0.01084848865866661, + 0.006705990992486477, + 0.007305877283215523, + 0.013782918453216553, + 0.01010658498853445, + 0.005119084380567074, + 0.0030864898581057787, + 0.003798137418925762, + 0.0036145937629044056, + 0.0029282220639288425, + 0.00423516146838665, + 0.003837529569864273, + 0.003144618822261691, + 0.00440358929336071, + 0.0044326926581561565, + 0.010145671665668488, + 0.013288558460772038, + 0.0037747721653431654, + 0.004747137427330017, + 0.009345565922558308, + 0.008591036312282085, + 0.0066889189183712006, + 0.005406627897173166, + 0.04296021908521652, + 0.01091785728931427, + 0.0063391188159585, + 0.006396569777280092, + 0.007108588702976704, + 0.04287979379296303, + 0.007129851263016462, + 0.004953073337674141, + 0.008992861025035381, + 0.011149848811328411, + 0.006134863942861557, + 0.006767552345991135, + 0.013927238993346691, + 0.008338549174368382, + 0.009760742075741291, + 0.01163798850029707, + 0.022134793922305107, + 0.0369880348443985, + 0.007444299291819334, + 0.00998512003570795, + 0.017753345891833305, + 0.015784025192260742, + 0.003384742420166731, + 0.004096307326108217, + 0.005801370367407799, + 0.005094735883176327, + 0.002405497245490551, + 0.0030931655783206224, + 0.0028630110900849104, + 0.004077611956745386, + 0.009257801808416843, + 0.011106189340353012, + 0.008092038333415985, + 0.00785217247903347, + 0.007772656157612801, + 0.015152014791965485, + 0.006569256074726582, + 0.008600492030382156, + 0.04031321033835411, + 0.11926309764385223, + 0.035895802080631256, + 0.022105908021330833, + 0.03414849936962128, + 0.029570389539003372, + 0.014741238206624985, + 0.08817965537309647, + 0.0023787678219377995, + 0.0022944207303225994, + 0.0019117422634735703, + 0.0019017943413928151, + 0.0016017898451536894, + 0.0020511059556156397, + 0.001763186533935368, + 0.0017593647353351116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.048651598393917084 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a low, ground-level camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull back and rise slightly during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are heavy ears of wheat visible in the foreground with rich textured details?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the ears of wheat sway rhythmically due to a gentle breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft evening sun creating an amber hue?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069082655012608 + ], + "opposite_logprob_tokens": [ + -4.511069297790527 + ], + "logprob_reward": -0.011069082655012608, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069082655012608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051452524960041046, + 0.05486993119120598, + 0.11953351646661758, + 0.3042069971561432, + 0.07974789291620255, + 0.04276812821626663, + 0.05663919821381569, + 0.06263691931962967, + 0.14703823626041412, + 0.08110663294792175 + ], + "credit_map": [ + 0.004786072764545679, + 0.0029553368221968412, + 0.010581535287201405, + 0.007159563712775707, + 0.00522035825997591, + 0.0044180285185575485, + 0.01065728347748518, + 0.005674346815794706, + 0.003954415675252676, + 0.0044198911637067795, + 0.00732256006449461, + 0.006296625826507807, + 0.0054998272098600864, + 0.011654450558125973, + 0.009072511456906796, + 0.006649650167673826, + 0.008326537907123566, + 0.00924464501440525, + 0.022423051297664642, + 0.013580339029431343, + 0.013318149372935295, + 0.02454979717731476, + 0.013221247121691704, + 0.014869746752083302, + 0.011902907863259315, + 0.007153435610234737, + 0.11177968978881836, + 0.024793017655611038, + 0.005787299480289221, + 0.011047868058085442, + 0.015413228422403336, + 0.11632956564426422, + 0.01057635061442852, + 0.006476265378296375, + 0.011615742929279804, + 0.010139037854969501, + 0.008594045415520668, + 0.01083176676183939, + 0.010810582898557186, + 0.010704108513891697, + 0.0032813737634569407, + 0.003753349184989929, + 0.0067202127538621426, + 0.012012169696390629, + 0.0023017541971057653, + 0.004304447211325169, + 0.00545277027413249, + 0.004942054860293865, + 0.0051393588073551655, + 0.0065596941858530045, + 0.011660892516374588, + 0.00980981346219778, + 0.0040833051316440105, + 0.006586097180843353, + 0.005273938179016113, + 0.007526095490902662, + 0.009610366076231003, + 0.009888346306979656, + 0.006468584295362234, + 0.007192525081336498, + 0.005437707994133234, + 0.010463191196322441, + 0.00519392266869545, + 0.008382278494536877, + 0.03681211546063423, + 0.027643242850899696, + 0.013675510883331299, + 0.012011428363621235, + 0.010173516348004341, + 0.016609270125627518, + 0.007877619937062263, + 0.022235531359910965, + 0.015683116391301155, + 0.013537265360355377, + 0.008547652512788773, + 0.008521046489477158, + 0.00699989777058363, + 0.012713217176496983, + 0.0067181955091655254, + 0.008386234752833843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.011069082655012608 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_2/index_tensor([1860, 1860])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_2/index_tensor([1860, 1860])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b00a82d63177f0db4ecc0c7590dca85f3f28375c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_2/index_tensor([1860, 1860])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium, eye-level shot initially frames a surreal oil painting of a magnificent floral arrangement, mounted in an ornate, dark wood frame against a dimly lit gallery wall. The painting itself depicts a vibrant bouquet, featuring lush, crimson roses with thick, thick petals, sunny yellow tulips, and deep purple irises, all set in a simple, clear glass vase that artfully refracts light. The texture of the oil painting is palpable, glossy under the spotlight focused from above. Then, the camera begins a slow, seamless advance, steadily approaching the canvas. As the frame tightens on the bouquet, a surreal transformation begins; the painted flowers begin to wilt in a time-lapse fashion. The vibrant crimson of the roses gradually fades to a dull, dusty maroon, the edges of the petals curling inward. The bright yellow of the tulips fades to a pale, sickly ochre, and their once-sturdy stems bend and droop with a palpable heaviness. The petals seem to lose their painterly structure, detaching and realistically sliding along the textured surface of the canvas, leaving faint, oily streaks in their path. The entire scene on the canvas shifts from a vibrant celebration of life to a somber depiction of decay, with the colors becoming muted and earthy as the bouquet collapses into a withered state, creating a poignant and melancholic atmosphere that contrasts sharply with its initial breathtaking beauty.", + "seed": 2114, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_2/index_1860_seed_2114_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05214891145005822, + "vlm_logprob": -0.05214891145005822, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.05214891145005822, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020160874351859093 + ], + "opposite_logprob_tokens": [ + -5.145160675048828 + ], + "logprob_reward": -0.020160874351859093, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.020160874351859093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12058115005493164, + 0.05266684293746948, + 0.035219207406044006, + 0.25000008940696716, + 0.12096726894378662, + 0.022554945200681686, + 0.2841919958591461, + 0.025644829496741295, + 0.03631383553147316, + 0.05185982584953308 + ], + "credit_map": [ + 0.029874287545681, + 0.007584479171782732, + 0.02530895546078682, + 0.0097480658441782, + 0.010815825313329697, + 0.014973612502217293, + 0.006651806645095348, + 0.01562411803752184, + 0.006062338128685951, + 0.005812720395624638, + 0.008430718444287777, + 0.004261557944118977, + 0.010311327874660492, + 0.006046295631676912, + 0.004456988535821438, + 0.007284891325980425, + 0.0037944659125059843, + 0.003205845132470131, + 0.005200237967073917, + 0.0038537343498319387, + 0.0031024652998894453, + 0.007204010616987944, + 0.002191209699958563, + 0.006667239125818014, + 0.02407793514430523, + 0.007393153849989176, + 0.08427540212869644, + 0.011239429004490376, + 0.006490261293947697, + 0.0856766402721405, + 0.0046240040101110935, + 0.026223260909318924, + 0.0038457417394965887, + 0.002438272349536419, + 0.004613256547600031, + 0.0033977054990828037, + 0.004074946511536837, + 0.05026799812912941, + 0.003021169686689973, + 0.0493081733584404, + 0.0021457509137690067, + 0.001956705469638109, + 0.004080303944647312, + 0.0027293239254504442, + 0.0020979272667318583, + 0.002553980564698577, + 0.0023591560311615467, + 0.0046317968517541885, + 0.006019956897944212, + 0.006964468397200108, + 0.011141413822770119, + 0.008727680891752243, + 0.02086961828172207, + 0.11683093011379242, + 0.007411319762468338, + 0.1062266156077385, + 0.002553728874772787, + 0.0027482123114168644, + 0.0036289128474891186, + 0.004121863283216953, + 0.002799021080136299, + 0.0038007916882634163, + 0.0030652612913399935, + 0.0029270367231220007, + 0.00329374591819942, + 0.002976699499413371, + 0.0044080354273319244, + 0.003892646636813879, + 0.0023159116972237825, + 0.00930548645555973, + 0.004266852512955666, + 0.005854458082467318, + 0.0061014574021101, + 0.005569117609411478, + 0.00760424742475152, + 0.005454891826957464, + 0.00644333939999342, + 0.00611641863361001, + 0.0064484248869121075, + 0.008121930062770844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.020160874351859093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020606130361557007 + ], + "opposite_logprob_tokens": [ + -3.89560604095459 + ], + "logprob_reward": -0.020606130361557007, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020606130361557007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06576403975486755, + 0.060340765863657, + 0.041806429624557495, + 0.12124650180339813, + 0.30697178840637207, + 0.02764696069061756, + 0.24054604768753052, + 0.0502859428524971, + 0.03537768870592117, + 0.05001378431916237 + ], + "credit_map": [ + 0.012665475718677044, + 0.004687946289777756, + 0.01437282469123602, + 0.004671941045671701, + 0.008916700258851051, + 0.008089862763881683, + 0.004539493005722761, + 0.007819796912372112, + 0.005685014650225639, + 0.00608054967597127, + 0.005928299389779568, + 0.006721890531480312, + 0.01342337392270565, + 0.005769812967628241, + 0.01084766536951065, + 0.005884163547307253, + 0.009130839258432388, + 0.0038957830984145403, + 0.009219354018568993, + 0.004702416248619556, + 0.004032823722809553, + 0.004621350206434727, + 0.0031324957963079214, + 0.0030713723972439766, + 0.013385528698563576, + 0.004301081877201796, + 0.04060022160410881, + 0.006668832153081894, + 0.004108719993382692, + 0.03735417127609253, + 0.0035948376171290874, + 0.011233111843466759, + 0.0086458008736372, + 0.005023842211812735, + 0.010587924160063267, + 0.006686433218419552, + 0.008580341935157776, + 0.13248957693576813, + 0.007092594169080257, + 0.12786531448364258, + 0.003578923176974058, + 0.002219545654952526, + 0.005092222709208727, + 0.002837558975443244, + 0.003463280852884054, + 0.0040664272382855415, + 0.002152848057448864, + 0.004236155655235052, + 0.007416021078824997, + 0.005240660160779953, + 0.009215853177011013, + 0.0070820036344230175, + 0.017781006172299385, + 0.09671546518802643, + 0.006278010550886393, + 0.09081703424453735, + 0.007672790437936783, + 0.005441433284431696, + 0.00729532353579998, + 0.007051559165120125, + 0.004682457074522972, + 0.007214048877358437, + 0.0051604523323476315, + 0.005767888389527798, + 0.0029560872353613377, + 0.0042175790295004845, + 0.004226284567266703, + 0.004315818659961224, + 0.002930620452389121, + 0.00787667091935873, + 0.0038553806953132153, + 0.004999247845262289, + 0.004632201977074146, + 0.005102113820612431, + 0.0077944411896169186, + 0.005604601930826902, + 0.005807782523334026, + 0.006340842694044113, + 0.006585683207958937, + 0.008146124891936779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.020606130361557007 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046506100334227085 + ], + "opposite_logprob_tokens": [ + -5.379650592803955 + ], + "logprob_reward": -0.0046506100334227085, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046506100334227085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022772572934627533, + 0.011799949221313, + 0.01289724837988615, + 0.243550106883049, + 0.08761373162269592, + 0.015339845791459084, + 0.5397152304649353, + 0.02063016965985298, + 0.0223503690212965, + 0.02333078719675541 + ], + "credit_map": [ + 0.0029980624094605446, + 0.00233479798771441, + 0.0039000518154352903, + 0.0025002253241837025, + 0.0034474751446396112, + 0.0024342508986592293, + 0.0019580197986215353, + 0.0031996879260987043, + 0.00130884547252208, + 0.001516793156042695, + 0.0016615347703918815, + 0.0011682909680530429, + 0.0018679476343095303, + 0.0013464823132380843, + 0.0012519138399511576, + 0.0016781395534053445, + 0.0014756666496396065, + 0.001274983398616314, + 0.0019380681915208697, + 0.0015660794451832771, + 0.0014434444019570947, + 0.001978030428290367, + 0.0012812433997169137, + 0.001939731533639133, + 0.024624677374958992, + 0.007012848276644945, + 0.08387986570596695, + 0.011185012757778168, + 0.0063985735177993774, + 0.08245282620191574, + 0.0036958851851522923, + 0.024300411343574524, + 0.002454758621752262, + 0.0011814578901976347, + 0.0030040168203413486, + 0.001930343802087009, + 0.0027131421957165003, + 0.037224624305963516, + 0.0020252789836376905, + 0.037080101668834686, + 0.0016987809212878346, + 0.001553533598780632, + 0.0026183035224676132, + 0.001886730082333088, + 0.0016288382466882467, + 0.0017511910991743207, + 0.00167821254581213, + 0.002524254610762, + 0.012820725329220295, + 0.011769970878958702, + 0.020218219608068466, + 0.011931218206882477, + 0.03943844512104988, + 0.2254558801651001, + 0.01166803389787674, + 0.20641276240348816, + 0.002706053201109171, + 0.0021254976745694876, + 0.0020724229980260134, + 0.0028198459185659885, + 0.003692569909617305, + 0.002697713440284133, + 0.0022438070736825466, + 0.002272256650030613, + 0.0023452681489288807, + 0.0037588237319141626, + 0.0034684319980442524, + 0.001632769824936986, + 0.0015723130200058222, + 0.004184702876955271, + 0.002263220027089119, + 0.003124838462099433, + 0.0030864826403558254, + 0.002785838209092617, + 0.00327141466550529, + 0.0016368207288905978, + 0.0028816480189561844, + 0.00370971136726439, + 0.002846046118065715, + 0.003112823236733675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.0046506100334227085 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035154461860657 + ], + "opposite_logprob_tokens": [ + -1.9103515148162842 + ], + "logprob_reward": -0.16035154461860657, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035154461860657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03618793189525604, + 0.015204453840851784, + 0.015224838629364967, + 0.4205826222896576, + 0.0662422776222229, + 0.026462504640221596, + 0.21279975771903992, + 0.04716518893837929, + 0.08570358157157898, + 0.07442691922187805 + ], + "credit_map": [ + 0.005358156282454729, + 0.0030929141212254763, + 0.006126605439931154, + 0.003545323619619012, + 0.004331630188971758, + 0.0047318898141384125, + 0.0030782974790781736, + 0.005923113785684109, + 0.0014982280554249883, + 0.0016630819300189614, + 0.0020544182043522596, + 0.0018137164879590273, + 0.0019435143331065774, + 0.0018243123777210712, + 0.0016048606485128403, + 0.0028023202903568745, + 0.0021122130565345287, + 0.0016678526299074292, + 0.003161810105666518, + 0.0018257726915180683, + 0.0014504332793876529, + 0.0016264523146674037, + 0.0015466756885871291, + 0.0018336273496970534, + 0.043012719601392746, + 0.01191170047968626, + 0.14654229581356049, + 0.018447542563080788, + 0.010473758913576603, + 0.14300790429115295, + 0.005640980787575245, + 0.04154565930366516, + 0.002301410771906376, + 0.0029718028381466866, + 0.0042143212631344795, + 0.0035350986290723085, + 0.003173557575792074, + 0.022527912631630898, + 0.0034686869475990534, + 0.02404947765171528, + 0.002861482324078679, + 0.002222413429990411, + 0.0039420342072844505, + 0.0031677407678216696, + 0.0024083333555608988, + 0.004545180592685938, + 0.0032746291253715754, + 0.004040685947984457, + 0.006340132560580969, + 0.005692091304808855, + 0.00839887373149395, + 0.00840761885046959, + 0.015448217280209064, + 0.08456449955701828, + 0.007286969106644392, + 0.07666132599115372, + 0.005255256313830614, + 0.005697856191545725, + 0.005164181813597679, + 0.007992449216544628, + 0.004730129148811102, + 0.007023918908089399, + 0.00639646640047431, + 0.004904923494905233, + 0.006241417024284601, + 0.013592514209449291, + 0.009843684732913971, + 0.013648159801959991, + 0.006186998449265957, + 0.014906753785908222, + 0.009909749031066895, + 0.01137429941445589, + 0.009246915578842163, + 0.012049414217472076, + 0.009706334210932255, + 0.007394357118755579, + 0.009249289520084858, + 0.008108225651085377, + 0.010078741237521172, + 0.008593632839620113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.16035154461860657 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054975397884845734 + ], + "opposite_logprob_tokens": [ + -2.9299755096435547 + ], + "logprob_reward": -0.054975397884845734, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054975397884845734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061332788318395615, + 0.02968643605709076, + 0.027049299329519272, + 0.08174297958612442, + 0.19767990708351135, + 0.021598808467388153, + 0.42464011907577515, + 0.03343180567026138, + 0.07613753527402878, + 0.046700332313776016 + ], + "credit_map": [ + 0.009010840207338333, + 0.004809633828699589, + 0.010520695708692074, + 0.006403719075024128, + 0.0051913135685026646, + 0.00853027869015932, + 0.004435320850461721, + 0.012430982664227486, + 0.003946526907384396, + 0.0031631686724722385, + 0.005029769614338875, + 0.003242316422984004, + 0.0038622820284217596, + 0.003170331474393606, + 0.004172780085355043, + 0.003099258756265044, + 0.0032928187865763903, + 0.0018388049211353064, + 0.003635136177763343, + 0.004276884254068136, + 0.0028638921212404966, + 0.004086707718670368, + 0.0027825357392430305, + 0.004272519610822201, + 0.009497243911027908, + 0.0038350617978721857, + 0.024826211854815483, + 0.004993761889636517, + 0.003986279014497995, + 0.023085396736860275, + 0.0031228973530232906, + 0.008396116085350513, + 0.005511834751814604, + 0.0023392143193632364, + 0.006834154017269611, + 0.004214445594698191, + 0.0056989192962646484, + 0.08533832430839539, + 0.004686364904046059, + 0.08305665105581284, + 0.0036891591735184193, + 0.0018267561681568623, + 0.0040964693762362, + 0.0016539235366508365, + 0.002467090031132102, + 0.0029190459754318, + 0.0014286700170487165, + 0.003517692908644676, + 0.011426649987697601, + 0.00890697818249464, + 0.015871888026595116, + 0.010042273439466953, + 0.03046233579516411, + 0.1753920316696167, + 0.009804921224713326, + 0.16273298859596252, + 0.003906326834112406, + 0.0036163758486509323, + 0.004151543602347374, + 0.0037798574194312096, + 0.0030863936990499496, + 0.005724471062421799, + 0.004218002315610647, + 0.004948835354298353, + 0.006278134882450104, + 0.006564083509147167, + 0.010084773413836956, + 0.009627100080251694, + 0.0043802326545119286, + 0.02071208693087101, + 0.00746948504820466, + 0.01102162804454565, + 0.006174125708639622, + 0.009027144871652126, + 0.007021926809102297, + 0.0042945099994540215, + 0.0049057272262871265, + 0.005960302893072367, + 0.004007984418421984, + 0.005308607127517462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.054975397884845734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, eye-level shot initially frames a surreal oil painting of a magnificent floral arrangement, mounted in an ornate, dark wood frame against a dimly lit gallery wall. The painting itself depicts a vibrant bouquet, featuring lush, crimson roses with thick, thick petals, sunny yellow tulips, and deep purple irises, all set in a simple, clear glass vase that artfully refracts light. The texture of the oil painting is palpable, glossy under the spotlight focused from above. Then, the camera begins a slow, seamless advance, steadily approaching the canvas. As the frame tightens on the bouquet, a surreal transformation begins; the painted flowers begin to wilt in a time-lapse fashion. The vibrant crimson of the roses gradually fades to a dull, dusty maroon, the edges of the petals curling inward. The bright yellow of the tulips fades to a pale, sickly ochre, and their once-sturdy stems bend and droop with a palpable heaviness. The petals seem to lose their painterly structure, detaching and realistically sliding along the textured surface of the canvas, leaving faint, oily streaks in their path. The entire scene on the canvas shifts from a vibrant celebration of life to a somber depiction of decay, with the colors becoming muted and earthy as the bouquet collapses into a withered state, creating a poignant and melancholic atmosphere that contrasts sharply with its initial breathtaking beauty.", + "seed": 2115, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_2/index_1860_seed_2115_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08977481871843337, + "vlm_logprob": -0.08977481871843337, + "vlm_yes_no_margin": 3.374999988079071, + "vlm_token_credit": -0.08977481871843337, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01971611939370632 + ], + "opposite_logprob_tokens": [ + -4.769716262817383 + ], + "logprob_reward": -0.01971611939370632, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01971611939370632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06986280530691147, + 0.24706970155239105, + 0.08685903996229172, + 0.06009755656123161, + 0.07764307409524918, + 0.06168629229068756, + 0.1711021065711975, + 0.08600498735904694, + 0.05788268893957138, + 0.08179179579019547 + ], + "credit_map": [ + 0.013487741351127625, + 0.006847138982266188, + 0.013060672208666801, + 0.007038247771561146, + 0.00635423231869936, + 0.00967694167047739, + 0.005244304891675711, + 0.008153526112437248, + 0.027029890567064285, + 0.022232234477996826, + 0.027127914130687714, + 0.028276942670345306, + 0.0243228692561388, + 0.04967770352959633, + 0.028240134939551353, + 0.04016200080513954, + 0.014831669628620148, + 0.006051564123481512, + 0.01828010007739067, + 0.0072664231993258, + 0.005318039562553167, + 0.01441574189811945, + 0.007619287818670273, + 0.01307620108127594, + 0.012109966017305851, + 0.004844179376959801, + 0.007350813131779432, + 0.007292685564607382, + 0.006618569605052471, + 0.007173649966716766, + 0.007208721712231636, + 0.0074989646673202515, + 0.009547708556056023, + 0.006619615945965052, + 0.013413454405963421, + 0.0065733278170228004, + 0.007231084164232016, + 0.01662866398692131, + 0.00749633926898241, + 0.010132867842912674, + 0.00810893066227436, + 0.005366934929043055, + 0.01187890488654375, + 0.006233596708625555, + 0.005815495736896992, + 0.007719283923506737, + 0.006681979633867741, + 0.009881160221993923, + 0.02490442804992199, + 0.014517119154334068, + 0.034569188952445984, + 0.01596151851117611, + 0.010816724970936775, + 0.02185148373246193, + 0.019896775484085083, + 0.028584862127900124, + 0.00921243242919445, + 0.011400883086025715, + 0.010132367722690105, + 0.01054462231695652, + 0.011344282887876034, + 0.007996346801519394, + 0.01618526689708233, + 0.009188772179186344, + 0.009245450608432293, + 0.0068490211851894855, + 0.010794411413371563, + 0.0068268682807683945, + 0.0049425531178712845, + 0.007277026306837797, + 0.004963567014783621, + 0.006983783096075058, + 0.015202962793409824, + 0.007315683178603649, + 0.01877562515437603, + 0.006714440882205963, + 0.0048290700651705265, + 0.008334948681294918, + 0.006524420343339443, + 0.014094643294811249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.01971611939370632 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637290023267269 + ], + "opposite_logprob_tokens": [ + -3.6513729095458984 + ], + "logprob_reward": -0.02637290023267269, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637290023267269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09549970924854279, + 0.17398673295974731, + 0.03957919403910637, + 0.14333222806453705, + 0.07095476984977722, + 0.193075031042099, + 0.14476390182971954, + 0.06479604542255402, + 0.04994285851716995, + 0.024069566279649734 + ], + "credit_map": [ + 0.011318455450236797, + 0.012909529730677605, + 0.01331633236259222, + 0.008526589721441269, + 0.01001744344830513, + 0.012405912391841412, + 0.014047704637050629, + 0.012957734987139702, + 0.01799941435456276, + 0.010527056641876698, + 0.02063341997563839, + 0.014253071509301662, + 0.011608421802520752, + 0.048503510653972626, + 0.02384760044515133, + 0.026614220812916756, + 0.004879225045442581, + 0.003958926070481539, + 0.005246900953352451, + 0.004281490109860897, + 0.0036105879116803408, + 0.005100126378238201, + 0.006817653309553862, + 0.00568427937105298, + 0.021821187809109688, + 0.017038939520716667, + 0.0162463691085577, + 0.01611555553972721, + 0.02001752331852913, + 0.016170205548405647, + 0.019469663500785828, + 0.01645277813076973, + 0.007430873811244965, + 0.0037291834596544504, + 0.009099461138248444, + 0.0044572846964001656, + 0.004187481477856636, + 0.020833073183894157, + 0.004034576937556267, + 0.017182838171720505, + 0.023663761094212532, + 0.01968868263065815, + 0.03762045130133629, + 0.016672326251864433, + 0.01778501830995083, + 0.020566897466778755, + 0.03425033390522003, + 0.022827545180916786, + 0.021198302507400513, + 0.012505000457167625, + 0.029531070962548256, + 0.013705399818718433, + 0.01205825712531805, + 0.018403509631752968, + 0.014889820478856564, + 0.022472525015473366, + 0.005611739587038755, + 0.007037696428596973, + 0.0062216538935899734, + 0.008613363839685917, + 0.008424794301390648, + 0.007183486595749855, + 0.0142919160425663, + 0.007411392405629158, + 0.006927275564521551, + 0.006183356512337923, + 0.005304149352014065, + 0.00569135881960392, + 0.004446977283805609, + 0.005218426696956158, + 0.007903764955699444, + 0.00826754979789257, + 0.002905910136178136, + 0.002873002551496029, + 0.003864372381940484, + 0.00294165569357574, + 0.0026286786887794733, + 0.00280221039429307, + 0.003155418671667576, + 0.0028983158990740776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.02637290023267269 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084172874689102 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011084172874689102, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084172874689102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04748370870947838, + 0.05717720836400986, + 0.025262387469410896, + 0.10661059617996216, + 0.2149246782064438, + 0.04355229064822197, + 0.3781609833240509, + 0.039787162095308304, + 0.06022237241268158, + 0.02681853622198105 + ], + "credit_map": [ + 0.011713096871972084, + 0.003331246552988887, + 0.0099605992436409, + 0.0034025979693979025, + 0.004053034819662571, + 0.006315867882221937, + 0.0041654277592897415, + 0.004541833885014057, + 0.00634415540844202, + 0.004224745091050863, + 0.007144251838326454, + 0.005535833537578583, + 0.0044749123044312, + 0.013513005338609219, + 0.005649494472891092, + 0.010290812700986862, + 0.0028652700129896402, + 0.0026946975849568844, + 0.002887612907215953, + 0.0028664651326835155, + 0.0028335319366306067, + 0.004189138300716877, + 0.0033739691134542227, + 0.003551703179255128, + 0.026015378534793854, + 0.008809121325612068, + 0.020413020625710487, + 0.0074638668447732925, + 0.008163832128047943, + 0.01134862843900919, + 0.00861087255179882, + 0.01578587107360363, + 0.017935022711753845, + 0.012883939780294895, + 0.012901616282761097, + 0.009446710348129272, + 0.015579245053231716, + 0.06708066165447235, + 0.018350875005126, + 0.06074662506580353, + 0.00630623335018754, + 0.004503665026277304, + 0.00819164328277111, + 0.0037518255412578583, + 0.004026947543025017, + 0.004393243230879307, + 0.005809230729937553, + 0.006569503806531429, + 0.051777955144643784, + 0.03312769904732704, + 0.06985633075237274, + 0.03752748668193817, + 0.03015102818608284, + 0.047831468284130096, + 0.047497861087322235, + 0.0603911429643631, + 0.004518858157098293, + 0.005463799461722374, + 0.00658660801127553, + 0.004088658839464188, + 0.005646653939038515, + 0.003788436995819211, + 0.005776630714535713, + 0.0039175148122012615, + 0.0069588725455105305, + 0.007086805533617735, + 0.009868862107396126, + 0.008657248690724373, + 0.004756529815495014, + 0.0065939538180828094, + 0.006518711801618338, + 0.00978139042854309, + 0.005108634941279888, + 0.003411303274333477, + 0.003802109742537141, + 0.0026339979376643896, + 0.0026637359987944365, + 0.00284846150316298, + 0.0029422889929264784, + 0.003408001968637109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.011084172874689102 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34862619638442993 + ], + "opposite_logprob_tokens": [ + -1.2236261367797852 + ], + "logprob_reward": -0.34862619638442993, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.34862619638442993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09318234771490097, + 0.17689859867095947, + 0.0444670170545578, + 0.09455347806215286, + 0.18830238282680511, + 0.06048577278852463, + 0.1378893256187439, + 0.0631742775440216, + 0.06167057901620865, + 0.079376220703125 + ], + "credit_map": [ + 0.015763111412525177, + 0.01240233238786459, + 0.015420179814100266, + 0.007368415594100952, + 0.015046613290905952, + 0.007151927798986435, + 0.011381194926798344, + 0.008648570626974106, + 0.01770513691008091, + 0.010704888962209225, + 0.021848151460289955, + 0.01622161641716957, + 0.013862702995538712, + 0.04965716972947121, + 0.020000802353024483, + 0.026898104697465897, + 0.007526238914579153, + 0.003938653506338596, + 0.007242659106850624, + 0.004461594391614199, + 0.003611269872635603, + 0.006188421975821257, + 0.005237019155174494, + 0.0062611582688987255, + 0.013915377669036388, + 0.010854714550077915, + 0.014942031353712082, + 0.009644846431910992, + 0.011120720766484737, + 0.011676965281367302, + 0.010418875142931938, + 0.01197995338588953, + 0.014372924342751503, + 0.013204355724155903, + 0.016856828704476357, + 0.01033761166036129, + 0.014232602901756763, + 0.05783496052026749, + 0.01846122182905674, + 0.04300187528133392, + 0.006696550641208887, + 0.008869197219610214, + 0.010612096637487411, + 0.004981707781553268, + 0.008237647823989391, + 0.008830024860799313, + 0.005544615909457207, + 0.006713930983096361, + 0.01780550740659237, + 0.013075122609734535, + 0.024131592363119125, + 0.01372289378196001, + 0.013659472577273846, + 0.015431897714734077, + 0.0172550231218338, + 0.022807804867625237, + 0.006291999481618404, + 0.006435437127947807, + 0.010364318266510963, + 0.006564827170222998, + 0.007337765768170357, + 0.007608416955918074, + 0.010579620487987995, + 0.007991889491677284, + 0.011684859171509743, + 0.005258576944470406, + 0.009045420214533806, + 0.005980342626571655, + 0.004106998909264803, + 0.006462681572884321, + 0.008445125073194504, + 0.010686575435101986, + 0.015728266909718513, + 0.008506176993250847, + 0.01048946287482977, + 0.008457968011498451, + 0.006567998789250851, + 0.008846155367791653, + 0.010538164526224136, + 0.010242026299238205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.34862619638442993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043074704706668854 + ], + "opposite_logprob_tokens": [ + -3.168074607849121 + ], + "logprob_reward": -0.043074704706668854, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043074704706668854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12135487049818039, + 0.2528795897960663, + 0.06612209230661392, + 0.07817547768354416, + 0.054923709481954575, + 0.16389204561710358, + 0.09265672415494919, + 0.07316429167985916, + 0.031129207462072372, + 0.06570202112197876 + ], + "credit_map": [ + 0.026364073157310486, + 0.009287171065807343, + 0.027016768231987953, + 0.010532704181969166, + 0.007391619961708784, + 0.014408511109650135, + 0.009950199164450169, + 0.016403816640377045, + 0.03070620447397232, + 0.01443968154489994, + 0.030925095081329346, + 0.022873325273394585, + 0.016370346769690514, + 0.07024448364973068, + 0.024502979591488838, + 0.04281746968626976, + 0.009010479785501957, + 0.00535856606438756, + 0.011129935272037983, + 0.00819506123661995, + 0.005919975694268942, + 0.008977217599749565, + 0.0077116647735238075, + 0.00981918815523386, + 0.014182413928210735, + 0.006896762177348137, + 0.01050811167806387, + 0.007041287142783403, + 0.010753682814538479, + 0.007765080779790878, + 0.01054706797003746, + 0.01048106886446476, + 0.006286801770329475, + 0.0037170539144426584, + 0.007272584829479456, + 0.003666309639811516, + 0.004017859697341919, + 0.01205866876989603, + 0.0036765867844223976, + 0.014227844774723053, + 0.019795281812548637, + 0.01651470921933651, + 0.03270093351602554, + 0.013311907649040222, + 0.014901546761393547, + 0.017205944284796715, + 0.026818884536623955, + 0.0226428359746933, + 0.011541290208697319, + 0.008071841672062874, + 0.015830617398023605, + 0.008879415690898895, + 0.008223208598792553, + 0.012245658785104752, + 0.011058331467211246, + 0.016806358471512794, + 0.010280397720634937, + 0.006115755531936884, + 0.017420679330825806, + 0.00793691910803318, + 0.007274073548614979, + 0.007201939821243286, + 0.00876070000231266, + 0.008173825219273567, + 0.006208152510225773, + 0.003118748776614666, + 0.004493165761232376, + 0.004007425159215927, + 0.0027034692466259003, + 0.00328255258500576, + 0.0040551344864070415, + 0.0032605594024062157, + 0.012970903888344765, + 0.00894876942038536, + 0.010840396396815777, + 0.005371826235204935, + 0.00674116937443614, + 0.007229023613035679, + 0.005637523252516985, + 0.007962402887642384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.043074704706668854 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, eye-level shot initially frames a surreal oil painting of a magnificent floral arrangement, mounted in an ornate, dark wood frame against a dimly lit gallery wall. The painting itself depicts a vibrant bouquet, featuring lush, crimson roses with thick, thick petals, sunny yellow tulips, and deep purple irises, all set in a simple, clear glass vase that artfully refracts light. The texture of the oil painting is palpable, glossy under the spotlight focused from above. Then, the camera begins a slow, seamless advance, steadily approaching the canvas. As the frame tightens on the bouquet, a surreal transformation begins; the painted flowers begin to wilt in a time-lapse fashion. The vibrant crimson of the roses gradually fades to a dull, dusty maroon, the edges of the petals curling inward. The bright yellow of the tulips fades to a pale, sickly ochre, and their once-sturdy stems bend and droop with a palpable heaviness. The petals seem to lose their painterly structure, detaching and realistically sliding along the textured surface of the canvas, leaving faint, oily streaks in their path. The entire scene on the canvas shifts from a vibrant celebration of life to a somber depiction of decay, with the colors becoming muted and earthy as the bouquet collapses into a withered state, creating a poignant and melancholic atmosphere that contrasts sharply with its initial breathtaking beauty.", + "seed": 2116, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_2/index_1860_seed_2116_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0232207827270031, + "vlm_logprob": -0.0232207827270031, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.0232207827270031, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01662578247487545 + ], + "opposite_logprob_tokens": [ + -5.516625881195068 + ], + "logprob_reward": -0.01662578247487545, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.01662578247487545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06705453991889954, + 0.11128558218479156, + 0.011940013617277145, + 0.056984011083841324, + 0.023284152150154114, + 0.010764826089143753, + 0.5722613334655762, + 0.12798932194709778, + 0.00936943106353283, + 0.009066827595233917 + ], + "credit_map": [ + 0.00641737412661314, + 0.006283543072640896, + 0.010084892623126507, + 0.008752112276852131, + 0.007948181591928005, + 0.010889231227338314, + 0.005860614590346813, + 0.010818583890795708, + 0.0060415142215788364, + 0.001226593623869121, + 0.0073717557825148106, + 0.002114515984430909, + 0.002495687687769532, + 0.04800639674067497, + 0.00336996978148818, + 0.04065914452075958, + 0.0010296788532286882, + 0.0011913940543308854, + 0.001280223368667066, + 0.000986878527328372, + 0.0014840629883110523, + 0.00215639709495008, + 0.0013003537897020578, + 0.0025110244750976562, + 0.005355512723326683, + 0.00693582883104682, + 0.006027828902006149, + 0.004962419159710407, + 0.011478141881525517, + 0.006316584534943104, + 0.00628989702090621, + 0.00961779709905386, + 0.0030321869999170303, + 0.0013433846179395914, + 0.003440119558945298, + 0.0029142212588340044, + 0.0018937364220619202, + 0.002460717922076583, + 0.0053887153044342995, + 0.0028110709972679615, + 0.001592734595760703, + 0.0009853907395154238, + 0.0014828646089881659, + 0.0009035346447490156, + 0.001058550551533699, + 0.0017248373478651047, + 0.0015740754315629601, + 0.0014428385766223073, + 0.03197808563709259, + 0.009083086624741554, + 0.056022800505161285, + 0.010293960571289062, + 0.012584337033331394, + 0.2388976365327835, + 0.009914213791489601, + 0.20348717272281647, + 0.011312589980661869, + 0.004444355145096779, + 0.010174036026000977, + 0.0036193907726556063, + 0.003976672422140837, + 0.04928075149655342, + 0.0035109242890030146, + 0.041670605540275574, + 0.0013378479052335024, + 0.0007499474450014532, + 0.002106140134856105, + 0.0007341742748394608, + 0.0007623209385201335, + 0.0012057340936735272, + 0.0009532679687254131, + 0.0015199986519291997, + 0.0010387698421254754, + 0.0007755651022307575, + 0.0016833051340654492, + 0.000976896146312356, + 0.0006411485956050456, + 0.001567077706567943, + 0.0012981343315914273, + 0.0010859309695661068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.01662578247487545 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216454699635506 + ], + "opposite_logprob_tokens": [ + -4.264216423034668 + ], + "logprob_reward": -0.014216454699635506, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216454699635506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08633953332901001, + 0.3660041391849518, + 0.012578243389725685, + 0.1252012997865677, + 0.026369204744696617, + 0.01899326778948307, + 0.11391075700521469, + 0.22647246718406677, + 0.012748075649142265, + 0.011383053846657276 + ], + "credit_map": [ + 0.008893664926290512, + 0.0075559187680482864, + 0.011843111366033554, + 0.005707112140953541, + 0.007700415328145027, + 0.02114364318549633, + 0.006083804648369551, + 0.0174118559807539, + 0.018224285915493965, + 0.0034750604536384344, + 0.029646417126059532, + 0.005985829513520002, + 0.007654751185327768, + 0.15965090692043304, + 0.010995668359100819, + 0.13037121295928955, + 0.0010313412640243769, + 0.001705488539300859, + 0.001426279777660966, + 0.0014560947893187404, + 0.0017791196005418897, + 0.0017277199076488614, + 0.0014455622294917703, + 0.0020066373981535435, + 0.011462354101240635, + 0.015709152445197105, + 0.016414226964116096, + 0.012197265401482582, + 0.020682115107774734, + 0.014062987640500069, + 0.014433315023779869, + 0.02023988962173462, + 0.0030642938800156116, + 0.002472523832693696, + 0.0044645341113209724, + 0.0025489286053925753, + 0.0035342881456017494, + 0.003925919067114592, + 0.0037208017893135548, + 0.0026379141490906477, + 0.0027246116660535336, + 0.0013218013336881995, + 0.0023925479035824537, + 0.001729366835206747, + 0.0019472356652840972, + 0.002312504453584552, + 0.004096592776477337, + 0.0024686090182513, + 0.006384050007909536, + 0.0020236445125192404, + 0.01135748066008091, + 0.002375774784013629, + 0.0027065810281783342, + 0.047055523842573166, + 0.0021929717622697353, + 0.03981473296880722, + 0.019345292821526527, + 0.005607800558209419, + 0.016172705218195915, + 0.0035363491624593735, + 0.005057129077613354, + 0.09295209497213364, + 0.004366500303149223, + 0.0794345885515213, + 0.0016793709946796298, + 0.0015490578953176737, + 0.0019204637501388788, + 0.0011905200080946088, + 0.0013782120076939464, + 0.0017257359577342868, + 0.0016972508747130632, + 0.0016074639279395342, + 0.0013265577144920826, + 0.001004949677735567, + 0.0015901539009064436, + 0.0012250792933627963, + 0.00079315435141325, + 0.002308985684067011, + 0.0017553690122440457, + 0.0013788038631901145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.014216454699635506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100605845451355 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.004100605845451355, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100605845451355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06371792405843735, + 0.1718931496143341, + 0.05824962630867958, + 0.04753931239247322, + 0.03417954593896866, + 0.03268415480852127, + 0.28644850850105286, + 0.24949927628040314, + 0.029720520600676537, + 0.026067979633808136 + ], + "credit_map": [ + 0.007047303020954132, + 0.0055345878936350346, + 0.006682767532765865, + 0.005151635967195034, + 0.006427413783967495, + 0.017754945904016495, + 0.004456615541130304, + 0.010662649758160114, + 0.008558896370232105, + 0.00262464489787817, + 0.012528364546597004, + 0.002902368316426873, + 0.004364983178675175, + 0.07391826808452606, + 0.005353263579308987, + 0.0616423562169075, + 0.00472622225061059, + 0.005556444171816111, + 0.005118526052683592, + 0.0069655789993703365, + 0.0056812455877661705, + 0.01100060623139143, + 0.008211962878704071, + 0.010989041067659855, + 0.0046884650364518166, + 0.006029848009347916, + 0.006186417303979397, + 0.004739703144878149, + 0.008061985485255718, + 0.005116770043969154, + 0.005394281819462776, + 0.007321842480450869, + 0.0035686062183231115, + 0.0031363964080810547, + 0.005493229255080223, + 0.003746517701074481, + 0.002786015858873725, + 0.0044113826006650925, + 0.006553788669407368, + 0.004483604803681374, + 0.0044901990331709385, + 0.0027721410151571035, + 0.004885421600192785, + 0.0025673648342490196, + 0.0032892723102122545, + 0.004330418072640896, + 0.005037784576416016, + 0.0053115529008209705, + 0.016552366316318512, + 0.004909879062324762, + 0.025975920259952545, + 0.005616323556751013, + 0.006532191764563322, + 0.11574626713991165, + 0.007179467938840389, + 0.1039360910654068, + 0.021623877808451653, + 0.009051965549588203, + 0.018736889585852623, + 0.006658582482486963, + 0.009422394447028637, + 0.09563140571117401, + 0.006835006643086672, + 0.08153915405273438, + 0.0031693740747869015, + 0.003194542834535241, + 0.006327209994196892, + 0.0032958625815808773, + 0.002568773226812482, + 0.003409670665860176, + 0.0045679896138608456, + 0.0031870994716882706, + 0.003077360102906823, + 0.0018272269517183304, + 0.00456589087843895, + 0.0019529565470293164, + 0.0016872251871973276, + 0.005531258415430784, + 0.004124943632632494, + 0.0033011180348694324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.004100605845451355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043058719485998154 + ], + "opposite_logprob_tokens": [ + -3.1680586338043213 + ], + "logprob_reward": -0.043058719485998154, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043058719485998154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08123882114887238, + 0.3438703119754791, + 0.0256529301404953, + 0.09606531262397766, + 0.04964454844594002, + 0.012717596255242825, + 0.21654188632965088, + 0.0276798028498888, + 0.03975844383239746, + 0.10683031380176544 + ], + "credit_map": [ + 0.009105396457016468, + 0.005242311395704746, + 0.009242418222129345, + 0.005434669554233551, + 0.004654328804463148, + 0.023614754900336266, + 0.007435597945004702, + 0.016509344801306725, + 0.017194420099258423, + 0.002932552946731448, + 0.025523168966174126, + 0.005238164681941271, + 0.007425790652632713, + 0.14913968741893768, + 0.010121334344148636, + 0.12629519402980804, + 0.0024350343737751245, + 0.0026372778229415417, + 0.003827917855232954, + 0.0035896077752113342, + 0.0025861056055873632, + 0.0030101484153419733, + 0.004176540765911341, + 0.003390297992154956, + 0.008648885414004326, + 0.012673595920205116, + 0.011112858541309834, + 0.010399690829217434, + 0.01350654661655426, + 0.011965778656303883, + 0.013321973383426666, + 0.014435982331633568, + 0.0054003954865038395, + 0.0027373181656003, + 0.007764373905956745, + 0.005665197502821684, + 0.004172435961663723, + 0.008033894933760166, + 0.010671737603843212, + 0.005199197679758072, + 0.0017973455833271146, + 0.0012902743183076382, + 0.0015970185631886125, + 0.0011579024139791727, + 0.00197417545132339, + 0.0018949389923363924, + 0.001281312550418079, + 0.001724627218209207, + 0.012790539301931858, + 0.004786977544426918, + 0.018766889348626137, + 0.004537632223218679, + 0.005235114600509405, + 0.0867735743522644, + 0.0047946046106517315, + 0.07885655760765076, + 0.003710842924192548, + 0.003217384684830904, + 0.003887931350618601, + 0.0029341201297938824, + 0.004213414620608091, + 0.0031992963049560785, + 0.002920615253970027, + 0.003596195951104164, + 0.004190814681351185, + 0.003065424272790551, + 0.008894409984350204, + 0.003512457711622119, + 0.003383084200322628, + 0.007545625325292349, + 0.0044100587256252766, + 0.004756568931043148, + 0.014006584882736206, + 0.006931765004992485, + 0.01745796948671341, + 0.007847500033676624, + 0.007254143711179495, + 0.023739898577332497, + 0.01680375449359417, + 0.012788704596459866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.043058719485998154 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810235112905502 + ], + "opposite_logprob_tokens": [ + -3.288102388381958 + ], + "logprob_reward": -0.03810235112905502, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810235112905502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.263353556394577, + 0.05751727148890495, + 0.037173207849264145, + 0.09686834365129471, + 0.017424769699573517, + 0.012284401804208755, + 0.09628956019878387, + 0.3396400809288025, + 0.02901517227292061, + 0.05043356120586395 + ], + "credit_map": [ + 0.02703641541302204, + 0.020909123122692108, + 0.03793711215257645, + 0.035032931715250015, + 0.02760443277657032, + 0.04122018814086914, + 0.02011074498295784, + 0.0535026378929615, + 0.004058694466948509, + 0.0020407643169164658, + 0.0090275714173913, + 0.002637720899656415, + 0.002716576214879751, + 0.021928507834672928, + 0.0028629344888031483, + 0.012244504876434803, + 0.003354552434757352, + 0.0034129098057746887, + 0.00499670160934329, + 0.0037479177117347717, + 0.003747601993381977, + 0.0045381407253444195, + 0.005542529281228781, + 0.00783285591751337, + 0.009933484718203545, + 0.012213003821671009, + 0.009413721039891243, + 0.00950457714498043, + 0.016600679606199265, + 0.010947664268314838, + 0.012872418388724327, + 0.01538279838860035, + 0.0020406891126185656, + 0.0024766111746430397, + 0.0024450323544442654, + 0.002201850526034832, + 0.00249266903847456, + 0.0018415345111861825, + 0.0018155868165194988, + 0.002110797679051757, + 0.0014821046497672796, + 0.0009552710107527673, + 0.0014940090477466583, + 0.0012139499885961413, + 0.0009764331625774503, + 0.0015651063295081258, + 0.0029038176871836185, + 0.0016937102191150188, + 0.0054549784399569035, + 0.0020715808495879173, + 0.009842072613537312, + 0.0024518617428839207, + 0.0032324332278221846, + 0.03821302205324173, + 0.002712827641516924, + 0.03231079876422882, + 0.029202038422226906, + 0.009599798358976841, + 0.024712225422263145, + 0.006441997829824686, + 0.008327162824571133, + 0.13718976080417633, + 0.007018019445240498, + 0.11714909225702286, + 0.0038993293419480324, + 0.001954283332452178, + 0.005406609736382961, + 0.002122529549524188, + 0.0017713659908622503, + 0.00638935063034296, + 0.0034396187402307987, + 0.004032086580991745, + 0.005239611957222223, + 0.002955120988190174, + 0.012771944515407085, + 0.004009466152638197, + 0.003415594808757305, + 0.009332101792097092, + 0.005733612459152937, + 0.006976113189011812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.03810235112905502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, eye-level shot initially frames a surreal oil painting of a magnificent floral arrangement, mounted in an ornate, dark wood frame against a dimly lit gallery wall. The painting itself depicts a vibrant bouquet, featuring lush, crimson roses with thick, thick petals, sunny yellow tulips, and deep purple irises, all set in a simple, clear glass vase that artfully refracts light. The texture of the oil painting is palpable, glossy under the spotlight focused from above. Then, the camera begins a slow, seamless advance, steadily approaching the canvas. As the frame tightens on the bouquet, a surreal transformation begins; the painted flowers begin to wilt in a time-lapse fashion. The vibrant crimson of the roses gradually fades to a dull, dusty maroon, the edges of the petals curling inward. The bright yellow of the tulips fades to a pale, sickly ochre, and their once-sturdy stems bend and droop with a palpable heaviness. The petals seem to lose their painterly structure, detaching and realistically sliding along the textured surface of the canvas, leaving faint, oily streaks in their path. The entire scene on the canvas shifts from a vibrant celebration of life to a somber depiction of decay, with the colors becoming muted and earthy as the bouquet collapses into a withered state, creating a poignant and melancholic atmosphere that contrasts sharply with its initial breathtaking beauty.", + "seed": 2117, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_2/index_1860_seed_2117_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03782473094761372, + "vlm_logprob": -0.03782473094761372, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03782473094761372, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014272513799369335 + ], + "opposite_logprob_tokens": [ + -5.764272689819336 + ], + "logprob_reward": -0.014272513799369335, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.014272513799369335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2535824179649353, + 0.05861685425043106, + 0.09779128432273865, + 0.12465258687734604, + 0.07661928981542587, + 0.06120264530181885, + 0.09078612923622131, + 0.05961964279413223, + 0.10322979837656021, + 0.07389932870864868 + ], + "credit_map": [ + 0.05568072944879532, + 0.01452808827161789, + 0.03022015653550625, + 0.014253465458750725, + 0.02280518412590027, + 0.05703568086028099, + 0.013285517692565918, + 0.04577358812093735, + 0.010160882025957108, + 0.004772084299474955, + 0.007561936508864164, + 0.004213577602058649, + 0.005353901535272598, + 0.014774235896766186, + 0.0029888139106333256, + 0.008791421540081501, + 0.020665869116783142, + 0.007468771189451218, + 0.01252774614840746, + 0.005627805832773447, + 0.007885009981691837, + 0.025220833718776703, + 0.004133603069931269, + 0.014261643402278423, + 0.014218602329492569, + 0.016708387061953545, + 0.014664744026958942, + 0.00727396784350276, + 0.025862840935587883, + 0.019393764436244965, + 0.010481401346623898, + 0.01604887843132019, + 0.02062186226248741, + 0.006582507397979498, + 0.00872093066573143, + 0.006973082199692726, + 0.0057601346634328365, + 0.011389690451323986, + 0.004817629233002663, + 0.011753453873097897, + 0.01154645811766386, + 0.00756227457895875, + 0.008153237402439117, + 0.005257561802864075, + 0.00842344667762518, + 0.006651935167610645, + 0.004701700061559677, + 0.008906031958758831, + 0.018476812168955803, + 0.01033661887049675, + 0.020963093265891075, + 0.004830686841160059, + 0.008317196741700172, + 0.00940550398081541, + 0.0075798239558935165, + 0.010876396670937538, + 0.014008905738592148, + 0.006817948073148727, + 0.008872130885720253, + 0.0031573933083564043, + 0.007600502576678991, + 0.005037469789385796, + 0.005929118487983942, + 0.008196171373128891, + 0.022545604035258293, + 0.01017683744430542, + 0.019372591748833656, + 0.00915816891938448, + 0.013795781880617142, + 0.01007601898163557, + 0.008979239501059055, + 0.009125567972660065, + 0.00898301787674427, + 0.008452347479760647, + 0.010040683671832085, + 0.007232224568724632, + 0.006780628580600023, + 0.009523170068860054, + 0.007429684512317181, + 0.015457568690180779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.014272513799369335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212576672434807 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.014212576672434807, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212576672434807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22375789284706116, + 0.10993166267871857, + 0.08843516558408737, + 0.08046030253171921, + 0.08710715919733047, + 0.047759994864463806, + 0.07093995809555054, + 0.09256602078676224, + 0.07214178889989853, + 0.1269001066684723 + ], + "credit_map": [ + 0.031018830835819244, + 0.010853963904082775, + 0.02916976436972618, + 0.011510957032442093, + 0.017472289502620697, + 0.05907169356942177, + 0.013923652470111847, + 0.05073671415448189, + 0.01913546398282051, + 0.010540888644754887, + 0.011322038248181343, + 0.008591456338763237, + 0.013718263246119022, + 0.02672569453716278, + 0.007560440804809332, + 0.01233739871531725, + 0.014409614726901054, + 0.009441375732421875, + 0.011609497480094433, + 0.007723917718976736, + 0.013521279208362103, + 0.013015605509281158, + 0.008283848874270916, + 0.010430016554892063, + 0.012250742875039577, + 0.01317511685192585, + 0.007705098018050194, + 0.004934241529554129, + 0.014557566493749619, + 0.011321037076413631, + 0.008205488324165344, + 0.008311000652611256, + 0.009520883671939373, + 0.013211783953011036, + 0.009631612338125706, + 0.006561266258358955, + 0.014059278182685375, + 0.012408815324306488, + 0.009053600020706654, + 0.012659910134971142, + 0.005253802984952927, + 0.006815379019826651, + 0.007994323968887329, + 0.0036153981927782297, + 0.007375387009233236, + 0.006091414485126734, + 0.004333910997956991, + 0.006280376110225916, + 0.01438954658806324, + 0.006844183895736933, + 0.011030236259102821, + 0.005740278400480747, + 0.008550872094929218, + 0.010993723757565022, + 0.006476453505456448, + 0.006914656609296799, + 0.017627965658903122, + 0.009815209545195103, + 0.015388941392302513, + 0.006633914541453123, + 0.016224971041083336, + 0.009005499072372913, + 0.008907819166779518, + 0.008961686864495277, + 0.013065672479569912, + 0.006847283337265253, + 0.014275779947638512, + 0.008110258728265762, + 0.006501051597297192, + 0.008509092032909393, + 0.00698096351698041, + 0.007851678878068924, + 0.017085526138544083, + 0.02000080794095993, + 0.017790162935853004, + 0.015077677555382252, + 0.01573886163532734, + 0.012846310622990131, + 0.013212194666266441, + 0.015148564241826534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.014212576672434807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643728025257587 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.004643728025257587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643728025257587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10797254741191864, + 0.09288568049669266, + 0.06961055845022202, + 0.08423473685979843, + 0.04424353688955307, + 0.11174952238798141, + 0.2033298909664154, + 0.14100439846515656, + 0.054059598594903946, + 0.09090954810380936 + ], + "credit_map": [ + 0.0147892776876688, + 0.010789206251502037, + 0.016525771468877792, + 0.010282385163009167, + 0.010948269627988338, + 0.01783493533730507, + 0.011401214636862278, + 0.015401484444737434, + 0.013837788254022598, + 0.006643495988100767, + 0.010922743938863277, + 0.007598480675369501, + 0.008880594745278358, + 0.02331024967133999, + 0.006360610015690327, + 0.01533171720802784, + 0.008696101605892181, + 0.006188464816659689, + 0.007611543871462345, + 0.00512283481657505, + 0.0072501664981245995, + 0.01247690711170435, + 0.00575095321983099, + 0.01651359163224697, + 0.006978283170610666, + 0.011882348917424679, + 0.009158603847026825, + 0.007628330495208502, + 0.02043706737458706, + 0.010151978582143784, + 0.008031065575778484, + 0.00996706448495388, + 0.009010576643049717, + 0.005080038215965033, + 0.005666228011250496, + 0.004028109833598137, + 0.0037011546082794666, + 0.005380128975957632, + 0.004904432222247124, + 0.0064728655852377415, + 0.015359969809651375, + 0.016226602718234062, + 0.01204895880073309, + 0.007686627097427845, + 0.018756825476884842, + 0.014214696362614632, + 0.012639367021620274, + 0.014816470444202423, + 0.0403728187084198, + 0.012909429147839546, + 0.057147420942783356, + 0.010109455324709415, + 0.013471633195877075, + 0.02879267744719982, + 0.013867507688701153, + 0.026658929884433746, + 0.029505588114261627, + 0.012395355850458145, + 0.023823518306016922, + 0.00876501202583313, + 0.012484090402722359, + 0.013929247856140137, + 0.01302630826830864, + 0.027075277641415596, + 0.009605628438293934, + 0.0065180170349776745, + 0.00783516839146614, + 0.005062277428805828, + 0.006781808566302061, + 0.006962877698242664, + 0.006035208702087402, + 0.005258613731712103, + 0.010822740383446217, + 0.011517359875142574, + 0.013971100561320782, + 0.009616805240511894, + 0.010164880193769932, + 0.014332780614495277, + 0.008441705256700516, + 0.012042170390486717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.004643728025257587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129259541630745 + ], + "opposite_logprob_tokens": [ + -2.2379260063171387 + ], + "logprob_reward": -0.1129259541630745, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129259541630745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10910927504301071, + 0.052601877599954605, + 0.039083514362573624, + 0.06873314082622528, + 0.08905241638422012, + 0.05984816700220108, + 0.15051838755607605, + 0.22741809487342834, + 0.12037845700979233, + 0.08325671404600143 + ], + "credit_map": [ + 0.024027163162827492, + 0.008572068996727467, + 0.014355441555380821, + 0.006719158496707678, + 0.01163325086236, + 0.021474963054060936, + 0.007714457344263792, + 0.014612758532166481, + 0.01027358416467905, + 0.0046755908988416195, + 0.006140320096164942, + 0.003677547676488757, + 0.006932211574167013, + 0.01088755577802658, + 0.0037737677339464426, + 0.006241294555366039, + 0.005897541996091604, + 0.004261365160346031, + 0.005635233595967293, + 0.002672650618478656, + 0.0056791710667312145, + 0.005736001301556826, + 0.00416901707649231, + 0.005032530520111322, + 0.010567170567810535, + 0.010579965077340603, + 0.00637153722345829, + 0.003681379370391369, + 0.014158844947814941, + 0.010123520158231258, + 0.005787680856883526, + 0.007463040295988321, + 0.016491077840328217, + 0.009222080931067467, + 0.011326275765895844, + 0.006795153953135014, + 0.008989168331027031, + 0.015626996755599976, + 0.006188403815031052, + 0.014413260854780674, + 0.009162883274257183, + 0.007313381880521774, + 0.008191543631255627, + 0.004876535851508379, + 0.007279587909579277, + 0.00827347207814455, + 0.007180504035204649, + 0.007570256013423204, + 0.03476768732070923, + 0.012622285634279251, + 0.023076528683304787, + 0.006446524523198605, + 0.011179042048752308, + 0.0237798560410738, + 0.011190586723387241, + 0.027455858886241913, + 0.0335555337369442, + 0.023825712502002716, + 0.04058701917529106, + 0.019229458644986153, + 0.024066392332315445, + 0.01940441131591797, + 0.024866430088877678, + 0.04188312217593193, + 0.017396850511431694, + 0.017761407420039177, + 0.02213718555867672, + 0.010394920594990253, + 0.020500343292951584, + 0.011653097346425056, + 0.00957705732434988, + 0.010957593098282814, + 0.009295637719333172, + 0.008415534161031246, + 0.012337548658251762, + 0.007643680088222027, + 0.0072052073664963245, + 0.010033195838332176, + 0.007456643041223288, + 0.020869260653853416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.1129259541630745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene an eye-level shot of an oil painting of a floral arrangement in a dark wood frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the painting initially feature vibrant crimson roses, sunny yellow tulips, and deep purple irises in a clear glass vase?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, seamless advance steadily approaching the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the flowers undergo a time-lapse transformation where they wilt, colors fade to dull earthy tones, and petals slide off the canvas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the oil painting texture visible and glossy under a spotlight focused from above?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306888207793236 + ], + "opposite_logprob_tokens": [ + -3.1680688858032227 + ], + "logprob_reward": -0.04306888207793236, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306888207793236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16275307536125183, + 0.2792803645133972, + 0.055777743458747864, + 0.040042445063591, + 0.03880034387111664, + 0.03989289700984955, + 0.13874571025371552, + 0.10574081540107727, + 0.048906002193689346, + 0.09006059169769287 + ], + "credit_map": [ + 0.02607364021241665, + 0.012021097354590893, + 0.020001865923404694, + 0.008529000915586948, + 0.013663681223988533, + 0.04269494116306305, + 0.008634893223643303, + 0.031133968383073807, + 0.05304102599620819, + 0.02465844713151455, + 0.03415383771061897, + 0.01531257014721632, + 0.022325634956359863, + 0.06462176144123077, + 0.021800819784402847, + 0.04336627572774887, + 0.00957332644611597, + 0.005415250547230244, + 0.00612260727211833, + 0.003142575966194272, + 0.0065918657928705215, + 0.01148253120481968, + 0.003886464051902294, + 0.009563117288053036, + 0.007963594980537891, + 0.0038479366339743137, + 0.004327499773353338, + 0.0027956608682870865, + 0.005411369726061821, + 0.007014574017375708, + 0.003084426047280431, + 0.005597384180873632, + 0.007159112486988306, + 0.0034281136468052864, + 0.00456860801205039, + 0.003624272532761097, + 0.0030652519781142473, + 0.006492269691079855, + 0.003750619012862444, + 0.006712093483656645, + 0.005607049912214279, + 0.005489716771990061, + 0.004624800756573677, + 0.004091271199285984, + 0.006103729363530874, + 0.00484626367688179, + 0.003418111242353916, + 0.005711955018341541, + 0.032410506159067154, + 0.00859883427619934, + 0.039241570979356766, + 0.0063043576665222645, + 0.0069720568135380745, + 0.01827257126569748, + 0.008150557056069374, + 0.01879524625837803, + 0.027681875973939896, + 0.00957426056265831, + 0.025879722088575363, + 0.00735334912315011, + 0.007990794256329536, + 0.008310197852551937, + 0.00966416671872139, + 0.009286448359489441, + 0.007602244149893522, + 0.005958647932857275, + 0.008098985999822617, + 0.004915645811706781, + 0.007917998358607292, + 0.004920654464513063, + 0.004911933559924364, + 0.004579894244670868, + 0.01119550596922636, + 0.010213526897132397, + 0.015372402034699917, + 0.010385640896856785, + 0.00997915305197239, + 0.012727387249469757, + 0.008272301405668259, + 0.011914679780602455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.04306888207793236 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_5/index_tensor([3810, 3810])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_5/index_tensor([3810, 3810])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ac860b67bc414454c348aa22ee3efc0eae793b32 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_5/index_tensor([3810, 3810])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An incredibly slow, steady, and deliberate upward tilt begins with a close-up shot of the hem of a red silk cheongsam, suspended against a dark, blurred background. A soft, focused sidelight casts a precise shadow on the cheongsam, highlighting the silk's smooth, watery texture and deep, vibrant red hue. As the camera steadily moves upward, it sweeps across the cheongsam's elegant side slits before focusing on a vast expanse of silk, where subtle shifts of light and shadow flow with the camera's movement. Then, an exquisitely detailed and lifelike golden phoenix embroidery gradually comes into view. The camera follows the phoenix's magnificent tail feathers as they ascend, each feather embroidered in shimmering metallic gold thread clearly visible, the fine, three-dimensional stitching appearing to stand out from the silk surface like a relief. As the camera continues to pan upward, fully revealing the phoenix's outstretched wings and raised head, its piercing eyes and gorgeous crown feathers appear with stunning realism in ultra-high-definition details. The golden thread reflects a glittering glow in the light, forming a gorgeous contrast with the warm luster of the red silk. The entire picture is full of oriental classical mystery and luxury, creating a tranquil atmosphere of focused gaze on top craftsmanship and artistic admiration.", + "seed": 11625, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_5/index_3810_seed_11625_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03295098182279617, + "vlm_logprob": -0.03295098182279617, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.03295098182279617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019314315170049667 + ], + "opposite_logprob_tokens": [ + -4.144314289093018 + ], + "logprob_reward": -0.019314315170049667, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019314315170049667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1351109892129898, + 0.10660597681999207, + 0.13478459417819977, + 0.19569984078407288, + 0.04448001831769943, + 0.16064351797103882, + 0.023486780002713203, + 0.03062254749238491, + 0.12214797735214233, + 0.04641786217689514 + ], + "credit_map": [ + 0.009424273855984211, + 0.017587751150131226, + 0.013163994997739792, + 0.015372334979474545, + 0.018984025344252586, + 0.016245190054178238, + 0.019758205860853195, + 0.024575190618634224, + 0.008831987157464027, + 0.007697810418903828, + 0.009092639200389385, + 0.021054763346910477, + 0.010324861854314804, + 0.025654636323451996, + 0.01180579885840416, + 0.012143472209572792, + 0.01130753569304943, + 0.01469061803072691, + 0.012729811482131481, + 0.01472142431885004, + 0.023902352899312973, + 0.01764436811208725, + 0.02001785673201084, + 0.019770609214901924, + 0.012605492025613785, + 0.01783978007733822, + 0.009862823411822319, + 0.012335414998233318, + 0.0601150318980217, + 0.024140220135450363, + 0.014546154998242855, + 0.04425489902496338, + 0.003395231906324625, + 0.004098278935998678, + 0.0030387314036488533, + 0.007840614765882492, + 0.0026863375678658485, + 0.006202747579663992, + 0.008719244040548801, + 0.008498827926814556, + 0.03620293363928795, + 0.013608729466795921, + 0.01702776551246643, + 0.035342782735824585, + 0.006575342267751694, + 0.013983338139951229, + 0.021403929218649864, + 0.01649867370724678, + 0.004066673573106527, + 0.003650954458862543, + 0.002694095950573683, + 0.002605492016300559, + 0.002589118666946888, + 0.0023000764194875956, + 0.0016987615963444114, + 0.003881604876369238, + 0.004802004434168339, + 0.002387339249253273, + 0.007829893380403519, + 0.0025959769263863564, + 0.004796926863491535, + 0.0029994796495884657, + 0.0028910443652421236, + 0.00231988076120615, + 0.037112150341272354, + 0.007328339386731386, + 0.041264694184064865, + 0.00698704831302166, + 0.008767773397266865, + 0.005498417187482119, + 0.008768225088715553, + 0.006421309430152178, + 0.007228855509310961, + 0.006138914730399847, + 0.007126969285309315, + 0.0049161906354129314, + 0.006513827946037054, + 0.005633693654090166, + 0.005006592720746994, + 0.0038528151344507933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.019314315170049667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012572435662150383 + ], + "opposite_logprob_tokens": [ + -4.387572288513184 + ], + "logprob_reward": -0.012572435662150383, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012572435662150383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11097011715173721, + 0.09031955897808075, + 0.1581737995147705, + 0.05106477066874504, + 0.03234769031405449, + 0.17282016575336456, + 0.04333457723259926, + 0.062397152185440063, + 0.20293506979942322, + 0.0756371021270752 + ], + "credit_map": [ + 0.011394891887903214, + 0.011457595974206924, + 0.01033890713006258, + 0.011325296945869923, + 0.018021879717707634, + 0.01477071177214384, + 0.018066363409161568, + 0.015594471246004105, + 0.006535651162266731, + 0.006838283035904169, + 0.006357922684401274, + 0.01825983263552189, + 0.009114458225667477, + 0.02599751204252243, + 0.008390270173549652, + 0.008825629949569702, + 0.012094113975763321, + 0.01623157039284706, + 0.01203025970607996, + 0.013134649954736233, + 0.03176974877715111, + 0.023978082463145256, + 0.029706960543990135, + 0.019228415563702583, + 0.005576028022915125, + 0.005839549470692873, + 0.004335301462560892, + 0.006977549754083157, + 0.00819053128361702, + 0.008447051048278809, + 0.005102057941257954, + 0.006596701219677925, + 0.0033168315421789885, + 0.0034672983456403017, + 0.004362395498901606, + 0.00531357666477561, + 0.0021959557197988033, + 0.0026712878607213497, + 0.0054192692041397095, + 0.005601075477898121, + 0.0362887866795063, + 0.015184806659817696, + 0.014385283924639225, + 0.04123423993587494, + 0.00731413159519434, + 0.014463982544839382, + 0.025560762733221054, + 0.018388178199529648, + 0.006307207979261875, + 0.006531818304210901, + 0.005365016870200634, + 0.006764482241123915, + 0.003743999870494008, + 0.003946569748222828, + 0.0034031670074909925, + 0.007272318005561829, + 0.012126393616199493, + 0.004332125652581453, + 0.01837632991373539, + 0.004009390249848366, + 0.009491446428000927, + 0.005655390676110983, + 0.00452107097953558, + 0.0038850002456456423, + 0.047085732221603394, + 0.016762331128120422, + 0.04863645136356354, + 0.015383830294013023, + 0.025387709960341454, + 0.02057017758488655, + 0.01604296639561653, + 0.013065885752439499, + 0.010640598833560944, + 0.011008162051439285, + 0.009538330137729645, + 0.00871348287910223, + 0.009607640095055103, + 0.012965094298124313, + 0.006431399378925562, + 0.0067323981784284115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.012572435662150383 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017254954436793923 + ], + "opposite_logprob_tokens": [ + -6.376725673675537 + ], + "logprob_reward": -0.0017254954436793923, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017254954436793923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10249918699264526, + 0.037372224032878876, + 0.07715961337089539, + 0.14495615661144257, + 0.06522662937641144, + 0.11500164121389389, + 0.0970211774110794, + 0.1245352029800415, + 0.19588406383991241, + 0.04034414887428284 + ], + "credit_map": [ + 0.009463812224566936, + 0.008606599643826485, + 0.011981228366494179, + 0.009013673290610313, + 0.022103453055024147, + 0.015600387938320637, + 0.014280086383223534, + 0.011449944227933884, + 0.00490829162299633, + 0.0036020518746227026, + 0.004296380095183849, + 0.004491335246711969, + 0.004591531120240688, + 0.005596479866653681, + 0.004404808394610882, + 0.005481344647705555, + 0.006067117676138878, + 0.008914642967283726, + 0.0058156223967671394, + 0.0066306572407484055, + 0.014704961329698563, + 0.010906784795224667, + 0.010989970527589321, + 0.013129854574799538, + 0.007556431461125612, + 0.012940209358930588, + 0.005733760539442301, + 0.006574377417564392, + 0.053053028881549835, + 0.016309823840856552, + 0.008842901326715946, + 0.03394562005996704, + 0.005275503732264042, + 0.006447825115174055, + 0.004555106163024902, + 0.011598796583712101, + 0.004682580009102821, + 0.006741868797689676, + 0.012584851123392582, + 0.013340093195438385, + 0.017202366143465042, + 0.012463186867535114, + 0.010832247324287891, + 0.02404811978340149, + 0.008337439969182014, + 0.014180801808834076, + 0.016086868941783905, + 0.011850610375404358, + 0.01249955128878355, + 0.012851973995566368, + 0.009181640110909939, + 0.008946957997977734, + 0.0132959745824337, + 0.01613849401473999, + 0.009182480163872242, + 0.014924097806215286, + 0.01919948309659958, + 0.010428067296743393, + 0.035147055983543396, + 0.012681122869253159, + 0.01872459053993225, + 0.009630738757550716, + 0.010076101869344711, + 0.008648049086332321, + 0.021868744865059853, + 0.02506829984486103, + 0.025656800717115402, + 0.03142636641860008, + 0.03076428920030594, + 0.019336199387907982, + 0.0231244545429945, + 0.018638893961906433, + 0.005070527084171772, + 0.005697318352758884, + 0.005389840342104435, + 0.00500791659578681, + 0.0058084880001842976, + 0.005902779754251242, + 0.003951328340917826, + 0.0035159506369382143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.0017254954436793923 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820041425526142 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.01820041425526142, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820041425526142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08629004657268524, + 0.03148721531033516, + 0.09100691974163055, + 0.13143913447856903, + 0.06742879748344421, + 0.1184556633234024, + 0.029082806780934334, + 0.04765181615948677, + 0.33323726058006287, + 0.06392040103673935 + ], + "credit_map": [ + 0.007792863063514233, + 0.008791611529886723, + 0.010931821539998055, + 0.007706677075475454, + 0.016560319811105728, + 0.013737702742218971, + 0.010765745304524899, + 0.01000330038368702, + 0.0039283838123083115, + 0.002730989595875144, + 0.003059113398194313, + 0.00310213933698833, + 0.004617536440491676, + 0.003680623136460781, + 0.00412699393928051, + 0.006241438444703817, + 0.008152649737894535, + 0.007670953869819641, + 0.008402443490922451, + 0.007454014848917723, + 0.018239300698041916, + 0.013271472416818142, + 0.012295598164200783, + 0.015520484186708927, + 0.008380575105547905, + 0.009854641743004322, + 0.007478725165128708, + 0.007106803357601166, + 0.0417582206428051, + 0.016177058219909668, + 0.007574677933007479, + 0.0331084281206131, + 0.005813458934426308, + 0.0053025418892502785, + 0.004897660575807095, + 0.013700675219297409, + 0.0026093097403645515, + 0.005911159794777632, + 0.015889840200543404, + 0.013304156251251698, + 0.022515738382935524, + 0.009596221148967743, + 0.011309187859296799, + 0.017207829281687737, + 0.007253322750329971, + 0.011809224262833595, + 0.01220384705811739, + 0.026560289785265923, + 0.003214724827557802, + 0.004020581487566233, + 0.005098395980894566, + 0.0036542180459946394, + 0.003008737228810787, + 0.0032755830325186253, + 0.0021098863799124956, + 0.004700678866356611, + 0.006157442461699247, + 0.0038957777433097363, + 0.010643607936799526, + 0.0038333269767463207, + 0.008991164155304432, + 0.005097414832562208, + 0.0060495734214782715, + 0.002983508165925741, + 0.03835683688521385, + 0.0452662818133831, + 0.03837460279464722, + 0.06501052528619766, + 0.04947683587670326, + 0.024785909801721573, + 0.0410601869225502, + 0.0309060700237751, + 0.0071562305092811584, + 0.00887168850749731, + 0.009334194473922253, + 0.006549912504851818, + 0.009780573658645153, + 0.009679505601525307, + 0.00673264404758811, + 0.005815649870783091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.01820041425526142 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294224858283997 + ], + "opposite_logprob_tokens": [ + -2.2379422187805176 + ], + "logprob_reward": -0.11294224858283997, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294224858283997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1609930843114853, + 0.1110132709145546, + 0.11412562429904938, + 0.24219122529029846, + 0.05271795019507408, + 0.11304739862680435, + 0.07283739745616913, + 0.0375400185585022, + 0.05385603755712509, + 0.041678085923194885 + ], + "credit_map": [ + 0.017600636929273605, + 0.01772676780819893, + 0.023541823029518127, + 0.016079315915703773, + 0.02001160942018032, + 0.024787427857518196, + 0.019409677013754845, + 0.0218358114361763, + 0.00614955136552453, + 0.010227940045297146, + 0.0059770443476736546, + 0.02156074345111847, + 0.01440697442740202, + 0.02903044782578945, + 0.011294618248939514, + 0.012365940026938915, + 0.009360070340335369, + 0.014239946380257607, + 0.011154862120747566, + 0.009486825205385685, + 0.024314023554325104, + 0.017896413803100586, + 0.015785370022058487, + 0.011888100765645504, + 0.010165960527956486, + 0.020950939506292343, + 0.008378992788493633, + 0.009587462991476059, + 0.09476187825202942, + 0.02507622167468071, + 0.013848096132278442, + 0.05942165106534958, + 0.004846012685447931, + 0.0038565516006201506, + 0.004333134740591049, + 0.01067906990647316, + 0.002710597589612007, + 0.004779026377946138, + 0.01240523625165224, + 0.009108316153287888, + 0.023543303832411766, + 0.008192065171897411, + 0.01292672660201788, + 0.02845635637640953, + 0.005340287461876869, + 0.008106491528451443, + 0.011591174639761448, + 0.014890986494719982, + 0.011903936043381691, + 0.009404054842889309, + 0.007498971652239561, + 0.00612865062430501, + 0.00820653885602951, + 0.010359679348766804, + 0.006088851019740105, + 0.013246708549559116, + 0.006430615205317736, + 0.00212779245339334, + 0.013020385056734085, + 0.00305931712500751, + 0.003714737482368946, + 0.0035670921206474304, + 0.0026961236726492643, + 0.002923953579738736, + 0.008616330102086067, + 0.0064930845983326435, + 0.008559600450098515, + 0.008481786586344242, + 0.007821447215974331, + 0.004115477669984102, + 0.005833654664456844, + 0.003934646025300026, + 0.005355700384825468, + 0.005265934392809868, + 0.005703301168978214, + 0.005387448240071535, + 0.005577518604695797, + 0.006408393383026123, + 0.004509176593273878, + 0.0034706122241914272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.11294224858283997 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An incredibly slow, steady, and deliberate upward tilt begins with a close-up shot of the hem of a red silk cheongsam, suspended against a dark, blurred background. A soft, focused sidelight casts a precise shadow on the cheongsam, highlighting the silk's smooth, watery texture and deep, vibrant red hue. As the camera steadily moves upward, it sweeps across the cheongsam's elegant side slits before focusing on a vast expanse of silk, where subtle shifts of light and shadow flow with the camera's movement. Then, an exquisitely detailed and lifelike golden phoenix embroidery gradually comes into view. The camera follows the phoenix's magnificent tail feathers as they ascend, each feather embroidered in shimmering metallic gold thread clearly visible, the fine, three-dimensional stitching appearing to stand out from the silk surface like a relief. As the camera continues to pan upward, fully revealing the phoenix's outstretched wings and raised head, its piercing eyes and gorgeous crown feathers appear with stunning realism in ultra-high-definition details. The golden thread reflects a glittering glow in the light, forming a gorgeous contrast with the warm luster of the red silk. The entire picture is full of oriental classical mystery and luxury, creating a tranquil atmosphere of focused gaze on top craftsmanship and artistic admiration.", + "seed": 11626, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_5/index_3810_seed_11626_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02780691203661263, + "vlm_logprob": -0.02780691203661263, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.02780691203661263, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011487646959722042 + ], + "opposite_logprob_tokens": [ + -4.7614874839782715 + ], + "logprob_reward": -0.011487646959722042, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011487646959722042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16513514518737793, + 0.19074098765850067, + 0.05886288359761238, + 0.08427822589874268, + 0.0705505833029747, + 0.029877034947276115, + 0.0360388346016407, + 0.0611104853451252, + 0.27165624499320984, + 0.03174962103366852 + ], + "credit_map": [ + 0.020236991345882416, + 0.025432106107473373, + 0.019258931279182434, + 0.02046227641403675, + 0.025402290746569633, + 0.02311576157808304, + 0.014082777313888073, + 0.017144011333584785, + 0.015707848593592644, + 0.028254752978682518, + 0.014475250616669655, + 0.020823024213314056, + 0.04655662924051285, + 0.02375996671617031, + 0.020793527364730835, + 0.02036997303366661, + 0.004122070036828518, + 0.009830400347709656, + 0.002794957719743252, + 0.010174386203289032, + 0.008726160041987896, + 0.009928219020366669, + 0.007285177707672119, + 0.006001514382660389, + 0.005552948918193579, + 0.01026251632720232, + 0.008106797933578491, + 0.00871447566896677, + 0.01776713691651821, + 0.012681826017796993, + 0.009508667513728142, + 0.011683855205774307, + 0.005643093958497047, + 0.004907610360532999, + 0.005597859621047974, + 0.00718542467802763, + 0.007410687860101461, + 0.013062683865427971, + 0.011860952712595463, + 0.014882267452776432, + 0.0031840947922319174, + 0.003286289982497692, + 0.003768328810110688, + 0.002299405401572585, + 0.003020247910171747, + 0.005136880092322826, + 0.003459254279732704, + 0.005722533445805311, + 0.004308661911636591, + 0.004903638269752264, + 0.007139033637940884, + 0.002960306592285633, + 0.0035883954260498285, + 0.002767059486359358, + 0.004830556455999613, + 0.005541183985769749, + 0.011221399530768394, + 0.0051232874393463135, + 0.006325435359030962, + 0.003640713868662715, + 0.005294820759445429, + 0.018693851307034492, + 0.00467643840238452, + 0.006134538911283016, + 0.07280969619750977, + 0.028479160740971565, + 0.046159762889146805, + 0.015680545940995216, + 0.05240349844098091, + 0.01749773696064949, + 0.01880946010351181, + 0.019816407933831215, + 0.005276170559227467, + 0.0037970200646668673, + 0.004967043176293373, + 0.00268785422667861, + 0.004047786816954613, + 0.004783082287758589, + 0.0031294652726501226, + 0.0030611995607614517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.011487646959722042 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008664277382194996 + ], + "opposite_logprob_tokens": [ + -4.758664131164551 + ], + "logprob_reward": -0.008664277382194996, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008664277382194996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10193268209695816, + 0.19311170279979706, + 0.03250552713871002, + 0.07996399700641632, + 0.06983483582735062, + 0.03241093084216118, + 0.04379189759492874, + 0.1272820085287094, + 0.2755279242992401, + 0.04363848641514778 + ], + "credit_map": [ + 0.010538705624639988, + 0.014420450665056705, + 0.010063350200653076, + 0.014201433397829533, + 0.013903830200433731, + 0.018660277128219604, + 0.008945198729634285, + 0.011199434287846088, + 0.016785986721515656, + 0.030031582340598106, + 0.01404576562345028, + 0.01955237053334713, + 0.047656480222940445, + 0.022042807191610336, + 0.02123618684709072, + 0.021760528907179832, + 0.0017927425215020776, + 0.00484621524810791, + 0.0016338367713615298, + 0.006295349448919296, + 0.005357622634619474, + 0.005217593628913164, + 0.004989135079085827, + 0.0023730306420475245, + 0.0058506326749920845, + 0.009957057423889637, + 0.008679607883095741, + 0.0119569581001997, + 0.013288293965160847, + 0.011525562033057213, + 0.007479736115783453, + 0.011226150207221508, + 0.006332050077617168, + 0.010693547315895557, + 0.00538259232416749, + 0.012233443558216095, + 0.00730985077098012, + 0.00817439891397953, + 0.009581504389643669, + 0.010127445682883263, + 0.0036098870914429426, + 0.003229216206818819, + 0.0030734101310372353, + 0.0032681503798812628, + 0.004409943241626024, + 0.004388483706861734, + 0.00608858885243535, + 0.004343251697719097, + 0.004767715930938721, + 0.007312112022191286, + 0.006770895328372717, + 0.004941400606185198, + 0.0043688369914889336, + 0.004893345758318901, + 0.0046779802069067955, + 0.00605961075052619, + 0.027872253209352493, + 0.008877009153366089, + 0.014178582467138767, + 0.008524658158421516, + 0.009927245788276196, + 0.040515027940273285, + 0.00677480036392808, + 0.010612434707581997, + 0.0757138580083847, + 0.027727238833904266, + 0.04872975870966911, + 0.015250423923134804, + 0.04729984700679779, + 0.01708405278623104, + 0.019021129235625267, + 0.02470158413052559, + 0.006361376028507948, + 0.005721991416066885, + 0.005663985386490822, + 0.003538428572937846, + 0.007204631809145212, + 0.006268263328820467, + 0.005262685939669609, + 0.003617123467847705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.008664277382194996 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002499791095033288 + ], + "opposite_logprob_tokens": [ + -6.002499580383301 + ], + "logprob_reward": -0.002499791095033288, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002499791095033288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10291773825883865, + 0.1169995665550232, + 0.033615756779909134, + 0.07048185169696808, + 0.07516869902610779, + 0.07087330520153046, + 0.032113853842020035, + 0.14641743898391724, + 0.280947744846344, + 0.07046406716108322 + ], + "credit_map": [ + 0.011776532046496868, + 0.011379671283066273, + 0.006591044366359711, + 0.010232997126877308, + 0.019806228578090668, + 0.020550183951854706, + 0.008838965557515621, + 0.0137421153485775, + 0.012588408775627613, + 0.014908653683960438, + 0.0110278045758605, + 0.010767950676381588, + 0.026994455605745316, + 0.013422133401036263, + 0.013554399833083153, + 0.013735761865973473, + 0.002377067692577839, + 0.0039032907225191593, + 0.0016965378308668733, + 0.004213468637317419, + 0.0060493540950119495, + 0.00690170144662261, + 0.004625785630196333, + 0.003848552703857422, + 0.0035131461918354034, + 0.011127188801765442, + 0.005301195662468672, + 0.008527067489922047, + 0.015272806398570538, + 0.009829756803810596, + 0.00817530695348978, + 0.008735381998121738, + 0.0098374979570508, + 0.007125485222786665, + 0.005950900260359049, + 0.008688793517649174, + 0.005900440271943808, + 0.009329491294920444, + 0.01564048044383526, + 0.012695610523223877, + 0.008875321596860886, + 0.006555795669555664, + 0.008568977937102318, + 0.006621256936341524, + 0.00807411503046751, + 0.0106681352481246, + 0.01192049402743578, + 0.00958920270204544, + 0.005140407010912895, + 0.004915029276162386, + 0.004873827565461397, + 0.002645620610564947, + 0.0031925321090966463, + 0.0030995600391179323, + 0.004173151217401028, + 0.004073723219335079, + 0.031771618872880936, + 0.017263544723391533, + 0.023318709805607796, + 0.011066984385251999, + 0.013318379409611225, + 0.01731838658452034, + 0.01404095720499754, + 0.01831885240972042, + 0.07171691954135895, + 0.02577345073223114, + 0.04772847890853882, + 0.014492481015622616, + 0.039925120770931244, + 0.02574891597032547, + 0.02073127031326294, + 0.03483109176158905, + 0.010171039961278439, + 0.00992409698665142, + 0.01110959704965353, + 0.005483683664351702, + 0.010561352595686913, + 0.009567244909703732, + 0.006327225361019373, + 0.007319832220673561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.002499791095033288 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078457236289978 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.016078457236289978, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078457236289978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06651657074689865, + 0.13261686265468597, + 0.039245352149009705, + 0.06585323065519333, + 0.08733724057674408, + 0.06947317719459534, + 0.07781407237052917, + 0.031342167407274246, + 0.35776233673095703, + 0.07203895598649979 + ], + "credit_map": [ + 0.008449412882328033, + 0.008155746385455132, + 0.007430970668792725, + 0.007118547335267067, + 0.009349733591079712, + 0.010774513706564903, + 0.007043165620416403, + 0.008194485679268837, + 0.013083426281809807, + 0.01849018968641758, + 0.01150462869554758, + 0.012655464000999928, + 0.029293730854988098, + 0.015394560992717743, + 0.01704411581158638, + 0.01515074260532856, + 0.00484014768153429, + 0.005594300571829081, + 0.0027052259538322687, + 0.005926522891968489, + 0.005792694166302681, + 0.0057234871201217175, + 0.004658765625208616, + 0.004004210699349642, + 0.004509808495640755, + 0.007731194607913494, + 0.007586391642689705, + 0.007842550985515118, + 0.012694889679551125, + 0.010129662230610847, + 0.006985173560678959, + 0.008373555727303028, + 0.012804059311747551, + 0.007544699124991894, + 0.005671580787748098, + 0.009852671064436436, + 0.005824465304613113, + 0.0108192665502429, + 0.019827574491500854, + 0.014992916956543922, + 0.011654963716864586, + 0.006769527215510607, + 0.011568265967071056, + 0.006499221082776785, + 0.007433462422341108, + 0.007369276136159897, + 0.010035980492830276, + 0.008142481558024883, + 0.0059838672168552876, + 0.008287876844406128, + 0.011469046585261822, + 0.005814378149807453, + 0.011158016510307789, + 0.0050444528460502625, + 0.01640775054693222, + 0.013648674823343754, + 0.004835076630115509, + 0.00450980244204402, + 0.00306794629432261, + 0.0029164429288357496, + 0.0033410973846912384, + 0.007310858462005854, + 0.0025724591687321663, + 0.002788482466712594, + 0.09395332634449005, + 0.03836914524435997, + 0.05788281559944153, + 0.01944330520927906, + 0.07826987653970718, + 0.023251233622431755, + 0.025136897340416908, + 0.021455762907862663, + 0.01031495165079832, + 0.009564081206917763, + 0.010530765168368816, + 0.005829975474625826, + 0.011187415570020676, + 0.010460744611918926, + 0.007109766826033592, + 0.007041252683848143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.016078457236289978 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030438750982285 + ], + "opposite_logprob_tokens": [ + -2.350304365158081 + ], + "logprob_reward": -0.10030438750982285, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030438750982285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.129485085606575, + 0.17405764758586884, + 0.057889632880687714, + 0.08122047781944275, + 0.12164190411567688, + 0.08218155056238174, + 0.06959546357393265, + 0.03511430695652962, + 0.2125847339630127, + 0.03622914478182793 + ], + "credit_map": [ + 0.01311205979436636, + 0.019108038395643234, + 0.010292756371200085, + 0.011207995936274529, + 0.02345535159111023, + 0.02528313733637333, + 0.01273962203413248, + 0.014286133460700512, + 0.012854133732616901, + 0.02398020029067993, + 0.01515725813806057, + 0.025264661759138107, + 0.034779142588377, + 0.01872580498456955, + 0.02271583303809166, + 0.020580634474754333, + 0.003726061200723052, + 0.008658827282488346, + 0.003441350068897009, + 0.007198039907962084, + 0.009452380239963531, + 0.01056925393640995, + 0.008166167885065079, + 0.006677560042589903, + 0.003925273660570383, + 0.008343242108821869, + 0.007102177944034338, + 0.007783303037285805, + 0.017024438828229904, + 0.011639024131000042, + 0.007576036266982555, + 0.017826993018388748, + 0.007160468026995659, + 0.012721010483801365, + 0.008169503882527351, + 0.019368205219507217, + 0.01342806313186884, + 0.024580713361501694, + 0.018576376140117645, + 0.017637569457292557, + 0.010775689035654068, + 0.008097480051219463, + 0.009853781200945377, + 0.007269161753356457, + 0.010694030672311783, + 0.010118896141648293, + 0.015640966594219208, + 0.00973154790699482, + 0.006994965020567179, + 0.009350300766527653, + 0.00991719588637352, + 0.006698026321828365, + 0.007669026497751474, + 0.005179295781999826, + 0.013052945025265217, + 0.010733711533248425, + 0.006192043423652649, + 0.003794947639107704, + 0.005711798090487719, + 0.0030275974422693253, + 0.002829092787578702, + 0.007273989729583263, + 0.00305371661670506, + 0.0032311254180967808, + 0.052138615399599075, + 0.023586509749293327, + 0.03164440393447876, + 0.012486487627029419, + 0.05208989232778549, + 0.013761884532868862, + 0.015003674663603306, + 0.01187328714877367, + 0.004957577213644981, + 0.0034161657094955444, + 0.005050012841820717, + 0.003163087647408247, + 0.004985745530575514, + 0.005754829850047827, + 0.004715830087661743, + 0.004185897298157215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.10030438750982285 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An incredibly slow, steady, and deliberate upward tilt begins with a close-up shot of the hem of a red silk cheongsam, suspended against a dark, blurred background. A soft, focused sidelight casts a precise shadow on the cheongsam, highlighting the silk's smooth, watery texture and deep, vibrant red hue. As the camera steadily moves upward, it sweeps across the cheongsam's elegant side slits before focusing on a vast expanse of silk, where subtle shifts of light and shadow flow with the camera's movement. Then, an exquisitely detailed and lifelike golden phoenix embroidery gradually comes into view. The camera follows the phoenix's magnificent tail feathers as they ascend, each feather embroidered in shimmering metallic gold thread clearly visible, the fine, three-dimensional stitching appearing to stand out from the silk surface like a relief. As the camera continues to pan upward, fully revealing the phoenix's outstretched wings and raised head, its piercing eyes and gorgeous crown feathers appear with stunning realism in ultra-high-definition details. The golden thread reflects a glittering glow in the light, forming a gorgeous contrast with the warm luster of the red silk. The entire picture is full of oriental classical mystery and luxury, creating a tranquil atmosphere of focused gaze on top craftsmanship and artistic admiration.", + "seed": 11627, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_5/index_3810_seed_11627_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02255676221102476, + "vlm_logprob": -0.02255676221102476, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02255676221102476, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020226646214723587 + ], + "opposite_logprob_tokens": [ + -4.39522647857666 + ], + "logprob_reward": -0.020226646214723587, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.020226646214723587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21282872557640076, + 0.12434400618076324, + 0.06129436567425728, + 0.030536547303199768, + 0.13432170450687408, + 0.14924356341362, + 0.15515808761119843, + 0.03868742659687996, + 0.05512677878141403, + 0.03845883905887604 + ], + "credit_map": [ + 0.03223027288913727, + 0.024740921333432198, + 0.04218290001153946, + 0.026375874876976013, + 0.021815666928887367, + 0.026749901473522186, + 0.015831585973501205, + 0.02290160581469536, + 0.014104082249104977, + 0.0138062359765172, + 0.01720518246293068, + 0.010249682702124119, + 0.018718035891652107, + 0.023927070200443268, + 0.011339793913066387, + 0.014993920922279358, + 0.00689608883112669, + 0.008245402947068214, + 0.009291715919971466, + 0.011852150782942772, + 0.008657127618789673, + 0.003933940082788467, + 0.0075006731785833836, + 0.0049172681756317616, + 0.003970954101532698, + 0.0042886510491371155, + 0.004086615983396769, + 0.003802909515798092, + 0.0036489833146333694, + 0.0041327220387756824, + 0.002649773610755801, + 0.003955936059355736, + 0.013913909904658794, + 0.0112380962818861, + 0.007104622200131416, + 0.019054176285862923, + 0.010820395313203335, + 0.01224951259791851, + 0.04310976341366768, + 0.01683122292160988, + 0.00952538475394249, + 0.026033200323581696, + 0.011404207907617092, + 0.0190858393907547, + 0.01713286153972149, + 0.03227482736110687, + 0.015448621474206448, + 0.018338611349463463, + 0.008937662467360497, + 0.01570386439561844, + 0.009626049548387527, + 0.02218562178313732, + 0.02593773417174816, + 0.02790498360991478, + 0.027813630178570747, + 0.01704852655529976, + 0.0036915154196321964, + 0.004591316916048527, + 0.004688647575676441, + 0.006078313570469618, + 0.00448646629229188, + 0.005481823813170195, + 0.003222799627110362, + 0.006446541752666235, + 0.006569690071046352, + 0.008027163334190845, + 0.004629077389836311, + 0.005189984571188688, + 0.01008224580436945, + 0.00833848211914301, + 0.005659960210323334, + 0.006630176678299904, + 0.0056082275696098804, + 0.0055832639336586, + 0.004867948591709137, + 0.005483894608914852, + 0.0033918116241693497, + 0.005002694204449654, + 0.003497175872325897, + 0.005023821257054806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.020226646214723587 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008671722374856472 + ], + "opposite_logprob_tokens": [ + -4.758671760559082 + ], + "logprob_reward": -0.008671722374856472, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008671722374856472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19162918627262115, + 0.1647825688123703, + 0.060210246592760086, + 0.06536893546581268, + 0.14826591312885284, + 0.04173269122838974, + 0.1387428641319275, + 0.06579720973968506, + 0.04681563749909401, + 0.07665479928255081 + ], + "credit_map": [ + 0.030799617990851402, + 0.022510068491101265, + 0.042770445346832275, + 0.017842972651124, + 0.021113360300660133, + 0.02648850344121456, + 0.014794488437473774, + 0.015309724025428295, + 0.012910007499158382, + 0.013804775662720203, + 0.014734318479895592, + 0.010600450448691845, + 0.036192890256643295, + 0.04308383911848068, + 0.016293056309223175, + 0.01716323383152485, + 0.007804658729583025, + 0.008523148484528065, + 0.009379173628985882, + 0.005496519152075052, + 0.00813351757824421, + 0.007091509643942118, + 0.007405429147183895, + 0.006376291159540415, + 0.0043030655942857265, + 0.013447641395032406, + 0.004842495080083609, + 0.009160563349723816, + 0.007131712976843119, + 0.006739683914929628, + 0.007826981134712696, + 0.011916791088879108, + 0.01788356713950634, + 0.020986182615160942, + 0.011667357757687569, + 0.029512716457247734, + 0.013821014203131199, + 0.007754590827971697, + 0.033511366695165634, + 0.013129121623933315, + 0.004238781053572893, + 0.005394039209932089, + 0.0040323627181351185, + 0.0050269970670342445, + 0.004707931075245142, + 0.005780600011348724, + 0.005801100749522448, + 0.006750880740582943, + 0.007954436354339123, + 0.01279295515269041, + 0.008001958020031452, + 0.018252434208989143, + 0.025134796276688576, + 0.02538529597222805, + 0.027835462242364883, + 0.013385520316660404, + 0.007730033714324236, + 0.007988078519701958, + 0.007791399024426937, + 0.011223001405596733, + 0.005782056134194136, + 0.009215964935719967, + 0.006395219825208187, + 0.00967145711183548, + 0.004689273424446583, + 0.008205917663872242, + 0.0031839776784181595, + 0.004398853052407503, + 0.00728850532323122, + 0.006878380663692951, + 0.0034175836481153965, + 0.00875314325094223, + 0.009560145437717438, + 0.012966562993824482, + 0.006467246916145086, + 0.00870911031961441, + 0.00845853891223669, + 0.010807842016220093, + 0.008334547281265259, + 0.011350797489285469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.008671722374856472 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642790675163269 + ], + "opposite_logprob_tokens": [ + -4.88264274597168 + ], + "logprob_reward": -0.007642790675163269, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642790675163269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13655045628547668, + 0.04148389771580696, + 0.026707390323281288, + 0.06016845628619194, + 0.09118854999542236, + 0.04439099133014679, + 0.29153433442115784, + 0.1666128933429718, + 0.1014825701713562, + 0.03988039866089821 + ], + "credit_map": [ + 0.011174595914781094, + 0.022080713883042336, + 0.028773926198482513, + 0.017324527725577354, + 0.015225104987621307, + 0.018392961472272873, + 0.010232103057205677, + 0.01334652304649353, + 0.003225957276299596, + 0.004575255326926708, + 0.0037118056789040565, + 0.0027985635679215193, + 0.010004864074289799, + 0.008574757725000381, + 0.004269307944923639, + 0.0043233847245574, + 0.0025133525487035513, + 0.002301927190274, + 0.0031403652392327785, + 0.007547020446509123, + 0.0032725296914577484, + 0.0021141215693205595, + 0.0037241417448967695, + 0.002093931194394827, + 0.004862183704972267, + 0.014014448039233685, + 0.005633485969156027, + 0.006309143733233213, + 0.0051749940030276775, + 0.005171492230147123, + 0.005172736011445522, + 0.013829976320266724, + 0.010285557247698307, + 0.007710212375968695, + 0.0063074869103729725, + 0.012343956157565117, + 0.007406897377222776, + 0.007999209687113762, + 0.026976631954312325, + 0.012158592231571674, + 0.002051111776381731, + 0.006066438276320696, + 0.0027940343134105206, + 0.0060409302823245525, + 0.004496464505791664, + 0.009084067307412624, + 0.005113693885505199, + 0.00874425284564495, + 0.018177704885601997, + 0.027513129636645317, + 0.01860528253018856, + 0.04413141310214996, + 0.06058671325445175, + 0.040856990963220596, + 0.04904577508568764, + 0.03261733427643776, + 0.016320841386914253, + 0.024784142151474953, + 0.01584656722843647, + 0.020056212320923805, + 0.019295504316687584, + 0.025733985006809235, + 0.01687881164252758, + 0.027696821838617325, + 0.011285767890512943, + 0.020153911784291267, + 0.007662506774067879, + 0.00934674870222807, + 0.015764987096190453, + 0.01328288670629263, + 0.0072733028791844845, + 0.016712455078959465, + 0.004426253493875265, + 0.006413758732378483, + 0.0028724740259349346, + 0.00698392977938056, + 0.0038933604955673218, + 0.004287362098693848, + 0.003795795841142535, + 0.0072074634954333305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.007642790675163269 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420646533370018 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.01420646533370018, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420646533370018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10473515093326569, + 0.05164986476302147, + 0.0915253683924675, + 0.05554712936282158, + 0.19327613711357117, + 0.08547011762857437, + 0.11834699660539627, + 0.0828966423869133, + 0.09490343928337097, + 0.12164913862943649 + ], + "credit_map": [ + 0.014399860985577106, + 0.012036358937621117, + 0.014013216830790043, + 0.020043279975652695, + 0.013155919499695301, + 0.010499953292310238, + 0.012322353199124336, + 0.008264209143817425, + 0.003699149237945676, + 0.004833877086639404, + 0.007078866008669138, + 0.004616499878466129, + 0.012147767469286919, + 0.009098473936319351, + 0.0051806322298944, + 0.004994602408260107, + 0.008282871916890144, + 0.011601562611758709, + 0.009257632307708263, + 0.021155882626771927, + 0.016871903091669083, + 0.0069884550757706165, + 0.010480772703886032, + 0.006886294577270746, + 0.005281522404402494, + 0.011871934868395329, + 0.00444110156968236, + 0.007126863580197096, + 0.0048250677064061165, + 0.005570642650127411, + 0.005070788785815239, + 0.011359209194779396, + 0.019512474536895752, + 0.01680501364171505, + 0.011039365082979202, + 0.02975161001086235, + 0.015893612056970596, + 0.014541874639689922, + 0.06716066598892212, + 0.0185715164989233, + 0.005084025673568249, + 0.009514556266367435, + 0.0045586940832436085, + 0.012250079773366451, + 0.007437247317284346, + 0.01772603765130043, + 0.010206452570855618, + 0.018693022429943085, + 0.007401023060083389, + 0.0099464301019907, + 0.007616829592734575, + 0.017298810184001923, + 0.022472629323601723, + 0.019371358677744865, + 0.022381562739610672, + 0.011858345940709114, + 0.008734449744224548, + 0.010400974191725254, + 0.010587893426418304, + 0.011849398724734783, + 0.00845959223806858, + 0.013586839661002159, + 0.006053201854228973, + 0.013224294409155846, + 0.011313608847558498, + 0.016445988789200783, + 0.008218144997954369, + 0.007942836731672287, + 0.015436888672411442, + 0.011961665004491806, + 0.007628621067851782, + 0.015955684706568718, + 0.014250666834414005, + 0.020064668729901314, + 0.009437245316803455, + 0.01664067432284355, + 0.013925591483712196, + 0.011244356632232666, + 0.014299959875643253, + 0.021785981953144073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.01420646533370018 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620361864566803 + ], + "opposite_logprob_tokens": [ + -2.8120362758636475 + ], + "logprob_reward": -0.0620361864566803, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620361864566803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1591024547815323, + 0.07294654846191406, + 0.05974401533603668, + 0.13430380821228027, + 0.2377932369709015, + 0.06349041312932968, + 0.056071121245622635, + 0.0841282457113266, + 0.09494896233081818, + 0.037471186369657516 + ], + "credit_map": [ + 0.020621422678232193, + 0.017096927389502525, + 0.03313502296805382, + 0.017295368015766144, + 0.015822477638721466, + 0.020220255479216576, + 0.013062318786978722, + 0.021848665550351143, + 0.006382766645401716, + 0.011342901736497879, + 0.011589409783482552, + 0.0065329596400260925, + 0.01192812342196703, + 0.0070326304994523525, + 0.010968215763568878, + 0.007169537246227264, + 0.00509403133764863, + 0.007844571024179459, + 0.008527500554919243, + 0.010754388757050037, + 0.007306888233870268, + 0.005544133018702269, + 0.009400981478393078, + 0.0052715172059834, + 0.012081019580364227, + 0.023341653868556023, + 0.011906245723366737, + 0.015346973203122616, + 0.015923211351037025, + 0.011561828665435314, + 0.017619552090764046, + 0.026523323729634285, + 0.02399079129099846, + 0.01740146242082119, + 0.011439655907452106, + 0.03699975833296776, + 0.018558558076620102, + 0.020267819985747337, + 0.08053162693977356, + 0.028603581711649895, + 0.005309511441737413, + 0.011717338114976883, + 0.010931698605418205, + 0.007919857278466225, + 0.0058355615474283695, + 0.005593857262283564, + 0.010008255951106548, + 0.006174334790557623, + 0.0034914410207420588, + 0.005644312594085932, + 0.003529905341565609, + 0.008815896697342396, + 0.011014421470463276, + 0.007967117242515087, + 0.009295042604207993, + 0.00631298404186964, + 0.0094320522621274, + 0.010291649959981441, + 0.010590601712465286, + 0.010560483671724796, + 0.010569548234343529, + 0.011404355056583881, + 0.008406431414186954, + 0.012873120605945587, + 0.012514298781752586, + 0.014516892842948437, + 0.009140816517174244, + 0.007856144569814205, + 0.014675001613795757, + 0.011700253933668137, + 0.008342072367668152, + 0.01620347797870636, + 0.005030447151511908, + 0.00524642551317811, + 0.003663248149678111, + 0.004300935659557581, + 0.004468675237149, + 0.006065584719181061, + 0.0035804291255772114, + 0.00511544244363904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0620361864566803 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An incredibly slow, steady, and deliberate upward tilt begins with a close-up shot of the hem of a red silk cheongsam, suspended against a dark, blurred background. A soft, focused sidelight casts a precise shadow on the cheongsam, highlighting the silk's smooth, watery texture and deep, vibrant red hue. As the camera steadily moves upward, it sweeps across the cheongsam's elegant side slits before focusing on a vast expanse of silk, where subtle shifts of light and shadow flow with the camera's movement. Then, an exquisitely detailed and lifelike golden phoenix embroidery gradually comes into view. The camera follows the phoenix's magnificent tail feathers as they ascend, each feather embroidered in shimmering metallic gold thread clearly visible, the fine, three-dimensional stitching appearing to stand out from the silk surface like a relief. As the camera continues to pan upward, fully revealing the phoenix's outstretched wings and raised head, its piercing eyes and gorgeous crown feathers appear with stunning realism in ultra-high-definition details. The golden thread reflects a glittering glow in the light, forming a gorgeous contrast with the warm luster of the red silk. The entire picture is full of oriental classical mystery and luxury, creating a tranquil atmosphere of focused gaze on top craftsmanship and artistic admiration.", + "seed": 11628, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_5/index_3810_seed_11628_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.028140408103354275, + "vlm_logprob": -0.028140408103354275, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.028140408103354275, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02926012873649597 + ], + "opposite_logprob_tokens": [ + -3.7792601585388184 + ], + "logprob_reward": -0.02926012873649597, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02926012873649597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19589708745479584, + 0.11083734780550003, + 0.18089640140533447, + 0.05715460330247879, + 0.10622777789831161, + 0.14610449969768524, + 0.04057350382208824, + 0.019018562510609627, + 0.015124328434467316, + 0.12816590070724487 + ], + "credit_map": [ + 0.019599711522459984, + 0.018911147490143776, + 0.030762871727347374, + 0.02752598747611046, + 0.026288552209734917, + 0.01719650998711586, + 0.016988996416330338, + 0.038623325526714325, + 0.013000795617699623, + 0.010369434021413326, + 0.006714932154864073, + 0.007894656620919704, + 0.01581222377717495, + 0.011464525014162064, + 0.036038048565387726, + 0.009542728774249554, + 0.005917635280638933, + 0.05686021223664284, + 0.00591878779232502, + 0.011513380333781242, + 0.016212131828069687, + 0.06508038192987442, + 0.009546167217195034, + 0.009847687557339668, + 0.003794287098571658, + 0.004089706111699343, + 0.003263927763327956, + 0.01821039989590645, + 0.009750491008162498, + 0.008997070603072643, + 0.003287785453721881, + 0.005760934669524431, + 0.003905736142769456, + 0.01112352591007948, + 0.005318755283951759, + 0.015261463820934296, + 0.01377212442457676, + 0.009246679954230785, + 0.018950920552015305, + 0.02864857017993927, + 0.01661962829530239, + 0.016328396275639534, + 0.022670937702059746, + 0.019138112664222717, + 0.0207330584526062, + 0.019617673009634018, + 0.017324956133961678, + 0.013671728782355785, + 0.0023071684408932924, + 0.0036865894217044115, + 0.003601226955652237, + 0.007940294221043587, + 0.00467121135443449, + 0.006134655326604843, + 0.005231000948697329, + 0.007001357153058052, + 0.0016187336295843124, + 0.0017673657275736332, + 0.001058446941897273, + 0.004482891410589218, + 0.002299301791936159, + 0.0020962804555892944, + 0.002729108789935708, + 0.0029664330650120974, + 0.0023580132983624935, + 0.001343645271845162, + 0.0019329198403283954, + 0.002177736023440957, + 0.002044440247118473, + 0.0021871868520975113, + 0.0013190305326133966, + 0.0017613550880923867, + 0.028408071026206017, + 0.006629284471273422, + 0.03165683522820473, + 0.01017153263092041, + 0.007958459667861462, + 0.007001260761171579, + 0.008581401780247688, + 0.02775905281305313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.02926012873649597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00866853166371584 + ], + "opposite_logprob_tokens": [ + -4.758668422698975 + ], + "logprob_reward": -0.00866853166371584, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00866853166371584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07565036416053772, + 0.12915033102035522, + 0.5588584542274475, + 0.0376356802880764, + 0.07022079825401306, + 0.01953871175646782, + 0.04834280163049698, + 0.013567131012678146, + 0.006672338582575321, + 0.0403633676469326 + ], + "credit_map": [ + 0.008722021244466305, + 0.005172072444111109, + 0.006226050667464733, + 0.00820164941251278, + 0.012242281809449196, + 0.006711876019835472, + 0.0047155339270830154, + 0.023658884689211845, + 0.013699375092983246, + 0.010910426266491413, + 0.006334747187793255, + 0.00897781178355217, + 0.023518158122897148, + 0.012532275170087814, + 0.04023192450404167, + 0.012945609167218208, + 0.012541319243609905, + 0.1828448325395584, + 0.012835153378546238, + 0.028341202065348625, + 0.04704377055168152, + 0.22430285811424255, + 0.01503818854689598, + 0.03591112047433853, + 0.0021532210521399975, + 0.0038580370601266623, + 0.0022589846048504114, + 0.0048034111969172955, + 0.00807200651615858, + 0.006961897015571594, + 0.0031328073237091303, + 0.006395313423126936, + 0.002598585095256567, + 0.004733198322355747, + 0.002976557007059455, + 0.016539691016077995, + 0.004840769339352846, + 0.006420132238417864, + 0.007292299065738916, + 0.02481956034898758, + 0.001084324554540217, + 0.002176701556891203, + 0.001269295229576528, + 0.002550744218751788, + 0.003647133009508252, + 0.0027931213844567537, + 0.003886162769049406, + 0.0021312283352017403, + 0.0033321601804345846, + 0.003375230822712183, + 0.0027084643952548504, + 0.009413285180926323, + 0.005279615055769682, + 0.006515657063573599, + 0.009502217173576355, + 0.008216171525418758, + 0.0010620758403092623, + 0.0013165930286049843, + 0.0013553788885474205, + 0.0026977071538567543, + 0.0011444612173363566, + 0.0012344735441729426, + 0.0026612551882863045, + 0.002095186384394765, + 0.0008512776112183928, + 0.0005825530970469117, + 0.0009968553204089403, + 0.0011559441918507218, + 0.0008339257910847664, + 0.0008531212224625051, + 0.000652182730846107, + 0.0007464785594493151, + 0.009611035697162151, + 0.004519072826951742, + 0.009436588734388351, + 0.00392120098695159, + 0.002971756039187312, + 0.00330704590305686, + 0.0036868436727672815, + 0.0029098261147737503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.00866853166371584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00119138287845999 + ], + "opposite_logprob_tokens": [ + -6.75119161605835 + ], + "logprob_reward": -0.00119138287845999, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.00119138287845999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09111572057008743, + 0.055999793112277985, + 0.09943582117557526, + 0.07556956261396408, + 0.11324560642242432, + 0.03194376081228256, + 0.1062244400382042, + 0.050885945558547974, + 0.0389321967959404, + 0.3366471230983734 + ], + "credit_map": [ + 0.010245829820632935, + 0.008320644497871399, + 0.007930099032819271, + 0.009937292896211147, + 0.013829998672008514, + 0.009069503284990788, + 0.007516831159591675, + 0.02426552213728428, + 0.003616223344579339, + 0.00447909627109766, + 0.004038736689835787, + 0.0038261101581156254, + 0.006823204457759857, + 0.004561330657452345, + 0.020115574821829796, + 0.008539516478776932, + 0.0067499130964279175, + 0.030762212350964546, + 0.0036081974394619465, + 0.007790792267769575, + 0.009212289936840534, + 0.03042166866362095, + 0.005411526653915644, + 0.005479224491864443, + 0.0052150241099298, + 0.005163630936294794, + 0.00488075939938426, + 0.025303252041339874, + 0.012335115112364292, + 0.01332154031842947, + 0.003902249038219452, + 0.005447996314615011, + 0.003127911826595664, + 0.009305656887590885, + 0.004265441559255123, + 0.02170325629413128, + 0.013517813757061958, + 0.011823983862996101, + 0.008556578308343887, + 0.04094495624303818, + 0.0024348206352442503, + 0.002574447076767683, + 0.0033058193512260914, + 0.004528403282165527, + 0.0041123200207948685, + 0.004998185206204653, + 0.006666174158453941, + 0.003323592944070697, + 0.010884522460401058, + 0.013772540725767612, + 0.008151048794388771, + 0.021539002656936646, + 0.014617782086133957, + 0.011759801767766476, + 0.012636878527700901, + 0.012862865813076496, + 0.003269367618486285, + 0.0039576017297804356, + 0.003219447797164321, + 0.011445111595094204, + 0.0053112865425646305, + 0.005185924004763365, + 0.008205503225326538, + 0.010291706770658493, + 0.005741894710808992, + 0.0029185188468545675, + 0.0040373001247644424, + 0.004229431971907616, + 0.005671004764735699, + 0.005716521758586168, + 0.003043915843591094, + 0.007573605980724096, + 0.07908346503973007, + 0.016401462256908417, + 0.08756919205188751, + 0.02559269778430462, + 0.019656896591186523, + 0.01566139981150627, + 0.01984548196196556, + 0.07283653318881989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.00119138287845999 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547007761895657 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012547007761895657, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547007761895657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10565955936908722, + 0.06293664872646332, + 0.2945997714996338, + 0.05230782926082611, + 0.14685815572738647, + 0.02704034559428692, + 0.060871291905641556, + 0.02430715411901474, + 0.03408641740679741, + 0.19133280217647552 + ], + "credit_map": [ + 0.011433454230427742, + 0.008703751489520073, + 0.00935071986168623, + 0.0076031493954360485, + 0.016724897548556328, + 0.010828849859535694, + 0.008011131547391415, + 0.03300360217690468, + 0.005694712977856398, + 0.005036099813878536, + 0.004693424329161644, + 0.005226782988756895, + 0.009528541937470436, + 0.004964946769177914, + 0.01802161894738674, + 0.009770513512194157, + 0.00750411581248045, + 0.09523380547761917, + 0.007537751458585262, + 0.015593374148011208, + 0.024377543479204178, + 0.11661367118358612, + 0.008351952768862247, + 0.019387545064091682, + 0.0030831473413854837, + 0.0038014890160411596, + 0.003433759091421962, + 0.016510961577296257, + 0.00924447737634182, + 0.00926801934838295, + 0.0029096438083797693, + 0.004056328907608986, + 0.003972450736910105, + 0.0119808129966259, + 0.004668604116886854, + 0.022194698452949524, + 0.012821908108890057, + 0.015385529957711697, + 0.020983342081308365, + 0.05485080927610397, + 0.001791716436855495, + 0.0029719602316617966, + 0.002172335283830762, + 0.004068094305694103, + 0.004417198244482279, + 0.004348009824752808, + 0.004413495305925608, + 0.002857533283531666, + 0.005355376750230789, + 0.008564370684325695, + 0.004121436737477779, + 0.010822818614542484, + 0.006825130898505449, + 0.007355656009167433, + 0.009002500213682652, + 0.008824000135064125, + 0.0020236491691321135, + 0.001961109461262822, + 0.0020245597697794437, + 0.004896524827927351, + 0.0026014239992946386, + 0.0018943684408441186, + 0.004602915607392788, + 0.004302600398659706, + 0.005691036581993103, + 0.0028877833392471075, + 0.003537082113325596, + 0.005478110630065203, + 0.004132137633860111, + 0.004662980791181326, + 0.0033519575372338295, + 0.004345325753092766, + 0.03214583173394203, + 0.021119486540555954, + 0.033682242035865784, + 0.01671699807047844, + 0.02018841914832592, + 0.017393555492162704, + 0.017956577241420746, + 0.032129690051078796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.012547007761895657 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a close-up shot of the hem of a red silk cheongsam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a slow, steady, and deliberate upward tilt throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a golden phoenix embroidery visible on the cheongsam with shimmering metallic gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the phoenix's outstretched wings, raised head, piercing eyes, and crown feathers clearly visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background dark and blurred with a soft, focused sidelight casting precise shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903498947620392 + ], + "opposite_logprob_tokens": [ + -2.4640350341796875 + ], + "logprob_reward": -0.08903498947620392, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903498947620392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1752208024263382, + 0.10166791826486588, + 0.3642998933792114, + 0.09413618594408035, + 0.1003817468881607, + 0.026494400575757027, + 0.03389034792780876, + 0.010763730853796005, + 0.009633001871407032, + 0.08351201564073563 + ], + "credit_map": [ + 0.014409935101866722, + 0.02120426669716835, + 0.020235072821378708, + 0.024793993681669235, + 0.02585049532353878, + 0.01834321953356266, + 0.013535561971366405, + 0.03684825450181961, + 0.014612858183681965, + 0.009161663241684437, + 0.01190900057554245, + 0.006973112002015114, + 0.015746213495731354, + 0.010485587641596794, + 0.023675596341490746, + 0.009103886783123016, + 0.008897324092686176, + 0.11715853959321976, + 0.008620833978056908, + 0.018645459786057472, + 0.031716205179691315, + 0.14469096064567566, + 0.010544783435761929, + 0.02402578294277191, + 0.0032272355165332556, + 0.006720016244798899, + 0.004319473635405302, + 0.020633842796087265, + 0.01872292347252369, + 0.020121527835726738, + 0.008305300958454609, + 0.012085853144526482, + 0.003612436354160309, + 0.009851524606347084, + 0.004062945954501629, + 0.010822669602930546, + 0.033560190349817276, + 0.009250138886272907, + 0.021974103525280952, + 0.007247732020914555, + 0.0017178545240312815, + 0.002580940490588546, + 0.0028908075764775276, + 0.003049355000257492, + 0.004728297702968121, + 0.003677894128486514, + 0.005216413643211126, + 0.0026328368112444878, + 0.0028072537388652563, + 0.00478637358173728, + 0.003107908181846142, + 0.005925263278186321, + 0.0036162170581519604, + 0.004567802418023348, + 0.00447285920381546, + 0.004606668371707201, + 0.00076109025394544, + 0.0008410679874941707, + 0.0008432306349277496, + 0.002661661244928837, + 0.0009129295940510929, + 0.0015467301709577441, + 0.0016696613747626543, + 0.0015273595927283168, + 0.0013855487341061234, + 0.0008543561561964452, + 0.0010224553989246488, + 0.0012443203013390303, + 0.001879509654827416, + 0.0014640504959970713, + 0.0008504687575623393, + 0.0009322923724539578, + 0.020273203030228615, + 0.0035522517282515764, + 0.022529419511556625, + 0.006110678426921368, + 0.004549335688352585, + 0.003891587955877185, + 0.00480511924251914, + 0.017800424247980118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.08903498947620392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_6/index_tensor([3107, 3107])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_6/index_tensor([3107, 3107])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b94180c7e1da830ab6097bd1d24db31b8c122fac --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_6/index_tensor([3107, 3107])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, sideways tracking shot closely follows a mythical white tiger as it gallops through an ancient, dense forest at night. Cold moonlight filters through the gaps between the tall, twisted canopies of ancient trees, casting a variegated, dynamic play of light and shadow across a ground covered in dead leaves and damp earth. The beast's snow-white, unblemished fur shimmers with a soft silver halo in the moonlight. Its muscular form ripples clearly beneath its fur as it gallops, and its icy blue eyes gleam with an otherworldly light in the darkness. Its limbs fully extended, it gallops with explosive force, its massive paws barely leaving a trace as it stamped across the ground. The trees in the background are reduced to blurred vertical lines by its breathtaking speed. As it approaches a massive, moss-covered fallen log that lies across the path, it doesn't slow at all. Its hind legs thrust forward, its entire body tracing an elegant, powerful white arc through the air. It effortlessly clears the obstacle, landing firmly on the other side and continuing its charge. Its breath becomes a visible white mist in the chill air. The image is dominated by deep midnight blue and inky black, with the white tiger itself as the most striking presence, creating an atmosphere imbued with a sense of primal power, mysterious elegance, and the beauty of speed.", + "seed": 8016, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_6/index_3107_seed_8016_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018656712863594292, + "vlm_logprob": -0.018656712863594292, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.018656712863594292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016165494918823242 + ], + "opposite_logprob_tokens": [ + -4.391165733337402 + ], + "logprob_reward": -0.016165494918823242, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016165494918823242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1407492309808731, + 0.12928013503551483, + 0.04582342505455017, + 0.035975366830825806, + 0.01563934050500393, + 0.07734838873147964, + 0.11477217078208923, + 0.10102207958698273, + 0.24731284379959106, + 0.09207699447870255 + ], + "credit_map": [ + 0.022243214771151543, + 0.015149952843785286, + 0.04275454580783844, + 0.008398805744946003, + 0.009711295366287231, + 0.022720953449606895, + 0.0065053915604949, + 0.013265067711472511, + 0.02055078186094761, + 0.016002889722585678, + 0.016308626160025597, + 0.025137847289443016, + 0.010064343921840191, + 0.01563332788646221, + 0.009143423289060593, + 0.01643889956176281, + 0.006039655767381191, + 0.004882416222244501, + 0.008648928254842758, + 0.004631425719708204, + 0.0038983067497611046, + 0.004558899439871311, + 0.004627432208508253, + 0.008536364883184433, + 0.004269185476005077, + 0.0038478032220155, + 0.0052378857508301735, + 0.004621751606464386, + 0.003260272089391947, + 0.003354887245222926, + 0.004582026042044163, + 0.006801558192819357, + 0.0017812635051086545, + 0.0015335498610511422, + 0.0016707094619050622, + 0.0018292156746611, + 0.0018224079394713044, + 0.002475056331604719, + 0.0018014098750427365, + 0.00272572785615921, + 0.01309975702315569, + 0.005146956071257591, + 0.0076608676463365555, + 0.005553014110773802, + 0.00819526705890894, + 0.016433965414762497, + 0.011067429557442665, + 0.010191135108470917, + 0.015537293627858162, + 0.011627773754298687, + 0.013059930875897408, + 0.012062261812388897, + 0.014128739014267921, + 0.016609404236078262, + 0.01842227205634117, + 0.013324490748345852, + 0.0266068447381258, + 0.005935876630246639, + 0.01398586854338646, + 0.007083606906235218, + 0.006204394623637199, + 0.007478886749595404, + 0.010579891502857208, + 0.02314671128988266, + 0.02247721515595913, + 0.016789717599749565, + 0.03147204965353012, + 0.0545901283621788, + 0.01749434880912304, + 0.017248330637812614, + 0.05671854317188263, + 0.030522530898451805, + 0.01427827961742878, + 0.008589374832808971, + 0.014498746953904629, + 0.017314624041318893, + 0.007682109251618385, + 0.007097349967807531, + 0.011455723084509373, + 0.011160784400999546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.016165494918823242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328958734869957 + ], + "opposite_logprob_tokens": [ + -3.773289680480957 + ], + "logprob_reward": -0.02328958734869957, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328958734869957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1665721982717514, + 0.11535145342350006, + 0.023909948766231537, + 0.02934185042977333, + 0.021734172478318214, + 0.15721145272254944, + 0.023056212812662125, + 0.29880407452583313, + 0.12677550315856934, + 0.037243183702230453 + ], + "credit_map": [ + 0.027070319280028343, + 0.010044961236417294, + 0.05501076206564903, + 0.011353646405041218, + 0.009045185521245003, + 0.029319779947400093, + 0.009887688793241978, + 0.014839865267276764, + 0.01676800288259983, + 0.012467087246477604, + 0.014446811750531197, + 0.023699255660176277, + 0.012809072621166706, + 0.011632065288722515, + 0.008221620693802834, + 0.015307535417377949, + 0.0022221768740564585, + 0.0015341606922447681, + 0.0039226035587489605, + 0.0018548137741163373, + 0.001636132481507957, + 0.0024065810721367598, + 0.0038865075912326574, + 0.006446970626711845, + 0.003172489581629634, + 0.004016004037111998, + 0.0031214512418955564, + 0.0054112705402076244, + 0.003556883195415139, + 0.0025453362613916397, + 0.004175554495304823, + 0.00334286130964756, + 0.0021355433855205774, + 0.0032794733997434378, + 0.0019335087854415178, + 0.0034270526375621557, + 0.003261611331254244, + 0.0024044562596827745, + 0.002173343440517783, + 0.0031191830057650805, + 0.03714374080300331, + 0.007625132333487272, + 0.015122070908546448, + 0.004744977690279484, + 0.00960537139326334, + 0.04659641534090042, + 0.011022119782865047, + 0.025351617485284805, + 0.00424193823710084, + 0.0023333467543125153, + 0.0020423864480108023, + 0.0024586357176303864, + 0.002554756123572588, + 0.0029921233654022217, + 0.0022375104017555714, + 0.004195517394691706, + 0.07490256428718567, + 0.014878896996378899, + 0.04882284626364708, + 0.015736933797597885, + 0.012689322233200073, + 0.015854349359869957, + 0.0371033251285553, + 0.07881583273410797, + 0.01160639151930809, + 0.008385497145354748, + 0.015582109801471233, + 0.028912747278809547, + 0.00880348589271307, + 0.007852466776967049, + 0.03078320622444153, + 0.014849592000246048, + 0.007331098895519972, + 0.0031127429101616144, + 0.004956128541380167, + 0.006244522985070944, + 0.0021095979027450085, + 0.002607241040095687, + 0.003934485372155905, + 0.006947366986423731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.02328958734869957 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195770025253296 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195770025253296, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195770025253296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19155587255954742, + 0.06456474959850311, + 0.07086778432130814, + 0.026614537462592125, + 0.01774270460009575, + 0.05795751512050629, + 0.015252368524670601, + 0.24104462563991547, + 0.2589949071407318, + 0.055404871702194214 + ], + "credit_map": [ + 0.028998341411352158, + 0.010450120083987713, + 0.06266027688980103, + 0.011085844598710537, + 0.014530342072248459, + 0.036542247980833054, + 0.010389966890215874, + 0.0168987475335598, + 0.008944264613091946, + 0.006507060956209898, + 0.008785867132246494, + 0.013813932426273823, + 0.0053652189671993256, + 0.00738843809813261, + 0.004278361797332764, + 0.009481613524258137, + 0.006908278446644545, + 0.005683980882167816, + 0.011635641567409039, + 0.005968618206679821, + 0.0058172037824988365, + 0.006869118195027113, + 0.010452830232679844, + 0.017532125115394592, + 0.0052929287776350975, + 0.002281275112181902, + 0.003834896720945835, + 0.002754700370132923, + 0.002205398865044117, + 0.0030196153093129396, + 0.003146866802126169, + 0.0040788608603179455, + 0.0022179805673658848, + 0.0017855813493952155, + 0.002609632443636656, + 0.002215938875451684, + 0.0019279307452961802, + 0.0026817885227501392, + 0.0021508035715669394, + 0.002153051784262061, + 0.012358215637505054, + 0.0030067393090575933, + 0.005752152763307095, + 0.0037180506624281406, + 0.0033289589919149876, + 0.014606896787881851, + 0.0069854892790317535, + 0.00820101983845234, + 0.0026506115682423115, + 0.00172690418548882, + 0.0013925298117101192, + 0.0014752858551219106, + 0.0021619766484946012, + 0.002227053977549076, + 0.0018816175870597363, + 0.0017363918013870716, + 0.03250059485435486, + 0.012194373644888401, + 0.06449227780103683, + 0.016126330941915512, + 0.01879878155887127, + 0.02171539142727852, + 0.023801680654287338, + 0.051415231078863144, + 0.021556654945015907, + 0.017641928046941757, + 0.03251158446073532, + 0.05852869525551796, + 0.01815076917409897, + 0.01783278025686741, + 0.06385667622089386, + 0.028915848582983017, + 0.00783802941441536, + 0.006075476296246052, + 0.010228127241134644, + 0.010460375808179379, + 0.004862288478761911, + 0.004540286958217621, + 0.006013898644596338, + 0.005386394914239645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.014195770025253296 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368911147117615 + ], + "opposite_logprob_tokens": [ + -3.408689022064209 + ], + "logprob_reward": -0.03368911147117615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368911147117615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36392027139663696, + 0.0962931364774704, + 0.14248788356781006, + 0.0556042343378067, + 0.028236133977770805, + 0.06668290495872498, + 0.021840479224920273, + 0.06667371839284897, + 0.08643150329589844, + 0.07182968407869339 + ], + "credit_map": [ + 0.054797910153865814, + 0.0351913720369339, + 0.11051730066537857, + 0.023071659728884697, + 0.02004319615662098, + 0.06104889139533043, + 0.02114824950695038, + 0.03810170292854309, + 0.022753801196813583, + 0.008066817186772823, + 0.01597112976014614, + 0.009015616960823536, + 0.005619892850518227, + 0.015817750245332718, + 0.0060478998348116875, + 0.013000234961509705, + 0.021630622446537018, + 0.00949215330183506, + 0.027037110179662704, + 0.010904095135629177, + 0.008059198968112469, + 0.015770068392157555, + 0.016361959278583527, + 0.03323269262909889, + 0.011663287878036499, + 0.00499134324491024, + 0.008646094240248203, + 0.0048072380013763905, + 0.005343710072338581, + 0.004991465248167515, + 0.007181765045970678, + 0.007979332469403744, + 0.004029445815831423, + 0.0033681239001452923, + 0.0034826791379600763, + 0.0033163693733513355, + 0.002668221015483141, + 0.0049338373355567455, + 0.002556295832619071, + 0.003881163662299514, + 0.012988610193133354, + 0.004088690038770437, + 0.007770399563014507, + 0.004987394902855158, + 0.005686281714588404, + 0.011166547425091267, + 0.011396098881959915, + 0.008598888292908669, + 0.004704844206571579, + 0.001876465161330998, + 0.002187945181503892, + 0.0024227844551205635, + 0.002246410818770528, + 0.00310420710593462, + 0.0020448637660592794, + 0.003252959344536066, + 0.015915971249341965, + 0.0035319565795361996, + 0.010143239982426167, + 0.005935057997703552, + 0.0037896214053034782, + 0.004826069343835115, + 0.0075649539940059185, + 0.014966849237680435, + 0.008633585646748543, + 0.005159887485206127, + 0.010717455297708511, + 0.020447738468647003, + 0.006210456136614084, + 0.0041825720109045506, + 0.022032931447029114, + 0.009046882390975952, + 0.009175794199109077, + 0.007692029699683189, + 0.011075391434133053, + 0.014291197061538696, + 0.00643721641972661, + 0.005839013960212469, + 0.009975898079574108, + 0.007343153469264507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03368911147117615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943600554019213 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943600554019213, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943600554019213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28642916679382324, + 0.1354280561208725, + 0.031719181686639786, + 0.01681176945567131, + 0.03145391494035721, + 0.07793030887842178, + 0.054017942398786545, + 0.21155484020709991, + 0.1151181235909462, + 0.0395367294549942 + ], + "credit_map": [ + 0.042827218770980835, + 0.01816616579890251, + 0.08981961756944656, + 0.019363826140761375, + 0.021794641390442848, + 0.05231606215238571, + 0.017307061702013016, + 0.024834565818309784, + 0.020319238305091858, + 0.010156085714697838, + 0.02444353699684143, + 0.017723601311445236, + 0.009498225525021553, + 0.023311937227845192, + 0.008729911409318447, + 0.021245505660772324, + 0.00560538936406374, + 0.0026017234195023775, + 0.005077141337096691, + 0.003301147138699889, + 0.0027226097881793976, + 0.0031706644222140312, + 0.003930388018488884, + 0.005310114473104477, + 0.0022484969813376665, + 0.002312843222171068, + 0.001795666292309761, + 0.002902328036725521, + 0.0015304327243939042, + 0.0017275979043915868, + 0.001958981854841113, + 0.002335422206670046, + 0.003072526538744569, + 0.003484776709228754, + 0.0029489092994481325, + 0.004477276932448149, + 0.003798736957833171, + 0.0035989994648844004, + 0.006043465808033943, + 0.004029220901429653, + 0.017537251114845276, + 0.004545479081571102, + 0.008764904923737049, + 0.0034540528431534767, + 0.004923646803945303, + 0.019977940246462822, + 0.007020852994173765, + 0.011706173419952393, + 0.006930225063115358, + 0.005808920133858919, + 0.0062574585899710655, + 0.005657467991113663, + 0.008382314816117287, + 0.006851919461041689, + 0.006619039922952652, + 0.007510589901357889, + 0.028092777356505394, + 0.014025374315679073, + 0.06985808163881302, + 0.014337620697915554, + 0.011413597501814365, + 0.012279094196856022, + 0.02162913605570793, + 0.03991913050413132, + 0.010369368828833103, + 0.007174885831773281, + 0.014160796999931335, + 0.026910139247775078, + 0.008067319169640541, + 0.008703065104782581, + 0.02642521820962429, + 0.01330731250345707, + 0.006812155246734619, + 0.0039037007372826338, + 0.005142869893461466, + 0.006479712203145027, + 0.003036730457097292, + 0.003038412658497691, + 0.0048620328307151794, + 0.006261113099753857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.005943600554019213 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, sideways tracking shot closely follows a mythical white tiger as it gallops through an ancient, dense forest at night. Cold moonlight filters through the gaps between the tall, twisted canopies of ancient trees, casting a variegated, dynamic play of light and shadow across a ground covered in dead leaves and damp earth. The beast's snow-white, unblemished fur shimmers with a soft silver halo in the moonlight. Its muscular form ripples clearly beneath its fur as it gallops, and its icy blue eyes gleam with an otherworldly light in the darkness. Its limbs fully extended, it gallops with explosive force, its massive paws barely leaving a trace as it stamped across the ground. The trees in the background are reduced to blurred vertical lines by its breathtaking speed. As it approaches a massive, moss-covered fallen log that lies across the path, it doesn't slow at all. Its hind legs thrust forward, its entire body tracing an elegant, powerful white arc through the air. It effortlessly clears the obstacle, landing firmly on the other side and continuing its charge. Its breath becomes a visible white mist in the chill air. The image is dominated by deep midnight blue and inky black, with the white tiger itself as the most striking presence, creating an atmosphere imbued with a sense of primal power, mysterious elegance, and the beauty of speed.", + "seed": 8017, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_6/index_3107_seed_8017_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019835827685892583, + "vlm_logprob": -0.019835827685892583, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.019835827685892583, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024204934015870094 + ], + "opposite_logprob_tokens": [ + -3.899204969406128 + ], + "logprob_reward": -0.024204934015870094, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024204934015870094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10248927772045135, + 0.054309844970703125, + 0.05114082247018814, + 0.09493500739336014, + 0.17383266985416412, + 0.06921468675136566, + 0.3239755928516388, + 0.03585708886384964, + 0.05371476337313652, + 0.04053020104765892 + ], + "credit_map": [ + 0.009996588341891766, + 0.012048687785863876, + 0.008139548823237419, + 0.01021741796284914, + 0.00784319918602705, + 0.008967307396233082, + 0.02455940656363964, + 0.02071712352335453, + 0.005019800737500191, + 0.004922384861856699, + 0.006270990706980228, + 0.007084066979587078, + 0.006457323208451271, + 0.006456737406551838, + 0.010418199002742767, + 0.0076803476549685, + 0.0032200487330555916, + 0.003431354183703661, + 0.00376114365644753, + 0.005767597816884518, + 0.003869161242619157, + 0.003239498008042574, + 0.014346195384860039, + 0.013505827635526657, + 0.011241812258958817, + 0.013031664304435253, + 0.011278646066784859, + 0.015041833743453026, + 0.007208729162812233, + 0.00962533988058567, + 0.013635627925395966, + 0.013871352188289165, + 0.01599825546145439, + 0.03725649043917656, + 0.01942339725792408, + 0.025063684210181236, + 0.015907933935523033, + 0.020754769444465637, + 0.018598457798361778, + 0.020829686895012856, + 0.004001442808657885, + 0.005070406012237072, + 0.006450305227190256, + 0.006002826616168022, + 0.012995411641895771, + 0.011746601201593876, + 0.012580404058098793, + 0.010367289185523987, + 0.015108850784599781, + 0.020799677819013596, + 0.02373610809445381, + 0.10981634259223938, + 0.02287101000547409, + 0.07774010300636292, + 0.02531255967915058, + 0.02859092317521572, + 0.0031585851684212685, + 0.0050391689874231815, + 0.0031960776541382074, + 0.0034471936523914337, + 0.003499145619571209, + 0.006538595072925091, + 0.005815277807414532, + 0.0051630460657179356, + 0.004476562608033419, + 0.005633077118545771, + 0.005374130792915821, + 0.006588751915842295, + 0.005296032410115004, + 0.006029560696333647, + 0.008365343324840069, + 0.011951307766139507, + 0.006667593028396368, + 0.009116711094975471, + 0.006381076294928789, + 0.003124106675386429, + 0.0035446686670184135, + 0.003575085662305355, + 0.003785285633057356, + 0.004335671663284302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.024204934015870094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634805254638195 + ], + "opposite_logprob_tokens": [ + -3.651348114013672 + ], + "logprob_reward": -0.02634805254638195, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634805254638195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0668526142835617, + 0.04242831841111183, + 0.03621216118335724, + 0.0296328105032444, + 0.029881641268730164, + 0.05630641058087349, + 0.6208771467208862, + 0.02054627425968647, + 0.05945124104619026, + 0.03781135380268097 + ], + "credit_map": [ + 0.006663053762167692, + 0.00844806618988514, + 0.009419040754437447, + 0.008920996449887753, + 0.00721750408411026, + 0.00653044180944562, + 0.008997077122330666, + 0.01065643411129713, + 0.004052065312862396, + 0.0040871091187000275, + 0.004789705853909254, + 0.00408438965678215, + 0.004673351533710957, + 0.00436374731361866, + 0.008496916852891445, + 0.007881030440330505, + 0.0016141930827870965, + 0.0019228486344218254, + 0.0015070383669808507, + 0.0019091322319582105, + 0.003421681933104992, + 0.0018411458004266024, + 0.012478793039917946, + 0.011517329141497612, + 0.0031985489185899496, + 0.004570462740957737, + 0.0029773511923849583, + 0.00427641486749053, + 0.003477381309494376, + 0.0032280997838824987, + 0.0037493316922336817, + 0.004155221860855818, + 0.002531713340431452, + 0.003141021355986595, + 0.002241437789052725, + 0.003145972965285182, + 0.004256811924278736, + 0.004651757888495922, + 0.005555356852710247, + 0.004357570316642523, + 0.004664683248847723, + 0.005566085688769817, + 0.009953509084880352, + 0.007017846219241619, + 0.005275695584714413, + 0.006205413490533829, + 0.009387223981320858, + 0.008235951885581017, + 0.021664289757609367, + 0.03422284871339798, + 0.025578763335943222, + 0.22796127200126648, + 0.03674060106277466, + 0.1937841922044754, + 0.032053399831056595, + 0.04887174814939499, + 0.0016142163658514619, + 0.0020252144895493984, + 0.0016295306850224733, + 0.002875942038372159, + 0.002215577056631446, + 0.003318196628242731, + 0.0036012420896440744, + 0.0032663557212799788, + 0.00575143750756979, + 0.007752487901598215, + 0.007215368095785379, + 0.008410018868744373, + 0.006746228318661451, + 0.006607992574572563, + 0.009648330509662628, + 0.007319377735257149, + 0.006294235587120056, + 0.009225337766110897, + 0.0036327820271253586, + 0.003689373377710581, + 0.003850420704111457, + 0.003927797079086304, + 0.00329725444316864, + 0.0038941537495702505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.02634805254638195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801908880472183 + ], + "opposite_logprob_tokens": [ + -3.529801845550537 + ], + "logprob_reward": -0.029801908880472183, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801908880472183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15091519057750702, + 0.03444905951619148, + 0.1601932942867279, + 0.04877255856990814, + 0.038047853857278824, + 0.03177467733621597, + 0.25100013613700867, + 0.09107153862714767, + 0.10683848708868027, + 0.08693722635507584 + ], + "credit_map": [ + 0.01174693088978529, + 0.011150268837809563, + 0.015512984246015549, + 0.013379295356571674, + 0.01735488325357437, + 0.0252543855458498, + 0.02129371091723442, + 0.03522271662950516, + 0.0026539787650108337, + 0.003422398352995515, + 0.0034126865211874247, + 0.0047384751960635185, + 0.004208384547382593, + 0.0037964079529047012, + 0.0064933886751532555, + 0.005723334848880768, + 0.005301530938595533, + 0.003991547506302595, + 0.00524660712108016, + 0.004430142231285572, + 0.011045074090361595, + 0.007406489923596382, + 0.06299123913049698, + 0.05978064984083176, + 0.005216890014708042, + 0.004288176540285349, + 0.0034739598631858826, + 0.0038223930168896914, + 0.006776357069611549, + 0.008256161585450172, + 0.008884853683412075, + 0.008053764700889587, + 0.004256885964423418, + 0.004919508937746286, + 0.00253878952935338, + 0.003089817939326167, + 0.0038438139017671347, + 0.005763945169746876, + 0.0074261948466300964, + 0.006208892911672592, + 0.0026794825680553913, + 0.002774284454062581, + 0.00298531842418015, + 0.003083400661125779, + 0.005391555838286877, + 0.0055061508901417255, + 0.004705104045569897, + 0.004649377427995205, + 0.008078702725470066, + 0.014905384741723537, + 0.011800494976341724, + 0.08911480754613876, + 0.015643388032913208, + 0.07652146369218826, + 0.013694755733013153, + 0.021241111680865288, + 0.006800148636102676, + 0.009530915878713131, + 0.00881883129477501, + 0.009725539945065975, + 0.00761302188038826, + 0.013181962072849274, + 0.019491033628582954, + 0.01591007597744465, + 0.011152185499668121, + 0.01079857163131237, + 0.011979254893958569, + 0.017275096848607063, + 0.009176487103104591, + 0.010318140499293804, + 0.025189334526658058, + 0.010949396528303623, + 0.01487548928707838, + 0.01271979883313179, + 0.017670094966888428, + 0.007911830209195614, + 0.0072191269136965275, + 0.006832167040556669, + 0.006429618690162897, + 0.013279092498123646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.029801908880472183 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190363697707653 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190363697707653, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190363697707653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1862747073173523, + 0.06307128071784973, + 0.037661340087652206, + 0.06456328928470612, + 0.02727155201137066, + 0.4647449254989624, + 0.11144847422838211, + 0.013427629135549068, + 0.011404883116483688, + 0.020131850615143776 + ], + "credit_map": [ + 0.017236026003956795, + 0.01212911494076252, + 0.016664257273077965, + 0.015071570873260498, + 0.024359922856092453, + 0.03065839782357216, + 0.028632789850234985, + 0.04152264446020126, + 0.00462922640144825, + 0.004528204444795847, + 0.006090132519602776, + 0.0044042798690497875, + 0.005494499113410711, + 0.0076782796531915665, + 0.017456652596592903, + 0.012790012173354626, + 0.002772072795778513, + 0.0023431177251040936, + 0.00222386815585196, + 0.002708333544433117, + 0.006271467078477144, + 0.0052133603021502495, + 0.007904380559921265, + 0.008224744349718094, + 0.008667553775012493, + 0.003978204447776079, + 0.0036334057804197073, + 0.0032962504774332047, + 0.012080756947398186, + 0.012516620568931103, + 0.010818167589604855, + 0.009572339244186878, + 0.0019187531433999538, + 0.0034317157696932554, + 0.0015303920954465866, + 0.0021726153790950775, + 0.0031484838109463453, + 0.0047635179944336414, + 0.005804711952805519, + 0.004501365125179291, + 0.018195653334259987, + 0.03015359304845333, + 0.023983726277947426, + 0.03570491448044777, + 0.08567391335964203, + 0.08965344727039337, + 0.09775660187005997, + 0.08362308889627457, + 0.005743114743381739, + 0.006322409491986036, + 0.005859883036464453, + 0.03967639058828354, + 0.006857542786747217, + 0.030522149056196213, + 0.00721585052087903, + 0.009251159615814686, + 0.0014854820910841227, + 0.0013287562178447843, + 0.0012952374527230859, + 0.0015357560478150845, + 0.0013162620598450303, + 0.0019134035101160407, + 0.0021467474289238453, + 0.0024059864226728678, + 0.001108496799133718, + 0.0011764338705688715, + 0.0012749365996569395, + 0.0019391874084249139, + 0.0008777404436841607, + 0.0011289844987913966, + 0.0019559741485863924, + 0.0019431308610364795, + 0.0036551544908434153, + 0.002614077879115939, + 0.0035778966266661882, + 0.0019033049466088414, + 0.0020241765305399895, + 0.001742761698551476, + 0.001954039093106985, + 0.002660442842170596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.014190363697707653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633879289031029 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004633879289031029, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633879289031029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22662806510925293, + 0.09587439894676208, + 0.11510927230119705, + 0.05412892997264862, + 0.02083020843565464, + 0.07562655955553055, + 0.2828723192214966, + 0.03875938430428505, + 0.04694521054625511, + 0.04322557523846626 + ], + "credit_map": [ + 0.021268999204039574, + 0.019266285002231598, + 0.017473064363002777, + 0.017226768657565117, + 0.027308916673064232, + 0.03029944747686386, + 0.04189830273389816, + 0.05188627913594246, + 0.007822155021131039, + 0.006185702048242092, + 0.0065394253470003605, + 0.006309641059488058, + 0.011418713256716728, + 0.009368856437504292, + 0.024915793910622597, + 0.02331411838531494, + 0.0032605675514787436, + 0.0027633756399154663, + 0.003241488942876458, + 0.003022952703759074, + 0.008507569320499897, + 0.007157865911722183, + 0.04386777803301811, + 0.04328767582774162, + 0.006651958450675011, + 0.0065771969966590405, + 0.004667346831411123, + 0.007124537136405706, + 0.00572243332862854, + 0.006537888664752245, + 0.008801049552857876, + 0.008046520873904228, + 0.0025279412511736155, + 0.0024903512094169855, + 0.0019609169103205204, + 0.0015766993165016174, + 0.0026765631046146154, + 0.0037409376818686724, + 0.0027590433601289988, + 0.003097754903137684, + 0.005424268078058958, + 0.0072795129381120205, + 0.006650389172136784, + 0.00804403331130743, + 0.014895474538207054, + 0.012290551327168941, + 0.010387090966105461, + 0.010655242949724197, + 0.009047390893101692, + 0.0157811027020216, + 0.010966564528644085, + 0.10262513905763626, + 0.016664572060108185, + 0.09056071937084198, + 0.01448306255042553, + 0.02274376153945923, + 0.002879116917029023, + 0.00399541761726141, + 0.0031427659559994936, + 0.00448704743757844, + 0.0029582809656858444, + 0.006898706313222647, + 0.005693372339010239, + 0.00870467722415924, + 0.005024462938308716, + 0.0033332034945487976, + 0.004481736570596695, + 0.008478927426040173, + 0.0028917447198182344, + 0.004040988627821207, + 0.012334899976849556, + 0.006359242368489504, + 0.006125458981841803, + 0.00856236182153225, + 0.005339379888027906, + 0.004303574562072754, + 0.003759463084861636, + 0.0035574459470808506, + 0.005088855512440205, + 0.006489035673439503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.004633879289031029 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, sideways tracking shot closely follows a mythical white tiger as it gallops through an ancient, dense forest at night. Cold moonlight filters through the gaps between the tall, twisted canopies of ancient trees, casting a variegated, dynamic play of light and shadow across a ground covered in dead leaves and damp earth. The beast's snow-white, unblemished fur shimmers with a soft silver halo in the moonlight. Its muscular form ripples clearly beneath its fur as it gallops, and its icy blue eyes gleam with an otherworldly light in the darkness. Its limbs fully extended, it gallops with explosive force, its massive paws barely leaving a trace as it stamped across the ground. The trees in the background are reduced to blurred vertical lines by its breathtaking speed. As it approaches a massive, moss-covered fallen log that lies across the path, it doesn't slow at all. Its hind legs thrust forward, its entire body tracing an elegant, powerful white arc through the air. It effortlessly clears the obstacle, landing firmly on the other side and continuing its charge. Its breath becomes a visible white mist in the chill air. The image is dominated by deep midnight blue and inky black, with the white tiger itself as the most striking presence, creating an atmosphere imbued with a sense of primal power, mysterious elegance, and the beauty of speed.", + "seed": 8018, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_6/index_3107_seed_8018_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12833011392503976, + "vlm_logprob": -0.12833011392503976, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.12833011392503976, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11605846881866455 + ], + "opposite_logprob_tokens": [ + -2.241058349609375 + ], + "logprob_reward": -0.11605846881866455, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11605846881866455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10016457736492157, + 0.02297787368297577, + 0.021212950348854065, + 0.03707415238022804, + 0.06325437873601913, + 0.5321345329284668, + 0.0338880680501461, + 0.09691885858774185, + 0.07382697612047195, + 0.018547654151916504 + ], + "credit_map": [ + 0.009998339228332043, + 0.00950043648481369, + 0.010411391966044903, + 0.017634596675634384, + 0.009004153311252594, + 0.016543937847018242, + 0.013371769338846207, + 0.013699952512979507, + 0.0014135047094896436, + 0.0015226206742227077, + 0.0016193340998142958, + 0.002057380275800824, + 0.0015952050453051925, + 0.0018142486223950982, + 0.006489963736385107, + 0.006465616635978222, + 0.0014196743723005056, + 0.00225727166980505, + 0.0020861439406871796, + 0.002848968608304858, + 0.0025376733392477036, + 0.003593356115743518, + 0.0033190969843417406, + 0.003150763688609004, + 0.0037033935077488422, + 0.003761430038139224, + 0.0033409297466278076, + 0.005438020918518305, + 0.004199857357889414, + 0.008490081876516342, + 0.002990440232679248, + 0.005150000564754009, + 0.004389850422739983, + 0.0026648961938917637, + 0.002684251870959997, + 0.003637191141024232, + 0.002694695256650448, + 0.023392602801322937, + 0.002860908629372716, + 0.020929982885718346, + 0.032865509390830994, + 0.02771979197859764, + 0.01678982563316822, + 0.019601445645093918, + 0.05257565528154373, + 0.19011825323104858, + 0.03301118686795235, + 0.15945285558700562, + 0.004965336062014103, + 0.0030162911862134933, + 0.005395009182393551, + 0.004604077432304621, + 0.002860395470634103, + 0.0036341084633022547, + 0.004453087225556374, + 0.0049597639590501785, + 0.015408293344080448, + 0.009887078776955605, + 0.012963593937456608, + 0.009881597943603992, + 0.009492075070738792, + 0.015358764678239822, + 0.012330456636846066, + 0.01159699261188507, + 0.0107646519318223, + 0.006703685037791729, + 0.00856629479676485, + 0.008864612318575382, + 0.007034836336970329, + 0.010349154472351074, + 0.008745620027184486, + 0.012798118405044079, + 0.003175031393766403, + 0.0021157977171242237, + 0.0020308138336986303, + 0.0018334059277549386, + 0.001925416523590684, + 0.0033875431399792433, + 0.002018310595303774, + 0.0020613358356058598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.11605846881866455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036495566368103 + ], + "opposite_logprob_tokens": [ + -1.9103649854660034 + ], + "logprob_reward": -0.16036495566368103, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036495566368103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1869657337665558, + 0.062390513718128204, + 0.03785987198352814, + 0.09135054051876068, + 0.1044008731842041, + 0.18006929755210876, + 0.1136576235294342, + 0.05357666313648224, + 0.13250784575939178, + 0.0372210294008255 + ], + "credit_map": [ + 0.015778392553329468, + 0.017048489302396774, + 0.017522048205137253, + 0.03084653615951538, + 0.01875065267086029, + 0.03598373755812645, + 0.024205921217799187, + 0.026829950511455536, + 0.005031416192650795, + 0.004388193134218454, + 0.005839120596647263, + 0.005749683827161789, + 0.00600332161411643, + 0.006452546920627356, + 0.015280583873391151, + 0.01364564523100853, + 0.003944545052945614, + 0.005023572128266096, + 0.004508308134973049, + 0.004641861654818058, + 0.004687945824116468, + 0.004527514800429344, + 0.005201622843742371, + 0.005324501544237137, + 0.008579879067838192, + 0.011612537316977978, + 0.008516985923051834, + 0.012360637076199055, + 0.011254347860813141, + 0.018736811354756355, + 0.00825957115739584, + 0.012029774487018585, + 0.011626273393630981, + 0.010631836950778961, + 0.007131719496101141, + 0.0154147082939744, + 0.00942301843315363, + 0.01695173978805542, + 0.01620599441230297, + 0.017015593126416206, + 0.010106867179274559, + 0.00716922152787447, + 0.007093202322721481, + 0.009534848853945732, + 0.013075885362923145, + 0.06762666255235672, + 0.00971947330981493, + 0.05574312433600426, + 0.017525112256407738, + 0.014185718260705471, + 0.0194235909730196, + 0.02040163055062294, + 0.00949848536401987, + 0.014035883359611034, + 0.010006430558860302, + 0.008580775931477547, + 0.009032382629811764, + 0.004071350209414959, + 0.006809703074395657, + 0.006225044839084148, + 0.006313582416623831, + 0.00719822384417057, + 0.0069803898222744465, + 0.0069459895603358746, + 0.019105233252048492, + 0.01267960760742426, + 0.017251573503017426, + 0.019847849383950233, + 0.012175369076430798, + 0.015174693427979946, + 0.01916787400841713, + 0.01710563711822033, + 0.007072666194289923, + 0.0031352946534752846, + 0.0031634129118174314, + 0.0035044692922383547, + 0.0034116872120648623, + 0.009888393804430962, + 0.003600278403609991, + 0.003444826463237405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.16036495566368103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22563792765140533 + ], + "opposite_logprob_tokens": [ + -1.6006379127502441 + ], + "logprob_reward": -0.22563792765140533, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22563792765140533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22957564890384674, + 0.02740435115993023, + 0.040584906935691833, + 0.040314797312021255, + 0.09466394037008286, + 0.17800728976726532, + 0.08215519785881042, + 0.047383230179548264, + 0.22141309082508087, + 0.03849757835268974 + ], + "credit_map": [ + 0.019321277737617493, + 0.020496584475040436, + 0.022409122437238693, + 0.03718318045139313, + 0.02285289205610752, + 0.043819770216941833, + 0.03136089816689491, + 0.032131925225257874, + 0.0017030412564054132, + 0.001595750916749239, + 0.0018488720525056124, + 0.0018629356054589152, + 0.0019146408885717392, + 0.0020715321879833937, + 0.008391362614929676, + 0.008016216568648815, + 0.0035913267638534307, + 0.004500722512602806, + 0.0036074344534426928, + 0.005157171748578548, + 0.005091347731649876, + 0.005874038208276033, + 0.006451509892940521, + 0.006311352364718914, + 0.0043618204072117805, + 0.004632631316781044, + 0.0031058769673109055, + 0.004011029377579689, + 0.005010392516851425, + 0.01008425373584032, + 0.0033882101997733116, + 0.005720582325011492, + 0.005741436034440994, + 0.00489339604973793, + 0.004043654538691044, + 0.005561566911637783, + 0.004025625064969063, + 0.0326019749045372, + 0.006217175628989935, + 0.03157909959554672, + 0.010422318242490292, + 0.007055036723613739, + 0.005602091550827026, + 0.007097678724676371, + 0.010336865670979023, + 0.069861501455307, + 0.010248430073261261, + 0.05738336592912674, + 0.011954386718571186, + 0.0068130409345030785, + 0.009352573193609715, + 0.012723254039883614, + 0.00597446970641613, + 0.011724494397640228, + 0.008320522494614124, + 0.015292453579604626, + 0.007361650932580233, + 0.004645358771085739, + 0.0054311007261276245, + 0.006024578586220741, + 0.006232349202036858, + 0.0054072607308626175, + 0.005249009933322668, + 0.007031922694295645, + 0.030893642455339432, + 0.02299121767282486, + 0.02733035571873188, + 0.028198881074786186, + 0.022845296189188957, + 0.026643948629498482, + 0.03034384548664093, + 0.03216592222452164, + 0.00799177773296833, + 0.0029824611265212297, + 0.003485151333734393, + 0.003280501812696457, + 0.002842066576704383, + 0.011362629942595959, + 0.00324069382622838, + 0.0033122943714261055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.22563792765140533 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705445289611816 + ], + "opposite_logprob_tokens": [ + -2.127054452896118 + ], + "logprob_reward": -0.12705445289611816, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705445289611816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3390777111053467, + 0.039089977741241455, + 0.08881645649671555, + 0.15854936838150024, + 0.08090320974588394, + 0.11839238554239273, + 0.02432829700410366, + 0.03673761337995529, + 0.08418723195791245, + 0.029917791485786438 + ], + "credit_map": [ + 0.021146640181541443, + 0.02963302470743656, + 0.033444516360759735, + 0.051290374249219894, + 0.03756878897547722, + 0.07150235027074814, + 0.04635819047689438, + 0.04813383147120476, + 0.0018714370671659708, + 0.0024658041074872017, + 0.0025262152776122093, + 0.0032492803875356913, + 0.006360325496643782, + 0.0064185489900410175, + 0.009470196440815926, + 0.006728166714310646, + 0.0072324201464653015, + 0.0054901461116969585, + 0.004905857145786285, + 0.006721870042383671, + 0.008740861900150776, + 0.0135875903069973, + 0.02076166495680809, + 0.02137604169547558, + 0.018582765012979507, + 0.016735022887587547, + 0.01001136377453804, + 0.011736921034753323, + 0.015849336981773376, + 0.04777545481920242, + 0.012099077925086021, + 0.02575943060219288, + 0.004477988462895155, + 0.004150950815528631, + 0.0023677691351622343, + 0.004065369721502066, + 0.004565708339214325, + 0.027820168063044548, + 0.004021412692964077, + 0.0294338446110487, + 0.00797832291573286, + 0.007346396800130606, + 0.004311813972890377, + 0.00579939316958189, + 0.014781486243009567, + 0.03836379572749138, + 0.008619861677289009, + 0.03119131363928318, + 0.00377518474124372, + 0.0014933086931705475, + 0.0045427437871694565, + 0.0030783636029809713, + 0.0015267470153048635, + 0.0023349232506006956, + 0.002465132623910904, + 0.005111893638968468, + 0.004881556611508131, + 0.0028709012549370527, + 0.00431779632344842, + 0.00474352715536952, + 0.0047010076232254505, + 0.005009044893085957, + 0.004875769838690758, + 0.005338012240827084, + 0.011185932904481888, + 0.007669859565794468, + 0.012006040662527084, + 0.013151466846466064, + 0.00822423305362463, + 0.008142837323248386, + 0.012796778231859207, + 0.011010080575942993, + 0.004924173466861248, + 0.0030425027944147587, + 0.0027616124134510756, + 0.0024896753020584583, + 0.002648642985150218, + 0.007350726518779993, + 0.003004152560606599, + 0.003696305677294731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.12705445289611816 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534764595329762 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534764595329762, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534764595329762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28256112337112427, + 0.10512281209230423, + 0.07190624624490738, + 0.08729445189237595, + 0.10527913272380829, + 0.09550053626298904, + 0.13913178443908691, + 0.03483001887798309, + 0.030255185440182686, + 0.04811875522136688 + ], + "credit_map": [ + 0.03282824531197548, + 0.021087827160954475, + 0.02040725015103817, + 0.035446684807538986, + 0.03127012401819229, + 0.04520802944898605, + 0.04459128901362419, + 0.05172165483236313, + 0.005449909251183271, + 0.005034333560615778, + 0.007819819264113903, + 0.009019932709634304, + 0.011786891147494316, + 0.017773978412151337, + 0.025202317163348198, + 0.023035624995827675, + 0.005176757927983999, + 0.006993087008595467, + 0.006640317849814892, + 0.008871628902852535, + 0.006081896834075451, + 0.0072274841368198395, + 0.014642521739006042, + 0.01627255044877529, + 0.008110140450298786, + 0.00953684002161026, + 0.0059190783649683, + 0.008170565590262413, + 0.010177850723266602, + 0.02467903308570385, + 0.006957030855119228, + 0.013743913732469082, + 0.0038502016104757786, + 0.006573621183633804, + 0.004198839887976646, + 0.008580107241868973, + 0.005476863123476505, + 0.034361112862825394, + 0.009455908089876175, + 0.03278246894478798, + 0.005876266397535801, + 0.00836885068565607, + 0.006839838344603777, + 0.01526723988354206, + 0.008516484871506691, + 0.02100382186472416, + 0.011058755218982697, + 0.018569281324744225, + 0.02163734845817089, + 0.011542669497430325, + 0.017938846722245216, + 0.019505903124809265, + 0.014704842120409012, + 0.016948416829109192, + 0.020746640861034393, + 0.016107112169265747, + 0.004209304694086313, + 0.002654952695593238, + 0.005340749863535166, + 0.00437242491170764, + 0.004696441348642111, + 0.003522365354001522, + 0.005835948511958122, + 0.004197829402983189, + 0.003989688120782375, + 0.0026209568604826927, + 0.0026843135710805655, + 0.003928950522094965, + 0.005036457441747189, + 0.003477799240499735, + 0.003734587924554944, + 0.004782430361956358, + 0.006716880481690168, + 0.0035211495123803616, + 0.005498633254319429, + 0.005568202119320631, + 0.004938863217830658, + 0.00883895717561245, + 0.006084000691771507, + 0.006952065974473953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.012534764595329762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, sideways tracking shot closely follows a mythical white tiger as it gallops through an ancient, dense forest at night. Cold moonlight filters through the gaps between the tall, twisted canopies of ancient trees, casting a variegated, dynamic play of light and shadow across a ground covered in dead leaves and damp earth. The beast's snow-white, unblemished fur shimmers with a soft silver halo in the moonlight. Its muscular form ripples clearly beneath its fur as it gallops, and its icy blue eyes gleam with an otherworldly light in the darkness. Its limbs fully extended, it gallops with explosive force, its massive paws barely leaving a trace as it stamped across the ground. The trees in the background are reduced to blurred vertical lines by its breathtaking speed. As it approaches a massive, moss-covered fallen log that lies across the path, it doesn't slow at all. Its hind legs thrust forward, its entire body tracing an elegant, powerful white arc through the air. It effortlessly clears the obstacle, landing firmly on the other side and continuing its charge. Its breath becomes a visible white mist in the chill air. The image is dominated by deep midnight blue and inky black, with the white tiger itself as the most striking presence, creating an atmosphere imbued with a sense of primal power, mysterious elegance, and the beauty of speed.", + "seed": 8019, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_6/index_3107_seed_8019_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.035014659725129606, + "vlm_logprob": -0.035014659725129606, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.035014659725129606, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04767858237028122 + ], + "opposite_logprob_tokens": [ + -3.1726784706115723 + ], + "logprob_reward": -0.04767858237028122, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04767858237028122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16742880642414093, + 0.02855263464152813, + 0.03221341222524643, + 0.11536985635757446, + 0.18070930242538452, + 0.03756061941385269, + 0.14163018763065338, + 0.03129100054502487, + 0.08621499687433243, + 0.17902922630310059 + ], + "credit_map": [ + 0.021296892315149307, + 0.014246241189539433, + 0.030566608533263206, + 0.017092319205403328, + 0.01678459160029888, + 0.02930659055709839, + 0.014402104541659355, + 0.023733466863632202, + 0.002896983874961734, + 0.0035194028168916702, + 0.003880141768604517, + 0.0036951692309230566, + 0.0033617187291383743, + 0.003927113488316536, + 0.003164971712976694, + 0.004107133485376835, + 0.003216231008991599, + 0.0037655874621123075, + 0.0035004534292966127, + 0.0026359311304986477, + 0.005222122184932232, + 0.00628175213932991, + 0.003753481898456812, + 0.003837850410491228, + 0.010196587070822716, + 0.012927216477692127, + 0.0124712735414505, + 0.010784424841403961, + 0.01350864116102457, + 0.017844628542661667, + 0.01475986372679472, + 0.022877216339111328, + 0.03389531746506691, + 0.01796126924455166, + 0.018029794096946716, + 0.012250159867107868, + 0.029101142659783363, + 0.02908918261528015, + 0.01784443110227585, + 0.02253800258040428, + 0.005608290899544954, + 0.003342399373650551, + 0.006558552850037813, + 0.004491351544857025, + 0.0036895701196044683, + 0.004193772561848164, + 0.004113471135497093, + 0.005563211161643267, + 0.02439737319946289, + 0.010230179876089096, + 0.022130101919174194, + 0.022598110139369965, + 0.016705872491002083, + 0.013849402777850628, + 0.013938250951468945, + 0.01778089627623558, + 0.0034527676180005074, + 0.007002420257776976, + 0.0025585226248949766, + 0.003526757238432765, + 0.002668829169124365, + 0.0018475073156878352, + 0.0030705323442816734, + 0.0071636647917330265, + 0.008198016323149204, + 0.009093698114156723, + 0.00885969027876854, + 0.012282580137252808, + 0.010449035093188286, + 0.010148745961487293, + 0.013282367959618568, + 0.013900866732001305, + 0.01175046805292368, + 0.011059380136430264, + 0.021287428215146065, + 0.01709502935409546, + 0.007728530094027519, + 0.04822353273630142, + 0.01769312098622322, + 0.04419173672795296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.04767858237028122 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062043242156505585 + ], + "opposite_logprob_tokens": [ + -2.8120431900024414 + ], + "logprob_reward": -0.062043242156505585, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062043242156505585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.261118084192276, + 0.04647914692759514, + 0.0945344939827919, + 0.06364090740680695, + 0.1775101125240326, + 0.09331725537776947, + 0.051729921251535416, + 0.0479666069149971, + 0.06678049266338348, + 0.09692297875881195 + ], + "credit_map": [ + 0.040779635310173035, + 0.01927269995212555, + 0.041714128106832504, + 0.020490670576691628, + 0.02449028007686138, + 0.05570805072784424, + 0.02100321277976036, + 0.03765939176082611, + 0.005416056606918573, + 0.0049179536290466785, + 0.006452612578868866, + 0.006990670692175627, + 0.004380505532026291, + 0.006817433051764965, + 0.005951068364083767, + 0.005552846938371658, + 0.008477162569761276, + 0.014901354908943176, + 0.010715107433497906, + 0.00894217099994421, + 0.018114855512976646, + 0.012120716273784637, + 0.008627388626337051, + 0.012635743245482445, + 0.005602184217423201, + 0.0071718161925673485, + 0.007044619414955378, + 0.005355264060199261, + 0.007273778319358826, + 0.010070225223898888, + 0.006845225114375353, + 0.014277792535722256, + 0.04196561872959137, + 0.012666203081607819, + 0.01599009893834591, + 0.012985640205442905, + 0.013326265849173069, + 0.037763386964797974, + 0.016020096838474274, + 0.02679281309247017, + 0.012885255739092827, + 0.011659090407192707, + 0.015683356672525406, + 0.008397313766181469, + 0.010410115122795105, + 0.012167412787675858, + 0.007983185350894928, + 0.014131525531411171, + 0.008467430248856544, + 0.004074995871633291, + 0.006304679438471794, + 0.008292817510664463, + 0.006095650140196085, + 0.005713402759283781, + 0.006239632144570351, + 0.006541315000504255, + 0.004972830414772034, + 0.00957948062568903, + 0.0040581864304840565, + 0.004231607541441917, + 0.004330494906753302, + 0.003525237087160349, + 0.005611545871943235, + 0.011657225899398327, + 0.0069359103217720985, + 0.006689789239317179, + 0.007826303131878376, + 0.008337709121406078, + 0.010436725802719593, + 0.008505914360284805, + 0.009963804855942726, + 0.008084336295723915, + 0.006172183435410261, + 0.0043779234401881695, + 0.010523754172027111, + 0.007397610694169998, + 0.004778015892952681, + 0.024118531495332718, + 0.010600452311336994, + 0.028954507783055305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.062043242156505585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370628505945206 + ], + "opposite_logprob_tokens": [ + -3.4087061882019043 + ], + "logprob_reward": -0.03370628505945206, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370628505945206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11003278195858002, + 0.04664847254753113, + 0.031244315207004547, + 0.14299526810646057, + 0.09452059864997864, + 0.031435124576091766, + 0.023670850321650505, + 0.01654505915939808, + 0.05971144512295723, + 0.4431960880756378 + ], + "credit_map": [ + 0.017234213650226593, + 0.00888091791421175, + 0.015870317816734314, + 0.009707512333989143, + 0.011802034452557564, + 0.02027098275721073, + 0.010918369516730309, + 0.015348434448242188, + 0.0033129809889942408, + 0.002169213956221938, + 0.003743838518857956, + 0.002208127873018384, + 0.0024902101140469313, + 0.01478141825646162, + 0.004297288134694099, + 0.013645391911268234, + 0.004077210556715727, + 0.002409732434898615, + 0.0033718289341777563, + 0.002292363904416561, + 0.0045814975164830685, + 0.007716653868556023, + 0.0023536013904958963, + 0.004441427532583475, + 0.012683187611401081, + 0.01421083603054285, + 0.0155220041051507, + 0.013648906722664833, + 0.015328514389693737, + 0.024646421894431114, + 0.019592909142374992, + 0.027362480759620667, + 0.018038587644696236, + 0.007519320119172335, + 0.009470525197684765, + 0.008859475143253803, + 0.00971189234405756, + 0.017536906525492668, + 0.011562546715140343, + 0.01182134635746479, + 0.004019292537122965, + 0.005520302336663008, + 0.0032995487563312054, + 0.004327138885855675, + 0.0032655061222612858, + 0.0034176327753812075, + 0.004013645928353071, + 0.003572057466953993, + 0.004337047692388296, + 0.0018249828135594726, + 0.002978464588522911, + 0.0040489924140274525, + 0.0026986096054315567, + 0.002379483310505748, + 0.0023350147530436516, + 0.0030682552605867386, + 0.0023894866462796926, + 0.0018202739302068949, + 0.002024964662268758, + 0.0013589368900284171, + 0.00193271622993052, + 0.0028782961890101433, + 0.002056635683402419, + 0.002083748346194625, + 0.005280076991766691, + 0.0061815218068659306, + 0.007532128598541021, + 0.006610615644603968, + 0.006008128169924021, + 0.006828167475759983, + 0.01014796830713749, + 0.011122833006083965, + 0.02415051870048046, + 0.013263529166579247, + 0.042749207466840744, + 0.022540319710969925, + 0.016604477539658546, + 0.1344064325094223, + 0.0461118221282959, + 0.14336979389190674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.03370628505945206 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577285438776016 + ], + "opposite_logprob_tokens": [ + -3.8955771923065186 + ], + "logprob_reward": -0.020577285438776016, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577285438776016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24418286979198456, + 0.022910041734576225, + 0.06951368600130081, + 0.3199940621852875, + 0.08399595320224762, + 0.03579353541135788, + 0.024486301466822624, + 0.03389934450387955, + 0.05830630660057068, + 0.10691794753074646 + ], + "credit_map": [ + 0.03285461291670799, + 0.01619960367679596, + 0.0520147979259491, + 0.023460783064365387, + 0.02291068620979786, + 0.04077661409974098, + 0.02277248725295067, + 0.03319326415657997, + 0.002331184223294258, + 0.0020179166458547115, + 0.0034528065007179976, + 0.002599815372377634, + 0.0022504329681396484, + 0.003648934420198202, + 0.0031688143499195576, + 0.003440136555582285, + 0.010117859579622746, + 0.006453663110733032, + 0.0067893825471401215, + 0.00430477736517787, + 0.007276900578290224, + 0.018548661842942238, + 0.005517665762454271, + 0.010504772886633873, + 0.028326701372861862, + 0.03014218620955944, + 0.03410110995173454, + 0.02749023772776127, + 0.03872229903936386, + 0.05802933871746063, + 0.037980470806360245, + 0.06520173698663712, + 0.015070881694555283, + 0.00726989796385169, + 0.00754252914339304, + 0.00633436581119895, + 0.011431879363954067, + 0.016315069049596786, + 0.009832896292209625, + 0.010198429226875305, + 0.0055978004820644855, + 0.002759771654382348, + 0.0034810651559382677, + 0.0036386989522725344, + 0.004067154135555029, + 0.005602496210485697, + 0.0036697741597890854, + 0.006976773496717215, + 0.0033005124423652887, + 0.0020057254005223513, + 0.005881504155695438, + 0.0026030559092760086, + 0.0027601507026702166, + 0.001992702018469572, + 0.002768014557659626, + 0.0031746369786560535, + 0.006506605073809624, + 0.003515078453347087, + 0.005417768377810717, + 0.003457177197560668, + 0.002877692924812436, + 0.004702401347458363, + 0.0033082026056945324, + 0.004114419687539339, + 0.004856481682509184, + 0.007544421125203371, + 0.0074930922128260136, + 0.006135110277682543, + 0.00532505800947547, + 0.008799683302640915, + 0.008780790492892265, + 0.00937167089432478, + 0.006613581907004118, + 0.0042472705245018005, + 0.011734113097190857, + 0.010627775453031063, + 0.0035117845982313156, + 0.029342591762542725, + 0.01093985140323639, + 0.029900971800088882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.020577285438776016 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical white tiger with snow-white, unblemished fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger galloping with explosive force, its limbs fully extended and hind legs thrusting forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger clear a massive, moss-covered fallen log without slowing down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient, dense forest at night with tall, twisted canopies and a ground covered in dead leaves and damp earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, sideways tracking shot where the background trees appear as blurred vertical lines due to speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067903600633144 + ], + "opposite_logprob_tokens": [ + -4.511067867279053 + ], + "logprob_reward": -0.011067903600633144, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067903600633144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38724246621131897, + 0.041014671325683594, + 0.04296157509088516, + 0.08690402656793594, + 0.09022832661867142, + 0.03330270200967789, + 0.02345450408756733, + 0.0393047071993351, + 0.04439631104469299, + 0.21119073033332825 + ], + "credit_map": [ + 0.058746177703142166, + 0.03357157111167908, + 0.06959418207406998, + 0.03754149377346039, + 0.032446712255477905, + 0.06382457911968231, + 0.03796536102890968, + 0.05355238914489746, + 0.0023781496565788984, + 0.003827566746622324, + 0.0034642687533050776, + 0.0032164573203772306, + 0.004526987671852112, + 0.01031885389238596, + 0.0038500071968883276, + 0.00943237915635109, + 0.00546212587505579, + 0.004874315578490496, + 0.004873168189078569, + 0.004270101431757212, + 0.006494532804936171, + 0.0077217938378453255, + 0.0038189291954040527, + 0.005446608178317547, + 0.009589610621333122, + 0.008514504879713058, + 0.010409376583993435, + 0.008038006722927094, + 0.010069184005260468, + 0.014372962526977062, + 0.010964436456561089, + 0.014945951290428638, + 0.010579786263406277, + 0.010752426460385323, + 0.00875756423920393, + 0.012039484456181526, + 0.013128690421581268, + 0.014147326350212097, + 0.01002509891986847, + 0.010797944851219654, + 0.004562421701848507, + 0.004690550733357668, + 0.0035695137921720743, + 0.003469443181529641, + 0.0048282565549016, + 0.0037543117068707943, + 0.0030961232259869576, + 0.005332078784704208, + 0.004030331503599882, + 0.0022112291771918535, + 0.0034608449786901474, + 0.002768420847132802, + 0.002417666371911764, + 0.0023820013739168644, + 0.00328341918066144, + 0.0029005908872932196, + 0.006952844560146332, + 0.005300670862197876, + 0.007056070026010275, + 0.003961298149079084, + 0.0029389068949967623, + 0.004507279023528099, + 0.00343867065384984, + 0.005148967728018761, + 0.0032180887646973133, + 0.003973385784775019, + 0.0034413097891956568, + 0.004897568374872208, + 0.0040131621062755585, + 0.004690955393016338, + 0.00929483026266098, + 0.010867007076740265, + 0.013611136935651302, + 0.008789120241999626, + 0.022250428795814514, + 0.015150473453104496, + 0.007716563530266285, + 0.06159444898366928, + 0.02242358773946762, + 0.05965495482087135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.011067903600633144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_8/index_tensor([7251, 7251])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_8/index_tensor([7251, 7251])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..797ca231d74e8998ac527a1242f1d8fcd148cd42 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_8/index_tensor([7251, 7251])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic close-up, static and precisely framed with a very shallow depth of field, captures a matte black metal candle snuff bottle standing upright on the rich, darkly textured surface of an old wooden table. High-contrast lighting defines the scene: a single, intense light source carves the snuff bottle's elegant form from the side, casting long, sharp shadows and creating stark highlights along one side of its slender handle and the curve of its bell-shaped head. The background is a complete, soft-focus blur, suggesting a dim, intimate space. The initial tranquility is shattered as a subtle vibration travels through the wooden surface, causing the snuff bottle to teeter for a brief, tense moment. It sways unsteadily, the handle wobbling slightly, before it decisively loses its center of gravity. The object then tilts, its initial descent slowly followed by a rapid acceleration as it spins around the edge of its base, finally colliding with the wood with a soft, gentle clink and coming to rest against its side. Its dark, elegant form now becomes a horizontal element in the composition, creating an atmosphere of quiet, contemplative, and visually appealing beauty.", + "seed": 21849, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_8/index_7251_seed_21849_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010458138398826122, + "vlm_logprob": -0.010458138398826122, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.010458138398826122, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03885507211089134 + ], + "opposite_logprob_tokens": [ + -3.6638550758361816 + ], + "logprob_reward": -0.03885507211089134, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03885507211089134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06569163501262665, + 0.04418295621871948, + 0.0316566526889801, + 0.028495611622929573, + 0.018008509650826454, + 0.02203560248017311, + 0.5061247944831848, + 0.032787665724754333, + 0.05650314688682556, + 0.1945134550333023 + ], + "credit_map": [ + 0.0071179927326738834, + 0.008474396541714668, + 0.010831288062036037, + 0.007582603953778744, + 0.008831076323986053, + 0.008121299557387829, + 0.007802169304341078, + 0.006930810399353504, + 0.004897433798760176, + 0.004378031007945538, + 0.007947571575641632, + 0.0042441776022315025, + 0.006865344941616058, + 0.005733416881412268, + 0.0053481715731322765, + 0.004768811631947756, + 0.002978062490001321, + 0.002909631934016943, + 0.005733830388635397, + 0.003764539957046509, + 0.005650576204061508, + 0.004019401967525482, + 0.0037985537201166153, + 0.0028020560275763273, + 0.0023913562763482332, + 0.003706848481670022, + 0.005222673062235117, + 0.0031169846188277006, + 0.003628065809607506, + 0.004610520787537098, + 0.0029847959522157907, + 0.002834367100149393, + 0.0016107799019664526, + 0.0022268989123404026, + 0.003055734094232321, + 0.0016906381351873279, + 0.0021856294479221106, + 0.0020322061609476805, + 0.002324227709323168, + 0.0028823951724916697, + 0.003148496150970459, + 0.003992573823779821, + 0.003958882763981819, + 0.002308697672560811, + 0.002129025524482131, + 0.002697335323318839, + 0.001701363711617887, + 0.002099227160215378, + 0.16407078504562378, + 0.01363812480121851, + 0.21353410184383392, + 0.023450836539268494, + 0.013783282600343227, + 0.03257991001009941, + 0.0248123686760664, + 0.020255358889698982, + 0.0025688703171908855, + 0.004821321927011013, + 0.0038803592324256897, + 0.00460903812199831, + 0.003655528649687767, + 0.004351464100182056, + 0.003933538682758808, + 0.004967544227838516, + 0.004367426969110966, + 0.009582964703440666, + 0.00647358363494277, + 0.008493266999721527, + 0.00533432699739933, + 0.008783918805420399, + 0.008016643114387989, + 0.0054510124027729034, + 0.028602782636880875, + 0.01549754198640585, + 0.025062493979930878, + 0.024737300351262093, + 0.011604712344706059, + 0.05143707990646362, + 0.01631242036819458, + 0.021259116008877754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.03885507211089134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011991222854703665 + ], + "opposite_logprob_tokens": [ + -6.751199245452881 + ], + "logprob_reward": -0.0011991222854703665, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011991222854703665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12387185543775558, + 0.055031079798936844, + 0.022713707759976387, + 0.02863873355090618, + 0.05999507009983063, + 0.07540199905633926, + 0.1787371188402176, + 0.06169943884015083, + 0.17047180235385895, + 0.22343918681144714 + ], + "credit_map": [ + 0.01594352163374424, + 0.020402126014232635, + 0.019035177305340767, + 0.010616620071232319, + 0.016963228583335876, + 0.012837310321629047, + 0.016748275607824326, + 0.011325591243803501, + 0.00878114439547062, + 0.008934461511671543, + 0.008000153116881847, + 0.0072402809746563435, + 0.005639564245939255, + 0.007025860249996185, + 0.003462650813162327, + 0.005946968216449022, + 0.0023619269486516714, + 0.00213405629619956, + 0.004110174719244242, + 0.002641374245285988, + 0.0024135441053658724, + 0.003599231829866767, + 0.0022574448958039284, + 0.003195954253897071, + 0.002409276319667697, + 0.0037176040932536125, + 0.006626418326050043, + 0.002976557007059455, + 0.0037123155780136585, + 0.004305348731577396, + 0.00236858194693923, + 0.0025226299185305834, + 0.005388831719756126, + 0.008329392410814762, + 0.00677314680069685, + 0.0073254723101854324, + 0.011698714457452297, + 0.007573902606964111, + 0.005688113626092672, + 0.007217496633529663, + 0.008198409341275692, + 0.010267852805554867, + 0.015117442235350609, + 0.009565574117004871, + 0.00860180240124464, + 0.01070625800639391, + 0.005905137397348881, + 0.00703952182084322, + 0.05986517295241356, + 0.005548128392547369, + 0.07176628708839417, + 0.007592751644551754, + 0.005736036226153374, + 0.011860544793307781, + 0.009754683822393417, + 0.006613514386117458, + 0.006267220247536898, + 0.010312668047845364, + 0.008102728985249996, + 0.007144202012568712, + 0.006227221339941025, + 0.010224765166640282, + 0.008428392000496387, + 0.004992238245904446, + 0.017816781997680664, + 0.026842549443244934, + 0.024983905255794525, + 0.02427915669977665, + 0.01663188263773918, + 0.028907395899295807, + 0.01725430227816105, + 0.013755816034972668, + 0.05372888594865799, + 0.011488178744912148, + 0.02372623234987259, + 0.014690722338855267, + 0.009078255854547024, + 0.07764091342687607, + 0.008767697960138321, + 0.024318281561136246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.0011991222854703665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052668387070298195 + ], + "opposite_logprob_tokens": [ + -5.2552666664123535 + ], + "logprob_reward": -0.0052668387070298195, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052668387070298195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20824867486953735, + 0.04587944969534874, + 0.02699938602745533, + 0.011328080669045448, + 0.07964833080768585, + 0.02046973630785942, + 0.46197110414505005, + 0.03413265198469162, + 0.04335619881749153, + 0.06796632707118988 + ], + "credit_map": [ + 0.03883199393749237, + 0.04484279453754425, + 0.04077661409974098, + 0.01404198445379734, + 0.016533035784959793, + 0.028061453253030777, + 0.009006192907691002, + 0.016154613345861435, + 0.005125908181071281, + 0.006510215811431408, + 0.0090824319049716, + 0.004588488955050707, + 0.003908294718712568, + 0.007787084672600031, + 0.0025022528134286404, + 0.006374770775437355, + 0.0035702651366591454, + 0.002508266596123576, + 0.008388490416109562, + 0.0024174386635422707, + 0.002225148258730769, + 0.0035133406054228544, + 0.0014020593371242285, + 0.00297437678091228, + 0.001183806569315493, + 0.0016288997139781713, + 0.001911868341267109, + 0.001471850322559476, + 0.0007742980378679931, + 0.0017193341627717018, + 0.0008506457670591772, + 0.0017873786855489016, + 0.008862625807523727, + 0.016866538673639297, + 0.010523179545998573, + 0.007961349561810493, + 0.006537712644785643, + 0.014071116223931313, + 0.004764033481478691, + 0.0100617790594697, + 0.0028032937552779913, + 0.0025741858407855034, + 0.003883209079504013, + 0.0018865113379433751, + 0.0021503802854567766, + 0.0030785142444074154, + 0.0018000672571361065, + 0.0022935743909329176, + 0.15098540484905243, + 0.011834587901830673, + 0.19723659753799438, + 0.021100947633385658, + 0.011132427491247654, + 0.029688067734241486, + 0.022555414587259293, + 0.01743767410516739, + 0.004008683376014233, + 0.005045109894126654, + 0.0074843429028987885, + 0.0032522273249924183, + 0.0026582665741443634, + 0.006178470328450203, + 0.002511140890419483, + 0.002994408831000328, + 0.004453578498214483, + 0.006449881009757519, + 0.004760180599987507, + 0.006229661405086517, + 0.004244994837790728, + 0.006489447318017483, + 0.005922600626945496, + 0.004805852193385363, + 0.013751289807260036, + 0.005247216206043959, + 0.005455310456454754, + 0.008029920980334282, + 0.0037559480406343937, + 0.019059346988797188, + 0.006211942061781883, + 0.006455353926867247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0052668387070298195 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525154173374176 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.00525154173374176, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525154173374176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15772561728954315, + 0.03424903005361557, + 0.013557675294578075, + 0.04582921043038368, + 0.048286810517311096, + 0.07096170634031296, + 0.30801787972450256, + 0.10862892121076584, + 0.15313473343849182, + 0.05960836261510849 + ], + "credit_map": [ + 0.03229943662881851, + 0.04166950285434723, + 0.026794785633683205, + 0.010186633095145226, + 0.011219070293009281, + 0.019141841679811478, + 0.007186672184616327, + 0.009227671660482883, + 0.004061519633978605, + 0.003831384005025029, + 0.004223927389830351, + 0.004087001085281372, + 0.004718829412013292, + 0.005148814991116524, + 0.004198880400508642, + 0.003978672903031111, + 0.0012440342688933015, + 0.0013335745315998793, + 0.002526237163692713, + 0.0011773736914619803, + 0.0019490880658850074, + 0.0018619056791067123, + 0.0017800630303099751, + 0.0016853992128744721, + 0.004149086773395538, + 0.0057706148363649845, + 0.008862589485943317, + 0.004712902940809727, + 0.007109303493052721, + 0.005370441824197769, + 0.0057937284000217915, + 0.004060541279613972, + 0.0046525513753294945, + 0.0055893585085868835, + 0.0076430970802903175, + 0.004473716951906681, + 0.008525480516254902, + 0.006063646171241999, + 0.004776532296091318, + 0.006562425289303064, + 0.00871800072491169, + 0.013773124665021896, + 0.010000880807638168, + 0.00906047597527504, + 0.005444695707410574, + 0.009362319484353065, + 0.006684473250061274, + 0.007917733862996101, + 0.10309469699859619, + 0.00991266593337059, + 0.1203046590089798, + 0.013593276962637901, + 0.008722178637981415, + 0.024034902453422546, + 0.015964072197675705, + 0.01239144615828991, + 0.011504505760967731, + 0.015790803357958794, + 0.013941292650997639, + 0.01055549643933773, + 0.011369794607162476, + 0.01799454167485237, + 0.0173416119068861, + 0.010130873881280422, + 0.015211636200547218, + 0.024083705618977547, + 0.01568460278213024, + 0.020781077444553375, + 0.013992935419082642, + 0.023254534229636192, + 0.02377905324101448, + 0.01634717546403408, + 0.0072137219831347466, + 0.006724865175783634, + 0.006381335202604532, + 0.009453284554183483, + 0.007782112341374159, + 0.006820403039455414, + 0.008847556076943874, + 0.006385085638612509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.00525154173374176 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001718117156997323 + ], + "opposite_logprob_tokens": [ + -6.376718044281006 + ], + "logprob_reward": -0.001718117156997323, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001718117156997323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18312418460845947, + 0.07190748304128647, + 0.07003915309906006, + 0.06731434166431427, + 0.05237439647316933, + 0.07835172861814499, + 0.20752310752868652, + 0.12005175650119781, + 0.07301906496286392, + 0.07629471272230148 + ], + "credit_map": [ + 0.02643934078514576, + 0.03564337268471718, + 0.04180578887462616, + 0.015502376481890678, + 0.011859163641929626, + 0.025340106338262558, + 0.00863382127135992, + 0.01790023036301136, + 0.007220687810331583, + 0.006317307241261005, + 0.018190834671258926, + 0.007718121167272329, + 0.005957580637186766, + 0.012854735367000103, + 0.004753468558192253, + 0.008894752711057663, + 0.009850911796092987, + 0.007223812863230705, + 0.020212067291140556, + 0.007075674831867218, + 0.004937044810503721, + 0.009435189887881279, + 0.0037293063942342997, + 0.0075751603581011295, + 0.006296782288700342, + 0.00854523479938507, + 0.012526359409093857, + 0.0059502883814275265, + 0.009162555448710918, + 0.009658431634306908, + 0.007106807082891464, + 0.008067883551120758, + 0.005418954882770777, + 0.00579212699085474, + 0.013085566461086273, + 0.005637948866933584, + 0.007052846252918243, + 0.006498793140053749, + 0.004629933275282383, + 0.004258232191205025, + 0.01137056015431881, + 0.018601179122924805, + 0.011329316534101963, + 0.008589126169681549, + 0.005605221260339022, + 0.009185326285660267, + 0.006207916419953108, + 0.0074630966410040855, + 0.05889975652098656, + 0.010833869688212872, + 0.0700807049870491, + 0.013685871846973896, + 0.009857626631855965, + 0.01853288896381855, + 0.01193277072161436, + 0.01369964424520731, + 0.014443952590227127, + 0.012448517605662346, + 0.01575452648103237, + 0.010517468675971031, + 0.011761585250496864, + 0.0237352903932333, + 0.02042040228843689, + 0.010970025323331356, + 0.00849162321537733, + 0.008455408737063408, + 0.01079462468624115, + 0.00772169278934598, + 0.0076156784780323505, + 0.01150503195822239, + 0.009583803825080395, + 0.008851203136146069, + 0.00894160196185112, + 0.014442851766943932, + 0.009926375932991505, + 0.010399983264505863, + 0.00574122229591012, + 0.010324647650122643, + 0.009448696859180927, + 0.007069343701004982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.001718117156997323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic close-up, static and precisely framed with a very shallow depth of field, captures a matte black metal candle snuff bottle standing upright on the rich, darkly textured surface of an old wooden table. High-contrast lighting defines the scene: a single, intense light source carves the snuff bottle's elegant form from the side, casting long, sharp shadows and creating stark highlights along one side of its slender handle and the curve of its bell-shaped head. The background is a complete, soft-focus blur, suggesting a dim, intimate space. The initial tranquility is shattered as a subtle vibration travels through the wooden surface, causing the snuff bottle to teeter for a brief, tense moment. It sways unsteadily, the handle wobbling slightly, before it decisively loses its center of gravity. The object then tilts, its initial descent slowly followed by a rapid acceleration as it spins around the edge of its base, finally colliding with the wood with a soft, gentle clink and coming to rest against its side. Its dark, elegant form now becomes a horizontal element in the composition, creating an atmosphere of quiet, contemplative, and visually appealing beauty.", + "seed": 21850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_8/index_7251_seed_21850_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016304513602517546, + "vlm_logprob": -0.016304513602517546, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.016304513602517546, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04789246618747711 + ], + "opposite_logprob_tokens": [ + -3.4228925704956055 + ], + "logprob_reward": -0.04789246618747711, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04789246618747711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2597123384475708, + 0.07705337554216385, + 0.05526287481188774, + 0.1619298905134201, + 0.033511701971292496, + 0.0673687756061554, + 0.1120975986123085, + 0.11580284684896469, + 0.04915628954768181, + 0.0681043341755867 + ], + "credit_map": [ + 0.033693574368953705, + 0.024936681613326073, + 0.06532116234302521, + 0.030798371881246567, + 0.024715842679142952, + 0.03488851338624954, + 0.013996554538607597, + 0.03136162832379341, + 0.008187095634639263, + 0.0077415574342012405, + 0.0188522320240736, + 0.009365594014525414, + 0.009487581439316273, + 0.009687989950180054, + 0.004618840292096138, + 0.009112484753131866, + 0.007921533659100533, + 0.00475431140512228, + 0.013734930194914341, + 0.006008585449308157, + 0.005387033335864544, + 0.008252616040408611, + 0.0026247166097164154, + 0.006579146720468998, + 0.01598479226231575, + 0.016153456643223763, + 0.043466243892908096, + 0.01928837038576603, + 0.01365193072706461, + 0.021657302975654602, + 0.008970266208052635, + 0.022757520899176598, + 0.0038398378528654575, + 0.003264036728069186, + 0.007079041097313166, + 0.0047327205538749695, + 0.003570714732632041, + 0.004467186983674765, + 0.002063872292637825, + 0.004494293127208948, + 0.008317170664668083, + 0.005988758523017168, + 0.012954170815646648, + 0.01139198336750269, + 0.006493819411844015, + 0.009543350897729397, + 0.0041466704569756985, + 0.008532846346497536, + 0.012791297398507595, + 0.015146365389227867, + 0.01965956948697567, + 0.01576818712055683, + 0.01411085668951273, + 0.01348259299993515, + 0.005349811166524887, + 0.015788910910487175, + 0.009785309433937073, + 0.01809314824640751, + 0.02173144929111004, + 0.009543106891214848, + 0.01768142729997635, + 0.01782519929111004, + 0.010632658377289772, + 0.010510549880564213, + 0.0068268002942204475, + 0.013650326058268547, + 0.00645360117778182, + 0.006835630163550377, + 0.003527740715071559, + 0.004666450433433056, + 0.003313555382192135, + 0.0038821864873170853, + 0.0066128382459282875, + 0.01343353372067213, + 0.009669971652328968, + 0.010977588593959808, + 0.006801285780966282, + 0.008275161497294903, + 0.005536460783332586, + 0.006797486450523138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.04789246618747711 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028410095255821943 + ], + "opposite_logprob_tokens": [ + -5.877840995788574 + ], + "logprob_reward": -0.0028410095255821943, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028410095255821943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16481417417526245, + 0.07197275757789612, + 0.0725431963801384, + 0.0866454690694809, + 0.03748751059174538, + 0.062371786683797836, + 0.07744567841291428, + 0.20203320682048798, + 0.10114080458879471, + 0.12354541569948196 + ], + "credit_map": [ + 0.021186502650380135, + 0.014509972184896469, + 0.03812847286462784, + 0.025368476286530495, + 0.0171223022043705, + 0.0223193671554327, + 0.010875563137233257, + 0.015303513966500759, + 0.008766567334532738, + 0.007540359627455473, + 0.015178967267274857, + 0.010198624804615974, + 0.006009479518979788, + 0.00879905465990305, + 0.006876661907881498, + 0.008603041991591454, + 0.013629820197820663, + 0.006525896023958921, + 0.016491366550326347, + 0.00584935350343585, + 0.004902003798633814, + 0.013594384305179119, + 0.0028252818156033754, + 0.008725089952349663, + 0.00989210419356823, + 0.008712136186659336, + 0.023334825411438942, + 0.00977884791791439, + 0.006459158379584551, + 0.013003775849938393, + 0.0044913021847605705, + 0.010973319411277771, + 0.0032301708124578, + 0.004540337715297937, + 0.007366781588643789, + 0.004745216108858585, + 0.006322398781776428, + 0.005042423028498888, + 0.002515039872378111, + 0.0037251408211886883, + 0.007789230439811945, + 0.005710918921977282, + 0.011088733561336994, + 0.007788558024913073, + 0.007816299796104431, + 0.008467324078083038, + 0.004350298549979925, + 0.009360420517623425, + 0.011763816699385643, + 0.0077748061157763, + 0.012854528613388538, + 0.010099610313773155, + 0.009500596672296524, + 0.011185476556420326, + 0.004687350243330002, + 0.009579486213624477, + 0.022752029821276665, + 0.025147346779704094, + 0.04591481760144234, + 0.019701620563864708, + 0.025662284344434738, + 0.026894301176071167, + 0.01729842834174633, + 0.01866237260401249, + 0.009202485904097557, + 0.022972621023654938, + 0.013138091191649437, + 0.011228580959141254, + 0.009758805856108665, + 0.014710983261466026, + 0.011686643585562706, + 0.008442592807114124, + 0.01135861873626709, + 0.01879054494202137, + 0.015386978164315224, + 0.012912631034851074, + 0.01572369784116745, + 0.016179116442799568, + 0.013812802731990814, + 0.019381022080779076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0028410095255821943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025022882036864758 + ], + "opposite_logprob_tokens": [ + -6.00250244140625 + ], + "logprob_reward": -0.0025022882036864758, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025022882036864758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1745261549949646, + 0.13946494460105896, + 0.12303003668785095, + 0.0963050052523613, + 0.04347113147377968, + 0.05145462602376938, + 0.042614005506038666, + 0.12364888191223145, + 0.07793597877025604, + 0.1275492012500763 + ], + "credit_map": [ + 0.029208825901150703, + 0.012345272116363049, + 0.042375463992357254, + 0.02055555209517479, + 0.012048191390931606, + 0.021849416196346283, + 0.010643990710377693, + 0.025499453768134117, + 0.01924079842865467, + 0.012052629142999649, + 0.03432367369532585, + 0.01997998170554638, + 0.012441027909517288, + 0.01740807294845581, + 0.00806726049631834, + 0.01595151610672474, + 0.01679525151848793, + 0.011448713950812817, + 0.03005889430642128, + 0.015749432146549225, + 0.010946999303996563, + 0.01623782515525818, + 0.007103706244379282, + 0.014689215458929539, + 0.01126299612224102, + 0.00915378239005804, + 0.02485153265297413, + 0.012383636087179184, + 0.007464277092367411, + 0.012840240262448788, + 0.006269171368330717, + 0.012079380452632904, + 0.005341504234820604, + 0.004183247219771147, + 0.009247641079127789, + 0.004918667953461409, + 0.004742578137665987, + 0.005895094946026802, + 0.0025357594713568687, + 0.006606640759855509, + 0.009170173667371273, + 0.004171621520072222, + 0.011384365148842335, + 0.007547819521278143, + 0.004366020206362009, + 0.007338204421103001, + 0.0028940944466739893, + 0.0045823282562196255, + 0.004679273813962936, + 0.00494733639061451, + 0.00663862656801939, + 0.00670319888740778, + 0.00761444354429841, + 0.004422497469931841, + 0.0024800654500722885, + 0.005128566175699234, + 0.012380472384393215, + 0.01992015168070793, + 0.02216501720249653, + 0.011882198974490166, + 0.015418832190334797, + 0.01939937099814415, + 0.011312716640532017, + 0.011170122772455215, + 0.011113394983112812, + 0.014166178181767464, + 0.017074689269065857, + 0.010130173526704311, + 0.0051397355273365974, + 0.006039595231413841, + 0.006775441113859415, + 0.007496768608689308, + 0.012210226617753506, + 0.024549756199121475, + 0.018698912113904953, + 0.015041696839034557, + 0.014482945203781128, + 0.019677167758345604, + 0.011570604518055916, + 0.011317887343466282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0025022882036864758 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342246681451797 + ], + "opposite_logprob_tokens": [ + -3.6513421535491943 + ], + "logprob_reward": -0.026342246681451797, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342246681451797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08947259187698364, + 0.07067565619945526, + 0.03733699023723602, + 0.11801257729530334, + 0.06128271296620369, + 0.030895086005330086, + 0.08210521191358566, + 0.25753650069236755, + 0.08459827303886414, + 0.16808439791202545 + ], + "credit_map": [ + 0.014737356454133987, + 0.008005150593817234, + 0.02069028466939926, + 0.009874537587165833, + 0.00694771483540535, + 0.0131039684638381, + 0.004711632616817951, + 0.011401944793760777, + 0.008873293176293373, + 0.006963262800127268, + 0.01609906740486622, + 0.009852687828242779, + 0.005648106336593628, + 0.009654452092945576, + 0.005337803158909082, + 0.008246984332799911, + 0.006127930246293545, + 0.0038277828134596348, + 0.009319665841758251, + 0.003670484758913517, + 0.003290719585493207, + 0.005500723607838154, + 0.0016787906643003225, + 0.003920892719179392, + 0.012899973429739475, + 0.011859183199703693, + 0.03218322619795799, + 0.013571787625551224, + 0.009889745153486729, + 0.015811026096343994, + 0.005709278862923384, + 0.016088364645838737, + 0.006207749247550964, + 0.006865880452096462, + 0.0153387151658535, + 0.008717124350368977, + 0.006851148325949907, + 0.0072576808743178844, + 0.004221300128847361, + 0.005823113024234772, + 0.00391597393900156, + 0.0028035060968250036, + 0.006214828696101904, + 0.00399215379729867, + 0.0032639142591506243, + 0.004112128634005785, + 0.002375963842496276, + 0.004216618370264769, + 0.011240445077419281, + 0.010650129988789558, + 0.0094349579885602, + 0.010499982163310051, + 0.014620587229728699, + 0.011083781719207764, + 0.005469358526170254, + 0.009105972945690155, + 0.025415925309062004, + 0.04474601522088051, + 0.03859651833772659, + 0.025189140811562538, + 0.03143618255853653, + 0.04828086867928505, + 0.019985029473900795, + 0.023886818438768387, + 0.009547636844217777, + 0.025107860565185547, + 0.008894339203834534, + 0.007439211010932922, + 0.006682185456156731, + 0.010704569518566132, + 0.007355306763201952, + 0.008867164142429829, + 0.015781335532665253, + 0.02965405397117138, + 0.016733922064304352, + 0.01772572100162506, + 0.020015772432088852, + 0.023661738261580467, + 0.017391838133335114, + 0.02712000533938408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.026342246681451797 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019445574143901467 + ], + "opposite_logprob_tokens": [ + -6.251944541931152 + ], + "logprob_reward": -0.0019445574143901467, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019445574143901467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11453896760940552, + 0.1301881968975067, + 0.06217564642429352, + 0.2883515954017639, + 0.06454403698444366, + 0.030978823080658913, + 0.06761211156845093, + 0.14966776967048645, + 0.044890038669109344, + 0.047052763402462006 + ], + "credit_map": [ + 0.016476241871714592, + 0.011606503278017044, + 0.02833220362663269, + 0.013084997422993183, + 0.008931530639529228, + 0.01760670728981495, + 0.007242060266435146, + 0.011258719488978386, + 0.01349753700196743, + 0.013413234613835812, + 0.03519397974014282, + 0.016917934641242027, + 0.011975493282079697, + 0.018417976796627045, + 0.007133795414119959, + 0.013638251461088657, + 0.010051976889371872, + 0.0053160120733082294, + 0.018303215503692627, + 0.006334327626973391, + 0.004281763453036547, + 0.009653553366661072, + 0.002252829959616065, + 0.005981965456157923, + 0.032988257706165314, + 0.027577275410294533, + 0.07743634283542633, + 0.032648518681526184, + 0.02224024012684822, + 0.041303981095552444, + 0.014267409220337868, + 0.03988957032561302, + 0.01089411973953247, + 0.00425285380333662, + 0.01809338666498661, + 0.008158302865922451, + 0.004251185804605484, + 0.008436959236860275, + 0.003255160991102457, + 0.007202066946774721, + 0.004116623662412167, + 0.003223692299798131, + 0.005865590646862984, + 0.0041379742324352264, + 0.0033057876862585545, + 0.003704305039718747, + 0.002111081499606371, + 0.004513769410550594, + 0.011416268534958363, + 0.006893726997077465, + 0.010966512374579906, + 0.008864893577992916, + 0.00805509090423584, + 0.008867698721587658, + 0.002970677800476551, + 0.009577246382832527, + 0.019190818071365356, + 0.01822863146662712, + 0.03730495646595955, + 0.01373848132789135, + 0.01704380102455616, + 0.019908804446458817, + 0.012103567831218243, + 0.012148717418313026, + 0.0040129404515028, + 0.007566908374428749, + 0.006815923843532801, + 0.0054497187957167625, + 0.0050469269044697285, + 0.005525762215256691, + 0.00550099415704608, + 0.004970860667526722, + 0.004963808227330446, + 0.007150397635996342, + 0.010164385661482811, + 0.006059307139366865, + 0.0037006677594035864, + 0.0061266301199793816, + 0.00361230387352407, + 0.005275262054055929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.0019445574143901467 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic close-up, static and precisely framed with a very shallow depth of field, captures a matte black metal candle snuff bottle standing upright on the rich, darkly textured surface of an old wooden table. High-contrast lighting defines the scene: a single, intense light source carves the snuff bottle's elegant form from the side, casting long, sharp shadows and creating stark highlights along one side of its slender handle and the curve of its bell-shaped head. The background is a complete, soft-focus blur, suggesting a dim, intimate space. The initial tranquility is shattered as a subtle vibration travels through the wooden surface, causing the snuff bottle to teeter for a brief, tense moment. It sways unsteadily, the handle wobbling slightly, before it decisively loses its center of gravity. The object then tilts, its initial descent slowly followed by a rapid acceleration as it spins around the edge of its base, finally colliding with the wood with a soft, gentle clink and coming to rest against its side. Its dark, elegant form now becomes a horizontal element in the composition, creating an atmosphere of quiet, contemplative, and visually appealing beauty.", + "seed": 21851, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_8/index_7251_seed_21851_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.009625114500522614, + "vlm_logprob": -0.009625114500522614, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.009625114500522614, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029220880940556526 + ], + "opposite_logprob_tokens": [ + -3.9042208194732666 + ], + "logprob_reward": -0.029220880940556526, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029220880940556526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036777056753635406, + 0.016992876306176186, + 0.015510587953031063, + 0.01685085892677307, + 0.12044111639261246, + 0.03393057733774185, + 0.1066184937953949, + 0.3833831250667572, + 0.2521345019340515, + 0.017360784113407135 + ], + "credit_map": [ + 0.005803741980344057, + 0.002668790752068162, + 0.0031102197244763374, + 0.0035256464034318924, + 0.005164851434528828, + 0.008240314200520515, + 0.004575380589812994, + 0.003688110038638115, + 0.0034768334589898586, + 0.0011016455246135592, + 0.0021492678206413984, + 0.0013669459149241447, + 0.0018106324132531881, + 0.0034613420721143484, + 0.0014837790513411164, + 0.0021424288861453533, + 0.0020830014254897833, + 0.0013701077550649643, + 0.001745383720844984, + 0.0018284565303474665, + 0.0028781150467693806, + 0.002265084069222212, + 0.0016607679426670074, + 0.0016796704148873687, + 0.0017567130271345377, + 0.0013256656238809228, + 0.0016117477789521217, + 0.002560880733653903, + 0.002708414103835821, + 0.0027969207148998976, + 0.0021382919512689114, + 0.0019522251095622778, + 0.014747900888323784, + 0.016148515045642853, + 0.008776861242949963, + 0.01341991126537323, + 0.03048093244433403, + 0.01778695173561573, + 0.011676331050693989, + 0.0074037169106304646, + 0.003151693381369114, + 0.0035437862388789654, + 0.004542575683444738, + 0.006542631890624762, + 0.004138729069381952, + 0.004850659519433975, + 0.004302992019802332, + 0.002857510233297944, + 0.006816082168370485, + 0.026480428874492645, + 0.007945121265947819, + 0.006159643642604351, + 0.009121838957071304, + 0.03349948674440384, + 0.007961755618453026, + 0.008634142577648163, + 0.017561793327331543, + 0.1527554839849472, + 0.014613388106226921, + 0.007530710659921169, + 0.015111718326807022, + 0.15243855118751526, + 0.009464447386562824, + 0.013907033018767834, + 0.013275439850986004, + 0.08713043481111526, + 0.00997726060450077, + 0.013280685059726238, + 0.09068765491247177, + 0.01618349365890026, + 0.010903152637183666, + 0.010696390643715858, + 0.0014693611301481724, + 0.004181501921266317, + 0.0020775930024683475, + 0.001918792026117444, + 0.0021758878137916327, + 0.0025913952849805355, + 0.0015544279012829065, + 0.0013918248005211353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.029220880940556526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019583588000386953 + ], + "opposite_logprob_tokens": [ + -6.25195837020874 + ], + "logprob_reward": -0.0019583588000386953, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019583588000386953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0741574615240097, + 0.022306067869067192, + 0.01922166533768177, + 0.02622378058731556, + 0.07458080351352692, + 0.26263245940208435, + 0.05704566836357117, + 0.40281206369400024, + 0.04026477783918381, + 0.020755236968398094 + ], + "credit_map": [ + 0.014001131989061832, + 0.005640039686113596, + 0.007794415578246117, + 0.005815931595861912, + 0.008605079725384712, + 0.012262660078704357, + 0.0078629394993186, + 0.012175261043012142, + 0.003888731822371483, + 0.0021602511405944824, + 0.002702222904190421, + 0.002242560498416424, + 0.0023062657564878464, + 0.003384245093911886, + 0.002216047141700983, + 0.003405740251764655, + 0.002661788370460272, + 0.001591361709870398, + 0.0018503222381696105, + 0.0022277419921010733, + 0.0034042787738144398, + 0.002433067886158824, + 0.0024953996762633324, + 0.0025577035266906023, + 0.003976667765527964, + 0.0016307965852320194, + 0.0025575291365385056, + 0.0032922900281846523, + 0.0036020835395902395, + 0.005105778109282255, + 0.002943377709016204, + 0.003115257481113076, + 0.00833052210509777, + 0.008391028270125389, + 0.006304827053099871, + 0.00893658958375454, + 0.01778983324766159, + 0.009794012643396854, + 0.00905398279428482, + 0.005980000831186771, + 0.012854776345193386, + 0.09088771790266037, + 0.013584788888692856, + 0.012079551815986633, + 0.016237478703260422, + 0.09436462819576263, + 0.011255336925387383, + 0.011368184350430965, + 0.003961209207773209, + 0.013125595636665821, + 0.0036923647858202457, + 0.00412321463227272, + 0.005496235564351082, + 0.016860011965036392, + 0.004037862177938223, + 0.005749166943132877, + 0.017561160027980804, + 0.16408580541610718, + 0.015914052724838257, + 0.009210356511175632, + 0.014144563116133213, + 0.15931908786296844, + 0.010663961991667747, + 0.011913090012967587, + 0.00403522327542305, + 0.005371909122914076, + 0.0034355048555880785, + 0.007810376118868589, + 0.0039736186154186726, + 0.004596760030835867, + 0.004856698680669069, + 0.006184684578329325, + 0.001954619539901614, + 0.0036491728387773037, + 0.0020428451243788004, + 0.0029902420938014984, + 0.002504640957340598, + 0.0033708035480231047, + 0.001488043344579637, + 0.002754869172349572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.0019583588000386953 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022072256542742252 + ], + "opposite_logprob_tokens": [ + -6.127207279205322 + ], + "logprob_reward": -0.0022072256542742252, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022072256542742252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14773768186569214, + 0.06173158064484596, + 0.027229243889451027, + 0.023115020245313644, + 0.056281305849552155, + 0.17879942059516907, + 0.2550846338272095, + 0.06938353925943375, + 0.11516444385051727, + 0.06547313928604126 + ], + "credit_map": [ + 0.024839797988533974, + 0.010833349078893661, + 0.02245396375656128, + 0.011656340211629868, + 0.01668459363281727, + 0.019138265401124954, + 0.013648606836795807, + 0.028482777997851372, + 0.010480892844498158, + 0.00474570645019412, + 0.009356172755360603, + 0.00556586217135191, + 0.0073915645480155945, + 0.007473365403711796, + 0.005576325114816427, + 0.01114169042557478, + 0.004374144598841667, + 0.0019033695571124554, + 0.003357547800987959, + 0.001983286114409566, + 0.003736478742212057, + 0.003971082158386707, + 0.0030655518639832735, + 0.004837781190872192, + 0.003456901293247938, + 0.001411782461218536, + 0.002981778932735324, + 0.0019013541750609875, + 0.003054107539355755, + 0.004406833089888096, + 0.002167673548683524, + 0.0037345876917243004, + 0.00812402181327343, + 0.008235828951001167, + 0.006602519191801548, + 0.006242631934583187, + 0.009895161725580692, + 0.007212942466139793, + 0.005471718031913042, + 0.004496482200920582, + 0.00842610839754343, + 0.0555318258702755, + 0.012032239697873592, + 0.010375246405601501, + 0.010925361886620522, + 0.0578044094145298, + 0.010220118798315525, + 0.01348409429192543, + 0.01589931733906269, + 0.03923838958144188, + 0.021979307755827904, + 0.016154684126377106, + 0.06889095157384872, + 0.0555681549012661, + 0.016043292358517647, + 0.021310538053512573, + 0.007404739502817392, + 0.007596075534820557, + 0.007445866242051125, + 0.007429970894008875, + 0.012158073484897614, + 0.010631334036588669, + 0.008516896516084671, + 0.008200577460229397, + 0.006294636987149715, + 0.03602795675396919, + 0.004663646221160889, + 0.004797263536602259, + 0.044072963297367096, + 0.011194299906492233, + 0.0047675129026174545, + 0.0033461591228842735, + 0.005039147101342678, + 0.017731962725520134, + 0.009504420682787895, + 0.005294318310916424, + 0.006157251540571451, + 0.011861992068588734, + 0.004042058251798153, + 0.005841988138854504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0022072256542742252 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537471950054169 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537471950054169, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537471950054169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07664797455072403, + 0.017644958570599556, + 0.01788797602057457, + 0.017585182562470436, + 0.03604290261864662, + 0.269782692193985, + 0.19365467131137848, + 0.06834650784730911, + 0.282857745885849, + 0.019549448043107986 + ], + "credit_map": [ + 0.016302848234772682, + 0.005884063895791769, + 0.008809265680611134, + 0.006001375149935484, + 0.010428795590996742, + 0.012928644195199013, + 0.007720306981354952, + 0.008572671562433243, + 0.0036869279574602842, + 0.0013837870210409164, + 0.0024584170896559954, + 0.0011475980281829834, + 0.0022351196967065334, + 0.0027396613731980324, + 0.0014862677780911326, + 0.0025071771815419197, + 0.00246982229873538, + 0.0015682432567700744, + 0.0022531715221703053, + 0.0025215446949005127, + 0.0027446742169559, + 0.0023343991488218307, + 0.0023933073971420527, + 0.0016028127865865827, + 0.001764641492627561, + 0.0011408058926463127, + 0.0018468055641278625, + 0.002472877036780119, + 0.0028731098864227533, + 0.002680230652913451, + 0.0026559375692158937, + 0.0021507747005671263, + 0.0043576802127063274, + 0.0033562735188752413, + 0.004635656252503395, + 0.003913503140211105, + 0.006831408478319645, + 0.004417526535689831, + 0.004957819357514381, + 0.0035730365198105574, + 0.013879773207008839, + 0.09165540337562561, + 0.013741045258939266, + 0.013023600913584232, + 0.016787752509117126, + 0.09414123743772507, + 0.013554321601986885, + 0.012999558821320534, + 0.01035236194729805, + 0.07133923470973969, + 0.009403644129633904, + 0.00647825188934803, + 0.013952960260212421, + 0.05893271043896675, + 0.009679649956524372, + 0.013515875674784184, + 0.0028470612596720457, + 0.005262595135718584, + 0.004696137737482786, + 0.0037037087604403496, + 0.020857689902186394, + 0.02249339036643505, + 0.003925404977053404, + 0.00456051854416728, + 0.014377394691109657, + 0.10114846378564835, + 0.01698887348175049, + 0.008989808149635792, + 0.09898021072149277, + 0.0192173570394516, + 0.011657297611236572, + 0.011498348787426949, + 0.0013642909470945597, + 0.0050374604761600494, + 0.0015233464073389769, + 0.002274908823892474, + 0.002418352058157325, + 0.002666993299499154, + 0.0022831903770565987, + 0.0019809056539088488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.012537471950054169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201635157689452 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201635157689452, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201635157689452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06522448360919952, + 0.018540706485509872, + 0.013775093480944633, + 0.009770927019417286, + 0.026776950806379318, + 0.07030628621578217, + 0.20619511604309082, + 0.3895460367202759, + 0.19103005528450012, + 0.008834341540932655 + ], + "credit_map": [ + 0.013876456767320633, + 0.0037589664570987225, + 0.010871028527617455, + 0.0036282383371144533, + 0.005191714037209749, + 0.012657782062888145, + 0.004650555085390806, + 0.010589740239083767, + 0.0034603909589350224, + 0.0013826035428792238, + 0.002806756878271699, + 0.0013607583241537213, + 0.001776225632056594, + 0.0033781384117901325, + 0.0016460760962218046, + 0.0027297574561089277, + 0.002242062008008361, + 0.0008334342855960131, + 0.002369524445384741, + 0.0011197335552424192, + 0.0012595315929502249, + 0.002531594829633832, + 0.001079380395822227, + 0.00233983201906085, + 0.0010612339247018099, + 0.0006309979362413287, + 0.0009107612422667444, + 0.0013570650480687618, + 0.0021532683167606592, + 0.001433500205166638, + 0.0013597566867247224, + 0.0008643438341096044, + 0.0032397129107266665, + 0.002921439940109849, + 0.003771768184378743, + 0.0028839365113526583, + 0.004314677324146032, + 0.0033428953029215336, + 0.0034935481380671263, + 0.002808971796184778, + 0.004535887390375137, + 0.01696164533495903, + 0.005224063526839018, + 0.007026524282991886, + 0.006866293493658304, + 0.019080180674791336, + 0.005824015010148287, + 0.004787680692970753, + 0.010514186695218086, + 0.07963576912879944, + 0.011877916753292084, + 0.008115618489682674, + 0.040438681840896606, + 0.03657100722193718, + 0.007607080973684788, + 0.011434860527515411, + 0.02065030299127102, + 0.16518864035606384, + 0.01456853374838829, + 0.007235880009829998, + 0.025012364611029625, + 0.13631214201450348, + 0.008692332543432713, + 0.011885851621627808, + 0.009555570781230927, + 0.06686560809612274, + 0.011678689159452915, + 0.007988108322024345, + 0.0657414123415947, + 0.012584008276462555, + 0.008434857241809368, + 0.008181816898286343, + 0.0011830907315015793, + 0.0013823341578245163, + 0.0013792620738968253, + 0.0010502533987164497, + 0.0006752791814506054, + 0.0010501311626285315, + 0.0009469061042182148, + 0.0011670844396576285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.002201635157689452 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic close-up, static and precisely framed with a very shallow depth of field, captures a matte black metal candle snuff bottle standing upright on the rich, darkly textured surface of an old wooden table. High-contrast lighting defines the scene: a single, intense light source carves the snuff bottle's elegant form from the side, casting long, sharp shadows and creating stark highlights along one side of its slender handle and the curve of its bell-shaped head. The background is a complete, soft-focus blur, suggesting a dim, intimate space. The initial tranquility is shattered as a subtle vibration travels through the wooden surface, causing the snuff bottle to teeter for a brief, tense moment. It sways unsteadily, the handle wobbling slightly, before it decisively loses its center of gravity. The object then tilts, its initial descent slowly followed by a rapid acceleration as it spins around the edge of its base, finally colliding with the wood with a soft, gentle clink and coming to rest against its side. Its dark, elegant form now becomes a horizontal element in the composition, creating an atmosphere of quiet, contemplative, and visually appealing beauty.", + "seed": 21852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_8/index_7251_seed_21852_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02133228280581534, + "vlm_logprob": -0.02133228280581534, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.02133228280581534, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06772643327713013 + ], + "opposite_logprob_tokens": [ + -3.3177263736724854 + ], + "logprob_reward": -0.06772643327713013, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06772643327713013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03974442929029465, + 0.014761888422071934, + 0.028781484812498093, + 0.02525804378092289, + 0.03591550141572952, + 0.05064870044589043, + 0.27676302194595337, + 0.39949432015419006, + 0.11051663011312485, + 0.018115954473614693 + ], + "credit_map": [ + 0.005749664735049009, + 0.002161942422389984, + 0.005738613661378622, + 0.003212038427591324, + 0.004303636495023966, + 0.012082387693226337, + 0.0032923491671681404, + 0.0032037978526204824, + 0.0017782115610316396, + 0.0012878125999122858, + 0.0024830421898514032, + 0.0017595132812857628, + 0.0018088468350470066, + 0.0022547950502485037, + 0.0016086004907265306, + 0.0017810658318921924, + 0.003533373586833477, + 0.003574544796720147, + 0.005280019249767065, + 0.004153957590460777, + 0.0025479274336248636, + 0.00374567206017673, + 0.0025763155426830053, + 0.003369675949215889, + 0.00406818138435483, + 0.0025051783304661512, + 0.008586039766669273, + 0.001840834622271359, + 0.0018919076537713408, + 0.0025637843646109104, + 0.0016785049811005592, + 0.0021236129105091095, + 0.0035973621997982264, + 0.010342931374907494, + 0.005365503020584583, + 0.002529581543058157, + 0.0033675283193588257, + 0.005351354833692312, + 0.0029020735528320074, + 0.002459164010360837, + 0.004889975301921368, + 0.006945404689759016, + 0.0069648087956011295, + 0.006808044854551554, + 0.004932706709951162, + 0.009727668948471546, + 0.004909136798232794, + 0.0054709515534341335, + 0.015919014811515808, + 0.11795878410339355, + 0.014387238770723343, + 0.005017872434109449, + 0.012801788747310638, + 0.10061544179916382, + 0.004100302699953318, + 0.005962559953331947, + 0.022627385333180428, + 0.1650126427412033, + 0.023420505225658417, + 0.010075204074382782, + 0.018209954723715782, + 0.14374414086341858, + 0.007132014725357294, + 0.009272498078644276, + 0.005375005304813385, + 0.04589538276195526, + 0.005516792647540569, + 0.0025653578341007233, + 0.0043522268533706665, + 0.04211550951004028, + 0.0018405790906399488, + 0.0028557812329381704, + 0.001960056833922863, + 0.002967383246868849, + 0.001964851515367627, + 0.0014689814997836947, + 0.0013381190365180373, + 0.005543834529817104, + 0.0012862575240433216, + 0.0015864706365391612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.06772643327713013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005285337567329407 + ], + "opposite_logprob_tokens": [ + -5.255285263061523 + ], + "logprob_reward": -0.005285337567329407, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005285337567329407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09118234366178513, + 0.02616305463016033, + 0.06536503881216049, + 0.029780671000480652, + 0.04594893380999565, + 0.10714206844568253, + 0.23263351619243622, + 0.3845658004283905, + 0.010650543496012688, + 0.006568024400621653 + ], + "credit_map": [ + 0.010691148228943348, + 0.007551864720880985, + 0.01606026478111744, + 0.009797601029276848, + 0.010840494185686111, + 0.01764548569917679, + 0.009513730183243752, + 0.009081761352717876, + 0.0036543025635182858, + 0.002272197278216481, + 0.0046583255752921104, + 0.003470153082162142, + 0.001888257684186101, + 0.004374958574771881, + 0.0022350975777953863, + 0.003609765786677599, + 0.00709237065166235, + 0.007704290095716715, + 0.016440462321043015, + 0.009417981840670109, + 0.006670104339718819, + 0.00691577373072505, + 0.004161882679909468, + 0.0069621773436665535, + 0.004210148938000202, + 0.004205992911010981, + 0.006450970191508532, + 0.00312118511646986, + 0.0031993461307138205, + 0.002978891134262085, + 0.0035226880572736263, + 0.0020914506167173386, + 0.006356661207973957, + 0.00635985704138875, + 0.01116284728050232, + 0.003901718882843852, + 0.003759498707950115, + 0.00703046377748251, + 0.003015973838046193, + 0.0043619186617434025, + 0.006846563890576363, + 0.036887552589178085, + 0.011508508585393429, + 0.007522799540311098, + 0.008526807650923729, + 0.02612374909222126, + 0.004940427374094725, + 0.0047856722958385944, + 0.01371185202151537, + 0.09667576849460602, + 0.013319917023181915, + 0.005298377014696598, + 0.010435115545988083, + 0.08310164511203766, + 0.004824659321457148, + 0.005266217980533838, + 0.02133679762482643, + 0.15890608727931976, + 0.022457050159573555, + 0.009985941462218761, + 0.017398864030838013, + 0.13874559104442596, + 0.006829431280493736, + 0.00890604592859745, + 0.0008520721457898617, + 0.0018100374145433307, + 0.0015892701921984553, + 0.0013307926710695028, + 0.0010482973884791136, + 0.0018601614283397794, + 0.001110177137888968, + 0.0010497354669496417, + 0.0007166115683503449, + 0.0009162813657894731, + 0.0010972936870530248, + 0.0007096022018231452, + 0.0007167025469243526, + 0.0007118401699699461, + 0.0007364178891293705, + 0.0009632761357352138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.005285337567329407 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004113546572625637 + ], + "opposite_logprob_tokens": [ + -5.504113674163818 + ], + "logprob_reward": -0.004113546572625637, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004113546572625637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23511309921741486, + 0.03780101239681244, + 0.022662311792373657, + 0.031057201325893402, + 0.03791915997862816, + 0.09233517944812775, + 0.3960280418395996, + 0.028281379491090775, + 0.03843565285205841, + 0.08036690205335617 + ], + "credit_map": [ + 0.035843707621097565, + 0.010431681759655476, + 0.04384748265147209, + 0.016978735104203224, + 0.020089607685804367, + 0.0650775283575058, + 0.02114972285926342, + 0.021694617345929146, + 0.005906124133616686, + 0.003427697578445077, + 0.003942099865525961, + 0.0036549000069499016, + 0.0024640713818371296, + 0.010531891137361526, + 0.002868971321731806, + 0.005005255341529846, + 0.0034194772597402334, + 0.002379365498200059, + 0.0038115184288471937, + 0.004300715867429972, + 0.0014980179257690907, + 0.0025261263363063335, + 0.0022024463396519423, + 0.0025246429722756147, + 0.0046618967317044735, + 0.0029988312162458897, + 0.008221405558288097, + 0.0030289662536233664, + 0.0024434179067611694, + 0.00352704175747931, + 0.002554833423346281, + 0.003620808245614171, + 0.0039003915153443813, + 0.005307621322572231, + 0.007110320497304201, + 0.004115943796932697, + 0.0035232158843427896, + 0.00496108690276742, + 0.003751298412680626, + 0.005249280482530594, + 0.011054999195039272, + 0.012833801098167896, + 0.016157742589712143, + 0.007423915434628725, + 0.009065440855920315, + 0.013857574202120304, + 0.006802621763199568, + 0.015139089897274971, + 0.021017184481024742, + 0.16359946131706238, + 0.023577546700835228, + 0.009238503873348236, + 0.017586607486009598, + 0.14381007850170135, + 0.007790569681674242, + 0.009408067911863327, + 0.0018566898070275784, + 0.0062338607385754585, + 0.002880728803575039, + 0.002569827949628234, + 0.0024854361545294523, + 0.007246304769068956, + 0.0026852053124457598, + 0.0023233259562402964, + 0.0015887526096776128, + 0.01511177234351635, + 0.0020331144332885742, + 0.0011115566594526172, + 0.0019518458284437656, + 0.014502760954201221, + 0.0008579430868849158, + 0.001277906820178032, + 0.004030690528452396, + 0.03232497349381447, + 0.003735183272510767, + 0.0022680489346385, + 0.003999494016170502, + 0.030131746083498, + 0.0018992810510098934, + 0.001977483509108424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.004113546572625637 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338299736380577 + ], + "opposite_logprob_tokens": [ + -3.6513383388519287 + ], + "logprob_reward": -0.026338299736380577, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338299736380577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04038265720009804, + 0.009164024144411087, + 0.015905341133475304, + 0.07020241022109985, + 0.03503105416893959, + 0.06245012581348419, + 0.4087373614311218, + 0.3109685778617859, + 0.03495064005255699, + 0.012207833118736744 + ], + "credit_map": [ + 0.004932676907628775, + 0.003345198230817914, + 0.005880878772586584, + 0.004657108336687088, + 0.006410911679267883, + 0.006688868626952171, + 0.004772783722728491, + 0.003694230690598488, + 0.0012857370311394334, + 0.0007993296603672206, + 0.0014583482407033443, + 0.0012410875642672181, + 0.00069890555460006, + 0.0019300282001495361, + 0.0006971512921154499, + 0.0010534367756918073, + 0.00208931602537632, + 0.002135158982127905, + 0.00178727472666651, + 0.0024050408974289894, + 0.001554817776195705, + 0.002082335064187646, + 0.0017156286630779505, + 0.002135767135769129, + 0.011484007351100445, + 0.007481350563466549, + 0.02368658408522606, + 0.00433133402839303, + 0.0059797256253659725, + 0.0075115323998034, + 0.006305105052888393, + 0.003422761568799615, + 0.003197631798684597, + 0.010449227876961231, + 0.005173444282263517, + 0.0029769649263471365, + 0.0028652858454734087, + 0.006099760066717863, + 0.0021573237609118223, + 0.002111415145918727, + 0.004764838609844446, + 0.017928633838891983, + 0.006950690411031246, + 0.005581749603152275, + 0.005128846503794193, + 0.01217945571988821, + 0.004102917388081551, + 0.0058129881508648396, + 0.022973770275712013, + 0.17119410634040833, + 0.023543838411569595, + 0.00860805343836546, + 0.01788083091378212, + 0.1482210010290146, + 0.007220651488751173, + 0.009095093235373497, + 0.016366755589842796, + 0.12499862164258957, + 0.020821819081902504, + 0.009166731499135494, + 0.014420436695218086, + 0.11044372618198395, + 0.006551969796419144, + 0.00819852203130722, + 0.0023700010497123003, + 0.013673939742147923, + 0.0025410931557416916, + 0.0011959249386563897, + 0.001780518563464284, + 0.010783549398183823, + 0.0012948481598868966, + 0.0013107657432556152, + 0.001001091324724257, + 0.001921736286021769, + 0.0019009840907528996, + 0.0011825037654489279, + 0.00140199507586658, + 0.0021440486889332533, + 0.0013548004208132625, + 0.0013006734661757946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.026338299736380577 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a matte black metal candle snuff bottle with a slender handle and bell-shaped head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle initially standing upright on a rich, darkly textured old wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single intense side light source creating high contrast, long sharp shadows, and stark highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bottle teeter, sway, and then tilt over to fall, spinning around its base before colliding with the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static close-up with a very shallow depth of field and a soft-focus blurred background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031977968756109476 + ], + "opposite_logprob_tokens": [ + -5.75319766998291 + ], + "logprob_reward": -0.0031977968756109476, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031977968756109476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08311574161052704, + 0.02836115099489689, + 0.03843989595770836, + 0.0443507581949234, + 0.07603956758975983, + 0.16699156165122986, + 0.1547948122024536, + 0.3564876914024353, + 0.03694543614983559, + 0.01447342149913311 + ], + "credit_map": [ + 0.011631133034825325, + 0.00623018154874444, + 0.016093507409095764, + 0.007094980217516422, + 0.006586045492440462, + 0.022176774218678474, + 0.0069613526575267315, + 0.006341759115457535, + 0.0033655930310487747, + 0.00161378039047122, + 0.005599734839051962, + 0.0025777919217944145, + 0.0019249080214649439, + 0.008764020167291164, + 0.00211352133192122, + 0.0024017994292080402, + 0.0049347057938575745, + 0.0044914186000823975, + 0.011891698464751244, + 0.00514466455206275, + 0.0031503408681601286, + 0.0032742854673415422, + 0.0022119602654129267, + 0.003340818453580141, + 0.007399731315672398, + 0.005123187322169542, + 0.015381266362965107, + 0.003192068776115775, + 0.003574277274310589, + 0.004569778684526682, + 0.0029516397044062614, + 0.002158805262297392, + 0.01099236961454153, + 0.013882570900022984, + 0.01774229295551777, + 0.004751807544380426, + 0.006902826018631458, + 0.011016531847417355, + 0.004376818425953388, + 0.006374354008585215, + 0.008316022343933582, + 0.06252779066562653, + 0.012147287838160992, + 0.0052256775088608265, + 0.008864997886121273, + 0.05722498521208763, + 0.006320239510387182, + 0.006364565342664719, + 0.009390573017299175, + 0.06469646841287613, + 0.009405733086168766, + 0.003913971595466137, + 0.007423889357596636, + 0.0535375252366066, + 0.002984435996040702, + 0.0034421992022544146, + 0.0200631245970726, + 0.14748777449131012, + 0.0208575539290905, + 0.00886212382465601, + 0.01574748195707798, + 0.1286337673664093, + 0.0066816722974181175, + 0.008154179900884628, + 0.0018314849585294724, + 0.014858261682093143, + 0.002010984579101205, + 0.0013202237896621227, + 0.0016362355090677738, + 0.013235275633633137, + 0.0008881448302417994, + 0.0011648223735392094, + 0.001931950100697577, + 0.004029067233204842, + 0.0016768366331234574, + 0.0012838731054216623, + 0.001426239381544292, + 0.001879130257293582, + 0.001081195310689509, + 0.001165128080174327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.0031977968756109476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file